Hydraulic plunger pump intelligent fault diagnosis method and system based on multi-source information fusion

Abstract

The invention relates to the technical field of fault diagnosis, and discloses a hydraulic plunger pump intelligent fault diagnosis method and system based on multi-source information fusion. The method comprises the following steps: (1) multi-source signal acquisition and preprocessing: acquiring multi-sensor signals under multiple working conditions based on a hydraulic plunger pump experiment bench, and constructing a data set with time-frequency characteristics by adopting continuous wavelet transform for noise reduction; (2) parallel depth feature extraction: designing an encoder fusing a residual network and a self-attention mechanism, and adaptively extracting key fault features from a time-frequency diagram; (3) multi-modal feature collaborative fusion: constructing a multi-scale fusion module, and realizing complementary enhancement of multi-source information in semantic and detail levels in combination with a variational auto-encoder and a cross attention and gating mechanism; and (4) self-adaptive evidence decision making: introducing a support weight mechanism to dynamically evaluate evidence reliability, and completing credible fusion and output of a diagnosis result. According to the method, the fault identification capability and the system reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of hydraulic piston pump fault diagnosis, and specifically to an intelligent fault diagnosis method and system for hydraulic piston pumps based on multi-source information fusion. Background Technology

[0002] Hydraulic piston pumps, as crucial components of hydraulic systems, are widely used in aerospace, deep-sea and marine engineering, and construction machinery, playing a key role in the normal operation of mechanical systems. Because hydraulic piston pumps frequently operate at high speeds and heavy loads, friction components inevitably experience varying degrees of wear. Therefore, real-time condition monitoring and fault diagnosis are essential to ensuring their stable operation.

[0003] Single sensor signals may lead to missing features or inaccurate information, affecting diagnostic results. Research on fault diagnosis methods based on multi-source information fusion theory has attracted widespread attention. The introduction of deep learning reduces the reliance of traditional fusion methods on manual features, making it easier to build higher-performance information fusion-based machine learning models. These models collaboratively analyze information from different sensors, extracting more comprehensive and complementary fault features, thereby significantly improving the accuracy and reliability of diagnostic results.

[0004] Existing information fusion methods have some limitations in addressing the fault diagnosis of hydraulic piston pumps. First, fault signals are susceptible to interference from noise and other factors, making it difficult to effectively separate and extract key fault features, thus affecting the model's ability to represent fused features. Furthermore, existing methods often employ a single-layer fusion architecture, which struggles to effectively extract latent features from multi-source data, resulting in a narrow scope of information utilization and insufficient adaptability to fault diagnosis under varying operating conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent fault diagnosis method and system for hydraulic piston pumps based on multi-source information fusion. By designing feature fusion, it solves the problem of feature loss or inaccuracy in fault information under a single mode. To further improve diagnostic accuracy, an adaptive decision framework is constructed, which can dynamically adjust the weights of each evidence source according to the reliability of signals from different modes. This method can maintain high diagnostic accuracy even when a sensor signal is subject to noise interference and has good cross-condition versatility, thus adapting to diverse fault diagnosis scenarios.

[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0007] A method for intelligent fault diagnosis of hydraulic piston pumps based on multi-source information fusion includes the following steps:

[0008] The time-domain signals of two different physical modes of the hydraulic piston pump under operating conditions are acquired synchronously; time-frequency analysis is performed on the time-domain signals of each mode, and the results are converted into two-dimensional time-frequency feature maps.

[0009] Deep learning networks are used to extract features from the time-frequency feature maps of each modality to obtain deep features for each modality. Attention mechanisms are used to weight and focus the deep features, and a classifier is used to output the single-modality fault prediction vector corresponding to each modality.

[0010] Deep features from two different modalities are extracted and fused to obtain a local feature fusion map and a global feature fusion map. After attention refinement and fusion processing of the local feature fusion map and the global feature fusion map, the fault prediction vector of the fused signal is output by a classifier.

[0011] Based on the single-mode fault prediction vector and the fault prediction vector of the fused signal, the information entropy of each prediction vector is calculated to obtain the inverse entropy weight, and the comprehensive similarity between any two prediction vectors is calculated to obtain the support weight. According to the inverse entropy weight and the support weight, the fusion weight of each prediction vector is dynamically allocated and weighted fusion is performed. Evidence theory is used to combine the evidence of the two single-mode fault prediction vectors.

[0012] The results of weighted fusion are combined with the results of evidence combination to obtain the probability distribution of fault types. The type corresponding to the maximum probability is used as the fault diagnosis result of the hydraulic piston pump.

[0013] Furthermore, the deep learning network is a lightweight residual network, which sequentially includes a convolutional layer, a first layer, a second layer, and a third layer. The first layer includes one residual block, using ReLU as the activation function. The second layer includes two residual blocks, using ReLU as the activation function. The third layer includes one residual block, using Maxout as the activation function. Each residual block is composed of a depthwise separable convolutional layer, a batch normalization layer, and an activation function in sequence. The attention mechanism is a multi-head self-attention mechanism.

[0014] Furthermore, deep features from two different modalities are extracted and fused to obtain a local feature fusion map and a global feature fusion map, as detailed below:

[0015] The time-frequency feature maps of each modality are extracted and output features are concatenated at the second level, and shallow fusion features are obtained. The shallow fusion features are then decoded by the first decoder to obtain a local feature fusion map.

[0016] The deep features of each modality output by the deep learning network are concatenated to obtain deep fusion features; the deep fusion features are then processed by a second decoder based on a variational autoencoder to obtain a global feature fusion map.

[0017] Furthermore, after attention refinement and dimensionality reduction processing of the local feature fusion map and the global feature fusion map, the fault prediction vector of the fused signal is output by the classifier, as follows:

[0018] By utilizing a cross-attention mechanism, the local feature fusion map and the global feature fusion map interact to obtain two sets of enhanced features, denoted as... and ;

[0019] Enhanced features and enhanced features The enhanced features and the original features are fused in the same proportion using residual connections to preserve the information of the original features, thus obtaining fused enhanced features. and fusion enhancement features ;

[0020] Multimodal fusion features containing both global and local features are obtained through the Sigmoid gating mechanism. The fusion method and weight gate calculation are as follows:

[0021] ,

[0022] ,

[0023] In the formula: Indicates fusion enhancement features and fusion enhancement features Perform the splicing operation; This indicates that the concatenated feature maps are subjected to a convolution operation; σ is the Sigmoid function.

[0024] Employing a convolutional attention module to fuse multimodal features F fusion Attention is refined in both channels and space, and then pooling and regularization are performed before the classifier outputs a multimodal fusion fault prediction vector.

[0025] Furthermore, the information entropy of each predicted vector is calculated to obtain the inverse entropy weight, and the comprehensive similarity between any two predicted vectors is calculated to obtain the support weight, as follows:

[0026] The fault prediction vectors of the two single-mode fault prediction vectors and the fault prediction vector of the fused signal are respectively processed by the Softmax activation function to obtain three basic probability assignments (BPAs), denoted as m. i When i=1, it represents the probability allocation corresponding to the first mode fault prediction vector; when i=2, it represents the probability allocation corresponding to the second mode fault prediction vector; when i=3, it represents the probability allocation corresponding to the multi-mode fusion fault prediction vector.

[0027] Calculate m i Information entropy According to information entropy Calculate its corresponding inverse entropy weight ;

[0028] Let set M = {m i} Calculate the JS divergence of any two predicted vectors in set M, convert it into JS similarity, combine it with cosine similarity to capture global feature alignment, and fuse them with weight α to obtain the similarity evaluation value of any two predicted vectors. , represented as:

[0029] ,

[0030] ,

[0031] In the formula, i represents a prediction vector in set M, and j represents another prediction vector in set M that is different from i. ; This represents the JS similarity between two predicted vectors; This represents the cosine similarity between two predicted vectors. This represents the JS divergence between two prediction vectors;

[0032] The support score of each predicted vector is obtained by summing its similarity to other predicted vectors, and then normalizing the result to obtain the support weights of each predicted vector. , represented as:

[0033] ,

[0034] In the formula, This represents the support weights of the i-th prediction vector; This represents the sum of the similarity evaluation values ​​of the i-th predicted vector and other predicted vectors; This represents the sum of the similarity evaluation values ​​of any two predicted vectors in set M.

[0035] Furthermore, based on the entropy inverse weight and support weight, the fusion weights of each prediction vector are dynamically allocated and weighted fusion is performed, as follows:

[0036] The final fusion weight w for each prediction vector is obtained by combining the inverse entropy weight and the support weight. i :

[0037] ,

[0038] Each prediction vector is represented as logit i The weighted sum of the three prediction vectors is obtained. .

[0039] Furthermore, by combining the weighted fusion results with the evidence combination results, the probability distribution of fault types is obtained, as follows:

[0040] Evidence theory is used to combine the evidence from two single-mode fault prediction vectors to obtain the combined evidence result m. DS ;

[0041] The weighted summation vector is normalized using the Softmax function to obtain a preliminary probability distribution. This preliminary probability distribution is then added to the DS fusion result scaled by the adjustment factor β to obtain the hydraulic piston pump fault diagnosis result.

[0042] ,

[0043] In the formula: β is the adjustment factor; L is the result of weighted summation of the three prediction vectors.

[0044] A multi-source information fusion-based intelligent fault diagnosis system for hydraulic piston pumps, used to implement the method, includes:

[0045] The signal acquisition and preprocessing module is used to acquire multimodal time-domain signals of the hydraulic piston pump and perform time-frequency transformation on the signals to generate a two-dimensional time-frequency feature map.

[0046] The single-mode feature extraction module is used to extract deep features from the time-frequency feature maps of each mode and output a single-mode fault prediction vector.

[0047] The multimodal feature fusion module is used to fuse deep features from different modalities and output a multimodal fused fault prediction vector.

[0048] An adaptive decision fusion module is used to dynamically weight and combine evidence theory with the single-modal and multi-modal fused fault prediction vectors;

[0049] The diagnostic output module is used to output the final fault diagnosis type based on the result of the dynamic weight fusion and evidence theory combination.

[0050] Furthermore, the multimodal feature fusion module includes:

[0051] Feature splicing unit, used to splice features of different modalities and different levels;

[0052] The decoder unit includes a first decoder for generating a local feature fusion map and a second decoder for generating a global feature fusion map;

[0053] Cross-attention and gating fusion units are used to achieve feature interaction and adaptive weighted fusion;

[0054] The attention refinement and output unit is used to refine the fused features and output the fused fault prediction vector.

[0055] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion.

[0056] The beneficial effects of this invention are as follows:

[0057] The intelligent fault diagnosis method and system for hydraulic piston pumps based on multi-source information fusion described in this invention utilizes multi-sensor technology to collect real-time signals of the hydraulic piston pump's operating status, ensuring that different modal data correspond to the same fault type. Continuous wavelet transform is used to denoise the original data, obtaining time-frequency diagrams of different modal signals, effectively separating the frequency components of fault features and noise background, thus enhancing fault feature representation. A residual network is used to mine deep fault features, and a multi-head self-attention mechanism is employed to increase the weight of key fault features, effectively suppressing noise interference. The feature fusion module enables complementary fault feature information between different modal signals, solving the problem of feature loss or inaccuracy that may exist in single-modal information. The adaptive decision framework effectively identifies highly reliable evidence sources, enhancing the robustness and accuracy of the diagnostic results. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are some embodiments of the present invention. For those skilled in the art, it is obvious that other drawings can be obtained from these drawings without creative effort.

[0059] Figure 1 This is a flowchart of the intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion as described in this invention.

[0060] Figure 2 This is a structural diagram of the model described in this invention.

[0061] Figure 3 This is a structural diagram of the encoder described in this invention.

[0062] Figure 4 This is a structural diagram of the second decoder described in this invention.

[0063] Figure 5 This is a structural diagram of the Feature Fusion Module (MSCA) described in this invention.

[0064] Figure 6 This is a diagram of the Adaptive Decision-Making Framework (AWDS) described in this invention.

[0065] Figure 7 This is a diagram showing the fault diagnosis results described in this invention. Detailed Implementation

[0066] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0067] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "axial," "radial," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0068] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0069] like Figure 1 As shown, the intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion described in this invention specifically includes the following steps:

[0070] S01: Collect multimodal signal data of hydraulic piston pump under different working loads. The multimodal signal data includes vibration signals and sound signals.

[0071] In this embodiment, a hydraulic piston pump test bench was constructed with a rated speed of 1470 r / min and a rated pressure of 31.5 MPa. A vibration acceleration sensor was installed along the z-axis of the hydraulic piston pump housing end cover, and a sound sensor was placed at the horizontal height of the pump's rotation axis, with a distance of no more than 0.5 m from the pump body. Time-domain signals of five typical states of the hydraulic piston pump under different operating pressures were collected as raw state data for analysis. The sampling frequency was 50 kHz, and each sample consisted of 1024 data points. The five typical states were: normal state, swashplate wear, slipper wear, slipper failure, and center spring failure.

[0072] S02: Denoising the multimodal signal data is performed using time-frequency analysis. Continuous Wavelet Transform (CWT) is used to denoise the original acquired signal, converting the one-dimensional time signal into a two-dimensional time-frequency graph. This results in a two-dimensional time-frequency graph that highlights hidden feature information. The complex Morlet wavelet (cmor3-3) is selected as the wavelet basis function, with both bandwidth and center frequency set to 3, and the scale sequence length set to 256.

[0073] Fault databases for vibration and sound signals were constructed separately. In this embodiment, for the vibration sensor, 1200 samples were collected for each of the five fault types under each working pressure, for a total of 6000 samples. The sample database was divided into a training sample set and a test sample set in a 3:7 ratio. For the sound sensor, the amount of data collected and the construction of the sample database were the same as those for the vibration sensor, as shown in Table 1.

[0074] Table 1. Dataset creation and corresponding labels

[0075]

[0076] S03: For multimodal signal data, the encoder is used to extract the deep features of the single-mode signal, the attention mechanism is used to obtain key fault feature information, and the fully connected layer outputs the fault prediction vectors of different modes of signal.

[0077] The encoder is a lightweight residual encoder network. The input to the lightweight residual encoder network is a 3-channel time-frequency plot generated by CWT. Figure 3As shown, the lightweight residual encoder network sequentially comprises convolutional layers, Layer 1, Layer 2, and Layer 3. The convolutional layers are divided into 64 7×7 convolutional kernels, a batch normalization (BN) layer, and a ReLU activation function. Layer 1 includes one residual block with 64 output channels, using the ReLU activation function. Layer 2 includes two residual blocks with 128 output channels, also using the ReLU activation function. Layer 3 includes one residual block with 256 output channels, using the Maxout activation function to enhance nonlinear fitting capability and avoid gradient saturation.

[0078] Each residual block consists of a depthwise separable convolutional layer, a batch normalization (BN) layer, and an activation function in sequence. Layers 1 and 2 use ReLU activation to avoid the vanishing gradient problem and simplify the computation process, while Layer 3 employs a learnable Maxout activation function to prevent insufficient model convergence caused by some inputs falling into hard saturation regions and output offset. Deep features of the vibration signal and deep features of the sound signal are extracted using encoders, respectively. In this embodiment, the encoder parameters are shown in Table 2.

[0079] Table 2 Model hyperparameters

[0080]

[0081] After obtaining the deep feature map, a multi-head self-attention mechanism is introduced for feature refinement. Specifically, the deep feature map is projected to generate multiple sets of query, key, and value vectors. By calculating the dot product of the query and key and scaling it, an attention weight score is obtained. This weight is then used to perform a weighted summation of the value vectors to obtain the key fault feature information. In this embodiment, the number of attention heads is set to 4.

[0082] The key fault feature information is reduced in dimensionality using global average pooling and then fed into a fully connected classifier to obtain the vibration modal fault prediction vector Logit. v Logit (Sound Modal Fault Prediction Vector) a The number of input channels in the fully connected layer is 256.

[0083] S04: The deep features of the vibration signal and the deep features of the sound signal are decoded by the decoder to obtain a local feature fusion map F1 and a global feature fusion map F2, respectively, for multimodal feature fusion. The local feature fusion map F1 and the global feature fusion map F2 are processed by the feature fusion module MSCA to obtain a fused feature that is complementary to the global and local features. The fault prediction vector of the fused signal is output by the fully connected layer, as follows:

[0084] S4.1: As Figure 2 As shown, the output features of vibration and sound signals in Layer 2 of the encoder are extracted and concatenated along the channel dimension to obtain shallow fusion features. These features are then decoded by the first decoder (Decoder1) to obtain a local feature fusion map F1. The first decoder uses a convolution, BN, and ReLU algorithm to downsample the input feature map and perform standardization and activation processing. This reduces the feature map size and enhances nonlinear expressive power while keeping the number of channels constant.

[0085] S4.2: As Figure 4 As shown, the output features of vibration and sound signals in Layer 3 of the encoder are extracted and concatenated to obtain deep fusion features. These features are then decoded by the second decoder (Decoder2) to obtain the global feature fusion map F2. The second decoder is a VAE (Variational Auto-Encoder) structure. The VAE encoder learns the latent distribution of the input features, estimates the distribution of the latent variable z using training set samples, calculates the mean μ and log-variance σ of the latent distribution through a fully connected layer, and uses the reparameter trick to sample ε, which satisfies the standard normal distribution, at the input layer. The latent variable z is obtained through the following linear transformation:

[0086] ,

[0087] The second decoder reconstructs z into a global feature fusion map F2 aligned with the size of F1. Specifically, the output latent variable z is mapped back to the shape of the original feature map through a fully connected layer, and the image is reconstructed through two layers of deconvolution to achieve effective alignment of multi-source features. This overcomes the problem that traditional single convolutional layers are difficult to align features due to their lack of ability to model the global distribution. At the same time, the original local information is preserved through skip connections, which improves the reconstruction effect.

[0088] S4.3: The local feature fusion map F1 and the global feature fusion map F2 are processed by the feature fusion module MSCA to obtain multimodal fusion features that are complementary to global and local features.

[0089] like Figure 5As shown, the feature fusion module MSCA includes a cross-attention mechanism, a convolutional attention module CBAM, a sigmoid gating mechanism, a GELU activation function, a dropout layer, and a fully connected layer;

[0090] A cross-attention mechanism is used to facilitate information exchange between the local feature fusion map F1 and the global feature fusion map F2. The global feature fusion map F2 is used as the query Q1 for the local feature fusion map F1, and F1 serves as the key (K1) and value (V1). Attention weights are calculated and F1 is updated to obtain the enhanced features. Conversely, using the local feature fusion map F1 as the query Q2 of the global feature fusion map F2, with the global feature fusion map F2 as the key (K2) and value (V2), the enhanced features are obtained. This process enhances the complementarity of features.

[0091] Enhanced features and enhanced features The enhanced features and original features are fused in the same proportion using residual connections to preserve the information of the original features. Layer normalization (LN) is then applied to eliminate feature distribution differences between different modalities. Finally, the GELU activation function is used to enhance the expressive power of the features, resulting in fused enhanced features. and fusion enhancement features .

[0092] Multimodal fusion features containing both global and local features are obtained through the Sigmoid gating mechanism. The fusion method and weight gate calculation are as follows:

[0093] ,

[0094] ,

[0095] In the formula, Indicates fusion enhancement features and fusion enhancement features Perform the splicing operation; This indicates that the concatenated feature maps are subjected to a convolution operation; σ is the Sigmoid function.

[0096] Multimodal fusion feature F fusion Processed via the Convolutional Attention Module (CBAM). CBAM sequentially performs channel attention (focusing on important feature channels) and spatial attention (focusing on important spatial regions) calculations to further enhance fault-sensitive features and suppress irrelevant noise.

[0097] Fusion output: The features refined by CBAM are subjected to global average pooling and Dropout regularization. Finally, a fully connected layer is passed to output the multimodal fused fault prediction vector Logit. f .

[0098] S05: Logit of vibration modal fault prediction vector v Logit, a vector for predicting audio modal faults. a Logit, a multimodal fusion fault prediction vector f The adaptive weighting decision framework (AWDS) is used to assess the reliability of each prediction vector, adjust the vector weights, and perform weighted fusion to obtain the fault diagnosis result of the hydraulic piston pump. Figure 6 As shown, the details are as follows:

[0099] S5.1: The vibration modal fault prediction vector Logit v Logit, a vector for predicting audio modal faults. a Logit, a multimodal fusion fault prediction vector f The three BPAs obtained by applying the Softmax activation function are denoted as m. v m a and m f Let set M = {m} v m a m f Let m be the prediction vector in set M. i When i=1, it represents the probability allocation m corresponding to the vibration mode fault prediction vector. v When i=2, it represents the probability allocation m corresponding to the sound modal fault prediction vector. a When i=3, it represents the probability allocation m corresponding to the multimodal fusion fault prediction vector. f ; Calculate m i Information entropy The smaller the entropy value, the higher the certainty of the prediction, according to information entropy. Calculate the inverse entropy weight of the i-th element. ;

[0100] ,

[0101] ,

[0102] In the formula, d represents the number of fault types; m i (k) represents the confidence or probability measure that the hydraulic piston pump is in the k-th type of fault state in the signal corresponding to the prediction vector i;

[0103] S5.2: Calculate the Jensen-Shannon divergence (JS) between any two predicted vectors in set M to measure conflict, transform it into JS similarity to reflect local prediction consistency, combine it with cosine similarity to capture global feature alignment, and fuse them with weight α to obtain the similarity evaluation value of any two predicted vectors. , represented as:

[0104] ,

[0105] ,

[0106] In the formula, i represents a prediction vector in set M, and j represents another prediction vector in set M that is different from i. ; This represents the JS similarity between two predicted vectors; This represents the cosine similarity between two predicted vectors. This represents the JS divergence between two prediction vectors;

[0107] For set M, weights are dynamically assigned based on the collaborative relationships between models. The support of each predicted vector is obtained by calculating the sum of its similarities with other predicted vectors, and the support weights of each predicted vector are obtained after normalization. , represented as:

[0108] ,

[0109] In the formula, This represents the support weights of the i-th prediction vector; This represents the sum of the similarity evaluation values ​​of the i-th predicted vector with other elements; It represents the sum of the similarity evaluation values ​​of any two predicted vectors in set M;

[0110] The final fusion weight w for each prediction vector is obtained by combining the inverse entropy weight and the support weight. i :

[0111] ,

[0112] The three original prediction vectors are weighted and summed: L = w1⋅logit v +w2⋅logit a +w3⋅logit f .

[0113] Logit, a prediction vector for vibration modal faults, is applied using the Dempster-Shafer Evidence Theory. v Logit, a vector for predicting audio modal faults. aBy fusing the evidence, we obtain the combined evidence result m. DS = .

[0114] The weighted summation vector is normalized using the Softmax function to obtain a preliminary probability distribution. This preliminary probability distribution is then added to the DS fusion result scaled by the adjustment factor β to obtain the hydraulic piston pump fault diagnosis result.

[0115]

[0116] In the formula, β is the adjustment factor.

[0117] The diagnostic results in the examples are as follows Figure 7 The confusion matrix is ​​shown.

[0118] like Figure 2 As shown, the present invention provides an intelligent fault diagnosis system for hydraulic piston pumps based on multi-source information fusion, comprising a signal acquisition and preprocessing module, a single-modal feature extraction and diagnosis module, a multi-modal feature fusion module, an adaptive decision fusion module, and a diagnostic output module. The signal acquisition and preprocessing module acquires multi-modal time-domain signals from the hydraulic piston pump and performs time-frequency transformation on the signals to generate a two-dimensional time-frequency feature map. The single-modal feature extraction and diagnosis module extracts deep features from the time-frequency feature maps of each modality and outputs a single-modal fault prediction vector. The multi-modal feature fusion module fuses deep features from different modalities and outputs a multi-modal fused fault prediction vector. The adaptive decision fusion module performs dynamic weight fusion and evidence theory combination on the single-modal and multi-modal fused fault prediction vectors. The diagnostic output module outputs the final fault diagnosis type based on the result of the dynamic weight fusion and evidence theory combination.

[0119] The multimodal feature fusion module includes a feature splicing unit, a decoder unit, a cross-attention and gating fusion unit, and an attention refinement and output unit. The feature splicing unit is used to splice features from different modalities and different levels. The decoder unit includes a first decoder for generating a local feature fusion map and a second decoder for generating a global feature fusion map. The cross-attention and gating fusion unit is used to realize feature interaction and adaptive weighted fusion. The attention refinement and output unit is used to refine the fused features and output a fused fault prediction vector.

[0120] A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion.

[0121] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0122] The detailed descriptions listed above are merely specific illustrations of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent fault diagnosis of hydraulic piston pumps based on multi-source information fusion, characterized in that, Includes the following steps: The time-domain signals of two different physical modes of the hydraulic piston pump under operating conditions are acquired synchronously; time-frequency analysis is performed on the time-domain signals of each mode, and the results are converted into two-dimensional time-frequency feature maps. Deep learning networks are used to extract features from the time-frequency feature maps of each modality to obtain deep features for each modality. Attention mechanisms are used to weight and focus the deep features, and a classifier is used to output the single-modality fault prediction vector corresponding to each modality. Deep features from two different modalities are extracted and fused to obtain a local feature fusion map and a global feature fusion map. After attention refinement and dimensionality reduction processing of the local feature fusion map and the global feature fusion map, the fault prediction vector of the fused signal is output by a classifier. Based on the single-mode fault prediction vector and the fault prediction vector of the fused signal, the information entropy of each prediction vector is calculated to obtain the inverse entropy weight, and the comprehensive similarity between any two prediction vectors is calculated to obtain the support weight. According to the inverse entropy weight and the support weight, the fusion weight of each prediction vector is dynamically allocated and weighted fusion is performed. Evidence theory is used to combine the evidence of the two single-mode fault prediction vectors. The results of weighted fusion are combined with the results of evidence combination to obtain the probability distribution of fault types. The type corresponding to the maximum probability is used as the fault diagnosis result of the hydraulic piston pump.

2. The intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion according to claim 1, characterized in that, The deep learning network is a lightweight residual network, which sequentially includes a convolutional layer, a first layer, a second layer, and a third layer. The first layer includes one residual block, using ReLU as the activation function. The second layer includes two residual blocks, using ReLU as the activation function. The third layer includes one residual block, using Maxout as the activation function. Each residual block consists of a depthwise separable convolutional layer, a batch normalization layer, and an activation function in sequence. The attention mechanism is a multi-head self-attention mechanism.

3. The intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion according to claim 2, characterized in that, Deep features from two different modalities are extracted and fused to obtain a local feature fusion map and a global feature fusion map, as detailed below: The time-frequency feature maps of each modality are extracted and output features are concatenated in the channel dimension to obtain shallow fusion features; the shallow fusion features are decoded by the first decoder to obtain a local feature fusion map. The deep features of each modality output by the deep learning network are concatenated to obtain deep fusion features; the deep fusion features are then processed by a second decoder based on a variational autoencoder to obtain a global feature fusion map.

4. The intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion according to claim 1, characterized in that, After attention refinement and dimensionality reduction processing of the local feature fusion map and the global feature fusion map, the fault prediction vector of the fused signal is output by the classifier, as follows: By utilizing a cross-attention mechanism, the local feature fusion map and the global feature fusion map interact to obtain two sets of enhanced features, denoted as... and ; Enhanced features and enhanced features The enhanced features and the original features are fused in the same proportion using residual connections to preserve the information of the original features, thus obtaining fused enhanced features. and fusion enhancement features ; A multimodal fusion feature containing both global and local features is obtained through the Sigmoid gating mechanism. The fusion method and weight gate calculation are as follows: ， ， In the formula: Indicates fusion enhancement features and fusion enhancement features Perform the splicing operation; This indicates that the concatenated feature maps are subjected to a convolution operation; σ is the Sigmoid function. Employing a convolutional attention module to fuse multimodal features F fusion Attention is refined in both channels and space, and then pooling and regularization are performed before the classifier outputs a multimodal fusion fault prediction vector.

5. The intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion according to claim 1, characterized in that, The information entropy of each predicted vector is calculated to obtain the inverse entropy weight, and the comprehensive similarity between any two predicted vectors is calculated to obtain the support weight, as follows: The fault prediction vectors of the two single-mode fault prediction vectors and the fault prediction vector of the fused signal are respectively processed by the Softmax activation function to obtain three basic probability assignments (BPAs), denoted as m. i When i=1, it represents the probability allocation corresponding to the first mode fault prediction vector; when i=2, it represents the probability allocation corresponding to the second mode fault prediction vector; when i=3, it represents the probability allocation corresponding to the multi-mode fusion fault prediction vector. Calculate m i Information entropy According to information entropy Calculate its corresponding inverse entropy weight ; Let set M = {m i } Calculate the JS divergence of any two predicted vectors in set M, convert it into JS similarity, combine it with cosine similarity to capture global feature alignment, and fuse them with weight α to obtain the similarity evaluation value of any two predicted vectors. , is represented as: ， ， In the formula, i represents a prediction vector in set M, and j represents another prediction vector in set M that is different from i. ; This represents the JS similarity between two predicted vectors; This represents the cosine similarity between two predicted vectors. This represents the JS divergence between two prediction vectors; The support score of each predicted vector is obtained by summing its similarity to other predicted vectors, and then normalizing the result to obtain the support weights of each predicted vector. , is represented as: ， In the formula, This represents the support weights of the i-th prediction vector; This represents the sum of the similarity evaluation values ​​of the i-th predicted vector and other predicted vectors; This represents the sum of the similarity evaluation values ​​of any two predicted vectors in set M.

6. The intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion according to claim 5, characterized in that, Based on the entropy inverse weight and support weight, the fusion weights of each prediction vector are dynamically allocated and weighted fusion is performed, as follows: The final fusion weight w for each prediction vector is obtained by combining the inverse entropy weight and the support weight. i : ， Each prediction vector is represented as logit i The weighted sum of the three prediction vectors is obtained. .

7. The intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion according to claim 1, characterized in that, By combining the weighted fusion results with the evidence combination results, the probability distribution of fault types is obtained, as follows: Evidence theory is used to combine the evidence from two single-mode fault prediction vectors to obtain the combined evidence result m. DS ; The weighted summation vector is normalized using the Softmax function to obtain a preliminary probability distribution. This preliminary probability distribution is then added to the DS fusion result scaled by the adjustment factor β to obtain the hydraulic piston pump fault diagnosis result. ， In the formula: β is the adjustment factor; L is the result of weighted fusion.

8. A hydraulic piston pump intelligent fault diagnosis system based on multi-source information fusion for implementing the method of any one of claims 1-7, characterized in that, include: The signal acquisition and preprocessing module is used to acquire multimodal time-domain signals of the hydraulic piston pump and perform time-frequency transformation on the signals to generate a two-dimensional time-frequency feature map. The single-mode feature extraction and diagnosis module is used to extract deep features from the time-frequency feature maps of each mode and output a single-mode fault prediction vector. The multimodal feature fusion module is used to fuse deep features from different modalities and output a multimodal fused fault prediction vector. An adaptive decision fusion module is used to dynamically weight and combine evidence theory with the single-modal and multi-modal fused fault prediction vectors; The diagnostic output module is used to output the final fault diagnosis type based on the result of the dynamic weight fusion and evidence theory combination.

9. The system according to claim 8, characterized in that, The multimodal feature fusion module includes: Feature splicing unit, used to splice features of different modalities and different levels; The decoder unit includes a first decoder for generating a local feature fusion map and a second decoder for generating a global feature fusion map; Cross-attention and gating fusion units are used to achieve feature interaction and adaptive weighted fusion; The attention refinement and output unit is used to refine the fused features and output the fused fault prediction vector.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the intelligent fault diagnosis method for hydraulic piston pumps based on multi-source information fusion as described in any one of claims 1-7.

Images