Internal Combustion Engine Fault Diagnosis Method and Equipment Based on Time-Domain and Frequency-Domain Feature Fusion

Abstract

This application discloses a method and device for internal combustion engine fault diagnosis based on the fusion of time-domain and frequency-domain features, relating to the field of fault detection technology. The method includes: acquiring the vibration acceleration signal of the internal combustion engine during operation; decomposing the vibration acceleration signal using variational mode decomposition based on a preset center frequency to obtain multiple signal components, wherein the sum of the bandwidths of the multiple signal components is minimized, and the reconstructed signal energy can yield the vibration acceleration signal; splicing the signal components; and inputting the spliced ​​signal and the vibration acceleration signal into a pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model. This application aims to solve the problem of low prediction accuracy in existing internal combustion engine fault diagnosis technologies, thereby improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of fault detection technology, and in particular to a method and device for diagnosing internal combustion engine faults based on the fusion of time-domain and frequency-domain features. Background Technology

[0002] As one of the core power sources for mechanical equipment, the internal combustion engine is widely used in transportation, agricultural and construction machinery, generator sets, and general aviation due to its high thermal efficiency and ease of movement. However, because the internal combustion engine is the core power source of mechanical equipment, a malfunction could lead to an accident.

[0003] Existing technologies improve equipment safety and minimize the impact of malfunctions through fault prediction. For example, machine learning and deep learning methods are used to diagnose internal combustion engine faults. Machine learning methods such as support vector machines and random forests are widely used; however, their feature extraction capabilities are poor, leading to low accuracy in the final fault diagnosis results. Deep learning methods, primarily convolutional neural networks, are suitable for image signals but have poor feature extraction capabilities for time-domain signals, resulting in low accuracy in the final fault diagnosis results. Summary of the Invention

[0004] To address the aforementioned problems and technical requirements, the applicant proposes a fault diagnosis method and device for internal combustion engines based on the fusion of time-domain and frequency-domain features. This method aims to solve the problem of low prediction accuracy in existing internal combustion engine fault diagnosis technologies and improve the accuracy of fault diagnosis.

[0005] This application provides a fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features, the method comprising:

[0006] Acquire vibration acceleration signals of internal combustion engines during operation;

[0007] Based on a preset center frequency, the vibration acceleration signal is decomposed using a variational mode decomposition method to obtain multiple signal components. Among these components, the sum of the bandwidths of the multiple signal components is minimized, and the vibration acceleration signal can be obtained by reconstructing the signal components.

[0008] The signal components are spliced ​​together, and the spliced ​​signal and the vibration acceleration signal are input into a pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model.

[0009] The process involves inputting the spliced ​​signal into an attention network to obtain a target attention matrix output by the attention network corresponding to the vibration acceleration signal; inputting the target attention matrix into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network; inputting the vibration acceleration signal into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal output by the frequency-domain feature extraction network; fusing the time-domain features and the frequency-domain features to obtain a fused feature; and performing fault diagnosis of the internal combustion engine based on the fused feature.

[0010] The fault prediction model includes: the attention network, the time-domain feature extraction network, and the frequency-domain feature extraction network;

[0011] The fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples, and fault category samples.

[0012] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion, the obtained spliced ​​signal is input into an attention network to obtain a target attention matrix output by the attention network corresponding to the vibration acceleration signal, including:

[0013] The spliced ​​signal is input into the attention network, and the weight of each signal component is calculated through the attention network. Based on the weight, the attention matrix corresponding to the spliced ​​signal is obtained. The product of the attention matrix and the spliced ​​signal is calculated to obtain the attention-weighted signal matrix, and the attention-weighted signal matrix is ​​determined as the target attention matrix.

[0014] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion, the method calculates the weight of each signal component through the attention network, and obtains the attention matrix corresponding to the spliced ​​signal based on the weight, including:

[0015] Extract the initial features of each of the signal components;

[0016] Assign a corresponding shallow neural network to each initial feature, and determine the probability that the initial feature is selected by the shallow neural network for classification.

[0017] The attention matrix is ​​obtained by combining the probabilities used for classification corresponding to each initial feature.

[0018] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion, the initial features of each signal component are extracted, including:

[0019] For each of the signal components, the following initial feature extraction process is performed:

[0020] The signal components are input into a preset feature extraction formula to obtain the initial features output by the feature extraction formula;

[0021] The feature extraction formula includes:

[0022]

[0023] Among them, E g Let represent the initial features corresponding to the g-th signal component, Tanh represent the activation function, W represent the coefficients, and b be a constant. Let g represent the g-th signal component, and ω represent the center frequency.

[0024] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion, a corresponding shallow neural network is assigned to each initial feature, and the probability that the initial feature is selected by the shallow neural network for classification is determined, including:

[0025] For each of the initial features, the following probability calculation process is performed:

[0026] The initial features are input into the corresponding shallow neural network. The selection probability is obtained through the selection formula in the shallow neural network, and the non-selection probability is obtained through the non-selection formula in the shallow neural network. The selection probability and the non-selection probability are input into the classification formula to obtain the probability that the initial features are selected for classification.

[0027] The selection formula includes:

[0028]

[0029] Where, p k This represents the probability of selection. This represents the network parameters of the k-th shallow neural network when the initial features are selected. This represents the network constant of the k-th shallow neural network when the initial features are selected;

[0030] The unselected formulas include:

[0031]

[0032] Where, n k This indicates the probability of not selecting a value. This represents the network parameters of the k-th shallow neural network when the initial feature is not selected. This represents the network constant of the k-th shallow neural network when the initial feature is not selected;

[0033] The classification formulas include:

[0034]

[0035] Among them, a k This represents the probability used for classification.

[0036] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion, the vibration acceleration signal is decomposed using variational mode decomposition based on a preset center frequency to obtain multiple signal components, including:

[0037] The vibration acceleration signal is decomposed using the variational mode decomposition method to obtain multiple signal components;

[0038] The signal components are converted into analytic signals based on the Hilbert transform and Dirac function.

[0039] The analytical signal is multiplied by an exponential term tuned to the corresponding center frequency to shift the spectrum of the signal component to the baseband, and the bandwidth of the signal component is estimated by Gaussian smoothing.

[0040] The decomposition is considered complete when the sum of the bandwidths of the multiple signal components is minimized, and the vibration acceleration signal can be obtained by reconstructing the signal components.

[0041] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion, the training process of the fault prediction model includes:

[0042] Obtain model training samples, wherein the model training samples include: the spliced ​​signal samples, the vibration acceleration signal samples, and the fault category samples;

[0043] The spliced ​​signal sample is input into the attention network of the fault prediction model to obtain a target attention matrix sample output by the attention network corresponding to the vibration acceleration signal sample; the target attention matrix sample is input into the time-domain feature extraction network of the fault prediction model to obtain the predicted time-domain features output by the time-domain feature extraction network; the vibration acceleration signal sample is input into the frequency-domain feature extraction network of the fault prediction model to obtain the predicted frequency-domain features output by the frequency-domain feature extraction network; the time-domain features and the frequency-domain features are fused to obtain fused features, and a predicted fault diagnosis result is obtained based on the fused features; the consistency between the fault category sample and the fault diagnosis result is determined; the objective function of the attention network, the network parameters of the time-domain feature extraction network, and the network parameters of the frequency-domain feature extraction network are adjusted based on the consistency until the consistency reaches a preset consistency.

[0044] According to an embodiment of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion according to this application, the objective function includes:

[0045]

[0046] Where A represents the attention matrix, θ = <θ l θ α >,θ l and θ α All represent the network parameters of the attention network, R(θ) represents the L2 norm, and λ represents the strength of the regularization. The network parameter is θ l The corresponding network layer, Y represents the predicted attention matrix, and Y is the actual attention matrix.

[0047] According to one embodiment of the internal combustion engine fault diagnosis method based on the fusion of time-domain and frequency-domain features, the time-domain feature extraction network includes a residual neural network structure;

[0048] The training process of the fault prediction model also includes:

[0049] During the training process, if the number of iterations exceeds the preset number and the consistency does not reach the preset consistency, a residual layer of the residual neural network structure is added, and the model is trained based on the modified fault prediction model.

[0050] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion as described in any of the preceding claims.

[0051] The internal combustion engine fault diagnosis method and device based on time-domain and frequency-domain feature fusion provided in this application acquires the vibration acceleration signal of the internal combustion engine during operation; determines the bandwidth corresponding to the vibration acceleration signal; and decomposes the vibration acceleration signal using variational mode decomposition based on the determined bandwidth and a preset center frequency to obtain multiple signal components. This application effectively separates the vibration acceleration signal data in the time domain using variational mode decomposition, providing an effective data foundation for subsequent extraction of time-domain features; and splices the signal components, inputting the spliced ​​signal and the vibration acceleration signal into a pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model.

[0052] Specifically, the attention network of the fault prediction model takes the spliced ​​signal as input and outputs a target attention matrix corresponding to the vibration acceleration signal. The target attention matrix is ​​then input into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network. This application utilizes an attention mechanism to quickly obtain the target attention matrix, thereby enabling the time-domain feature extraction network to quickly obtain the time-domain features. Simultaneously, the vibration acceleration signal is input into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal output by the frequency-domain feature extraction network. By processing the time-domain and frequency-domain features separately, this application can accurately extract the signal features of the vibration acceleration signal, providing an accurate data foundation for subsequent fault diagnosis. Finally, the time-domain and frequency-domain features are fused to obtain fused features, which are then used for fault diagnosis of the internal combustion engine. The entire process improves the accuracy, timeliness, and comprehensiveness of signal feature extraction, thereby improving the overall accuracy of fault diagnosis results. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is one of the flowcharts illustrating the fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features provided in this application embodiment;

[0055] Figure 2 This is the second flowchart of the fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features provided in this application embodiment;

[0056] Figure 3 This is a schematic diagram of the structure of the internal combustion engine fault diagnosis device based on the fusion of time-domain and frequency-domain features provided in the embodiments of this application;

[0057] Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0059] This application provides a fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features. This method can be applied to smart terminals or servers. Other descriptions in the embodiments of this application are illustrative and not intended to limit the scope of protection of this application, and will not be described in detail thereafter. The specific implementation of the method is as follows: Figure 1 As shown:

[0060] Step 101: Obtain the vibration acceleration signal of the internal combustion engine during operation.

[0061] Vibration acceleration signals can be collected by an acceleration signal acquisition device installed at the internal combustion engine.

[0062] Step 102: Determine the bandwidth corresponding to the vibration acceleration signal. Based on the determined bandwidth and the preset center frequency, use the variational mode decomposition method to decompose the vibration acceleration signal to obtain multiple signal components.

[0063] Step 103: Splice the signal components. Input the spliced ​​signal and vibration acceleration signal into the pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model.

[0064] The process involves using an attention network to obtain the spliced ​​signal, which outputs a target attention matrix corresponding to the vibration acceleration signal. This target attention matrix is ​​then input into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal. The vibration acceleration signal is then input into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal. Finally, the time-domain and frequency-domain features are fused to obtain a fused feature, which is used for fault diagnosis of the internal combustion engine.

[0065] The fault prediction model includes: an attention network, a time-domain feature extraction network, and a frequency-domain feature extraction network.

[0066] The fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples, and fault category samples.

[0067] The internal combustion engine fault diagnosis method based on the fusion of time-domain and frequency-domain features provided in this application acquires the vibration acceleration signal of the internal combustion engine during operation; determines the bandwidth corresponding to the vibration acceleration signal; and decomposes the vibration acceleration signal using variational mode decomposition based on the determined bandwidth and a preset center frequency to obtain multiple signal components. This application effectively separates the vibration acceleration signal data in the time domain using variational mode decomposition, providing an effective data foundation for subsequent extraction of time-domain features; and splices the signal components, inputting the spliced ​​signal and the vibration acceleration signal into a pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model.

[0068] Specifically, the attention network of the fault prediction model takes the spliced ​​signal as input and outputs a target attention matrix corresponding to the vibration acceleration signal. The target attention matrix is ​​then input into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network. This application utilizes an attention mechanism to quickly obtain the target attention matrix, thereby enabling the time-domain feature extraction network to quickly obtain the time-domain features. Simultaneously, the vibration acceleration signal is input into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal output by the frequency-domain feature extraction network. By processing the time-domain and frequency-domain features separately, this application can accurately extract the signal features of the vibration acceleration signal, providing an accurate data foundation for subsequent fault diagnosis. Finally, the time-domain and frequency-domain features are fused to obtain fused features, which are then used for fault diagnosis of the internal combustion engine. The entire process improves the accuracy, timeliness, and comprehensiveness of signal feature extraction, thereby improving the overall accuracy of fault diagnosis results.

[0069] In one specific embodiment, the vibration acceleration signal is decomposed into multiple signal components based on a preset center frequency using a variational mode decomposition method. The specific implementation includes:

[0070] The vibration acceleration signal is decomposed using variational mode decomposition to obtain multiple signal components. The signal components are then converted into analytic signals based on Hilbert transform and Dirac function. The analytic signals are multiplied by the exponential term tuned to the corresponding center frequency to shift the spectrum of the signal components to the baseband, and the bandwidth of the signal components is estimated by Gaussian smoothing. The decomposition is considered complete when the sum of the bandwidths of the multiple signal components is minimized and the reconstructed signal components can yield the vibration acceleration signal.

[0071] Specifically, the signal component (IMF) is a set of amplitude-modulated and frequency-modulated signals. According to Carson's rule, the bandwidth of the IMF is limited, and the IMF is close to the center frequency.

[0072] To facilitate and accurately estimate the bandwidth information of each IMF, the Hilbert transform is used to represent the actual signal (signal components) in the form of an analytic signal, as detailed in formula (1):

[0073]

[0074] in, Let ω represent the analytic signal, ω represent the center frequency, j represent the imaginary unit, and the square of j equals -1.

[0075] Among them, the one-dimensional Hilbert transform (1DHT) is a linear translation-invariant operator that can cause a 90° phase shift after mapping a one-dimensional signal.

[0076] Therefore, the imaginary part of the signal can be constructed using 1DHT, thereby representing the signal components as analytic signals, as detailed in formula (2):

[0077] u g (t)=u g (t)+jH(u g (t))=A g (t)e jφ(t) =A g (t)(cos(φ(t))+j sin(φ(t)))…………(2)

[0078] Among them, u g (t) represents the g-th signal component, A g (t) represents the amplitude, φ(t) represents the phase, and e jφ(t) This represents the rotation vector of a complex signal.

[0079] The instantaneous frequency of the complex signal can be obtained by differentiating the phase, as shown in formula (3):

[0080] ω(t)=dφ(t) / dt…………………………………………(3)

[0081] Where ω(t) represents the instantaneous frequency.

[0082] Here, complex signal refers to a complex signal, because variational mode decomposition is performed in the frequency domain and has real and imaginary components, which are complex signal components.

[0083] Furthermore, the original signal can be obtained by taking the real part of the analytic signal. Therefore, when performing the Hilbert transform on the signal components, the Dirac function is introduced, and the analytic signal obtained is shown in formula (4):

[0084]

[0085] Here, δ(t) represents the Dirac function.

[0086] In one specific embodiment, the specific implementation of inputting the obtained spliced ​​signal into an attention network to obtain the target attention matrix output by the attention network corresponding to the vibration acceleration signal includes:

[0087] The spliced ​​signal is input into an attention network, which calculates the weight of each signal component. Based on these weights, the attention matrix corresponding to the spliced ​​signal is obtained. The product of the attention matrix and the spliced ​​signal is calculated to obtain the attention-weighted signal matrix, which is then used as the target attention matrix.

[0088] The spliced ​​signal is represented by X.

[0089] Among them, attention networks calculate the weights of input data based on the attention mechanism, focusing on the larger weights and treating them as important information to improve the efficiency and accuracy of the data.

[0090] In one specific embodiment, the specific implementation of calculating the weight of each signal component through an attention network and obtaining the attention matrix corresponding to the spliced ​​signal based on the weights includes:

[0091] Extract initial features for each signal component. Assign a corresponding shallow neural network to each initial feature and determine the probability that the shallow neural network will select the initial feature for classification. Combine the classification probabilities corresponding to each initial feature to obtain the attention matrix.

[0092] In one specific embodiment, the specific implementation of extracting the initial features of each signal component includes:

[0093] For each signal component, perform the following initial feature extraction process:

[0094] Input the signal components into the preset feature extraction formula to obtain the initial features output by the feature extraction formula.

[0095] The feature extraction formula is shown in formula (5):

[0096]

[0097] Among them, E g Let represent the initial features corresponding to the g-th signal component, Tanh represent the activation function, W represent the coefficients, and b be a constant. Let g represent the g-th signal component, and ω represent the center frequency.

[0098] Specifically, the initial features of the signal components are extracted using formula (5), which compresses the initial features into a smaller vector while retaining most of the information (discarding some redundant features and noise). The Tanh function, which has both positive and negative values, can retain important information during the extraction process.

[0099] In one specific embodiment, a corresponding shallow neural network is assigned to each initial feature, and the specific implementation of determining the probability that the initial feature is selected by the shallow neural network for classification includes:

[0100] For each initial feature, perform the following probability calculation process:

[0101] The initial features are input into the corresponding shallow neural network. The selection probability is obtained through the selection formula in the shallow neural network, and the non-selection probability is obtained through the non-selection formula in the shallow neural network. The selection probability and the non-selection probability are then input into the classification formula to obtain the probability that the initial features are selected for classification.

[0102] The selection formula is shown in formula (6):

[0103]

[0104] Where, p k This represents the probability of selection. This represents the network parameters of the k-th shallow neural network given that the initial features have been selected. This represents the network constant of the k-th shallow neural network when the initial features are selected.

[0105] For formulas not selected, see formula (7):

[0106]

[0107] Where, n k This indicates the probability of not selecting a value. This represents the network parameters of the k-th shallow neural network when the initial features are not selected. This represents the network constant of the k-th shallow neural network when the initial features are not selected.

[0108] The classification formula is shown in formula (8):

[0109]

[0110] Among them, a k This represents the probability used for classification.

[0111] Specifically, initial features are input into an attention network, and each initial feature is assigned a shallow neural network to determine its probability of being selected. Each shallow neural network generates two values: a selection probability and a non-selection probability. Furthermore, only the initial features whose probability for classification is greater than a preset probability are considered. Then, the probabilities for classification corresponding to each initial feature are combined to obtain the attention matrix A.

[0112] in,

[0113] Where m represents the number of iterations during the training of the attention network, and d represents the number of initial features with a probability greater than a preset probability used for classification.

[0114] Finally, after obtaining the attention matrix A, the product of A and X is calculated to obtain G, i.e., G = A * X. At this point, G is the target attention matrix.

[0115] In one specific embodiment, the specific implementation of inputting the target attention matrix into the temporal feature extraction network to obtain the temporal features of the vibration acceleration signal output by the temporal feature extraction network includes:

[0116] In the temporal domain, the target attention matrix is ​​input into the temporal feature extraction network, which transforms the target attention matrix into the input size of the adaptive residual layer. Then, multi-level feature extraction is performed through the residual layer of the temporal feature extraction network to obtain intermediate features. Subsequently, a convolutional layer is used to transform the dimension of the intermediate features and perform further feature extraction. After batch normalization, non-linear mapping is performed through an activation function, and temporal features are extracted using a self-attention mechanism. Finally, the features from the self-attention mechanism are globally averaged to obtain the temporal features.

[0117] In one specific embodiment, the specific implementation of inputting the vibration acceleration signal into a frequency domain feature extraction network to obtain the frequency domain features of the vibration acceleration signal output by the frequency domain feature extraction network includes:

[0118] In the frequency domain, data propagates synchronously with the time domain. The vibration acceleration signal is input into a frequency domain feature extraction network, where a Fast Fourier Transform (FFT) transforms the signal from the time domain into a single-sided frequency domain signal. Frequency domain features are then extracted through multiple convolutional layers, residual layers, and pooling layers. Since frequency domain signal features have lost their original temporal sequence, convolutional neural networks can effectively extract them.

[0119] After feature extraction is complete, a fully connected layer is used to expand the feature layer, so that the scale of the frequency domain features is consistent with the scale of the time domain features.

[0120] In one specific embodiment, time-domain features and frequency-domain features are fused to obtain fused features. The specific implementation of fault diagnosis of internal combustion engines based on fused features includes:

[0121] The time-domain features and frequency-domain features are concatenated and input into a classifier consisting of two fully connected layers. The classifier then outputs the fault diagnosis results.

[0122] Specifically, almost all convolutional layers in the aforementioned network structure have corresponding pooling layers and activation functions, including the ReLU activation function. When faced with input data less than 1, this function results in zero gradient propagation during both forward and backward propagation, which can deactivate some neurons. Therefore, this application utilizes a smooth Swish function to improve the accuracy and efficiency of fault diagnosis.

[0123] In one specific embodiment, the training process of the fault prediction model includes:

[0124] Obtain model training samples; input the spliced ​​signal samples into the attention network of the fault prediction model to obtain the target attention matrix sample output by the attention network corresponding to the vibration acceleration signal sample; input the target attention matrix sample into the time-domain feature extraction network of the fault prediction model to obtain the predicted time-domain features output by the time-domain feature extraction network; input the vibration acceleration signal samples into the frequency-domain feature extraction network of the fault prediction model to obtain the predicted frequency-domain features output by the frequency-domain feature extraction network; fuse the time-domain features and frequency-domain features to obtain the fused features, and obtain the predicted fault diagnosis result based on the fused features; determine the consistency between the fault category samples and the fault diagnosis result; adjust the objective function of the attention network, the network parameters of the time-domain feature extraction network, and the network parameters of the frequency-domain feature extraction network based on the consistency until the consistency reaches the preset consistency.

[0125] The model training samples include: spliced ​​signal samples, vibration acceleration signal samples, and fault category samples.

[0126] Specifically, the model training samples can be divided according to a preset ratio, such as 4:1. Four-fifths of the data can be used to train the fault prediction model, and one-fifth of the data can be used to test the prediction accuracy of the fault prediction model.

[0127] In addition, the vibration acceleration signal of the internal combustion engine can be acquired in real time in the form of a time window and added to the model training samples to update the fault prediction model.

[0128] In one specific embodiment, the objective function is shown in formula (9):

[0129]

[0130] Where A represents the attention matrix, θ = <θ lθ α >,θ l and θ α All represent the network parameters of the attention network, R(θ) represents the L2 norm, and λ represents the strength of the regularization. The network parameter is θ l The corresponding network layer, Y represents the predicted attention matrix, and Y is the actual attention matrix.

[0131] In one specific embodiment, the temporal feature extraction network includes a residual neural network structure. The training process of the fault prediction model further includes: during the training process, if the number of iterations exceeds a preset number and the consistency does not reach a preset consistency, adding a residual layer to the residual neural network structure, and training based on the modified fault prediction model.

[0132] Furthermore, during model training, due to multiple iterations of training the attention network, the output signal components after n iterations are...

[0133] Specifically, the initial number of residual layers is determined in advance based on the complexity of the fault and the number of training samples of the model, and the residual layers are adjusted based on the actual training process.

[0134] In addition, during model training, Dropout layers are used to randomly mask some neurons to prevent the model from becoming overly dependent on certain neurons and thus prevent overfitting.

[0135] Below, through Figure 2 The implementation of this application is described in detail below:

[0136] Step 201: Obtain the vibration acceleration signal of the internal combustion engine during operation.

[0137] Step 202: Based on the preset center frequency, the vibration acceleration signal is decomposed using the variational mode decomposition method to obtain multiple signal components.

[0138] Step 203: Splice the signal components and input the spliced ​​signal into a preset attention network to obtain the target attention matrix output by the attention network corresponding to the vibration acceleration signal.

[0139] Step 204: Input the target attention matrix into the preset time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network.

[0140] Step 205: Input the vibration acceleration signal into a preset frequency domain feature extraction network to obtain the frequency domain features of the vibration acceleration signal output by the frequency domain feature extraction network.

[0141] Step 206: Fuse time-domain features and frequency-domain features to obtain fused features. Input the fused features into a preset classifier to obtain the fault diagnosis results of the internal combustion engine output by the classifier.

[0142] This application leverages the advantages of convolutional neural networks, residual neural networks, and self-attention mechanisms to enhance the model's ability to represent temporal vibration signals, effectively improving the accuracy of fault diagnosis. Furthermore, it employs a parallel feature extraction method in the time and frequency domains, preserving temporal features while simultaneously obtaining frequency domain features. Combining these time and frequency domain features for classification effectively improves the model's training efficiency and accuracy. Additionally, using the Switch function as the activation function further enhances the model's nonlinear fitting ability and convergence efficiency. Finally, an incremental learning update strategy is employed to update the model, enabling it to continuously adapt to changes in the internal combustion engine throughout its lifecycle and maintain stable diagnostic results.

[0143] This application also provides an internal combustion engine fault diagnosis device based on the fusion of time-domain and frequency-domain features. The specific implementation of this device can be found in the description of the internal combustion engine fault diagnosis method based on the fusion of time-domain and frequency-domain features; repeated details will not be elaborated upon. Figure 3 As shown, the device includes:

[0144] The acquisition module 301 is used to acquire the vibration acceleration signal of the internal combustion engine during operation;

[0145] The signal decomposition module 302 is used to decompose the vibration acceleration signal based on a preset center frequency using the variational mode decomposition method to obtain multiple signal components, wherein the sum of the bandwidths of the multiple signal components is minimized, and the reconstructed signal components can obtain the vibration acceleration signal.

[0146] Fault prediction 303 is used to splice signal components. The spliced ​​signal and vibration acceleration signal are input into the pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model.

[0147] The process involves inputting the spliced ​​signal into an attention network to obtain a target attention matrix corresponding to the vibration acceleration signal. This target attention matrix is ​​then input into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal. The vibration acceleration signal is further input into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal. Finally, the time-domain and frequency-domain features are fused to obtain a fused feature, which is used for fault diagnosis of the internal combustion engine.

[0148] The fault prediction model includes: an attention network, a time-domain feature extraction network, and a frequency-domain feature extraction network;

[0149] The fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples, and fault category samples.

[0150] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 401, a communication interface 402, a memory 403, and a communication bus 404. The processor 401, communication interface 402, and memory 403 communicate with each other via the communication bus 404. The processor 401 can call logical instructions in the memory 403 to execute a fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features. This method includes: acquiring the vibration acceleration signal of the internal combustion engine during operation; decomposing the vibration acceleration signal using a variational mode decomposition method based on a preset center frequency to obtain multiple signal components, wherein the sum of the bandwidths of the multiple signal components is minimized, and the reconstructed signal components can obtain the vibration acceleration signal; splicing the signal components; inputting the obtained spliced ​​signal and vibration acceleration signal into a pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model; wherein the obtained spliced ​​signal is input into an attention network to obtain the attention... The attention matrix corresponding to the vibration acceleration signal is output by the attention network; the target attention matrix is ​​input into the time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network; the vibration acceleration signal is input into the frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal output by the frequency-domain feature extraction network; the time-domain features and frequency-domain features are fused to obtain fused features, and fault diagnosis of internal combustion engines is performed based on the fused features; the fault prediction model includes: attention network, time-domain feature extraction network and frequency-domain feature extraction network; the fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples and fault category samples.

[0151] Furthermore, the logical instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by the computer, the computer can execute the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion provided by the above methods. The method includes: acquiring the vibration acceleration signal of the internal combustion engine during operation; decomposing the vibration acceleration signal using a variational mode decomposition method based on a preset center frequency to obtain multiple signal components, wherein the sum of the bandwidths of the multiple signal components is minimized, and the reconstructed signal components can obtain the vibration acceleration signal; splicing the signal components, and inputting the obtained spliced ​​signal and vibration acceleration signal into a pre-trained fault prediction model to obtain the fault prediction signal. The fault diagnosis results output by the model are measured. Specifically, the spliced ​​signal is input into an attention network to obtain a target attention matrix corresponding to the vibration acceleration signal. The target attention matrix is ​​then input into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network. The vibration acceleration signal is then input into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal output by the frequency-domain feature extraction network. The time-domain and frequency-domain features are fused to obtain a fused feature, which is used for fault diagnosis of the internal combustion engine. The fault prediction model includes an attention network, a time-domain feature extraction network, and a frequency-domain feature extraction network. The fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples, and fault category samples.

[0153] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program performs the internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion provided in the above embodiments. The method includes: acquiring a vibration acceleration signal of the internal combustion engine during operation; decomposing the vibration acceleration signal using a variational mode decomposition method based on a preset center frequency to obtain multiple signal components, wherein the sum of the bandwidths of the multiple signal components is minimized, and the reconstructed signal components can yield a vibration acceleration signal; splicing the signal components; and inputting the obtained spliced ​​signal and vibration acceleration signal into a pre-trained fault prediction model to obtain a fault diagnosis result output by the fault prediction model. The spliced ​​signal is input into an attention network to obtain a target attention matrix corresponding to the vibration acceleration signal. The target attention matrix is ​​then input into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal. The vibration acceleration signal is then input into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal. The time-domain and frequency-domain features are fused to obtain a fused feature, which is used for fault diagnosis of the internal combustion engine. The fault prediction model includes an attention network, a time-domain feature extraction network, and a frequency-domain feature extraction network. The fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples, and fault category samples.

[0154] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0155] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0156] Finally, it should be noted that the above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.

Claims

1. A fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features, characterized in that, The method includes: Acquire vibration acceleration signals of internal combustion engines during operation; Based on a preset center frequency, the vibration acceleration signal is decomposed using a variational mode decomposition method to obtain multiple signal components. Among these components, the sum of the bandwidths of the multiple signal components is minimized, and the vibration acceleration signal can be obtained by reconstructing the signal components. The signal components are spliced ​​together, and the spliced ​​signal and the vibration acceleration signal are input into a pre-trained fault prediction model to obtain the fault diagnosis result output by the fault prediction model. The process involves inputting the spliced ​​signal into an attention network to obtain a target attention matrix output by the attention network corresponding to the vibration acceleration signal; inputting the target attention matrix into a time-domain feature extraction network to obtain the time-domain features of the vibration acceleration signal output by the time-domain feature extraction network; inputting the vibration acceleration signal into a frequency-domain feature extraction network to obtain the frequency-domain features of the vibration acceleration signal output by the frequency-domain feature extraction network; fusing the time-domain features and the frequency-domain features to obtain a fused feature; and performing fault diagnosis of the internal combustion engine based on the fused feature. The fault prediction model includes: the attention network, the time-domain feature extraction network, and the frequency-domain feature extraction network; The fault prediction model is trained based on spliced ​​signal samples, vibration acceleration signal samples, and fault category samples. The attention grid performs the following initial feature extraction process for each signal component: The signal components are input into a preset feature extraction formula to obtain the initial features output by the feature extraction formula; The feature extraction formula includes: Among them, E g Let represent the initial features corresponding to the g-th signal component, Tanh represent the activation function, W represent the coefficients, and b be a constant. Let g represent the g-th signal component, and ω represent the center frequency. The attention grid performs the following probability calculation process for each of the initial features: The initial features are input into the corresponding shallow neural network. The selection probability is obtained through the selection formula in the shallow neural network, and the non-selection probability is obtained through the non-selection formula in the shallow neural network. The selection probability and the non-selection probability are input into the classification formula to obtain the probability that the initial features are selected for classification. The selection formula includes: Where, p k This represents the probability of selection. This represents the network parameters of the k-th shallow neural network when the initial features are selected. This represents the network constant of the k-th shallow neural network when the initial features are selected; The unselected formulas include: Where, n k This indicates the probability of not selecting. This represents the network parameters of the k-th shallow neural network when the initial feature is not selected. This represents the network constant of the k-th shallow neural network when the initial feature is not selected; The classification formulas include: Among them, a k This represents the probability used for classification; The objective function of the attention network includes: Where A represents the attention matrix, θ = <θ l θ α >,θ l and θ α All represent the network parameters of the attention network, R(θ) represents the L2 norm, and λ represents the strength of the regularization. The network parameter is θ l The corresponding network layer, Y represents the predicted attention matrix, and Y is the actual attention matrix.

2. The fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features according to claim 1, characterized in that, The obtained spliced ​​signal is input into an attention network to obtain a target attention matrix output by the attention network corresponding to the vibration acceleration signal, including: The spliced ​​signal is input into the attention network, and the weight of each signal component is calculated through the attention network. Based on the weight, the attention matrix corresponding to the spliced ​​signal is obtained. The product of the attention matrix and the spliced ​​signal is calculated to obtain the attention-weighted signal matrix, and the attention-weighted signal matrix is ​​determined as the target attention matrix.

3. The fault diagnosis method for internal combustion engines based on the fusion of time-domain and frequency-domain features according to claim 2, characterized in that, The attention network calculates the weights of each signal component, and the attention matrix corresponding to the concatenated signal is obtained based on the weights, including: Extract the initial features of each of the signal components; Assign a corresponding shallow neural network to each initial feature, and determine the probability that the initial feature is selected by the shallow neural network for classification. The attention matrix is ​​obtained by combining the probabilities used for classification corresponding to each initial feature.

4. The internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion according to claim 1, characterized in that, Based on a preset center frequency, the vibration acceleration signal is decomposed using a variational mode decomposition method to obtain multiple signal components, including: The vibration acceleration signal is decomposed using the variational mode decomposition method to obtain multiple signal components; The signal components are converted into analytic signals based on the Hilbert transform and Dirac function. The analytical signal is multiplied by an exponential term tuned to the corresponding center frequency to shift the spectrum of the signal component to the baseband, and the bandwidth of the signal component is estimated by Gaussian smoothing. The decomposition is considered complete when the sum of the bandwidths of the multiple signal components is minimized, and the vibration acceleration signal can be obtained by reconstructing the signal components.

5. The internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion according to claim 1, characterized in that, The training process of the fault prediction model includes: Obtain model training samples, wherein the model training samples include: the spliced ​​signal samples, the vibration acceleration signal samples, and the fault category samples; The spliced ​​signal sample is input into the attention network of the fault prediction model to obtain a target attention matrix sample output by the attention network corresponding to the vibration acceleration signal sample; the target attention matrix sample is input into the time-domain feature extraction network of the fault prediction model to obtain the predicted time-domain features output by the time-domain feature extraction network; the vibration acceleration signal sample is input into the frequency-domain feature extraction network of the fault prediction model to obtain the predicted frequency-domain features output by the frequency-domain feature extraction network; the time-domain features and the frequency-domain features are fused to obtain fused features, and a predicted fault diagnosis result is obtained based on the fused features; the consistency between the fault category sample and the fault diagnosis result is determined; the objective function of the attention network, the network parameters of the time-domain feature extraction network, and the network parameters of the frequency-domain feature extraction network are adjusted based on the consistency until the consistency reaches a preset consistency.

6. The internal combustion engine fault diagnosis method based on time-domain and frequency-domain feature fusion according to claim 5, characterized in that, The temporal feature extraction network includes a residual neural network structure; The training process of the fault prediction model also includes: During the training process, if the number of iterations exceeds the preset number and the consistency does not reach the preset consistency, a residual layer of the residual neural network structure is added, and the model is trained based on the modified fault prediction model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the internal combustion engine fault diagnosis method based on the fusion of time-domain and frequency-domain features as described in any one of claims 1 to 6.

Images