Misfire Detection in Spark Ignition Engine Using Transfer Learning

Misfire detection in an internal combustion engine is an important activity. Any undetected misfire can lead to loss of fuel and power in the automobile. As the fuel cost is more, one cannot afford to waste money because of the misfire. Even if one is ready to spend more money on fuel, the power of the engine comes down; thereby, the vehicle performance falls drastically because of the misfire in IC engines. Hence, researchers paid a lot of attention to detect the misfire in IC engines and rectify it. Drawbacks of conventional diagnostic techniques include the requirement of high level of human intelligence and professional expertise in the field, which made the researchers look for intelligent and automatic diagnostic tools. There are many techniques suggested by researchers to detect the misfire in IC engines. This paper proposes the use of transfer learning technology to detect the misfire in the IC engine. First, the vibration signals were collected from the engine head and plots are made which will work as input to the deep learning algorithms. The deep learning algorithms have the capability to learn from the plots of vibration signals and classify the state of the misfire in the IC engines. In the present work, the pretrained networks such as AlexNet, VGG-16, GoogLeNet, and ResNet-50 are employed to identify the misfire state of the engine. In the pretrained networks, the effect of hyperparameters such as back size, solver, learning rate, and train-test split ratio was studied and the best performing network was suggested for misfire detection.


Introduction
In the modern era, transportation has become an integral part in the day-to-day life of every individual. People have to commute to places for either personal or official work. In one's income, transportation cost takes a reasonable amount of share. Also, from the industrial point of view, a huge amount of money is spent on transportation. While this is inevitable, there are few things that are in our hands by which one can save some amount of money. at can be done by keeping the vehicle in the proper working condition and thus saving the waste of money due to the faulty components and their working. Misfire is one such fault that takes away a lot of money and engine power. e causes for misfire may be due to any one of the reasons such as cylinder gasket leak, electrical fault, sensor malfunction, and defective spark plug. Misfire can result in loss of power and produce unwanted vibration that can impart wear and tear in allied components. Finding the misfire at an early stage and correcting it will save a lot of money for the owners of the vehicle. Misfire in an engine is not only a matter of money but also has the other technical impacts on the engine. In a multicylinder, if one cylinder is misfiring, then the remaining cylinders have to work extra in order to take the load of the engine. Apart from money and technical impact, misfire has an adverse effect on the environment also. Occurrence of misfire in a cylinder of the IC engine infers that the fuel has gone to the cylinder and the process of combustion did not initiate. Misfire can sometimes cause knocking, and the unburnt gases can be emitted as a pollutant into the atmosphere.
ere are various misfire detection methods that are proposed by researchers and that can be categorized as follows [1]: (1) Temperature monitoring: One can monitor the catalyst temperature [2] or exhaust temperature [3] and utilize these data for misfire detection; however, this method will work well only when the ignition-induced misfire rate is above certain percentage. (2) Wheel speed sensor: Many researchers have worked on wheel speed sensor because the wheel speed sensor is already available in the automobile. Based on the patients in the crank speed, researchers find misfire in the engine [4].

(3) Oxygen sensor:
It is mounted in the exhaust manifold, and from the fluctuation of the differentiated oxygen sensor signal, the misfire information can be found [5]. (4) Pressure monitoring: e misfire can be easily monitored effectively using the pressure inside the engine cylinder. However, the reliability and cost of the pressure sensor are the challenges [6][7][8]. (5) Vibration monitoring: Vibration signal acquired from the engine head is used for extracting features that can be used with machine learning algorithms. Babu Deva Senapati et al. have used how machine learning techniques along with features extracted from vibration signals and classified the misfire state of the IC engine [9]. ere are researchers who have used the time domain in signals directly, and other researchers have transformed the time domain signals into frequency domain signals using fast Fourier transform, and then, they have applied machine learning for classification of the misfire states. However, vibration signals are highly dynamic in nature and complex.
is throws a lot of challenges in the signal processing of vibrations.
Researchers have extensively used statistical features, histogram features, autoregressive moving average features, and wavelet features for misfire detection. While using these features, they have used a number of feature selection techniques including principal component analysis and decision tree. e classifiers they have used include decision tree support vector machine K Star algorithm and many more. While there are so many research works reported in the literature using machine learning in misfire detection, the effectiveness of machine learning depends upon the effectiveness of the features. e features engineering is very challenging especially when the signal is very complex like vibration signals. In this scenario, researchers look for some automated technology that can learn to classify without extracting explicit features from the signals. Deep learning is one such technology that has the capability to learn from the images without explicitly extracting features from them. However, one cannot take the pictures of the misfire inside the cylinder. It is unsafe and very costly affair.
Taking vibration signals from the engine cylinder head is a family and established activity. e vibration signals can be plotted and that can be used as a signature which can be used to detect the misfire in IC engines. Deep learning convolutional neural networks have a capability to understand the images and classify them as per the predefined classes.
is study is our humble attempt to check whether the signatures of vibration signals plot are good enough to train deep learning algorithm for misfire detection. In the present study, pretrained networks such as AlexNet, VGG-16, GoogLeNet, and ResNet-50 were used. As these networks have been trained already for some other data set, they are called pretrained networks. It is easy to train pretrained network than a brand-new network from scratch. ese pretrained networks have some knowledge, and we have to adapt that knowledge to the problem at hand (transfer the knowledge to other domain). is is called transfer learning, a subset of deep learning.
Later, significant developments in artificial intelligence (AI) and "Internet of ings (IoT)" have gained the attention of the scientific community. AI is considered as a powerful tool in the field of big data processing and data exploration. Modern intelligent fault diagnosis techniques are built on theories and concepts based on AI. Deep learning, an AI strategy first proposed by Hinton et al., in the field of science initiated the wave of research into different fields of study. Deep learning strategy (also termed as deep neural networks) consists of a number of neural layers stacked in a hierarchical structure that extracts information from the input. e architecture is called "deep" since the raw data information is learnt through multiple levels of layer-bylayer procedure. Starting from the raw data input, Deep Neural Network (DNN) discovers the structure of complicated data sets and automatically learns the most significant features through several layers. Automatic feature learning capability and improved nonlinear regression ability have made deep learning models to be widely used in language processing, object detection, visual inspections, surveillance, and robotics. Hence, there is a critical need and great potential to utilize the automatic learning capability of DNN in the field of mechanical systems fault diagnosis.
In the past few years, DNN was adopted by many researchers to either perform classification or feature selection in the field of fault diagnosis imitating traditional fault diagnosis methods. Initially, the features from acquired signals were extracted using various feature extraction techniques that are utilized to classify DNN models. Several authors reported on the above-mentioned strategy that is discussed as follows. A deep belief network was proposed by Li et al. to diagnose the condition of bearings and gearboxes using statistical features obtained from frequency, time, and time-frequency domains [10]. Chen et al. adopted a convolutional neural network (CNN) to detect the condition of the gearbox using frequency and time features [11]. An air compressor fault diagnosis was carried out by Verma et al. using a sparse autoencoder [12]. Shao et al. used an optimized deep belief network to diagnose bearing faults using 18 time features [13]. In this stage of fault diagnosis, DNN was used only as a replacement for the classifier, and the prime advantage of DNN feature learning was not equipped completely ( Figure 1(a)).
After 2015, the researcher started using DNN for feature learning and feature selection along with classification.
is provided a complete suite for all the activities in one go. Here, DNNs have taken images as input and corresponding classes as output. e DNNs are learnt without explicit need for feature extraction, feature selection, etc. e literatures based on the above-mentioned stage are discussed as follows. Condition monitoring of roller bearings was carried out by Guo et al. using deep convolutional neural networks consisting of two ensembles. Feature extraction and fault pattern recognition were carried out in one CNN while the fault classification was performed in the other CNN [14]. Zhao et al. performed tool condition monitoring using a convolutional long short-term memory (C-LSTM) [15]. Owing to the feature learning capability of DNN, the process of intelligent fault diagnosis grows into an effective and more automated approach (Figure 1(b)).
CNN forms the basic blocks of deep learning that have been adopted to learn complex features from image data. CNN is considered as one of the most dominant approaches in fault diagnosis, object detection, and speech recognition. However, very limited studies have been carried out in misfire detection of IC engines. Additionally, the application of CNN in the misfire detection was not attempted. Some related works representing the application of CNN in mechanical systems are presented in Table 1. In the present study, the performance of various state-of-the-art pretrained networks like AlexNet [23], VGG-16 [24], GoogLeNet [25], and ResNet-50 [26] was evaluated for detecting the misfire state from images acquired from vibration signals. Experiments were carried out on the pretrained networks by varying the train-test split ratio and various hyperparameters such as batch size, learning rate, and solver. e derived results are compared and tabulated to identify the suitable network for misfire detection in IC engine. e technical contributions in the present study are provided as follows.
(i) e present study considers four misfire conditions (c1 mis, c2 mis, c3 mis, and c4 mis) and one normal condition in a four-cylinder petrol engine. (ii) Vibration signals were obtained from the head of the engine, and the plots of vibration signals were utilized as input for pretrained networks.
(iii) Four pretrained networks namely, AlexNet, VGG-16, GoogLeNet, and ResNet-50, were considered in the study to perform classification of misfire state. (iv) Various hyperparameters such as batch size, solver, learning rate, and train-test split ratio were altered, and the performance of the pretrained networks was assessed. (v) e best performing network for misfire detection is identified based on the results obtained. e rest of the paper is organized as follows. In Section 2, the experimental description includes experimental setup, data acquisition, and procedure. Section 3 presents the overview of the convolution neural network (CNN). Section 4 describes the process of transfer learning for misfire detection using vibration signal. Result and discussion are provided in Section 5, and the conclusion is presented at the end of the manuscript.

Experimental Studies
e present section discusses about the experimental setup fabricated for the study, data acquisition method, and experimental procedure followed in the study. e experimental setup comprises of a spark-assisted gasolinepowered IC engine. Initially, the accelerometer was placed (using adhesive technique) on the engine to acquire the vibration signals for good condition. e same procedure is repeated further to acquire the vibration signals of misfire in the IC engine by cutting off the spark in the cylinder one by one. e complete methodology involved in the process of misfire detection in IC Engine is depicted in Figure 2.

Experimental Setup.
e present study utilizes a sparkignited IC engine that has been equipped with provisions to create the misfire manually in a specific cylinder. e complete setup is connected to a data acquisition system and an accelerometer to acquire the vibration signals for every engine condition. An analog to digital converter helps in converting the acquired analog vibration signals into digital values that are stored in a storage device. Figure 3 shows the experimental setup used in the present study. A 10hp-rated four-cylinder four-stroke spark-ignited (SI) engine is adopted in the present study. e misfire in the cylinders is performed by cutting off the power supply to the spark plugs present in each cylinder. A nut and screw mechanism is used to fix the accelerator in the engine at the specific position. e engine accelerator is firmly attached at the desired position via a screw and nut mechanism. e speed of the engine is kept constant in no-load condition. e specifications of the engine are provided in Table 2.

Data Acquisition.
Data acquisition (DAQ) is the process of creating digital values from the world around such that it can be visualized, analyzed, and stored in a computer. In the present study, misfire detection in an IC engine is carried out by acquiring vibration signals with the help of an Computational Intelligence and Neuroscience accelerometer (piezo-electric sensor with sensitivity of 10.26 mV/G). Accelerometers are capable of detecting both large and small vibrations. e working principle of the accelerometer states that output voltage magnitude is directly proportional to vibration signal intensity. e accelerometer is fixed on the centre of the engine block using adhesive techniques such that the vibration data are recorded for all conditions. A dactron FFT analyzer is used to convert the analog vibration signal into digital form. e acquired vibration data are further processed, and vibration plots are stored in a computer where a transfer learning approach is applied to identify the best performing pretrained network for misfire detection.

Experimental Procedure.
Initially, electrical cranking was applied to start the engine at no-load condition. A buffer time of 15 minutes is provided to the engine for achieving the optimal operating condition. e FFT analyzer connected to the engine is turned on after engine stabilization. e engine speed was maintained at 1500 rpm, and the vibration data for normal operating condition are measured at 24 kHZ sampling frequency and 8192 sampling length. Five test conditions, namely, misfire in either of the cylinder 1, 2, 3, or 4 and normal condition, are considered in the study. Data acquisition for all the test conditions was performed at 1500 rpm. e sample vibration signals for misfire and normal operating conditions are presented in Figure 4. e two waveforms acquired display significant differences aligning with the fact that unique patterns are displayed by vibration signals for each condition. e parameters considered during signal collection are described as follows: (i) Sampling length: 8192 steps (ii) Sampling frequency: 24 kHz (iii) Number of instances for each condition: 100 (iv) Speed: 1500 rpm

Description of Convolutional Neural Networks (CNN)
CNN formulates a connection between the image and image features by creating a set of biases and weights. e architecture is designed in a hierarchical manner that works on an automatic feature learning algorithm. e classification ability of CNN is based on the features that were learnt during the process of convolutions. e CNN architecture is composed of sequential stacking of several layers, namely, convolution layer, pooling layer, and fully connected layer. e general architecture of CNN architecture is provided in Figure 5.
e basic function of CNN is involved around the abovementioned layer groups as given below: (a) e input for a convolution layer is provided with the help of input layer which stores the image pixel values. (b) e convolution layer is stacked next to the input layer that is filled with neurons of different weights and biases. e output of every neuron is calculated based on the product of volume and weights provided by the input. Rectified linear units (ReLU) is adopted widely as the activation function such that the nonlinearity among the problem is sustained. (c) e immediate layer present next to convolution layer is the pooling layer which acts as the downsampling layer. Features of high dimension are sampled down to achieve the spatial dimensionality such that the number of parameters can be reduced. (d) e fully connected layers complete the CNN architecture by providing the classification results for a particular problem. e classification is carried out by assigning a particular range of values to every class such that improved performance is achieved. Matrix form of image features is converted into vector form through the fully connected layers.
CNN is able to transform the original input layer by layer to generate class scores for classification and regression purposes using convolutional and down sampling techniques. It is therefore important to remember that it will not be enough to simply determine the overall design of CNN architecture. ese models can take some time to build and refine. Now, let us analyze the individual layers in detail, describing their hyperparameters and connectivity.

Convolution Layer.
e learning process of CNN is instrumented by means of the convolution layer. e collection of several learnable kernels or filters is dependent on the parameters assigned to the layers. e kernels function in such a way that they spread across the wide range of input by consuming lower spatial dimensionality. In the process of convolution, two-dimensional activation maps are created for every image fed into the filter throughout the spatial dimensionality. e scalar product of weights and volume are calculated for every image data point that passes through the kernel present in the convolution layer. e values created through each filter (generally termed as activations) triggers the network to learn significant features that are available in the spatial domain. e midpoint of the kernel is placed over the input vector through which the weighted sum of itself and any neighboring pixels are calculated and replaced. Also, the model complexity can be reduced by convolution layers by means of hyperparameter optimization. Depth (no. of filters), stride (movement of filter in one direction), and zero-padding (adding zeros around the border of input image) are the three hyperparameters that can optimize the performance of convolution layers.  Deep learning technique Mechanical system [16] CNN with wavelet transform Motor bearing [17] Hierarchical CNN Roller bearing [18] CNN [12] Sparse autoencoder and deep belief network [19] Recurrent neural network [20] Stacked autoencoder Gear box [21] Generative adversarial network [22] CNN Centrifugal pump Computational Intelligence and Neuroscience 5 have a destructive nature and are classified into two forms depending upon their function, namely, average and max pooling. Also, max pooling is the most widely used pooling layers due to the efficient performance on all data types. e size of the filter and stride length are commonly fixed at 2 × 2 to allow the pooling layer to expand throughout the input spatial dimensionality.

Fully Connected
Layer. e ultimate layer of any CNN network is composed of fully connected layers. e output from the final convolution or pooling layer acts as the input to the fully connected layer. e output of convolution or pooling layers will be of matrix form which must be flattened before being fed into the fully connected layer. Sigmoid or softmax are adopted as the activation functions fully connected to perform the classification task for the given input data.

Misfire Detection in IC Engines Using
Pretrained Models e present section discusses about the various pretrained networks considered in the study to detect misfire in IC engines. Initially, the vibration signals are acquired for various conditions of IC engines and stored in the form of vibration plots (images). e acquired images are further resized and preprocessed to a size of 227 × 227 or 224 × 224 according to the input requirement of the adopted pretrained model. Further, various renowned pretrained network models like VGG-16, AlexNet, ResNet-50, and GoogLeNet were utilized to perform image classification and identify the condition of the IC engine. Transfer learning is adopted in this study in which the initial weights of the networks trained on ImageNet are restored. Also, to apply the networks for custom data set, the final output layers are replaced with new layers corresponding to the number of classes defined by the user. Figure 6 represents the overall workflow of misfire detection in IC engines using pretrained networks. A brief description of the pretrained models considered in the study is provided as follows.

Data Set Formation and Preprocessing.
In the present study, a data set of images containing the conditions of the IC engine was created from the vibration signals acquired. Five test conditions, namely, c1 mis, c2 mis, c3 mis, c4 mis, and normal condition, were considered. A total of 500 images (100 images for every class) were created using the acquired vibration signals. e acquired images were resized to a 224 × 224 or 227 × 227 such that the input size of the image is acceptable for the pretrained network adopted. Figure 7 represents the vibration plots of various conditions of IC engine conditions collected for the study.

AlexNet Pretrained Network.
Alex Krizhevesky introduced one of the best performing networks in the annual ImageNet Large Scale Visual Recognition Challenge (ILSVRC) that has been trained over 1.2 million images consisting of 1000 image classes. e network paved way for various advancements in the field of computer vision using deep learning. AlexNet is a simple network composed of 8 layers and 61 million learnable parameters. e input of the network accepts images of size 227 × 227. e input image is fed into the first convolution layer, where 96 different filters with a size of 11 × 11 try to convolve, normalize, and max pool the image to a size of 55 × 55. e output of the first   Computational Intelligence and Neuroscience convolution layer is fed into the second convolution layers consisting of 256 receptive filters followed by a max pooling layer of 3 × 3 size. e output image size is further reduced to a size of 27 × 27. e image is further reduced to a size of 13 × 13 while passing through the consecutive convolution layers (third, fourth, and fifth). Rectified linear units (ReLU) are adopted as the activation function in each layer to resolve nonlinear problems. Post the final convolution layer, two fully connected layers with 4096 learnable parameters is stacked to convert the image matrices into vector form. e architecture is completed with an output layer comprising of softmax activation function to perform classification for the adopted problem. To avoid the problem of model overfitting, a dropout layer of ratio 0.5 is added before the final fully connected layer.

VGG-16 Pretrained
Network. e VGG-16 architecture was introduced by Karen Simonyan and Andrew Zisserman in the annual ILSVRC 2014 that was eventually claimed as the best performing network. e VGG-16 also termed as Visual Geometry Group-16 (also known as Oxford Net) was developed by a group of researchers from Oxford University. e network was constructed with 13 convolution layers, 5 max pooling layers, 3 fully connected layers, and a classification layer. e convolution layers are stacked in a unique pattern that is focused towards image classification problems. e working of the VGG-16 architecture can be understood by considering two learnable parameters "B" and "d" along with a filter of size 3 × 3. A mathematical operation is carried out when the image passes through the convolution layer which makes the filter to move about the image of "z" pixels such that the convolution operation is performed to produce an output image "y". e following equation, y � f(By + d), represents the working of the convolution operation. Simple image features like edges are learnt by initial layers of VGG-16 architecture while complex image features are learnt by deep layers. e convolution process followed by max pooling helps in breaking down and resizing images such that highly contributing features are extracted along with limited  memory consumption. e stride value and size of the filter can influence the image output from a convolution layer. Every convolution layer is provided with ReLU activation function, and a dropout of 0.5 is attached before the fully connected layer.

GoogLeNet Pretrained Network.
In the annual ILSVRC 2014, Szegedy et al. proposed a network architecture named GoogLeNet that was aimed at solving image classification and object detection. e architecture was composed of 22 layers that had its application extended into the fields of facial recognition, robotics, adversarial training, etc. e network is arranged with nine inception modules connected to four convolution, four max pooling, three average pooling, five fully connected, and three softmax layers. ReLU is the activation function adopted in fully connected layers that are supported by a dropout layer of 0.5 ratio. Inception modules present in GoogLeNet architecture enable the network to solve more complex computer vision problems. Identification of complex features through variations in convolution layer filter size is the prime advantage of adopting inception modules. Such process can help in reducing the computational time and dimensional complexity. Even though the architecture of GoogLeNet looks robust with 22 layers, the volume of trainable parameters is less in comparison with AlexNet.

ResNet Pretrained Network. Residual network (ResNet)
was the most successful and efficient network in the annual ILSVRC 2015 developed by He et al. e major advantages of using ResNet architecture is high convergence rate with accurate classification. e common objects in context (COCO) data set were used to train the ResNet architecture. Residual units were stacked together to form the ResNet architecture. ResNet architectures come in many forms depending on the number of residual units present and the variations in number of layers. ResNet architecture's success was influenced by the application of identity shortcuts in which the value of output identity mimics the input values identity. Like other networks, ResNet is also composed of convolution pooling and fully connected layers. e ResNet architecture resembles VGG network architecture. However, the former is eight times deeper than the latter resulting in a greater number of learnable features. Overall, the ResNet-50 architecture considered in the study consists of 49 convolution layers and 1 fully connected layer. Table 3 represents the various characteristic features of the adopted pretrained networks.

Results and Discussion
In this section, the performance of the pretrained models (AlexNet, VGG-16, GoogLeNet, and ResNet-50) for misfire detection in IC Engine is evaluated. A total of four experiments were carried out based on variations in train-test split ratio, optimizer (or) solver, initial learning rate, and batch size. e overall experimentation was carried out in the desktop version of Matlab 2019b using the deep learning toolbox and transfer learning package. e detailed experimental study is explained as follows.

Effect of Train-Test Split Ratio.
Train-test split ratio is the process of dividing the collected data set into two subsets, namely, training data and testing data. Training data are utilized to train our pretrained network while the testing data are used to evaluate the trained model for the custom data set. In the present study, five train-test split ratios were experimented for all the pre-trained networks by fixing default values for certain hyperparameters such that uniform evaluation is carried out. e hyperparameters like solver (SGDM), batch size (10), and learning rate (0.0001) were kept to default values to identify the best train-test split ratio. Table 4 describes the performance of various pretrained networks for varying train-test split ratio.
From Table 4, one can infer that the performance of each pretrained network varies with a change in the train-test split ratio. e observations made from Table 4 state that AlexNet achieved a maximum accuracy of 90.70% for the train-test split ratios of 0.7. However, GoogLeNet and ResNet-50 achieve higher classification accuracy of 96.00% and 93.60% for train-test ratio of 0.75 and 0.8, respectively. Additionally, one can observe that VGG-16 pretrained network achieved 97.30% classification accuracy for 0.85 train-test split ratio. Computing the overall accuracy of the pretrained networks, it can be observed that VGG-16 produces the best classification accuracy of 95.76%.

Effect of Optimizers.
Optimizers or solvers are the algorithms that are adopted during the training process such that the loss value is minimized to achieve an improved performance of the model. In the present study, an experimentation was carried out by varying the solvers, namely, stochastic gradient descent (SGDM), adaptive moment estimation (Adam), and root mean square propagation (RMSprop) to evaluate the performance of the model. e best performing train-test split ratio for each model was fixed based on the experimental results depicted in Section 5.1, i.e., 0.7 for AlexNet, 0.75 for GoogLeNet, 0.8 for ResNet-50, and 0.85 for VGG-16. Table 5 depicts the performance of pretrained models for different solvers and load conditions. Changing the optimizers can have an impact on the performance of the pretrained networks which is evident from Table 5. e observations state that the adoption of SGDM solver produced better classification accuracy for three pretrained networks, namely, AlexNet with 90.70%, VGG-16 with 97.30%, and GoogLeNet with 96.00%. However, ResNet-50 achieved a maximum accuracy of 97.20% for Adam optimizer. GoogLeNet displayed poor performance for RMSprop optimizer.

Effect of Learning Rate.
Learning rate is one critical parameter that monitors the change in the training model with respect to the estimated error during every instance of model weight upgradation. Selecting the best learning rate Computational Intelligence and Neuroscience can be challenging since small learning rate will elevate computational time while higher learning rates will result in improper training. In the present study, three values of learning weight like 0.001, 0.0001, and 0.0003 were used to evaluate the performance of the model. e other hyperparameters like train test ratio and optimizer are fixed for the pretrained networks that are depicted as follows: AlexNet (0.7 train-test split, SGDM solver), VGG-16 (0.85 train-test split, SGDM solver), GoogLeNet (0.8 train-test split, SGDM solver), and ResNet 50 (0.75 train-test split, Adam solver). Table 6 depicts the performance of pretrained models for different learning rates.
From Table 6, one can infer that each pretrained network's performance varies with a change in learning rate. e observations state that AlexNet and VGG-16 achieved a maximum accuracy of 94.00% and 98.70% for 0.0003 learning rate, respectively. However, GoogLeNet and ResNet-50 achieved higher classification accuracy of 98.00% and 97.60% for learning rates 0.001 and 0.0001, respectively. Higher classification accuracy states that the model has learnt the features well and that the error value is minimal.

Effect of Batch Size.
Batch size represents the number of samples that are propagated into a training network prior to model weight upgradation. e performance of the pretrained models in the present study is evaluated for different batch sizes (8,10,16,24, and 32) with the selected best performing hyperparameters based on the experimental results mentioned in previous sections. e other hyperparameters like train test ratio, optimizer, and learning rate are fixed for the pretrained networks that are depicted as follows: AlexNet (0.7 train-test split, SGDM solver, and 0.0003 learning rate), VGG-16 (0.85 train-test split, SGDM solver, and 0.0003 learning rate), GoogLeNet (0.8 train-test split, SGDM solver, and 0.001 learning rate), and ResNet 50 (0.75 train-test split, Adam solver, and 0.0001 learning rate). Table 7 depicts the performance of pretrained models for different batch sizes and load conditions. From Table 7, one can observe that there are minimal changes incurred by changing the batch size. However, all the pretrained networks resulted better classification accuracy when the batch size selected was 10. Among all the networks, VGG-16 produced an overall accuracy of 98.08%. Increase in batch size will reduce the training time, thereby leading to accelerated training progress. However, the generalization capability of the network reduces with higher batch sizes.
us, in the present study, batch size 10 is suggested as the optimal value for training the pretrained networks.

Comparative Study of Pretrained Models.
In the present section, the performance evaluation of pretrained networks is discussed. Based on the experimental results obtained Every pretrained network performs differently for the learning rates considered. e values highlighted in bold signify the best classification accuracies displayed by the pretrained networks for the variations in learning rate. e best performing network with highest overall classification accuracy is presented in the last column.   from previous sections, the optimal hyperparameters that enhance the performance of the pretrained models were identified. e list of best hyperparameters that produced an improved performance of pretrained models is provided in Table 8. e comparative study on the performance of pretrained networks with best hyperparameters is depicted in Table 9.
From Table 9, one can infer that VGG-16 established the utmost performance for the optimal hyperparameters selected. Based on the overall classification accuracy, superior classification, and lower computational time consumed, one can suggest VGG-16 for misfire detection in IC engines. e training progress and confusion matrix of best performing networks are presented in Figures 8 and 9, respectively.
From the training progresses represented in Figure 8, one can observe that the training process reaches saturation post 18 epochs. e sign of saturation represents that the VGG-16 network is trained effectively for the given IC engine misfire data set. e overall loss during the training progress for all the networks has drastically reduced representing the selection of optimal hyperparameters. Additionally, Figure 9 depicts the confusion matrix of VGG-16 architecture for misfire detection in IC engine. Confusion matrix in general depicts the performance level of a    Bolded values represent the best classification accuracy obtained by the individual pretrained networks for experimentation with different batch sizes. e pretrained network with best overall classification accuracy is represented in the final column. particular model or algorithm. e evaluation of the confusion matrix is carried out in regards to the instances classified in the major diagonal. e elements present in the major diagonal represent the correctly classified instances while the other nondiagonal elements depict the misclassified instances. e observations made from the confusion matrices state that VGG-16 architecture produced accurate classification accuracy for all conditions without any misclassifications except c4 mis (one instance is misclassified as c3 mis). Misclassification can occur due to various reasons like noise interruption, poor signal quality, and similarity between acquired signals. Lack of misclassification instances infers that the network has learnt all the features effectively and that the loss during training is negligible. us, from the observation made, VGG-16 is suggested as the best performing network for misfire detection in the IC engine. Additionally, VGG16 architecture consumed a computational training time of 2400 seconds in a low hardware system comprising of 8 GB RAM with no graphical card. However, the usage of sophisticated hardware systems (128 GB RAM with graphical card) can drastically reduce the computational time while elevating the capital cost of the system (up to 4 times).

Comparative Study with Other State-of-the-Art Works.
A comparative study is carried out in this section to display the superiority of the proposed technique with other stateof-the-art works presented in the literature. Table 10 displays the performance comparison of various techniques with the proposed technique. From Table 10, one can infer that the proposed method outclassed every other state-of-the-art works by displaying a classification accuracy of 98.7%. Classifier k nearest neighbor (KNN) exhibited the secondbest classification accuracy with 95.8% next to that support vector machines (SVM), logistic model tree (LMT), and K-star algorithm with the value of 91.20, 89.40, and 82.60%, respectively.

Conclusion
In the present paper, four pretrained deep learning models, namely, AlexNet, VGG-16, GoogLeNet, and ResNet-50, were applied to diagnose the misfire in IC engine form vibration plots acquired. Four fault conditions (c1 mis, c2 mis, c3 mis, and c4 mis) and one normal condition were considered in the study. e pretrained networks are composed of CNN layers that perform an integrated approach of feature extraction, selection, and classification, thereby formulating an end-to-end machine learning approach. e pretrained networks are capable of processing the vibration plots and provide accurate classification results. e experimental results enumerate that the adopted networks are capable of learning complex features and produce convincing classification results for misfire detection in the IC engine. Various hyperparameters like traintest split ratio, optimizer, learning rate, and batch size were altered, and the optimal hyperparameters were identified for all the networks. VGG-16 was the best performing networks with 98.70% accuracy over AlexNet (94.00%), GoogLeNet (98.10%), and ResNet-50 (97.60%). Based on the comparative studies, VGG-16 is selected as the best performing network among the other networks considered in the study and is suggested for real-time application of misfire detection in IC engines. As a future scope, the position of the accelerometer and its effectiveness in classifying the cylinder correctly can be studied. Also, the proposed fault diagnosis system can be installed in real-time operation to perform onboard diagnosis.

Data Availability
e data used to support the findings of this study are included within the article and further data or information can be obtained from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.

12
Computational Intelligence and Neuroscience