An Intelligent Approach for Cooling Radiator Fault Diagnosis Based on Infrared Thermal Image Processing Amin Taheri-Garavanda,d*, Hojjat Ahmadia, Mahmoud Omida, Seyed Saeid Mohtasebia, Kaveh Mollazadeb, Alan John Russell Smithc, Giovanni Maria Carlomagnod a Department of Mechanical Engineering of Agricultural Machinery, University of Tehran, Karaj, Iran of Biosystems Engineering, University of Kurdistan, Sanandaj, Iran c Department of Mechanical Engineering, University of Melbourne, Melbourne, Australia dDepartment of Industrial Engineering, Naples University of Federico II, Naples, Italy b Department Abstract This research presents a new intelligent fault diagnosis and condition monitoring system for classification of different conditions of mechanical equipment that produces distinct thermal signatures for different fault conditions. This will be illustrated by considering the classification of six types of cooling radiator conditions: radiator tubes blockage, radiator fins blockage, loose connection between fins and tubes, radiator door failure, coolant leakage, and normal conditions. The proposed system consists of several distinct procedures including thermal image acquisition, image preprocessing, image processing, twodimensional discrete wavelet transform (2D-DWT), feature extraction, feature selection using a genetic algorithm (GA), and finally classification by artificial neural networks (ANNs). The 2D-DWT is implemented to decompose the thermal images. Subsequently, statistical texture features are extracted from the original images and are decomposed into thermal images. The significant selected features are used to enhance the performance of the designed ANN classifier for the 6 types of cooling radiator conditions (output layer) in the next stage. For the tested system, the input layer consisted of 16 neurons based on the feature selection operation. The best performance of ANN was obtained with a 16-6-6 topology. The classification results demonstrated that this system can be employed satisfactorily as an intelligent condition monitoring and fault diagnosis for a class of cooling radiator. Keywords: Cooling Radiator, Condition Monitoring, Thermal Images, Discrete Wavelet Transform, Genetic Algorithm, Artificial Neural Networks. Introduction The radiator is a key component of an engine’s cooling system, playing an important role in maintaining the operating temperature of the engine. The temperature of an internal combustion engine can reach 2700°C (combustion gases) when it is operating at full throttle. Most engine component materials are not be able to endure this temperature and would rapidly fail if they are not properly cooled. Overheating of the engine can cause oil to thin, engine parts to expand, lubrication to break down, and engine moving parts to be damaged. Therefore removing heat from an engine is indispensable for the appropriate operation of engine. Most of the heat is removed by convection [1, 2] to environmental air. The radiator is a kind of heat exchanger and important element in the cooling system of vehicles. Its main purpose is moving the excessive heat from the engine block to the surrounding air, which ensures reliable operation of the engine [3-5]. The importance of thermal studies of radiators arises principally from the acknowledged difficulty of detecting the root cause of crack-induced leakage and other types of failures in radiators [6]. Condition monitoring aims to prevent unplanned breakdowns, make the most of the plant availability and decrease associated hazards. There are some non-destructive techniques that are often used for condition monitoring such as vibration analysis, eddy-current testing, radiography, ultrasonic testing, and acoustic emission[7].Temperature is one of the most useful parameters that indicates the structural health of a machine. Hence, temperature monitoring of equipment or processes has been identified as one of the best predictive maintenance methodologies [8]. Infrared radiation is emitted from the surface of any physical object with temperature above absolute zero. The infrared (IR) energy is not visual since IR radiation is not in the visible range of the electromagnetic spectrum for human and regular cameras. Infrared thermography is a technique used for converting invisible heat energy into a visual thermal image that shows the thermal energy emitting from the object surface. Based on this trait, thermography is currently applied to machine condition monitoring and diagnosis fields where the temperature represents a key parameter [9]. IR thermography has been used for the nondestructive evaluation of joints[10].Kim et al.[11] used IR thermography for fault diagnosis of ball bearings when rotational machinery had foreign material inside the bearing sunder a dynamic loading condition. Lee and Kim [12] employed thermal imaging for the early detection and condition monitoring of the leakage from the closure plug of heavy water reactors during on-site inspections. They reported that the location of the leakages could be identified and the leak status could be monitored in real-time with IR thermography. Ge et al. [13] inspected the temperature distributions of air-cooled condensers and calculated the influence of ambient air temperature, natural air flow, and surface defects on the performance of the units by IR thermography. The thermography technique has evolved as a useful method for real-time temperature monitoring of machines or processes in a non-contact and non-intrusive way for various condition monitoring applications, which can decrease breakdowns or emergency shutdowns, maintenance costs and risk of accidents, augment the performance and increase productivity. By applying modern image processing methods to the acquired IR thermal images with artificial intelligence (AI) based approaches, better decisions may be made rapidly without human intervention [8]. Younus and Yang [14] presented an intelligent fault diagnosis system for classification of different rotary machine conditions that utilized the processing of IR thermal images. They used a two-dimensional discrete wavelet transformation (2D-DWT) to decompose the thermal image. In order to assist in diagnosing the different machine conditions, they utilized support vector machines (SVMs) and linear discriminant analysis (LDA) methods as classifiers. Artificial neural networks (ANNs) are robust, adaptive and strong numerical models for pattern recognition and classification [15]. ANNs are very powerful tools that can be trained to solve complex non-linear classification problems. Huda and Taib [9] applied IR thermography for predictive and preventive maintenance of thermal defects in electrical equipment. They utilized statistical features, a multilayer perceptron (MLP) neural network, and a discriminant analysis classifier to allocate the hotspot thermal status into ‘defect’ and ‘no defect’ categories. Abu-Mahfouz [16] used ANNs for the detection and classification of malfunction, wear and damage of a gearbox operating under steady state conditions. ANNs were applied for the damage indices classification of aerospace structures with the use of Lamb waves [17]. Fault diagnosis of machinery can be handled as a task of pattern recognition and classification that includes data acquisition, feature extraction, feature selection and final condition classification steps which are the demandable tasks of fault diagnosis. To obtain correct data (normal or abnormal), it is important to complete all steps of signal processing whatever the signal type such as vibration, thermal image, current, ultrasonic, or acoustic. Moreover, fault pattern classification from images typically consists of these steps: image acquisition, pre-processing, segmentation, feature extraction or dimension reduction, feature selection, classification and decision [18, 19].In order to use condition monitoring and fault detection of machines, signal processing techniques are initially required to process the data acquired from the machinery. Wavelet transform is an early technique that has been employed for one-dimensional signal processing. Recently, 2D-DWT is frequently considered as a decomposition algorithm in the image processing field. 2D-DWT is a tool that is applied for analysis of 2D signals such as X–ray images, magnetic resonance images (MRI), synthetic aperture radar (SAR), and RGB images [20].However, the data obtained from the decomposition process of a wavelet transform are seldom practical because of the huge dimensionality which causes difficulties of data storage and data mining for the next procedures. Representing data as features or dimensionality reductions is a process of extracting the functional information from the dataset to remove artifacts and reduce the dimensionality. However, it must protect the characteristic features which show faults and conditions of the machinery as far as possible. Dimensionality reduction is an important data preprocessing procedure for classification tasks and commonly falls into two categories: feature compression and feature selection [14]. The engine cooling system and the radiator, in particular, as its main component to maintain the temperature of engine are vitally important to the operation of an engine. A radiator that is defective will cause the engine to be stopped or it will reduce engine performance. Thus fault diagnosis and condition monitoring of a radiator is very important. In this paper, a thermography-based technique is considered for fault detection and condition monitoring of radiators because temperature is a key parameter in defining a radiator’s condition. Accordingly, the aim of this work is the development and implementation of a new intelligent fault diagnosis and condition monitoring system for classification of common faults occurring in a cooling radiator using IR thermal images. In the present study, the intelligent condition monitoring system has a number of procedures that must be applied sequentially, including: IR thermal image acquisition, preprocessing, image processing via2D-DWT, feature extraction, feature selection, and classification. The 2D-DWT is applied to thermal image decomposition. Consequently, statistical texture features are extracted from the original and decomposed thermal images. In the next step, the significant features are selected based on a genetic algorithm (GA) to enhance the performance of the ANN classifier in the final stage. Details of the methodology adopted for collecting the infrared images, and for analyzing these are provided in the next section. 2. Materials and Methods 2.1. Test setup and experimental procedure To simulate faults in the cooling radiator a test apparatus, as shown inFig.1, was prepared. The setup consists of a radiator, thermal infrared camera, cooling fan, flow meter, reservoir with heating elements, pump, thermocouple with control circuit, velocity sensor, temperature sensors, a PC for sensor data acquisition and another PC for image capture from the thermography camera with analyzing software(See Figs. 1 and 2).More information about experimental setup is summarized in Table 1. Table 1. Details of experimental setup. Thermography camera (ULIRvision TI160) Radiator Fan Water pump Flow meter Reservoir water heating system velocity sensor temperature sensor Focal plane array (FPA) uncooled micro bolometer detector, 8-14 m spectral range ±2% accuracy <65 mk at 30 C thermal Sensitivity Emissivity correction Adjustable from 0.1 to 1.0 copper radiator for U.T.B. Diesel engine suction fan with 3-phase electromotor that controlled by inverter Centrifugal pump Rotameter with Metering range 8-70 lit/min and ±2 lit/min accuracy 9 kW heating element heating elements fixed into the reservoir container with 100 liter volume, water is heated up in the range of 50-120C hot wire velocity sensor with ±2% accuracy LM35 with ±1°C accuracy Fig. 1 A schematic diagram of experimental setup for thermographic fault diagnosis of cooling radiator Fig. 2 Experimental setup for thermal fault diagnosis and condition monitoring of cooling radiator In this setup test, the water heating system acts as a supply of heat which simulates the operation of an engine. The heating element heats up water to a temperature range of 50 OC -110OCin which a simple algorithm is used to control and adjust the temperature of the coolant. After heating, hot coolant is pumped by the centrifugal water pump into the radiator. The rotameter is installed between the pump and radiator and measures the flow rate of the hot coolant. The rotameter's flow is controlled by controlling bypass valves achieving different flow rates of the hot coolant. Then the inlet water temperature to the radiator is sensed by a temperature sensor. The hot coolant flows through the radiator core. The cold air is sucked in by a fan and decreases the temperature of the coolant flowing through the radiator. The velocity of air is controlled with a 3-phase inverter electromotor. Then the outlet coolant is returned to the reservoir where the heating element heats it up again and is recirculated in the flow circuit to maintain the continuity of flow. Six types of radiator condition including radiator tubes blockage (TB), radiator fins blockage (FB), loose connections between Fins & tubes (LC), radiator door failure (DF), coolant leakage (CL) and normal (N) conditions were investigated. Fault diagnosis experiments were conducted for these conditions at three coolant temperatures (70, 80 and 90°C), three flow rates (40, 55 and 70 lit/min), and two suction air velocities (2.0 and3 m/s).Figure 3 shows some of the acquired IR thermal images of the radiator conditions. Fig. 3. Sample of the acquired infrared thermal images of the six condition of radiator (A: radiator fins blockage, B: loose connections between fins &tubes, C: coolant leakage, D: radiator door failure, e radiator tubes blockage: and F: normal) In order for the thermography camera to provide reliable temperature measurements, certain parameters must be set. The most important of these parameters is the emissivity of the object; the other parameters are scale temperature, relative humidity, focal length of camera, and distance. According to the experimental conditions all of these parameters were adjusted accordingly. 2.2. Feature extraction Wavelets, as mathematical functions, decompose data into different frequency components and then analysis every component with a resolution matched to its scale (multi-resolution time-scale analysis). Two dimensional discrete wavelet transform (2D-DWT) is a useful tool to image processing in a multiscale representation structure and to capture details of localized image in space and frequency domains together[21]. Mi and Lan [22]proposed Haar wavelet for image processing that has good advantages for image analysis such as fast, simple processing, high ratio of image compression, good de-noising effect and good image features to maintain the characteristics. In this study, the 2D-DWT with Haar wavelet was used to decompose the original image. The results in each thermal images was decomposed into a first level approximation component, and detailed components that include of horizontal, vertical and diagonal details. Diagram of wavelet decomposed according to decomposition algorithm of Gonzalez et al [20] is shown in Fig. 4. Texture analysis is important in several applications of digital image analysis for classification, detection or segmentation of images based on local spatial patterns of intensity or color. Texture features play an important role in the classification at many machine vision applications such as biomedical image analyzing, automated visual inspection, content based image retrieval, remote sensing applications, etc. Most of the textural features are generally obtained from the application of a local operator, statistical analysis, or measurement in a transformed domain like Wavelet [23, 24]. Wavelet transform attains consistently good performance and ranks among the best approaches. Commonly, the result of wavelet transform cannot be used for character calculation, just statistics result from the result of wavelet transform can be used to indicate the texture character. Accordingly, texture features like mean, variance, moments, gradient vectors and energy density strings are extracted from wavelet sub-images at wavelet decomposition achieve the efficient results [25]. Fig 4. Two dimensional discrete wavelet decomposition using filters bank for thermal image. Wavelet transform attains consistently good performance and ranks among the best approaches. Commonly, the result of wavelet transform cannot be used for character calculation, just statistics result from the result of wavelet transform can be used to indicate the texture character. Accordingly, texture features like mean, variance, moments, gradient vectors and energy density strings are extracted from wavelet sub-images at wavelet decomposition achieve the efficient results [25]. Discrete wavelet transform was used for multiple resolutions image processing. The 2D-WTand first decomposition level were applied in the decomposition of IR thermal image data from different conditions of the radiator. By performing decomposition, four types of wavelet coefficients could be obtained from each IR thermal image. Original thermal image and four kinds of wavelet (approximation component, and detailed components that include horizontal, vertical and diagonal details) were considered for feature extraction because they may be useful for radiator fault diagnosis. The aim of feature extraction is to find a simple and effective transform signal or image to fault diagnosis and condition monitoring. An approach which is used frequently for feature extraction is based on statistical properties of the intensity histogram [20].The histogram features are considered as the most basic feature extraction method of texture analysis which present information in relation to the characteristic of the gray level distribution for the image. Histogram on gray scale images is defined as follows: � = � �� � (1) where z is a random variable indicating intensity, H(zi) is image histogram, N is the number of all pixels in a grey level image and P(zi) is the normalized histogram. First order statistical features, and useful texture features of the image can be obtained from the histogram mean which is the average value of the intensity. It gives some information about general brightness of the image. The variance expresses the intensity variation around the mean and measures average contrast, Smoothness measures the relative smoothness of the intensity in a region. Skewness measures the asymmetry about the mean in the gray level distribution. The energy measures uniformity, when all intensity values are equal (maximally uniform) its value is maximum and decreases from there, in fact it represents about how the gray levels are distributed. The entropy is a measure of randomness of the histogram. The following equations are the first order statistical features of the image that were obtained using the histogram [20, 26]: Mean: Standard deviation: Smoothness: Skewness: Uniformity (Energy): Entropy: = ∑�− = � = √∑�− = � − = 1 − 1⁄ 1 + � = ∑�− = � = ∑�− = � (2) − � = − ∑�− = P � � (3) (4) (5) (6) P (7) Textural features were measured from thermal original image and sub-images of the wavelet decomposition by histogram based features. Therefore, altogether 30 features were extracted for each thermal image. 2.3. Feature Selection After completing the feature extraction process the superior features were chosen from the extracted feature vector. All features may not be relevant to the problem, and some of them might reduce accuracy or cause over fitting. Thus in order to have a sufficient feature vector, features which do not improve classification accuracy should be discarded from the feature vector [27]. Feature selection aims to select a small subset of the relevant features from the original ones according to certain relevance evaluation criterion, which commonly leads to higher learning accuracy for classification, lower computational cost, and better model interpretability [28]. So, feature selection may be viewed as an important pre-processing technique to remove irrelevant and redundant data. It can be applied in both supervised and unsupervised learning methods. In supervised learning, feature selection aims to maximize classification accuracy[29].Genetic algorithm (GA)is a biological approach for performing subset selection and is based on the wrapper method but it needs a classifier such as ANN, SVM, KNN, etc to evaluate each individual of the population [30, 31]. Samanta et al. [32] used GA and ANN as feature selection and classifier for fault detection of bearing. Gamarra and Quintero [33] applied GA feature selection in an ANN classification system for image pattern recognition. The proposed approach is based on a hybrid system that uses both GA and ANN. GA was used to select subsets of genes (features) and ANN classifiers were used to classify cases and returned a metric of the error which was used as a fitness function for selected subset of genes. GAs are iterative processes in which the best individuals of a population are selected to reproduce and pass to the next generation. These steps stop after many generations, optimal or near optimal solutions are described as follows: 1. The feature selection method starts with random generation of an initial population of chromosomes. 2. The evaluation of the fitness function for an individual population begins with the application of the chromosome’s mask. For each chromosome representing selected features, training dataset is used to train the ANN classifier, while selected features of the testing dataset are used to calculate classification accuracy. When the classification accuracy is obtained, each chromosome is evaluated with cost function according to the following formula: Cost function = 1 − overall classification accuracy (8) 3. Parent selection: Select parent chromosomes from population (among the current population and the generated children) according to their fitness. 4. Crossover: With a crossover probability, generate the parents to form new offspring, that is, children. Crossover operator with some constraints has been implemented in order to maintain acceptability of the solution. Crossover probability was set to 80% 5. Mutation: The generated children in the previous stage can suffer mutations in some of their mask’s pixels with a probability of 5%. 6. Place new offspring in the new population. Use new generated population for a further run of the algorithm if the end condition is satisfied, stop, and return the best solution in current population. Otherwise the algorithm pass to the next generation, beginning at step 2. So the feature mask for the best solution is determined as a string like “0010010110011011…”. Here, values of “1” or “0” suggest that the feature is selected or removed, respectively. Fig 5. Proposed framework for intelligent fault diagnosis and condition monitoring of cooling radiator Six kinds of texture features (mean, standard deviation, smoothness, skewness, energy and entropy) were extracted from all thermal images and wavelet data. So altogether, 30 features were extracted for each sample. After completing feature extraction process, the size of data matrix was 1620*30*6 (1620 samples, 30 features and 6 classes). Therefore, in order to reduce dimensions, lower computational cost and higher learning accuracy for classification, feature selection was performed by combining GA and ANN techniques. The proposed framework for intelligent fault diagnosis and condition monitoring of cooling radiator is shown in Fig. 5. In order to do feature selection, a MATLAB program was developed to select the best features by GAs, and then the radiator condition diagnosis classification was done by ANN classifier. 2.4. Intelligent classification and condition monitoring of the radiator Classification is the final step in the condition monitoring process of the radiator. Moreover classification is the process of training in order to assign a sample to pre-determined classes. The aim of classification is to find a rule, based on selected features or training elements, that allows assigning each thermal image of radiator to any possible classes. The classification process includes training, cross-validation, and testing steps. Therefore the features data have to be divided into three subsets: training set, cross-validation set, and testing set. The training set is used to train the ANN, while crossvalidation set is used to prevent the overtraining and the testing set is assigned to test validity of the classifier. Here, 60% of dataset was randomly selected as training set (972 samples), 20% for crossvalidation (324 samples), and the remaining(20%of data set or 324samples) were used as testing set. Multilayer perceptron (MLP) is one of the most popular ANN architectures which is based on supervised learning algorithm and is used for classification. This network consists of input, hidden and output layers. The input layer has nodes which represent the normalized extracted and selected features from the measured thermal images. The number of input nodes was varied from 1 to 30 according to selected feature by GAs. So the number of input nodes is equal to number of the selected features. The number of nodes in the output layer corresponds to the number of target classes. As previously described, there are six classes for condition monitoring of the radiator. The MLP network was built, trained and implemented using MATLAB's neural network toolbox (2013b). Levenberg-Marquardt (LM) algorithm was used as training algorithm. The ANN was iteratively trained to minimize the performance function (MSE) between the ANN outputs and the corresponding target data. The gradient of performance function of MSE was used to adjust the network weights and biases in each iteration. In this study, a MSE of 10-4, a minimum gradient of 10-10, a maximum validation number increase of 10, and maximum iteration number (epoch) of 1000 were used as stopping criterion. The training process would stop if any of these conditions were met. The initial weights and biases of the network were generated automatically by the program. Table 2: confusion matrix for classification of six classes C*1 n11 ... n16 ... n66 C*6 n61 ... C6 ... ... ... C1 The performance analysis of classification can be evaluated by the semi-global performance matrix, known as the confusion matrix. This matrix contains information about actual and estimated classification data obtained by the ANN classifier. Table 2 shows the confusion matrix for a six class classifier which represents how the instances are distributed over actual (columns) and estimated (rows) classes. The terms (nij) correspond to the pixels that are classified into class number i by the ANN classifier (i.e. C*i), when they actually belong to class number j (i.e. Cj). Accordingly, the right diagonal elements (i=j) correspond to correctly classifiedinstances, while off-diagonal terms (i≠j) represent incorrectly classified ones. When considering one class i in particular, one may distinguish four kinds of instances: true positives (TP) and false positives (FP) are instances correctly and incorrectly identified as C *i, whereas true negatives (TN) and false negatives (FN) are instances correctly and incorrectly rejected as C*i, respectively.The corresponding counts are determined as �� = , , �� = ,+ − , , �� = − �� − �� − �� where ,+ � +, − , and �� = +, are the sums of the confusion matrix elements over row i and column j, respectively (Labatut and Cheri, 2011). The classification performances were measured based on the values of the confusion matrix, such as percentage of specificity, sensitivity, precision, accuracy and area under the curve (AUC). The following equations present them for classification: � � � � � �� � � � � ��� = =� ��� +��� �� +��� +��� +��� ��� =� �� +��� =� ��� � ��� =� �� +��� ��� +��� +� ��� �� +��� ��� �� +��� (9) (10) (11) (12) (13) Accuracy focuses on overall effectiveness of ANN classifier. Precision evaluates class agreement of the data labels with the positive labels defined by the classifier. Sensitivity shows the effectiveness of the ANN classifier to recognize positive labels and how effectively a classifier identifies negative labels and AUC is indicative ability of classifier to avoid false classification (Sokolova and Lapalme, 2009). 3. Results and Discussions Thermal IR image acquisition from the different condition of radiator was done by thermal camera at three coolant temperatures (70, 80 and 90°C),three flow rates(40, 55 and 70 lit/min) and two suction air velocities (2.0 and3 m/s), respectively. Altogether, 1620 samples were obtained from IR thermal images of all different conditions of the radiator and all experimental conditions. After gray scaling and auto-cropping to eliminate the background, some of the acquired thermal images of different conditions of radiator are presented in Fig.6. When radiator fins are blocked in different areas, the air flow is severally restricted. So, irrespective of the type of material that blocks the radiator's surface, the heat transfer rate in these zones is reduced.So there will be hot spots in the thermal image in these areas (higher intensity in the gray level) compared to the normal condition (compare Figs. 6A and 6F).Where there are loose connections between fins and tubes in different areas, the fins do not participate in the heat transfer. Thus, there will be cold spots in the thermal image in these areas (lower intensity of gray level) compared with normal condition (see Fig. 6B). When the radiator has a leakage, that is a gradual loss of coolant, the coolant slightly flows out; because of the higher heat transfer coefficient of water than air or may be due to evaporation of the fluid. There will be cold spots in leakage areas (see Fig. 6C). Figure 6D shows the effects of radiator door failure in the IR thermal image in which the temperatures of the top parts are hotter than in the normal condition. The radiator tube blockage condition in different areas leads to clogged coolant flow in these tubes. Therefore, these tubes do not participate in heat transfer. So there will be cold spots in the thermal image in these regions compared with the normal condition (see Fig. 6E).The proposed system is expected to be able to detect the difference between the IR thermal images for each condition of the cooling radiator and also to classify and recognize the different conditions of radiator. Fig. 6. Some of the IR thermal images of the six conditions of radiator after gray scale and auto-cropping (A: radiator fins blockage, B: loose connections between fins &tubes, C: coolant leakage, D: radiator door failure, E: radiator tubes blockage: and F: normal) For feature selection, finding the best features manually is very time consuming so it is necessary to build and investigate Q__ networks to check all the cases. The GA assisted in the selection of the best features in a very short time and with checking 1020 networks only. Running the program took almost 30 min with a PC (core i5 CPU, TM4440 @ 3.10 GH).The result of program was a MLP network with only one hidden layer that has 6 neurons. The final selected features were used as inputs of classifier by optimizing accuracy of ANN's classifier. The selected features used as the inputs of ANN classifier are shown in Table 3. Table. 3. Final selected features, based on GA Images/ Statistical texture features Original Thermal Image Wavelet Approximation Image Wavelet horizontal Image Wavelet vertical Image Wavelet diagonal Image Mean 1 0 1 0 1 Standard Deviation 0 0 0 0 1 Smoothness Skewness 1 1 0 1 1 0 1 1 0 0 Energy 1 0 0 1 0 Entropy 0 1 1 1 1 The final and important stage for fault detection is classification. For the cooling radiator, after creation the MLP network was trained by the LM back propagation training algorithm. Mean square error (MSE) was used as the performance function and selected features were considered as the inputs of the network. The trained network was implemented for classifying the test data. The topology of the ANN is the main component in designing an optimal classifier, because structure impacts on the learning ability and the accuracy of the final network in classifying data. The number of hidden layers and their neurons are major factors for designing MLP networks. The number of neurons in the input and output layers were fixed because they are dependent on the feature vector and the number of classes, respectively. The input layer consisted of 16 nodes based on the feature selection operation (see Table 3). The output layer consisted of six neurons which were related to the six classes, i.e., radiator tube blockage, radiator fin blockage, loose connection between fins &tubes, radiator door failure, coolant leakage and normal, for fault diagnosis and condition monitoring of the radiator. The number of hidden layers and their neurons depend on the difficulty of the investigated problem. Generally, one hidden layer with a lower number of neurons is preferred, because it leads to a reduction in the network size and an increase in the network’s learning ability. This matter is very important for online condition monitoring. Several combinations of the number of neurons in the hidden layer varying from 2 to 15 and the number of epochs varying from100 to 1000were investigated by a trial and error method. To find the best combination, the total classification accuracy was used as the selection criterion. The results showed that the hidden layer with six neurons (i.e., a 16-6-6 topology) had the smallest size with the highest total classification accuracy. Thus, the 16-6-6 network was selected as the best topology for fault diagnosis and condition monitoring of the cooling radiator. This MLP network as the radiator classifier is shown in Fig. 7. Table 4 shows the confusion matrix as a result of ANN with primary feature vector inputs (without any feature selection) using experimental data. Table 5 gives the performance parameters of the classifier according to the above-mentioned confusion matrix, including classification accuracy, precision, sensitivity, specificity and AUC for radiator tube blockage (TB), radiator fin blockage (FB), loose connection between fins &tubes (LC), radiator door failure (DF), coolant leakage (CL) and normal (N) classes. The classification accuracy of the ANN classifier for FB, LC, CL, DF, TB and N classes were 95.99%, 98.45, 94.13, 91.66, 99.38 and 97.53%, respectively. The overall accuracy of the ANN classifier obtained was 88.58 %. The average per class accuracy, precision, sensitivity, specificity, AUC were 96.62%, 87.38.68, 87.47, 97.77 and 92.62%, respectively. Tabele.4. Confusion matrix obtained from the evaluation of 30-6-6 ANN classifier (without feature selection). Estimated/ Actual FB LC CL DF TB N FB LC CL DF TB N 57 0 3 4 0 0 0 58 1 2 0 0 1 0 38 4 0 0 3 2 7 28 0 3 0 0 0 2 58 0 2 0 3 0 0 48 Table.5.Performance measurements of 30-6-6 ANN classifier (without feature selection) Class Accuracy Precision Sensitivity Specificity AUC FB 95.99 90.47 89.06 97.69 93.37 LC 98.45 96.66 95.08 99.23 97.16 CL 94.13 73.08 88.37 95.01 91.69 DF 91.66 70 66.17 95.73 80.42 TB 99.38 100 96.66 100 98.33 N 97.53 94.11 90.56 98.9 94.73 Average per-class 96.62 87.38 87.47 97.77 92.62 Table 6 shows the confusion matrix as a result of the ANN classifier with the 16-6-6 topology using selected features as inputs (Table 3, feature selection based on GA) for the experimental dataset. Table 7 gives the performance measurements of the classifier according to the confusion matrix (Table 6), including classification accuracy, precision, sensitivity, specificity and AUC for all classes. The classification accuracy of the ANN classifier for FB, LC, CL, DF, TB and N classes were 99.07%, 99.07, 95.37, 95.37, 99.38 and 98.76%, respectively. The overall accuracy of the optimum ANN classifier was obtained as 93.83%. The average per class of accuracy, precision, sensitivity, specificity, AUC were 97.84%, 92.35, 92.93, 98.75 and 95.66%, respectively. Tabele.6. Confusion matrix obtained from the evaluation of 16-6-6 ANN classifier (with GA feature selection). Estimated/ FB LC CL DF TB N Actual FB 0 1 0 0 0 60 LC 0 1 0 0 0 60 CL 0 0 9 0 1 36 DF 0 2 3 1 0 34 TB 0 0 0 0 0 66 N 2 0 0 0 1 48 Table.7.Performance measures of 30-6-6 ANN classifier (with GA feature selection). Class Accuracy Precision Sensitivity Specificity AUC FB 99.07 98.36 96.77 99.62 98.2 LC 99.07 98.36 96.77 99.62 98.2 CL 95.37 78.26 87.8 96.47 92.14 DF 95.37 85 79.07 97.86 88.47 TB 99.38 100 97.06 100 98.53 N 98.76 94.12 97.96 98.91 98.43 Average per-class 97.84 92.35 92.93 98.75 95.66 Comparing the results of for the two classifiers (classifying using primary feature vector and selected feature vector via GA) indicted significant improvement of classification performance using the GA feature selection. The GA feature selection method selected a small subset of the relevant features from the original ones according to certain relevant evaluation criterion. This led to a higher classification performance, lower computational cost and better model interpretability. Figure 8 shows the changes in MSE with changing epoch number when MLP was used for determining the selected feature inputs, over a training period of 203 epochs. The best performance of the ANN is seen at epoch 193 for MLP with16-6-6 topology. Fig 7. The best ANN model with 16-6-6 topology for fault diagnosis of cooling radiator. Fig 8. Changes of MSE related to epoch numbers when MLP for selected feature inputs, is trained with 203 epochs Therefore, the proposed intelligent system could classify and recognize IR thermal images for the different conditions of radiator with high accuracy. This provides confidence that the system can be employed for the intelligent condition monitoring and fault diagnosis of a cooling radiator. 4. Conclusions This paper has presented a useful application of thermography for intelligent fault diagnosis and condition monitoring, and applied it to a cooling radiator. In general, the application of intelligent condition monitoring and fault diagnosis to detect fault types precisely is very complex and difficult, but by combining image processing, genetic algorithm (GA) and artificial neural network (ANN) techniques provides both diagnosis efficiency and accuracy gains. In this study, a new intelligent diagnosis system has been developed and applied to the classification of six types of cooling radiator conditions via using of infrared thermal images; namely, radiator tube blockage, radiator fin blockage, loose connections between fins and tubes, radiator door failure, coolant leakage and normal. The proposed system consisted of several subsequent procedures including thermal image acquisition, preprocessing, image processing via two dimensional discrete wavelet transform (2D-DWT), feature extraction, feature selection, and classification. The 2D-DWT was implemented to decompose the thermal images. Subsequently, statistical texture features were extracted from the original and decomposed thermal images. The feature selection based on GA was used in selecting significant features in order to enhance the performance of the ANN in the final stage. The classification results demonstrated that this system could be employed for the intelligent condition monitoring and fault diagnosis of mechanical equipment that has a strong thermal signature indicating its operating state. Since there is some initial training of the system, it is better suited to the ongoing and periodic maintenance of multiple pieces of similar equipment. In such cases, there should be a significant positive return on the investment. Nomenclature Cj C*i Actually Classes Classifier Classes FP FN False Positives False Negatives TP TN H(zi) p(zi) z N True Positives True Negatives Image Histogram Normalized Histogram Random variable of intensity Number of all pixels  m S U E Mean Standard deviation Skewness Uniformity Entropy References [1] T.Astarita, G.M. Carlomagno, Infrared thermography for thermos-fluid-dynamics, Springer, 2013. [2] G.M. Carlomagno, G. Cardone, Infrared thermography for convective heat transfer measurements, Exp. Fluids. 49 (2010) 1187-1218. [3] Y.Cao, K. Kengskool, An automotive radiator employing wickless heat pipes, MAESC Conference, Memphis, TN. 2009. [4] S.D. Oduro,J. Ampofo,Effect of radiator fins blockage by clay soil on the engine cooling temperature, Int. J. Eng. Tech.2 (6)(2012) 1002-1009. [5] S.C. Pang, M.A. Kalam, H.H. Masjuki, M.A.Hazrat, A review on air flow and coolant flow circuit in vehicles’ cooling systemInt. J. Heat Mass Transfer 55 (2012) 6295–6306. [6] S. Maoa, C. Cheng, X. Li, E.E. Michaelides, Thermal/structural analysis of radiators for heavy-duty trucks. Appl. Therm. Eng. 30 (2010) 1438-1446. [7] R.A. Epperly, G.E. Heberlein, L.G. Eads, A tool for reliability and safety: predict and prevent equipment failures with thermography, in: IEEE Petroleum and Chemical Industry Conference, Banff, Alta, 1997, pp. 59–68. [8] S. Bagavathiappan, B.B. Lahiri, T. Saravanan, J. Philip, T. Jayakumar, Infrared thermography for condition monitoring – A review. Infr. Phys. & Tech.60 (2013) 35–55. [9] A.S. Nazmul Huda, S. Taib. Application of infrared thermography for predictive/preventive maintenance of thermal defect in electrical equipment. Appl. Therm. Eng.61 (2013) 220-227. [10] C. Meola, G.M. Carlomagno, A. Squillace, G. Giorleo. The use of infrared thermography for nondestructive evaluation of joints.Infr. Phys. & Tech.46 (2004) 93–99. [11] W.T. Kim, S. Ranjit, D.H. Park, M. Y. Choi, Infrared Thermographic NDT for the Fault Diagnosis of Bearing with Foreign Substances inside under Loading Condition. 11th European Conference on Non-Destructive Testing (ECNDT 2014), Prague, Czech Republic, 2014. [12] S.G. Lee, W.T. Kim, Inspection of the leakage for the closure plug of heavy water reactor using infrared thermography in nuclear on-site application, 10th International Conference on Quantitative Infra Red Thermography, Québec (Canada),2010. [13] Z. Ge, X. Du, L.Yang, Y. Yang, Y. Li, Y. Jin, Performance monitoring of direct air-cooled power generating unit with infrared thermography, Appl. Therm. Eng.31 (2011) 418-424. [14] A.M. Younus, B.S. Yang, Intelligent fault diagnosis of rotating machinery using infrared thermal image, Expert Syst. Appl. 39 (2012) 2082–2091. [15] A. Darabi, X. Maldague. Neural network based defect detection and depth estimation in TNDE, NDT&E International 35 (2002) 165-175. [16] I.A. Abu-Mahfouz, A comparative study of three artificial neural networks for the detection and classification of gear faults. In. J. Gen. Syst. 34 (3) (2005) 261-277. [17] Z, Dworakowski, L, Ambrozinski, P. Packo,K. Dragan, T. Stepinski, T. Uh, Application of artificial neural networks for damage indices classification with the use of Lamb waves for the aerospace structures. Key Eng. Mat, 588 (2014) 12-21. [18] S.E. Umbaugh, Computer Imaging:Digital Image analysis and processing, Taylor & Francis, UK, 2005. [19] A.M. Younus, A. Widodoa, B.S. Yang, Application of thermal image: machine fault diagnosis using PCA and ICA combine with SVM, International Conference on Mechanical Engineering (ICME2009) Dhaka, Bangladesh, 2009. [20] R.C. Gonzalez, R.E. Woods,S.L. Eddins, Digital image processing using MATLAB. Pearson Prentice Hall, New Jersey, USA, 2004. [21] S. Mallat, Theory for multi resolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell. 11(7) (1989) 674-693. [22] C. Mi W. Lan,Haar Wavelet Transform application in image processing, Ningxia Engineering Technology, 2003. [23] N. Sebe, M. Lew, Wavelet based texture classification, in Proc. 15th Int. Conf. Pattern Recogn. (ICPR), Barcelona, Spain, 3(2000) 3959–3962. [24] P.S. Hiremath, S. Shivashankar. Wavelet based features for texture classification, GVIP Journal, 6, (3) (2006) 55-58. [25] P.W. Huang, S.K. Dai, Image retrieval by texture similarity. Pattern Recogn. 36 (2003) 665 – 679. [26] A. Materka, M. Strzelecki, Texture Analysis Methods – A Review, Technical University of Lodz, Institute of Electronics, COST B11 report, Brussels 1998. [27] K. Mollazade, M. Omid, A. Arefi. Comparing data mining classifiers for grading raisins based on visual features. Comp. Elec. Agr. 84 (2012) 124–131. [28] J. Tang, S. Alelyani, H. Liu. Feature selection for classification: A review. In Data Classification: Algorithms and Applications. CRC Press, 2014. [29] J. Dy, C.E.Brodley, Feature selection for unsupervised learning, J. Mach. Learn. Res. 5 (2004) 845-889. [30] A.A. Altun, N. Allahverdi, Neural network based recognition by using genetic algorithm for feature selection of enhanced fingerprints. In ICANNGA, 2007. [31] K. Kim, I. Han, Genetic algorithms approach to feature discretization in artificial 19 neural networks for the prediction of stock price index, Expert Syst. Appli. 19(2) (2000) 125–132. [32]B. Samanta, K.R. Al-Balushi, S.A. Al-Araimi, Bearing Fault Detection Using Artificial Neural Networks and Genetic Algorithm. EURASIP J. Appl. Sig. Process.3 (2004)366–377. [33] M.R.A. Gamarra,C.G.M.Quintero, Using genetic algorithm feature selection in neural classification systems for image pattern recognition, Ingeniería E Investigación, 33(1) (2013) 52-58. [34] V. Labatut, H. Cheri, Accuracy Measures for the Comparison of Classifiers. Al-Dahoud Ali. The 5th International Conference on Information Technology, May 2011, Amman, Jordan. Al-Zaytoonah University of Jordan, 2011. [35] M. Sokolova, G. Lapalme, A systematic analysis of performance measures for classification tasks. Inform. Process. Manag. 45 (2009) 427–437.