Planetary gearbox intelligent fault diagnosis method based on ITD and LSTM-CNN

Abstract

The invention provides a planetary gearbox intelligent fault diagnosis method based on ITD and LSTM-CNN, and relates to the technical field of planetary gearbox fault diagnosis. The planetary gear box intelligent fault diagnosis method based on the I TD and the LSTM-CNN comprises the following steps that firstly, vibration signals of a planetary gear box are collected through a sensor, I TD decomposition is carried out on each sampling signal to obtain a plurality of PR components, the PR components with large instantaneous frequency fluctuation are selected through Hilbert transform, and the PR components with large instantaneous frequency fluctuation are obtained; converting the reconstructed signal into a frequency spectrum function through Fourier transform, extracting the amplitude of a feature frequency, constructing a feature vector, inputting the feature vector into an LSTM-CNN for intelligent diagnosis, and combining a classification cross entropy loss function to ensure the stability of calculation and the high efficiency of training. A comprehensive and efficient solution is provided for fault type and degree diagnosis of the sun gear of the gear transmission system, and compared with a classic neural network model, the diagnosis accuracy of different types of gear faults and different degrees of faults is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of planetary gearbox fault diagnosis, and specifically to an intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN. Background Art

[0002] Planetary gearboxes (PGs) have excellent transmission efficiency and load-bearing capacity and are widely used in fields such as wind power generation, industrial equipment, and aerospace. However, under long-term harsh working environments and complex working conditions, the key components of planetary gearboxes are prone to wear and fatigue, posing potential failure risks. If these risks are not diagnosed in a timely manner, it may lead to equipment downtime, causing serious economic losses and even potential safety hazards.

[0003] Time-frequency analysis methods, such as WT, EMD, ITD, etc., are widely used in fault feature analysis and can effectively extract fault features. However, methods such as ITD rely too much on manual interpretation and expert experience. With the rise of deep learning, models such as LSTM and CNN have shown good performance in fault diagnosis, capable of automatically extracting features and classifying, reducing the dependence on human experience. However, research based on LSTM has deficiencies in capturing the spatial features of complex signals, and the accuracy of CNN-based methods fluctuates and the computational efficiency is low when the training data is complex. Therefore, it is of great significance to design an intelligent diagnosis method that can effectively identify both the fault types and severity of planetary gearboxes. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides an intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN, solving the problem of defects in the fault diagnosis of planetary gearboxes in the prior art.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN, comprising the following steps:

[0006] S1. Obtain several different types of vibration signals of the sun gear of the test bench planetary gearbox through vibration sensors;

[0007] S2. Decompose each sampled signal into PR components according to ITD;

[0008] S3. Select the PR components with relatively large instantaneous frequency fluctuations through Hilbert transform to reconstruct the signal;

[0009] S4. Transform the reconstructed signal into a frequency spectrum through Fourier transform, then extract the amplitudes of the characteristic frequencies, construct a feature vector, and establish a data set;

[0010] S5. Input the dataset into the LSTM-CNN model for training, combine it with the categorical cross-entropy loss function, and use the Adam optimizer and ReduceLROnPlateau learning rate scheduler to enhance the training process.

[0011] Preferably, step S1 further includes the following steps:

[0012] S1.1. Classify the health status of the sun gear of the planetary gearbox of the dynamic transmission chain simulator into the following categories: normal, broken tooth, missing tooth, wear; classify the health status of the sun gear of the planetary gearbox of the rotary fault simulation test bench into the following categories: normal, pitting, tooth surface wear, slight crack, moderate crack, severe crack.

[0013] S1.2. Set the speed of the DDS motor to 1800 RPM, the sampling time to 5 seconds, and the sampling frequency to 7680 Hz. Collect 50 groups of vibration data for each type of gear fault, with each group of vibration signals containing 38400 data points. Set the speed of the RFS motor to 3000 RPM, the sampling time to 5 seconds, and the sampling frequency to 10240 Hz. Collect 50 groups of vibration data for each type of gear fault, with each group of vibration signals containing 51200 data points.

[0014] S1.3. Collect four different types of vibration signals X = [X1, X2, … X n of the sun gear of the planetary gearbox, where X is the vibration signal dataset and Xi is the vibration signal.

[0015] Preferably, step S2 further includes the following steps:

[0016] S2.1. Decompose the vibration signal according to ITD. Define L as the baseline extraction operator for extracting low-frequency signals. The decomposition expression of the original signal X t is:

[0017] X t = LX t + (1 - L)X t = L t + H t ,

[0018] where L t = LX t is the baseline signal and H t = (1 - L)X t is the PR component;

[0019] S2.2. Extract the extreme values of the signal X t , that is, X k , and the corresponding time is τ k , where k = 0, 1, 2…, τ0 = 0. Define a piecewise baseline extraction operator L as Let X k be X(τ k ), and L k be L(τ k ). Assume that L t and H t are defined in the interval [0, τ k , and X t is applicable to the interval [0, τ k+2 . Then in (τ k , τ k+1 , where 0 < ∝ < 1 and ∝ is 0.5;

[0020] S2.3. Define the PR extraction operator H as where represents the PR component with the highest frequency on the baseline, as the next input signal, repeat the above steps until the limit signal becomes a monotonic function or a constant, and finally decompose the original signal into multiple PR components with different frequencies and a residual signal, In the formula, p represents the total number of iterations, represents the i-th PR component, represents the residual signal.

[0021] Preferably, in S3, the PR component with a large instantaneous power fluctuation is selected through Hilbert transform, In the formula, H[z(t)] represents the value of the Hilbert transform signal at time t, and z(t) and Z(τ) represent the values of the original signal at time t and time τ respectively. The selected PR components are superimposed to reconstruct the signal.

[0022] Preferably, S4 further includes the following steps:

[0023] S4.1. Transform the reconstructed signal f(t) into the frequency spectrum function F(w) through Fourier transform, where w and t represent the frequency and time of the reconstructed signal respectively, and e -iwt represents the complex exponential function, which is e -iwt = cos(wt) - isin(wt);

[0024] S4.2. Extract the amplitude A(w) = |F(w)| of the characteristic frequency, construct the characteristic vector, and each characteristic vector represents a sample signal to construct a data set.

[0025] Preferably, S5 further includes the following steps:

[0026] S5.1. Construct an LSTM-CNN model, which includes 1 LSTM layer, 2 convolutional layers, and 2 pooling layers;

[0027] S5.2. Input the dataset into the LSTM-CNN model for training and summation testing to obtain the diagnostic accuracy. The softmax function is used to classify the fault types of the output layer of the LSTM-CNN, which is expressed as: where σ(x i ) represents the softmax output of the i-th component, and x i is an element of the input vector, represents the sum of the exponents of all input elements, ensuring that the output is normalized to a valid probability distribution;

[0028] S5.3. Combine it with the categorical cross-entropy loss function to ensure computational stability and training efficiency: where L represents the cross-entropy loss, n is the number of classes, y i is the true label of the i-th class, is the predicted probability of the i-th class;

[0029] S5.4. Use the Adam optimizer and the ReduceLROnPlateau learning rate scheduler to enhance the training process, ensure fast convergence and efficient handling of sparse gradients, and dynamically adjust the learning rate according to the validation results to prevent training stagnation and optimize its performance over multiple training epochs.

[0030] The present invention provides an intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN. It has the following beneficial effects:

[0031] 1. The present invention studies the problem of intelligent fault diagnosis of planetary gearboxes, and combines signal demodulation, feature extraction, and deep learning-based fault classification into a framework. The signals are analyzed using ITD transform, Hilbert transform, and Fourier transform, and classified using the LSTM-CNN model, providing a comprehensive and efficient solution for diagnosing the fault types and degrees of the sun gear in the gear transmission system.

[0032] 2. Compared with the LSTM-CNN and classical neural network models, the present invention improves the diagnostic accuracy for different types and degrees of gear faults in the case of small samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flowchart of the implementation of the present invention;

[0034] Figure 2 is an example diagram of vibration signals of different types of faults on the DDS used in the present invention;

[0035] Figure 3 Example diagram of vibration signals of different types of faults on the RFS used in the present invention;

[0036] Figure 4 In the present invention, the vibration signal is decomposed into 5 PR sub-diagrams using ITD;

[0037] Figure 5 Example diagram of the PR1, PR2, and PR3 components with relatively large instantaneous frequency fluctuations among the 5 selected PR components in the present invention;

[0038] Figure 6 The corresponding spectrogram obtained by performing Fourier transform on the reconstructed signal in the present invention;

[0039] Figure 7 Confusion matrix of four models in Experiment 1 of the present invention;

[0040] Figure 8 Comparison diagram of the diagnostic accuracy rates of four models in Experiment 1 of the present invention;

[0041] Figure 9 Confusion matrix of four models in Experiment 2 of the present invention;

[0042] Figure 10 Comparison diagram of the diagnostic accuracy rates of four models in Experiment 2 of the present invention. Detailed implementation manners

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0044] Embodiment:

[0045] As Figure 1 shown, the embodiment of the present invention provides an intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN, including the following steps:

[0046] S1. Obtain several different types of vibration signals of the sun gear of the test bench planetary gearbox through vibration sensors.

[0047] The present invention can be used for intelligent fault diagnosis of rotating machinery such as planetary gearboxes. In this embodiment, taking the planetary gearbox as an example, experimental analysis is carried out using vibration signal data of different types of the sun gear of the planetary gearbox. The vibration signals of the sun gear of the planetary gearbox are obtained through the installed sensors, and the vibration signal data of the planetary gearboxes on the two test benches are collected respectively under 4 and 6 working conditions, specifically as follows:

[0048] The vibration signal data used in this embodiment are obtained from the datasets of the Dynamic Drive Chain Simulator (DDS) in the Equipment Reliability Prediction and Health Management Laboratory of the University of Science and Technology of China and the Rotating Fault Simulation Test Bench (RFS) in the Key Laboratory of Advanced Mechatronic System Design and Intelligent Control, School of Mechanical Engineering, Tianjin University of Technology.

[0049] The sun gear of the planetary gearbox of the Dynamic Drive Chain Simulator (DDS) is divided into the following 4 health conditions: normal, tooth breakage, tooth missing, and wear.

[0050] The DDS test bench collects vibration signals at a motor speed of 1800 RPM, with a sampling time of 5 seconds and a sampling frequency of 7680 Hz. 50 sets of vibration data are collected for each type of gear fault, and each set of vibration signals contains 38400 data points. Examples of vibration signals for different types of faults on the DDS are as Figure 2 shown. The characteristic frequency selection of the DDS experimental platform is shown in Table 1.

[0051] Table 1 Characteristic Frequencies on the DDS Platform

[0052]

[0053] Among them, fc, fshaft, and fmesh represent the rotational frequencies of the planetary carrier, shaft, and gear meshing frequencies, respectively.

[0054] The sun gear of the planetary gearbox of the Rotating Fault Simulation Test Bench (RFS) is divided into the following 6 health conditions: normal, pitting, tooth surface wear, slight crack, moderate crack, and severe crack.

[0055] The RFS test bench collects vibration signals at a motor speed of 3000 RPM, with a sampling time of 5 seconds and a sampling frequency of 10240 Hz. 50 sets of vibration data are collected for each type of gear fault, and each set of vibration signals contains 51200 data points. Examples of vibration signals for different types of faults on the RFS platform are as Figure 3 shown. The characteristic frequency selection of the DDS experimental platform is shown in Table 2.

[0056] Table 2 Characteristic Frequencies on the RFS Platform

[0057]

[0058] Various different types of vibration signals X = [X1, X2, … X n of the sun gear of the planetary gearbox are collected, where X is the vibration signal dataset and Xi is the vibration signal;

[0059] S2. Each sampling signal is decomposed into PR components according to ITD.

[0060] S2.1. Decompose the vibration signal using ITD, extract multiple PR components, define L as the baseline extraction operator for extracting low-frequency signals, and the decomposition expression of the original signal Xt is X t = LX t +(1 - L)X t = L t + H t , where L t = LX t is the baseline signal, and H t =(1 - L)X t is the PR component;

[0061] S2.2. Extract the extreme values Xk of the signal Xt, and the corresponding time is τk, where k = 0, 1, 2…, generally τ0 = 0. Define a piecewise baseline extraction operator L as Let X k be X(τ k ), L k be L(τ k ), assuming that L t and H t are defined in the range of the interval [0, τ k , and X t is defined in the range of the interval [0, τ k+2 . In the continuous extreme value range (τ k , τ k+1 , set a piecewise limit extraction operator L: where 0 < ∝ < 1. Generally, ∝ is 0.5;

[0062] S2.3. Define the PR extraction operator H: where represents the highest frequency of the PR component. Repeat this process for the next input signal until the limit signal becomes a monotonic function or a constant. Finally, the original signal is decomposed into several PR components with different frequencies and a residual signal. In the formula, p represents the total number of iterations, represents the i-th PR component, represents the residual signal. Decompose the collected vibration signal samples according to the ITD method, and decompose each vibration signal into 5 PR components, as shown in Figure 4 .

[0063] S3. Select the PR component with relatively large instantaneous frequency fluctuations through Hilbert transform to reconstruct the signal.

[0064] Since the core fault characteristics of the original signal are mainly concentrated on the initial PR components, the PR components with larger instantaneous power fluctuations are selected through Hilbert transform, that is, the Hilbert transform is performed on the first 5 PR components. The corresponding instantaneous spectrum is obtained. In the formula, H[z(t)] represents the value of the Hilbert transform signal at time t, and z(t) and Z(τ) represent the values of the original signal at time t and time τ respectively. Select the PR1, PR2, and PR3 components with larger instantaneous frequency fluctuations as Figure 5 shown:

[0065] Superimpose them as sensitive components to obtain a reconstructed signal, and more sensitive fault information can be obtained from the reconstructed signal.

[0066] S4. Transform the reconstructed signal into a spectrum through Fourier transform, then extract the amplitude of the characteristic frequency, construct a feature vector, and establish a data set.

[0067] S4.1. Perform Fourier transform on the reconstructed signal to obtain the spectrum function F(w), where F(w) represents the Fourier transform of the time-domain signal f(t), w and t are the frequency and time of the reconstructed signal, and e -iwt represents the complex exponential function, which is e -iwt = cos(wt) - isin(wt), and e is the base of the natural logarithm. The corresponding spectrum is obtained after Fourier transform as Figure 6 shown;

[0068] S4.2. The signal characteristic frequencies fc, fshaft, and fmesh are extracted to construct a feature vector, and each feature vector represents a sample signal to construct a data set, where fc, fshaft, and fmesh represent the rotational frequencies of the planetary carrier, shaft, and gear meshing frequencies respectively.

[0069] S5. Input the data set into the LSTM-CNN model for training, combine it with the categorical cross-entropy loss function, and use the Adam optimizer and ReduceLROnPlateau learning rate scheduler to enhance the training process.

[0070] S5.1. Construct the LSTM-CNN model. The LSTM-CNN model combines long short-term memory (LSTM) layers and convolutional neural network (CNN) layers. The model contains 1 LSTM layer, 2 convolutional layers, and 2 pooling layers, and its parameters are shown in Table 3.

[0071] Table 3 LSTM-CNN model parameters

[0072]

[0073] The model structure starts with an LSTM layer, which processes sequential inputs to extract temporal features. These features are then passed through a series of two-dimensional convolutional and pooling layers to further extract and refine spatial information. The model also includes an unfolding layer and a fully connected layer to combine these extracted features and learn complex patterns.

[0074] S5.2. Input the dataset into the LSTM-CNN model for training to obtain the diagnostic accuracy rate. For classification, the output layer uses softmax activation to predict the classification probability, which can be expressed as: where σ(x i ) represents the softmax output of the i-th component, and x i is an element of the input vector, represents the sum of the exponents of all input elements, ensuring that the output is normalized to a valid probability distribution;

[0075] S5.3. The model is trained with categorical cross-entropy as the loss function to ensure computational stability and training efficiency: where L represents the cross-entropy loss, n is the number of classes, y i is the true label of the i-th class, is the predicted probability of the i-th class;

[0076] S5.4. Use the Adam optimizer and the ReduceLROnPlateau learning rate scheduler to enhance the training process, ensure fast convergence and efficient handling of sparse gradients, and dynamically adjust the learning rate according to the validation results to prevent training stagnation and optimize its performance over multiple training epochs.

[0077] The following combines two specific experiments to elaborate in detail on the technical effects of the present invention.

[0078] 1. Experiment 1

[0079] Experiment 1 uses the data of the Dynamic Drive Chain Simulator (DDS) in the Equipment Reliability, Prediction, and Health Management Laboratory of the University of Science and Technology of China to verify the effectiveness of the ITD and LSTM-CNN methods proposed in this paper, which is described in detail as follows:

[0080] (1) Use ITD to reconstruct the signal samples of the DDS experimental platform, construct a feature vector dataset, and divide it into a training set and a test set according to a ratio of 3:7. The sample division of the DDS dataset is shown in Table 4;

[0081] Table 4 Sample Division of the DDS Dataset

[0082]

[0083]

[0084] (2) Use the LSTM-CNN model proposed by the present invention to test and verify the data set. The batch size of this model is set to 20, the learning rate is set to 0.001, and the training frequency is set to 100 times;

[0085] (3) Construct the characteristic frequency points corresponding to the PG original fault signal, build the original signal data set, and input the same set parameters into the LSTM-CNN model, the CNN and DNN models for identification to obtain the comparison results of several models;

[0086] (4) The confusion matrices of the four models are as Figure 7 shown, and the diagnostic accuracies of the four models are compared in Table 5 and Figure 8 .

[0087] Table 5 Accuracies of the Four Models on DDS

[0088]

[0089] Compared with the other three models, the ITD-LSTM-CNN method proposed by the present invention significantly extracts the core fault features from the vibration signal and improves the diagnostic accuracy under various fault conditions. This method greatly improves the calculation efficiency and reduces the training time. The experimental results show that the ITD-LSTM-CNN method is effective.

[0090] 2. Experiment 2

[0091] Experiment 2 uses the data of the Rotating Fault Simulation Test Bench (RFS) of the School of Mechanical Engineering, Tianjin University of Technology and the Tianjin Key Laboratory of Advanced Mechatronic System Design and Intelligent Control to verify the effectiveness of the ITD and LSTM-CNN methods proposed in this paper, which is described in detail as follows:

[0092] (1) Use ITD to reconstruct the signal samples of the RFS experimental platform, build the feature vector data set, and divide it into a training set and a test set according to the ratio of 3:7. The sample division of the RFS data set is shown in Table 6;

[0093] Table 6 Sample Division of RFS Data Set

[0094]

[0095] (2) Use the LSTM-CNN model proposed by the present invention to test and verify the data set. The batch size of this model is set to 20, the learning rate is set to 0.001, and the training frequency is set to 100 times;

[0096] (3) Construct the characteristic frequency points corresponding to the original PG fault signal, construct the original signal data set, and input the same set parameters into the LSTMCNN model, the CNN and DNN models for identification to obtain the comparison results of several models;

[0097] (4) The confusion matrices of the four models are as Figure 9 shown, and the diagnostic accuracies of the four models are compared in Table 7 and Figure 10 .

[0098] Table 7 Accuracies of the Four Models on DDS

[0099]

[0100] Compared with the other three models, the ITD-LSTM-CNN method proposed by the present invention significantly extracts the core fault features from the vibration signal and improves the diagnostic accuracy under various fault conditions. This method greatly improves the computational efficiency and reduces the training time. The experimental results show that the ITD-LSTM-CNN method is effective.

[0101] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made in these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN, characterized in that The method includes the following steps: S1. Obtain several different types of vibration signals of the sun gear of the planetary gearbox on the test bench through a vibration sensor; S2. Decompose each sampling signal into PR components according to ITD; S3. Select the PR components with relatively large instantaneous frequency fluctuations through Hilbert transform to reconstruct the signal; S4. Transform the reconstructed signal into a frequency spectrum through Fourier transform, then extract the amplitudes of the characteristic frequencies, construct a feature vector, and establish a data set; S5. Input the data set into the LSTM-CNN model for training, combine it with the categorical cross-entropy loss function, and use the Adam optimizer and the ReduceLROnPlateau learning rate scheduler to enhance the training process.

2. The intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN according to claim 1, wherein, The S1 further includes the following steps: S1.

1. Divide the health conditions of the sun gear of the planetary gearbox of the dynamic transmission chain simulator into the following types: normal, broken teeth, missing teeth, wear, and divide the health conditions of the sun gear of the planetary gearbox of the rotary fault simulation test bench into the following types: normal, pitting, tooth surface wear, slight crack, moderate crack, severe crack; S1.

2. Set the DDS motor speed to 1800 RPM, the sampling time to 5 seconds, and the sampling frequency to 7680 Hz. Collect 50 groups of vibration data for each type of gear fault. Each group of vibration signals contains 38400 data points. Set the RFS motor speed to 3000 RPM, the sampling time to 5 seconds, and the sampling frequency to 10240 Hz. Collect 50 groups of vibration data for each type of gear fault. Each group of vibration signals contains 51200 data points; S1.

3. Four different types of vibration signals X = [X1, X2, … X n of the sun gear of the planetary gearbox are collected, where X is the vibration signal data set and Xi is the vibration signal.

3. The intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN according to claim 1, characterized in that, The S2 further includes the following steps: S2.

1. Decompose the vibration signal according to ITD, define L as the baseline extraction operator for extracting low-frequency signals, and the original signal X t The decomposition expression is as follows: X t = LX t + (1 - L)X t = L t + H t , where L t = LX t is the baseline signal, and H t = (1 - L)X t is the PR component; S2.

2. Extract signal X t 's extreme value, i.e., X k , and the corresponding time is τ k , where k = 0, 1, 2…, τ0 = 0, define a piecewise baseline extraction operator L as Let X k be X(τ k ), L k be L(τ k ), assuming that L t and H t are defined in the interval [0, τ k , X t applies to the interval [0, τ k+2 , then in (τ k , τ k+1 , where 0 < ∝ < 1, and ∝ is 0.5; S2.

3. Define the PR extraction operator H as where represents the PR component with the highest frequency on the baseline. As the next input signal, repeat the above steps until the limit signal becomes a monotonic function or a constant. Finally, decompose the original signal into multiple PR components with different frequencies and a residual signal. In the formula, p represents the total number of iterations, represents the i-th PR component, and represents the residual signal.

4. The intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN according to claim 1, wherein, In S3: The PR component with large instantaneous power fluctuation is selected by Hilbert transform. Where H[z(t)] represents the value of the Hilbert transform signal at time t, z(t) and Z(τ) represent the values ​​of the original signal at time t and time τ, respectively. The selected PR components are superimposed to reconstruct the signal.

5. The intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN according to claim 1, wherein The S4 further includes the following steps: S4.

1. Transform the reconstructed signal f(t) into the frequency spectrum function F(ω) through Fourier transform, where ω and t represent the frequency and time of the reconstructed signal respectively, and e -iwt represents the complex exponential function, which is e -iwt = cos(ωt) - i sin(ωt); S4.

2. Extract the amplitude of the characteristic frequency A(w)=|F(w)|, construct a feature vector, and each feature vector represents a sample signal to construct a data set.

6. The intelligent fault diagnosis method for planetary gearboxes based on ITD and LSTM-CNN according to claim 1, wherein The S5 further includes the following steps: S5.

1. Construct an LSTM-CNN model, including 1 LSTM layer, 2 convolutional layers, and 2 pooling layers; S5.

2. Input the dataset into the LSTM-CNN model for training and summation testing to obtain the diagnostic accuracy rate. The softmax function is used to classify the fault types of the output layer of the LSTM-CNN, which is expressed as: where σ(x i ) represents the softmax output of the i-th component, and x i is an element of the input vector, represents the sum of the exponents of all input elements, ensuring that the output is normalized to a valid probability distribution; S5.

3. Combine with the categorical cross-entropy loss function to ensure computational stability and training efficiency: where L represents the cross-entropy loss, n is the number of classes, and y i is the true label of class i, and is the predicted probability of class i; S5.

4. Use the Adam optimizer and the ReduceLROnPlateau learning rate scheduler to enhance the training process, ensure fast convergence and efficient processing of sparse gradients, and dynamically adjust the learning rate according to the verification results to prevent training stagnation.

Images