Online fault diagnosis method for RV speed reducer

Abstract

The invention provides an RV reducer online fault diagnosis method, and belongs to the technical field of reducer fault diagnosis based on computer data processing. The method comprises the following steps: firstly, constructing a fault data set covering multi-modal signals and multiple working conditions, and integrating vibration, current and high-frequency elastic stress wave data; performing segmentation, time-frequency conversion and multi-modal feature fusion on the original time series data to generate a three-channel time-frequency feature map, and strengthening visual expression of fault features; then constructing a domain adversarial attention neural network, and realizing cross-working-condition fault feature migration and accurate classification through adversarial training of a feature extractor, a fault classifier and a domain classifier; and finally, deploying an online diagnosis model, generating a health indicator in combination with the health state baseline model, predicting the remaining service life through a long short-term memory network, and completing fault early warning and life evaluation. According to the method, accurate fault identification and residual service life prediction under complex working conditions are realized, and reliable technical support is provided for the RV speed reducer.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of reducer fault diagnosis based on computer data processing, and particularly relates to an RV reducer online fault diagnosis method. BACKGROUND

[0002] In the field of RV reducer fault diagnosis, existing methods mostly rely on single modal data or model training under fixed working conditions, which makes it difficult to cope with complex and variable operating scenarios in actual applications. The fault features of RV reducers vary significantly with operating conditions such as speed and load, and the fault signals often exhibit non-stationary transient characteristics. Meanwhile, the fault signals of different components such as planetary gear trains, cycloid gear trains, and supporting bearings are superimposed on each other, which poses strict requirements on the operating condition adaptability and multi-modal feature fusion capability of the diagnosis model. Existing RV reducer fault diagnosis methods include the following: (1) Fault diagnosis method based on traditional signal processing: features are extracted from single vibration or current signals through Fourier transform, wavelet analysis, etc., and traditional classifiers such as support vector machines and random forests are used to realize fault recognition. This method relies on manual feature design and has poor adaptability to non-stationary signals, and does not consider the impact of operating condition changes on diagnosis results; (2) Fault diagnosis method based on single modal deep learning: models such as convolutional neural networks and recurrent neural networks are constructed to learn features directly from single modal data such as vibration or acoustic emission and complete fault classification. This method can automatically extract features, but lacks multi-source information complementation, making it difficult to fully capture the essence of faults, and has limited cross-condition generalization ability; (3) Fault diagnosis method based on multi-modal fusion: multiple modal data such as vibration and current are integrated, and feature vectors are constructed through simple concatenation or weighted fusion, and then input into the model for diagnosis. This method utilizes multi-source information, but the feature fusion method is rough and does not solve the domain shift problem caused by operating condition differences, and the diagnosis performance of the model under actual complex operating conditions is unstable.

[0003] Therefore, existing RV reducer fault diagnosis methods still have obvious limitations in operating condition adaptability, diagnosis accuracy, and practicality. Traditional signal processing methods rely on manual experience to design features and have insufficient recognition ability for complex faults; single modal deep learning methods lack multi-source information collaboration and are difficult to cope with multi-component compound faults; multi-modal fusion methods do not effectively solve the cross-condition domain shift problem and have poor generalization ability, and most of them can only achieve fault classification and cannot provide equipment health state evaluation and remaining useful life prediction, making it difficult to meet the full life cycle management needs in industrial scenarios. SUMMARY

[0004] To solve the above problems, the present application provides an RV reducer online fault diagnosis method, which includes the following processes: S1, collecting vibration time series data and high-frequency elastic stress wave data of the reducer, and three-phase current time series signals of the motor driving the reducer in real time; S2, segmenting, time-frequency converting and multi-modal feature fusing the collected three kinds of data to generate three-channel time-frequency feature maps; S3, inputting the three-channel time-frequency feature maps into an online fault diagnosis model based on a domain adversarial attention network to output a probability distribution corresponding to a fault label; The online fault diagnosis model comprises a feature extractor and a fault classifier, and the feature extractor realizes key fault feature extraction through a channel and spatial dual attention mechanism; during model training, a domain classifier is combined to realize adversarial training of different working conditions at a feature level through a gradient reversal layer; S4, inputting the three-channel time-frequency feature maps into a health state baseline model trained with health data to calculate a reconstruction error of the current input, construct a health indicator degradation time series, input the degradation time series into a pre-trained long short-term memory network model to obtain a degradation time series prediction curve, and then calculate a time required to reach a failure threshold as a current remaining service life of the reducer.

[0005] Preferably, the three kinds of data collected in S2 are segmented, and the specific process is as follows: A sampling window with a fixed length T is set, an overlap rate M is set, and the three kinds of modal time series data are synchronously and slidingly cut; each time of sliding, three synchronous signal segments are intercepted from the three signals, and the three signal segments are combined to form a group of input data.

[0006] Preferably, the time-frequency conversion in S2 is as follows: The three one-dimensional signal segments respectively from the vibration time series data, the three-phase current time series signals and the high-frequency elastic stress wave data are subjected to continuous wavelet transform; Morlet wavelet is selected to convert each one-dimensional signal segment into a high-resolution two-dimensional time-frequency map, which can show the distribution and transient characteristics of signal energy in time and frequency dimensions, and finally three two-dimensional time-frequency maps are obtained, including a vibration time-frequency map, a current time-frequency map and a sound emission time-frequency map.

[0007] Preferably, the multi-modal feature fusion in S2 is as follows: The three two-dimensional time-frequency maps after time-frequency conversion, i.e., the vibration time-frequency map, the current time-frequency map and the sound emission time-frequency map, are stacked to form a three-channel fused time-frequency map by using a channel stacking strategy.

[0008] Preferably, the feature extractor comprises three convolutional blocks, an attention mechanism module and a global average pooling layer in sequence. Firstly, the first convolutional block contains a 3x3 two-dimensional convolutional layer using 32 filters and a ReLU function, followed by a 2x2 max-pooling layer; the second convolutional block contains a 3x3 two-dimensional convolutional layer using 64 filters and a ReLU activation function, followed by a 2x2 max-pooling layer; the third convolutional block contains a 3x3 two-dimensional convolutional layer using 128 filters and a ReLU activation function, followed by a 2x2 max-pooling layer; The feature map output by the third convolutional block is sent to an attention mechanism module composed of a channel attention module and a spatial attention module in sequence; firstly, in the channel attention module, the input feature map is subjected to parallel pooling, including global average pooling and global maximum pooling, and then enters a multi-layer perceptron for processing, and the output thereof generates channel attention weights via a Sigmoid activation function; the weights are subjected to element-level multiplication with the original input feature map to obtain a channel-weighted feature map; secondly, the channel-weighted feature map enters the spatial attention module, is subjected to channel pooling and then spliced, and is processed by a convolutional layer, and the output thereof generates spatial attention weights via a Sigmoid activation function; finally, the spatial attention weights are subjected to element-level multiplication with the channel-weighted feature map to generate a double-weighted feature map; Finally, the feature map weighted by the attention module is converted into a fixed-length high-dimensional feature vector by a global average pooling layer.

[0009] Preferably, the fault classifier has an input of the high-dimensional feature vector output by the feature extractor, and its structure includes a fully connected layer and an output layer in sequence, the fully connected layer contains 256 neurons and uses a ReLU activation function, and a dropout layer is additionally provided to prevent overfitting; the output layer contains 4 neurons corresponding to four categories of fault labels, i.e., normal healthy state, planetary gear train fault, cycloid gear train fault, and support bearing fault; the output layer uses a Softmax activation function to output a probability distribution corresponding to the fault label.

[0010] Preferably, the domain classifier is a working condition predictor, and its input is connected to the feature vector output by the feature extractor through a gradient reversal layer; its structure includes a fully connected layer and an output layer in sequence, the fully connected layer contains 256 neurons and uses a ReLU activation function; the output layer contains a number of neurons equal to the total number of working condition combinations; the output layer uses a Softmax activation function to output a probability distribution corresponding to the working condition label.

[0011] Preferably, the S3 adversarial training has the following specific process: The fault classification loss calculates the difference between the prediction result of the fault classifier and the real fault label through a classification cross-entropy function, and optimizing this loss updates the parameters of the feature extractor and the fault classifier, forcing the feature extractor to extract a feature vector containing sufficient fault recognition information; at the same time, the domain classification loss also calculates the difference between the prediction result of the domain classifier and the real working condition label through a classification cross-entropy function; in the training, the domain classifier itself is optimized to minimize this domain classification loss, so that it can accurately distinguish the working condition; and when back-propagating, the gradient inversion layer reverses the gradient sign returned by the domain classifier and passes it to the feature extractor, so that the adversarial goal of the feature extractor is to maximize this domain classification loss; finally, the total optimization function of the feature extractor is represented as the fault classification loss minus the domain classification loss multiplied by the adversarial balance coefficient.

[0012] Preferably, the health state baseline model adopts a variational autoencoder composed of a convolutional encoder and a transposed convolutional decoder; the convolutional encoder receives a health three-channel fused time-frequency graph as input, and its structure successively includes three convolutional blocks, which are the same as the three convolutional blocks of the feature extractor, to compress the time-frequency graph; then a fully connected layer is used to output the parameters of the latent distribution, i.e., the mean vector and the log-variance vector corresponding to each latent dimension; the decoder samples a latent vector from the latent distribution, maps it back to a high-dimensional feature space through a fully connected layer, and then successively passes it through three transposed convolutional blocks symmetric to the convolutional blocks of the encoder for upsampling, and finally reconstructs a three-channel fused time-frequency graph with the same size as the original input time-frequency graph; the reconstruction error of the variational autoencoder for the current input is calculated, and this reconstruction error value is defined as the quantized health indicator.

[0013] Preferably, in S4, the continuously generated quantized health indicator values are recorded and stored in chronological order to form a health indicator degradation time sequence reflecting the degradation process of the reducer from health to degradation; then the constructed health indicator degradation time sequence is input into a pre-trained long short-term memory network model, which will automatically extrapolate the future growth trend of the sequence; by setting a predefined failure threshold, when the health indicator value is greater than the failure threshold, it indicates that the reducer cannot meet the service requirements, and the future time required for the predicted health indicator growth curve to reach the failure threshold is calculated, which is output as the current remaining service life of the reducer.

[0014] Compared with the prior art, the present application has the following beneficial effects: (1) Multi-modal collaborative perception fault feature representation mechanism: the present application integrates multi-modal data of vibration, three-phase current and high-frequency elastic stress wave, generates a three-channel time-frequency graph through continuous wavelet transform and channel stacking fusion, fully excavates the complementary information of different modal data, comprehensively depicts the multi-dimensional features of faults, and improves the recognition ability of compound faults; (2) Domain adversarial attention-driven working condition adaptive modeling: Introducing domain adversarial training and channel-space dual attention mechanism, cross-working condition feature transfer is achieved through gradient inversion layer, while strengthening the extraction of key fault features, effectively solving the domain shift problem caused by working condition changes, and significantly improving the generalization ability of the model under low speed, medium speed, high speed and different load combination working conditions. (3) Full life cycle health management system: It not only realizes accurate identification and timely early warning of faults, but also builds a health status baseline model through variational autoencoder, generates quantitative health indicators, and combines long short-term memory network to predict the remaining service life, forming a full-process management of fault diagnosis-health assessment-life prediction, which meets the actual needs of equipment maintenance in industrial scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall method of the present invention.

[0016] Figure 2 This is a flowchart illustrating the construction process of the multimodal time-series dataset of this invention.

[0017] Figure 3 This is a flowchart of the signal preprocessing and multimodal feature map construction process of the present invention.

[0018] Figure 4 This is a schematic diagram of the domain adversarial neural network architecture and training loss function of this invention.

[0019] Figure 5 This is a schematic diagram illustrating the online diagnostic deployment and health status assessment of the present invention.

[0020] Figure 6 This is a diagram showing the ablation experiment results in an embodiment of the present invention.

[0021] Figure 7 This is a fault classification confusion matrix diagram in an embodiment of the present invention.

[0022] Figure 8 This is a timing diagram showing the health state evolution of the RV reducer in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] The online fault diagnosis method for RV reducers provided in this embodiment includes four core stages in its overall process, such as... Figure 1As shown: first, in the S1 stage, a fault dataset covering multi-modal signals and multi-conditions is constructed, integrating vibration, current, and high-frequency elastic stress wave data, labeling working conditions and fault labels, to provide a basis for subsequent modeling; then, in the S2 stage, the original time series data is segmented, time-frequency converted, and multi-modal features are fused to generate a three-channel time-frequency feature map, strengthening the visual expression of fault features; in the S3 stage, a domain adversarial attention neural network is constructed, through the adversarial training of the feature extractor, fault classifier, and domain classifier, realizing the transfer and accurate classification of fault features across conditions; finally, in the S4 stage, the online diagnosis model is deployed, the health indicator is generated by combining the health state baseline model, and the remaining useful life is predicted by the long short-term memory network, completing fault warning and life assessment.

[0025] S1, RV reducer multi-modal sensing and multi-condition fault dataset construction: This step aims to integrate structured multi-modal time series data with condition and fault labels, providing a data basis for subsequent model training. The flowchart of this step is shown in Figure 2 .

[0026] S1.1 Multi-modal sensor deployment and data acquisition: On a standard RV reducer fault diagnosis test bench, multi-physical field sensors are deployed on key parts of the measured reducer to achieve comprehensive information perception. First, acceleration sensors are installed radially and axially near the input flange and output flange of the reducer housing to collect vibration time series data. Second, high-precision Hall current sensors are installed on the three-phase AC input of the servo motor driving the RV reducer to collect three-phase current time series signals. Finally, an acoustic emission sensor is attached to the surface of the reducer housing near the main bearing to collect high-frequency elastic stress wave data.

[0027] S1.2 Multi-condition experiment matrix design: To ensure that the model can adapt to the real and variable working conditions of the reducer in precision transmission applications, this step simulates typical working condition combinations on the test bench. Through the servo motor controller and the magnetic powder brake, the RV reducer is controlled to run stably at multiple different speeds, including low speed 500 rpm, medium speed 1000 rpm, and high speed 1500 rpm, and multiple different loads, including light load 20% rated torque, medium load 50%, and heavy load 80%. First, fault implantation and data collection are performed. Under each of the above speed and load condition combinations, multi-modal sensor time series data of the RV reducer in different health states are collected, and each data is accurately labeled with its corresponding condition label and health state label: 1. Normal health state: Collect baseline data of a new or well-running healthy reducer under all working conditions; 2. Typical fault conditions: Typical faults of RV reducer are pre-fabricated on replaceable components by means of electrical discharge machining, artificial wear and fatigue test. These faults include: (1) planetary gear train faults: pitting corrosion, root crack or tooth fracture of planetary gear tooth surface. (2) cycloid gear train faults: non-uniform wear of cycloid tooth profile, raceway spalling or rolling element damage of crank bearing. (3) support bearing faults: inner ring, outer ring or rolling element faults of main support bearing at the input or output end of the reducer.

[0028] S1.3 Data integration: This step integrates the data of S1.1 and S1.2, and finally outputs a structured multi-modal time series dataset with working condition and fault labels, which includes two types of data: multi-modal time series data and label data. Multi-modal time series data includes vibration time series data, three-phase current time series signal, and high-frequency elastic stress wave data. Label data includes working condition label and fault label. The working condition label annotates the specific speed and load corresponding to each time series data collection. The fault label annotates the health status of the reducer corresponding to each time series data, including normal health status, planetary gear train fault, cycloid gear train fault, and support bearing fault.

[0029] S2, Signal preprocessing and multi-modal feature map construction: This step aims to preprocess and feature engineer the multi-modal time series dataset output by S1.3. First, the original one-dimensional time series signal is segmented by fixed length sliding window. Continuous wavelet transform is performed on vibration, current and acoustic emission signals to generate corresponding two-dimensional time-frequency maps. Second, the channel stacking strategy is used to fuse the three time-frequency maps into a unified three-channel time-frequency feature map. Finally, a preprocessed feature dataset containing three-channel fused feature maps and corresponding labels is output, preparing for deep model input. The flow chart of this step is shown in Figure 3 .

[0030] S2.1 Time series data segmentation and sample construction: This step first segments the three types of original signals: vibration time series data, three-phase current time series signal, and high-frequency elastic stress wave data, from the multi-modal time series dataset of S1.3, using the fixed length overlapping sliding window method. Specifically, a fixed length T sampling window is set, and the overlap rate is set to M. The multi-modal time series data is cut synchronously by sliding. Each sliding cuts three synchronous signal segments from the three types of signals. The three signal segments are combined to form an original multi-modal sample. At the same time, according to the label data in the multi-modal time series dataset, the working condition label and fault label corresponding to the collection time of the original multi-modal sample are assigned.

[0031] S2.2 Time-frequency analysis based on continuous wavelet transform: Considering that the RV reducer operates under a variety of different rotational speed conditions, and that the fault usually exhibits non-stationary transient impact characteristics, this step adopts a time-frequency analysis technique to convert one-dimensional signal segments into two-dimensional images. Specifically, for each original multi-modal sample in S2.1, three one-dimensional signal segments from the vibration time series data, three-phase current time series signals, and high-frequency elastic stress wave data, respectively, are subjected to continuous wavelet transform. Specifically, the Morlet wavelet is selected to convert each one-dimensional signal segment into a high-resolution two-dimensional time-frequency image. The two-dimensional time-frequency image can clearly show the distribution and transient characteristics of signal energy in both time and frequency dimensions, which is more conducive to subsequent model automatic extraction of deep fault features than the original time series signal. This step ultimately obtains three two-dimensional time-frequency images, including a vibration time-frequency image, a current time-frequency image, and an acoustic emission time-frequency image.

[0032] S2.3 Multi-modal feature map fusion: This step fuses the separated two-dimensional time-frequency images converted from multi-source sensor information in S2.2 into a unified feature representation rich in multi-dimensional information. Specifically, a channel stacking strategy is adopted to stack the three two-dimensional time-frequency images corresponding to the same original multi-modal sample in S2.1, i.e., the vibration time-frequency image, the current time-frequency image, and the acoustic emission time-frequency image, to form a three-channel fused time-frequency image. In this way, the subsequent convolutional neural network model can simultaneously perceive and fuse the composite fault features from vibration, current, and acoustic emission at the same spatial location.

[0033] S2.4 Preprocessed feature dataset output: This step finally outputs a structured preprocessed feature dataset, which will serve as the final input for the S3 deep learning model. The dataset consists of two parts: three-channel fused feature data and label data. The three-channel fused feature data includes a large number of three-channel fused time-frequency images generated in S2.3, and the label data includes the corresponding working condition label and fault label retained in S2.1 for each three-channel fused time-frequency image. Figure 1

[0034] S3, Condition-adaptive fault diagnosis model construction based on domain adversarial attention network: This step is the core of the present application, and it constructs a domain adversarial attention neural network composed of a feature extractor, a fault classifier, and a domain classifier. This step enhances key feature extraction through an attention mechanism, and simultaneously realizes adversarial training of different working conditions at the feature level using a gradient reversal layer, so that the model maintains high fault classification accuracy while being invariant to working conditions. The training data for this step is the preprocessed feature dataset output by S2.4. After adversarial training converges, the domain classifier is stripped off, and only the feature extractor and fault classifier for inference are retained to generate an online diagnosis model.

[0035] ​S3.1 Domain adversarial attention model architecture design: This step builds a domain adversarial neural network architecture composed of three core components, including a feature extractor, a fault classifier, and a domain classifier. The domain adversarial attention model architecture and the training loss function are as shown in Figure 4

[0036] 1. Feature extractor: composed of a deep convolutional attention network, the input of which is the three-channel fused time-frequency graph in S2.4. Its structure includes three convolutional blocks, a channel and spatial attention module, and a global average pooling layer in turn. First, convolutional block one contains a 3x3 two-dimensional convolution layer using 32 filters and a ReLU function, followed by a 2x2 max pooling layer; convolutional block two contains a 3x3 two-dimensional convolution layer using 64 filters and a ReLU activation function, followed by a 2x2 max pooling layer; convolutional block three contains a 3x3 two-dimensional convolution layer using 128 filters and a ReLU activation function, followed by a 2x2 max pooling layer. Second, the feature map output by convolutional block three is fed into an attention mechanism composed of a channel attention module and a spatial attention module in sequence. First, in the channel attention module, the input feature map is processed through parallel pooling, including global average pooling and global maximum pooling, and then enters a multi-layer perceptron for processing, and the output is passed through a Sigmoid activation function to generate channel attention weights; the weights are multiplied element-wise with the original input feature map to obtain a channel-weighted feature map. Second, the channel-weighted feature map enters the spatial attention module, is processed through channel pooling, then concatenated, and then processed through a convolution layer, and the output is passed through a Sigmoid activation function to generate spatial attention weights; finally, the spatial attention weights are multiplied element-wise with the channel-weighted feature map to generate a double-weighted feature map. Finally, the attention module weighted feature map is converted into a fixed-length high-dimensional feature vector through a global average pooling layer.

[0037] 2. Fault classifier: its input is the high-dimensional feature vector output by the feature extractor. Its structure includes a fully connected layer and an output layer in turn, the fully connected layer contains 256 neurons, uses a ReLU activation function, and is attached with a dropout layer to prevent overfitting. The output layer contains 4 neurons, corresponding to the four categories of fault labels in S1.3, i.e. normal healthy state, planetary gear train fault, cycloid gear train fault, and support bearing fault. This layer uses a Softmax activation function to output a probability distribution corresponding to the fault label.

[0038] ​3. Domain classifier: This classifier is a working condition predictor, whose input is connected to the feature vector output from the feature extractor through a gradient reversal layer. Its structure includes a fully connected layer and an output layer in turn. The fully connected layer contains 256 neurons and uses the ReLU activation function. The output layer contains a number of neurons equal to the total number of unique combinations of working condition labels in S1.3. This layer uses the Softmax activation function and outputs a probability distribution corresponding to the working condition label.

[0039] S3.2 Adversarial training strategy and optimization: This step uses the preprocessed feature dataset in S2.4 to perform end-to-end adversarial training on the model defined in S3.1. The training process optimizes a composite loss function that includes two mutually antagonistic objectives: minimizing the fault classification loss and maximizing the domain classification loss. Specifically, the fault classification loss is calculated by the classification cross-entropy function to measure the difference between the predicted results of the fault classifier and the true fault labels. Optimizing this loss updates the parameters of the feature extractor and the fault classifier, forcing the feature extractor to extract feature vectors that contain sufficient fault recognition information. At the same time, the domain classification loss is also calculated by the classification cross-entropy function to measure the difference between the predicted results of the domain classifier and the true working condition labels. During training, the domain classifier itself is optimized to minimize this domain classification loss, enabling it to accurately distinguish working conditions. When backpropagating, the gradient reversal layer reverses the gradient sign returned by the domain classifier and passes it to the feature extractor, which makes the feature extractor's adversarial objective maximize this domain classification loss. This adversarial force forces the feature extractor to extract feature vectors that actively discard working condition-related information, making it impossible for the domain classifier to distinguish the source. Ultimately, the total optimization function of the feature extractor is the fault classification loss minus the domain classification loss multiplied by the adversarial balance coefficient, achieving high fault classification accuracy and strong working condition invariance.

[0040] S3.3 Solidification and output of online fault diagnosis model: After the adversarial training in S3.2 converges, i.e., the fault classification loss converges to a minimum, and the classification accuracy of the domain classifier on the working condition approaches random guessing, the trained S3.1 domain-adversarial attention model architecture is solidified. Specifically, the components used only for training, including the domain classifier and its associated gradient reversal layer, are detached from the computation graph and discarded. Only the trained and optimized parameters of the working condition adaptive feature extractor and the fault classifier are retained and exported. The cascade of these two components, i.e., the feature extractor connected to the fault diagnosis classifier, is encapsulated and saved as a separate, forward inference-only model, which is the final output of the online fault diagnosis model of the present invention and is delivered for subsequent online deployment.

[0041] S4, Online diagnosis deployment and health status assessment: This step aims to deploy the online fault diagnosis model output by S3.3 into practical applications, generate real-time three-channel fusion time-frequency graphs in combination with real-time collected multi-modal signals, and realize immediate identification and alarm of faults. At the same time, the reconstruction error of the input sample is calculated by using the health state baseline variational autoencoder to generate a health indicator, which monitors the degradation process of the equipment. Based on the time series of the health indicator, the long short-term memory network is used to predict its future trend, and the remaining useful life of the reducer is determined by the failure threshold. The schematic diagram of this step is shown in Figure 5

[0042] S4.1 Health state baseline model construction: This step aims to construct a baseline model for quantifying the health state. First, the health data is screened, and from the pre-processed feature data set output by S2.4, the three-channel fusion feature data with normal health state in the fault label of all label data is selected. Then the baseline model is trained, a variational autoencoder model is constructed, and only the above-mentioned health three-channel fusion time-frequency graph is used to train the variational autoencoder. The model structure is as follows: The model is composed of a convolutional encoder and a transposed convolutional decoder. The convolutional encoder receives the health three-channel fusion time-frequency graph as input, and its structure includes three convolutional blocks in turn, which are the same as the first, second and third structures of the convolutional blocks of the feature extractor in S3.1, to compress the time-frequency graph; then a fully connected layer is used to output the parameters of the latent distribution, i.e. the mean vector and the log variance vector corresponding to each latent dimension; the decoder then samples a latent vector from the latent distribution, maps it back to a high-dimensional feature space through a fully connected layer, and then performs upsampling through three transposed convolutional blocks symmetric to the convolutional blocks of the encoder, and finally reconstructs a three-channel fusion time-frequency graph with the same size as the original input time-frequency graph. After training, the model has the ability to fit the feature distribution of the health data, and can reproduce the input data under normal health state with very low reconstruction error. The trained health state baseline model is output.

[0043] ​S4.2 Online diagnosis and health indicator generation: This step is the real-time running stage of the online fault diagnosis model and the health state baseline model, deployed in the edge computing device or the industrial host computer. Firstly, this step performs real-time data processing procedures: real-time acquisition of vibration time series data, three-phase current time series signals and high-frequency elastic stress wave data through the acceleration sensor, Hall current sensor and acoustic emission sensor laid in S1.1; then, the collected real-time data stream is continuously generated into real-time three-channel fusion time-frequency maps in strict accordance with the complete preprocessing procedures defined by S2.1 time series segmentation, S2.2 time-frequency analysis and S2.3 multi-modal feature map fusion. Then, real-time fault diagnosis is performed, and the generated real-time three-channel fusion time-frequency maps are input into the online fault diagnosis model solidified in S3.3, and the fault classifier component will immediately output a probability distribution corresponding to the fault label, thereby realizing immediate fault diagnosis and early warning. At the same time of performing real-time fault diagnosis, health indicator generation is performed, and the same real-time three-channel fusion time-frequency maps are input into the health state baseline model trained in S4.1, and the reconstruction error of the variational autoencoder for the current input is calculated. This reconstruction error value is defined as the quantitative health indicator, which is close to zero when the reducer is in a normal healthy state, and significantly increases when a typical fault state occurs.

[0044] S4.3 Remaining useful life prediction: This step aims to predict the future failure risk of the reducer based on the historical trend of the health indicator generated in S4.2. Firstly, degradation sequence construction is performed, and the continuously generated quantitative health indicator values are recorded and stored in chronological order in the online diagnosis process of S4.2, forming a health indicator degradation time series reflecting the process of the reducer from health to degradation. Then, the remaining useful life prediction is performed, and a long short-term memory network model is constructed and pre-trained. This model is specifically used to learn and predict the future trend of non-stationary time series. The constructed health indicator degradation time series is input into the long short-term memory network model, which will automatically extrapolate the future growth trend of the sequence. By setting a predefined failure threshold, i.e. the health indicator value greater than the failure threshold, representing that the reducer cannot meet the service requirements, the future time required for the predicted health indicator growth curve to reach the failure threshold is calculated, which is output as the current remaining useful life of the reducer.

[0045] Simulation experiment: 1. Simulation experiment setup: The experimental platform is built based on the standard RV reducer fault diagnosis test bench, and the core parameters are as follows: RV reducer model: RV-40E, transmission ratio: 1:81, rated torque: 50 N·m; Servo motor: 1.5 kW permanent magnet synchronous motor, speed range: 0-2000 rpm, supporting speed closed-loop control; Loading device: magnetic powder brake, load adjustment range: 0-80% rated torque; Sensor configuration: acceleration sensor: PCB 352C33, range: ±50g, sampling frequency: 20kHz, arranged along the radial / axial direction; Hall current sensor: ACS758, accuracy: ±1.5%, sampling frequency: 10kHz, collecting three-phase current of the motor; Acoustic emission sensor: PAC R15a, resonant frequency: 150kHz, sampling frequency: 1MHz, attached to the shell near the bearing; The experimental data covers 9 working condition combinations of 3 speeds of 500 rpm, 1000 rpm, and 1500 rpm, and 3 loads of 20%, 50%, and 80% rated torque, including 4 types of healthy states: normal, planetary gear train fault, cycloid gear train fault, and support bearing fault, 1000 time series samples are collected for each type of state, and the single sample duration is 10s.

[0046] 2. Experimental parameter setting: sliding window segmentation: window length T=1024 sampling points, overlap rate M=50%; continuous wavelet transform: Morlet wavelet is selected, scale range is 1-64, time-frequency graph resolution is 256x256 pixels; feature map fusion: three-channel stacking, including vibration time-frequency graph, flow time-frequency graph, and acoustic emission time-frequency graph; Model training parameters: hardware environment: NVIDIA 4090 GPU (24GB video memory); software environment: Python 3.8, PyTorch 1.12, Scikit-learn 1.2, Matplotlib 3.7; Domain adversarial attention network parameters: feature extractor: 3 convolutional blocks, each convolutional block contains 3x3 convolution, ReLU, and 2x2 max pooling, channel-spatial attention module, which are realized by PyTorch; fault classifier: fully connected layer contains 256 neurons, ReLU activation function and Dropout, Dropout inactivation parameter is set to 0.5, output layer contains 4 neurons and Softmax; domain classifier: fully connected layer contains 256 neurons and ReLU, output layer contains 9 neurons and Softmax; training parameters: batch size is 32, learning rate is 1e-4, Adam optimizer is used, adversarial balance coefficient is 0.3, training rounds are 100; Variational autoencoder: symmetric structure with feature extractor convolutional block, latent dimension is 64, reconstruction loss is MSE, training rounds are 80; LSTM: 2-layer LSTM with 128 neurons in the hidden layer and a fully connected layer, 30 time steps, 10 prediction steps, and 50 training rounds.

[0047] 3. Comparison model parameters: Model 1: only use vibration time-frequency diagram to train CNN classifier; Model 2: simple splicing of vibration, current, acoustic emission time-frequency diagram is input into CNN; Model 3: directly train multi-modal CNN; Model 4: domain adversarial model without attention mechanism.

[0048] 4. Simulation experiment results Table 1. Experimental results

[0049] The experimental results are shown in Table 1 and Figure 6 The ablation experiment results show that the method has the highest classification accuracy, proving the effectiveness of the network structure and training strategy of the application.

[0050] The fault classification confusion matrix is shown in Figure 7 As can be seen from the figure, the number of sample classification errors is small, proving that the application can effectively distinguish the features of the four types of faults.

[0051] Figure 8 The health state evolution process of the RV reducer in the 1000-hour full life cycle is shown, the fault warning point is triggered at 692 hours, which is 208 hours earlier than the complete failure time of 900h, leaving sufficient time for maintenance planning and spare parts procurement, proving that the application can effectively maintain the health of the equipment and reduce the damage risk of the RV reducer.

[0052] The above only describes the preferred embodiments of the application and is not intended to limit the application. Those skilled in the art can make various modifications and changes to the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

[0053] Although the specific embodiments of the application have been described above, they are not intended to limit the scope of protection of the application. Those skilled in the art should understand that various modifications or changes made to the technical solutions of the application without creative labor are still within the protection scope of the application.

Claims

1. An online fault diagnosis method for RV reducers, characterized in that, Includes the following processes: S1, real-time acquisition of reducer vibration timing data and high-frequency elastic stress wave data, as well as the three-phase current timing signal of the motor driving the reducer; S2, the three types of collected data are segmented, time-frequency converted and multimodal feature fused to generate a three-channel time-frequency feature map; S3, input the three-channel time-frequency feature map into the online fault diagnosis model based on domain adversarial attention network, and output a probability distribution corresponding to the fault label; The online fault diagnosis model includes a feature extractor and a fault classifier. The feature extractor extracts key fault features through a dual attention mechanism of channel and space. During model training, a domain classifier is combined with a gradient inversion layer to achieve adversarial training of the feature layer under different working conditions. S4. Input the three-channel time-frequency feature map into the health status baseline model trained with health data, calculate the reconstruction error of the current input, construct the health indicator degradation time series, input it into the pre-trained long short-term memory network model, obtain the degradation time series prediction curve, and then calculate the time required to reach the failure threshold, which is used as the current remaining service life of the reducer.

2. The online fault diagnosis method for an RV reducer as described in claim 1, characterized in that: In step S2, the three types of collected data are segmented, and the specific process is as follows: Set a sampling window of fixed length T and an overlap rate of M, and perform synchronous sliding cuts on the three modal time series data; each slide extracts three synchronous signal segments from the three signals, and combines these three signal segments to form a set of input data.

3. The online fault diagnosis method for an RV reducer as described in claim 2, characterized in that: The time-frequency conversion in S2 is specifically as follows: For three one-dimensional signal segments derived from vibration time-series data, three-phase current time-series signals, and high-frequency elastic stress wave data, continuous wavelet transforms are performed on each segment. Morlet wavelets are then used to convert each one-dimensional signal segment into a high-resolution two-dimensional time-frequency plot. This two-dimensional time-frequency plot can display the distribution and transient characteristics of signal energy in both time and frequency dimensions, ultimately resulting in three two-dimensional time-frequency plots: vibration time-frequency plot, current time-frequency plot, and acoustic emission time-frequency plot.

4. The online fault diagnosis method for an RV reducer as described in claim 3, characterized in that: The multimodal feature fusion process in S2 is as follows: A channel stacking strategy is adopted to stack the three two-dimensional time-frequency maps after time-frequency conversion, namely the vibration time-frequency map, the current time-frequency map, and the acoustic emission time-frequency map, to form a three-channel fused time-frequency map.

5. The online fault diagnosis method for an RV reducer as described in claim 1, characterized in that: The feature extractor comprises, in sequence, three convolutional blocks, an attention mechanism module, and a global average pooling layer; First, convolutional block one contains a 3x3 2D convolutional layer using 32 filters and the ReLU function, followed by a 2x2 max pooling layer; convolutional block two contains a 3x3 2D convolutional layer using 64 filters and the ReLU activation function, followed by a 2x2 max pooling layer; convolutional block three contains a 3x3 2D convolutional layer using 128 filters and the ReLU activation function, followed by a 2x2 max pooling layer. The feature map of the three outputs of the convolution block is fed into an attention mechanism module consisting of a channel attention module and a spatial attention module in sequence. First, in the channel attention module, the input feature map is subjected to parallel pooling, including global average pooling and global max pooling, and then enters a multilayer perceptron for processing. Its output is used to generate channel attention weights through the Sigmoid activation function. The weights are multiplied element-wise with the original input feature map to obtain a channel-weighted feature map. Then, the channel-weighted feature map enters the spatial attention module, is concatenated after channel pooling, and then processed by a convolutional layer. Its output is then activated by the Sigmoid activation function to generate spatial attention weights. Finally, the spatial attention weights are multiplied element-wise with the channel-weighted feature map to generate a double-weighted feature map; Finally, the weighted feature map from the attention module is passed through a global average pooling layer to transform it into a fixed-length high-dimensional feature vector.

6. The online fault diagnosis method for an RV reducer as described in claim 1, characterized in that: The fault classifier takes a high-dimensional feature vector as input from the feature extractor. Its structure includes a fully connected layer and an output layer. The fully connected layer contains 256 neurons, uses the ReLU activation function, and has an additional dropout layer to prevent overfitting. The output layer contains 4 neurons, corresponding to the four categories of fault labels: normal health, planetary gear train fault, cycloidal gear train fault, and support bearing fault. The output layer uses the Softmax activation function to output a probability distribution corresponding to the fault label.

7. The online fault diagnosis method for an RV reducer as described in claim 1, characterized in that: The domain classifier is a working condition predictor. Its input is connected to the feature vector output by the feature extractor through a gradient inversion layer. The structure includes a fully connected layer and an output layer. The fully connected layer contains 256 neurons and uses the ReLU activation function. The output layer contains a total number of neurons for each working condition combination. The output layer uses the Softmax activation function and outputs a probability distribution corresponding to the working condition label.

8. The online fault diagnosis method for an RV reducer as described in claim 1, characterized in that: The adversarial training in S3 is specifically performed as follows: The fault classification loss is calculated using the classification cross-entropy function to determine the difference between the fault classifier's prediction and the actual fault label. Optimizing this loss updates the parameters of both the feature extractor and the fault classifier, forcing the feature extractor to extract feature vectors containing sufficient fault identification information. Similarly, the domain classification loss is calculated using the classification cross-entropy function to determine the difference between the domain classifier's prediction and the actual operating condition label. During training, the domain classifier itself is optimized to minimize this domain classification loss, enabling it to accurately distinguish operating conditions. The gradient reversal layer, during backpropagation, reverses the sign of the gradient returned by the domain classifier before passing it to the feature extractor, making the feature extractor's adversarial objective to maximize this domain classification loss. Ultimately, the overall optimization function of the feature extractor is the fault classification loss minus the adversarial balance coefficient multiplied by the domain classification loss.

9. The online fault diagnosis method for an RV reducer as described in claim 5, characterized in that: The health status baseline model employs a variational autoencoder, consisting of a convolutional encoder and a transposed convolutional decoder. The convolutional encoder receives a fused three-channel time-frequency map of health as input. Its structure includes three convolutional blocks, identical to the three convolutional blocks of the feature extractor, compressing the time-frequency map. Subsequently, it outputs the parameters of the latent distribution through a fully connected layer, namely the mean vector and log-variance vector corresponding to each latent dimension. The decoder samples a latent vector from the latent distribution, maps it back to the high-dimensional feature space through a fully connected layer, and then upsamples it through three transposed convolutional blocks symmetrical to the encoder's convolutional block structure. Finally, it reconstructs a three-channel fused time-frequency map with the same size as the original input time-frequency map. The reconstruction error of the variational autoencoder to the current input is calculated, and this reconstruction error value is defined as the quantized health indicator.

10. The online fault diagnosis method for an RV reducer as described in claim 1, characterized in that: In step S4, the continuously generated quantitative health indicator values ​​are recorded and stored in chronological order to form a health indicator degradation time series reflecting the decelerator's process from health to degradation. Subsequently, the constructed health indicator degradation time series is used as input to a pre-trained long short-term memory network model, which will automatically extrapolate the future growth trend of the sequence. By setting a predefined failure threshold, when the health indicator value exceeds the failure threshold, it indicates that the reducer can no longer meet the service requirements. The future time required for the predicted health indicator growth curve to reach the failure threshold is calculated, and this future time is output as the current remaining service life of the reducer.

Images