Signal processing method, sample set construction method, fault diagnosis model construction method and fault diagnosis method

Abstract

The invention discloses a signal processing method, a sample set construction method, a fault diagnosis model construction method and a fault diagnosis method, and the method comprises the steps: firstly, collecting an original monitoring signal of equipment through a set sensor; then, a penalty factor and a modal number of variational empirical mode decomposition can be adaptively optimized by adopting a simulated annealing mayfly naiad algorithm, and an original monitoring signal is decomposed according to the optimal penalty factor and modal number, so that a more accurate modal function is obtained. And finally, signal reconstruction can be carried out according to the modal function with higher accuracy, and the accuracy of the monitoring signal obtained through reconstruction is enhanced. And constructing a training sample set on the basis of the reconstructed monitoring signal, and performing diagnosis model training on the basis of the training sample set, thereby obtaining a fault diagnosis model capable of accurately identifying equipment faults, and quickly and accurately identifying anomalies and faults of the gearbox.

Description

Technical Field

[0001] The present invention relates to the technical field of feature extraction, and in particular to a signal processing, sample set construction, fault diagnosis model construction and a fault diagnosis method. Background Art

[0002] The gearbox is one of the key components of a wind turbine. Gearbox fault detection usually involves installing an accelerometer on the transmission chain, collecting vibration signals through the accelerometer, and extracting the time domain, frequency domain, and time-frequency domain features of the vibration signals. Finally, the gearbox is judged based on the features to determine whether there is a fault. However, there will be mechanical vibration noise in the collected vibration signal, which will have a significant impact on the accuracy of the feature extraction results.

[0003] The existing technology usually uses empirical mode decomposition, variational empirical mode decomposition or wavelet denoising to reduce the noise of vibration signals. Among them, variational empirical mode decomposition has stronger robustness to sampling and noise, and can effectively avoid modal aliasing, so it is widely used in signal noise reduction. When using variational empirical mode decomposition to reduce the noise of vibration signals, it is necessary to predetermine the values ​​of the two parameters, the mode number K and the penalty factor α. These parameters determine the accuracy of each component after decomposition. Therefore, how to obtain the optimal variational empirical mode decomposition parameters is particularly important for gearbox fault diagnosis. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a signal processing, sample set construction, fault diagnosis model construction and fault diagnosis method, which can improve the accuracy of monitoring signals. The specific technical solutions are as follows:

[0005] In a first aspect, a signal processing method is provided. In a first implementable manner of the first aspect, the method includes:

[0006] Obtaining original monitoring signals;

[0007] Using a simulated annealing ephemera algorithm to optimize the parameters of the variational empirical mode decomposition, and performing variational empirical mode decomposition on the original monitoring signal based on the optimized parameters;

[0008] The monitoring signal is reconstructed according to the decomposed modal function.

[0009] In combination with the first implementable manner of the first aspect, in a second implementable manner of the first aspect, a simulated annealing ephemera algorithm is used to optimize parameters of variational empirical mode decomposition, including:

[0010] The optimal parameters of the variational empirical mode decomposition are searched by using an ephemeral algorithm, and during the search process, the acceptance probability of the searched optimal parameters is determined by using a simulated annealing algorithm according to the fitness function value of the ephemeral algorithm.

[0011] In combination with the first implementable manner of the first aspect, in a third implementable manner of the first aspect, the local minimum envelope entropy is used as the fitness function of the simulated annealing ephemera algorithm.

[0012] In combination with the first implementable manner of the first aspect, in a fourth implementable manner of the first aspect, reconstructing the monitoring signal according to the decomposed modal function includes:

[0013] The monitoring signal is reconstructed based on the maximum correlation coefficient-kurtosis criterion.

[0014] In a second aspect, a sample set construction method is provided. In a first implementable manner of the second aspect, the method includes:

[0015] Using the monitoring signal processing method as described in any one of the first to fourth possible implementations of the first aspect, the acquired original signal is processed to obtain an enhanced monitoring signal;

[0016] A sample signal is intercepted from the monitoring signal through a set sliding window, and a training sample set is constructed according to the intercepted sample signal.

[0017] In combination with the first implementable manner of the second aspect, in a second implementable manner of the second aspect, constructing a training sample set according to the intercepted sample signal includes:

[0018] Perform compressed sensing processing on the intercepted sample signal.

[0019] In combination with the first implementable manner of the second aspect, in a third implementable manner of the second aspect, it further includes: normalizing all samples in the training sample set.

[0020] In a third aspect, a method for constructing a fault diagnosis model is provided. In a first implementable manner of the third aspect, the method includes:

[0021] Adopting the training sample set construction method as described in the first implementable manner of the second aspect, respectively constructing a source domain sample set and a target domain sample set;

[0022] Training the AIDBN model using the source domain sample set;

[0023] Based on the target domain sample set, the trained AIDBN model parameters are migrated to the target domain to obtain the AIDBN diagnostic model corresponding to the target domain.

[0024] In combination with the first implementable manner of the third aspect, in a second implementable manner of the third aspect, training the AIDBN model using the source domain sample set includes:

[0025] The simulated annealing ephemera algorithm is used to optimize the hyperparameters of the AIDBN model.

[0026] In a fourth aspect, a device fault diagnosis method is provided, comprising:

[0027] Adopting the device fault diagnosis model construction method as described in the first implementable manner of the third aspect to construct an AIDBN diagnosis model;

[0028] Acquire the current monitoring signal of the device, and extract features of the monitoring signal through the AIDBN diagnostic model to obtain the current feature label of the device;

[0029] The feature tag is matched with the fault feature database, and the device fault is determined based on the matching result.

[0030] Beneficial effects: By adopting the signal processing, sample set construction, fault diagnosis model construction and fault diagnosis method of the present invention, the parameters of the variational empirical mode decomposition can be optimized through the simulated annealing ephemera algorithm, and the optimal parameters can be obtained. The variational empirical mode decomposition of the original monitoring signal is performed with the optimal parameters, which can improve the accuracy of the modal function obtained by decomposition, and then enhance the accuracy of the reconstructed monitoring signal, providing a basis for the subsequent accurate identification of equipment faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the specific implementation of the present invention, the following will briefly introduce the drawings required for use in the specific implementation. In all the drawings, each element or part is not necessarily drawn according to the actual scale.

[0032] Figure 1 A flowchart of a signal processing method provided by an embodiment of the present invention;

[0033] Figure 2 A flow chart of a sample set construction method provided by an embodiment of the present invention;

[0034] Figure 3 A flowchart of a method for constructing a fault diagnosis model provided by an embodiment of the present invention;

[0035] Figure 4 A flowchart of a fault diagnosis method provided by an embodiment of the present invention;

[0036] Figure 5 The time domain diagram of the source domain vibration signal and the target domain vibration signal obtained by monitoring;

[0037] Figure 6 Optimization curve of variational empirical mode decomposition parameters

[0038] Figure 7This is a fault characteristic diagram identified by the fault diagnosis method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0040] The following table shows the fault codes corresponding to different fault types of wind turbine gearboxes:

[0041]

[0042] Table 1 Fault categories

[0043] like Figure 1 The signal processing method shown in FIG. 1 includes:

[0044] Step 1: Obtain the original monitoring signal;

[0045] Step 2: Optimizing the parameters of the variational empirical mode decomposition using a simulated annealing ephemera algorithm, and performing variational empirical mode decomposition on the original monitoring signal based on the optimized parameters;

[0046] Step 3: Reconstruct the monitoring signal based on the decomposed modal function.

[0047] Specifically, first, the original monitoring signal of the equipment can be collected by the set sensors. Then, the penalty factor and mode number of the variational empirical mode decomposition can be adaptively optimized by using the simulated annealing ephemera algorithm, and then the original monitoring signal can be decomposed according to the optimal penalty factor and mode number to obtain a more accurate modal function. Finally, the signal can be reconstructed according to the modal function with higher accuracy to enhance the quality of the reconstructed monitoring signal, so that the reconstructed monitoring signal is more distinguishable and the recognition of variable working conditions and multiple models is more stable, thereby providing a basis for the subsequent accurate identification of equipment failures.

[0048] In this embodiment, optionally, a simulated annealing mayfly algorithm is used to optimize the parameters of the variational empirical mode decomposition, including: using the mayfly algorithm to search for the optimal parameters of the variational empirical mode decomposition, and using the simulated annealing algorithm to determine the reception probability of the search result according to the fitness function value of the mayfly algorithm during the search process.

[0049] Specifically, the mayfly algorithm is used to search for the optimal parameters of the variational empirical mode decomposition. Specifically, first, the positions of male and female mayflies in the d-dimensional space are represented as x = (x1, x2, x3, L, x d ), the speed of the mayfly is expressed as v = (v1, v2, v3, L, vd ), the position of the male mayfly is adjusted according to its own experience and that of its neighbors, is the search space position of the i-th mayfly at the t-th iteration, and the t+1-th position is updated to

[0050]

[0051] Its speed is updated as follows:

[0052]

[0053] in, is the speed of mayfly i in the t-th iteration of the j-dimension, is the position of mayfly i in the tth iteration in the j dimension, a1 and a2 are the attraction coefficients of the mayfly's swimming behavior, pbest is the best position recorded, gbest is the global optimal position, β is the visibility coefficient, r p is the distance between the current position and pbest, r g The distance between the current position and gbest.

[0054] Then, in order to better obtain the optimal position, the mayfly must constantly update its speed. The specific calculation formula is as follows:

[0055]

[0056] Among them, d is the dance coefficient, r is the random coefficient, r∈[-1,1].

[0057] Afterwards, assuming is the position of the female mayfly i at the tth iteration, and its position is updated as follows:

[0058]

[0059] The speed of its female mayflies has been updated as follows:

[0060]

[0061] Among them, fl is the random walk coefficient, f(y i ) is the fitness function value corresponding to the position of female mayfly i, f(x i ) is the fitness function value corresponding to the position of male mayfly i.

[0062] Finally, the mating of male and female mayflies is random. The best male and female individuals are used for reproduction, and the second-best male and second-best female are used for reproduction. After crossing, two offspring are obtained. The offspring are:

[0063] offs1=L*male+(1-L)*female

[0064] offs2=L*female+(1-L)*male;

[0065] Wherein, L is a random number, L = [-1, 1], male is the male parent, and female is the female parent. The fitness values ​​corresponding to the two offspring obtained by crossover are recalculated to update the positions of male and female mayflies.

[0066] Although it has good convergence speed and accuracy, it is easy to find local optimal solutions in the solution search process. Therefore, when the Mayfly algorithm searches for the optimal position, the simulated annealing algorithm can be used to receive the poor position found by the Mayfly algorithm with a certain probability, thereby effectively avoiding the local optimal solution of the Mayfly algorithm.

[0067] Specifically, according to the fitness value of the current mayfly and the fitness value of the previous generation of mayflies, the increment △f=f(S2)-f(S1) is calculated, where f(S1) is the fitness value of the previous generation of mayflies;

[0068] If △f<0, the current mayfly position is accepted. Otherwise, the probability of the current mayfly position being accepted needs to be calculated. The specific calculation formula is as follows:

[0069]

[0070] Using the method based on fitness and acceptance probability, the temperature T is determined by the following formula:

[0071] T=-f / lnP r

[0072] Among them, f is the fitness value, P r is the initial acceptance probability.

[0073] In this embodiment, optionally, the local minimum envelope entropy is used as the fitness function of the simulated annealing ephemera algorithm, and the specific calculation formula is as follows:

[0074]

[0075] In the formula, E e is the envelope entropy, e j is the normalized form of a(j); a(j) is the envelope signal of the original monitoring signal after Hilbert transform, and N is the number of zero-mean signals.

[0076] It should be understood that after the bearing vibration signal of the wind turbine gearbox undergoes variational mode decomposition, the envelope entropy value of each component can reflect the sparse characteristics of the component. If the component contains more noise, it will mask the impact characteristics of the signal, then the sparsity of the component is weaker and the envelope entropy is larger; otherwise, the envelope entropy is smaller. The smallest envelope entropy value among all components after each decomposition is selected as the local minimum envelope entropy, and the component corresponding to the minimum entropy value contains the best feature information.

[0077] The simulated annealing ephemera algorithm, ephemera algorithm and particle swarm algorithm are used to optimize the parameters of variational empirical mode decomposition. The optimization curves of these three algorithms for variational empirical mode decomposition are as follows: Figure 6 As shown. Figure 6 It can be seen that compared with the Mayfly algorithm and the particle swarm algorithm, the simulated annealing Mayfly algorithm has a faster iteration speed, is easier to jump out of the local extreme value and reach convergence, and the optimization effect is better than the standard Mayfly algorithm and the particle swarm algorithm.

[0078] In this embodiment, optionally, reconstructing the monitoring signal according to the modal function obtained by decomposition includes:

[0079] The monitoring signal is reconstructed based on the maximum correlation coefficient-kurtosis criterion.

[0080] Specifically, the correlation coefficient is a statistic that measures the degree of correlation between variables. The larger the correlation coefficient between two signals, the stronger the correlation, and vice versa. Kurtosis is used to describe the peak degree of each component and is only related to the impact signal in each component. By combining the two methods, the maximum correlation coefficient-kurtosis can be established to reconstruct the signal more accurately.

[0081] like Figure 2 The flowchart of the sample set construction method shown in FIG. 1 includes:

[0082] Step S1, using the above-mentioned signal processing method to process the acquired original signal to obtain an enhanced monitoring signal;

[0083] Step S1: intercepting sample signals from the monitoring signal through a set sliding window, and constructing a training sample set according to the intercepted sample signals.

[0084] Specifically, first, the above signal processing method is used to process the collected equipment monitoring original signal to obtain an enhanced monitoring signal. Then, a set sliding window is used to intercept multiple signal segments from the monitoring signal as sample signals, and a training sample set with sufficient training samples is constructed based on all the intercepted sample signals to provide a basis for subsequent training of an effective equipment fault diagnosis model.

[0085] In this embodiment, optionally, a training sample set is constructed according to the intercepted sample signal, including:

[0086] Perform compressed sensing processing on the intercepted sample signal.

[0087] Specifically, after the sample signal is intercepted, in order to further enhance the signal quality, the sample signal can also be subjected to compressed sensing processing. When performing compressed sensing processing, the compression rate CR can be determined according to the dimension A of the set sample signal and the dimension B of the monitoring signal. The specific calculation formula is as follows:

[0088]

[0089] In this embodiment, optionally, all samples in the training sample set are normalized.

[0090] After constructing the training sample set, the mapminmax function can be used to normalize each sample in the training sample set to the range of [-1, 1] for subsequent training of the model.

[0091] like Figure 3 The flowchart of the fault diagnosis model construction method shown in FIG. 1 includes:

[0092] Step D1: Use the above sample set construction method to construct a source domain sample set and a target domain sample set respectively;

[0093] Step D2, training the AIDBN model using the source domain sample set;

[0094] Step D3: Based on the target domain sample set, the trained AIDBN model parameters are migrated to the target domain to obtain the AIDBN diagnostic model corresponding to the target domain.

[0095] Specifically, first, the above construction method can be used to collect the monitoring time domain signals of the source device and the target device respectively, such as Figure 5 As shown. The source domain sample set and the target domain sample set are constructed according to the corresponding monitoring signals. Then, the AIDBN model can be trained with the constructed source domain sample set. Finally, the knowledge in the AIDBN model trained based on the source domain sample set can be transferred to the AIDBN diagnostic model of the target device by using the transfer learning method through the target domain sample set. This effectively reduces the training time of the diagnostic model of the target device, improves the fault diagnosis efficiency of the target device, and reduces the deployment management and operation and maintenance costs of the target device.

[0096] Among them, the AIDBN model is an improved model based on the DBN model. The energy function of the restricted Boltzmann machine (RBM) of the DBN model is:

[0097]

[0098] The visible layer unit is v = {v1, v2, Lv i}, the hidden layer unit is h = {h1, h2, Lh j}, ω is the weight matrix, under the condition that the thresholds of the visible layer unit and the hidden layer unit are a and b respectively, j has v i ∈(0,1),h j ∈(0,1).

[0099] The activation probabilities corresponding to the hidden layer units and the visible layer are:

[0100]

[0101] By maximizing the log-likelihood function, we can train the RBM and obtain:

[0102]

[0103] The stochastic gradient ascent method is used to solve the above formula and obtain the parameter θ.

[0104] The update criteria for solving the parameters of the maximizing likelihood function based on the gradient ascent algorithm becomes the formula:

[0105] △ω ij =ε( <v i h j > p(h|v) - <v i h j > recon );

[0106] △b j =ε( <v j > p(h|v) - <v j > recon );

[0107] △a i =ε( <h i > p(h|v) - <h i > recon );

[0108] Among them, ε is the learning rate, <> recon is the expected value of the partial derivative function under the reconstructed model distribution.

[0109] In order to further enhance the robustness of the DBN diagnosis model, improve the training accuracy and reduce the model training time, the Gaussian distributed continuous restricted Boltzmann machine (CRBM) is used. At this time, the energy function formula is:

[0110]

[0111] Among them, σ i is the Gaussian noise standard deviation vector.

[0112] In order to further realize the sparse features of each hidden layer of CRBM, a sparse regularization term is introduced in the maximum likelihood function, which is adapted by adjusting the gap between the activation probability of the hidden layer nodes and the sparse coefficient. At this time, the definition of the unsupervised pre-training model with sparse regularization is added:

[0113] f=f C +λf S

[0114] Among them, f C is the maximum likelihood function of CRBM, λ is the regularization coefficient, and f s is the sparse regularization function.

[0115] The improved deep confidence model (IDBN) is expressed as follows:

[0116]

[0117] In this embodiment, optionally, the AIDBN model is trained using the source domain sample set, including: optimizing the hyperparameters of the AIDBN model using a simulated annealing ephemera algorithm.

[0118] Specifically, during the training process of the AIDBN model, the simulated annealing ephemera algorithm can be used to optimize the hyperparameters of the AIDBN model. The specific steps and principles are basically the same as the steps and principles of using the simulated annealing ephemera algorithm to optimize the optimal parameters of the variational empirical mode decomposition. The main difference is that the fitness function is the minimum error rate of the AIDBN model training. The specific calculation formula is as follows:

[0119]

[0120] Among them, y F is the number of samples with identification errors, y T To identify the correct number of samples.

[0121] like Figure 4 The flowchart of the fault diagnosis method shown in FIG. 1 includes:

[0122] Step C1, using the above fault diagnosis model construction method to build an AIDBN diagnosis model;

[0123] Step C2, obtaining the current monitoring signal of the device, and extracting features of the monitoring signal through the AIDBN diagnostic model to obtain the current feature label of the device;

[0124] Step C3: Match the feature tag with the fault feature database, and determine the equipment fault based on the matching result.

[0125] Specifically, first, the above fault diagnosis model construction method is used to build an AIDBN diagnosis model for diagnosing the target device. Then, the current original monitoring signal of the target device can be obtained, and the original monitoring signal can be processed using the above signal processing method to enhance the quality of the monitoring signal input to the AIDBN diagnosis model. The feature label of the current device can be obtained through the AIDBN diagnosis model, such as Figure 7 Finally, the obtained feature label is matched with the feature labels of various faults in the fault feature database to determine the current fault type of the target device.

[0126] In order to verify the technical effect of the fault diagnosis method provided in this embodiment, another 7 diagnostic methods are used to diagnose the wind turbine gearbox under different working conditions, and the diagnostic results are shown in the following table:

[0127] Table 2 Comparison of diagnostic results of different models

[0128] Model 1 is an AIDBN diagnostic model constructed using the construction method provided in this embodiment.

[0129] Model 2 is a diagnostic model built based on the standard VMD, CS and AIDBN models.

[0130] Model three is a diagnostic model built based on the optimized VMD and AIDBN models, which reconstructs the time domain signal without compressed sensing.

[0131] Model 4 is a diagnostic model built based on the optimized VMD and AIDBN models.

[0132] Model 5 is a diagnostic model built based on the optimized VMD, CS and standard DBN models;

[0133] Model 6 is a diagnostic model based on the optimized VMD, CS and multi-layer perceptron models;

[0134] Model 7 is a diagnostic model based on time domain indicators, CS and SVM, in which SVM uses RBF kernel function and its key parameters are optimized by SAMA to obtain the best combination;

[0135] Model eight is a diagnostic model constructed based on time domain indicators, CS and BPNN models.

[0136] It can be seen from the above table that compared with other models, the AIDBN diagnostic model constructed in this embodiment has higher accuracy for various fault conditions of wind turbine gearboxes, and can accurately identify different fault types, fault degrees, fault locations, etc. of wind farm gearboxes, and its stability and accuracy are better than traditional identification methods.

[0137] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A signal processing method, characterized in that: include: Obtaining original monitoring signals; Using a simulated annealing ephemera algorithm to optimize the parameters of the variational empirical mode decomposition, and performing variational empirical mode decomposition on the original monitoring signal based on the optimized parameters; The monitoring signal is reconstructed according to the decomposed modal function.

2. The signal processing method according to claim 1, characterized in that: The simulated annealing ephemera algorithm is used to optimize the parameters of the variational empirical mode decomposition, including: The optimal parameters of the variational empirical mode decomposition are searched by using an ephemeral algorithm, and during the search process, the acceptance probability of the searched optimal parameters is determined by using a simulated annealing algorithm according to the fitness function value of the ephemeral algorithm.

3. The signal processing method according to claim 1, characterized in that: The local minimum envelope entropy is used as the fitness function of the simulated annealing mayfly algorithm to optimize the variational empirical mode decomposition.

4. The signal processing method according to claim 1, characterized in that: Reconstruct the monitoring signal according to the decomposed modal function, including: The monitoring signal is reconstructed based on the maximum correlation coefficient-kurtosis criterion.

5. A sample set construction method, characterized in that: include: Using the signal processing method according to any one of claims 1 to 4, the acquired original signal is processed to obtain an enhanced monitoring signal; A sample signal is intercepted from the monitoring signal through a set sliding window, and a training sample set is constructed according to the intercepted sample signal.

6. The sample set construction method according to claim 5, characterized in that: A training sample set is constructed based on the intercepted sample signal, including: Perform compressed sensing processing on the intercepted sample signal.

7. The sample set construction method according to claim 5, characterized in that: Also includes: All samples in the training sample set are normalized.

8. A method for constructing a fault diagnosis model, characterized in that: include: Using the sample set construction method as claimed in claim 5, constructing a source domain sample set and a target domain sample set respectively; Training the AIDBN model using the source domain sample set; Based on the target domain sample set, the trained AIDBN model parameters are migrated to the target domain to obtain the AIDBN diagnostic model corresponding to the target domain.

9. The fault diagnosis model construction method according to claim 8, characterized in that: The AIDBN model is trained using the source domain sample set, including: The simulated annealing ephemera algorithm is used to optimize the hyperparameters of the AIDBN model.

10. A fault diagnosis method, characterized in that: include: Adopting the fault diagnosis model construction method as claimed in claim 8 to construct an AIDBN diagnosis model; Acquire the current monitoring signal of the device, and extract features of the monitoring signal through the AIDBN diagnostic model to obtain the current feature label of the device; The feature tag is matched with the fault feature database, and the device fault is determined based on the matching result.

Images