Fault diagnosis method for high-speed hydraulic dynamometer

Abstract

The invention discloses a high-speed hydraulic dynamometer fault diagnosis method, and the method comprises the steps: building a dual-channel signal parallel convolutional neural network structure, carrying out the deeper feature extraction of data from multiple dimensions and multiple scales, designing an LSSA-SVM classifier, and carrying out the fault diagnosis of a high-speed hydraulic dynamometer. A chaotic lens initialization method and adaptive parameters are adopted to improve the search precision of a penalty factor C and a kernel function g, falling into a local optimal position is avoided, key parameters of a classifier are set more accurately, and finally fault diagnosis is performed according to online collected data. According to the method, fault classification of the hydraulic dynamometer is quicker and more accurate, the feature classification effect is better, and the problems that an existing fault prediction method of the hydraulic dynamometer is low in diagnosis precision and deep feature extraction is difficult are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed hydraulic dynamometer testing, and more specifically, to a high-speed hydraulic dynamometer fault diagnosis method. Background Art

[0002] As a power loading device for ground testing of aircraft engines, the high-speed hydraulic dynamometer mainly uses an electro-hydraulic servo valve to accurately adjust the opening of the inlet and outlet valves, control the increase or decrease of the dynamometer water volume, adjust the width of the water vortex ring, and thus control the dynamometer rotor torque, and at the same time control the engine torque connected to it. Therefore, when the aircraft engine is running at high speed, if the high-speed hydraulic dynamometer system fails, it will have a direct impact on the dynamometer torque, causing the engine to run unsteadily or even out of control, which will not only reduce the efficiency of scientific research and production, but also may cause serious safety hazards. The most commonly used solution at present is to disassemble and inspect the high-speed hydraulic dynamometer system to troubleshoot the fault by stopping the ground test mission of the aircraft engine. Disassembly and inspection will waste a lot of time and manpower, and may introduce new faults, so this method is inefficient and has poor real-time performance. With the continuous improvement of the level of independent research and development of aircraft engines, the demand for high-speed hydraulic dynamometers is becoming more and more urgent.

[0003] The prior art uses model training methods to detect faults in a timely manner during fault detection. For example, CN202311337905.3, a method for modeling and detecting aircraft engine fault modes based on IWOA-optimized BP, establishes an engine fault mode model based on a BP neural network, and optimizes the BP neural network based on IWOA to obtain the optimal weights and thresholds for the BP neural network. The aircraft engine fault mode model of the optimized BP neural network is constructed with the optimal weights and thresholds. The fault data is detected for abnormal data through the optimized model to achieve engine fault mode detection. However, the hydraulic dynamometer is an energy damping system that converts the input mechanical energy into heat energy and dissipates it to the outside world through the vortex loss effect in the form of a stator and a rotor. The factors that affect the rotation state of the rotor are more complex and diverse. The existing model prediction method has low fault diagnosis accuracy and is difficult to extract deep features, making it difficult to directly apply it to the hydraulic dynamometer. Summary of the invention

[0004] The main technical problem to be solved by the present invention is to provide a high-speed hydraulic dynamometer fault diagnosis method in view of the shortcomings of the fault diagnosis model in the prior art, such as low fault detection accuracy and difficulty in deep feature extraction.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A high-speed hydraulic dynamometer fault diagnosis method, comprising the steps of:

[0007] S1. The pressure signal data of each node of the high-speed hydraulic dynamometer under different fault conditions is used as the original data, the 1D time series signal data set is enhanced by overlapping sample segmentation technology, and the 1D time series signal is normalized and reconstructed into a 2D image signal;

[0008] S2. Construct a convolutional neural network structure, train it through back propagation, complete feature extraction, and obtain the optimal parameters;

[0009] S3. Solve the data through the LSSA algorithm to obtain the penalty factor C and the kernel function g, and obtain the SVM classifier;

[0010] S31. Randomly generate a population with N individuals, first map to generate a chaotic solution, then generate a chaotic reverse solution from the chaotic solution, sort the fitness according to the elitist strategy, and select the first N individuals to form the initial solution population;

[0011] The chaotic solution generation formula is:

[0012] X′ j =Ub j +Lb j / 2+Ub j +Lb j / 2K+X j / K,j=1,2,…,d

[0013]

[0014] The chaotic reverse solution generation formula is:

[0015]

[0016] γ∈(0,1),γ≠0.5 and γ≠x0

[0017] x j =Ub j +Lb j -x j ,j=1,2,…,d

[0018] Among them, Ub j With Lb j represent the upper and lower limits in the decision space, K is the scaling factor, and X j ′ represents the individual after mapping;

[0019] S32. Classify the initial population according to the warning threshold R2, safety threshold ST and fitness sequence, and divide the population individuals into discoverers and joiners. The golden sine algorithm is introduced to update the discoverers, which is expressed as:

[0020]

[0021] Where t represents the current iteration number, X i is the position vector of the ith individual, X best represents the current global optimal position, C1∈[0,2π] is the step size control parameter, C2∈[0,π] is the direction control parameter, A1 and A2 are the constants The calculated split coefficient;

[0022] Adaptive optimization of joiner updates is expressed as:

[0023]

[0024] Where X p and X worst are the finder's optimal position and the global worst position respectively; A + =A T (AA T ) -1 , A represents a 1×d matrix in which each element is randomly assigned a value of 1 or -1;

[0025] S33. Introduce a nonlinear increasing adaptive weight based on gamma distribution, reconstruct the control parameters and random factors to obtain its adaptive adjustment coefficient η, the expression is as follows:

[0026]

[0027] Where Γ(λ, 1-μ) is the incomplete gamma coefficient; λ is the random search factor, which ranges from (0,1); μ represents the ratio of the current iteration number to the maximum iteration number;

[0028] S34. Randomly select 10% to 20% of the total population as scouts to perform position correction. The position update is expressed as:

[0029]

[0030] Among them, f i 、f g and f w are the fitness value of the current individual and the global best and worst fitness values, respectively; K is a random number between [-1,1]; β is the step size control parameter; and ε is a constant;

[0031] S35. Calculate the fitness function, expressed as:

[0032]

[0033] y i (x) = y i (wx i )≥1

[0034] Where w is the inertia weight, which is used to adjust the search range of the solution space. i (x) represents the predicted value, y i (x) represents the true value, M is the total number of iterations;

[0035] S36. Update the state of the penalty factor C and the kernel function g until the maximum number of iterations and fitness requirements are met, and the optimal penalty factor C and kernel function g are obtained and applied to the SVM classifier, which is expressed as:

[0036]

[0037] Where K(x,x i ) is the kernel function of the classification problem; largrange coefficient a>0, subscript i=1,2,…,l; x r and x s is any pair of support vectors among the classified features;

[0038] S4. Input the features extracted by the neural network into the SVM classifier for fault diagnosis.

[0039] Furthermore, the original data is segmented by regional samples, and the number of samples N obtained by segmentation is:

[0040]

[0041] Where [.] is rounded down, n is the total number of signal sampling points, l is the segmentation width, and s is the step size, which is smaller than the segmentation width.

[0042] Furthermore, the convolutional neural network structure processing steps include: first performing convolution calculations through a convolution layer, using an activation function for nonlinear transformation, then reducing the dimension of the data in a pooling layer, converting a two-dimensional vector into a one-dimensional vector through a Flatten layer, fusing the one-dimensional and two-dimensional features, and then using a fully connected layer to rearrange and splice the multi-dimensional feature vector into a one-dimensional feature vector, and performing weight offset processing on the eigenvalue s and inputting it into the activation function f(.), and finally applying a Dropout strategy to randomly delete some neurons according to the configuration ratio.

[0043] Furthermore, the one-dimensional convolution calculation in the convolution layer is expressed as:

[0044]

[0045] The two-dimensional convolution calculation is expressed as:

[0046]

[0047] Where k(i) represents the feature extracted by the i-th convolution kernel; w and b represent weights and biases; r xRepresents the input 1D time series signal; r xy Represents the input 2D image signal; m and n depend on r x With r xy The number of data z;

[0048] The activation function uses the Relu activation function, which is expressed as:

[0049] f[k(i)]=min{0,k(i)}.

[0050] Furthermore, the pooling layer adopts maximum pooling, which is expressed as:

[0051]

[0052] In the formula, t l(a，u) represents the u-th neuron in the a-th feature map in the l-th layer; v represents the width of the convolution kernel; j represents the j-th pooling kernel.

[0053] Furthermore, the transformation process of the Flatten layer is expressed as:

[0054]

[0055]

[0056] Furthermore, the feature fusion is expressed as:

[0057]

[0058] where q i1 is a 1D feature vector; q i2 is the 2D feature vector after Flatten expansion; s is the feature vector The characteristic values ​​in .

[0059] Furthermore, the operation of the fully connected layer is expressed as:

[0060] s m =f(w m s m-1 +b m ).

[0061] Furthermore, the back propagation training adopts the Softmax function, which is expressed as:

[0062]

[0063] Where X represents the input feature vector; j represents the jth predicted category; K represents the total number of categories; β j Represents the connection weight with the upper layer.

[0064] Furthermore, in the finder update process, the division coefficients A1 and A2 are calculated as follows:

[0065] When f(X best ) <f(X t+1 ), the value of the original A2 is assigned to a2, and then the value of the original A1 is assigned to the current A2.

[0066] Now A1 is iterated according to A1=a1τ+a2(1-τ);

[0067] Otherwise, assign the original value of A1 to a1, and then assign the original value of A2 to the current A1. The current A2 is calculated according to A2=

[0068] a1(1-τ)+a2τ iteration;

[0069] If A1 and A2 have the same value, a1 and a2 will be assigned random values.

[0070] Compared with the prior art, the beneficial effects are:

[0071] The present invention establishes a dual-channel signal parallel convolutional neural network structure, which can perform deeper feature extraction on data from multiple dimensions and scales to cope with complex system failure conditions of high-speed hydraulic dynamometers; the present invention designs an LSSA-SVM classifier, improves and optimizes the algorithm steps, establishes a new mathematical model, and sets the key parameters of the classifier more accurately. The penalty factor C and kernel function g are more accurate and more suitable for hydraulic dynamometer fault classification, thereby making the feature classification effect better, making the hydraulic dynamometer fault classification more rapid and accurate, thereby making the feature classification effect better.

[0072] The fault diagnosis method of the present invention has practical significance. The method can be deployed in the control system of the aircraft engine test bench to monitor the hydraulic dynamometer fault in the aircraft engine test bench in real time, which can effectively prevent accidents and improve the safety of aircraft engines, hydraulic dynamometers and other related equipment. It can assist technicians to find problems at the early stage of failure, reduce maintenance cycles, and reduce maintenance costs; ensure that the aircraft engine test mission maintains efficient and stable operation, and improve scientific research and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a flow chart of the fault diagnosis method described in Example 1.

[0074] Figure 2 This is a structural diagram of the fault diagnosis model described in Example 1.

[0075] Figure 3 This is a flow chart of the chaotic lens initialization algorithm described in Example 1.

[0076] Figure 4This is a flow chart of the simulation experiment described in Example 2.

[0077] Figure 5 This is a schematic diagram of a simulated fault in Example 2;

[0078] a is the simulation of oil contamination by changing the gas content, b is the leakage of the actuator, c is the blockage of the reversing valve, and d is the leakage of the hydraulic pump;

[0079] Figure 6 This is a graph showing changes in accuracy and loss rate in Example 2;

[0080] Figure 7 This is the t-SNE clustering effect diagram in Example 2. DETAILED DESCRIPTION

[0081] The present invention will be further explained and illustrated below in conjunction with the embodiments, but the specific embodiments do not limit the present invention in any form.

[0082] Example 1

[0083] This embodiment provides a high-speed hydraulic dynamometer fault diagnosis method, such as Figures 1-2 , the steps include:

[0084] S1. According to the control principle of hydraulic dynamometer system, a hydraulic dynamometer system model is constructed based on AMESim. By collecting data of pressure signals of the hydraulic dynamometer system of the aircraft engine test bench, the pressure signals of relevant nodes are taken, and the key data of the corresponding node positions of the dynamometer system in the actual aircraft engine test bench are collected to verify the reliability of the model.

[0085] S2. By changing the corresponding component parameters or adding components to simulate different fault states, the pressure signal of each node is taken as the original data, and the 1D time series signal is enhanced by regional sample segmentation technology. The number of samples N obtained by segmentation is:

[0086]

[0087] Where [.] is rounded down, n is the total number of signal sampling points, l is the segmentation width, s is the step length, and the step length is smaller than the segmentation width. This can avoid the problem of feature loss in the process of segmenting samples and can retain the correlation of adjacent time samples of 1D time series signals to the greatest extent.

[0088] S3. Normalize the 1D time series signal to the range of [0,1] to reduce the redundant data generated by the reconstruction of the 2D image signal. By converting the 1D value change into the grayscale brightness between 0 and 1, and matching the data points to the gray scale color blocks, and then stacking the obtained 2D image signal, 1024 data points can be reconstructed and stacked into a 32×32 2D image signal. The 2D image signal represents higher-dimensional feature information, which can improve the reliability of the data.

[0089] S4. Construct a convolutional neural network structure, use the training set to perform back propagation training, complete feature extraction, and obtain optimal parameters; the back propagation (BP) training process uses the Softmax function, which is expressed as:

[0090]

[0091] Where X represents the input feature vector; j represents the jth predicted category; K represents the total number of categories; β j Represents the connection weight with the upper layer.

[0092] S5. The penalty factor C and the kernel function g are obtained by solving the test set through the LSSA algorithm, and the SVM classifier is obtained;

[0093] S51. Figure 3 , randomly generate a population of N individuals, first map to generate a chaotic solution, then generate a chaotic reverse solution from the chaotic solution, sort the fitness according to the elitist strategy, and select the first N individuals to form the initial solution population;

[0094] The chaotic solution generation formula is:

[0095] X j j=Ub j +Lb j / 2+Ub j +Lb j / 2K+X j / K,j=1,2,…,d

[0096]

[0097] The chaotic reverse solution generation formula is:

[0098]

[0099] γ∈(0,1),γ≠0.5 and γ≠x0

[0100] x j =Ub j +Lb j -x j ,j=1,2,…,d

[0101] Among them, Ub j With Lb j represent the upper and lower limits in the decision space, K is the scaling factor, and X j is the individual after mapping, γ is a random number between 0 and 1, and is not equal to 0.5 nor the initial value;

[0102] S52. Classify the initial population according to the warning threshold R2, safety threshold ST and fitness sequence, and divide the population individuals into discoverers and joiners. The golden sine algorithm is introduced for the update of discoverers, which is expressed as:

[0103]

[0104] Where T represents the current iteration number, X i is the position vector of the ith individual, X best represents the current global optimal position, C1∈[0,2π] is the step size control parameter, C2∈[0,π] is the direction control parameter, A1 and A2 are the constants The calculated partition coefficient, Q×L, represents a matrix whose elements are all random numbers Q that satisfy the normal distribution;

[0105] Assume that the initial values ​​of a1 and a2 are -π and π, when f(X best ) <f(X t+1 ), assign the original value of A2 to a2, then assign the original value of A1 to the current A2, and the current A1 iterates according to A1=a1τ+a2(1-τ); otherwise, assign the original value of A1 to a1, then assign the original value of A2 to the current A1, and the current A2 iterates according to A2=a1(1-τ)+a2τ. In addition, if A1 and A2 have the same value, a1 and a2 are randomly assigned.

[0106] Adaptive optimization of joiner updates is expressed as:

[0107]

[0108] Where X p and X worst are the finder's optimal position and the global worst position respectively; A + =A T (AA T ) -1 , A represents a 1×d matrix in which each element is randomly assigned a value of 1 or -1, and η is the adaptive adjustment coefficient;

[0109] S53. Introduce a nonlinear increasing adaptive weight based on gamma distribution, reconstruct its control parameters and random factors to obtain its adaptive adjustment coefficient η, the expression is as follows:

[0110]

[0111] Where Γ(λ, 1-μ) is the incomplete gamma coefficient; λ is the random search factor, which ranges from (0,1); μ represents the ratio of the current iteration number to the maximum iteration number. This method can effectively adjust the speed at which the joiner converges to the discoverer, and avoid the sudden drop in population diversity at the end of the iteration, which causes the algorithm to fall into a local optimum.

[0112] S54. Randomly select 10% to 20% of the total population as scouts to perform position correction to avoid converging to the wrong position and being unable to escape. The position update is expressed as:

[0113]

[0114] Among them, f i 、f g and f w are the fitness value of the current individual and the global best and worst fitness values, respectively; K is a random number between [-1,1]; β is the step size control parameter; and ε is a constant;

[0115] S55. Calculate the fitness function, expressed as:

[0116]

[0117] y i (x) = y i (wx i )≥1

[0118] Where w is the inertia weight, which is used to adjust the search range of the solution space. i (x) represents the predicted value, y i (x) represents the true value, M is the total number of iterations;

[0119] S56. Update the status of the penalty factor C and the kernel function g until the maximum number of iterations and the fitness requirements are met, and the optimal penalty factor C and the kernel function g are obtained, which are applied to the SVM classifier and are expressed as:

[0120]

[0121] Where K(x,x i ) is the kernel function of the classification problem; largrange coefficient a>0, subscript i=1,2,…,l; x r and x s is any pair of support vectors among the classified features;

[0122] S4. The features of the training set and test set extracted by the neural network are input into the SVM classifier and deployed in the actual aircraft engine test bench for online diagnosis. By analyzing and processing the collected signal parameters of a certain length of the dynamometer, the instability fault conditions of the components related to the types trained by the neural network training set can be directly determined.

[0123] Example 2

[0124] This embodiment provides the processing steps of the convolutional neural network structure in step S4 based on embodiment 1, including:

[0125] S41. Establish a dual-channel signal parallel convolutional neural network structure, one channel is a 1DCNN structure to process one-dimensional time series signals, and the other channel is a 2DCNN structure to process two-dimensional image signals, so as to achieve multi-scale and deeper feature extraction. In the 1DCNN structure, the one-dimensional convolution is implemented by the conv1d function, the maximum value pooling is used, and the "Same" padding is used to ensure that the size of the feature matrix after convolution remains unchanged. In the 2DCNN structure, the two-dimensional convolution is implemented by the conv2d function, and the pooling and padding methods are consistent with the 1DCNN structure.

[0126] S42. In the convolution layer, the input information is convolved, and each feature vector output by the convolution kernel corresponds to a class of features. The 1D convolution calculation process is expressed as:

[0127]

[0128] The 2D convolution calculation process is expressed as:

[0129]

[0130] Where k(i) represents the feature extracted by the i-th convolution kernel; w and b represent weights and biases; r x Represents the input 1D time series signal; r xy Represents the input 2D image signal; m and n depend on r x With r xy The number of data z.

[0131] S43. Then, a nonlinear transformation is performed through an activation function to transform the linearly inseparable multidimensional features into new features with strong linear separability. The activation function adopts the Relu activation function, which is expressed as:

[0132] f[k(i)]=min{0,k(i)}

[0133] S44. Downsample in the pooling layer to achieve data dimension reduction, and filter to extract deeper feature information. The pooling layer uses maximum pooling, which is expressed as:

[0134]

[0135] In the formula, t l(a，u) represents the u-th neuron in the a-th feature map in the l-th layer; v represents the width of the convolution kernel; j represents the j-th pooling kernel.

[0136] S45. The two-dimensional vector is converted into a one-dimensional vector through the Flatten layer. The expansion process is expressed as:

[0137]

[0138] The 1D and 2D features are fused, and the feature fusion is expressed as:

[0139]

[0140] where q i1 is a 1D feature vector; q i2 is the 2D feature vector after Flatten expansion; s is the feature vector The characteristic values ​​in .

[0141] S46. The feature vector Input to the fully connected layer, rearrange the multi-dimensional feature vector in order to splice it into a one-dimensional feature vector, and perform weight shift processing on the feature value s and input it to the activation function f(.) to obtain a better classification effect. The operation is expressed as:

[0142] s m =f(w m s m-1 +b m )

[0143] S47. Apply the Dropout strategy to randomly delete some neurons according to the configuration ratio to enhance the robustness and generalization of the model.

[0144] Example 3

[0145] This embodiment provides a simulation experiment example of the method described in Embodiment 1, and the parameters of the convolutional neural network structure constructed therein are set as shown in Table 1 below:

[0146] Table 1

[0147] Network Layer size Number of channels Fiter size Step Length Input layer / 1D 1×1024 3 The first convolutional layer 1×1024 24 1×3 1 First pooling layer 1×512 24 1×2 1 The second convolutional layer 1×512 48 1×3 1 Second pooling layer 1×256 48 1×2 1 Flatten layer 1×4096 3 Input layer / 2D 32×32 3 The first convolutional layer 32×32 24 3×3 1 First pooling layer 16×16 24 2×2 1 The second convolutional layer 16×16 48 3×3 1 Second pooling layer 8×8 48 2×2 1 Flatten layer 1×1024 3 Feature fusion layer 1×5120 3 Fully connected layer 1×512 3 SSA-SVM 1×5 1

[0148] In addition, the initial learning rate is 0.001, the number of feature extractor iterations is 1000, the batch size is 16, and the loss function uses cross entropy and Adam optimizer; Dropout is set to 0.5 in the training phase and 1.0 in the test phase. R2=0.8, ST is a random number between 0 and 1, the discoverers account for 20% of the total proportion, and the rest are joiners. The LSSA population size is set to 20 for 200 iterations, K=1000, R2=0.8, ST is a random number between 0 and 1, the discoverers account for 20% of the total proportion, and the rest are joiners. The simulation process is as follows Figure 4 According to the hydraulic dynamometer water inlet and outlet valve drive control system, the AMESim simulation model is established. Some models are shown in Figure 5 As shown, several typical faults such as hydraulic pump leakage, cylinder leakage, overflow valve blockage, etc. are simulated, and one-dimensional time domain signals are obtained by simulation. The simulation outputs the changes of one-dimensional time series signals of each node under different states and enhances them to obtain two-dimensional image signals. The two data sets are input into the neural network model for adaptive feature extraction, and the detected fault category is output.

[0149] Gaussian radial basis function (RBF) is used as the kernel function and input into LSSA-SVM for classification learning. The changes in accuracy and loss rate during training are shown in the following figure: Figure 6 As shown, the algorithm model proposed in the present invention has begun to converge after about 20 iterations, and the fault diagnosis accuracy of the model's test set is stabilized at 99.37%, with higher stability and higher classification accuracy.

[0150] In order to distinguish the effects of sample features before and after feature extraction and feature fusion at each layer, the t-SNE algorithm is used to perform visual analysis on the feature extraction and fusion stages, as shown in the figure. Figure 7 As shown in the T-SNE clustering effect diagram, it can be seen that the fault feature points are clustered separately, and the boundaries are clear and separable. This shows that the degree of mixing between different features after the neural network model proposed by the present invention is smaller, proving that the model is easier to converge and can be better classified.

[0151] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A high-speed hydraulic dynamometer fault diagnosis method, characterized in that the steps include: S1. The pressure signal data of each node of the high-speed hydraulic dynamometer under different fault conditions is used as the original data, the 1D time series signal data set is enhanced by overlapping sample segmentation technology, and the 1D time series signal is normalized and reconstructed into a 2D image signal; S2. Construct a convolutional neural network structure, train it through back propagation, complete feature extraction, and obtain the optimal parameters; S3. Solve the data through the LSSA algorithm to obtain the penalty factor C and the kernel function g, and obtain the SVM classifier; S31. Randomly generate a population with N individuals, first map to generate a chaotic solution, then generate a chaotic reverse solution from the chaotic solution, sort the fitness according to the elitist strategy, and select the first N individuals to form the initial solution population; The chaotic solution generation formula is: X j ′=Ub j +Lb j / 2+Ub j +Lb j / 2K+X j / K,j=1,2,…,d The chaotic reverse solution generation formula is: γ∈(0,1),γ≠0.5 and γ≠x0 x j =Ub j +Lb j -x j ,j=1,2,Ω,d Among them, Ub j With Lb j represent the upper and lower limits in the decision space, K is the scaling factor, and X j ′ represents the individual after mapping; S32. Classify the initial population according to the warning threshold R2, safety threshold ST and fitness sequence, and divide the population individuals into discoverers and joiners. The golden sine algorithm is introduced to update the discoverers, which is expressed as: Where t represents the current iteration number, X i is the position vector of the ith individual, X best represents the current global optimal position, C1∈[0,2π] is the step size control parameter, C2∈[0,π] is the direction control parameter, A1 and A2 are the constants The calculated split coefficient; Adaptive optimization of joiner updates is expressed as: Where X p and X worst are the finder's optimal position and the global worst position respectively; A + =A T (AA T ) -1 , A represents a 1×d matrix in which each element is randomly assigned a value of 1 or -1; S33. Introduce a nonlinear increasing adaptive weight based on gamma distribution, reconstruct the control parameters and random factors to obtain its adaptive adjustment coefficient η, the expression is as follows: Where Γ(λ, 1-μ) is the incomplete gamma coefficient; λ is the random search factor, which ranges from (0,1); μ represents the ratio of the current iteration number to the maximum iteration number; S34. Randomly select 10% to 20% of the total population as scouts to perform position correction. The position update is expressed as: Among them, f i 、f g and f w are the fitness value of the current individual and the global best and worst fitness values, respectively; K is a random number between [-1,1]; β is the step size control parameter; and ε is a constant; S35. Calculate the fitness function, expressed as: y i (x)=y i (wx i )≥1 Where w is the inertia weight, which is used to adjust the search range of the solution space. i (x) represents the predicted value, y i (x) represents the true value, M is the total number of iterations; S36. Update the state of the penalty factor C and the kernel function g until the maximum number of iterations and fitness requirements are met, and the optimal penalty factor C and kernel function g are obtained and applied to the SVM classifier, which is expressed as: Where K(x,x i ) is the kernel function of the classification problem; largrange coefficient a>0, subscript i=1,2,…,l; x r and x s is any pair of support vectors among the classified features; S4. Input the features extracted by the neural network into the SVM classifier for fault diagnosis.

2. The high-speed hydraulic dynamometer fault diagnosis method according to claim 1, characterized in that: The original data is segmented by regional samples, and the number of samples N obtained by segmentation is: Where [.] is rounded down, n is the total number of signal sampling points, l is the segmentation width, and s is the step size, which is smaller than the segmentation width.

3. The high-speed hydraulic dynamometer fault diagnosis method according to claim 1, characterized in that: The convolutional neural network structure processing steps include: first performing convolution calculation through a convolution layer, using an activation function for nonlinear transformation, then reducing the dimension of the data in a pooling layer, converting a two-dimensional vector into a one-dimensional vector through a Flatten layer, fusing the one-dimensional and two-dimensional features, and then using a fully connected layer to rearrange and splice the multi-dimensional feature vector into a one-dimensional feature vector, and performing weight offset processing on the eigenvalue s and inputting it into the activation function f(.), and finally applying a Dropout strategy to randomly delete some neurons according to the configuration ratio.

4. The high-speed hydraulic dynamometer fault diagnosis method according to claim 3, characterized in that: The one-dimensional convolution calculation in the convolution layer is expressed as: The two-dimensional convolution calculation is expressed as: Where k(i) represents the feature extracted by the i-th convolution kernel; w and b represent weights and biases; r x Represents the input 1D time series signal; r xy Represents the input 2D image signal; m and n depend on r x With r xy The number of data z; The activation function uses the Relu activation function, which is expressed as: f[k(i)]=min{0,k(i)}.

5. The high-speed hydraulic dynamometer fault diagnosis method according to claim 3, characterized in that: The pooling layer adopts maximum pooling, which is expressed as: Where, t l(a，u) represents the u-th neuron in the a-th feature map in the l-th layer; v represents the width of the convolution kernel; j represents the j-th pooling kernel.

6. The high-speed hydraulic dynamometer fault diagnosis method according to claim 3, characterized in that: The transformation process of the Flatten layer is expressed as:

7. The high-speed hydraulic dynamometer fault diagnosis method according to claim 3, characterized in that: The feature fusion is expressed as: where q i1 is a 1D feature vector; q i2 is the 2D feature vector after Flatten expansion; s is the feature vector The characteristic values ​​in .

8. The high-speed hydraulic dynamometer fault diagnosis method according to claim 3, characterized in that: The operation of the fully connected layer is expressed as: s m =f(w m s m-1 +b m )。 9. The high-speed hydraulic dynamometer fault diagnosis method according to claim 1, characterized in that: The back propagation training adopts the Softmax function, which is expressed as: Where X represents the input feature vector; j represents the jth predicted category; K represents the total number of categories; β j Represents the connection weight with the upper layer.

10. The high-speed hydraulic dynamometer fault diagnosis method according to claim 1, characterized in that: In the discoverer update process, the division coefficients A1 and A2 are calculated as follows: When f(X best ) <f(X t+1 ), assign the original value of A2 to a2, then assign the original value of A1 to the current A2, and the current A1 is iterated according to A1=a1τ+a2(1-τ); Otherwise, assign the original value of A1 to a1, and then assign the original value of A2 to the current A1. The current A2 is A2=a1(1-τ)+a2τ iteration; If A1 and A2 have the same value, a1 and a2 will be assigned random values.

Images