Fault diagnosis method and device

Abstract

The present invention provides a fault diagnosis method and device, comprising: preprocessing a vibration signal of a planetary gearbox to obtain at least two sets of training data; training a target neural network model using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer to the input layer, and the second receiving layer is a feedback structure of the output layer to the hidden layer; integrating the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model; and diagnosing planetary gearbox faults according to the integrated diagnostic model. In the solution of the present invention, the two receiving layers can improve the dynamic processing capability of the planetary gearbox timing signals, making the integrated diagnostic model applicable to different operating conditions and environments. Furthermore, the integrated training of the diagnostic models using the integrated algorithm can ensure the stability and diagnostic accuracy of the integrated diagnostic model.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault diagnosis method and device. Background Art

[0002] Planetary gearboxes are widely used in a wide range of applications, including aerospace, automotive, and heavy industry, due to their high transmission ratios, compact design, and high load capacity. However, due to their complex structure and harsh operating environments, they are often susceptible to damage such as corrosion and abrasion. As damage accumulates, gears are prone to serious faults such as broken and missing teeth, which can lead to equipment downtime, high maintenance costs, and safety concerns. Therefore, planetary gearbox fault diagnosis can effectively improve equipment reliability and service levels.

[0003] Since planetary gearboxes have different operating conditions and environments in actual operation, existing diagnostic models are difficult to ensure diagnostic stability and accuracy when performing fault diagnosis. Summary of the Invention

[0004] The object of the present invention is to provide a fault diagnosis method and device to solve the problem that it is difficult to ensure diagnostic stability and accuracy when the existing diagnostic model performs fault diagnosis on a planetary gearbox.

[0005] To achieve the above objectives, an embodiment of the present invention provides a fault diagnosis method, comprising:

[0006] Preprocessing the vibration signal of the planetary gearbox to obtain at least two sets of training data;

[0007] The target neural network model is trained using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is a feedback structure of the output layer of the target neural network model to the hidden layer;

[0008] Performing integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model;

[0009] Perform fault diagnosis on the planetary gearbox according to the integrated diagnostic model and output a diagnostic result.

[0010] Optionally, performing integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model includes:

[0011] Obtaining a plurality of training subsets from the at least two sets of training data by using a bootstrap algorithm;

[0012] Based on each of the training subsets, the diagnostic model is trained multiple times using an iterative AdaBoost algorithm to obtain an AdaBoost ensemble model, wherein the AdaBoost algorithm updates the weight of the sample according to the diagnostic error of the sample during each model training process;

[0013] An average integration process is performed on the multiple AdaBoost integration models to obtain an integrated diagnostic model.

[0014] Optionally, the iterative AdaBoost algorithm is used to perform multiple model training on the diagnostic model to obtain an AdaBoost integrated model, including:

[0015] According to the relationship between the sample error and the adaptive change in the t-th model training process, the sample weight used in the t+1-th model training is determined;

[0016] Determine the AdaBoost training model obtained from the t+1th model training based on the sample weights used in the t+1th model training;

[0017] Based on the multiple AdaBoost training models, an AdaBoost integration model is obtained.

[0018] Optionally, determining the sample weights used for the t+1th model training according to the relationship between the sample error and the adaptive variation in the tth model training process includes:

[0019] When the sample error in the t-th model training process is greater than the adaptive change amount, the product of the sample weight in the t-th model training process and the inverse of the error of the diagnostic model is used as the sample weight for the t+1-th model training;

[0020] Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change, the product of the sample weight in the t-th model training process and the error of the diagnostic model is used as the sample weight used for the t+1-th model training.

[0021] Optionally, the adaptive change amount is obtained by the following formula:

[0022]

[0023] in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

[0024] In order to achieve the above object, an embodiment of the present invention further provides a fault diagnosis device, comprising:

[0025] A first processing module is used to pre-process the vibration signal of the planetary gearbox to obtain at least two sets of training data;

[0026] A first training module is configured to train a target neural network model using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is a feedback structure of the output layer of the target neural network model to the hidden layer;

[0027] A second training module is configured to perform integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model;

[0028] The diagnostic module is used to perform fault diagnosis on the planetary gearbox according to the integrated diagnostic model and output a diagnostic result.

[0029] Optionally, the second training module includes:

[0030] A first acquisition submodule is configured to acquire a plurality of training subsets from the at least two sets of training data by using a bootstrap algorithm;

[0031] A first training submodule is configured to perform multiple model training on the diagnostic model using an iterative AdaBoost algorithm based on each training subset to obtain an AdaBoost ensemble model, wherein the AdaBoost algorithm updates the weight of the sample according to the diagnostic error of the sample during each model training process;

[0032] The first processing submodule is used to perform average integration processing on the multiple AdaBoost integration models to obtain an integrated diagnostic model.

[0033] Optionally, the first training submodule includes:

[0034] The first determining unit is used to determine the sample weight used for the t+1th model training according to the relationship between the sample error and the adaptive change amount during the tth model training process;

[0035] The second determining unit is used to determine the AdaBoost training model obtained from the t+1th model training according to the sample weights used in the t+1th model training;

[0036] The acquisition unit is used to obtain an AdaBoost integrated model based on the multiple AdaBoost training models.

[0037] Optionally, the first determining unit is configured to, when the sample error in the t-th model training process is greater than the adaptive change amount, use the product of the sample weight in the t-th model training process and the inverse of the error of the diagnostic model as the sample weight for the t+1-th model training;

[0038] Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change, the product of the sample weight in the t-th model training process and the error of the diagnostic model is used as the sample weight used for the t+1-th model training.

[0039] Optionally, the adaptive change amount is obtained by the following formula:

[0040]

[0041] in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

[0042] In order to achieve the above-mentioned purpose, an embodiment of the present invention also provides a fault diagnosis device, including: a transceiver, a processor, a memory, and a program or instruction stored on the memory and executable on the processor; wherein, when the processor executes the program or instruction, the steps in the above-mentioned fault diagnosis method are implemented.

[0043] In order to achieve the above objectives, an embodiment of the present invention further provides a readable storage medium having a program or instruction stored thereon, which implements the steps in the above fault diagnosis method when executed by a processor.

[0044] The beneficial effects of the above technical solution of the present invention are as follows:

[0045] In an embodiment of the present invention, planetary gearbox fault diagnosis is performed using an integrated diagnostic model. This integrated diagnostic model is first trained using at least two sets of training data to obtain a target neural network model having the first and second receiving layers, and then integratedly trained using an integrated algorithm. Specifically, the integrated diagnostic model includes a dual feedback mechanism, from the input layer to the hidden layer and from the output layer to the input layer. This dual feedback mechanism improves the dynamic processing capability of the planetary gearbox timing signals, making the integrated diagnostic model adaptable to various operating conditions and environments. Furthermore, integrated training of the diagnostic model using an integrated algorithm further ensures the generalization capability, diagnostic stability, and accuracy of the integrated diagnostic model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of a fault diagnosis method according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of the OIF Elman neural network;

[0048] Figure 3 It is a schematic diagram of the structure of the OHF Elman neural network;

[0049] Figure 4 Schematic diagram of the structure of the OIHF Elman neural network according to an embodiment of the present invention;

[0050] Figure 5 Schematic diagram of the relationship between the number of hidden layer neurons and MSE of the neural network model;

[0051] Figure 6 This is one of the diagnostic effect diagrams on all test set samples under 20Hz-0V working conditions;

[0052] Figure 7 Schematic diagram of the diagnostic results of the test set under 20Hz-0V working conditions;

[0053] Figure 8 Schematic diagram of the distribution of the diagnosis results of the Bagging algorithm in terms of the number of weak learners;

[0054] Figure 9 This is the second diagnostic effect on all test set samples under 20Hz-0V working condition;

[0055] Figure 10 Schematic diagram of the diagnostic effect on all test set samples under 30Hz-2V working conditions;

[0056] Figure 11 A schematic diagram of a module of a fault diagnosis device according to an embodiment of the present invention;

[0057] Figure 12 This is one of the structural block diagrams of the fault diagnosis device according to an embodiment of the present invention;

[0058] Figure 13 This is the second structural block diagram of the fault diagnosis device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0060] It should be understood that references throughout this specification to "one embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present invention. Therefore, the appearances of "in one embodiment" or "in an embodiment" throughout this specification do not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0061] In various embodiments of the present invention, it should be understood that the size of the serial numbers of the following processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0062] Additionally, the terms "system" and "network" are often used interchangeably herein.

[0063] In the embodiments provided herein, it should be understood that "B corresponding to A" means that B is associated with A and B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B based solely on A; B can also be determined based on A and / or other information.

[0064] like Figure 1 As shown, an embodiment of the present invention provides a fault training method, including:

[0065] Step 101: Preprocess the vibration signal of the planetary gearbox to obtain at least two sets of training data.

[0066] In this embodiment of the present invention, the vibration signals of the planetary gears in the x-direction are first collected through the gearbox, and actual operating data sets are obtained under two operating conditions (the speed and load configuration settings are 20Hz-0V and 30Hz-2V, respectively). The data sets include four fault conditions: tooth erosion, missing teeth, root fault, tooth surface fault, and planetary gear vibration signals under normal conditions. Each column of signals contains 1,048,560 data points. Optionally, in order to effectively display the fault characteristics of the planetary gears, a length of 4369 is selected for signal segmentation, and 240 groups of samples are obtained for each column of signals, of which 140 groups are used as training sets and 100 groups are used as test sets.

[0067] Next, the vibration signal is subjected to noise reduction and feature extraction. Due to the characteristics of the gearbox such as harsh working environment and complex working conditions, the collected vibration signal is strongly affected by background noise, and effective noise reduction and feature extraction of the original signal are required. After that, Ensemble Empirical Mode Decomposition (EEMD) is used to decompose, reduce noise and reconstruct the signal. Among them, kurtosis and correlation coefficient are used as the judgment criteria for screening the basic modal components (Intrinsic Mode Function, IMF) obtained by EEMD modal decomposition, and six parameters in the time domain are selected: mean, standard deviation, root mean square value, skewness, kurtosis, and margin as the fault feature vector of the signal. The expected output form of the sample is [0,…,0,1,0…,0] T In the output result, position i corresponding to 1 represents the i-th failure mode, and the rest are 0. The mean and standard deviation of the mean square error are selected as evaluation indicators.

[0068] Step 102: Train the target neural network model using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is a feedback structure of the output layer of the target neural network model to the hidden layer.

[0069] In an embodiment of the present invention, the target neural network model is an OIHF Elman neural network.

[0070] First, the Elman neural network is explained as follows.

[0071] Elman neural network adds a connection layer to the traditional three-layer network structure (input layer, hidden layer and output layer). Assume that the input u is an r-dimensional vector, the hidden layer output x and the connection layer output x c are all n-dimensional vectors, and the output y of the output layer is an m-dimensional vector. Connection weight matrix W I1 , W I2 and W I3 are n×n, n×r and m×n dimensional matrices respectively. The nonlinear state space mathematical model of the Elman network is as follows:

[0072] x(k)=f(W I1 x c (k)+W I2 u(k-1));

[0073] x c (k) = αx c (k-1)+x(k-1);

[0074] y(k)=W I3 x(k);

[0075] Where α is the self-connection feedback gain factor, f(·) is usually taken as a sigmoid function, the value range of α is (0, 1), and k represents the number of iterations.

[0076] Preferably, considering that the feedback information of each layer of the neural network will affect the accuracy and speed of the network's data processing, a feedback mechanism of the output layer is further designed based on the Elman neural network, which is the feedback of the output layer to the input layer and the feedback of the output layer to the hidden layer, namely OIF Elman neural network and OHF Elman neural network. The structures of OIF Elman neural network and OHF Elman neural network are as follows: Figure 2 and Figure 3 As shown in the figure, the feedback structure of the output layer added to the input layer of the OIF Elman neural network is called the connecting layer 2, and its connection weight matrix is ​​W I4 The feedback structure of the output layer added to the OHF Elman neural network to the hidden layer is called the connecting layer 3, and its connection weight matrix is ​​W I5 The nonlinear state space mathematical model of the OIF Elman neural network is:

[0077] x(k)=f(W I1 x c (k)+W I2 u(k-1)+W I4 y c (k));

[0078] x c (k) = αx c (k-1)+x(k-1);

[0079] y c (k) = γy c (k-1)+y(k-1);

[0080] y(k)=W I3 x(k);

[0081] Nonlinear state space mathematical model of OHF Elman network:

[0082] x(k)=f(W I1 x c (k)+W I2 u(k-1));

[0083] x c (k) = αx c (k-1)+x(k-1);

[0084] y c (k) = γy c (k-1)+y(k-1);

[0085] y(k)=W I3 x(k)+W I5 y c (k);

[0086] Among them, α and γ are self-connection feedback gain factors, f(·) is usually taken as a sigmoid function, and the value range of α and γ is [0,1).

[0087] In view of the inherent similarity of planetary gearbox fault signals, a dual feedback mechanism of the output layer to the hidden layer and the output layer to the input layer is designed, namely the OIHF Elman neural network (the above target neural network model). The proposed OIHFElman neural network topology is as follows Figure 4 As shown, where the connection weight matrix W I4 、W I5 Respectively with the weight matrix W in the OIF Elman neural network I4 , the weight matrix W in the OHF Elman neural network I5 The mathematical model of the OIHF Elman neural network is:

[0088] x(k)=f(W I1 x c (k)+W I2 u(k-1)+W I4 y c (k));

[0089] x c (k) = αx c (k-1)+x(k-1);

[0090] y c (k) = γy c (k-1)+y(k-1);

[0091] y(k)=W I3 x(k)+W I5 y c (k);

[0092] Among them, α and γ are the self-connection feedback gain factors, f(·) is usually taken as a sigmoid function, and the value range of α and γ is [0,1).

[0093] The OIHF Elman neural network has self-feedback gain factors: α and γ, which improve the network's processing ability for gearbox timing data.

[0094] The OIHF Elman network structure is r×n×m. The dimension of its output layer ensures the complexity of the research problem. It can consider the composite fault of rolling bearing (the situation where multiple faults occur at the same time). For example, the output format is [0,…,0,1,0…,0] T Each element can indicate the possibility that the sample belongs to this type of fault. If a compound fault occurs, it can be represented by multiple elements being 1, and then the fault form and state of multiple faults occurring at the same time can be diagnosed.

[0095] Step 103: Perform integration training on the at least two diagnostic models according to the integration algorithm to obtain an integrated diagnostic model.

[0096] Here, the above-mentioned integrated algorithms include the AdaBoost algorithm and the Bagging algorithm.

[0097] Step 104: Perform fault diagnosis on the planetary gearbox according to the integrated diagnostic model and output a diagnostic result.

[0098] The fault diagnosis method of an embodiment of the present invention utilizes an integrated diagnostic model to diagnose planetary gearbox faults. This integrated diagnostic model is first trained using at least two sets of training data to obtain a target neural network model having the first and second receiving layers, and then integratedly trained using an integrated algorithm. Specifically, the integrated diagnostic model includes a dual feedback mechanism, from the input layer to the hidden layer and from the output layer to the input layer. This dual feedback mechanism improves the dynamic processing capability of the planetary gearbox timing signals, making the integrated diagnostic model adaptable to various operating conditions and environments. Furthermore, integrated training of the diagnostic model using an integrated algorithm further ensures the generalization capability, diagnostic stability, and diagnostic accuracy of the integrated diagnostic model.

[0099] Optionally, performing integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model includes:

[0100] Obtaining a plurality of training subsets from the at least two sets of training data by using a bootstrap algorithm;

[0101] Based on each of the training subsets, the diagnostic model is trained multiple times using an iterative AdaBoost algorithm to obtain an AdaBoost ensemble model, wherein the AdaBoost algorithm updates the weight of the sample according to the diagnostic error of the sample during each model training process;

[0102] An average integration process is performed on the multiple AdaBoost integration models to obtain an integrated diagnostic model.

[0103] The AdaBoost algorithm updates the weight of the sample according to the diagnostic error of the sample during each model training process.

[0104] In the embodiment of the present invention, a training subset is obtained by a self-service bootstrap algorithm (Bagging algorithm), that is, sampling with replacement is performed by Bagging, which reduces the impact of difficult-to-classify samples on the AdaBoost algorithm, effectively ensures the generalization ability and stability of the diagnostic model, and performs model training by the AdaBoost algorithm, which can effectively improve the accuracy of the diagnostic model.

[0105] Optionally, the iterative AdaBoost algorithm is used to perform multiple model training on the diagnostic model to obtain an AdaBoost integrated model, including:

[0106] According to the relationship between the sample error and the adaptive change in the t-th model training process, the sample weight used in the t+1-th model training is determined;

[0107] Determine the AdaBoost training model obtained from the t+1th model training based on the sample weights used in the t+1th model training;

[0108] Based on the multiple AdaBoost training models, an AdaBoost integration model is obtained.

[0109] Here, t is greater than or equal to 1.

[0110] Specifically, according to the sample weights used in the t+1th model training, the error of the diagnostic model of the t+1th model training is obtained, and according to the error of the diagnostic model of the t+1th model training, the AdaBoost training model obtained by the t+1th model training is obtained; then, the AdaBoost training model obtained by the 1st model training and the AdaBoost training model obtained by the t+1th model training are weightedly integrated to obtain the AdaBoost integration model.

[0111] In an embodiment of the present invention, the adaptive variation is related to the sample error and the sample number, rather than a fixed threshold. By replacing the fixed threshold with the adaptive variation, the sample weights are updated based on the relationship between the sample error and the adaptive variation. This allows the AdaBoost algorithm to be applied to different sample data and scenarios without the need to determine an effective threshold through preliminary experiments. This also controls the weight distribution of AdaBoost on erroneous samples, slowing down its iterative variation amplitude and preventing the expansion of weights due to some difficult-to-distinguish samples, thereby reducing the possibility of AdaBoost overfitting. Furthermore, the sample iteration formula is changed to control the expansion rate of sample weights and achieve a reasonable distribution.

[0112] Further optionally, determining the sample weights used for the t+1th model training according to the relationship between the sample error and the adaptive change amount during the tth model training process includes:

[0113] When the sample error in the t-th model training process is greater than the adaptive change amount, normalizing the first sample weight in the t-th model training process to obtain a sample weight for the t+1-th model training, where the first sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the inverse of the error of the diagnostic model;

[0114] Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change, the second sample weight in the t-th model training process is normalized to obtain the sample weight used for the t+1-th model training, where the second sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the error of the diagnostic model.

[0115] Further optionally, the adaptive change amount is obtained by the following formula:

[0116]

[0117] in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

[0118] The AdaBoost algorithm according to the embodiment of the present invention is described below.

[0119] Input: Original sample set: Weak learner: H(x), number of weak learners: l;

[0120] Initialization: The initial weight of the training sample is Dt (i) = 1 / N, i = 1, 2, ..., N, error rate: ε t =0;

[0121] Training: For t = 1, 2, ..., l, cyclic training is performed to establish a regression model based on the sample set T and the weak learner H(x): H t (x);

[0122] Compute the error for each sample:

[0123] Among them, T i is the real output, O ti is the model output, and s is the dimension of each set of sample output.

[0124] Adaptive change:

[0125] Calculate the error of the weak learner: n can be 1, 2, or 3, but is usually 1;

[0126] Sample weight update: And normalize it:

[0127] Output: After l rounds of training, we get l weak learners, which are combined into a strong learner through weighted combination:

[0128]

[0129] In the embodiment of the present invention, some optimizations are performed in the AdaBoost algorithm, replacing the fixed threshold with an adaptive change. In the sample weight iteration formula, when the sample error is greater than the threshold, its change coefficient is changed from 1 to the inverse of the weak learner error of that round, which can improve the diagnostic accuracy of the entire sample.

[0130] In a specific embodiment of the present invention, the Bagging algorithm is combined with the aforementioned AdaBoost algorithm. First, multiple training subsets are obtained using the bootstrap method. Multiple rounds of neural network model training are performed on each training subset. In each round, the AdaBoost algorithm is used to adjust the sample weights to obtain model weights. The models are then weighted ensembled. Finally, the AdaBoost ensemble models from the multiple training subsets are averaged to obtain the AdaBoost-Bagging ensemble model.

[0131] AdaBoost controls the weight of each sample by calculating the sample's diagnostic error, thereby improving the diagnostic accuracy of the entire sample. Bagging effectively controls the training samples of the diagnostic model through sampling with replacement, thereby obtaining a diagnostic model that is applicable to different samples, and then obtains a strong learner through weighting. Because AdaBoost is highly sensitive to difficult samples, it can cause sample weights to expand uncontrollably. The AdaBoost-Bagging dual ensemble algorithm model, which introduces bagging, reduces the impact of difficult samples on AdaBoost by changing the training samples through bagging.

[0132] The following is an explanation of the algorithm flow of the AdaBoost-Bagging integration model:

[0133] Input: Original sample set: Weak learner: H(x);

[0134] The number of weak learners in the AdaBoost algorithm: l;

[0135] The number of weak learners in the Bagging algorithm: k;

[0136] Training: for i=1, 2, ..., k, cyclic training;

[0137] The bootstrap method is used to obtain a sample set with a sample size of N: T i As a training subset, and initialize the sample weight to D i，1 (t)=1 / N,t=1,2,…,N,error rate ε t =0; training: when j=1, 2, ..., l, cyclic training;

[0138] Based on the training subset T i Build a regression model with weak learner H(x): H ij (x); calculate the error for each sample: Among them, T t is the real output, O ijt is the model output, s is the dimension of each set of sample output;

[0139] Adaptive change:

[0140] Calculate H ij (x) Error:

[0141] n can be 1, 2, or 3, but is usually 1;

[0142] Sample weight update:

[0143] And normalize it:

[0144] Output: AdaBoost strong learner;

[0145] After l rounds of training, l weak learners are obtained, which are combined with strong learners through weighting:

[0146]

[0147] Output: AdaBoost-Bagging strong learner:

[0148] The following is an explanation based on the experimental data of the neural network model.

[0149] Based on the sample of 20Hz-0V working condition, the following network parameters are set: 6 input neurons, 5 output neurons, hidden layer activation function is tansig, output layer is logsig, training function is trainlm, and the error function is MSE (Mean Square Error). In order to effectively select the number of hidden layer neurons, an integer set in the interval [5,16] is selected for analysis based on the Elman neural network, and the average value of the diagnostic error MSE is selected as the evaluation target. Figure 5 , the number of neurons in the hidden layer is selected as 6.

[0150] According to the training set and test set divided and extracted in step 101 and the above parameter design, numerical experiments of Elman neural network, OIF Elman neural network, OHF Elman neural network and OIHF Elman neural network are carried out respectively. The diagnostic effects on all test set samples under 20Hz-0V working conditions are as follows: Figure 6 The overall diagnostic effect on the test set is shown in Table 1. It can be seen that the feedback effect of the output layer is more obvious, among which the OIHF Elman neural network achieves better improvement effect.

[0151] Table 1

[0152] Elman OHF Elman OIF Elman OIHF Elman MSE - Mean 0.0350 0.0057 0.0063 0.0023 MSE - Standard Deviation 0.0885 0.0236 0.0278 0.0217

[0153] Numerical experiments of AdaBoost-Bagging ensemble model:

[0154] Based on the samples of 20Hz-0V working condition, the following ensemble algorithm parameters are set: the number of weak learners of AdaBoost ensemble algorithm is 15, and the number of weak learners of Bagging is 30. Considering the iteration time and effectiveness of ensemble learning, the OIHFElman neural network is used as a weak learner, and the diagnosis analysis of AdaBoost ensemble algorithm on the integer interval [5,16] is performed. The diagnosis results of the test set under 20Hz-0V working condition are as follows: Figure 7 As shown. For the Bagging ensemble algorithm, the OIHF Elman neural network is also used as a weak learner. First, a numerical experiment with 100 weak learners is conducted to determine that the number of weak learners for the Bagging algorithm ranges from 1 to 100. Then, 1 to 100 weak learners are randomly selected with replacement for combination. This is repeated 30 times and the average result is obtained. 110% of the MSE of the diagnostic error under 100 weak learners is selected as the upper limit of the error. The distribution of the diagnostic results of the Bagging algorithm in terms of the number of weak learners is obtained as shown below. Figure 8 As shown. Figure 7 and Figure 8 , it can be found that the optimal number of weak learners for AdaBoost algorithm and Bagging algorithm are 15 and 30 respectively.

[0155] According to the previously extracted training set, test set and the above parameter design, the OHF Elman, OIF Elman and OIHF Elman neural networks were selected to conduct numerical experiments on AdaBoost integration, Bagging integration and AdaBoost-Bagging dual integration respectively. The diagnostic effects on all test set samples of 20Hz-0V working condition are as follows: Figure 9 The overall diagnostic effect on the 20Hz-0V working condition test set is shown in Table 2. It can be seen that the AdaBoost-Bagging dual integration model achieves more accurate diagnosis on most samples and achieves the most stable and accurate diagnostic effect on the entire test set.

[0156] Table 2

[0157]

[0158] Retain the relevant parameters of the AdaBoost-Bagging integrated model and perform fault diagnosis on the 30Hz-2V working condition samples. The diagnostic effect on all test set samples of the 30Hz-2V working condition is as follows: Figure 10 The overall diagnosis results on the 30Hz-2V working condition test set are shown in Table 3. From the data in the table, we can see that the OIHF Elman AdaBoost-Bagging model achieves high diagnostic accuracy and stability on the 30Hz-2V working condition samples.

[0159] Table 3

[0160]

[0161] In the embodiment of the present invention, the expected output format of the sample is defined as: [0, ..., 0, 1, 0 ..., 0] T , the position i corresponding to 1 represents the i-th fault mode, and the rest are 0. Therefore, each group of test samples will obtain an absolute error matrix. If 0.5 is defined as the upper limit of all elements in the absolute error matrix, according to the diagnosis results of the OIHF Elman AdaBoost-Bagging model on the 20Hz-0V and 30Hz-2V working condition samples, it can be obtained that the diagnostic accuracy of the OIHF Elman AdaBoost-Bagging model on the 20Hz-0V and 30Hz-2V working condition samples is 99.6% and 96.2%, respectively, which verifies the effectiveness of the OIHF ElmanAdaBoost-Bagging model in planetary gear fault diagnosis under variable working conditions.

[0162] The fault diagnosis method of an embodiment of the present invention utilizes an integrated diagnostic model to diagnose planetary gearbox faults. This integrated diagnostic model is first trained using at least two sets of training data to obtain a target neural network model having the first and second receiving layers, and then integratedly trained using an integrated algorithm. Specifically, the integrated diagnostic model includes a dual feedback mechanism, from the input layer to the hidden layer and from the output layer to the input layer. This dual feedback mechanism improves the dynamic processing capability of the planetary gearbox timing signals, making the integrated diagnostic model adaptable to various operating conditions and environments. Furthermore, integrated training of the diagnostic model using an integrated algorithm further ensures the generalization capability, diagnostic stability, and diagnostic accuracy of the integrated diagnostic model.

[0163] like Figure 11 As shown, the embodiment of the present invention further provides a fault diagnosis device 1100, including:

[0164] A first processing module 1101 is configured to pre-process the vibration signal of the planetary gearbox to obtain at least two sets of training data;

[0165] A first training module 1102 is configured to train a target neural network model using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is a feedback structure of the output layer of the target neural network model to the hidden layer;

[0166] The second training module 1103 is configured to perform integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model;

[0167] The diagnosis module 1104 is configured to perform fault diagnosis on the planetary gearbox according to the integrated diagnosis model and output a diagnosis result.

[0168] Optionally, the second training module includes:

[0169] A first acquisition submodule is configured to acquire a plurality of training subsets from the at least two sets of training data by using a bootstrap algorithm;

[0170] A first training submodule is configured to perform multiple model training on the diagnostic model using an iterative AdaBoost algorithm based on each training subset to obtain an AdaBoost ensemble model, wherein the AdaBoost algorithm updates the weight of the sample according to the diagnostic error of the sample during each model training process;

[0171] The first processing submodule is used to perform average integration processing on the multiple AdaBoost integration models to obtain an integrated diagnostic model.

[0172] Optionally, the first training submodule includes:

[0173] The first determining unit is used to determine the sample weight used for the t+1th model training according to the relationship between the sample error and the adaptive change amount during the tth model training process;

[0174] The second determining unit is used to determine the AdaBoost training model obtained from the t+1th model training according to the sample weights used in the t+1th model training;

[0175] The acquisition unit is used to obtain an AdaBoost integrated model based on the multiple AdaBoost training models.

[0176] Optionally, the first determining unit is configured to, when the sample error in the t-th model training process is greater than the adaptive change amount, perform normalization processing on the first sample weight in the t-th model training process to obtain a sample weight for the t+1-th model training, where the first sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the inverse of the error of the diagnostic model;

[0177] Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change, the second sample weight in the t-th model training process is normalized to obtain the sample weight used for the t+1-th model training, where the second sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the error of the diagnostic model.

[0178] Optionally, the adaptive change amount is obtained by the following formula:

[0179]

[0180] in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

[0181] The fault diagnosis device can implement all the implementations of the above-mentioned fault diagnosis method embodiments, and to avoid repetition, they are not described here again.

[0182] like Figure 12 As shown, the embodiment of the present invention further provides a fault diagnosis device 1200, including: a processor 1210:

[0183] The processor 1210 is used to preprocess the vibration signal of the planetary gearbox to obtain at least two sets of training data; train the target neural network model through the at least two sets of training data to obtain at least two diagnostic models, the target neural network model includes a first receiving layer and a second receiving layer, the first receiving layer is the feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is the feedback structure of the output layer of the target neural network model to the hidden layer; according to the integration algorithm, the at least two diagnostic models are integrated trained to obtain an integrated diagnostic model; the planetary gearbox fault is diagnosed according to the integrated diagnostic model, and the diagnostic results are output.

[0184] The processor 1210 is further configured to: obtain a plurality of training subsets from the at least two sets of training data using a bootstrap algorithm; perform multiple model training on the diagnostic model using an iterative AdaBoost algorithm based on each of the training subsets to obtain an AdaBoost integrated model, wherein the AdaBoost algorithm updates the weight of the sample according to the diagnostic error of the sample during each model training process; and perform average integration processing on the multiple AdaBoost integrated models to obtain an integrated diagnostic model.

[0185] The processor 1210 is further configured to: determine the sample weight used for the t+1th model training based on the relationship between the sample error and the adaptive change during the tth model training; determine the AdaBoost training model obtained from the t+1th model training based on the sample weight used for the t+1th model training; and obtain an AdaBoost integrated model based on multiple AdaBoost training models.

[0186] The processor 1210 is further configured to: when the sample error in the t-th model training process is greater than the adaptive change amount, normalize the first sample weight in the t-th model training process to obtain a sample weight for the t+1-th model training, where the first sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the inverse of the error of the diagnostic model;

[0187] Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change, the second sample weight in the t-th model training process is normalized to obtain the sample weight used for the t+1-th model training, where the second sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the error of the diagnostic model.

[0188] Optionally, the adaptive change amount is obtained by the following formula:

[0189]

[0190] in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

[0191] The fault diagnosis device can implement all the implementations of the above-mentioned fault diagnosis method embodiments, and to avoid repetition, they are not described here again.

[0192] Another embodiment of the present invention further provides a fault diagnosis device, such as Figure 13 As shown, it includes a transceiver 1310, a processor 1300, a memory 1320, and a program or instruction stored in the memory 1320 and executable on the processor 1300; when the processor 1300 executes the program or instruction, the steps of the above-mentioned fault diagnosis method are implemented.

[0193] The transceiver 1310 is configured to receive and send data under the control of the processor 1300 .

[0194] Among them, Figure 13In the embodiment, the bus architecture may include any number of interconnected buses and bridges, specifically various circuits of one or more processors represented by processor 1300 and memory represented by memory 1320. The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 1310 may be a plurality of elements, i.e., a transmitter and a receiver, providing a unit for communicating with various other devices on a transmission medium. The processor 1300 is responsible for managing the bus architecture and general processing, and the memory 1320 may store data used by the processor 1300 when performing operations.

[0195] An embodiment of the present invention also provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps in the fault diagnosis method described above are implemented and the same technical effect can be achieved. To avoid repetition, they will not be described here.

[0196] The processor is the processor in the fault diagnosis device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0197] In embodiments of the present invention, modules can be implemented in software so that they can be executed by various types of processors. For example, an identified executable code module can include one or more physical or logical blocks of computer instructions, for example, which can be constructed as objects, procedures, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but can include different instructions stored in different locations, which, when logically combined together, constitute the module and achieve the specified purpose of the module.

[0198] In fact, executable code module can be a single instruction or many instructions, and can even be distributed on a plurality of different code segments, distributed in the middle of different programs, and distributed across a plurality of memory devices.Similarly, operating data can be identified in the module, and can be implemented and organized in the data structure of any appropriate type according to any appropriate form.Described operating data can be collected as a single data set, or can be distributed in different locations (including on different storage devices), and can only be present on a system or network as an electronic signal at least in part.

[0199] When a module can be implemented using software, given the current state of hardware technology, those skilled in the art can build corresponding hardware circuits to implement the corresponding functions of the module, regardless of cost. The hardware circuits may include conventional very large scale integration (VLSI) circuits or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules may also be implemented using programmable hardware devices, such as field programmable gate arrays, programmable array logic, or programmable logic devices.

[0200] The above exemplary embodiments are described with reference to the accompanying drawings. Many different forms and embodiments are possible without departing from the spirit and teachings of the present invention. Therefore, the present invention should not be construed as limited to the exemplary embodiments set forth herein. Rather, these exemplary embodiments are provided so that this disclosure will be complete and perfect and will convey the scope of the invention to those skilled in the art. In the drawings, component sizes and relative sizes may be exaggerated for clarity. The terminology used herein is for purposes of describing specific exemplary embodiments only and is not intended to be limiting. As used herein, the singular forms "a," "an," and "the" are intended to include plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "comprising" and / or "including," when used in this specification, indicate the presence of stated features, integers, steps, operations, components, and / or elements, but do not preclude the presence or addition of one or more other features, integers, steps, operations, components, elements, and / or groups thereof. Unless otherwise indicated, when stated, a range of values ​​includes the upper and lower limits of that range and any subranges therebetween.

[0201] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A fault diagnosis method, characterized in that: include: Preprocessing the vibration signal of the planetary gearbox to obtain at least two sets of training data; The target neural network model is trained using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is a feedback structure of the output layer of the target neural network model to the hidden layer; Performing integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model; Performing fault diagnosis on the planetary gearbox according to the integrated diagnostic model and outputting the diagnostic result; The step of performing integrated training on the at least two diagnostic models according to the integrated algorithm to obtain an integrated diagnostic model includes: Based on each training subset, when the sample error in the t-th model training process is greater than the adaptive change amount, normalize the first sample weight in the t-th model training process to obtain the sample weight used for the t+1-th model training, where the first sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the inverse of the error of the diagnostic model; Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change amount, the second sample weight in the t-th model training process is normalized to obtain a sample weight for the t+1-th model training, where the second sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the error of the diagnostic model; Determine the AdaBoost training model obtained from the t+1th model training based on the sample weights used in the t+1th model training; Performing average integration processing on multiple AdaBoost integration models to obtain the integrated diagnostic model; The adaptive change is obtained by the following formula: in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

2. The method according to claim 1, characterized in that The step of performing integrated training on the at least two diagnostic models according to the integrated algorithm to obtain an integrated diagnostic model further includes: A plurality of training subsets are obtained from the at least two sets of training data using a bootstrap algorithm.

3. A fault diagnosis device, characterized in that: include: A first processing module is used to pre-process the vibration signal of the planetary gearbox to obtain at least two sets of training data; A first training module is configured to train a target neural network model using the at least two sets of training data to obtain at least two diagnostic models, wherein the target neural network model includes a first receiving layer and a second receiving layer, wherein the first receiving layer is a feedback structure of the output layer of the target neural network model to the input layer, and the second receiving layer is a feedback structure of the output layer of the target neural network model to the hidden layer; A second training module is used to perform integrated training on the at least two diagnostic models according to an integrated algorithm to obtain an integrated diagnostic model; a diagnostic module, configured to perform fault diagnosis on the planetary gearbox according to the integrated diagnostic model and output a diagnostic result; Wherein, the second training module includes: a first determining unit configured to normalize, based on each training subset, a first sample weight in the t-th model training process when the sample error in the t-th model training process is greater than the adaptive change amount, to obtain a sample weight for the t+1-th model training, where the first sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the inverse of the error of the diagnostic model; Alternatively, when the sample error in the t-th model training process is less than or equal to the adaptive change amount, the second sample weight in the t-th model training process is normalized to obtain a sample weight for the t+1-th model training, where the second sample weight is the sample weight obtained by multiplying the sample weight in the t-th model training process by the error of the diagnostic model; The second determining unit is used to determine the AdaBoost training model obtained from the t+1th model training according to the sample weights used in the t+1th model training; An acquisition unit, configured to obtain an AdaBoost integrated model based on the plurality of AdaBoost training models; The adaptive change is obtained by the following formula: in, Represents the adaptive change during the t-th model training process, MSE ti It represents the sample error of the i-th sample in the t-th model training, N represents the number of samples, 1≤i≤N.

4. The device according to claim 3, characterized in that The second training module also includes: The first acquisition submodule is configured to acquire a plurality of the training subsets from the at least two sets of training data by using a bootstrap algorithm.

5. A fault diagnosis device comprising: A transceiver, a processor, a memory, and a program or instruction stored in the memory and executable on the processor; wherein the processor implements the steps of the fault diagnosis method according to any one of claims 1 to 2 when executing the program or instruction.

6. A readable storage medium having a program or instruction stored thereon, characterized in that: When the program or instruction is executed by a processor, the steps of the fault diagnosis method according to any one of claims 1 to 2 are implemented.

Images