PreDyn-ST model-based EHA performance degradation prediction method and system

Abstract

The invention relates to an EHA performance degradation prediction method and system based on a PreDys-ST model, and provides a performance degradation prediction method and system of an EHA hydraulic system based on a deep learning method (PreDys-ST) model. The method comprises the following steps: S1, building an EHA test bed; s2, signal acquisition; s3, obtaining a degradation label; s4, a PreDyn-ST model is constructed; s5, carrying out PreDyn-ST model training; and S6, obtaining a service life prediction result of model prediction. On the other hand, the system comprises a data acquisition module, a degradation label definition module, a model construction module, a data division module and a model verification prediction module. According to the method, dynamic changes of the space-time characteristics in the degradation process can be visualized, the changes are explained in combination with a physical degradation mechanism, the interpretability of the model is enhanced, and the credibility of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the field of EHA hydraulic system performance degradation prediction, and in particular relates to an EHA performance degradation prediction method and system based on a PreDyn-ST model. Background Art

[0002] As the main power unit for flight control, the aircraft hydraulic system can provide power for aircraft takeoff, maneuvering, landing gear retraction and braking, etc. It is a key system closely related to flight safety. At present, the most widely used system on civil airliners is the traditional centralized hydraulic system, but it still has disadvantages such as large energy loss and low efficiency. The high-performance electro-hydrostatic actuator (EHA) has the advantages of small size, light weight, fast response speed, high power-to-weight ratio, etc. It can use a compact and reliable independent direct drive method to replace the advantages of the traditional centralized hydraulic system that is prone to defects. It has been used as a redundant system on large civil airliners to a certain extent, and has great application prospects.

[0003] However, currently airlines only record the onboard QAR data and EICAS hydraulic system data as independent parameters, but they have no direct relationship with the health status of the aircraft hydraulic system. Since the changes in the health status of the system still need to be obtained through indirect analysis, correlation analysis or damage conversion of alternating loads, it is difficult to accurately characterize the health status of the aircraft hydraulic system in real time.

[0004] In practical applications, the degradation process of the system will cause changes in the interaction between time characteristics and space characteristics, and the simple superposition of time and space characteristics cannot adapt to these dynamic changes. Especially in different degradation stages, the weights of time and space characteristics should be flexibly adjusted to effectively reflect the transition of the system state. How to dynamically adapt to the model of changes in time and space characteristics to improve the modeling accuracy and robustness of multi-sensor time series data with complex degradation patterns and degradation processes is a key issue that needs to be addressed. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an EHA performance degradation prediction method and system based on the PreDyn-ST model. The method can visualize the dynamic changes of spatiotemporal characteristics during the degradation process, and explain these changes in combination with the physical degradation mechanism, thereby enhancing the interpretability of the model and improving the credibility of the model.

[0006] To achieve the above object, the present invention discloses the following technical solution:

[0007] In a first aspect, the present invention relates to a method for predicting EHA performance degradation based on a PreDyn-ST model, comprising the following steps:

[0008] S1. Build the EHA test bench;

[0009] S2, signal acquisition: collect piston rod displacement sensor signals, force sensor signals, flow and pressure signals of two chambers of hydraulic servo cylinder, servo motor current signals and servo motor speed signals through data acquisition sensors;

[0010] S3, establishing a component function signal set with degradation labels: obtaining system efficiency according to sensor signal data, taking the system efficiency as a degradation label, and establishing a component function signal set with degradation labels;

[0011] S4. Construct PreDyn-ST model: Use the SimCLR architecture of contrastive learning, the GCN spatial feature extraction module based on physical priors, and the Transformer temporal feature extraction module to construct the PreDyn-ST model. The output of the PreDyn-ST model is used as the life prediction result of the system. It includes the following sub-steps:

[0012] S41. Use the SimCLR architecture of contrastive learning to shorten the distance between positive samples in the sliding window between time steps in the multivariate time series of sensor signal data, increase the distance between negative samples, and generate enhanced data.

[0013] S42, inputting the enhanced data into a feature extraction module to perform preliminary feature extraction;

[0014] S43, the extracted features are respectively input into the Transformer temporal feature extraction module and the GCN spatial feature extraction module based on physical prior;

[0015] S44. As the degradation stage changes, the PreDyn-ST model adaptively adjusts the weights of the Transformer temporal feature extraction module and the physical prior-based GCN spatial feature extraction module and outputs the system life prediction results;

[0016] S5, PreDyn-ST model training: After dividing the component function signal set with degradation labels in S3 into a training set, a validation set, and a test set, the PreDyn-ST model is trained and validated, and the model parameters are adjusted to obtain a trained PreDyn-ST model;

[0017] S6. Obtain the life prediction result predicted by the model: collect the multivariate time series sequence of sensor signal data in real time and input it into the PreDyn-ST model trained in S5, output the predicted system life prediction result, judge whether the system life prediction result is lower than the set threshold, and determine the performance degradation state of EHA according to the judgment result.

[0018] Further, in step S4, the SimCLR architecture represents the loss of the overall sample by averaging the losses of all positive sample pairs;

[0019] The loss for a positive pair is defined as:

[0020]

[0021] Among them, sim(z i ,z j ) is z i and z j The cosine similarity between is defined as τ is a temperature parameter; Z is expressed as the feature vector of the positive sample pair; i, j and k are all parameters; 1 [k≠i] is an indicator function that takes the value 1 when k≠i and 0 otherwise; the denominator is summed over all 2N possible pairs in the batch, excluding positive sample pairs;

[0022] The total loss is calculated by averaging the losses of all positive pairs in the batch:

[0023]

[0024] Further, in step S4, in the graph G=(V,ξ,A) formed by the EHA data in the GCN spatial feature extraction module based on physical prior, V is the number of nodes, i.e., the number of sensors; ξ is the edge set between each pair of connected nodes; is the adjacency matrix showing the spatial relationship between nodes; L = I n -D -1 / 2 AD -1 / 2 is the symmetric normalized graph Laplacian matrix, and the diagonal matrix D is calculated based on the adjacency matrix, that is, is the eigenvalue of L;

[0025] The convolution of vertex embedding and filter f is expressed as:

[0026] h=x* g f=U((U T x)⊙(U T f))

[0027] in, is the eigenvector matrix of the Laplace matrix, ⊙ represents the Hadamard product;

[0028] Diagonal matrix g θ , g θ For U T The diagonal matrix representation of f is:

[0029] gθ =diag(U T f)=diag(θ0,θ1,…,θ n-1 )

[0030] Among them, diag is the operation of converting a vector into a diagonal matrix, θ i is each characteristic component after being projected into the frequency domain;

[0031] The initial graph convolution formula can be obtained:

[0032] h=x* g f=Ug θ U T x;

[0033] Simplify graph convolution calculation using Chebyshev polynomials:

[0034]

[0035] in, and are the rescaled eigenvalues ​​and Laplacian matrices, respectively. is the Chebyshev coefficient, is a Chebyshev polynomial of order k, is an explicitly defined convolution kernel.

[0036] Furthermore, in step S4, in the Transformer temporal feature extraction module:

[0037] The self-attention calculation mechanism is expressed as:

[0038]

[0039] Among them, Q, K, and V are query, key, and value matrices respectively. Then the similarity between each element in the sequence and other elements is calculated to determine its attention weight;

[0040] The multi-head attention function is expressed as:

[0041] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0042] head i For each attention head in the multi-head attention (i=1,2,3...k), W o is the output linear transformation matrix;

[0043] Among them, the calculation method of each attention head is:

[0044]

[0045] The formula for preserving the position information of the sequence is:

[0046]

[0047] Among them, pos is the position index, i is the dimension index, and d model is the dimension of the model.

[0048] Furthermore, in said S1, the constructed EHA test bench includes an active part and a load part driven by the EHA, wherein the active part includes a servo motor, a gear pump, a servo hydraulic cylinder and a plurality of data acquisition sensors;

[0049] The load part adopts a top cylinder structure including a motor pump, a relief valve and a plurality of one-way valves.

[0050] On the other hand, the present invention relates to a system for predicting EHA performance degradation based on the PreDyn-ST model, which comprises: a data acquisition module, a degradation label definition module, a model construction module, a data partitioning module and a model verification prediction module:

[0051] Data acquisition module: collects piston rod displacement sensor signals, force sensor signals for displaying load size, flow and pressure signals of the two chambers of the hydraulic servo cylinder, servo motor current signals, and servo motor speed signals through sensors of the EHA test bench;

[0052] Degradation label definition module: organizes the signals collected in the EHA test bench and outputs them into system efficiency as degradation labels;

[0053] Model building module: build PreDyn-ST model;

[0054] Data division module: divide the data in the collected EHA data set into training set, validation set and test set;

[0055] Model verification and prediction module: adjust the model parameters, train and verify the extracted feature labels with the PreDyn-ST model, and perform time series prediction to obtain the performance indicator results of the model prediction.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) Aiming at the temporal dependence and spatial structural changes of multi-sensor time series (MTS) data caused by different stages of the degradation process, this paper proposes a method to capture the changes of spatiotemporal characteristics in the degradation process with dynamic weights. This method can visualize the dynamic changes of spatiotemporal characteristics in the degradation process and interpret these changes in combination with the physical degradation mechanism, which enhances the interpretability of the model and improves the credibility of the model.

[0058] (2) The present invention proposes a PreDyn-ST method based on a pre-training strategy. First, the contrastive learning method is used to perform data enhancement on the sliding window of the time series. Then, the GCN framework is used to build a spatial feature module in the downstream task, and the Transformer is used to build a temporal feature module. Finally, dynamic weights are used to adaptively adjust the spatiotemporal ratio according to changes in the degradation process to achieve degradation prediction.

[0059] (3) The present invention simulated the working conditions of the aircraft hydraulic system by using an electrostatic hydraulic actuator (EHA) performance degradation test bench, carried out accelerated degradation tests, and obtained the degradation data of the EHA system. Based on the proposed method, these degradation data were effectively predicted, demonstrating the applicability and accuracy of the method under actual working conditions. In addition, the generalization ability of the model was verified on the C-MAPSS dataset, and the results showed that the model not only performed well in the EHA system, but also had high accuracy and robustness in degradation prediction tasks in different fields. This experiment further demonstrated the model's potential for wide application in a variety of complex systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flow chart of the EHA performance degradation prediction method based on the PreDyn-ST model of the present invention;

[0061] Figure 2 This is the principle diagram of the EHA test bench of the present invention;

[0062] Figure 3 A schematic diagram of the structure of the comparative learning model for processing MTS data in the present invention;

[0063] Figure 4 It is a schematic diagram of the structure of the GCN spatial feature extraction module based on physical priors of the present invention;

[0064] Figure 5 This is a schematic diagram of the structure of the Transformer time feature extraction module of the present invention;

[0065] Figure 6 This is a schematic diagram of the PreDyn-ST framework flow of the present invention;

[0066] Figure 7Schematic diagram of the overall framework of EHA performance degradation prediction of the present invention;

[0067] Figure 8 Schematic diagram of feature distribution before and after EHA data pre-training of the present invention;

[0068] Fig. 9 This is a comparison chart of the predicted value and the actual value of the EHA efficiency in one day of the present invention;

[0069] Fig.10 This is a comparison chart of the actual RUL and predicted RUL of a single engine in the C-MAPSS test set of the present invention;

[0070] Fig.11 Schematic diagram of the EHA data training and validation loss curve of the present invention;

[0071] Fig.12 It is a schematic diagram of the overall CSI curve of the C-MAPSS training data set of the present invention;

[0072] Fig.13 This is the experimental CSI curve and dynamic weight change diagram of the EHA data set within one day of the present invention. DETAILED DESCRIPTION

[0073] The exemplary embodiments, features and aspects of the present invention will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0074] On the one hand, the present invention provides a method for predicting EHA performance degradation based on the PreDyn-ST model, such as Figure 1-Figure 13 As shown, the EHA performance degradation prediction method based on the PreDyn-ST model includes the following steps:

[0075] S1. Build an EHA test bench.

[0076] S2. Signal acquisition: The piston rod displacement sensor signal, force sensor signal, flow and pressure signals of the two chambers of the hydraulic servo cylinder, servo motor current signal and servo motor speed signal are collected through data acquisition sensors.

[0077] S3. Obtain degradation labels: Obtain corresponding system efficiency according to component function signals, use the system efficiency as a degradation label, and establish a component function signal set with degradation labels.

[0078] In this embodiment, the system efficiency is the efficiency of the motor pump of the EHA hydraulic system, which is stable at 25%-30% under normal working conditions. If the system degrades, the efficiency will decrease. The efficiency of the motor pump of the EHA hydraulic system can be calculated by extracting from the component function signal data.

[0079] S4. Construct PreDyn-ST model: Use the SimCLR architecture of contrastive learning, the GCN spatial feature extraction module based on physical priors, and the Transformer temporal feature extraction module to construct the PreDyn-ST model, and use the output degradation label as the warning value.

[0080] Among them, the GCN spatial feature extraction module based on physical prior and the construction of the adjacency matrix are combined with the similarity between sensor signal nodes and the physical connection information of each component in the EHA system.

[0081] S41. Use the SimCLR architecture of contrastive learning to shorten the distance between positive samples in the sliding window between time steps in the multivariate time series of sensor signals, increase the distance between negative samples, and generate enhanced data.

[0082] S42, inputting the generated enhanced data into the feature extraction layer module for preliminary feature extraction.

[0083] S43. The extracted features are respectively input into the Transformer temporal feature extraction module and the GCN spatial feature extraction module based on physical prior.

[0084] S44. As the degradation stage changes, the change in sensor signal correlation has an important impact on the spatial information transmission capability of the GCN graph. By observing the sensor correlation CSI value, the increase in the CSI value makes the connection between sensors increase the weight of the spatial module. By adaptively adjusting the weights of the time module and the spatial module, the life prediction result of the system is obtained.

[0085] S5, data division: divide the data collected in S2 into training set, validation set, and test set.

[0086] S6. Obtain the performance indicator results predicted by the model: adjust the model parameters, train and verify the extracted feature labels with the PreDyn-ST model, and perform time series prediction to obtain the performance indicator results predicted by the model and determine whether they are lower than the degraded labels.

[0087] The SimCLR architecture of contrastive learning in S4: adopts the time window input method, the input data is a window; a window includes all sensor signals and the specified number of time steps. Contrastive learning can use random transformation to enhance each time series window sample x in the data set twice to generate two related time series windows and Then, the encoder network f(·) is used to extract feature representations from the enhanced time series window and generate a feature vector and Finally, the feature vector is passed to the projection head g(·), g(·). Positive samples are the time steps corresponding to the two time windows; otherwise, they are negative samples.

[0088] In a second aspect, the present invention provides a system for predicting an EHA performance degradation method based on a PreDyn-ST model, which includes: a data acquisition module, a degradation label definition module, a model construction module, a data partitioning module, and a model verification prediction module:

[0089] Data acquisition module: The sensors of the EHA test bench collect piston rod displacement sensor signals, force sensor signals (used to display load size), flow and pressure signals of the two chambers of the hydraulic servo cylinder, servo motor current signals, and servo motor speed signals.

[0090] Degradation label definition module: The signals collected in the EHA test bench are sorted and output into system efficiency as degradation labels.

[0091] Model building module: Build the PreDyn-ST model.

[0092] Data division module: divides the data in the collected EHA dataset into training set, validation set, and test set.

[0093] Model verification and prediction module: adjust the model parameters, train and verify the extracted feature labels with the PreDyn-ST model, and perform time series prediction to obtain the performance indicator results of the model prediction.

[0094] Specifically, the data collection method includes:

[0095] EHA performance degradation test bench, the test bench is divided into two parts, namely the active part of EHA drive and the load part. The active part includes servo motor, gear pump, servo hydraulic cylinder and multiple data acquisition sensors.

[0096] The load part adopts a top cylinder structure, including a motor pump, a relief valve and multiple check valves. The working process of EHA is that the servo motor drives the high-speed gear pump to convert electrical energy into hydraulic energy. The pump pushes the hydraulic oil into the piston chamber of the hydraulic servo cylinder and generates thrust to drive the piston movement. Figure 2shown.

[0097] When the control system issues a command, the servo motor adjusts the output flow and pressure of the pump according to the required movement direction and speed, thereby controlling the movement of the piston. The load part uses a relief valve to adjust the system pressure to achieve the purpose of controlling the load size. The sensor signal is collected by the data acquisition card combined with the written data acquisition software;

[0098] The experiment collected signals from eight sensors. Figure 8 The data collected in the previous ten days are shown, including the piston rod displacement sensor signal, the force sensor signal used to display the load size, the flow and pressure signals of the two chambers of the hydraulic servo cylinder, the servo motor current signal and the servo motor speed signal.

[0099] During the test, the waveform control used a sine wave, the EHA's operating condition was a reciprocating motion with a frequency of 1HZ and a piston stroke of 30cm, and the data sampling frequency was 1000HZ. The test was conducted for 3 hours a day for 84 days, fully recording the performance of the equipment throughout its life cycle.

[0100] In step S3, the system efficiency is used as a degradation label, including:

[0101] Since system efficiency is an important indicator to measure the degree of system degradation, system efficiency is used as a degradation label. If the system efficiency is lower than a certain threshold, it indicates that the system is faulty.

[0102] The system efficiency is calculated from the raw data using a self-written Python program. In order to correspond to the training data, the system average efficiency is calculated every 10 cycles.

[0103] Furthermore, the method of constructing the PreDyn-ST model includes:

[0104] The PreDyn-ST model includes the SimCLR architecture for contrastive learning, a GCN spatial feature extraction module based on physical priors, and a Transformer temporal feature extraction module.

[0105] A physical prior-based GCN spatial feature extraction module is proposed, in which the construction of the adjacency matrix is ​​not only based on the similarity between nodes (sensor signals), but also combines the physical connection information of each component in the EHA system.

[0106] The contrastive learning of the SimCLR model represents the loss of the entire sample by averaging the losses of all positive sample pairs in a batch.

[0107] The loss for a positive pair is defined as:

[0108]

[0109] Among them, sim(z i ,z j ) is z i and z j The cosine similarity between is defined as τ is a temperature parameter; 1 [k≠i] is an indicator function that takes the value 1 when k≠i and 0 otherwise; the denominator is summed over all 2N possible pairs in the batch, excluding positive sample pairs.

[0110] Among them, sim(z i ,z j ) is z i and z j The cosine similarity between is defined as τ is a temperature parameter; 1 [k≠i] is an indicator function that takes the value 1 when k≠i and 0 otherwise; the denominator is summed over all 2N possible pairs in the batch, excluding positive sample pairs.

[0111] The total loss is calculated by averaging the losses of all positive pairs in the batch:

[0112]

[0113] By minimizing this loss, the model can learn feature representations that are insensitive to the data augmentation process. This representation remains consistent under different data augmentation transformations (such as rotation, scaling, adding noise, etc.), thereby improving the robustness and generalization ability of the model and being able to more stably cope with diverse input data, such as Figure 3 shown.

[0114] The GCN spatial feature extraction module based on physical prior is mainly used to capture the spatial structural characteristics of the data, and adjust the influence of spatial characteristics in degradation prediction through the dynamic weight method, so as to improve the analysis accuracy of degradation data.

[0115] The GCN spatial feature extraction module based on physical priors includes: a graph G = (V, ξ, A) formed by EHA data: V is the number of nodes, i.e., sensors; ξ is the edge set between each pair of connected nodes; is the adjacency matrix showing the spatial relationship between nodes; L = I n -D -1 / 2 AD -1 / 2 is the symmetric normalized graph Laplacian matrix, and the diagonal matrix D is calculated based on the adjacency matrix, that is, is the eigenvalue of L.

[0116] (1) The convolution of vertex embedding and filter f is expressed as:

[0117] h=x* g f=U((U T x)⊙(U T f))

[0118] in, is the eigenvector matrix of the Laplace matrix, and ⊙ represents the Hadamard product.

[0119] (2) Diagonal matrix g θ , g θ For U T The diagonal matrix representation of f is:

[0120] Among them, diag is the operation of converting a vector into a diagonal matrix, θ i is each characteristic component after being projected into the frequency domain.

[0121] (3) The initial graph convolution formula can be obtained:

[0122] h=x* g f=Ug θ U T x.

[0123] (4) Use Chebyshev polynomials to simplify graph convolution calculations:

[0124]

[0125] in, and are the rescaled eigenvalues ​​and Laplacian matrices, respectively. is the Chebyshev coefficient, is a Chebyshev polynomial of order k, is an explicitly defined convolution kernel.

[0126] GCN is mainly used to capture the spatial structure characteristics of data and adjust the influence of spatial characteristics in degradation prediction through dynamic weight method, so as to improve the analysis accuracy of degradation data. Figure 4 shown.

[0127] The Transformer temporal feature extraction module includes: (1) self-attention calculation mechanism, (2) multi-head attention function, and (3) formula for preserving the position information of the sequence.

[0128] Transformer is used to focus on the temporal characteristics of the data. In this process, multi-head attention and position encoding mechanisms are used simultaneously, and the dynamic weight method is used to adjust the impact of the degradation time characteristics in degradation prediction.

[0129] (1) Self-attention calculation mechanism:

[0130]

[0131] Among them, Q, K, and V are query, key, and value matrices respectively. The similarity between each element in the sequence and other elements is calculated to determine its attention weight.

[0132] (2) The mathematical representation of the multi-head attention function is:

[0133] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O

[0134] Among them, the calculation method of each attention head is:

[0135]

[0136] (3) The formula for retaining the position information of the sequence is:

[0137]

[0138] Among them, pos is the position index, i is the dimension index, and d model is the dimension of the model.

[0139] Transformer is used to focus on the temporal characteristics of the data. In this process, multi-head attention and position encoding mechanisms are used simultaneously, and the dynamic weight method is used to adjust the impact of degradation time characteristics in degradation prediction, such as Figure 5 shown.

[0140] The constructed PreDyn-ST framework process is shown below: Figure 6 shown.

[0141] The constructed PreDyn-ST model is used to predict the actual RUL value at the failure point of four C-MAPSS test sets and the comparison results between the predicted RUL value show that the proposed model performs well on the four C-MAPSS test data sets and has good prediction effects, such as Fig.10 shown.

[0142] The constructed PreDyn-ST model has good fitting ability and can accurately predict the changes in the efficiency of the EHA system within a day and in the long term, such as Fig. 9 shown.

[0143] In this embodiment, the training, verification and testing of the model include:

[0144] On the constructed PreDyn-ST model, for the pre-training model, set n_length to 50, n_h id to 512, projection_d im to 64, temperature to 0.5, num_levels to 2, lr to 1.5e-4, Epochs to 200, and batch_size to 100.

[0145] The feature points processed by contrastive learning have a wider distribution range and a larger distance between feature points, which is conducive to distinguishing between feature points. Figure 8 shown.

[0146] After the contrast learning preprocessed data, the training and validation loss curves drop rapidly, and only a small number of Epochs are needed to converge quickly, such as Fig.11 shown.

[0147] For the overall CSI curve of the C-MAPSS training data set, Fig.12 shown.

[0148] After visualizing the CSI value, the proportion of the time module and the space module is adaptively adjusted, such as Fig.13 shown.

[0149] For the prediction task, set nfeat to 16, nh id to 128, GCN num_levels to 2, num_node to 14, Transformer num_levels to 2h idden_d im to 8, num_windows to 5, window_sample to 50, batch_size to 200, Epochs to 100, and lr to 1.5e-4;

[0150] In this example, different model variants are evaluated to verify the superiority of the combination. The experiments are divided into the following configurations: GCN model, Transformer model, pre-trained GCN model, pre-trained Transformer model, dynamic weight GCN-Transformer temporal feature extraction module, and the proposed pre-trained-dynamic weight GCN-Transformer temporal feature extraction module (PreDyn-ST). The results are shown in Table 1.

[0151] Table 1

[0152]

[0153] The results show that both pre-training and dynamic weights improve the performance of the model, and the proposed PreDyn-ST method has the best prediction effect.

[0154] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for predicting EHA performance degradation based on the PreDyn-ST model, characterized in that: It includes the following steps: S1. Build the EHA test bench; S2, signal acquisition: collect piston rod displacement sensor signals, force sensor signals, flow and pressure signals of two chambers of hydraulic servo cylinder, servo motor current signals and servo motor speed signals through data acquisition sensors; S3, establishing a component function signal set with degradation labels: obtaining system efficiency according to sensor signal data, taking the system efficiency as a degradation label, and establishing a component function signal set with degradation labels; S4. Construct PreDyn-ST model: Use the SimCLR architecture of contrastive learning, the GCN spatial feature extraction module based on physical priors, and the Transformer temporal feature extraction module to construct the PreDyn-ST model. The output of the PreDyn-ST model is used as the life prediction result of the system. It includes the following sub-steps: S41. Use the SimCLR architecture of contrastive learning to shorten the distance between positive samples in the sliding window between time steps in the multivariate time series of sensor signal data, increase the distance between negative samples, and generate enhanced data. S42, inputting the enhanced data into a feature extraction module to perform preliminary feature extraction; S43, the extracted features are respectively input into the Transformer temporal feature extraction module and the GCN spatial feature extraction module based on physical prior; S44. As the degradation stage changes, the PreDyn-ST model adaptively adjusts the weights of the Transformer temporal feature extraction module and the physical prior-based GCN spatial feature extraction module and outputs the system life prediction results; S5, PreDyn-ST model training: After dividing the component function signal set with degradation labels in S3 into a training set, a validation set, and a test set, the PreDyn-ST model is trained and validated, and the model parameters are adjusted to obtain a trained PreDyn-ST model; S6. Obtain the life prediction result predicted by the model: collect the multivariate time series of sensor signal data in real time and input it into the PreDyn-ST model trained in S5, output the predicted system life prediction result, judge whether the system life prediction result is lower than the set threshold, and determine the performance degradation state of EHA according to the judgment result.

2. The EHA performance degradation prediction method based on the PreDyn-ST model according to claim 1, characterized in that: In step S4, the SimCLR framework represents the loss of the entire sample by averaging the losses of all positive sample pairs; The loss for a positive pair is defined as: Among them, sim(z i ,z j ) is z i and z j The cosine similarity between is defined as τ is a temperature parameter; Z is expressed as the feature vector of the positive sample pair; i, j and k are all parameters; 1 [k≠i] is an indicator function that takes the value 1 when k≠i and 0 otherwise; the denominator is summed over all 2N possible pairs in the batch, excluding positive sample pairs; The total loss is calculated by averaging the losses of all positive pairs in the batch:

3. The EHA performance degradation prediction method based on the PreDyn-ST model according to claim 1, characterized in that: In step S4, in the graph G=(V,ξ,A) formed by the EHA data in the GCN spatial feature extraction module based on physical prior, V is the number of nodes, i.e., the number of sensors; ξ is the edge set between each pair of connected nodes; is the adjacency matrix showing the spatial relationship between nodes; L = I n -D -1 / 2 AD -1 / 2 is the symmetric normalized graph Laplacian matrix, and the diagonal matrix D is calculated based on the adjacency matrix, that is, is the eigenvalue of L; The convolution of vertex embedding and filter f is expressed as: h=x* g f=U((U T x)⊙(U T f)) in, is the eigenvector matrix of the Laplace matrix, ⊙ represents the Hadamard product; Diagonal matrix g θ , g θ For U T The diagonal matrix representation of f is: g θ =diag(U T f)=diag(θ0,θ1,…,θ n-1 ) Among them, diag is the operation of converting a vector into a diagonal matrix, θ i is each characteristic component after being projected into the frequency domain; The initial graph convolution formula can be obtained: h=x* g f=Ug θ U T x; Simplify graph convolution calculation using Chebyshev polynomials: in, and are the rescaled eigenvalues ​​and Laplacian matrices, respectively. is the Chebyshev coefficient, is a Chebyshev polynomial of order k, is an explicitly defined convolution kernel.

4. The EHA performance degradation prediction method based on the PreDyn-ST model according to claim 1, characterized in that: In step S4, in the Transformer temporal feature extraction module: The self-attention calculation mechanism is expressed as: Among them, Q, K, and V are query, key, and value matrices respectively. Then the similarity between each element in the sequence and other elements is calculated to determine its attention weight; The multi-head attention function is expressed as: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O head i For each attention head in the multi-head attention (i=1,2,3...k), W o is the output linear transformation matrix; Among them, the calculation method of each attention head is: head i =Attention(QW i Q ,KW i K ,VW i V ); The formula for preserving the position information of the sequence is: Among them, pos is the position index, i is the dimension index, and d model is the dimension of the model.

5. The EHA performance degradation prediction method based on the PreDyn-ST model according to claim 1, characterized in that: In step S1, the constructed EHA test bench includes an active part and a load part of the EHA drive, wherein the active part includes a servo motor, a gear pump, a servo hydraulic cylinder and a plurality of data acquisition sensors; The load part adopts a top cylinder structure, and the load part includes a motor pump, a relief valve and a plurality of one-way valves.

6. A system for the EHA performance degradation prediction method based on the PreDyn-ST model according to any one of claims 1 to 5, characterized in that: It includes data acquisition module, degradation label definition module, model construction module, data partitioning module and model verification and prediction module: Data acquisition module: collects piston rod displacement sensor signals, force sensor signals for displaying load size, flow and pressure signals of the two chambers of the hydraulic servo cylinder, servo motor current signals, and servo motor speed signals through sensors of the EHA test bench; Degradation label definition module: organizes the signals collected in the EHA test bench and outputs them into system efficiency as degradation labels; Model building module: build PreDyn-ST model; Data partitioning module: divides the data in the collected EHA dataset into training set, validation set and test set; Model verification and prediction module: adjust the model parameters, train and verify the extracted feature labels with the PreDyn-ST model, and perform time series prediction to obtain the performance indicator results of the model prediction.

Images