Planetary gearbox fault diagnosis method based on deep learning and transfer learning

Abstract

The invention provides a planetary gearbox fault diagnosis method based on deep learning and transfer learning. The method comprises the steps of 1, collecting vibration acceleration signals of a planetary gearbox in different states under different working conditions; 2, dividing the vibration acceleration signal into source domain data and target domain data; step 3, constructing a deep learning model in which CNN-LSTM is fused with CAM (Content Addressable Memory) and Self-Attention; 4, debugging and optimizing the deep learning model by using the source domain data, and determining an optimal pre-training model; 5, expanding and finely adjusting the sample; 6, designing and comparing fine-tuning migration schemes, and determining an optimal fine-tuning migration scheme; step 7, utilizing the optimal fine-tuning migration scheme to finely tune the migration model according to the expanded fine-tuning sample; and step 8, performing fault diagnosis on the planetary gearbox according to the target domain data by using the fine-tuned migration model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a planetary gearbox fault diagnosis method based on deep learning and transfer learning. Background Art

[0002] Planetary gearboxes have the advantages of high efficiency, compact design, high torque density and multi-speed ratio selectivity, and are often used in many fields such as aerospace, wind power generation, industrial machinery, and robotics. However, due to their complex structure and harsh service conditions, they are prone to failure. Deep learning technology can autonomously identify features and directly obtain complex functions from input to output from data. It has been applied by many scholars to the fault diagnosis tasks of gears and bearings.

[0003] Traditional planetary gearbox fault diagnosis algorithms are based on fixed-condition fault diagnosis scenarios and are not applicable to cross-condition fault diagnosis scenarios. Fine-tuning migration models is considered an effective way to diagnose cross-condition faults, and the performance of pre-trained models is a key factor affecting the efficiency of cross-condition fault diagnosis. Secondly, if the number of fine-tuning samples is too small, it is easy to cause problems such as model overfitting, decreased stability, and poor cross-condition fault diagnosis results.

[0004] Therefore, a solution is urgently needed. Summary of the invention

[0005] One of the purposes of the present invention is to provide a planetary gearbox fault diagnosis method based on deep learning and transfer learning, which improves the performance of the pre-trained model and increases the number of fine-tuning samples, and can effectively improve the cross-operating condition fault diagnosis performance of the planetary gearbox.

[0006] An embodiment of the present invention provides a planetary gearbox fault diagnosis method based on deep learning and transfer learning, comprising:

[0007] Step 1: Collect vibration acceleration signals of the planetary gearbox in different states under different working conditions;

[0008] Step 2: Divide the vibration acceleration signal into source domain data and target domain data;

[0009] Step 3: Build a deep learning model of CNN-LSTM integrating CAM and Self-Attention;

[0010] Step 4: Use source domain data to debug and optimize the deep learning model and determine the optimal pre-training model;

[0011] Step 5: Expand and fine-tune samples;

[0012] Step 6: Design and compare fine-tuning migration plans to determine the optimal fine-tuning migration plan;

[0013] Step 7: Use the optimal fine-tuning migration solution to fine-tune the migration model according to the expanded fine-tuning samples;

[0014] Step 8: Use the fine-tuned migration model to perform fault diagnosis on the planetary gearbox according to the target domain data.

[0015] Optionally, the vibration acceleration signal is collected in step 1 by an acceleration sensor;

[0016] The vibration acceleration signal is the signal of the vibration acceleration of the planetary gearbox in the X direction, Y direction, and Z direction.

[0017] Optionally, the deep learning model constructed in step three includes 1 input layer B1, 4 convolutional layers C1, C2, C3, C4, 4 pooling layers D1, D2, D3, D4, 1 channel attention mechanism layer E1, 1 long short-term memory network layer F1, 1 self-attention mechanism layer G1, 1 flattening layer H1, 2 fully connected layers I1, I2, 3 discard layers J1, J2, J3 and 1 output classification layer K1; the order of each layer is: B1, C1, D1, C2, D2, E1, F1, C3, D3, C4, D4, G1, H1, J1, I1, J2, I2, J3, K1.

[0018] Optionally, before debugging and optimizing the deep learning model using the source domain data in step 4, 1024 consecutive data points are continuously intercepted from the source domain data as a sample, the overlap between two adjacent samples is 0, 900 samples are taken for each state, and they are divided into a source domain training set, a source domain verification set, and a source domain test set in a ratio of 7:1:1.

[0019] Optionally, before debugging and optimizing the deep learning model using the source domain data in step 4, the optimal pre-trained model is determined based on the model evaluation system, according to samples intercepted by the deep learning model in different source domain data and the model loss accuracy curves exhibited by different optimizers.

[0020] Optionally, when expanding the fine-tuning samples in step five, a random noise sequence of a certain length is generated, and a random position is generated on the original fine-tuning sample, and some data points of the original fine-tuning sample are replaced with the generated random noise sequence to expand the number of fine-tuning samples; the length of the random noise sequence is α, and the value of α includes at least: 128, 256, 512, and the replacement is repeated 9 times to obtain 500 expanded fine-tuning samples of each type.

[0021] Optionally, the fine-tuning migration solution includes:

[0022] Solution 1: Migrate the pre-trained model to the target domain task, expand the fine-tuning samples, and fine-tune the entire migration model;

[0023] Solution 2: Execute Solution 1 and freeze the migration model to the shallow CNN (D2).

[0024] Solution 3: Execute Solution 1 and migrate the frozen model to CAM;

[0025] Solution 4: Execute Solution 1 and migrate the model to LSTM.

[0026] Solution 5: Execute Solution 1 and migrate the model to the deep CNN (D4).

[0027] Solution 6: Execute Solution 1 and freeze the migration model to Self-Attention.

[0028] Solution 7: Execute Solution 1 and freeze the entire model (except the classification layer).

[0029] Solution 8: The model is not fine-tuned and transferred, and the target domain samples are used to directly train the model;

[0030] Solution 9: Execute Solution 1, migrate the pre-trained model to the target domain task, and fine-tune the entire migration model using target domain samples.

[0031] Optionally, in step eight, when the fine-tuned migration model is used to perform fault diagnosis on the planetary gearbox according to the target domain data, 1024 consecutive data points are continuously intercepted from the target domain data as a sample, the overlap between two adjacent samples is 0, and 250 samples are taken for each type of state in the target domain, which are divided into corresponding target domain training set, validation set and test set in a ratio of 1:2:2.

[0032] Optionally, the planetary gearbox fault diagnosis method based on deep learning and transfer learning further includes:

[0033] Obtaining a fault diagnosis history for troubleshooting a planetary gearbox;

[0034] Necessary expansion of fault diagnosis history;

[0035] Generate a digital model based on the fault diagnosis history expanded by necessity;

[0036] Based on the digital model, interactively determine the diagnostic optimization plan with the user;

[0037] Execute diagnostic optimization program.

[0038] Optionally, the fault diagnosis history is expanded as necessary, including:

[0039] The fault diagnosis history is represented by a time axis to obtain a historical time axis; wherein different historical items are set at different time nodes on the historical time axis, and the generation time of the historical item corresponds to the time node at which it is located;

[0040] Determine a historical item combination from the historical time axis; wherein the on-axis spacing between adjacent historical items in the historical item combination does not exceed a spacing threshold and has a preset standard historical item association relationship; and the total number of historical items in the same historical item combination exceeds a number threshold;

[0041] Based on the combination selection rule, historical item combinations are selected in sequence on the historical time axis;

[0042] Each time a selection is made, the selected historical item combination is characterized to obtain a characteristic entity;

[0043] Matching the feature entity with the index feature entity;

[0044] When the match is met, based on the situation analysis template corresponding to the index feature entity that matches the match, the situation analysis is performed on the historical item combination selected on the historical time axis and other historical item combinations within the preset axis distance before and after the selected historical item combination to obtain the historical situation;

[0045] Match historical situations with standard situations;

[0046] When the match is met, the selected historical item combination is expanded based on the historical expansion template corresponding to the standard case of the match;

[0047] When all combinations of history items are selected, the expanded fault diagnosis history is determined based on the current history time axis.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The present invention constructs a CNN-LSTM fusion CAM and SelfAttention fault diagnosis algorithm model, which shows good fixed-condition fault diagnosis performance under vibration acceleration signals in different directions of the planetary gearbox, and the model has strong generalization;

[0050] (2) The present invention obtains a pre-trained model with better performance by fully training and debugging the model on the source domain data and selecting the optimizer type and the number of training samples based on the performance of the model on the source domain validation set;

[0051] (3) The present invention introduces a sample expansion method, which generates a random noise sequence of a certain length to replace some data points of the sample, and replaces them multiple times to expand the sample quantity;

[0052] (4) The present invention combines the transfer learning theory and uses a sample expansion method to expand the number of fine-tuning samples. The expanded fine-tuning samples are used to fine-tune the migration model to complete the cross-operating fault diagnosis of the planetary gearbox. This method solves the problem that the model is prone to overfitting, stability is reduced, and the cross-operating diagnosis accuracy is not high due to the small number of fine-tuning samples.

[0053] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0054] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0056] Figure 1 It is a schematic diagram of the overall structure of the model of the present invention.

[0057] Figure 2 Schematic diagram of the source domain and target domain sample capture process of the present invention.

[0058] Figure 3 This is a comparison curve of the model validation set loss under different numbers of training samples in the present invention.

[0059] Figure 4 This is a comparison curve of the model validation set accuracy under different numbers of training samples in the present invention.

[0060] Figure 5 This is a comparison curve of the model validation set loss under different optimizers of the present invention.

[0061] Figure 6 This is a comparison curve of the model validation set accuracy under different optimizers of the present invention.

[0062] Figure 7 This is a bar graph of the diagnostic accuracy of the present invention under different working conditions and sensor data at different positions.

[0063] Figure 8 This is the classification confusion matrix diagram of the model output under the X-direction vibration data of the planetary gearbox in the working condition of 20Hz-0V.

[0064] Fig. 9 This is a T-SNE visualization diagram of the model output layer under the X-direction vibration data of the planetary gearbox in the working condition of 20Hz-0V.

[0065] Fig.10 Schematic diagram of the target domain fine-tuning sample expansion method of the present invention.

[0066] Fig.11 This is a comparison curve of the model validation set loss under different random noise lengths of the present invention.

[0067] Fig.12 This is a comparison curve of the accuracy of the model validation set under different random noise lengths of the present invention.

[0068] Fig.13 This is the comparison curve of the validation set loss under different migration strategies of the model of the present invention.

[0069] Fig.14 This is the comparison curve of the validation set accuracy of the model under different migration strategies.

[0070] Fig.15 This is a bar graph of the diagnostic accuracy of the present invention when the working condition 20Hz-0V is migrated to the working condition 30Hz-2V, and when the working condition 30Hz-2V is migrated to the working condition 20Hz-0V.

[0071] Fig.16 The model classification confusion matrix diagram is a diagram of the present invention when the working condition of 20Hz-0V is migrated to the working condition of 30Hz-2V.

[0072] Fig.17 This is the T-SNE visualization diagram of the model output layer when the working condition of 20Hz-0V is migrated to the working condition of 30Hz-2V. DETAILED DESCRIPTION

[0073] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0074] The embodiment of the present invention provides a planetary gearbox fault diagnosis method based on deep learning and transfer learning, such as Figure 1 As shown, including:

[0075] Step 1: Collect vibration acceleration signals of the planetary gearbox in different states under different working conditions;

[0076] Step 2: Divide the vibration acceleration signal into source domain data and target domain data;

[0077] Step 3: Build a deep learning model of CNN-LSTM integrating CAM and Self-Attention;

[0078] Step 4: Use source domain data to debug and optimize the deep learning model and determine the optimal pre-training model;

[0079] Step 5: Expand and fine-tune samples;

[0080] Step 6: Design and compare fine-tuning migration plans to determine the optimal fine-tuning migration plan;

[0081] Step 7: Use the optimal fine-tuning migration solution to fine-tune the migration model according to the expanded fine-tuning samples;

[0082] Step 8: Use the fine-tuned migration model to perform fault diagnosis on the planetary gearbox according to the target domain data.

[0083] In step 1, the vibration acceleration signal is collected through an acceleration sensor;

[0084] The vibration acceleration signal is the signal of the vibration acceleration of the planetary gearbox in the X direction, Y direction, and Z direction.

[0085] The deep learning model constructed in step three includes 1 input layer B1, 4 convolutional layers C1, C2, C3, C4, 4 pooling layers D1, D2, D3, D4, 1 channel attention mechanism layer E1, 1 long short-term memory network layer F1, 1 self-attention mechanism layer G1, 1 flattening layer H1, 2 fully connected layers I1, I2, 3 dropout layers J1, J2, J3 and 1 output classification layer K1; the order of each layer is: B1, C1, D1, C2, D2, E1, F1, C3, D3, C4, D4, G1, H1, J1, I1, J2, I2, J3, K1.

[0086] In step 4, before debugging and optimizing the deep learning model using the source domain data, 1024 consecutive data points are continuously intercepted from the source domain data as a sample, with the overlap between two adjacent samples being 0. 900 samples are taken for each state and divided into source domain training set, source domain verification set, and source domain test set in a ratio of 7:1:1.

[0087] In step 4, before debugging and optimizing the deep learning model using source domain data, the optimal pre-training model is determined based on the model evaluation system, according to the samples intercepted by the deep learning model in different source domain data and the model loss accuracy curves under different optimizers.

[0088] When expanding the fine-tuning samples in step 5, a random noise sequence of a certain length is generated, and a random position is generated on the original fine-tuning sample. The generated random noise sequence is used to replace some data points of the original fine-tuning sample to expand the number of fine-tuning samples; the length of the random noise sequence is α, and the value of α includes at least: 128, 256, 512. The replacement is repeated 9 times to obtain 500 expanded fine-tuning samples of each type.

[0089] The fine-tuning migration solution includes:

[0090] Solution 1: Migrate the pre-trained model to the target domain task, expand the fine-tuning samples, and fine-tune the entire migration model;

[0091] Solution 2: Execute Solution 1 and freeze the migration model to the shallow CNN (D2).

[0092] Solution 3: Execute Solution 1 and migrate the frozen model to CAM;

[0093] Solution 4: Execute Solution 1 and migrate the model to LSTM.

[0094] Solution 5: Execute Solution 1 and migrate the model to the deep CNN (D4).

[0095] Solution 6: Execute Solution 1 and freeze the migration model to Self-Attention.

[0096] Solution 7: Execute Solution 1 and freeze the entire model (except the classification layer).

[0097] Solution 8: The model is not fine-tuned and transferred, and the target domain samples are used to directly train the model;

[0098] Solution 9: Execute Solution 1, migrate the pre-trained model to the target domain task, and fine-tune the entire migration model using target domain samples.

[0099] In step eight, the fine-tuned migration model is used to perform fault diagnosis on the planetary gearbox according to the target domain data. 1024 consecutive data points are continuously intercepted from the target domain data as a sample, and the overlap between two adjacent samples is 0. 250 samples are taken from each state in the target domain and divided into the corresponding target domain training set, validation set, and test set in a ratio of 1:2:2.

[0100] The working principle and beneficial effects of the above technical solution are:

[0101] Step 1: Use an acceleration sensor to collect vibration acceleration signals of the planetary gearbox in different states (different states refer to the 4 types of bearing faults, 4 types of gear faults and 2 types of normal states introduced below) under different working conditions (such as speed-load configuration of 1200rpm-0Nm, i.e. 20Hz-0V; speed-load configuration of 1800rpm-7.32Nm, i.e. 30Hz-2V). The vibration acceleration signals cover 4 types of bearing faults (i.e. inner ring fault, outer ring fault, rolling element fault, inner and outer ring mixed fault), 4 types of gear faults (i.e. gear defect, broken tooth fault, gear root crack, tooth surface wear) and 2 types of normal states (i.e. normal bearing, normal gear).

[0102] Step 2: divide the vibration acceleration signal obtained in step 1 into source domain data with sufficient data volume and target domain data with less data volume; specifically, the technical personnel can set the division rules in advance, or the data with more data can be divided into source domain data and the data with less data volume can be divided into target domain data.

[0103] Step 3: Divide the source domain data and the target domain data according to a certain ratio to obtain the corresponding training set, validation set, and test set.

[0104] Among them, in the source domain, the training set contains rich annotated data, which is used for model initialization training so that the model can learn the features and patterns in the source domain; the validation set is used to tune the model's hyperparameters and select the optimal model, and does not participate in the model's training process; the test set is used to evaluate the performance of the model in the source domain and help understand the model's generalization ability.

[0105] In the target domain, the training set is used for adaptive training of the model in the target domain. Usually, there is less data, and the goal is to apply the knowledge learned in the source domain to the target domain through methods such as transfer learning. The validation set is used to tune the performance of the model in the target domain and help judge the effect of the source domain knowledge transfer. The test set is used to evaluate the final performance of the model in the target domain and verify the robustness of the model under different working conditions.

[0106] Step 4: Construct a model structure that combines CNN (convolutional neural network)-LSTM (long short-term memory network) with CAM (channel attention mechanism) and Self-Attention (self-attention mechanism).

[0107] Step 5: Fully train and debug the model built in step 4 in the source domain (referring to the dataset and feature space on which the model was initially trained, which usually contains rich annotated data and can provide enough information to help the model learn effective feature representations and decision boundaries). After training and debugging, save the pre-trained model.

[0108] Step 6: Migrate the pre-trained model to the diagnostic task in the target domain (the target domain is the new data set and feature space to which the model is applied) (the diagnostic task in the target domain refers to the task of identifying, classifying or locating faults in the target domain in cross-operating condition fault diagnosis).

[0109] Step 7: Generate a random noise sequence of a certain length, generate a random position on the original fine-tuning sample (referring to the training set on the target domain), and use the generated random noise sequence to replace some data points of the original fine-tuning sample to expand the number of fine-tuning samples.

[0110] Step 8. Design, compare and fine-tune the migration plan.

[0111] Step 9: Use the fine-tuning samples expanded in step 7 to fine-tune the migration model (the target domain training set is adjusted to the pre-trained model after source domain data debugging and optimization). After fine-tuning, save the model and call the target domain test set to complete the cross-operating fault diagnosis task of the planetary gearbox.

[0112] In step 1, the vibration acceleration signal is the vibration acceleration signal of the planetary gearbox in the X direction, Y direction, and Z direction.

[0113] In step 3, the sample is intercepted as follows Figure 2 As shown. 1024 continuous data points are continuously intercepted in the source domain and the target domain as a sample, and the overlap between two adjacent samples is 0. 900 samples are taken for each state type (inner ring fault, outer ring fault, rolling element fault, inner and outer ring mixed fault, gear defect, broken tooth fault, gear root crack, tooth surface wear and normal bearing, normal gear) in the source domain, and 9000 samples are taken for 10 states, and they are divided into the corresponding source domain training set, source domain validation set and source domain test set with a ratio of 7:1:1; 250 samples are taken for each state type (inner ring fault, outer ring fault, rolling element fault, inner and outer ring mixed fault, gear defect, broken tooth fault, gear root crack, tooth surface wear and normal bearing, normal gear) in the target domain, and 2500 samples are taken for 10 states, and they are divided into the corresponding target domain training set, validation set and test set with a ratio of 1:2:2. The labels of all source domain data samples and part of the target domain data (target domain training set and validation set) are processed, that is, normal bearing, inner ring fault, outer ring fault, rolling element fault, inner and outer ring mixed fault, normal gear, gear defect, broken tooth fault, gear root crack, tooth surface wear, a total of 10 types of status corresponding to the labels are 0, 1, 2, 3, 4, 5, 6, 7, 8, 9; the meaning of the numbers in brackets is that different fault types correspond to different labels. The details of sample division are shown in Table 1.

[0114] Table 1. Division of source domain and target domain samples

[0115]

[0116] In step 4, the model structure of CNN-LSTM fusion CAM and Self-Attention is constructed. The model includes 1 input layer (B1), 4 convolutional layers (C1, C2, C3, C4), 4 pooling layers (D1, D2, D3, D4), 1 channel attention mechanism layer (E1), 1 long short-term memory network layer (F1), 1 self-attention mechanism layer (G1), 1 flattening layer (H1), 2 fully connected layers (I1, I2), 3 drop layers (J1, J2, J3) and 1 output classification layer (K1). The order of each layer is: "B1→C1→D1→C2→D2→E1→F1→C3→D3→C4→D4→G1→H1→J1→I1→J2→I2→J3→K1". The model network parameters are shown in Table 2.

[0117] Table 2 Model network parameters

[0118]

[0119] In step 5, the source domain data is used to debug and optimize the model built in step 4. The number of training samples (i.e., source domain sample training samples) (i.e., the number of training samples of each type is set to 100, 300, 500, and 700); the optimizer types are Adam optimizer (Ir = 0.01, 0.001, 0.0001), SGD optimizer (Ir = 0.01, 0.001, 0.0001), and adaptive optimizer Adadelta. The influence on the model determines the number of training samples and the type of optimizer.

[0120] The model loss accuracy curve under different numbers of training samples is as follows Figure 3 , Figure 4 As shown in the figure. The construction process of the model loss-precision curve includes: recording the loss and precision values ​​after each training cycle (epoch), taking the number of iterations as the horizontal axis, and constructing the curve with the loss and precision values ​​of the model after each iteration as the vertical axis, and specifically constructing it based on the source domain validation set. As the number of training samples increases, the convergence of the model loss-precision curve accelerates and the fluctuation decreases. When the number of training samples in each category is 700, the loss curve tends to 0 and the precision curve tends to 1. Considering that too many training samples will increase the risk of overfitting of the model to a certain extent, when achieving better diagnostic performance, the number of training samples will no longer increase, that is, the training set is composed of 700 samples of each category.

[0121] The model loss accuracy curves under different optimizers are as follows Figure 5 , Figure 6 As shown in the figure. When the Adam optimizer is at Ir = 0.001 and the SGD optimizer is at Ir = 0.01, the model loss accuracy curve is relatively smooth. However, under the Adadelta optimizer, the model converges after about 12 iterations, the convergence speed is relatively fast, and the overall fluctuation is small during the iteration process. This is because the Adadelta optimizer adaptively adjusts the learning rate during the training process to avoid the problem that the model cannot effectively learn the feature information related to the fault due to the learning rate being set too high; the model falls into the local optimal solution when the learning rate is set too low.

[0122] The pre-trained model was tested under different sensor vibration acceleration directions, and 10 verifications were performed in each direction. The average value of the 10 experiments was taken as the final result. Figure 7 As shown in the figure, the classification confusion matrix and model output layer T-SNE visualization of a certain experiment are as follows: Figure 8 , Fig. 9 shown.

[0123] Analysis of diagnostic accuracy bar chart: The diagnostic accuracy of the model for 1000 test samples (the entire source domain test set) under the working condition of 20Hz-0V planetary gearbox X-direction data and 30-2 planetary gearbox X and Z-direction data reached 100%, and the lowest accuracy at other positions was 99.87%, and the overall average diagnostic accuracy was 99.96%. Analysis of classification confusion matrix and model output layer T-SNE visualization: The model misjudged 2 samples out of 1000 test samples under the working condition of 20Hz-0V planetary gearbox X-direction data. Under the SUFD data set, the various states of the model output layer are clustered in strips, and there is almost no aliasing between states, indicating that the proposed fault diagnosis model has a strong learning ability and can accurately identify the operating status of the planetary gearbox.

[0124] In one embodiment of the present invention, in step seven, the sample expansion method schematic diagram is as follows: Fig.10 As shown. Replace some data points of the fine-tuning samples with random noise sequence lengths of 128, 256, and 521, and repeat the replacement 9 times (generate a certain length of random noise to replace some data points of the original fine-tuning samples. Repeated replacement means repeating the process nine times), and get 500 fine-tuning samples of each category after expansion, and a total of 5000 samples for 10 categories of states. Under different random noise lengths, the model's cross-condition fault diagnosis loss accuracy curve on the validation set is shown as follows Fig.11 , Fig.12 shown.

[0125] Analysis of the model loss accuracy curve under different random noise data: When the random noise length is 256, compared with 128 and 512 random noise data points, the loss accuracy curve is more stable and smooth, the loss tends to be lower, and the accuracy tends to be higher. The reason is that when the random noise length is too short, the diversity of the expanded fine-tuning samples is insufficient; when the random noise length is too long, the correlation between the expanded fine-tuning samples and the original samples is reduced. Therefore, the random noise data length is set to 256.

[0126] In one embodiment of the present invention, in step eight, the migration fine-tuning strategy is specifically as follows:

[0127] 1) Solution 1: After fully training the source domain samples (source domain training set), the pre-trained model is obtained, the pre-trained model is transferred to the target domain task, the fine-tuning samples are expanded and the transfer model is fine-tuned as a whole (the pre-trained model transferred to the target domain diagnosis task is fine-tuned using the target domain training set);

[0128] 2) Scheme 2: The previous steps are the same as those in Scheme 1, and the migration model is frozen to the shallow CNN (D2);

[0129] 3) Scheme 3: The previous steps are the same as those in Scheme 1, and the migration model is frozen to CAM;

[0130] 4) Solution 4: The previous steps are the same as Solution 1, and the migration model is frozen to LSTM;

[0131] 5) Scheme 5: The previous steps are the same as those in Scheme 1, and the migration model is frozen to the deep CNN (D4);

[0132] 6) Solution 6: The previous steps are the same as Solution 1, and the migration model is frozen to Self-Attention;

[0133] 7) Scheme 7: The previous steps are the same as those in Scheme 1, freezing the entire model (K1);

[0134] 8) Solution 8: The model is not fine-tuned and transferred, and the target domain samples are used to directly train the model;

[0135] 9) Solution 9: The previous steps are the same as Solution 1. The pre-trained model is migrated to the target domain task, and the target domain samples are used to fine-tune the migration model as a whole.

[0136] Under different schemes, the model cross-operating condition fault diagnosis loss accuracy curve is as follows: Fig.13 , Fig.14 The loss accuracy statistics are shown in Table 3.

[0137] Table 3 Statistics of model loss accuracy under different migration strategies

[0138]

[0139] Analysis of the loss accuracy curve of model cross-condition fault diagnosis under different fine-tuning migration schemes: The model loss accuracy curve under Scheme 1 is better than that of other schemes, which verifies the effectiveness of the proposed method. Analysis of loss accuracy statistics: Taking average accuracy as an indicator for analysis, the average accuracy of Scheme 1 is 99.40%, which is 7.09% and 1.78% higher than that of no model migration and fine-tuning (Scheme 8) and no fine-tuning sample expansion after migration (Scheme 9); Taking average loss as an indicator for analysis, the average loss of Scheme 1 is 1.75%, which is 18.47% and 8.05% higher than that of Schemes 8 and 9, respectively, which verifies the effectiveness of model migration and fine-tuning sample expansion. Moreover, compared with the other migration model freezing schemes (Schemes 2-7), the loss values ​​of Scheme 1 are reduced to varying degrees, and the accuracy values ​​are improved to varying degrees, indicating that the fine-tuning model as a whole is more applicable in the proposed fine-tuning migration model.

[0140] In one embodiment of the present invention, in step nine, the fine-tuning sample expanded in step six is ​​used to fine-tune the migration model. After analysis in step eight, the optimal fine-tuning migration strategy is determined as scheme one. Comparison of fault diagnosis accuracy of the model across working conditions in scheme one Fig.15 As shown, the T-SNE visualization of the classification matrix and model output of a certain experiment is as follows Fig.16 , Fig.17shown.

[0141] Analysis of the cross-condition fault diagnosis accuracy comparison bar chart: The diagnostic accuracy of 20Hz-0V to 30Hz-2V migration is 99.40%, and the diagnostic accuracy of 30Hz-2V to 20Hz-0V migration is 99.08%. The overall average diagnostic accuracy across working conditions is 99.24%. Analysis of the output classification confusion matrix and T-SNE visualization: During the migration from 20Hz-0V to 30Hz-2V. Under 1000 test samples, the model only misclassified 5 samples, with a diagnostic accuracy of 99.50%, and the various states were clearly clustered with less aliasing.

[0142] In one embodiment, the planetary gearbox fault diagnosis method based on deep learning and transfer learning further includes:

[0143] Step nine, obtaining a fault diagnosis history of a planetary gearbox fault diagnosis;

[0144] In step nine, the fault diagnosis history is the diagnosis history record generated when a large number of different planetary gearboxes are diagnosed by using the fine-tuned migration model according to the target domain data, and at least includes the diagnosed fault type, diagnosis time, etc.;

[0145] Step 10: Necessarily expand the fault diagnosis history;

[0146] In step 10, after necessary expansion, the content of the fault diagnosis history will be richer, which is more conducive to generating a digital model to interactively determine the diagnosis optimization plan with the user;

[0147] Step 11: Generate a digital model based on the fault diagnosis history expanded by necessity;

[0148] In step 11, when generating the digital model, the data type of the fault diagnosis history expanded according to necessity is represented by a preset corresponding digital template. For example, if the data type is the diagnosis fault type, the corresponding digital model is a visual table of the fault type. Then, the results after various representations are integrated to obtain the digital model.

[0149] Step 12: Based on the digital model, interactively determine the diagnostic optimization plan with the user;

[0150] In step 12, when interactively determining the diagnosis optimization plan with the user based on the digital model, the digital model is first displayed to the user. The user will decide how to perform the diagnosis optimization in the process of viewing the displayed digital model, and feedback the diagnosis optimization plan. For example, the user finds that some fault types of the planetary gearbox are associated faults (one fault will induce another fault) by viewing the digital model. After making a decision, the feedback diagnosis optimization plan is to automatically provide a warning reminder of the corresponding associated fault when the fault type is diagnosed;

[0151] Step 13: Execute the diagnostic optimization plan.

[0152] In step thirteen, after the user provides feedback on the diagnostic optimization plan, the system executes it to achieve further performance improvement of the system's planetary gearbox fault diagnosis.

[0153] In one embodiment, step ten, necessarily expanding the fault diagnosis history, includes:

[0154] S1. Represent the fault diagnosis history in a time axis to obtain a history time axis; wherein different history items are set at different time nodes on the history time axis, and the generation time of the history item corresponds to the time node at which it is located;

[0155] In step S1, there are multiple time nodes on the time axis, and there are multiple history items in the fault diagnosis history. Each history item is a historical record. When the time axis is represented, the history item is set to the corresponding time node according to the generation time of the history item.

[0156] S2. Determine a historical item combination on the historical time axis; wherein the on-axis spacing between adjacent historical items in the historical item combination does not exceed a spacing threshold and has a preset standard historical item association relationship; and the total number of historical items in the same historical item combination exceeds a number threshold;

[0157] In step S2, the on-axis spacing is the interval between adjacent historical items on the historical time axis, which represents the interval between the two times of generation. The spacing threshold can be 100 seconds to ensure that the spacing is not too large and constrain the historical items to be generated in the same stage. The preset standard historical item association relationship includes: the fault type in the two historical items is the same, the fault cause in the two historical items is the same, the fault impact range in the two historical items is the same, etc. When the two historical items have the preset standard historical item association relationship, it can be further explained that necessary expansion may be required. For example: the fault type in the two historical items is the same, and the fault cause in the two historical items is the same, then batches of the same fault type will be collaboratively solved in history, and relevant records of the collaborative solution need to be supplemented, such as solution experience, solution process information, etc.; the number threshold can be 4; ensuring that the on-axis spacing between adjacent historical items in the historical item combination does not exceed the spacing threshold and has a preset standard historical item association relationship, and the total number of historical items in the same historical item combination exceeds the number threshold, can make the determined historical item combination can fully and individually reflect the diagnosis situation that may need to be necessary to expand in a continuous fault diagnosis stage;

[0158] S3. Based on the combination selection rule, select the combination of historical items in sequence on the historical time axis;

[0159] S4. Each time a selection is made, the selected historical item combination is characterized to obtain a characteristic entity;

[0160] In step S4, when performing the characterization process, firstly, the multimodal features of the selected historical item combination are extracted, and the multimodal features include: content features of the historical items, associations between different content items (for example, the same fault cause), relationship types and corresponding relationship numbers of associations between different content items, and average generation time intervals between historical items, etc.; then, the multimodal features are represented as feature entities in the form of vectors;

[0161] S5. Matching the feature entity with the index feature entity;

[0162] In step S5, the index feature entity is pre-represented in the form of a vector by a multimodal feature that represents the need for necessity expansion in the selected historical item combination. The multimodal feature that represents the need for necessity expansion in the selected historical item combination may be, for example: the average generation time interval between historical items is 10 seconds, and the fault types and fault causes in most historical items are the same, which means that the same fault has been frequently diagnosed in the planetary gearbox in history. In the history, the planetary gearboxes with the same fault must have been batch-coordinated to solve the problem, and it is necessary to obtain information such as solving experience, which means that the selected historical item combination needs to be expanded;

[0163] S6. When the match is met, based on the situation analysis template corresponding to the index feature entity that matches the match, situation analysis is performed on the selected historical item combination on the historical time axis and other historical item combinations within a preset axis distance before and after the selected historical item combination to obtain the historical situation;

[0164] In step S6, the index feature entity is preset with a corresponding situation analysis template, and the situation analysis template is a template for analyzing multiple historical item combinations to determine historical situations; when it is determined that the selected historical item combination needs to be expanded by necessity, the historical situation is determined to facilitate accurate determination of how to expand the selected historical item combination by necessity; the preset axis distance range can be within 5 minutes; when performing situation analysis, not only the selected historical item combination is analyzed, but also other historical item combinations within the preset axis distance before and after the selected historical item combination are analyzed, and the historical situation is comprehensively determined to more comprehensively determine how the selected historical item combination needs to be expanded by necessity; situation feature extraction rules will be set on the situation analysis template, such as extracting the diagnosis frequency of different types of diagnostic faults, extracting the total number of fault diagnoses, the total number of historical item combinations analyzed by the situation, etc., extracting situation features based on situation feature extraction rules, and forming historical situations based on different situation features;

[0165] S7, matching historical situations with standard situations;

[0166] In step S7, the standard situation is a situation where the preset representative historical item combination needs to be necessarily expanded;

[0167] S8. When the match is met, the selected historical item combination is expanded based on the historical expansion template corresponding to the standard case of the match;

[0168] In step S8, different standard situations are preset with corresponding historical expansion templates. For example, if the situation characteristics forming the standard situation are that the diagnosis frequency of a certain type of diagnostic fault is 100 Hz, the total number of fault diagnoses is 20 times, and the total number of historical item combinations analyzed by the situation is 3, then the standard situation indicates that the type of diagnostic fault occurs frequently and appears in multiple historical item combinations, etc. The corresponding historical expansion template is a processing record of expanding the historical item combination to process the type of diagnostic fault, such as processing time, processing experience information, etc.; when the match is met, the selected historical item combination is expanded based on the historical expansion template corresponding to the standard situation that matches the match;

[0169] S9, when all historical item combinations are selected, determining an expanded fault diagnosis history based on the current historical time axis;

[0170] In step S9, when all historical item combinations are selected, some historical item combinations on the current historical time axis have been expanded, so the contents of all historical item combinations on the current historical time axis can be integrated to obtain an expanded fault diagnosis history;

[0171] The combination selection rules include:

[0172] When making the first selection, select the first historical item combination on the historical timeline;

[0173] When making the nth selection, if there is a historical situation that matches the standard situation in the previous n-1 selections, then the last one of the historical item combination items involving the last historical situation that matches the standard situation in the previous n-1 selections is selected; if it does not exist, then the next one of the historical item combination selected for the n-1th selection is selected; n>1.

[0174] In the combination selection rule, when the selection starts, the earliest historical item combination is selected, that is, the historical item combination at the front of the historical timeline; when the selection is made subsequently, if the historical item combination item has been used (involved) in determining the historical situation that matches the standard situation, it means that it is used to jointly reflect the situation that requires the expansion of the historical item combination, and no further selection is required. Then, the last one of the historical item combination items involved in the historical situation that matches the standard situation in the previous n-1 selections is selected; if no historical item combination item has been used to determine the historical situation that matches the standard situation, then the next one of the historical item combinations selected for the n-1th time is selected.

[0175] When the embodiment of the present invention performs necessity expansion on the fault diagnosis history, a combination of historical items on the historical timeline is introduced, and based on the combination selection rule, the combination of historical items is selected in sequence on the historical timeline, the feature entity is determined, and the feature entity is matched with the index feature entity. When the match is met, the historical situation is determined based on the situation analysis template corresponding to the matched index feature entity, and the historical situation is matched with the standard situation. When the match is met, the selected combination of historical items is expanded based on the history expansion template corresponding to the matched standard situation, which greatly improves the comprehensiveness, accuracy and applicability of the necessity expansion of the fault diagnosis history; the introduction of the combination selection rule greatly improves the rationality of the selection of the combination of historical items, improves the working efficiency of the system, and is also more intelligent.

[0176] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A planetary gearbox fault diagnosis method based on deep learning and transfer learning, characterized in that: include: Step 1: Collect vibration acceleration signals of the planetary gearbox in different states under different working conditions; Step 2: Divide the vibration acceleration signal into source domain data and target domain data; Step 3: Build a deep learning model of CNN-LSTM integrating CAM and Self-Attention; Step 4: Use source domain data to debug and optimize the deep learning model and determine the optimal pre-training model; Step 5: Expand and fine-tune samples; Step 6: Design and compare fine-tuning migration plans to determine the optimal fine-tuning migration plan; Step 7: Use the optimal fine-tuning migration solution to fine-tune the migration model according to the expanded fine-tuning samples; Step 8: Use the fine-tuned migration model to perform fault diagnosis on the planetary gearbox according to the target domain data.

2. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: In step 1, the vibration acceleration signal is collected through an acceleration sensor; The vibration acceleration signal is the signal of the vibration acceleration of the planetary gearbox in the X direction, Y direction, and Z direction.

3. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: The deep learning model constructed in step three includes 1 input layer B1, 4 convolutional layers C1, C2, C3, C4, 4 pooling layers D1, D2, D3, D4, 1 channel attention mechanism layer E1, 1 long short-term memory network layer F1, 1 self-attention mechanism layer G1, 1 flattening layer H1, 2 fully connected layers I1, I2, 3 dropout layers J1, J2, J3 and 1 output classification layer K1; the order of each layer is: B1, C1, D1, C2, D2, E1, F1, C3, D3, C4, D4, G1, H1, J1, I1, J2, I2, J3, K1.

4. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: In step 4, before debugging and optimizing the deep learning model using the source domain data, 1024 consecutive data points are continuously intercepted from the source domain data as a sample, with the overlap between two adjacent samples being 0. 900 samples are taken for each state and divided into source domain training set, source domain verification set, and source domain test set in a ratio of 7:1:

1.

5. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 4, characterized in that: In step 4, before debugging and optimizing the deep learning model using source domain data, the optimal pre-training model is determined based on the model evaluation system, according to the samples intercepted by the deep learning model in different source domain data and the model loss accuracy curves under different optimizers.

6. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: When expanding the fine-tuning samples in step 5, a random noise sequence of a certain length is generated, and a random position is generated on the original fine-tuning sample, and some data points of the original fine-tuning sample are replaced with the generated random noise sequence to expand the number of fine-tuning samples; The length of the random noise sequence is α, and the value of α includes at least 128, 256, and 512. The replacement is repeated 9 times to obtain 500 expanded fine-tuning samples of each category.

7. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: The fine-tuning migration solution includes: Solution 1: Migrate the pre-trained model to the target domain task, expand the fine-tuning samples, and fine-tune the entire migration model; Solution 2: Execute Solution 1 and freeze the migration model to the shallow CNN (D2). Solution 3: Execute Solution 1 and migrate the frozen model to CAM; Solution 4: Execute Solution 1 and migrate the model to LSTM. Solution 5: Execute Solution 1 and migrate the model to the deep CNN (D4). Solution 6: Execute Solution 1 and freeze the migration model to Self-Attention. Solution 7: Execute Solution 1 and freeze the entire model (except the classification layer). Solution 8: The model is not fine-tuned and transferred, and the target domain samples are used to directly train the model; Solution 9: Execute Solution 1, migrate the pre-trained model to the target domain task, and fine-tune the entire migration model using target domain samples.

8. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: In step eight, the fine-tuned migration model is used to perform fault diagnosis on the planetary gearbox according to the target domain data. 1024 consecutive data points are continuously intercepted from the target domain data as a sample, and the overlap between two adjacent samples is 0. 250 samples are taken from each state in the target domain and divided into the corresponding target domain training set, validation set, and test set in a ratio of 1:2:

2.

9. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: Also includes: Obtaining a fault diagnosis history for troubleshooting a planetary gearbox; Necessary expansion of fault diagnosis history; Generate a digital model based on the fault diagnosis history expanded by necessity; Based on the digital model, interactively determine the diagnostic optimization plan with the user; Execute diagnostic optimization program.

10. The planetary gearbox fault diagnosis method based on deep learning and transfer learning according to claim 1, characterized in that: Necessary expansion of fault diagnosis history, including: The fault diagnosis history is represented by a time axis to obtain a historical time axis; wherein different historical items are set at different time nodes on the historical time axis, and the generation time of the historical item corresponds to the time node at which it is located; Determine a historical item combination from the historical time axis; wherein the on-axis spacing between adjacent historical items in the historical item combination does not exceed a spacing threshold and has a preset standard historical item association relationship; and the total number of historical items in the same historical item combination exceeds a number threshold; Based on the combination selection rule, historical item combinations are selected in sequence on the historical time axis; Each time a selection is made, the selected historical item combination is characterized to obtain a characteristic entity; Matching the feature entity with the index feature entity; When the match is met, based on the situation analysis template corresponding to the index feature entity that matches the match, the situation analysis is performed on the historical item combination selected on the historical time axis and other historical item combinations within the preset axis distance before and after the selected historical item combination to obtain the historical situation; Match historical situations with standard situations; When the match is met, the selected historical item combination is expanded based on the historical expansion template corresponding to the standard case of the match; When all combinations of history items are selected, the expanded fault diagnosis history is determined based on the current history time axis.

Images