Research on the application of artificial intelligence method in automobile engine fault diagnosis

In recent years, the development of artificial intelligence (AI) has made great progress. It has been widely and deeply applied in the intelligent fault diagnosis of large machinery such as complex mechanical systems, rotating machinery, large fans, and generator sets, and has achieved good diagnostic results. The original expert system and Bayesian network enabled the intelligent system to possess not only reasoning ability, but also expert knowledge, so it has been widely used in engineering [1, 2]. However, there are many types of faults in automobile engines and they are related to each other, making it difficult to summarize knowledge, so the application of expert systems encounters a 'bottleneck'. Therefore, artificial intelligence began to find ways to allow machines to learn knowledge on their own. In recent years, SVM and neural networks have been widely used in the field of machine learning [3, 4].

Artificial neural network is a 'black box', and its non-linear mapping ability is often applied to pattern recognition. Liang Feng et al [5] used BP neural network to diagnose the faults of high-pressure common-rail electronically controlled diesel engines. In view of the faster response and faster convergence of the RBF network than the BP network, Lu Huaimin et al [6] used the RBF network to realize the diagnosis of EFI engine faults. The PNN network responds equally quickly, and on the basis of the RBF network, competing neurons are added to make the output more accurate. It also has obvious advantages in processing large amounts of data. Therefore, Zhang Xixi and Du Zhendong et al [7, 8] used PNN to realize the fault diagnosis of motor bearing and plunger pump respectively, and obtained better classification results. Furthermore, Murphey YL et al proposed an incremental learning neural network that can realize offline and online misfire fault diagnosis [9]. Compared with BP neural network, it has the ability of online incremental learning, and does not forget the existing expert knowledge of the system, and has better fire detection capabilities.

In the case of fewer learning samples, the SVM classification method has stronger adaptability, better classification ability and higher computational efficiency than the neural network classification method [10, 11]. Xu Yuxiu et al [12] extracted the time-domain energy value of the vehicle engine vibration signal according to the cycle of the crankshaft angle, and used SVM to classify the fault features. The results proved that the method is feasible. Liu Changyuan et al [13] also proposed an engine fault diagnosis method based on Twin Support Vector Machine (TWSVM), which can improve the pertinence and effectiveness of engine fault diagnosis.

The combination of multiple algorithms for engine fault diagnosis is a way to improve accuracy. Ilkivová M R et al [14] compared the Kalman filter autoregressive linear model with the nonlinear model combined with the RBF neural network, and the results showed that the nonlinear model has a better misfire diagnosis effect. According to the theory of multi-sensor information fusion, Qiao Xinyong et al [15] established an integrated neural network information fusion model to realize the diagnosis of cylinder misfire faults, and the accuracy and reliability were improved. Chen J [16] used genetic algorithm to extract the optimal amplitude characteristics of the vibration signal on the cylinder surface, and input it into the neural network to realize fault detection and fault type and degree discrimination. The scattering rate function is also used to extract the best phase characteristics of the signal for fault location.

Yuan RD et al [17] pointed out that the misfire fault diagnosis algorithm based on particle swarm optimization support vector machine (PSO-SVM) will fall into the local optimal situation. To solve this problem, an improved particle swarm optimization algorithm was proposed to support vector machine optimization, compared with PSO-SVM and GA-SVM algorithm, this diagnosis method has higher accuracy. In recent years, as an important branch of artificial intelligence machine learning, deep learning methods have many successful application cases in the fault diagnosis of complex systems [18–20]. However, the structural framework of deep learning methods is more complex than traditional shallow neural networks, and the amount of calculations larger, longer learning and training time. For fault classification problems with fewer patterns, shallow neural networks are more applicable. In addition to the above-mentioned cases where some monitoring parameters and the method of combining intelligent algorithms, there are many other monitoring parameters for automobile engines. Therefore, it is necessary to study the application of new diagnostic parameters and the methods of combining multiple algorithms to form a fast and effective artificial intelligence fault diagnosis method.

According to the two fault diagnosis methods of SVM and PNN, the collected vibration signals of automobile engine under normal and abnormal conditions are processed in two different ways: one is obtained through spectrum analysis and calculation of time-domain statistics to describe the working status. The feature parameter is used as the learning sample of the CV-SVM model; the other is divided into the time domain, and the original vibration acceleration signal is directly divided into multiple pieces of signal data of equal length as the input vector of the PSO-PNN model. In this way, two fault diagnosis models are designed and established, and they are applied in automobile engine fault diagnosis. On this basis, the diagnostic effects of several diagnostic methods in practical applications are compared.

The specific method is shown in figure 1, including:

• 1.  
  Collect the vibration acceleration signal of the engine, and extract the indicators describing the working state through time-domain statistical analysis and spectrum analysis to provide learning samples for the analysis of engine vibration signals and the identification and diagnosis of fault signals.
• 2.  
  Analyze the SVM algorithm and its recognition accuracy, obtain better performance parameters, and enhance the aggregation of learning samples, thereby improving the diagnostic performance of the classifier.
• 3.  
  By utilizing PNN's learning characteristics of large-volume samples, the original vibration signal is directly input into the neural network. And combine PNN with particle swarm algorithm to design and optimize a multi-class classification model based on vibration acceleration signal to ensure its diagnostic performance in practical applications.
• 4.  
  Through visual processing and analysis, the feasibility, accuracy and effectiveness of artificial intelligence fault diagnosis methods such as SVM and PNN are compared more clearly and effectively.

Zoom In Zoom Out Reset image size

The experiment includes engine fault setting and vibration acceleration signal extraction. The tested object is the gasoline engine of Beijing Hyundai Yakunt car (Vertical; Four-cylinder; Water-cooled; Four-stroke; Rated power is 69.9KW; Rated speed is 6000 r min⁻¹). First of all, by taking out the spark plug of a cylinder of the engine and reducing the gap to set the spark plug gap too small fault, and by unplugging the sensor signal plug to set the throttle position sensor signal interruption (Referred to as TPS signal interruption) and intake manifold absolute pressure sensor control signal interruption (Referred to as MAP signal interruption) [21]. Using the detection system composed of BBM's MKII signal collector, PAK test analysis software and PCB three-way acceleration sensor to obtain the vibration acceleration signals of multiple measuring points under the normal and three abnormal states of the automobile engine. The composition of the test system is shown in figure 2, and the acquisition frequency is 8192 Hz.

Zoom In Zoom Out Reset image size

Since the transmission path between the measuring point in the middle of the cylinder and the main excitation force is short, the vibration acceleration response signal is more direct and sensitive, so the vertical acceleration signal of the measuring point is used for analysis.

The collected vibration acceleration time-domain signal is shown in figure 3 (Speed is 1600 rpm, Load is 40 Nm). It can be seen that the vibration time-domain signal law of the engine in these four states is roughly the same, and on the surface, the degree of distinction is not significant. The time-domain waveform graph under normal conditions is relatively regular, and there is no obvious shock and fluctuation; When the spark plug gap is small, the combustion of the mixture is affected, resulting in a difference in the vibration signal of the engine casing; Similarly, when the engine is in TPS signal interruption and MAP signal interruption, the engine enters the parameter failure state. At this time, the combustion situation of the mixture is different from the normal state, and the high-frequency shock is more obvious. Judging from the waveform alone, the overall laws are very similar and difficult to distinguish, especially when the speed fluctuates and the load changes, the time-domain signal fault characteristics are not obvious. The waveforms of these states are difficult to distinguish. Therefore, it is necessary to establish an artificial intelligence model for recognition.

Zoom In Zoom Out Reset image size

SVM can use the classification hyperplane to perform pattern recognition and classification. Although it does not involve the intrinsic domain of the problem, it can provide good generalization performance in fault diagnosis. The selection of the core parameters (Penalty parameter c, kernel function parameter g) is the key to obtain the ideal classification accuracy. The idea of cross-validation (CV) can be used to obtain the optimal parameters, and can effectively avoid the occurrence of over-learning and under-learning.

In addition, in the selection of learning samples, based on the precise learning ability of SVM itself for small samples, the characteristic amplitudes in the frequency spectrum of the vibration acceleration signal in each state are extracted, and other characteristic quantities (Mean value and Energy value) obtained from the time domain signal as a training sample, which has a strong ability to characterize the working state of the engine.

3.1. Define the characteristic quantity

3.1.1. Subsubsection heading

When the fault signal is relatively weak and the fault characteristics are easily submerged in the noise, it is difficult to obtain an effective diagnosis result by pure time domain analysis. At this time, it is necessary to further combine the frequency spectrum analysis of the vibration signal [22], and use the frequency domain feature index as the original learning data for the SVM model to learn [23]. The frequency spectrum of the vibration acceleration signal of the engine in the four states is shown in figure 4.

Zoom In Zoom Out Reset image size

It can be seen from the figure 4 that the main components of the acceleration spectrum in the low frequency range in various states are the rotational frequency (26.62 Hz) and its multiplication components (53.3 Hz, 106.5 Hz, 159.7 Hz). Among them, the double frequency has the largest amplitude, which is much higher than other frequency components, which corresponds to the frequency of impact of engine gas explosive force. Comparing and analyzing the spectrograms of the four states, it is found that the amplitudes of 1, 2, 4 and 6 times of frequency spectrum have obvious characteristics: the rotational frequency amplitude of the normal state spectrum is obviously smaller than that of the three fault states, while The frequency doubling amplitude is obviously larger than that in the three failure states; the 1, 2, 4, and 6 frequency doubling amplitudes in the TPS signal interruption state spectrum are significantly smaller than the other two failure states; the frequency spectrum of the two states of spark gap being too small and MAP signal interruption are similar, but the values are still different. Therefore, the 1, 2, 4, 6-fold frequency amplitudes in the four states are extracted as the key feature learning samples of the spectrum feature quantity, so as to compose the learning sample data of the SVM model, and realizing the diagnosis of the engine fault state.

3.1.2. Time domain signal feature extraction

When the signal obviously contains simple harmonic, periodic or instantaneous impact components, the time-domain signal analysis method can be used directly to extract the characteristic index [24]. The full-spectrum diagrams under the four states are shown in figure 5, and it can be seen that the vibration energy of the vibration signal in the high frequency range is significantly different. Therefore, two characteristic quantities of the time-domain signal are extracted as needed: the Mean value and the Energy value, which are used to characterize the potential impact energy of the vibration signal in each state.

• (1)  
  Mean value: It describes the average change of the signal, characterizes the swing center of the signal amplitude, and represents the DC or static part of the signal. Used to evaluate whether the signal is stable, it is the constant component of the signal, and its expression is as follows.

  $$\begin{eqnarray}&&{X}_{a}=\displaystyle \frac{1}{N}\displaystyle \sum _{i=0}^{N-1}{X}_{i}^{2}\end{eqnarray} \tag{ 1 }$$

  In the formula, ${X}_{a}$ represents the average value of the acceleration of a certain segment of the vibration signal; N represents the number of vibrations of the segment of the vibration signal; X_i is the acceleration value of each vibration.
• (2)  
  Energy value: It is one of the important indicators of vibration signal, reflecting the intensity of its vibration. The energy of the signal is the integral value of the instantaneous power, that is, the integral of the square of the signal amplitude, and its expression is as follows.

$$\begin{eqnarray}&&{E}_{x}=\displaystyle {\int }_{-\infty }^{\infty }| x(t){| }^{2}dt\end{eqnarray} \tag{ 2 }$$

Zoom In Zoom Out Reset image size

In the formula, E_x represents the total energy in the positive and negative directions in a certain segment of vibration signal; x(t) represents the instantaneous vibration shock amount.

3.2. Implementation of SVM classifier

There are many kinds of kernel functions of SVM, such as linear kernel, polynomial kernel, two-layer perceptron kernel and radial basis (RBF) kernel. Among them, the RBF kernel can map the sample to a higher-dimensional space, and can deal with the problem when the relationship between the class label and the feature is nonlinear. In addition, the RBF kernel parameters include penalty parameters (c) and kernel function parameters (g), which have the advantage of fewer parameters compared with other kernels [25, 26]. Therefore, the RBF kernel is selected as the kernel function of SVM as follows:

$$\begin{eqnarray}&&{\rm{K}}\left({\rm{x}},{x}_{i}\right)=\exp \left(-\gamma {\left|\left|{\rm{x}}-{x}_{i}\right|\right|}^{2}\right),\gamma \gt 0\end{eqnarray} \tag{ 3 }$$

Where k(x, x_i ) is the representation method of the SVM inner product kernel, where xi is the support vector; x is the feature vector extracted from the input space; γ is the Gaussian kernel bandwidth.

By using the RBF kernel function, a classification hyperplane can be established as a decision surface, which maximizes the separation edge between positive and negative examples:

$$\begin{eqnarray}&&\mathop{\min }\limits_{\alpha }\displaystyle \frac{1}{2}\displaystyle \sum _{i=1}^{j}\displaystyle \sum _{j=1}^{l}{y}_{i}{y}_{j}{\alpha }_{i}{\alpha }_{j}K\left({x}_{i},{x}_{j}\right)-\displaystyle \sum _{j=1}^{l}{\alpha }_{j}\end{eqnarray} \tag{ 4 }$$

$$\begin{eqnarray*}&&{\rm{s}}.{\rm{t}}.\displaystyle \sum _{i=1}^{l}{y}_{i}{\alpha }_{i}=0,0\leqslant {\alpha }_{i}\leqslant c,i=1,\,\ldots ,\,l\end{eqnarray*}$$

The formula expresses that the original sample is mapped to a high-dimensional space for classification through a dual relationship. (x_i , x_j ) and (y_i , y_j ) represent different types of feature vectors, a is a parameter of the model, and the range c of a needs to be adjusted and determined by us. After obtaining the optimal solution α* of the model parameters, the bias threshold b* of the model will be calculated on this basis, and the decision function f(x) will be constructed, as follows:

$$\begin{eqnarray}&&{b}^{* }={y}_{i}-\displaystyle \sum _{i=1}^{l}{y}_{i}{{\alpha }_{i}}^{* }K({x}_{i}-{x}_{j})\end{eqnarray} \tag{ 5 }$$

$$\begin{eqnarray}&&{\rm{f}}\left({\rm{x}}\right)=\mathrm{sgn}\left(\displaystyle \sum _{i=1}^{l}{{\alpha }_{i}}^{* }{y}_{i}{\rm{K}}\left({\rm{x}},{x}_{i}\right)+{b}^{* }\right)\end{eqnarray} \tag{ 6 }$$

In the case of separable mode, SVM first classifies samples of a certain category into one class, and classifies the remaining samples into another class, so that K classes form K-SVM. In actual application, the sample to be tested will be classified as the type with the largest classification function value, and the architecture of SVM is shown in figure 6.

Zoom In Zoom Out Reset image size

3.3. Train and analysis of CV-SVM classification model

Three states with large differences in operating conditions and four states with similar operating conditions are selected as diagnostic objects. Constructing an attribute matrix: [Amplitude of one-fold frequency; Amplitude of two-fold frequency; Amplitude of four-fold frequency; Amplitude of six-fold frequency; Mean value; Energy value]. There are 15 sets of data for each state, a total of 60 sets, and the fractal dimension distribution of the six types of attribute values is shown in figure 7. It can be seen that in the formed feature vector, the values of different feature attributes are quite different, which will cause the objective function to become 'flat'. In this way, when performing gradient descent, many detours will be taken, which will increase the training time and affect the training accuracy to a certain extent.

Zoom In Zoom Out Reset image size

Therefore, in order to improve the speed and accuracy of training, it is necessary to normalize the samples, as shown in table 1.

Table 1. Comparison of different normalization methods.

Normalized mode	Accuracy of three categories (%)	Accuracy of four categories (%)
No normalization	33.3% (9/30)	25% (10/40)
[−1, 1] normalization	46.7% (13/30)	40% (16/40)
[0, 1] normalization	90.0% (27/30)	85% (34/40)

It can be seen from the table that the highest accuracy rate of the SVM classifier at this time is only 90%, because the two key parameters of SVM: the penalty coefficient (c) and the kernel function parameter (g) affect the performance of the classifier. Among them, if c is too high, it is easy to overfit; if c is too small, it is easy to underfit. The larger the g, the fewer support vectors, the smaller the g, the more support vectors, and the number of support vectors affects the speed of training and prediction. By using the K-CV method, the original data can be equally divided into K sets, and each subset of data will be used as a verification set, and the remaining K-1 subsets of data will be used as the training set, so that K models will be obtained. Using the average of the classification accuracy of the final validation set of the K models as the performance index of the classifier under this K-CV, and the final result is also more convincing [27].

The parameters c and g are predicted in the range of [−10, 10], and the best parameter combination is obtained. The selection results of the parameters c and g are shown in figure 8.

Zoom In Zoom Out Reset image size

Finally, the best parameters for the three states and the four states are c = 0.7613, g = 0.1894 and c = 4.776, g = 1.5157, respectively. The training and testing results of SVM are shown in figure 9.

Zoom In Zoom Out Reset image size

The test results of the classification and recognition of the fault state by the SVM model are shown in figure 10.

Zoom In Zoom Out Reset image size

The results show that the selection of parameters c and g directly affects the classification accuracy of SVM. For the recognition of the three states of normal state, too small spark gap, and TPS signal interruption, all 30 test samples are correctly diagnosed.

After adding the failure state of the MAP signal interruption, the main reason for the abnormal combustion of the mixture caused by this failure is the concentration of the mixture, which is similar to the phenomenon caused by the TPS signal interruption, and the vibration response signal and frequency spectrum are relatively similar. Therefore, the accuracy of the diagnostic model is slightly reduced, and the accuracy rate is only 97.5%. It can be seen from the classification scatter diagram that the errors are mainly concentrated in the failure of MAP signal interruption. The TPS signal interruption state and the MAP signal interruption state overlap and are difficult to distinguish. On the other hand, the time for model diagnosis is 0.00792 s and 0.008414s respectively, which is short.

In the realization of pattern classification, PNN has a good advantage in processing large-volume samples. The collected original vibration acceleration time-domain signal can be directly divided into the time domain, and multiple sets of one-dimensional signal data can be intercepted and used as the training sample of the network model. On this basis, a PNN fault identification model is established, and the PSO parameter optimization is used to train and optimize the model.

4.1. PNN model

The PNN network is a four-layer network (Input layer, Pattern layer, Summation layer and Output layer), and its structure is shown in figure 11.

Zoom In Zoom Out Reset image size

In this experiment, take the identification of three fault states as an example. The input layer combines N segments of continuous vibration signals into a multi-dimensional vector X = {X₁ ,X₂ ,...,X₁₀₀₀} and inputs it into the neural network as the original learning data. The input vector length is 1000. Then, the input vector passes through the mode layer and is weighted:

$$\begin{eqnarray}&&{Z}_{i}=x{\omega }_{i}\end{eqnarray} \tag{ 7 }$$

In the formula, Z_i is the weighted input vector, and the weighting coefficient is W_i .

In the mode layer, one input sample corresponds to one neuron, and each neuron will form a node center. If there is a diagnosis data input, the Euclidean distance from these centers will be calculated and a scalar value will be obtained, as follows:

$$\begin{eqnarray}&&{\varnothing }_{ij}\left(x\right)=\displaystyle \frac{1}{{\left(2\pi \right)}^{\displaystyle \frac{1}{2}}{\sigma }^{d}}{e}^{-\tfrac{\left(x-{x}_{ij}\right){\left(x-{x}_{ij}\right)}^{T}}{{\sigma }^{2}}}\end{eqnarray} \tag{ 8 }$$

${\varnothing }_{ij}\left(x\right)$ represents the input/output relationship between the jth neuron of the i-th pattern in the mode layer, and the sensitivity σ of pattern classification is the smoothing factor [28, 29]. The scalar value obtained from the mode layer will enter the summation layer. The output scalar value of the model layer neurons of the same type will be weighted and averaged, and the probability estimation will be performed:

$$\begin{eqnarray}&&{v}_{i}=\displaystyle \frac{\displaystyle {\sum }_{j=1}^{L}{\varnothing }_{ij}}{L}\end{eqnarray} \tag{ 9 }$$

Where v_i is the output of the i-th category, and L is the number of neurons in the i-th category.

Finally, the output layer receives the probability density functions from the summation layer, and extracts the largest one as the output in a competitive manner to complete fault diagnosis.

4.2. Training of PNN and optimization based on PSO

The classification accuracy of a probabilistic neural network is mainly related to the number of training samples and the spread parameter. Moreover, enough samples are needed to ensure the accuracy of the network model. On the other hand, the selection of the spread constant is also very critical. If the spread value is too large, the effect of each training sample will become larger. If the spread value is too small, the discrimination between training samples will become smaller [30]. Therefore, selecting the appropriate spread value can achieve the best network operating speed and diagnostic effect.

4.2.1. Determination of sample size

Different from the learning of small samples by SVM, PNN can directly input the original vibration signal into the probabilistic neural network to construct a fault diagnosis model. The effects of the three state diagnosis models in different capacities of training samples are shown in table 2.

Table 2. Comparison of classification results of different capacities of training samples.

Training sample size	Accuracy ratio(%)
60 × 1000	83.3% (25/30)
75 × 1000	90.0% (27/30)
90 × 1000	96.7% (29/30)

It can be seen from the data that when the training sample size increases from 60 groups to 90 groups, the accuracy of PNN increases from 83.3% to 96.7%. Therefore, selecting the original vibration signal data with a reasonable capacity is an important guarantee for the PNN network to diagnose engine faults, but too many samples will reduce the operating speed. Therefore, it is also necessary to optimize the selection of PNN's core parameters (spread) to improve the fault diagnosis speed while improving its performance.

4.2.2. PSO-based parameter (spread)optimization

Finding the best spread value is necessary for the establishment of the PNN model. In terms of parameter selection, compared with the trial method and the empirical method, the particle swarm optimization (PSO) is used to directly find the optimal value of spread, which is more accurate and efficient [31].

Set the individual numerical range N of the initial particle swarm P(x) to [0, 20], the maximum particle velocity Vmax to 10, and the maximum evolutionary algebra K to 100. The particles will imitate the predation of a flock of birds, randomly select the initial path, and extend to n-dimensional space. The position of the particle in the n-dimensional space is expressed as a vector Xi = (x₁ , x₂ ...x_n ), and the flying speed is expressed as a vector Vi = (v₁ , v₂ ,...,v_n ). Each particle has an fitness value determined by the objective function, and knows the best position (Pbest) it has found so far and the current position Xi, which is the particle's own flight experience. In addition, each particle also knows the best position (gbest) found by all particles in the entire group so far, which can be regarded as the experience of the particle's companion. Particles use their unique memory methods to communicate and share the best experience of themselves and their companions, and finally determine an optimal spread individual [32], which can be boiled down to the following two formulas.

$$\begin{eqnarray}&&{V}_{id}^{k+1}=\omega {V}_{id}^{k}+{c}_{1}{r}_{1}({P}_{id}^{k}-{X}_{id}^{k})+{c}_{2}{r}_{2}({P}_{id}^{k}-{X}_{id}^{k})\end{eqnarray} \tag{ 10 }$$

$$\begin{eqnarray}&&{X}_{id}^{k+1}={X}_{id}^{k}+{V}_{id}^{k}\end{eqnarray} \tag{ 11 }$$

In the formula, ω is the inertia weight factor; d = (1, 2,...,n); i = (1, 2,...,n); K is the current iteration number; V_id is the velocity of the particle; c₁ and c₂ are non-negative constants called acceleration factors; r₁ and r₂ are random numbers distributed in the interval [0, 1].

In 100 iterations, the optimization process of the dispersion constant (spread) value and the fitness evolution curve are shown in figure 12.

Zoom In Zoom Out Reset image size

The lower the fitness value, the better the spread value. It can be seen from the figure that the spread value in the range of [0.8891, 2.234] has better adaptability in the three states, and the spread value in the range of [2.54, 4.696] has better adaptability in the four states. The best spread values corresponding to the three-state diagnosis and the four-state diagnosis are 0.8891 and 2.986. At this time, the best spread value is assigned to the network for prediction, and the network has the highest calculation and classification accuracy.

Select 90 and 120 (30 in each state) training samples respectively for network training. The test samples are 30 and 40 (10 in each state), and the classification results are shown in figure 13.

Zoom In Zoom Out Reset image size

When distinguishing the samples to be diagnosed in the three states with large differences in fault conditions, all the 30 groups of test samples were correctly identified. After adding the failure state of the MAP signal interruption whose failure type is similar to the TPS signal interruption state, 4 of the 40 randomly selected verification samples failed to be correctly identified, with an accuracy rate of 90%, and the errors are mainly concentrated in the state of the MAP signal interruption and the state of TPS signal interruption. In terms of diagnosis speed, the two diagnosis times were 0.04126 s and 0.04713 s respectively. Therefore, as long as the training samples are sufficient, the PSO-PNN model can directly learn the original engine vibration signal data, and realize efficient and stable fault state identification.

In order to prove the fault identification performance of the above-mentioned CV-SVM and PSO-PNN two parameter optimization methods, it is compared with the non-optimized SVM method, the non-optimized PNN method and the traditional BP neural network method. The diagnosis results are shown in table 3. The results show that the two methods, CV-SVM and PSO-PNN, obtained by combining with the parameter optimization algorithm, can accurately diagnose the normal state, the spark gap is too small state and TPS signal interruption state according to the characteristic quantity of the engine vibration signal, and the accuracy rate is 100%. For the identification of the four states, the accuracy rates are 97.5% and 90%, respectively. The diagnostic accuracy rates of the non-optimized SVM and non-optimized PNN methods have decreased. The Identification accuracy rates of the three states and the four states are only 90%, 85% and 83.3%, 72.5% respectively. The accuracy rate is significantly lower than that of CV-SVM and PSO-PNN. At the same time, compared with the BP neural network, the fault learning ability and fault recognition performance of the SVM and PNN models are stronger, and the accuracy rate is higher. From another perspective, when the MAP signal interruption, which has a similar fault phenomenon as the TPS signal interruption, is added, the diagnostic accuracy of these methods has decreased. Among them, the least decrease in accuracy is the SVM model, and the diagnostic performance of SVM is more stable. However, the SVM modeling requires artificial feature selection and preprocessing. Moreover, the selection of feature quantities will have a great impact on the diagnostic accuracy of the SVM model, which relies more on manual experience. In terms of diagnosis speed, whether it is SVM, PNN method or BP network, the diagnosis time does not exceed 0.05 s, and fault classification and identification can be completed quickly.

The time domain and frequency spectrum of the vibration acceleration signal of the engine body surface are analyzed in detail. The classification methods based on CV-SVM and PSO-PNN are applied to automobile engine fault diagnosis. The results show that both methods can effectively diagnose the fault state we set. In terms of diagnosis time, accuracy and stability, SVM performs higher, but the PNN method does not need to manually analyze and process the original vibration signals, which simplifies the modeling process and is more conducive to the realization of automation and intelligence. Both of these methods can be applied to other types of engines by updating the training samples, and the reconstruction process of the diagnostic model is convenient and will not cause damage to the engine. They are effective methods for diagnosing engine faults.

In future work, this method will continue to be tested on more different types of engine fault identification. In addition, this sensitive, convenient and efficient fault diagnosis method can use artificial intelligence algorithms to detect and identify faults based on engine surface vibration signals without disassembling the engine, which cannot be done manually. Combining this method with real-time monitoring technology can detect engine failures in time, detect abnormalities even before the engine is seriously damaged, and predict the failures. Now we have continued to study this aspect in depth. Next, we will try to realize real-time acquisition of vibration signals, wireless signal transmission, intelligent diagnosis of vibration signals, and early warning of fault conditions, so that this fault diagnosis method can achieve higher value and functions.

This research has been supported by the Natural Science Foundation of Guangdong Province, China (2018A030313947 and 2019A1515011779), and Guangdong Provincial College Youth Innovative Talents Program(2019KQNCX066), as well as Science and Technology Planning Project of Guangzhou, China(201803030041 and 201905010007).

The data that support the findings of this study are available upon reasonable request from the authors.