A reciprocating compressor air valve fault diagnosis method

Abstract

The application discloses a kind of reciprocating compressor air valve fault diagnosis methods, it is related to reciprocating compressor fault diagnosis technical field.The method steps are:S1, collection reciprocating compressor air valve vibration acceleration signal;S2, based on the window length of Hankel matrix in dynamic modal decomposition is optimized in chaos evolution optimization algorithm;S3, air valve vibration acceleration signal is decomposed and reconstructed;S4, the two-dimensional multiscale attention entropy of reconstructed signal is calculated, and two-dimensional multiscale attention entropy feature matrix is constructed;S5, two-dimensional multiscale attention entropy feature matrix is input least square support vector machine optimized by particle swarm optimization algorithm, and the fault diagnosis of reciprocating compressor air valve is realized.The application solves the problem that parameter selection of traditional method has no universality, feature extraction is not comprehensive and diagnostic precision is insufficient, realizes the high reliability and high accuracy of reciprocating compressor air valve fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of reciprocating compressor fault diagnosis technology, and in particular to a method for diagnosing faults in reciprocating compressor valves. Background Technology

[0002] Reciprocating compressors, as core equipment in key industrial sectors such as petroleum, chemical, and natural gas, are characterized by their wide applicable pressure range and compact structure. However, they often operate under harsh conditions of high temperature, high pressure, and high load, and their numerous components and complex structure lead to frequent failures. As a critical component of reciprocating compressors, valve failures (such as valve plate breakage, spring failure, or valve gaps) can cause unplanned equipment shutdowns and even lead to safety accidents such as leaks and explosions, resulting in significant economic losses and the risk of personal injury.

[0003] The vibration signal of the valve in a reciprocating compressor exhibits nonlinear and non-stationary characteristics, with weak fault features easily affected by noise, posing a challenge to fault diagnosis. Existing diagnostic methods, such as wavelet analysis, suffer from the complexity of wavelet basis function selection; adaptive decomposition methods like EMD, LMD, and VMD are sensitive to parameter settings and the number of components, prone to overfitting or incomplete feature characterization. To address these issues—too many components leading to overfitting, and too few components failing to fully characterize the signal—this invention proposes Dynamic Mode Decomposition (DMD), which extracts the frequency, amplitude, and phase features of each mode, enabling analysis and prediction of dynamic systems. However, as an effective signal decomposition method, the effectiveness of DMD is affected by the length of the Hankel matrix window. L The impact is significant, but currently this parameter is mostly selected based on experience and lacks universality.

[0004] In feature extraction, entropy methods can effectively capture signal complexity and nonlinear features, but traditional one-dimensional entropy methods can only reflect single-dimensional information and are difficult to comprehensively characterize the joint time-frequency features of complex fault signals. In the fault identification stage, least squares support vector machines (LSSVM) have good small-sample learning capabilities, but their diagnostic accuracy is greatly affected by kernel parameters and penalty parameters; improper parameter selection can lead to a decline in model performance. Therefore, there is an urgent need for a fault diagnosis method for reciprocating compressor valves that balances parameter optimization, comprehensive feature extraction, and high-precision identification. Summary of the Invention

[0005] The purpose of this invention is to propose a fault diagnosis method for reciprocating compressor valves, in order to solve the problems of lack of universality in the selection of Hankel matrix window length, incomplete extraction of one-dimensional entropy features, and the influence of LSSVM model parameters on diagnostic accuracy in the prior art, so as to achieve accurate and efficient diagnosis of reciprocating compressor valve faults.

[0006] To achieve the above objectives, this invention proposes a method for diagnosing valve faults in a reciprocating compressor, the specific steps of which are as follows: Step S1: Collect vibration acceleration signals of the reciprocating compressor valves; Step S2: Introduce the Chaotic Evolutionary Optimization Algorithm (CEO) to optimize the size of the Hankel matrix window in Dynamic Mode Decomposition (DMD); Step S3: Determine the truncation rank of the optimized DMD by using singular value hard thresholding (SVHT), and decompose and reconstruct the valve vibration acceleration signal to obtain the reconstructed signal. Step S4: Calculate the two-dimensional multi-scale attention entropy MATE of the reconstructed signal. 2D Construct a two-dimensional multi-scale attention entropy feature matrix; Step S5: Input the two-dimensional multi-scale attention entropy feature matrix into the least squares support vector machine (LSSVM) optimized by the particle swarm optimization algorithm (PSO) to realize the fault diagnosis of the reciprocating compressor valve.

[0007] Preferably, the Dynamic Mode Decomposition (DMD) is optimized based on the Chaotic Evolutionary Optimization Algorithm (CEO), and the specific steps are as follows: Step S221: Initialize the CEO algorithm parameters, including population size, maximum number of iterations, and define the Hankel matrix window length. L Search scope [ L min , L max ];in, L min This is the minimum window length. L max This represents the maximum window length. Step S222: Generate chaotic variables using chaotic mapping; Step S223: Linearly map the chaotic variable from the interval (0,1) to [ L min , L max The candidate window length is obtained by rounding down the interval. L ; Step S224: Based on the vibration acceleration signal and the candidate window length L Constructing the Hankel matrix H ; Step S225: For the Hankel matrix... H Perform singular value decomposition, calculate DMD modes and eigenvalues, and complete DMD decomposition; Step S226: Use the root mean square error (RMSE) as the fitness function to evaluate the current candidate window length.L Performance; Step S227: Check if the termination condition is met. The termination condition is reaching the maximum number of iterations or the fitness function value meets the preset accuracy. If met, output the optimal window length. L and the corresponding DMD model; if not satisfied, update the chaotic variables and return to step S221; Preferably, in step S3, the specific process of decomposition and reconstruction is as follows: [The text abruptly shifts to a different topic] ...optimal window length... L Substituting the data into the DMD, the truncation rank is adaptively selected using SVHT to decompose the vibration acceleration signal, and the reconstructed signal is obtained after removing noise components.

[0008] Preferably, in step S4, the two-dimensional multi-scale attention entropy MATE of the reconstructed signal is calculated. 2D The specific steps are as follows: Step S41: Define the two-dimensional matrix form of the reconstructed signal, as shown in the following formula: ; in, To reconstruct the two-dimensional signal matrix, the first... i Line number j Column elements, N and M These represent the number of sampling points for the signal in the vertical and horizontal directions, respectively. Given a scale factor The signal matrix is ​​divided into The nth sub-block is calculated, and the average value of each sub-block is obtained. Two-dimensional coarse-grained matrix at scale The formula is as follows: ; ; in, For scale Next ( p,q ) coarse-grained units, p The row index for spatial location. q Index to the column representing the spatial location; Step S42, at scale Below, the coarsened feature matrix... An attention mechanism is introduced to calculate the attention score of the feature vector at each location, as shown in the following formula: ; in, For scale Next (p, q Attention scores for each spatial unit. It is a non-linear activation function. For learnable weight matrix, For bias terms; The attention weight is obtained by normalizing the attention score, as shown in the following formula: ; in, For scale Next (p, q The weights of ) spatial units; Step S43, at each scale Next, the normalized attention weight matrix is ​​treated as a probability distribution, and the Shannon entropy of the probability distribution is calculated as follows: ; ; ; in, For scale Attention entropy For the original two-dimensional data at scale After coarsening, the length along the vertical direction is... For the original two-dimensional data at scale The length along the horizontal direction after coarsening; Step S44: Employ an attention-weighted fusion strategy to sum the Shannon entropies at all scales to obtain the two-dimensional multi-scale attention entropy, as shown in the following formula: ; in, For scale Two-dimensional multi-scale attention entropy, For scale The fusion weight, S This represents the total number of scales.

[0009] Preferably, in step S5, the specific process of PSO optimizing LSSVM is as follows: Step S51: Initialize PSO parameters, including particle swarm size, maximum number of iterations, inertia weight, and acceleration coefficient. Randomly generate the initial position and velocity of particles in the solution space. The particle position corresponds to the kernel parameters and penalty parameters of LSSVM. Step S52: Using the fault diagnosis accuracy of LSSVM as the fitness function, calculate the fitness value of each particle. Step S53: Compare the particle's current fitness value with its own historical best fitness value. p best ,renew p best Comparison of all particles pbest With the group's historical best fitness value g best ,renew g best ; Step S54: Adjust the particle velocity and position according to the PSO velocity and position update formula, and apply boundary constraints to avoid exceeding the solution space; Step S55: Check whether the termination condition is met, wherein the termination condition is reaching the maximum number of iterations or g best If the corresponding fitness value meets the preset accuracy, output the optimal kernel parameters and penalty parameters, and construct the PSO-LSSVM model; otherwise, return to step S52.

[0010] Therefore, this invention proposes a method for diagnosing valve faults in reciprocating compressors, which has the following beneficial effects: (1) This invention optimizes the window length of the Hankel matrix of DMD through the CEO algorithm and realizes adaptive parameter optimization with RMSE as the fitness function, which overcomes the limitations of traditional empirical selection and improves the stability and universality of DMD decomposition. (2) The two-dimensional multi-scale attention entropy (MATE) proposed in this invention 2D This method extends one-dimensional time series into a two-dimensional matrix form. By combining multi-scale coarse-grained processing and attention mechanisms, it can simultaneously capture the complex time-domain and frequency-domain features of fault signals. This overcomes the limitation of traditional one-dimensional entropy methods, which cannot fully characterize the essence of nonlinear and non-stationary fault signals in a single dimension, and significantly improves the extraction capability and discriminative power of weak fault features.

[0011] (3) The present invention uses the particle swarm optimization algorithm (PSO) to adaptively optimize the kernel parameters and penalty parameters of the least squares support vector machine (LSSVM), which effectively solves the problem that improper LSSVM parameter settings affect diagnostic accuracy. The constructed PSO-LSSVM model has stronger generalization ability and classification performance.

[0012] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0013] Figure 1 This is a flowchart of a method for diagnosing valve faults in a reciprocating compressor according to the present invention; Figure 2 This is a schematic diagram of vibration signals of the reciprocating compressor valve under various states in this invention; wherein, (a) is a schematic diagram of vibration signals of the valve under normal state, (b) is a schematic diagram of vibration signals of the valve plate under broken state, (c) is a schematic diagram of vibration signals of the valve under notched state, and (d) is a schematic diagram of vibration signals of the valve under notched spring state. Figure 3 This is a flowchart of the DMD method optimized by the chaotic evolutionary optimization algorithm in this invention; Figure 4 The figure shows the CEO optimization window length iteration curve and the optimization results for different states of the corresponding air valve in the embodiments of the present invention; wherein, (a) is the optimization result of the air valve in normal state, (b) is the optimization result of the air valve plate in broken state, (c) is the optimization result of the air valve in notched state, and (d) is the optimization result of the air valve in insufficient spring state. Figure 5 The above are the effect diagrams of the DMD reconstruction signal after optimization in the embodiments of the present invention; wherein, (a) is the effect diagram of the reconstruction signal under the normal state of the air valve, (b) is the effect diagram of the reconstruction signal under the broken state of the air valve plate, (c) is the effect diagram of the reconstruction signal under the notched state of the air valve, and (d) is the effect diagram of the reconstruction signal under the missing spring state of the air valve. Figure 6 These are entropy curves for different entropy methods in embodiments of the present invention; wherein, (a) is the entropy curve for the CEO-DMD-MES entropy method, (b) is the entropy curve for the CEO-DMD-MATE entropy method, (c) is the entropy curve for the CEO-DMD-MATE entropy method, and (d) is the entropy curve for the CEO-DMD-MATE entropy method. 2D Entropy curve of the entropy method; Figure 7 The diagrams illustrate the feature extraction capabilities of different entropy methods for the optimized DMD reconstructed signal in embodiments of the present invention; wherein, (a) is a diagram illustrating the feature extraction capability of the MES entropy method for the optimized DMD reconstructed signal, (b) is a diagram illustrating the feature extraction capability of the MATE entropy method for the optimized DMD reconstructed signal, (c) is a diagram illustrating the feature extraction capability of the CMATE entropy method for the optimized DMD reconstructed signal, and (d) is a diagram illustrating the feature extraction capability of the MATE entropy method for the optimized DMD reconstructed signal. 2D A schematic diagram illustrating the feature extraction capability of the entropy method for the optimized DMD reconstructed signal; Figure 8 This is a flowchart of the PSO-optimized LSSVM algorithm in this invention; Figure 9 This is a schematic diagram comparing the LSSVM classification and diagnosis results of different models in this embodiment of the invention; where (a) is DMD-MATE. 2D - The classification and diagnosis results of the LSSVM model, (b) is for DMD-MATE 2D The classification and diagnostic results of the PSO-LSSVM model are shown in (c) and (d) respectively. The classification and diagnostic results of the CEO-DMD-CMATE-LSSVM model are shown in (e) respectively. 2D- The classification and diagnostic results of the LSSVM model, (f) is for CEO-DMD-MATE 2D -Classification and diagnostic results of the PSO-LSSVM model. Detailed Implementation

[0014] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0015] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0016] Example like Figure 1 As shown, the present invention provides a method for diagnosing valve faults in a reciprocating compressor, the specific steps of which are as follows: Step S1, as follows Figure 2 As shown, the vibration acceleration signal of the reciprocating compressor valve is collected; Step S2: Introduce the Chaotic Evolutionary Optimization Algorithm (CEO) to optimize the size of the Hankel matrix window in Dynamic Mode Decomposition (DMD); like Figure 3 As shown, the specific steps of optimizing Dynamic Mode Decomposition (DMD) based on the Chaotic Evolutionary Optimization Algorithm (CEO) are as follows: Step S221: Initialize the CEO algorithm parameters, including population size, maximum number of iterations, and define the Hankel matrix window length. L Search scope [ L min , L max ];in, L min This is the minimum window length. L max This represents the maximum window length. Step S222: Generate chaotic variables using chaotic mapping; Step S223: Linearly map the chaotic variable from the interval (0,1) to [ L min , L max The candidate window length is obtained by rounding down the interval. L ; Step S224: Based on the vibration acceleration signal and candidate window length LConstructing the Hankel matrix H ; Step S225: For the Hankel matrix... H Perform singular value decomposition, calculate DMD modes and eigenvalues, and complete DMD decomposition; Step S226: Use the root mean square error (RMSE) as the fitness function to evaluate the current candidate window length. L Performance; Step S227: Check if the termination condition is met. The termination condition is reaching the maximum number of iterations or the fitness function value meets the preset accuracy. If met, output the optimal window length. L And the corresponding DMD model; if not satisfied, update the chaotic variables and return to step S221.

[0017] like Figure 4 As shown in Table 1, the optimal Hankel matrix window length L was obtained through optimization.

[0018] Table 1 Optimized parameters L size

[0019] Step S3: Optimal window length L Substituting the data into the DMD, the truncated rank is adaptively selected using SVHT to decompose the vibration acceleration signal. After removing noise components, the reconstructed signal is obtained. The optimized DMD reconstructed signal is as follows: Figure 5 As shown.

[0020] Step S4: Calculate the two-dimensional multi-scale attention entropy MATE of the reconstructed signal. 2D Construct a two-dimensional multi-scale attention entropy feature matrix.

[0021] Vibration signals from four states of a reciprocating compressor were selected, and each data point was divided into 40 non-overlapping groups with a sample length of 6024. Analysis showed that the noise reduction and reconstruction of the vibration signals preserved the vibration characteristics and key information of the original signals. Therefore, the reconstructed vibration signal data was used as the research object to calculate the two-dimensional multi-scale attention entropy. Vibration signals from four states of the compressor—normal valve, broken valve plate, valve notch, and valve spring shortage—were selected. The specific steps are as follows: Step S41: Define the two-dimensional matrix form of the reconstructed signal, as shown in the following formula: ; in, To reconstruct the two-dimensional signal matrix, the first... i Line number j Column elements, N and M These represent the number of sampling points for the signal in the vertical and horizontal directions, respectively. Given a scale factor The signal matrix is ​​divided into The nth sub-block is calculated, and the average value of each sub-block is obtained. Two-dimensional coarse-grained matrix at scale The formula is as follows: ; ; in, For scale Next ( p,q ) coarse-grained units, p The row index for spatial location. q Index to the column representing the spatial location; Step S42, at scale Below, the coarsened feature matrix... An attention mechanism is introduced to calculate the attention score of the feature vector at each location, as shown in the following formula: ; in, For scale Next (p, q Attention scores for each spatial unit. It is a non-linear activation function. For learnable weight matrix, For bias terms; The attention weight is obtained by normalizing the attention score, as shown in the following formula: ; in, For scale Next (p, q The weights of ) spatial units; Step S43, at each scale Next, the normalized attention weight matrix is ​​treated as a probability distribution, and the Shannon entropy of the probability distribution is calculated as follows: ; ; ; in, For scale Attention entropy For the original two-dimensional data at scale After coarsening, the length along the vertical direction is... For the original two-dimensional data at scale The length along the horizontal direction after coarsening; Step S44: Employ an attention-weighted fusion strategy to sum the Shannon entropies at all scales to obtain the two-dimensional multi-scale attention entropy, as shown in the following formula: ; in, For scale Two-dimensional multi-scale attention entropy, For scale The fusion weight, S This represents the total number of scales.

[0022] To verify the two-dimensional multiscale attention entropy (MATE) 2D To assess the effectiveness of this method, this paper compares traditional multiscale entropy (MSE), multiscale attention entropy (MATE), and composite multiscale attention entropy (CMATE) with two-dimensional multiscale attention entropy (MATE). 2D The comparison results are as follows: Figure 6 As shown.

[0023] Depend on Figure 6 As shown in Table 2, MATE achieves good results while ensuring the integrity of the extracted feature information. 2D Although it takes longer than the other two methods, its ability to capture feature information is the most outstanding.

[0024] Table 2 Feature extraction time for different entropy methods

[0025] To verify the two-dimensional multi-scale attention entropy (MATE) 2D The feature extraction capability of the method is demonstrated by using t-SNE to reduce the dimensionality of features extracted by different entropy methods at 20 scales, followed by two-dimensional visualization. The clustering ability of points for each fault type represents the feature extraction capability of the method; the smaller the distance between points within the same region and the larger the distance between different points, the better the feature extraction capability. The feature extraction capability results are shown below. Figure 7 As shown.

[0026] Step S5: Select 100 groups for each state of the air valve, with 80 groups as the training set and 20 groups as the test set. Substitute the entropy value features of the reciprocating compressor air valve under different states into the least squares support vector machine (LSSVM) optimized by the particle swarm optimization algorithm (PSO) for classification and recognition, thereby realizing the fault diagnosis of the reciprocating compressor air valve; Figure 8 As shown, the specific process of PSO optimizing LSSVM is as follows: Step S51: Initialize PSO parameters, including particle swarm size, maximum number of iterations, inertia weight, and acceleration coefficient. Randomly generate the initial positions and velocities of particles in the solution space. The particle positions correspond to the kernel parameters and penalty parameters of LSSVM. Step S52: Using the fault diagnosis accuracy of LSSVM as the fitness function, calculate the fitness value of each particle. Step S53: Compare the particle's current fitness value with its own historical best fitness value. p best ,renew p best Comparison of all particles p best With the group's historical best fitness value g best ,renew g best ; Step S54: Adjust the particle velocity and position according to the PSO velocity and position update formula, and apply boundary constraints to avoid exceeding the solution space; Step S55: Check if the termination condition is met. The termination condition is reaching the maximum number of iterations or g best If the corresponding fitness value meets the preset accuracy, output the optimal kernel parameters and penalty parameters, and construct the PSO-LSSVM model; otherwise, return to step S52.

[0027] like Figure 9 As shown, the CEO-DMD-MATE proposed in this invention... 2D The PSO-LSSVM model significantly outperforms other comparative diagnostic models in LSSVM classification and diagnosis, fully verifying the superiority of the model of this invention.

[0028] One hundred sets of entropy feature matrix data were selected for each of the following conditions: normal valve state, broken valve plate, valve notch, and valve spring shortage. Eighty sets were randomly selected for each health state category, totaling 320 sets, as the training set. The remaining 20 sets for each health state, totaling 80 sets, were used as the test set to further validate the CEO-DMD-MATE protocol. 2D The superior performance of the PSO-LSSVM model was compared using various fault diagnosis models. The number of training and test samples was the same as above. Multiple tests were conducted, and the average value of the results was calculated. The comparison results are shown in Table 3.

[0029] Table 3 Comparison of Different Diagnostic Models

[0030] As shown in Table 3, the CEO-DMD-MATE proposed in this invention... 2D The -PSO-LSSVM diagnostic model has a significantly higher state recognition rate than other comparative diagnostic models, which fully verifies the effectiveness and accuracy of the diagnostic model proposed in this invention.

[0031] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0032] Therefore, this invention provides a method for diagnosing valve faults in reciprocating compressors, which solves the problems of non-universality in parameter selection, incomplete feature extraction, and insufficient diagnostic accuracy in traditional methods. It achieves high-precision and high-reliability diagnosis of valve faults in reciprocating compressors, which can effectively ensure industrial production safety and reduce economic losses.

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing valve faults in a reciprocating compressor, characterized in that, The specific steps are as follows: Step S1: Collect vibration acceleration signals of the reciprocating compressor valves; Step S2: Introduce the Chaotic Evolutionary Optimization Algorithm (CEO) to optimize the size of the Hankel matrix window in Dynamic Mode Decomposition (DMD); Step S3: Determine the truncation rank of the optimized DMD by using singular value hard thresholding (SVHT), and decompose and reconstruct the valve vibration acceleration signal to obtain the reconstructed signal. Step S4: Calculate the two-dimensional multi-scale attention entropy MATE of the reconstructed signal. 2D Construct a two-dimensional multi-scale attention entropy feature matrix; Step S5: Input the two-dimensional multi-scale attention entropy feature matrix into the least squares support vector machine (LSSVM) optimized by the particle swarm optimization algorithm (PSO) to realize the fault diagnosis of the reciprocating compressor valve.

2. The method for diagnosing valve faults in a reciprocating compressor according to claim 1, characterized in that, In step S2, the dynamic mode decomposition (DMD) is optimized based on the chaotic evolutionary optimization algorithm CEO. The specific steps are as follows: Step S221: Initialize the CEO algorithm parameters, including population size, maximum number of iterations, and define the Hankel matrix window length. L Search scope [ L min , L max ]; in, L min This is the minimum window length. L max This represents the maximum window length. Step S222: Generate chaotic variables using chaotic mapping; Step S223: Linearly map the chaotic variable from the interval (0,1) to [ L min , L max The candidate window length is obtained by rounding down the interval. L ; Step S224: Based on the vibration acceleration signal and the candidate window length L Constructing the Hankel matrix H ; Step S225: For the Hankel matrix... H Perform singular value decomposition, calculate DMD modes and eigenvalues, and complete DMD decomposition; Step S226: Use the root mean square error (RMSE) as the fitness function to evaluate the current candidate window length. L Performance; Step S227: Check if the termination condition is met. The termination condition is reaching the maximum number of iterations or the fitness function value meets the preset accuracy. If met, output the optimal window length. L And the corresponding DMD model; if not satisfied, update the chaotic variables and return to step S221.

3. The method for diagnosing valve faults in a reciprocating compressor according to claim 2, characterized in that, In step S3, the specific process of decomposition and reconstruction is as follows: The optimal window length is... L Substituting the data into the DMD, the truncation rank is adaptively selected using SVHT to decompose the vibration acceleration signal, and the reconstructed signal is obtained after removing noise components.

4. The method for diagnosing valve faults in a reciprocating compressor according to claim 3, characterized in that, In step S4, the two-dimensional multi-scale attention entropy MATE of the reconstructed signal is calculated. 2D The specific steps are as follows: Step S41: Define the two-dimensional matrix form of the reconstructed signal, as shown in the following formula: ； in, To reconstruct the two-dimensional signal matrix, the first... i Line number j Column elements, N and M These represent the number of sampling points for the signal in the vertical and horizontal directions, respectively. Given a scale factor The signal matrix is ​​divided into The nth sub-block is calculated, and the average value of each sub-block is obtained. Two-dimensional coarse-grained matrix at scale The formula is as follows: ； ； in, For scale Next ( p,q ) coarse-grained units, p The row index for spatial location. q Index to the column representing the spatial location; Step S42, at scale Below, the coarsened feature matrix... An attention mechanism is introduced to calculate the attention score of the feature vector at each location, as shown in the following formula: ； in, For scale Next (p, q Attention scores for each spatial unit. It is a non-linear activation function. For learnable weight matrix, For bias terms; The attention weight is obtained by normalizing the attention score, as shown in the following formula: ； in, For scale Next (p, q The weights of ) spatial units; Step S43, at each scale Next, the normalized attention weight matrix is ​​treated as a probability distribution, and the Shannon entropy of the probability distribution is calculated as follows: ； ； ； in, For scale Attention entropy For the original two-dimensional data at scale After coarsening, the length along the vertical direction is... For the original two-dimensional data at scale The length along the horizontal direction after coarsening; Step S44: Employ an attention-weighted fusion strategy to sum the Shannon entropies at all scales to obtain the two-dimensional multi-scale attention entropy, as shown in the following formula: ； in, For scale Two-dimensional multi-scale attention entropy, For scale The fusion weight, S This represents the total number of scales.

5. The method for diagnosing valve faults in a reciprocating compressor according to claim 4, characterized in that, In step S5, the specific process of PSO optimizing LSSVM is as follows: Step S51: Initialize PSO parameters, including particle swarm size, maximum number of iterations, inertia weight, and acceleration coefficient. Randomly generate the initial position and velocity of particles in the solution space. The particle position corresponds to the kernel parameters and penalty parameters of LSSVM. Step S52: Using the fault diagnosis accuracy of LSSVM as the fitness function, calculate the fitness value of each particle. Step S53: Compare the particle's current fitness value with its own historical best fitness value. p best ,renew p best Comparison of all particles p best With the group's historical best fitness value g best ,renew g best ; Step S54: Adjust the particle velocity and position according to the PSO velocity and position update formula, and apply boundary constraints to avoid exceeding the solution space; Step S55: Check whether the termination condition is met, wherein the termination condition is reaching the maximum number of iterations or g best If the corresponding fitness value meets the preset accuracy, output the optimal kernel parameters and penalty parameters, and construct the PSO-LSSVM model; otherwise, return to step S52.

Images