A high-power gearbox operating condition recognition and risk early warning method

Abstract

The application provides a high-power gear box operation condition recognition and risk early warning method, and belongs to the technical field of gear box risk early warning based on deep learning; first, operation data of a high-power gear box under multiple conditions is collected based on multiple sensors, and preprocessed to form a time series data set; then, a multi-branch feature extraction model is designed, and deep features of specific conditions such as broken teeth, bearing faults, axial displacement and tooth side gap changes are extracted respectively under the guidance of prior knowledge; further, a gear box operation condition recognition model is constructed, multi-label condition recognition and confidence evaluation are realized through feature fusion based on an attention mechanism, Transformer enhanced modeling and graph reasoning network; finally, a multi-level dynamic early warning module is designed, and grading early warning is triggered by comprehensively considering instantaneous state, evolution trend and prognosis evaluation, so as to guarantee the safe and stable operation of the high-power gear box. The application improves the recognition accuracy and generalization performance of complex conditions and multiple fault concurrent conditions.

Description

Technical Field

[0001] This invention belongs to the field of gearbox risk warning technology based on deep learning, and particularly relates to a method for identifying the operating conditions and providing risk warnings for high-power gearboxes. Background Technology

[0002] As the core transmission component of a system, the operating status of a high-power gearbox directly affects the reliability, efficiency, and lifespan of the entire system. Common failure modes of high-power gearboxes include broken teeth, bearing failure, and axial displacement. If these failures are not detected and addressed in a timely manner, they may lead to cascading equipment damage, production stoppages, or even safety accidents. Therefore, accurate identification and risk warning of the operating conditions of high-power gearboxes are key technologies for achieving predictive maintenance, avoiding sudden downtime, and improving operational safety.

[0003] Existing methods for identifying operating conditions and diagnosing faults in high-power gearboxes can be mainly divided into three categories: methods based on physical models, methods based on signal processing and feature extraction, and methods based on machine learning and deep learning.

[0004] The physical model-based approach relies on establishing an accurate mathematical model of the gear transmission system. It simulates the system behavior under normal and fault conditions using dynamic equations, vibration response models, or thermodynamic models. Fault detection and localization are achieved by comparing the differences between actual measured data and model outputs. For example, a stiffness excitation model of the gear pair can be established to analyze the vibration response changes caused by tooth root cracks. While this method has clear physical meaning, it depends heavily on the accuracy of the model. For complex operating conditions and scenarios with multiple coupled faults, model establishment is difficult, computation is complex, and it struggles to adapt to individual differences and changes in the operating environment, limiting its engineering applicability.

[0005] The signal processing and feature extraction-based method analyzes sensor signals such as vibration, temperature, and acoustic emission to extract features in the time domain (e.g., root mean square, kurtosis), frequency domain (e.g., Fourier transform, envelope spectrum), and time-frequency domain (e.g., wavelet transform, empirical mode decomposition). These features are then used to construct a fault index system or to perform state identification using classifiers (e.g., support vector machine, decision tree). For example, the local fault frequency of a bearing can be identified by analyzing the spectral characteristics of the vibration signal. This method does not require a precise physical model, but its effectiveness heavily depends on the quality of feature extraction and expert experience. Furthermore, it lacks the ability to handle early, weak fault features and the fusion of multi-source heterogeneous sensor data, making it difficult to achieve adaptive and accurate multi-label condition identification.

[0006] Machine learning / deep learning-based methods: In recent years, with the development of sensor technology and computing power, data-driven methods have gradually become mainstream. These methods utilize machine learning (such as Random Forest and XGBoost) or deep learning models (such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs)) to automatically learn features and patterns from raw or simply preprocessed data. For example, one-dimensional CNNs can be used to directly extract features from vibration signals for fault classification, or LSTMs can be used to model the temporal dependencies of sensor signals. Deep learning models have demonstrated powerful capabilities in complex pattern recognition, avoiding the limitations of manual feature extraction. However, existing methods still have many shortcomings: most studies focus on single-type sensor data (such as vibration), failing to fully integrate multi-source heterogeneous sensor information (such as temperature fields, oil particles, and displacement); models are mostly single-network structures, making it difficult to specifically adapt to the unique representation requirements of different fault modes; furthermore, existing methods focus primarily on the accuracy of fault classification, while failing to adequately capture the evolutionary trends of operating states, lacking dynamic decision-making mechanisms for risk warning, and struggling to meet the comprehensive requirements of real-time performance, robustness, and interpretability in industrial settings.

[0007] Although existing methods have made progress in gearbox condition monitoring, the following major defects still exist: (1) Insufficient ability to fuse multi-source heterogeneous data, failing to systematically integrate multi-modal sensing information such as vibration, temperature, displacement, and oil, and lacking an effective alignment method for asynchronous sampling data from multiple sensors, which restricts the full characterization of fault features; (2) Limited ability to model time-series dynamics, most methods fail to fully explore the long-range dependencies and dynamic evolution patterns in sensor data, and have low sensitivity to early and progressive faults; (3) Poor model generalization and interpretability, the black-box nature of deep learning makes the diagnostic decision-making process difficult to understand, and the models are often designed for specific equipment or operating conditions, making it difficult to adapt to different operating conditions and equipment variations; (4) Imperfect early warning mechanism, most existing methods focus on identification after the fault occurs or in the severe stage, lacking a multi-level early warning strategy based on confidence assessment and trend prediction, and unable to achieve true early warning and predictive maintenance. Summary of the Invention

[0008] To address the above problems, this invention proposes a method for identifying the operating conditions and providing risk warnings for high-power gearboxes, comprising the following steps:

[0009] S1 collects real-time operating data of a high-power gearbox based on multiple types of sensors, including bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data.

[0010] S2. Input the data collected in S1 into the trained multi-branch feature extraction model. First, extract the time-dependent depth features of the multi-source data of the gearbox working condition based on the gearbox working condition feature extraction backbone network. Then, share the time-dependent depth features of the multi-source data based on the multi-branch extraction network and output the time-series enhancement features of the broken tooth working condition, the depth features of the bearing failure working condition, the depth features of the axial displacement working condition, and the depth features of the tooth flank clearance change working condition.

[0011] S3. Input the four features obtained in S2 into the trained gearbox operating condition recognition model. The model uses a multi-feature fusion mechanism and a multi-label classification task modeling method to predict the current operating condition of the gearbox and obtain four prediction results and their corresponding confidence levels: broken teeth, bearing failure, axial displacement and tooth backlash.

[0012] S4, based on the prediction results and confidence level obtained from S3, provides corresponding early warnings through a preset risk warning strategy.

[0013] Preferably, an accelerometer is arranged at the gearbox bearing to collect vibration signal data during bearing operation as bearing vibration signal data; displacement sensors are arranged on the input and output shafts of the gearbox to collect axial displacement data during bearing operation; an accelerometer is arranged at the gear meshing point to collect vibration signal data during gear operation as gear vibration signal data; a thermistor sensor array is arranged at the gearbox bearing to collect temperature data during bearing operation as bearing temperature field data; a thermocouple sensor array is arranged at the gear meshing point to collect temperature data during gear operation as gear temperature field data; and an oil particle counter is installed at the oil outlet of the gearbox to collect the number of metal particles during gearbox operation.

[0014] Preferably, the multi-branch feature extraction model includes a gearbox operating condition feature extraction backbone network, a broken tooth operating condition feature extraction branch network, a bearing failure operating condition feature extraction branch network, an axial displacement operating condition feature extraction branch network, and a tooth flank clearance change operating condition feature extraction branch network.

[0015] The gearbox operating condition feature extraction backbone network takes multi-source data of gearbox operating conditions, consisting of bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data, as input to obtain time-dependent depth features of the multi-source data of gearbox operating conditions, including time-dependent depth features of gear vibration signal data, time-dependent depth features of gear temperature field data, time-dependent depth features of bearing vibration signal data, time-dependent depth features of bearing temperature field data, time-dependent depth features of metal particle quantity data, and time-dependent depth features of axial displacement data.

[0016] The inputs of the tooth breakage condition feature extraction branch network are the time-dependent depth features of gear vibration signal data and the time-dependent depth features of gear temperature field data, and the output is the time-series enhancement features of the tooth breakage condition.

[0017] The bearing fault condition feature extraction branch network takes as input the time-dependent depth features of bearing vibration signal data, bearing temperature field data, and metal particle quantity data, and outputs the bearing fault condition depth features.

[0018] The input of the axial displacement condition feature extraction branch network is the time-dependent depth feature of bearing vibration signal data, and the output is the axial displacement condition depth feature.

[0019] The input of the tooth backlash variation feature extraction branch network is the time-dependent depth feature of gear vibration signal data, and the output is the depth feature of tooth backlash variation.

[0020] Preferably, the gearbox operating condition feature extraction backbone network includes a feature alignment layer and two sequentially connected feature modeling networks. The feature modeling network includes a fully connected layer, a long short-term memory network layer, and a ReLU activation function layer. First, the multi-source data of gearbox operating conditions, consisting of bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data, is input into the feature alignment layer for time-scale feature alignment. The feature alignment layer uses an autoregressive differential moving average model to align the multi-source data of gearbox operating conditions in the time dimension. Second, the aligned gearbox operating conditions... Multi-source data is input into the first feature modeling network for feature extraction. After feature extraction in the fully connected layer, the data is input into the ReLU activation function layer for feature truncation, discarding feature points with a feature value of 0. Then, the output of the ReLU activation function layer is input into the long short-term memory network layer. The long short-term memory network layer is used to model the time dependency of the features, capturing the potential forward time dependency of each working condition data in the multi-source data of gearbox working conditions, and obtaining the time dependency features of the multi-source data of gearbox working conditions. The time dependency features output by the first feature modeling network are input into the second feature modeling network to obtain the time dependency depth features of the multi-source data of gearbox working conditions.

[0021] Preferably, the gearbox operating condition identification model includes a feature weighted fusion network, an operating condition perception enhancement network, and a multi-label operating condition identification network;

[0022] The feature-weighted fusion network, serving as the pre-layer of the gearbox operating condition recognition model, receives temporal enhancement features of tooth breakage, bearing fault condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features from the multi-branch feature extraction model. Based on the attention mechanism and feature weighting, it obtains attention features for tooth breakage, bearing fault, axial displacement, and tooth flank clearance change conditions. The attention features of each condition output from the feature-weighted fusion network are then input into the operating condition perception enhancement network, and processed... A stacked Transformer encoder performs layer-by-layer feature enhancement on the working condition attention features, outputting enhanced features for tooth breakage, bearing failure, axial displacement, and tooth flank clearance changes. The four enhanced features are input into a multi-label working condition recognition network, which outputs the working condition prediction results and their confidence scores. The multi-label working condition recognition network first constructs a working condition association graph, and then uses a graph attention network for working condition recognition and prediction.

[0023] Preferably, the feature weighted fusion network includes a feature splicing layer, a self-attention mechanism layer, and a Softmax activation function layer. First, the obtained temporal enhancement features of broken tooth conditions, bearing fault conditions, axial displacement conditions, and tooth flank clearance change conditions are input into the feature splicing layer for feature splicing to form an aggregated feature matrix. Second, the aggregated feature matrix is ​​input into the self-attention mechanism layer. The self-attention mechanism uses three learnable weight matrices to map the input aggregated feature matrix into three different representation spaces: query Q, key K, and value V. By calculating the dot product of Q and K and performing a scaling operation, the temporal enhancement features of broken teeth are obtained. Attention scores are calculated among the features, bearing failure condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features. These attention scores are then input into a Softmax activation function layer for normalization, generating attention weights ranging from 0 to 1, including attention weights for tooth breakage time-series enhancement features, bearing failure condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features. Finally, the four condition features are dot-producted with their corresponding weights to obtain the attention features for tooth breakage condition, bearing failure condition, axial displacement condition, and tooth flank clearance change condition.

[0024] Preferably, the condition perception enhancement network includes A stacked Transformer encoder, each encoder layer containing a multi-head attention mechanism layer and a feedforward neural network layer; the multi-head attention mechanism layer interacts with information between various dimensions within each working condition attention feature, learning the dependencies between features from different representation subspaces through multiple parallel attention heads, enhancing the expressive power of the working condition attention features, and obtaining working condition context-aware features; the feedforward neural network layer performs a nonlinear transformation on the context-aware features output by the multi-head attention mechanism layer, enhancing the model's expressive power; after... The stacked Transformer encoders perform layer-by-layer feature enhancement on the working condition attention features. The working condition perception enhancement network finally outputs working condition perception enhancement features for broken teeth, bearing failure, axial displacement, and tooth backlash change.

[0025] Preferably, the multi-label working condition recognition network first constructs a working condition association graph. In the working condition graph, each node represents a working condition label. The initial side length between nodes is the attention score between the working conditions. The output enhanced features of broken tooth working condition perception, enhanced features of bearing fault working condition perception, enhanced features of axial displacement working condition perception, and enhanced features of tooth flank clearance change working condition perception are used as query information. Each working condition node in the graph is given an initial feature representation through linear transformation.

[0026] The working condition association graph is input into a graph attention network for inference. During the inference process, the graph attention network continuously updates the side lengths and features of each node in the working condition graph, thereby explicitly quantifying the instantaneous mutual influence strength between various working conditions. After information propagation and aggregation through multiple layers of graph attention network, each node obtains a final feature representation that integrates relevant working condition information. Finally, these updated node features are passed through an independent linear layer and a sigmoid activation function to output the working condition prediction results and their confidence scores.

[0027] Preferably, the operating condition prediction result is a four-dimensional vector. Each dimension of the vector stores the prediction result and confidence level of four operating conditions: tooth breakage, bearing failure, axial displacement, and tooth backlash. If the operating condition prediction result of a certain dimension is 1, it indicates that the gearbox is currently in the operating condition corresponding to that dimension of the vector. If it is 0, it indicates that the gearbox is not currently in the operating condition corresponding to that dimension of the vector. The confidence level is a value between 0 and 1, which reflects the probability that the model is in a certain operating condition. The larger the value, the higher the probability. If the value of the operating condition prediction result is 0 in all dimensions, the gearbox is currently in normal operating condition.

[0028] Preferably, S4 specifically includes:

[0029] A threshold will be set for the confidence level of different operating conditions. When the confidence level of a certain operating condition output by S3 is within a certain time frame... When all values ​​exceed this threshold, the operating condition is identified as an event to be observed that requires the initiation of an early warning process;

[0030] If a certain operating condition is identified as an event to be observed, then first calculate the time of that operating condition. The gradient of confidence level for each operating condition is used to determine the operating time based on the gradient. The rate of change of confidence within a given period; secondly, the operating condition over time. The operating condition confidence input to the long short-term memory network output future to The predicted confidence values ​​of the operating condition at multiple time points; if the predicted confidence value of the operating condition at any time point exceeds the threshold of interest, and the operating condition is within the time frame... An alarm is triggered if the gradient of the confidence level of the working condition exceeds a preset threshold.

[0031] The innovative aspects of this invention include:

[0032] (1) By deploying various sensors such as vibration, displacement, temperature, and oil particles to collect gearbox operation data, a high-quality time-series dataset covering five typical working conditions is constructed. This method ensures the comprehensiveness of the data and the consistency of the time series, providing a foundation for subsequent deep feature extraction.

[0033] (2) Multi-branch temporal feature extraction architecture: Based on the physical prior knowledge of different fault modes of gearbox, a dedicated branch network is designed to extract deep temporal features of heterogeneous multi-source sensor data, so as to realize the targeted enhancement and representation of fault features and overcome the limitation of traditional single model in adapting to multiple types of fault features.

[0034] (3) Multi-label working condition identification mechanism based on attention and graph reasoning: The feature weighted fusion network is used to integrate multi-branch features, and cross-dimensional feature interaction is achieved through the improved Transformer encoder. Furthermore, the graph attention network is introduced to explicitly model the correlation between working condition labels, thereby improving the identification accuracy and interpretability of multiple concurrent faults.

[0035] (4) Multi-level early warning strategy that integrates instantaneous confidence and evolution trend: A dual confirmation mechanism based on confidence duration and gradient change is proposed, and confidence prediction is performed by combining Long Short-Term Memory Network (LSTM) to realize the upgrade of early warning from state recognition to trend judgment, effectively suppressing false alarms and improving the reliability of early warning.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) Multi-branch temporal feature extraction architecture: Based on the physical prior knowledge of different fault modes of large gearbox, a dedicated branch network is designed to extract deep temporal features of heterogeneous multi-source sensor data, so as to realize the targeted enhancement and representation of fault features and overcome the limitation of traditional single model in adapting to multiple types of fault features.

[0038] (2) Multi-label working condition identification mechanism based on attention and graph reasoning: The feature weighted fusion network is used to integrate multi-branch features, and cross-dimensional feature interaction is achieved through the improved Transformer encoder. Furthermore, the graph attention network is introduced to explicitly model the correlation between working condition labels, thereby improving the identification accuracy and interpretability of multiple concurrent faults.

[0039] (3) Multi-level early warning strategy that integrates instantaneous confidence and evolution trend: A dual confirmation mechanism based on confidence duration and gradient change is proposed, and confidence prediction is performed by combining Long Short-Term Memory Network (LSTM) to realize the upgrade of early warning from state recognition to trend judgment, effectively suppressing false alarms and improving the reliability of early warning. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0041] Figure 2 This is a diagram illustrating the overall framework of a multi-branch feature extraction model.

[0042] Figure 3 A schematic diagram of the gearbox operating condition identification model structure.

[0043] Figure 4 This is a schematic diagram of a multi-level dynamic early warning process for a gearbox.

[0044] Figure 5 This is a comparison chart showing the recognition delay of early faults by different models in the embodiments.

[0045] Figure 6 This is a performance comparison chart of different models in the gearbox condition recognition task in the embodiment.

[0046] Figure 7 The image shows a radar chart illustrating the early warning performance in this embodiment.

[0047] Figure 8 This is a confusion matrix diagram for a multi-fault concurrent scenario in the embodiment. Detailed Implementation

[0048] This invention proposes a method for identifying the operating conditions and providing early warning of risks in high-power gearboxes based on temporal deep networks, aiming to achieve accurate perception of the operating status of high-power gearboxes and proactive warning of fault risks. The overall process is as follows: Figure 1As shown: First, operating data of a high-power gearbox under multiple operating conditions is collected based on multiple types of sensors, and missing values ​​are filled in using interpolation methods to form a high-quality time-series dataset. Then, a multi-branch feature extraction model is designed to extract deep features of specific operating conditions such as broken teeth, bearing failure, axial displacement, and tooth flank clearance changes, guided by prior knowledge. Next, a gearbox operating condition identification model is constructed, which realizes multi-label operating condition identification and confidence assessment through feature fusion based on attention mechanism, Transformer enhanced modeling, and graph inference network. Finally, a multi-level dynamic early warning module is designed to trigger graded early warnings by comprehensively considering instantaneous status, evolution trend, and prognostic assessment to ensure the safe and stable operation of the high-power gearbox.

[0049] The specific implementation process of the present invention will be described in detail below with reference to specific embodiments.

[0050] I. Data Acquisition for High-Power Gearbox Operating Condition Identification

[0051] Gearbox operating conditions are categorized into five types: normal, broken tooth, bearing failure, axial displacement, and tooth backlash variation. Broken tooth manifests as sudden impact vibration and abnormal temperature rise; bearing failure is characterized by high-frequency vibration, localized temperature changes, and wear particles in the oil; axial displacement is identified through vibration analysis and axial displacement sensors; tooth backlash variation affects the gear meshing process, leading to increased vibration and noise. Guided by the above prior knowledge, this invention designs the following data acquisition method to cover the five operating conditions of the gearbox.

[0052] Accelerometers are installed at the gearbox bearings to collect vibration signal data during bearing operation, which is recorded as bearing vibration signal data. Displacement sensors are installed on the input and output shafts of the gearbox to collect axial displacement data during bearing operation. Accelerometers are installed at the gear meshing points to collect vibration signal data during gear operation, which is recorded as gear vibration signal data. A thermistor sensor array is installed at the gearbox bearings to collect temperature data during bearing operation, which is recorded as bearing temperature field data. A thermocouple sensor array is installed at the gear meshing points to collect temperature data during gear operation, which is recorded as gear temperature field data. An oil particle counter is installed at the oil outlet of the gearbox to collect data on the number of metal particles during gearbox operation.

[0053] Missing values ​​in the collected data are filled in, specifically by using cubic spline interpolation to fill in the missing values ​​in the working condition data.

[0054] II. Design of Multi-branch Feature Extraction Model

[0055] Guided by prior knowledge that "tooth breakage utilizes gear vibration signal data and gear temperature field data; bearing failure utilizes bearing vibration signal data, bearing temperature field data, and metal particle quantity data; axial displacement utilizes bearing vibration signal data and axial displacement data; and tooth flank clearance variation utilizes gear vibration signal data," a multi-branch feature extraction model was designed. This model includes a main network for gearbox operating condition feature extraction, branch networks for tooth breakage, bearing failure, axial displacement, and tooth flank clearance variation. All branch networks share the operating condition features extracted by the main network. The overall structure is as follows: Figure 2 As shown.

[0056] 1. Gearbox Operating Condition Feature Extraction Backbone Network: To effectively extract common features from multi-source data on gearbox operating conditions, this invention constructs a gearbox operating condition feature extraction backbone network, including a feature alignment layer and two sequentially connected feature modeling networks. The feature modeling network includes a fully connected layer, a long short-term memory (LSTM) network layer, and a ReLU activation function layer. First, the multi-source data on gearbox operating conditions, consisting of bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data, is input into the feature alignment layer for time-scale feature alignment. The feature alignment layer uses an autoregressive differential moving average (ARMA) model to align the multi-source data on gearbox operating conditions in the time dimension, solving the problem of data misalignment caused by differences in sensor sampling frequencies. Second, the aligned multi-source data on gearbox operating conditions is input into the first feature modeling network for feature extraction. After feature extraction, the data is input into a ReLU activation function layer for feature truncation, discarding feature points with a feature value of 0. The output of the ReLU activation function layer is then input into a Long Short-Term Memory (LSTM) network layer. The LSM network layer models the temporal dependencies of the features, capturing the potential forward temporal dependencies of each type of data in the multi-source data of gearbox operating conditions, thus obtaining the time dependency features of the multi-source data of gearbox operating conditions. The time dependency features output from the first feature modeling network are input into a second feature modeling network to obtain the time dependency depth features of the multi-source data of gearbox operating conditions. These time dependency depth features of the multi-source data of gearbox operating conditions consist of the time dependency depth features of gear vibration signal data, gear temperature field data, bearing vibration signal data, bearing temperature field data, metal particle quantity data, and axial displacement data.

[0057] 2. Tooth Breakage Condition Feature Extraction Branch Network: To effectively fuse multi-source deep features related to tooth breakage conditions, this invention constructs a tooth breakage condition feature extraction branch network. Its input consists of the time-dependent deep features of gear vibration signal data and gear temperature field data output from the gearbox condition feature extraction backbone network. This branch network comprises a feature fusion layer and a temporal feature enhancement module. First, the two deep features are input into the feature fusion layer, where they are concatenated through a fully connected layer and nonlinearly transformed using the ReLU activation function to obtain the fused preliminary tooth breakage features. Then, the fused features are input into the temporal feature enhancement module, which is composed of a fully connected layer, a bidirectional long short-term memory network layer, and a ReLU activation function layer connected in series. The fully connected layer further extracts higher-order features, the ReLU function removes redundant information, and the Bi-LSTM layer simultaneously captures forward and backward temporal dependencies, outputting the temporal enhanced features of the tooth breakage condition.

[0058] 3. Bearing Fault Condition Feature Extraction Branch Network: Addressing the multi-source, heterogeneous deep features of bearing faults, this invention designs a bearing fault condition feature extraction branch network. Its inputs include time-dependent deep features of bearing vibration signal data, bearing temperature field data, and metal particle quantity data. This branch network comprises a feature aggregation layer, a gated recurrent unit layer, and a convolutional enhancement module. First, the three deep features are input to the feature aggregation layer, and then weighted and fused through a fully connected layer to output the aggregated preliminary bearing fault features. Subsequently, the preliminary features are input to the gated recurrent unit layer to capture temporal dynamic changes and output the bearing fault temporal features. Finally, the temporal features are input to the convolutional enhancement module to further extract local temporal patterns and enhance feature robustness, outputting the bearing fault condition deep features. The convolutional enhancement module consists of a one-dimensional convolutional layer, a ReLU activation function, and a max-pooling layer connected in series.

[0059] 4. Axial Displacement Condition Feature Extraction Branch Network: To efficiently model the temporal characteristics of axial displacement conditions, this invention constructs an axial displacement condition feature extraction branch network, whose input is the time-dependent deep features of bearing vibration signal data. This branch network adopts a lightweight structure, consisting of two cascaded temporal modeling units and skip connections. Each temporal modeling unit includes a fully connected layer, a long short-term memory network layer, and a ReLU activation function layer: the fully connected layer performs feature transformation, the long short-term memory network layer models the time dependency, and the ReLU function enhances nonlinearity. The first temporal modeling unit outputs primary temporal features, which are added to the output of the second temporal modeling unit through skip connections, and then the dimensions are adjusted by the fully connected layer to finally output the deep features of the axial displacement condition. This design alleviates the gradient vanishing problem in deep networks and improves feature reuse capabilities.

[0060] 5. Gear Backlash Variation Feature Extraction Branch Network: To address the nonlinear temporal characteristics of gear backlash variation, this invention designs a gear backlash variation feature extraction branch network, whose input is the time-dependent depth feature of gear vibration signal data. This branch network consists of an attention mechanism module and a temporal convolutional network module connected in series. First, the input features are fed into the attention mechanism module, which includes a self-attention layer and a fully connected layer. It calculates feature weights to focus on key temporal information and outputs weighted attention features. Subsequently, the weighted attention features are fed into the temporal convolutional network module, which consists of causal convolutional layers, dilated convolutional layers, and ReLU activation function layers. The causal convolutional layers and dilated convolutions capture long-term temporal dependencies, outputting the depth feature of gear backlash variation. This structure avoids recursive computation, improving parallelization efficiency and long-sequence modeling capabilities.

[0061] III. Design of Wheelbox Operating Condition Identification Model

[0062] After the multi-branch feature extraction model generates deep features for each operating condition, this invention designs a gearbox operating condition identification model that integrates features from different branch networks. It employs a multi-feature fusion mechanism and a multi-label classification task modeling method to identify and predict the current operating condition of the gearbox, while providing confidence assessments for each prediction result. The gearbox operating condition identification module uses an attention-based feature weighted fusion network as a front-end layer, receiving outputs from each branch network of the multi-branch feature extraction model, including tooth breakage timing enhancement features, bearing fault operating condition depth features, axial displacement operating condition depth features, and tooth flank clearance change operating condition depth features. Based on the attention mechanism and feature weighting, attention features for tooth breakage, bearing fault, axial displacement, and tooth flank clearance change are obtained. The attention features for each operating condition output by the feature weighted fusion network are input into the operating condition perception enhancement network; after... A stacked Transformer encoder performs layer-by-layer feature enhancement on the operating condition attention features, outputting enhanced features for tooth breakage, bearing failure, axial displacement, and tooth flank clearance changes. These four enhanced features are then input into a multi-label operating condition recognition network, which outputs the operating condition prediction results and their confidence scores. The overall structure is as follows: Figure 3 As shown.

[0063] 1. Feature Weighted Fusion Network Based on Attention Mechanism: The feature weighted fusion network based on attention mechanism includes a feature splicing layer, a self-attention mechanism layer, and a Softmax activation function layer. First, the obtained temporal enhancement features of broken tooth condition, bearing fault condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features are input into the feature splicing layer for feature splicing to form an aggregated feature matrix. Second, the aggregated feature matrix is ​​input into the self-attention mechanism layer. The self-attention mechanism maps the input aggregated feature matrix into three different representation spaces: query Q, key K, and value V, respectively, through three learnable weight matrices. By calculating the dot product of Q and K and performing a scaling operation, attention scores are obtained among the tooth breakage time-series enhancement features, bearing failure condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features. These attention scores are then input into a Softmax activation function layer for normalization, generating attention weights ranging from 0 to 1. These weights include the attention weights for the tooth breakage time-series enhancement features, bearing failure condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features. These weights reveal the model's level of attention to different condition features. Finally, the four condition features are dot-producted with their corresponding weights to obtain the attention features for the tooth breakage condition, bearing failure condition, axial displacement condition, and tooth flank clearance change condition.

[0064] 2. Condition-Aware Enhancement Network: After the feature weighted fusion network outputs attention features for broken teeth, bearing failure, axial displacement, and tooth flank clearance changes, this invention improves the encoder layer (a core component of the Transformer architecture) to create a condition-aware enhancement network. This network enables deep interaction between different dimensions of features, generating feature representations with higher information density and stronger context awareness. The condition-aware enhancement network includes... The model employs a stacked Transformer encoder, with each encoder layer containing a multi-head attention mechanism layer and a feedforward neural network layer. The multi-head attention mechanism layer facilitates information interaction between various dimensions within each working condition attention feature. Through multiple parallel attention heads, it learns the dependencies between features from different representation subspaces, enhancing the expressive power of the working condition attention features and obtaining context-aware features. The feedforward neural network layer further performs nonlinear transformations on the context-aware features output by the multi-head attention mechanism layer, further enhancing the model's expressive power. The stacked Transformer encoders perform layer-by-layer feature enhancement on the working condition attention features. The working condition perception enhancement network finally outputs working condition perception enhancement features for broken teeth, bearing failure, axial displacement, and tooth backlash change.

[0065] 3. Multi-label working condition recognition network based on graph reasoning: To explicitly model the inherent correlations between various working condition features of the gearbox, this invention designs a multi-label working condition recognition network based on graph reasoning. First, a working condition correlation graph is constructed, and then a graph attention network is used for working condition recognition. 1) Constructing the working condition correlation graph: Each node in the working condition graph represents a working condition label. The initial edge length between nodes is the attention score between working conditions obtained by the feature weighted fusion network. The working condition perception enhancement features output by the working condition perception enhancement network, such as broken tooth working condition perception enhancement features, bearing fault working condition perception enhancement features, axial displacement working condition perception enhancement features, and tooth flank clearance change working condition perception enhancement features, are used as query information. Each working condition node in the graph is assigned an initial feature representation through linear transformation. 2) Inputting the working condition correlation graph into the graph attention network for reasoning: During the reasoning process, the graph attention network continuously updates the edge length and features of each node in the working condition graph, thereby explicitly quantifying the instantaneous mutual influence intensity between various working conditions. After information propagation and aggregation through multiple layers of graph attention network, each node obtains a final feature representation that integrates relevant working condition information. Finally, these updated node features are passed through an independent linear layer and a sigmoid activation function to output the operating condition prediction results and their confidence scores. The operating condition prediction results are four-dimensional vectors, with each dimension storing the prediction results and confidence scores for four operating conditions: broken tooth, bearing failure, axial displacement, and tooth flank clearance. A prediction result of 1 in a certain dimension indicates that the gearbox is currently in the operating condition corresponding to that dimension, while a value of 0 indicates that the gearbox is not currently in the operating condition corresponding to that dimension. The confidence score is a value between 0 and 1, reflecting the probability that the model represents a certain operating condition, with a higher value indicating a higher probability. If the value of the operating condition prediction result is 0 in all dimensions, the gearbox is currently in normal operating condition.

[0066] IV. Design of Gearbox Early Warning Module

[0067] After achieving accurate identification and confidence assessment of gearbox operating conditions, a multi-level dynamic early warning system integrating instantaneous status, evolutionary trends, and prognostic evaluation was designed. This system transforms these instantaneous diagnostic conclusions into early warning strategies. The overall process is as follows: Figure 4 As shown.

[0068] 1) Determination of Observable Operating Condition Events: To avoid false alarms caused by instantaneous sensor noise or random model fluctuations, the initial triggering of the warning adopts a dual confirmation mechanism of instantaneous confidence level and duration. Specifically, the system will set a concern threshold for the confidence level of different operating conditions. When the confidence level of a certain operating condition output by S3 exceeds a certain threshold within a certain time frame... When all values ​​exceed this threshold, the operating condition is identified as an event to be observed that requires the initiation of an early warning process;

[0069] 2) Fault Evolution Trend Analysis and Early Warning Method: If a certain operating condition is identified as an event to be observed, then step 2) will be executed. First, the operating condition will be calculated over time. The gradient of confidence level for each operating condition is used to determine the operating time based on the gradient. The rate of change of confidence within a given period; secondly, the operating condition over time. The operating condition confidence input to the long short-term memory network output future to If the predicted confidence values ​​of the operating condition at multiple time points exceed the attention threshold described in S4-1, and the operating condition is within a certain time frame... An alarm is triggered if the gradient of the confidence level of the working condition exceeds a preset threshold.

[0070] V. Analysis of Experimental Results

[0071] The experiment used a 1.5MW wind turbine gearbox as the object, and vibration, displacement, temperature, and oil sensors were arranged according to the patented scheme to collect 100 hours of data, including early and concurrent faults. After ARIMA alignment and interpolation completion, the experiment compared the performance of traditional schemes, including support vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), long short-term memory network (LSTM), and convolutional neural network-long short-term memory network combined model (CNN-LSTM), with that of the present invention.

[0072] like Figure 5 As shown, in the early fault identification scenario of gearboxes, the identification latency of SVM is 15.8 hours, while that of 1D-CNN, LSTM, and CNN-LSTM are 12.5 hours, 8.3 hours, and 5.8 hours, respectively. This invention achieves an average accuracy of 0.85 within 8 hours, with an identification latency of only 2.1 hours, a reduction of 64% compared to CNN-LSTM, effectively solving the problem of "low sensitivity to early faults" mentioned in the patent background.

[0073] like Figure 6 As shown, the present invention outperforms SVM, 1D-CNN, LSTM, and CNN-LSTM in several key performance indicators for evaluating gearbox condition recognition, including accuracy, precision, recall, and F1 score, fully demonstrating its superior performance in gearbox condition recognition tasks.

[0074] like Figure 7 As shown, in terms of early warning time and average accuracy for various types of faults such as bearing failure, broken teeth, axial displacement, and tooth flank clearance, this invention outperforms existing methods such as SVM, 1D-CNN, LSTM, and CNN-LSTM in all dimensions, demonstrating superior gearbox fault early warning performance in all aspects.

[0075] Figure 8 By comparing the confusion matrix of CNN-LSTM and the present invention in a multi-fault concurrent scenario, the present invention has higher accuracy in identifying various faults, lower probability of false identification, and can more accurately deal with the complex situation of multiple faults in the gearbox.

[0076] 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.

[0077] 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. A method for identifying operating conditions and providing risk warnings for high-power gearboxes, characterized in that, Includes the following steps: S1 collects real-time operating data of a high-power gearbox based on multiple types of sensors, including bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data; S2. Input the data collected in S1 into the trained multi-branch feature extraction model. First, extract the time-dependent depth features of the multi-source data of the gearbox working condition based on the gearbox working condition feature extraction backbone network. Then, share the time-dependent depth features of the multi-source data based on the multi-branch extraction network and output the time-series enhancement features of the broken tooth working condition, the depth features of the bearing failure working condition, the depth features of the axial displacement working condition, and the depth features of the tooth flank clearance change working condition. The multi-branch feature extraction model includes a gearbox operating condition feature extraction backbone network, a broken tooth operating condition feature extraction branch network, a bearing failure operating condition feature extraction branch network, an axial displacement operating condition feature extraction branch network, and a tooth flank clearance change operating condition feature extraction branch network. The gearbox operating condition feature extraction backbone network takes multi-source data of gearbox operating conditions, consisting of bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data, as input to obtain time-dependent depth features of the multi-source data of gearbox operating conditions, including time-dependent depth features of gear vibration signal data, time-dependent depth features of gear temperature field data, time-dependent depth features of bearing vibration signal data, time-dependent depth features of bearing temperature field data, time-dependent depth features of metal particle quantity data, and time-dependent depth features of axial displacement data. The inputs of the tooth breakage condition feature extraction branch network are the time-dependent depth features of gear vibration signal data and the time-dependent depth features of gear temperature field data, and the output is the time-series enhancement features of the tooth breakage condition. The bearing fault condition feature extraction branch network takes as input the time-dependent depth features of bearing vibration signal data, bearing temperature field data, and metal particle quantity data, and outputs the bearing fault condition depth features. The input of the axial displacement condition feature extraction branch network is the time-dependent depth feature of bearing vibration signal data, and the output is the axial displacement condition depth feature. The input of the tooth backlash variation working condition feature extraction branch network is the time-dependent depth feature of gear vibration signal data, and the output is the depth feature of tooth backlash variation working condition. S3. Input the four features obtained in S2 into the trained gearbox operating condition recognition model. The model uses a multi-feature fusion mechanism and a multi-label classification task modeling method to predict the current operating condition of the gearbox and obtain four prediction results and their corresponding confidence levels: broken teeth, bearing failure, axial displacement and tooth backlash. The gearbox operating condition identification model includes a feature weighted fusion network, an operating condition perception enhancement network, and a multi-label operating condition identification network. The feature weighted fusion network serves as the front layer of the gearbox operating condition identification model. It receives tooth breakage time sequence enhancement features, bearing fault condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features from the multi-branch feature extraction model. Based on the attention mechanism and feature weighting, it obtains tooth breakage condition attention features, bearing fault condition attention features, axial displacement condition attention features, and tooth flank clearance change condition attention features. The attention features of each working condition output from the feature weighted fusion network are input into the working condition perception enhancement network, and then... A stacked Transformer encoder performs layer-by-layer feature enhancement on the working condition attention features, and outputs enhanced features for tooth breakage, bearing failure, axial displacement, and tooth backlash change. Four enhancement features are input into a multi-label working condition recognition network, which outputs the working condition prediction results and their confidence scores. The multi-label working condition recognition network first constructs a working condition association graph, and then uses a graph attention network for working condition recognition and prediction. S4, based on the prediction results and confidence level obtained from S3, provides corresponding early warnings through a preset risk warning strategy.

2. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 1, characterized in that: Accelerometers are installed at the gearbox bearings to collect vibration signal data during bearing operation. Displacement sensors are installed on the input and output shafts of the gearbox to collect axial displacement data during bearing operation. Accelerometers are installed at the gear meshing points to collect vibration signal data during gear operation. A thermistor sensor array is installed at the gearbox bearings to collect temperature data during bearing operation. A thermocouple sensor array is installed at the gear meshing points to collect temperature data during gear operation. An oil particle counter is installed at the oil outlet of the gearbox to collect the number of metal particles during gearbox operation.

3. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 1, characterized in that: The gearbox operating condition feature extraction backbone network includes a feature alignment layer and two sequentially connected feature modeling networks. The feature modeling networks consist of a fully connected layer, a long short-term memory network layer, and a ReLU activation function layer. First, multi-source data on gearbox operating conditions, comprising bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data, and metal particle quantity data, is input into the feature alignment layer for time-scale feature alignment. The feature alignment layer uses an autoregressive differential moving average model to align the multi-source data on gearbox operating conditions along the time dimension. Second, the aligned multi-source data on gearbox operating conditions... The data is input into the first feature modeling network for feature extraction. After feature extraction in the fully connected layer, the data is input into the ReLU activation function layer for feature truncation, discarding feature points with a feature value of 0. Then, the output of the ReLU activation function layer is input into the long short-term memory network layer. The long short-term memory network layer is used to model the time dependency of the features, capturing the potential forward time dependency of each type of gearbox operating condition data from multiple sources, and obtaining the time dependency features of the gearbox operating condition data from multiple sources. The time dependency features output from the first feature modeling network are input into the second feature modeling network to obtain the time dependency depth features of the gearbox operating condition data from multiple sources.

4. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 1, characterized in that: The feature weighted fusion network comprises a feature splicing layer, a self-attention mechanism layer, and a Softmax activation function layer. First, the obtained temporal enhancement features of broken tooth condition, bearing fault condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features are input into the feature splicing layer for feature splicing, forming an aggregated feature matrix. Second, the aggregated feature matrix is ​​input into the self-attention mechanism layer. The self-attention mechanism uses three learnable weight matrices to map the input aggregated feature matrix into three different representation spaces: query Q, key K, and value V. By calculating the dot product of Q and K and performing a scaling operation, the attention scores among the broken tooth temporal enhancement features, bearing fault condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features are obtained. The attention score is then input into the Softmax activation function layer for normalization, generating attention weights ranging from 0 to 1, including the attention weights for tooth breakage time-series enhancement features, bearing failure condition depth features, axial displacement condition depth features, and tooth flank clearance change condition depth features. Finally, the four operating condition features are multiplied by their corresponding weights to obtain the attention features for broken tooth condition, bearing failure condition, axial displacement condition, and tooth flank clearance change condition.

5. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 4, characterized in that: The condition awareness enhancement network includes A stacked Transformer encoder, each encoder layer containing a multi-head attention mechanism layer and a feedforward neural network layer; the multi-head attention mechanism layer interacts with each dimension within the attention feature of each working condition, and learns the dependencies between features from different representation subspaces through multiple parallel attention heads, thereby enhancing the expressive power of the working condition attention features and obtaining working condition context-aware features. The feedforward neural network layer performs a nonlinear transformation on the context-aware features output by the multi-head attention mechanism layer, enhancing the model's expressive power; after... The stacked Transformer encoders perform layer-by-layer feature enhancement on the working condition attention features. The working condition perception enhancement network finally outputs working condition perception enhancement features for broken teeth, bearing failure, axial displacement, and tooth backlash change.

6. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 5, characterized in that: The multi-label working condition recognition network first constructs a working condition association graph. Each node in the working condition graph represents a working condition label. The initial side length between nodes is the attention score between the working conditions. The output enhanced features of broken tooth working condition perception, enhanced features of bearing fault working condition perception, enhanced features of axial displacement working condition perception, and enhanced features of tooth flank clearance change working condition perception are used as query information. Each working condition node in the graph is given an initial feature representation through linear transformation. The working condition association graph is input into a graph attention network for inference. During the inference process, the graph attention network continuously updates the side lengths and features of each node in the working condition graph, thereby explicitly quantifying the instantaneous mutual influence strength between various working conditions. After information propagation and aggregation through multiple layers of graph attention network, each node obtains a final feature representation that integrates relevant working condition information. Finally, these updated node features are passed through an independent linear layer and a sigmoid activation function to output the working condition prediction results and their confidence scores.

7. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 6, characterized in that: The operating condition prediction result is a four-dimensional vector. Each dimension of the vector stores the prediction result and confidence level for four operating conditions: broken tooth, bearing failure, axial displacement, and tooth backlash. A prediction result of 1 for a certain dimension indicates that the gearbox is currently in the operating condition corresponding to that dimension, while a result of 0 indicates that the gearbox is not currently in the operating condition corresponding to that dimension. The confidence level is a value between 0 and 1, reflecting the probability that the model represents a certain operating condition. The higher the value, the higher the probability. If the value of the operating condition prediction result is 0 in all dimensions, the gearbox is currently in normal operating condition.

8. The method for identifying operating conditions and providing risk warnings for high-power gearboxes as described in claim 1, characterized in that: Specifically, S4 is: A threshold will be set for the confidence level of different operating conditions. When the confidence level of a certain operating condition output by S3 is within a certain time frame... When all values ​​exceed this threshold, the operating condition is identified as an event to be observed that requires the initiation of an early warning process; If a certain operating condition is identified as an event to be observed, then first calculate the time of that operating condition. The gradient of confidence level for each operating condition is used to determine the operating time based on the gradient. The rate of change of confidence within a given period; secondly, the operating condition over time. The operating condition confidence input to the long short-term memory network output future to The predicted confidence values ​​of the operating condition at multiple time points; if the predicted confidence value of the operating condition at any time point exceeds the threshold of interest, and the operating condition is within the time frame... An alarm is triggered if the gradient of the confidence level of the working condition exceeds a preset threshold.

Images