Diesel engine fault diagnosis method, system, equipment and medium

Abstract

The invention discloses a diesel engine fault diagnosis method, system and device and a medium, and relates to the technical field of diesel engine fault diagnos.The method comprises the steps that vibration signal data of a diesel engine are collected; constructing a fault diagnosis model in parallel based on the improved one-dimensional convolutional neural network; vibration signal data are input into a fault diagnosis model, time domain features and frequency domain features of vibration signals are extracted through a dual-channel convolution layer, the time domain features and the frequency domain features are input into a gating circulation unit layer after passing through a three-layer pooling layer for time sequence modeling, and dynamic change information of the time domain features and the frequency domain features is obtained; key information of dynamic change information is extracted through an attention mechanism layer and fused, the key information passes through a global maximum pooling layer and then is input into a Softmax layer, and a classification result is output to diagnose faults of the diesel engine; according to the method, key information of dynamic change information is extracted by introducing a gating circulation unit and an attention mechanism, and the fault detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to a diesel engine fault diagnosis method, system, equipment and medium. Background Art

[0002] As an internal combustion drive machine, the core task of a diesel engine is to convert the heat energy generated by oil into mechanical energy. It is an indispensable power source in modern society. The structure of a diesel engine is relatively complex, and the operating conditions are constantly changing. These factors often lead to failures and irreparable losses. Analyzing the structural composition, working principle and fault mechanism of a diesel engine is crucial to completing the fault diagnosis task. A fault diagnosis method based on vibration analysis has been proposed and widely used. The results obtained by using a computer to simulate the vibration signal of a diesel engine for analysis are not very accurate.

[0003] As a classic type of deep learning model, CNN has shown excellent recognition performance in many fields. For example, a fault diagnosis model based on one-dimensional densely connected CNN is proposed, which can keep the accuracy of wind turbine gearbox fault diagnosis at a high level. A deep CNN model designed by combining wide and narrow convolution kernels is also proposed, which shows accurate diagnosis results in rolling bearing fault diagnosis. A fault diagnosis model based on one-dimensional CNN is proposed for common bearing faults. This model integrates feature extraction and classification, and demonstrates the effectiveness of one-dimensional CNN in the field of bearing fault diagnosis. A WDCNN fault diagnosis model with a large convolution kernel as the first layer of the convolution layer is also proposed. The large convolution kernel is used as a low-pass filter. The model has good noise resistance through experiments. A parallel one-dimensional convolutional neural network model is proposed, which uses large and small convolution kernels to obtain the frequency domain and time domain features of the vibration signal respectively, making full use of the characteristic information of the vibration signal and improving the diagnostic accuracy of the model. It is also proposed that on the basis of the deep convolutional neural network, a gated recurrent unit is introduced to solve the gradient explosion problem in the neural network, and good generalization and noise resistance are achieved under high noise background. The deep one-dimensional convolutional neural network and dual-channel information fusion method are used for diesel engine fault diagnosis, and good results have been achieved.

[0004] However, existing diesel engine fault diagnosis methods mainly face the limitations of traditional time-frequency analysis methods in processing non-stationary and complex signals, the challenges of adaptive decomposition methods such as EMD, EEMD and VMD in modal aliasing, boundary effects and parameter selection, and the dependence of machine learning and artificial neural networks on manual operations and large data volume requirements in feature extraction and processing. At present, when deep learning models, especially recurrent neural networks, are used for fault diagnosis, gradient vanishing or explosion, as well as long-term dependency problems, are prone to occur when processing long sequence data, which reduces the accuracy and efficiency of fault diagnosis. Summary of the invention

[0005] In view of the shortcomings of the prior art that gradient vanishing or exploding, as well as long-term dependence problems, reduce the accuracy and efficiency of fault diagnosis when processing long sequence data, the present invention proposes a diesel engine fault diagnosis method, system, equipment and medium. By proposing a fault diagnosis model based on an improved parallel one-dimensional convolutional neural network, high-precision identification of diesel engine faults is achieved, thereby solving the problems existing in the prior art.

[0006] A diesel engine fault diagnosis method comprises the following steps:

[0007] Collect vibration signal data of diesel engines;

[0008] A fault diagnosis model is constructed based on the improved parallel one-dimensional convolutional neural network; the vibration signal data is input into the fault diagnosis model, and the time domain features and frequency domain features of the vibration signal are extracted through the dual-channel convolution layer, and the time domain features and frequency domain features are respectively input into the gated recurrent unit layer after three layers of convolution pooling for time series modeling to obtain the dynamic change information of the time domain features and frequency domain features; the key features related to diesel engine faults in the dynamic change information of the time domain features and frequency domain features are respectively extracted through the attention mechanism layer, and the two extracted key features are spliced ​​and fused, and the fused features are input into the Softmax layer after passing through the global maximum pooling layer, and the classification results are output;

[0009] The diesel engine fault is diagnosed according to the classification results.

[0010] Furthermore, the vibration signal data is subjected to noise reduction processing before being input into the fault diagnosis model, specifically comprising the following steps:

[0011] The minimum envelope entropy is used as the fitness function of the dung beetle optimization algorithm DBO, and DBO is used to optimize two important parameters of the variational mode decomposition algorithm VMD.

[0012] The optimized VMD is used to adaptively decompose the vibration signal to obtain multiple modal components;

[0013] Signal reconstruction of multiple modal components is performed according to the relevant kurtosis criterion;

[0014] The signal-to-noise ratio (SNR), mean square error (MSE) and mean absolute error (MAE) are used to evaluate the noise reduction of the reconstructed signal. The larger the SNR value, the smaller the MSE and MAE values, and the less noise signal there is in the reconstructed signal.

[0015] The vibration signal data after noise reduction is input into the fault diagnosis model.

[0016] Furthermore, the time domain features and the frequency domain features are respectively subjected to three layers of convolutional pooling and then input into the gated recurrent unit layer for time series modeling to obtain dynamic change information of the time domain features and the frequency domain features; specifically, the following steps are included:

[0017] By introducing a reset gate r t and update gate z t Capture the long-term dependencies in the time series; the reset gate r t It is expressed as:

[0018] r t =σ(W r [h t-1 , X t ]),

[0019] Where W r represents the weight parameter that can be learned; h t-1 Indicates the hidden state of the previous moment; X t represents the current input signal characteristics, and t represents the current time;

[0020] The update gate z t It is expressed as:

[0021] Z t =σ(W Z [h t-1 , X t ]),

[0022] Where W Z Represents the weight parameter that can be learned; σ represents the activation function; there is an update gate at each moment, and the output value is calculated by the Sigmoid function. When the value is close to 1, it means that the convolutional neural network remembers the previous information, and when the value is close to 0, it means that the convolutional neural network ignores the previous state information;

[0023] Reset Gate t By controlling the degree of integration of the current moment’s input information and the previous moment’s output, the candidate state is obtained. It is expressed as:

[0024]

[0025] Where W h Represents the parameters that can be learned;

[0026] Candidate status With update gate z t Combined, the information output at the final moment is h t , which is expressed as:

[0027]

[0028] Furthermore, extracting key features related to diesel engine failure from the dynamic change information of time domain features and frequency domain features through the attention mechanism layer specifically includes the following steps:

[0029] The characteristic information F∈R of the one-dimensional time series signal C×1×W In the input attention mechanism layer CBAM, maximum pooling and average pooling operations are performed in the channel attention module;

[0030] The feature information after pooling is changed from C×1×W to C×1×1, and then the number of channels is compressed and then expanded, and the two activated results are output through the Relu activation function;

[0031] The two output results are added element by element and then a Simogid activation function is used to obtain the channel attention M. C ∈R C×1×1 ; Then the feature information F′ adjusted by the channel attention module is expressed as:

[0032]

[0033] Perform maximum pooling and average pooling operations on the feature information F′ along the channel dimension, and then concatenate the pooled feature information along the channel dimension to generate spatial attention weights;

[0034] Apply the Sigmoid activation function to the generated spatial attention weights, limit the weights between 0 and 1, apply the spatial attention weights to the original feature information, and weight the features of each spatial position to obtain the spatial attention M. S ∈R C×1×W ; Then the characteristic information F″ after CBAM adjustment is expressed as:

[0035]

[0036] Furthermore, the collecting of vibration signal data of the diesel engine specifically includes selecting a cylinder head of the diesel engine as a measuring point, and installing a vibration sensor at the measuring point to collect acceleration information generated by vibration during operation of the diesel engine to obtain a vibration signal.

[0037] The present invention also provides a diesel engine fault diagnosis system, comprising:

[0038] An acquisition module, used for acquiring vibration signal data of a diesel engine;

[0039] The data processing module is used to build a fault diagnosis model based on the improved parallel one-dimensional convolutional neural network; the vibration signal data is input into the fault diagnosis model, and the time domain features and frequency domain features of the vibration signal are respectively extracted through the dual-channel convolution layer, and the time domain features and frequency domain features are respectively input into the gated recurrent unit layer for time series modeling after being subjected to three-layer convolution pooling to obtain the dynamic change information of the time domain features and frequency domain features; the key features related to the diesel engine fault in the dynamic change information of the time domain features and frequency domain features are respectively extracted through the attention mechanism layer, and the two extracted key features are spliced ​​and fused, and the fused features are input into the Softmax layer after passing through the global maximum pooling layer, and the classification results are output;

[0040] The diagnosis module is used to diagnose the fault of the diesel engine according to the classification result.

[0041] The present invention also proposes a diesel engine fault diagnosis computer device, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the diesel engine fault diagnosis method when executing the computer program.

[0042] The present invention further proposes a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the diesel engine fault diagnosis method.

[0043] The present invention provides a diesel engine fault diagnosis method, system, device and medium, which have the following beneficial effects:

[0044] The present invention builds an improved parallel one-dimensional convolutional neural network fault diagnosis model, constructs a dual-channel multi-layer convolution and pooling network structure, uses a gated recurrent unit to optimize the network model, reduces the gradient explosion problem that may be caused by the deep network structure, introduces an attention mechanism, solves the forgetting problem that may occur in the gated recurrent unit during the back propagation process of model training, adopts a global maximum pooling layer to prevent overfitting that may occur during the training process, and finally uses a Softmax layer for output classification, thereby completing the construction of the fault diagnosis model; the method introduces a gated recurrent unit and an attention mechanism, solves the gradient problem and the long-term dependency problem through the reset gate and update gate mechanism of the GRU, and uses the attention mechanism to improve the model's attention to key information, reduce the number of parameters, and enhance the model's generalization ability and the ability to process complex data, thereby significantly improving the accuracy and efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of the test bench structure in an embodiment of the present invention;

[0046] Figure 2This is a flow chart of a signal noise reduction method based on DBO algorithm to optimize VMD parameters in an embodiment of the present invention;

[0047] Figure 3 Schematic diagram of the receptive field in an embodiment of the present invention;

[0048] Figure 4 A parallel one-dimensional convolutional neural network structure diagram in an embodiment of the present invention;

[0049] Figure 5 Schematic diagram of the structure of the convolutional attention module in an embodiment of the present invention;

[0050] Figure 6 A schematic diagram of the model structure of an improved parallel one-dimensional convolutional neural network in an embodiment of the present invention;

[0051] Figure 7 A schematic diagram of a fault diagnosis process of an improved parallel one-dimensional convolutional neural network in an embodiment of the present invention;

[0052] Figure 8 Schematic diagram of the accuracy change curve of each model validation set in the embodiment of the present invention;

[0053] Fig. 9 Schematic diagram of the loss rate change curve of each model validation set in an embodiment of the present invention;

[0054] Fig.10 This is a schematic diagram of the accuracy of the test set in an embodiment of the present invention; Fig.10 a in the figure is a schematic diagram of the test set accuracy of the improved parallel one-dimensional CNN fault diagnosis model; Fig.10 b in the figure is a schematic diagram of the test set accuracy of the parallel one-dimensional CNN; Fig.10 The c in the figure is a schematic diagram of the test set accuracy of the single-channel small convolution kernel CNN; Fig.10 d in the figure is a schematic diagram of the test set accuracy of a single-channel large convolution kernel CNN;

[0055] Fig.11 A schematic diagram of resampling in an embodiment of the present invention;

[0056] Fig.12 It is a schematic diagram of an iterative curve of accuracy and loss rate of a fault diagnosis model in an embodiment of the present invention;

[0057] Fig.13 Schematic diagram of the verification set accuracy of each model in the training process in an embodiment of the present invention;

[0058] Fig.14 Schematic diagram of the validation set loss rate of each model during the training process in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0060] The present invention proposes a diesel engine fault diagnosis method, such as Figure 7 As shown; the method specifically comprises the following steps:

[0061] S1. Collect vibration signals.

[0062] During the operation of a diesel engine, the interaction and movement between components will generate a wealth of vibration data, which carries a lot of key characteristic information that can reveal the performance status of the diesel engine. By analyzing these vibration data, the operating status of the diesel engine can be effectively monitored without disassembling it, which is of great significance for the safe operation of the diesel engine under complex and harsh conditions. Building a diesel engine vibration signal acquisition test bench is an effective way to obtain such data. The test bench structure is as follows: Figure 1 As shown in the figure, such a test bench can capture and analyze vibration information, and then identify potential operating problems, which helps to implement preventive maintenance measures, reduce the occurrence of failures and extend the service life of the engine. The vibration signal of the diesel engine is collected by the sensor, and the collected signal is input into the DAQ board through the sensor. In the DAQ, the vibration signal collected by the signal acquisition system is pre-processed by the signal conditioning circuit to obtain a series of continuously changing analog signals. The continuously changing analog signals are converted into a series of digital signals through the analog-to-digital converter to facilitate the processing and analysis of the digital system. The digital signal is input into the data processing module to complete the processing of the digital signal for use by the fault diagnosis module.

[0063] The vibration sensor can understand the operating status of the diesel engine and detect abnormal vibration and faults in time by monitoring the amplitude, frequency and direction of the diesel engine vibration. The vibration signal of the diesel engine is collected by the sensor. During the collection process, since the diesel engine generates vibration when working, the sensor can collect the acceleration information generated by the vibration, and after processing, the vibration signal can be obtained.

[0064] Selecting the measurement point of diesel engine vibration signal is very important for accurate monitoring and analysis of vibration. The selection of vibration measurement point should be based on the principle of the shortest path between the vibration source and the measurement point to obtain a signal containing rich system status information. At the same time, the measurement point should be convenient for installing the sensor without hindering the normal operation of the diesel engine. The combustion gas force during the operation of the unit will directly act on the cylinder head, causing the cylinder head to vibrate, so the surface of the cylinder head contains rich information on the operating status of the diesel engine, so it is most appropriate to select the measurement point at the cylinder head.

[0065] In order to better diagnose the faults of the diesel engine's operating state, it is first necessary to analyze the differences between the vibration signals in each state, and perform time domain and frequency domain analysis on the vibration signals in order to select appropriate signal processing methods and fault diagnosis methods. The diesel engine completes a cycle according to four strokes during operation, so the vibration signal it generates is periodic.

[0066] S2. Optimize VMD parameters based on the dung beetle algorithm, use the optimized VMD to adaptively decompose the vibration signal, screen and reconstruct the decomposed component signals, and obtain clearer characteristic parameters and information. The overall framework of VMD (variational mode decomposition algorithm) is based on solving a variational problem. Compared with the traditional recursive decomposition mode, VMD decomposes the signal into multiple modal components in a non-recursive form, thereby effectively reducing the modal aliasing effect. This decomposition method has good noise robustness and is helpful for signal noise reduction and feature extraction.

[0067] S2.1、VMD variational problem construction: VMD realizes signal decomposition by constructing a series of variational problems. By optimizing and solving these variational problems, the modal components of the signal can be obtained. The input signal is decomposed into k components with a certain center frequency ω k The limited bandwidth modal function u k (t), x(t) is iterated to find the optimal solution, which minimizes the sum of the estimated bandwidths of each mode and satisfies that the sum of all modes is equal to the original signal. When constructing the variational problem, the Hilbert transform is used to obtain the single-sided spectrum of the signal, and the spectrum of each mode is modulated to the corresponding estimated baseband, and the bandwidth of each mode is estimated. The construction process of the variational problem is shown in formula (1).

[0068]

[0069] In the formula, {u k}={u1,u2,…,u k} are k IMF components; {ω k}={ω1,ω2,…,ω k ) is the center frequency of each component; * is the convolution operation; is the gradient operation; t represents the time variable, which means the time series of the signal; j represents the index of the modal component, which refers to the jth modal component (IMF); st represents the constraint condition, which refers to the constraint condition on the modal component or other parameters in the variational problem; x represents the input signal or variable, which means the signal data to be decomposed.

[0070] S2.1.1. Solution of variational problem: After completing the construction of variational problem, solving the problem requires converting it from a constrained problem to an unconstrained problem. This process can be achieved by adding the quadratic penalty term α and the Lagrange multiplier λ(t). This method can ensure the accuracy of signal reconstruction even under noise interference, and strictly abide by the constraints by adopting the improved Lagrange formula. Its expression is shown in formula (2).

[0071]

[0072] Where λ represents the Lagrange multiplier, which is used to incorporate constraints into the objective function; f(t) represents a function of time t.

[0073] Then the variational problem is solved iteratively by the alternating multiplier (ADMM), and the modal component u k The solution is shown in formula (3), the center frequency ω k The solution of is shown in formula (4), and the solution of λ is shown in formula (5).

[0074]

[0075] Where n and i are the indices in the iteration process, which are used to indicate the number of iterations or steps; ω represents the frequency of the signal; f(ω) represents the spectrum of the original signal; u i (ω) represents the spectrum of the i-th modal component; λ(ω) represents the Lagrange multiplier or penalty factor, which is used to incorporate constraints into the objective function. The Lagrange multiplier is used to ensure that specific constraints are met during the optimization process; when the condition is met When the iteration stops, the variational problem is solved, and k modal components with finite bandwidth are obtained. ε represents a small positive number, which is used to determine the condition for stopping the iteration. That is, when the residual or error is less than this value, the problem is considered to have been solved.

[0076] S2.2, Dung Beetle Optimizer (DBO) simulates the behavior of dung beetles when looking for food, and searches for the optimal solution by simulating the foraging process of dung beetle populations. The basic principle of the dung beetle optimization algorithm is to search for the optimal solution by continuously updating the position of dung beetles. In each step of the algorithm, dung beetles will update their positions based on their own information and the information of their neighbors, and evaluate the objective function value based on the updated position. Through the iterative search process, dung beetles gradually gather near the optimal solution and finally converge to the optimal solution. During the rolling process, dung beetles need to keep rolling in a straight line according to celestial clues, and the position of dung beetles changes. The calculation of the position of dung beetles is shown in formula (6).

[0077]

[0078] In the formula, x i (t) represents the position information of the i-th dung beetle at the t-th iteration; k∈(0, 0.2] is the deflection coefficient; b∈(0, 1) is a constant; a represents the natural coefficient; X ω is the global worst position; Δx represents the change in light intensity.

[0079] When the dung beetle encounters an obstacle during its forward movement, it will reposition itself by dancing to obtain a new route. After obtaining the new route, the dung beetle will roll the ball to continue moving forward. The position update definition of the dung beetle is shown in formula (7).

[0080] x i (t+1)=x i (t)+tan(θ)|x i (t)-x i (t-1)| (7);

[0081] Where θ is the deflection angle; it belongs to [0, π].

[0082] The egg-laying behavior of dung beetles is restricted by the regional boundaries, which are defined as shown in formula (8).

[0083]

[0084] In the formula, X * is the best location for spawning at present; Lb * Indicates the lower limit of the spawning area; Ub * represents the upper limit of the spawning area; Lb represents the lower limit of the optimization problem; Ub represents the upper limit of the optimization problem; R = 1-t / T max , T max Indicates the maximum number of iterations.

[0085] It can be seen from formula (8) that the boundary of the spawning area changes due to the influence of R. At this time, the position of the egg ball is shown in formula (9).

[0086] B i (t+1)=X * +b1×(B i (t)-Lb * )+b2×(B i (t)-Ub * ) (9);

[0087] In the formula, B i (t) is the position information of the i-th egg ball at the t-th iteration; b1 and b2 represent two 1×D-dimensional independent random variables, and D represents the dimension of the search space. At this time, the egg ball is confined to a certain area. When the egg ball grows up to become a small dung beetle, it will also forage. In order to guide the small dung beetle to forage, its activity area needs to be divided and limited. The definition of the regional boundary is shown in formula (10).

[0088]

[0089] In the formula, X b Indicates the location of the best foraging area; Lb b Indicates the lower limit of the area; Ub b represents the upper limit of the region; at this time, the position of the small dung beetle is shown in formula (11).

[0090] x i (t+1)=x i (t)+C1×(x i (t)-Lb b )+C2×(x i (t)-Ub b ) (11);

[0091] In the formula, x i (t) represents the position information of the i-th dung beetle at the t-th iteration; C1 represents a random number that obeys the normal distribution; C2 is a random variable belonging to (0,1).

[0092] Dung beetles that steal food from other dung beetles are called thieving dung beetles. From equation (10), we can see that X b is the best foraging area. Therefore, the thieving dung beetle will steal in this area to obtain food. At this time, the position of the thieving dung beetle is shown in formula (12).

[0093] x i (t+1)=X b +s×g×(|x i (t)-X * |+|x i (t)-X b |) (12);

[0094] In the formula, g represents a random vector of size 1×D following a normal distribution; s represents a constant value. The position of the thieving dung beetle is constantly updated during the optimization process, and the optimal position X is finally output. b .

[0095] S2.3. Using the minimum envelope entropy as the fitness function of the DBO algorithm, DBO is used to optimize the two important parameters of VMD. After obtaining the parameter combination, it is input into the VMD algorithm. The vibration signal is adaptively decomposed to obtain multiple modal components. The modal components are screened and reconstructed according to the relevant kurtosis criterion, thereby completing the noise reduction processing of the original vibration signal.

[0096] S2.3.1. Fitness function design.

[0097] Envelope entropy is a method that calculates the complexity of a pattern based on its frequency. Its significance is to decompose a complex signal (or pattern) to find the frequency characteristics of the local signal, so that effective signal processing can be performed, signals can be classified, patterns can be discovered, etc. Envelope entropy calculation is shown in formula (13).

[0098]

[0099] In the formula, p j is the normalized form of a(j); a(j) is the envelope signal obtained after Hilbert demodulation, E p It stands for Envelope Entropy, which is an indicator to measure the complexity of a signal. In the calculation formula of Envelope Entropy, N represents the number of samples, which is the number of samples in the envelope signal obtained after the signal is transformed by Hilbert.

[0100] Envelope entropy can be used as one of the indicators to evaluate candidate solutions in the search process of the optimization algorithm. By using the minimum envelope entropy as the fitness function, the optimization algorithm can tend to select solutions that perform well in terms of minimum envelope entropy, thereby guiding the search direction towards a better solution. When searching for the best parameter combination, when the envelope entropy value reaches the minimum, it means that the parameter combination is the optimal solution. At this time, the modal components obtained by VMD decomposition contain more fault features, the waveform is a regular pulse, and the signal shows a strong sparse characteristic. Therefore, selecting the minimum envelope entropy as the fitness function allows the decomposed signal to obtain more useful feature information.

[0101] S2.3.2, related kurtosis criterion; kurtosis is one of the statistics that describes the sharpness of probability distribution or the distribution form of data points in a data set. It is a dimensionless parameter. In signal processing, kurtosis is particularly sensitive to impulse signals and can be used as an indicator to describe the characteristics and distribution of signals. When a diesel engine fails, the vibration signal will contain more impact components, and the kurtosis will increase significantly. The calculation of kurtosis K is shown in formula (14).

[0102]

[0103] Where σ is the mean difference of the signal; is the average value of the signal.

[0104] The value of the correlation coefficient can show the changing trend between two variables. Its value range is usually between -1 and 1. When the correlation coefficient is 1, it means that the two variables have a strong positive correlation. When the correlation coefficient is -1, it means that the two variables have a strong negative correlation. When the correlation coefficient is close to 0, it means that the correlation between the two variables is very low.

[0105] The correlation kurtosis is a statistical measure used to measure the degree of association between different data components and a given variable (usually the original data). It combines the kurtosis value of each data component with the correlation coefficient of its given variable, thereby quantifying the contribution of each component to the original data and taking into account their correlation with the original data.

[0106] S2.3.3. Evaluation index of noise reduction effect: After the vibration signal is decomposed by DBO-VMD, the IMF components after decomposition are screened by the correlation kurtosis criterion, and the components with large correlation kurtosis values ​​are selected for signal reconstruction. In order to evaluate the advantages of the method used in noise reduction, three commonly used noise reduction evaluation indexes are introduced. The signal-to-noise ratio (SNR) is usually used to describe the proportional relationship between the useful signal and the background noise. When the ratio is larger, it means that the noise signal in the reconstructed signal is less and the noise reduction effect is better. The mean square error (MSE) can be used to measure the difference between the reconstructed signal and the original signal. The smaller the MSE value, the better the noise reduction effect. The mean absolute error (MAE) represents the average value of the absolute value of the error between the reconstructed signal and the original signal. The smaller the MAE value, the better the noise reduction effect. The calculation formulas of the three noise reduction effect evaluation indexes are shown in equations (15), (16), and (17), respectively.

[0107]

[0108]

[0109] In the formula, x n is the original vibration signal; x m (n) is the reconstructed signal after denoising, and m represents the total number of sample points considered.

[0110] Considering that the fault characteristic information displayed by the diesel engine when a fault occurs is easily covered by noise and internal invalid information and becomes difficult to extract, the present invention proposes a signal noise reduction method based on the DBO algorithm to optimize the VMD parameters. The overall flow chart is as follows: Figure 2 As shown, the specific steps include the following steps

[0111] (1) Initialize the parameters of the DBO algorithm, set the selection range of the VMD parameters [k, a], select the minimum envelope entropy as the fitness function, iteratively optimize the vibration signals under different fault conditions, and output the optimal parameter combination [k, a].

[0112] (2) Substitute the parameter combination [k, a] optimized by the DBO algorithm into the VMD algorithm, and use VMD to decompose the original vibration signal into k modal components.

[0113] (3) Calculate the relevant kurtosis values ​​of the k modal components obtained respectively, select the indexes according to the relevant kurtosis criterion dimension, and select the modal components with larger relevant kurtosis values ​​for reconstruction.

[0114] S4. An improved parallel one-dimensional convolutional neural network model is built by introducing a gated recurrent unit, and a public data set and a preprocessed simulated vibration signal are input as sample data into the model for diesel engine fault diagnosis. Convolutional neural networks can process data samples of different dimensions, such as one-dimensional data such as time and signal, two-dimensional data such as image and audio, and three-dimensional data such as video. Since the diesel engine vibration signal is one-dimensional data, the present invention uses a one-dimensional convolutional neural network.

[0115] S4.1. A fault diagnosis model based on a parallel one-dimensional convolutional neural network. This model consists of two channels, using large and small convolution kernels as the first feature extraction layer. Subsequent convolution layers all use smaller convolution kernels. Small convolution kernels can not only build a deeper network, but also better capture local feature information and reduce training time. Large convolution kernels usually have a wider receptive field, can cover a larger area on the input, and are more suitable for capturing global features. The receptive field refers to the size of the area where a point on the signal output by each layer of CNN is mapped on the input signal. If a one-dimensional convolution layer contains multiple convolution kernels, and the width of each convolution kernel is k, then the receptive field of each neuron is k, which means that each neuron only focuses on a local area of ​​length k in the input sequence when calculating the output. The receptive field diagram is shown below. Figure 3 Compared with the single-channel one-dimensional convolutional neural network model, this model can fully extract the time-frequency domain feature information of the vibration signal while speeding up the network training speed. The parallel one-dimensional convolutional neural network structure is shown in Figure 4As shown in the figure, it includes convolutional layers, BN layers, activation layers, pooling layers, and fully connected layers. The convolutional layer extracts and learns features from the input data, providing more informative and abstract representations for subsequent layers. The convolutional layer traverses the input data with a fixed step length through the convolution kernel to achieve weight sharing, which can not only reduce the number of parameters but also improve the calculation speed. The BN layer calculates its mean and variance, then normalizes the input, and finally applies a learnable scaling factor and offset parameter to restore the model fitting ability. Doing so can accelerate the convergence of the model, reduce the problem of gradient disappearance or explosion, and also improve the generalization ability of the model. During the training process, the BN layer calculates the mean and variance by counting each mini-batch, and normalizes the input data in the forward propagation. In the inference phase, the mean and variance obtained in the training phase are usually used for normalization instead of calculating the statistical information again. The activation layer is a linear combination of the input data, which limits its ability to capture complex features. The activation function enables the neural network to learn and simulate complex nonlinear relationships by performing nonlinear transformations on the input data, thereby enhancing the understanding and expression of the data. The pooling layer reduces the dimensionality and compresses the input data while retaining the most important features. The pooling layer is usually used after the convolution layer. When the input signal passes through the convolution layer, a lot of data will be generated, including some redundant information. The pooling operation can retain the most significant features in the data and help the network better learn and understand the data. There are two main types of pooling layers. The maximum pooling uses the maximum value of the calculated area as the value after pooling, while the average pooling uses the average value of the calculated area as the value after pooling. The fully connected layer is at the end of the convolutional neural network. The function of the fully connected layer is to convert the features learned in the network into the category space of the corresponding samples. Starting from the multi-dimensional features obtained by the last layer of pooling, this feature is first converted into a one-dimensional vector and then input into the fully connected layer. In this layer, the output is generated by weighting and calculating each input feature and then processing it through an activation function. The fully connected layer integrates the local features extracted by the convolution layer or pooling layer to form a fully interconnected network structure.

[0116] S4.2, based on improved parallel one-dimensional convolutional neural network model.

[0117] S4.2.1, Gated recurrent unit GRU, by introducing the reset gate r t and update gate z t , these two gating vectors determine which information can be used as the output of the gated recurrent unit, which can better capture the long-term dependency problems in the time series and reduce the complexity of the calculation.

[0118] At time t, reset gate r t It can determine whether the network ignores the previous state information. The calculation process is shown in formula (18).

[0119] r t =σ(W r [h t-1 , X t ]) (18);

[0120] Where W r represents the weight parameter that can be learned; h t-1 Indicates the hidden state of the previous moment; X t Represents the current input signal characteristics, and t represents the current time.

[0121] At each moment, there is an update gate that calculates the output value through the Sigmoid function. When the value is close to 1, it means that the network needs to remember the previous information. When the value is close to 0, it means that the network ignores the previous state information. t The calculation process of is shown in formula (19).

[0122] Z t =σ(W Z [h t-1 , X t ]) (19);

[0123] Where W Z Represents the weight parameter that can be learned; σ represents the activation function; at each moment there is an update gate, and the output value is calculated by the Sigmoid function. When the value is close to 1, it means that the convolutional neural network remembers the previous information, and when the value is close to 0, it means that the convolutional neural network ignores the previous state information.

[0124] Reset Gate t By controlling the degree of integration of the current moment’s input information and the previous moment’s output, the candidate state is obtained. The calculation process is shown in formula (20).

[0125]

[0126] Where W h represents learnable parameters.

[0127] Combined with update gate z t The final output information is h t , and its calculation process is shown in formula (21).

[0128]

[0129] S4.2.2, Attention Mechanism; CBAM (Convolutional Block Attention Module) is a lightweight convolutional attention module that can perform attention operations in two dimensions: spatial dimension and channel dimension. Figure 5 It can be seen that CBAM consists of input, channel attention, spatial attention and output. Such a structure can not only save parameters and reduce calculations, but also can be flexibly embedded into the existing network architecture to optimize the network model. Input a one-dimensional time series signal feature information F∈R C×1×W , perform maximum pooling and average pooling operations in the channel attention module, change the pooled feature information from C×1×W to C×1×1, then compress and expand the number of channels, and obtain two activated results through the Relu activation function. Add these two output results element by element, and then use a Simogid activation function to obtain the channel attention M C ∈R C×1×1 The calculation of feature information after adjustment by the channel attention mechanism is shown in formula (22).

[0130]

[0131] The feature information F′ is subjected to maximum pooling and average pooling operations along the channel dimension, and then the pooled feature information is concatenated along the channel dimension. The feature information is processed through the convolution layer to generate the spatial attention weight. The generated spatial attention weight is applied with the Sigmoid activation function, the weight is limited to between 0 and 1, the spatial attention weight is applied to the original feature information, and the spatial attention M is obtained after weighting the features of each spatial position. S ∈R C×1×W The feature information F″ after CBAM adjustment is calculated as shown in formula (23).

[0132]

[0133] S4.2.3, Fault diagnosis process based on improved parallel one-dimensional convolutional neural network:

[0134] The parallel one-dimensional convolutional neural network is a deep convolutional neural network. As the number of network layers increases, it may cause the problem of gradient vanishing. In order to change this situation, GRU and attention mechanism are introduced in the fully connected layer. GRU effectively solves the problems of gradient vanishing and long-term dependency in recurrent neural networks by introducing reset gates and update gates. When GRU processes signals, it is easy to forget when the signal exceeds a certain length. To solve this problem, the attention mechanism is introduced to compress the data, thereby alleviating the forgetting problem of GRU in the process of processing signals, so that the model can quickly lock the key points of the target in a large amount of information. Improve the model structure of the parallel one-dimensional convolutional neural network, such as Figure 6 As shown. Figure 6 It can be seen that the improved parallel one-dimensional convolutional neural network model uses dual-channel large and small convolution kernels to process vibration signals. The two channels extract the time domain features and frequency domain features of the vibration signal respectively, which can fully utilize the characteristic information of the signal while speeding up the training speed of the network model. After the characteristic information of the vibration signal is extracted through three layers of convolution pooling, the GRU layer performs time series modeling on the extracted features to learn the relationships and dynamic changes in the sequence. The model will generate a large number of high-dimensional time features, which are difficult to distinguish and understand intuitively. By adding an attention mechanism layer after the GRU layer, important information can be extracted, so that it can quickly lock the key points in the target from a large amount of information, thereby eliminating useless information and completing the training task more efficiently. Finally, the global maximum pooling layer is used instead of the fully connected layer, which improves the computational efficiency while reducing the number of parameters, and also reduces the risk of overfitting. The fault diagnosis process of the improved parallel one-dimensional convolutional neural network is as follows: Figure 7 As shown, the fault diagnosis algorithm process includes the following steps:

[0135] (1) The DBO-VMD algorithm is used to decompose the original vibration signal and then screen and reconstruct it. The reconstructed signal is divided into training set, test set and validation set.

[0136] (2) Build an improved parallel one-dimensional convolutional neural network fault diagnosis model, construct a dual-channel multi-layer convolution and pooling network structure, use gated recurrent units to optimize the network model, reduce the gradient explosion problem that may be caused by the deep network structure, introduce the attention mechanism, solve the forgetting problem that may occur in the gated recurrent unit during the back propagation process of model training, use the global maximum pooling layer to prevent overfitting that may occur during the training process, and finally use the Softmax layer for output classification to complete the construction of the fault diagnosis model.

[0137] (3) Input the divided training set data samples into the improved one-dimensional convolutional neural network model, perform forward propagation and loss function calculation, and then input the validation set into the model for model verification. Compare the obtained results with the true values, calculate the error between the two, and determine whether the model meets the stopping criteria. If the stopping criteria are met, stop training and save the model. If not, perform error backpropagation to each layer of the model to update the parameters and continue training until the termination condition is met.

[0138] (4) Input the divided test set data samples into the trained network model, output the confusion matrix of the test set, and verify the diagnosis results.

[0139] S5. Experimental verification and result analysis.

[0140] S5.1. Experimental verification and result analysis based on public datasets.

[0141] In order to verify the effectiveness of the improved one-dimensional convolutional neural network algorithm proposed in this paper, the CWRU bearing open data set was selected for experimental analysis. This data set can be used to test the reliability of mechanical fault diagnosis algorithms. A two-horsepower motor was used for fault experiments. Single-point faults of different diameters were introduced at multiple locations of the motor, and signals were collected through vibration sensors installed on the motor housing.

[0142] The motor uses two types of bearings, namely the drive-end bearing and the fan-end bearing. This test uses the vibration data collected from the drive-end bearing, sets the signal sampling frequency to 12kHz, sets the speed to 1797r / min, selects the fault locations as the inner ring fault, outer ring fault, and rolling element fault, sets the fault defect diameters to 0.1778mm, 0.3556mm, and 0.5334mm, respectively, uses 1 normal vibration signal and 9 fault vibration signals, a total of 10 categories of data, and the data set sample length is 1200. The selected information is shown in Table 1.

[0143] Table 1 Classification of bearing data labels

[0144]

[0145] Tensorflow-GPU (version 2.0.0) and Python (version 3.6.3) are used to implement the programming and operation of the model. The computer processor (CPU) is Inter(R) Core(TM) i7-10870H, and the graphics card (GPU) is NVIDIA GeForce RTX 2060. In order to better demonstrate the advantages of the model, the performance of the network model is optimized, and the parameter indicators are shown in Table 2.

[0146] Table 2 Model parameter indicators

[0147]

[0148] The learning rate is used to control the size of each parameter update to make the training process more stable and reliable. Dropout can randomly inactivate the activation values ​​of certain neurons during training. The Relu activation function is used to measure model performance and guide model optimization to prevent model overfitting. Batch Size is used to update the number of samples of model parameters during iterative training to improve the generalization ability and stability of the model. A reasonable setting of the number of iterations can effectively train a network model with excellent performance and strong generalization ability.

[0149] The proposed DBO-VMD algorithm is used to perform noise reduction preprocessing on the divided CWRU public data set, and the processed data samples are divided according to the ratio of training set: test set: validation set of 7:2:1. The divided training set samples are input into the improved parallel one-dimensional CNN model for training, and the forward propagation loss function is calculated. The validation set samples are input into the model to verify the model. Single-channel large convolution kernel CNN, single-channel small convolution kernel CNN and parallel one-dimensional CNN are selected for comparison with the algorithm proposed in this invention. The accuracy change curve of each model validation set during training is shown in the figure below. Figure 8 As shown. Figure 8 It can be seen that the change curve of the dual-channel CNN network is more stable than that of the single-channel CNN network, and the accuracy is higher. Compared with the parallel one-dimensional CNN, the algorithm model proposed in the present invention converges faster. After 20 rounds of training, the accuracy is basically stable, which verifies the accuracy and stability of the model. The loss rate change curve of each model validation set during the training process is shown in Fig. 9 shown.

[0150] The divided test set samples are input into the above-trained improved parallel one-dimensional CNN fault diagnosis model for verification. The obtained test set accuracy is as follows: Fig.10 As shown in a, the horizontal axis represents the number of real labels, and the vertical axis represents the number of predicted labels. From the label correspondence, the algorithm proposed in the present invention has a higher accuracy in the test set. The same data set samples are input into parallel one-dimensional CNN, single-channel small convolution kernel CNN, and single-channel large convolution kernel CNN for testing. The results are shown in Figure 1. Fig.10 b in Fig.10 c in Fig.10 As shown in d, by comparison, it can be seen that when the algorithm model of the present invention is used to diagnose the test set, the accuracy is significantly higher than that of the other models.

[0151] By sorting out the above confusion matrix, the accuracy of different fault diagnosis models is obtained, as shown in Table 3. It can be seen that the accuracy of the algorithm proposed in the present invention is as high as 99.62%, which is higher than the accuracy of other models. The accuracy of large convolution kernel CNN is lower than that of small convolution kernel CNN. The main reason is that the small convolution kernel has fewer parameters, which can better realize parameter sharing, make the model more concise, and reduce the risk of overfitting. Compared with the single-channel network, the accuracy of the dual-channel network is improved, which verifies that the dual-channel network can more fully contain feature information and help enhance the recognition accuracy of the model. The accuracy of the algorithm proposed in the present invention is significantly improved compared with the dual-channel convolutional neural network model, which shows that the algorithm proposed in the present invention can make full use of feature information, reduce training time and improve model accuracy, and verify the effectiveness of the model.

[0152] Table 3 Accuracy of different fault diagnosis model test sets

[0153]

[0154] S5.1. Experimental verification and result analysis based on the vibration signal of the test bench.

[0155] Deep learning network models often require a large number of data samples when training. This section uses overlapping sampling to enhance the sample data. Data enhancement can not only expand the data set, but also help the model learn different representations of the data, reduce overfitting, and improve the model's ability to adapt to new data. Fig.11 shown.

[0156] The vibration signal data collected under four conditions of the diesel engine were selected to complete the fault diagnosis experiment. The vibration data after noise reduction preprocessing by the DBO-VMD method in Section 3.4.2 was resampled as the input sample of the fault diagnosis model. The data set was divided according to the ratio of training set: test set: validation set = 7:2:1, and the network was built according to the same computer configuration and optimization method as in the previous section. The iterative curve of the accuracy and loss rate of the fault diagnosis model proposed in the present invention is shown in Figure 2. Fig.12 shown.

[0157] from Fig.12 It can be seen that after 25 iterations of training, the accuracy and loss rate of the model are both in a stable state, indicating that the model can perfectly extract the features in the data samples. And the curve fitting degree is high, indicating that the model does not have serious overfitting or underfitting, proving the accuracy of the proposed method. In order to verify the superiority of the algorithm proposed in the present invention in diesel engine fault diagnosis, parallel one-dimensional CNN, single-channel small convolution kernel CNN, and single-channel large convolution kernel CNN are still selected as comparison models for testing. The verification set accuracy and loss rate of each model during the training process are shown as follows: Fig.13 and Fig.14 shown.

[0158] from Fig.13 and Fig.14 It can be seen that the curve of the single-channel one-dimensional CNN is very unstable, and the accuracy is also lower than that of the dual-channel one-dimensional CNN. The algorithm proposed in the present invention has a faster convergence speed, and the waveform curve has no large fluctuations, and has good stability. The test results of each model are shown in Table 4. The fault diagnosis method proposed in the present invention has the highest diagnosis rate of 99.56%. Compared with the parallel one-dimensional CNN, it has increased by 0.62%, and the diagnosis results of the single-channel large convolution kernel CNN and the single-channel small convolution kernel CNN have been significantly improved. It shows that the algorithm of the present invention has good fault identification accuracy and can better meet the monitoring needs of vibration signals.

[0159] Table 4 Accuracy and loss value of each model

[0160]

[0161] The present invention constructs a fault diagnosis model of an improved parallel one-dimensional convolutional neural network, uses a dual-channel large and small convolution kernel to extract the frequency domain and time domain features of the vibration signal respectively, and introduces a gated recurrent unit to ensure that the gradient explosion phenomenon does not occur in the neural network as the number of layers increases. An attention mechanism is added to ensure that the gated recurrent unit does not forget during back propagation. The method is experimentally verified using the Case Western Reserve University rolling bearing data set and the actual vibration signal of a diesel engine, and experimentally compared with the existing fault diagnosis model, confirming the reliability and superiority of the method.

[0162] Based on the same inventive concept, the present invention also proposes a diesel engine fault diagnosis system, comprising:

[0163] The acquisition module is used to collect vibration signal data of the diesel engine.

[0164] The data processing module is used to build a fault diagnosis model based on the improved parallel one-dimensional convolutional neural network; the vibration signal data is input into the fault diagnosis model, and the time domain features and frequency domain features of the vibration signal are extracted respectively through the dual-channel convolution layer, and the time domain features and frequency domain features are respectively input into the gated recurrent unit layer for time series modeling after being subjected to three-layer convolution pooling to obtain the dynamic change information of the time domain features and frequency domain features; the key features related to the diesel engine fault in the dynamic change information of the time domain features and frequency domain features are respectively extracted through the attention mechanism layer, the two extracted key features are spliced ​​and fused, the fused features are input into the Softmax layer after passing through the global maximum pooling layer, and the classification results are output.

[0165] The diagnosis module is used to diagnose the fault of the diesel engine according to the classification result.

[0166] The present invention also provides a diesel engine fault diagnosis computer device, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the diesel engine fault diagnosis method when executing the computer program.

[0167] The present invention further provides a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the steps of the diesel engine fault diagnosis method are executed.

[0168] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A diesel engine fault diagnosis method, characterized in that: The following steps are involved: Collect vibration signal data of diesel engine; A fault diagnosis model is constructed based on the improved parallel one-dimensional convolutional neural network; the vibration signal data is input into the fault diagnosis model, and the time domain features and frequency domain features of the vibration signal are extracted through the dual-channel convolution layer, and the time domain features and frequency domain features are respectively input into the gated recurrent unit layer after three layers of convolution pooling for time series modeling to obtain the dynamic change information of the time domain features and frequency domain features; the key features related to diesel engine faults in the dynamic change information of the time domain features and frequency domain features are respectively extracted through the attention mechanism layer, and the two extracted key features are spliced ​​and fused, and the fused features are input into the Softmax layer after passing through the global maximum pooling layer, and the classification results are output; The diesel engine fault is diagnosed according to the classification results.

2. A diesel engine fault diagnosis method according to claim 1, characterized in that: The step of performing noise reduction processing on the vibration signal data before inputting the vibration signal data into the fault diagnosis model specifically comprises the following steps: The minimum envelope entropy is used as the fitness function of the dung beetle optimization algorithm DBO, and DBO is used to optimize two important parameters of the variational mode decomposition algorithm VMD. The optimized VMD is used to adaptively decompose the vibration signal to obtain multiple modal components; Signal reconstruction of multiple modal components is performed according to the relevant kurtosis criterion; The signal-to-noise ratio (SNR), mean square error (MSE) and mean absolute error (MAE) are used to evaluate the noise reduction of the reconstructed signal. The larger the SNR value, the smaller the MSE and MAE values, and the less noise signal there is in the reconstructed signal. The vibration signal data after noise reduction is input into the fault diagnosis model.

3. A diesel engine fault diagnosis method according to claim 1, characterized in that: The time domain features and the frequency domain features are respectively subjected to three layers of convolutional pooling and then input into the gated recurrent unit layer for time series modeling to obtain dynamic change information of the time domain features and the frequency domain features; specifically, the following steps are included: By introducing a reset gate r t and update gate z t Capture the long-term dependencies in the time series; the reset gate r t It is expressed as: r t =σ(W r [h t-1 ,X t ]), Where W r represents the weight parameter that can be learned; h t-1 Indicates the hidden state of the previous moment; X t represents the current input signal characteristics, and t represents the current time; The update gate z t It is expressed as: Z t =σ(W z [h t-1 ,X t ]), Where W Z Represents the weight parameter that can be learned; σ represents the activation function; there is an update gate at each moment, and the output value is calculated by the Sigmoid function. When the value is close to 1, it means that the convolutional neural network remembers the previous information, and when the value is close to 0, it means that the convolutional neural network ignores the previous state information; Reset Gate t By controlling the degree of integration of the current moment’s input information and the previous moment’s output, the candidate state is obtained. It is expressed as: Where W h represents the parameters that can be learned; Candidate status With update gate z t Combined, the information output at the final moment is h t , which is expressed as:

4. A diesel engine fault diagnosis method according to claim 1, characterized in that: The method of extracting key features related to diesel engine failure from the dynamic change information of time domain features and frequency domain features through the attention mechanism layer specifically includes the following steps: The characteristic information F∈R of the one-dimensional time series signal C×l×W In the input attention mechanism layer CBAM, maximum pooling and average pooling operations are performed in the channel attention module; The feature information after pooling is changed from C×1×W to C×1×1, and then the number of channels is compressed and then expanded, and the two activated results are output through the Relu activation function; The two output results are added element by element and then a Simogid activation function is used to obtain the channel attention M. c ∈R C ×1×1 ; Then the feature information F′ adjusted by the channel attention module is expressed as: Perform maximum pooling and average pooling operations on the feature information F′ along the channel dimension, and then concatenate the pooled feature information along the channel dimension to generate spatial attention weights; Apply the Sigmoid activation function to the generated spatial attention weights, limit the weights between 0 and 1, apply the spatial attention weights to the original feature information, and weight the features of each spatial position to obtain the spatial attention M. s ∈R C ×1×W ; Then the characteristic information F″ after CBAM adjustment is expressed as:

5. A diesel engine fault diagnosis method according to claim 1, characterized in that: The collecting of vibration signal data of the diesel engine specifically includes selecting a cylinder head of the diesel engine as a measuring point, and installing a vibration sensor at the measuring point to collect acceleration information generated by vibration during the operation of the diesel engine to obtain a vibration signal.

6. A diesel engine fault diagnosis system, characterized in that: include: An acquisition module, used for acquiring vibration signal data of a diesel engine; The data processing module is used to build a fault diagnosis model based on the improved parallel one-dimensional convolutional neural network; the vibration signal data is input into the fault diagnosis model, and the time domain features and frequency domain features of the vibration signal are respectively extracted through the dual-channel convolution layer, and the time domain features and frequency domain features are respectively input into the gated recurrent unit layer for time series modeling after being subjected to three-layer convolution pooling to obtain the dynamic change information of the time domain features and frequency domain features; the key features related to the diesel engine fault in the dynamic change information of the time domain features and frequency domain features are respectively extracted through the attention mechanism layer, and the two extracted key features are spliced ​​and fused, and the fused features are input into the Softmax layer after passing through the global maximum pooling layer, and the classification results are output; The diagnosis module is used to diagnose the fault of the diesel engine according to the classification result.

7. A diesel engine fault diagnosis computer device, characterized in that: include: A memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of the diesel engine fault diagnosis method according to any one of claims 1 to 5 when executing the computer program.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the diesel engine fault diagnosis method according to any one of claims 1 to 5.

Images