An online fault diagnosis method for RV reducer

Abstract

The application provides an RV reducer online fault diagnosis method, and belongs to the technical field of reducer fault diagnosis based on computer data processing; firstly, a fault data set covering multi-modal signals and multiple working conditions is constructed, and vibration, current and high-frequency elastic stress wave data are integrated; the original time series data are segmented, time-frequency conversion and multi-modal feature fusion are performed, a three-channel time-frequency feature map is generated, and the visual expression of fault features is strengthened; then, a domain adversarial attention neural network is constructed, through the adversarial training of the feature extractor, the fault classifier and the domain classifier, the fault feature migration and accurate classification across working conditions are realized; finally, the online diagnosis model is deployed, the health indicator is generated in combination with the health state baseline model, the remaining useful life is predicted through the long short-term memory network, and fault early warning and life assessment are completed. The application realizes accurate identification of faults under complex working conditions and prediction of the remaining useful life, and provides reliable technical support for the RV reducer.

Description

Technical Field

[0001] This invention belongs to the field of speed reducer fault diagnosis technology based on computer data processing, and particularly relates to an online fault diagnosis method for RV speed reducers. Background Technology

[0002] In the field of RV reducer fault diagnosis, existing methods mostly rely on single-modal data or model training under fixed operating conditions, making it difficult to cope with complex and ever-changing operating scenarios in practical applications. The fault characteristics of RV reducers vary significantly with changes in operating conditions such as speed and load, and fault signals often exhibit non-stationary transient characteristics. Furthermore, fault signals from different components such as planetary gear trains, cycloidal gear trains, and support bearings are superimposed, placing stringent demands on the diagnostic model's adaptability to operating conditions and its ability to fuse multimodal features. Existing RV reducer fault diagnosis methods include the following:

[0003] (1) Fault diagnosis method based on traditional signal processing: Features are extracted from a single vibration or current signal using techniques such as Fourier transform and wavelet analysis, and then traditional classifiers such as support vector machine and random forest are used to identify the fault. This method relies on manually designed features, has poor adaptability to non-stationary signals, and does not consider the impact of changes in operating conditions on the diagnostic results;

[0004] (2) Fault diagnosis method based on single-modal deep learning: Convolutional neural networks, recurrent neural networks and other models are constructed to learn features directly from single-modal data such as vibration or acoustic emission and complete fault classification. Although this method can automatically extract features, it lacks multi-source information complementarity, makes it difficult to fully characterize the essence of the fault, and has limited generalization ability across working conditions;

[0005] (3) Fault diagnosis method based on multimodal fusion: Integrate multiple modal data such as vibration and current, construct feature vectors through simple splicing or weighted fusion, and then input them into the model for diagnosis. Although this method utilizes multi-source information, the feature fusion method is coarse and does not solve the domain shift problem caused by differences in working conditions. The diagnostic performance of the model under actual complex working conditions is unstable.

[0006] Therefore, existing RV reducer fault diagnosis methods still have significant limitations in terms of operating condition adaptability, diagnostic accuracy, and practicality. Traditional signal processing methods rely on manual experience to design features, resulting in insufficient ability to identify complex faults; single-modal deep learning methods lack multi-source information coordination, making it difficult to cope with complex faults involving multiple components; multi-modal fusion methods have not effectively solved the problem of cross-operating condition domain offset, have poor generalization ability, and most can only achieve fault classification, failing to provide equipment health status assessment and remaining service life prediction, thus failing to meet the full life cycle management needs in industrial scenarios. Summary of the Invention

[0007] To address the above problems, this invention proposes an online fault diagnosis method for RV reducers, comprising the following steps:

[0008] S1, real-time acquisition of vibration timing data and high-frequency elastic stress wave data of the reducer, as well as the three-phase current timing signal of the motor driving the reducer;

[0009] 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;

[0010] 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;

[0011] 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.

[0012] 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.

[0013] Preferably, in step S2, the three types of collected data are segmented, and the specific process is as follows:

[0014] 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.

[0015] Preferably, the time-frequency conversion in S2 specifically involves the following process:

[0016] 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.

[0017] Preferably, the multimodal feature fusion in S2 specifically involves the following process:

[0018] 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.

[0019] Preferably, the feature extractor comprises three convolutional blocks, an attention mechanism module, and a global average pooling layer.

[0020] 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.

[0021] The feature map output from the convolutional 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 undergoes parallel pooling, including global average pooling and global max pooling, and then enters a multilayer perceptron for processing. Its output is activated by a sigmoid activation function to generate channel attention weights. These weights are multiplied element-wise with the original input feature map to obtain a channel-weighted feature map. Second, this channel-weighted feature map enters the spatial attention module, is concatenated after channel pooling, and then processed by a convolutional layer. Its output is activated by a sigmoid activation function to generate spatial attention weights. Finally, these spatial attention weights are multiplied element-wise with the channel-weighted feature map to generate a double-weighted feature map.

[0022] 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.

[0023] Preferably, the fault classifier takes a high-dimensional feature vector output by the feature extractor as input and 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 healthy state, 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.

[0024] Preferably, the domain classifier is a working condition predictor, whose 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 in sequence. The fully connected layer contains 256 neurons and uses the ReLU activation function; the output layer contains a total number of neurons for the working condition combinations; the output layer uses the Softmax activation function and outputs a probability distribution corresponding to the working condition label.

[0025] Preferably, the adversarial training in S3 specifically involves the following process:

[0026] 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.

[0027] Preferably, 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, and 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 structure of the encoder's convolutional blocks, finally reconstructing a fused three-channel time-frequency map of 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.

[0028] Preferably, 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 process of the reducer 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 is greater than 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.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) Fault feature characterization mechanism of multimodal collaborative sensing: This invention integrates multimodal data of vibration, three-phase current and high-frequency elastic stress wave, and generates a three-channel time-frequency map by continuous wavelet transform and channel stacking fusion, fully explores the complementary information of different modal data, comprehensively depicts the multidimensional features of faults, and improves the ability to identify composite faults.

[0031] (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.

[0032] (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

[0033] Figure 1 This is a flowchart of the overall method of the present invention.

[0034] Figure 2 This is a flowchart illustrating the construction process of the multimodal time-series dataset of this invention.

[0035] Figure 3 This is a flowchart of the signal preprocessing and multimodal feature map construction process of the present invention.

[0036] Figure 4 This is a schematic diagram of the domain adversarial neural network architecture and training loss function of this invention.

[0037] Figure 5This is a schematic diagram illustrating the online diagnostic deployment and health status assessment of the present invention.

[0038] Figure 6 This is a diagram showing the ablation experiment results in an embodiment of the present invention.

[0039] Figure 7 This is a fault classification confusion matrix diagram in an embodiment of the present invention.

[0040] 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

[0041] 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.

[0042] The online fault diagnosis method for RV reducers provided in this embodiment includes four core stages in its overall process, such as... Figure 1 As shown: First, in stage S1, a fault dataset covering multimodal signals and multiple operating conditions is constructed, integrating vibration, current, and high-frequency elastic stress wave data, and labeling operating conditions and faults to provide a foundation for subsequent modeling. Next, in stage S2, the original time-series data is segmented, time-frequency transformed, and multimodal features are fused to generate a three-channel time-frequency feature map, enhancing the visual representation of fault features. In stage S3, a domain adversarial attention neural network is constructed, and through adversarial training of the feature extractor, fault classifier, and domain classifier, fault feature transfer and accurate classification across operating conditions are achieved. Finally, in stage S4, an online diagnostic model is deployed, combining a health status baseline model to generate a health indicator, and a long short-term memory network is used to predict the remaining service life, completing fault warning and life assessment.

[0043] Construction of S1 and RV reducer multimodal sensing and multi-condition fault dataset:

[0044] This step aims to integrate and form a structured multimodal time-series dataset with operating conditions and fault labels, providing a data foundation for subsequent model training. The flowchart for this step is as follows: Figure 2 As shown.

[0045] S1.1 Multimodal Sensor Deployment and Data Acquisition: On a standard RV reducer fault diagnosis test bench, multi-physics sensors were deployed on key components of the reducer under test to achieve comprehensive information sensing. First, accelerometers were installed radially and axially near the input and output flanges of the reducer housing to collect vibration timing data. Second, high-precision Hall current sensors were installed at the three-phase AC input of the servo motor driving the RV reducer to collect the three-phase current timing signal of the motor. Finally, acoustic emission sensors were attached to the surface of the reducer housing, near the main load-bearing bearings, to collect high-frequency elastic stress wave data.

[0046] S1.2 Multi-condition Experimental Matrix Design: To ensure the model can adapt to the real and varied operating conditions of the reducer in precision transmission applications, this step simulates typical operating condition combinations on a test bench. Using a servo motor controller and magnetic powder brake, the RV reducer is controlled to operate stably at various speeds, including low speed (500 rpm), medium speed (1000 rpm), and high speed (1500 rpm), and various loads, including light load (20% of rated torque), medium load (50%), and heavy load (80%). First, fault implantation and data acquisition are performed. Under each of the above speed and load combinations, multi-modal sensor time-series data of the RV reducer are collected at different health states, and each data point is precisely labeled with its corresponding operating condition label and health state label:

[0047] 1. Normal health status: Collect baseline data of brand new or well-functioning healthy reducers under all operating conditions;

[0048] 2. Typical Fault Conditions: Typical faults of the RV reducer are prefabricated on replaceable parts using electrical discharge machining, artificial wear, and fatigue testing. These faults include: (1) Planetary gear train faults: pitting on the planetary gear tooth surface, tooth root cracks, or tooth breakage. (2) Cycloidal gear train faults: uneven wear of the cycloidal tooth profile, raceway spalling of the crank bearing, or damage to the rolling elements. (3) Support bearing faults: inner ring faults, outer ring faults, or rolling element faults of the main support bearing at the input or output end of the reducer.

[0049] S1.3 Data Integration: This step integrates the data from S1.1 and S1.2, ultimately outputting a structured multimodal time-series dataset with operating condition and fault labels. Specifically, it includes two types of data: multimodal time-series data, three-phase current time-series signals, and high-frequency elastic stress wave data. The label data includes operating condition labels and fault labels. The operating condition labels indicate the specific speed and load corresponding to each time-series data acquisition. The fault labels indicate the reducer health status corresponding to each time-series data, including normal health status, planetary gear train fault, cycloidal gear train fault, and support bearing fault.

[0050] S2. Signal preprocessing and multimodal feature map construction:

[0051] This step aims to preprocess and feature-engineer the multimodal time-series dataset output from S1.3. First, the original one-dimensional time-series signal is segmented using a fixed-length sliding window. Continuous wavelet transforms are performed on the vibration, current, and acoustic emission signals to generate corresponding two-dimensional time-frequency maps. Second, a channel stacking strategy is used to fuse the three time-frequency maps into a unified three-channel time-frequency feature map. The final output is a preprocessed feature dataset containing the fused three-channel feature map and corresponding labels, preparing it for input to the deep model. The flowchart for this step is shown below. Figure 3 As shown.

[0052] S2.1 Time Series Data Segmentation and Sample Construction: This step first segments the multimodal time series dataset from S1.3 using a fixed-length overlapping sliding window method, targeting the three raw signals: vibration time series data, three-phase current time series signals, and high-frequency elastic stress wave data. Specifically, a sampling window of fixed length T is set, with an overlap rate of M, and the multimodal time series data is synchronously slid-cut. Each slide extracts three synchronous signal segments from the three signals. These three signal segments are combined to form a raw multimodal sample. Simultaneously, based on the label data in the multimodal time series dataset, the raw multimodal sample is assigned the operating condition label and fault label corresponding to its acquisition time.

[0053] S2.2 Time-Frequency Analysis Based on Continuous Wavelet Transform: Considering that the RV reducer operates under various speed conditions and that faults typically exhibit non-stationary transient impact characteristics, this step employs time-frequency analysis technology to convert one-dimensional signal segments into two-dimensional images. Specifically, for each of the three one-dimensional signal segments in each original multimodal sample from S2.1, derived from vibration time-series data, three-phase current time-series signals, and high-frequency elastic stress wave data respectively, continuous wavelet transform is performed. Specifically, Morlet wavelets are selected to convert each one-dimensional signal segment into a high-resolution two-dimensional time-frequency plot. This two-dimensional time-frequency plot can clearly display the distribution of signal energy and transient characteristics in both time and frequency dimensions, which is more conducive to the subsequent model's automatic extraction of deep fault features compared to the original time-series signal. This step ultimately yields three two-dimensional time-frequency plots, including a vibration time-frequency plot, a current time-frequency plot, and an acoustic emission time-frequency plot.

[0054] S2.3 Multimodal Feature Map Fusion: This step fuses the separate two-dimensional time-frequency maps derived 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 maps corresponding to the same original multimodal sample in S2.1—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. 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.

[0055] S2.4 Preprocessed Feature Dataset Output: This step ultimately outputs a structured preprocessed feature dataset, which will serve as the final input to the S3 deep learning model. This 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 maps generated in S2.3, and the label data includes the labels retained in S2.1 for each three-channel fused time-frequency map. Figure 1 One corresponding operating condition label and fault label.

[0056] S3. Construction of a condition-adaptive fault diagnosis model based on domain adversarial attention networks:

[0057] This step is the core of the invention. It constructs a domain adversarial attention neural network consisting of a feature extractor, a fault classifier, and a domain classifier. This step enhances key feature extraction through an attention mechanism and utilizes a gradient inversion layer to achieve adversarial training of the feature layer under different operating conditions, enabling the model to maintain high fault classification accuracy while possessing operating condition invariance. The training data for this step is the preprocessed feature dataset output by S2.4. After the adversarial training converges, the domain classifier is stripped, retaining only the feature extractor and fault classifier used for inference, generating an online diagnostic model.

[0058] S3.1 Domain Adversarial Attention Model Architecture Design: This step constructs a domain adversarial neural network architecture consisting of three core components: a feature extractor, a fault classifier, and a domain classifier. The domain adversarial attention model architecture and training loss function are as follows: Figure 4 As shown.

[0059] 1. Feature Extractor: Composed of a deep convolutional attention network, its input is the three-channel fused time-frequency map from S2.4. Its structure consists of three convolutional blocks, a channel and spatial attention module, and a global average pooling layer. First, convolutional block one contains a 3x3 2D convolutional layer using 32 filters and a ReLU function, followed by a 2x2 max-pooling layer; convolutional block two contains a 3x3 2D convolutional layer using 64 filters and a ReLU activation function, followed by a 2x2 max-pooling layer; convolutional block three contains a 3x3 2D convolutional layer using 128 filters and a ReLU activation function, followed by a 2x2 max-pooling layer. Next, the feature map output from convolutional block three is fed into an attention mechanism consisting of a channel attention module and a spatial attention module, arranged sequentially. First, within the channel attention module, the input feature map undergoes parallel pooling, including global average pooling and global max pooling, and then is processed by a multilayer perceptron. Its output is then processed by a sigmoid activation function to generate channel attention weights. These weights are element-wise multiplied with the original input feature map to obtain a channel-weighted feature map. Second, this channel-weighted feature map enters the spatial attention module, where it is concatenated after channel pooling and then processed by a convolutional layer. Its output is again processed by a sigmoid activation function to generate spatial attention weights. Finally, these spatial attention weights are element-wise multiplied 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.

[0060] 2. Fault Classifier: Its input is the high-dimensional feature vector output by the feature extractor. Its structure consists of a fully connected layer and an output layer. The fully connected layer contains 256 neurons, uses the ReLU activation function, and includes a dropout layer to prevent overfitting. The output layer contains 4 neurons, corresponding to the four categories of fault labels in S1.3: normal healthy state, planetary gear train fault, cycloidal gear train fault, and support bearing fault. This layer uses the Softmax activation function and outputs a probability distribution corresponding to the fault label.

[0061] 3. Domain Classifier: This classifier is a work condition predictor. Its input is connected to the feature vector output by the feature extractor through a gradient inversion layer. Its structure consists of 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 representing the total number of work condition combinations, equal to the total number of unique combinations of work condition labels in S1.3. This layer uses the Softmax activation function and outputs a probability distribution corresponding to the work condition label.

[0062] S3.2 Adversarial Training Strategy and Optimization: This step utilizes the preprocessed feature dataset from S2.4 to perform end-to-end adversarial training on the model defined in S3.1. The training process optimizes a composite loss function containing two adversarial objectives: minimizing the fault classification loss and maximizing the domain classification loss. Specifically, the fault classification loss calculates the difference between the fault classifier's prediction and the true fault label using the classification cross-entropy function. Optimizing this loss updates the parameters of the feature extractor and the fault classifier, forcing the feature extractor to extract feature vectors containing sufficient fault identification information. Simultaneously, the domain classification loss also calculates the difference between the domain classifier's prediction and the true operating condition label using the classification cross-entropy function. During training, the domain classifier itself is optimized to minimize this domain classification loss, enabling it to accurately distinguish operating conditions; while the gradient reversal layer, during backpropagation, reverses the sign of the gradient returned by the domain classifier before passing it to the feature extractor. This makes the adversarial objective of the feature extractor to maximize this domain classification loss. This adversarial approach forces the feature vectors extracted by the feature extractor to actively discard operating condition-related information, making it impossible for the domain classifier to discern their origin. Ultimately, the overall optimization function of the feature extractor is expressed as the fault classification loss minus the adversarial balance coefficient multiplied by the domain classification loss, thereby simultaneously achieving high fault classification accuracy and strong operating condition invariance.

[0063] S3.3 Consolidation and Output of the Online Fault Diagnosis Model: After the adversarial training in S3.2 converges, i.e., the fault classification loss stably converges to a minimum, and the domain classifier's classification accuracy for the working conditions approaches random guessing, the trained domain adversarial attention model architecture in S3.1 is consolidated. Specifically, components used only for training, including the domain classifier and its associated gradient inversion layer, are removed from the computation graph and discarded. Only the trained and optimized working condition adaptive feature extractor and the parameter weights of the fault classifier are retained and exported. The cascade of these two components—the feature extractor connected to the fault diagnosis classifier—is encapsulated and saved as an independent model dedicated to forward inference. This model is the final output of the online fault diagnosis model of this invention and is delivered to subsequent steps for online deployment.

[0064] S4. Online diagnostic deployment and health status assessment:

[0065] This step aims to deploy the online fault diagnosis model output by S3.3 into practical applications. It combines real-time acquired multimodal signals to generate a real-time three-channel fused time-frequency map, enabling immediate fault identification and alarm. Simultaneously, a health indicator is generated by calculating the reconstruction error of the input samples using a health status baseline variational autoencoder to monitor the equipment degradation process. Based on the time series of the health indicator, a long short-term memory network is used to predict its future trend, and the remaining service life of the reducer is determined through a failure threshold. A schematic diagram of this step is shown below. Figure 5 As shown.

[0066] S4.1 Health Status Baseline Model Construction: This step aims to construct a baseline model for quantifying health status. First, health data is screened. From the preprocessed feature dataset output in S2.4, three-channel fused feature data with fault labels indicating normal health status is selected from all labeled data. Then, the baseline model is trained, constructing a variational autoencoder model and training it using only the selected health three-channel fused time-frequency plots. The model structure is as follows: This model consists of a convolutional encoder and a transposed convolutional decoder. The convolutional encoder receives a fused three-channel time-frequency image of health as input. Its structure consists of three convolutional blocks, identical to the structure of convolutional blocks one, two, and three in the feature extractor in S3.1, compressing the time-frequency image. Then, a fully connected layer outputs the parameters of the latent distribution, namely the mean vector and log-variance vector corresponding to each latent dimension. The decoder samples a latent vector from this latent distribution, maps it back to a high-dimensional feature space through a fully connected layer, and then sequentially upsamples it through three transposed convolutional blocks symmetrical to the encoder's convolutional block structure, ultimately reconstructing a fused three-channel time-frequency image of the same size as the original input time-frequency image. After training, the model possesses the ability to fit the feature distribution of health data, reproducing input data under normal health conditions with extremely low reconstruction error. This step outputs the trained baseline model of the health state.

[0067] S4.2 Online Diagnosis and Health Indicator Generation: This step is the real-time operation phase of the online fault diagnosis model and health status baseline model, deployed on edge computing devices or industrial host computers. This step first executes a real-time data processing flow: vibration time-series data, three-phase current time-series signals, and high-frequency elastic stress wave data are acquired in real time using the accelerometer, Hall current sensor, and acoustic emission sensor deployed in S1.1. Subsequently, the acquired real-time data stream is processed strictly according to the complete preprocessing flow defined in S2.1 time-series data segmentation, S2.2 time-frequency analysis, and S2.3 multimodal feature map fusion, continuously generating a real-time three-channel fused time-frequency map. Next, real-time fault diagnosis is performed. The generated real-time three-channel fused time-frequency map is input into the online fault diagnosis model solidified in S3.3. Its fault classifier component immediately outputs a probability distribution corresponding to the fault label, thereby achieving immediate fault diagnosis and early warning. While performing real-time fault diagnosis, a health indicator is generated by inputting the same real-time three-channel fused time-frequency map into the health state baseline model trained in S4.1, and calculating the reconstruction error of the variational autoencoder for the current input. This reconstruction error value is defined as the quantized health indicator, which is close to zero when the reducer is in a normal healthy state, but rises significantly when a typical fault condition occurs.

[0068] S4.3 Remaining Service Life Prediction: This step aims to predict the future failure risk of the reducer based on the historical trends of the health indicators generated in S4.2. First, a degradation sequence is constructed. During the online diagnostic process in S4.2, continuously generated quantified health indicator values ​​are recorded and stored chronologically, forming a health indicator degradation time series reflecting the reducer's transition from health to degradation. Next, remaining service life prediction is performed by constructing and pre-training a Long Short-Term Memory (LSTM) network model. This model is specifically designed to learn and predict the future trends of non-stationary time series. Using the constructed health indicator degradation time series as input, the LSM network model automatically extrapolates the future growth trend of the sequence. By setting a predefined failure threshold—that is, when the health indicator value exceeds this threshold, indicating that the reducer can no longer meet service requirements—the future time required for the predicted health indicator growth curve to reach this failure threshold is calculated. This time is output as the reducer's current remaining service life.

[0069] Simulation experiment:

[0070] 1. Simulation Experiment Setup: An experimental platform was built based on a standard RV reducer fault diagnosis test bench. The core parameters are as follows: RV reducer model: RV-40E, transmission ratio: 1:81, rated torque: 50 N·m; servo motor: 1.5kW permanent magnet synchronous motor, speed range: 0-2000 rpm, supporting closed-loop speed control; loading device: magnetic powder brake, load adjustment range: 0-80% of rated torque; sensor configuration: accelerometer: PCB 352C33, range: ±50g, sampling frequency: 20kHz, radial / axial arrangement; Hall current sensor: ACS758, accuracy: ±1.5%, sampling frequency: 10kHz, collecting three-phase motor current; acoustic emission sensor: PAC R15α, resonant frequency: 150kHz, sampling frequency: 1MHz, attached to the housing near the bearing.

[0071] The experimental data covers nine operating condition combinations with three speeds of 500 rpm, 1000 rpm, and 1500 rpm, and three loads of 20%, 50%, and 80% of the rated torque. It includes four health states: normal, planetary gear train failure, cycloidal gear train failure, and support bearing failure. 1000 time-series samples were collected for each state, with each sample lasting 10 seconds.

[0072] 2. Experimental parameter settings: 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 plot resolution is 256×256 pixels; Feature map fusion: three channels are stacked, including vibration time-frequency plot, flow time-frequency plot and acoustic emission time-frequency plot;

[0073] Model training parameters: Hardware environment: NVIDIA 4090 GPU (24GB VRAM); Software environment: Python 3.8, PyTorch 1.12, Scikit-learn 1.2, Matplotlib 3.7;

[0074] Domain adversarial attention network parameters: Feature extractor: 3 convolutional blocks, each containing 3×3 convolutions, ReLU and 2×2 max pooling, and a channel-space attention module, all implemented in PyTorch; Fault classifier: Fully connected layer containing 256 neurons, ReLU activation function and Dropout, with Dropout deactivation parameter set to 0.5, output layer containing 4 neurons and Softmax; Domain classifier: Fully connected layer containing 256 neurons and ReLU, output layer containing 9 neurons and Softmax; Training parameters: Batch size 32, learning rate 1e-4, Adam optimizer, adversarial balance coefficient 0.3, training epochs 100;

[0075] Variational autoencoder: a structure symmetrical to the convolutional blocks of the feature extractor, with a latent dimension of 64, a reconstruction loss of MSE, and 80 training epochs;

[0076] LSTM: 2-layer LSTM, with 128 hidden neurons and a fully connected layer, a time step of 30, a prediction step of 10, and 50 training epochs.

[0077] 3. Comparison of model parameters: Model 1: CNN classifier trained only with vibration time-frequency graphs; Model 2: Vibration, current, and acoustic emission time-frequency graphs simply stitched together and input into CNN; Model 3: Multimodal CNN directly trained; Model 4: Domain adversarial model without attention mechanism.

[0078] 4. Simulation Experiment Results

[0079] Table 1 Experimental Results

[0080]

[0081] The experimental results are shown in Table 1 and Figure 6 As shown, the ablation experiment results demonstrate that the method of this invention has the highest classification accuracy, proving the effectiveness of the network structure and training strategy of this invention.

[0082] Fault classification confusion matrix as follows Figure 7 As shown in the figure, the present invention reduces the number of misclassified samples, proving that the present invention can effectively distinguish the characteristics of four types of faults.

[0083] Figure 8The invention demonstrates the health status evolution of the RV reducer over its 1000-hour lifespan. The fault warning point is triggered at 692 hours, 208 hours earlier than the 900-hour time for complete equipment failure. This provides ample time for developing maintenance plans and procuring spare parts, proving that the method of this invention can effectively maintain equipment health and reduce the risk of RV reducer damage.

[0084] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0085] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An online fault diagnosis method for RV reducers, characterized in that, Includes the following processes: S1, for RV reducers, collects multi-modal sensor timing data in real time under the current speed and load conditions, including 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 fused time-frequency feature map; S3, input the three-channel fused 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. 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 transformed into a fixed-length high-dimensional feature vector through a global average pooling layer. The fault classifier takes a high-dimensional feature vector output by the feature extractor as input. Its structure consists of 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. S4. Input the three-channel fused 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. 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.

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 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.

6. 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.

7. 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