High-power gearbox operation condition identification and risk early warning method

Abstract

The invention provides a high-power gearbox operation condition identification and risk early warning method, and belongs to the technical field of gearbox risk early warning based on deep learning. The method comprises the following steps: firstly, acquiring operation data of the high-power gear box under multiple working conditions based on multiple types of sensors, and preprocessing to form a time sequence data set; then, a multi-branch feature extraction model is designed, and depth features of specific working conditions such as tooth breakage, bearing faults, axial displacement and tooth backlash changes are extracted with priori knowledge as guidance; further constructing a gearbox operation condition recognition model, and realizing multi-label condition recognition and confidence evaluation through feature fusion based on an attention mechanism, Transform enhanced modeling and a graph reasoning network; finally, a multi-stage dynamic early warning module is designed, graded early warning is triggered by integrating the instantaneous state, the evolution trend and prognosis evaluation, and safe and stable operation of the high-power gearbox is guaranteed. According to the method, the recognition accuracy and generalization performance of complex working conditions and multi-fault concurrent situations are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of gearbox risk early warning based on deep learning, and particularly relates to a high-power gearbox operating condition recognition and risk early warning method. BACKGROUND

[0002] As the core transmission component of the system, the operating state of the high-power gearbox directly affects the reliability, efficiency and service life of the entire system. Common failure modes of high-power gearboxes include broken teeth, bearing failure, axial displacement, etc. If these faults are not discovered and handled in time, they may cause equipment chain damage, production stoppage and even safety accidents. Therefore, accurately identifying the operating conditions of high-power gearboxes and providing risk early warning is a key technology for realizing predictive maintenance, avoiding sudden shutdowns and improving operational safety.

[0003] Existing high-power gearbox operating condition recognition and fault diagnosis methods can be mainly divided into the following 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] Methods based on physical models: This method relies on the establishment of an accurate mathematical model of the gear transmission system, which simulates the behavior of the system under normal and fault conditions through dynamic equations, vibration response models or thermodynamic models, etc. By comparing the differences between actual measurement data and model output, fault detection and positioning are achieved. For example, by establishing a stiffness excitation model of the gear pair, the vibration response changes caused by tooth root cracks are analyzed. This method has clear physical meaning, but its accuracy depends on the model. For complex operating conditions and multiple fault coupling scenarios, model establishment is difficult, computation is complex, and it is difficult to adapt to individual differences and changes in operating environment, limiting its engineering practicality.

[0005] Methods based on signal processing and feature extraction: This method analyzes vibration, temperature, acoustic emission and other sensor signals, extracts time domain (such as root mean square, kurtosis), frequency domain (such as Fourier transform, envelope spectrum) and time-frequency domain (such as wavelet transform, empirical mode decomposition) features, and uses these features to construct a fault index system or use a classifier (such as support vector machine, decision tree) for state recognition. For example, by analyzing the frequency spectrum characteristics of the vibration signal, the local fault frequency of the bearing is identified. This method does not require an accurate physical model, but its effectiveness depends heavily on the quality of feature extraction and expert experience. It also lacks the ability to process early weak fault features and multi-source heterogeneous sensor data fusion, making it difficult to achieve adaptive and accurate multi-label operating condition recognition.

[0006] Machine learning / deep learning based method: In recent years, with the development of sensor technology and computing power, data-driven methods have gradually become the mainstream. This method uses machine learning (such as random forest, XGBoost) or deep learning models (such as convolutional neural network CNN, long short-term memory network LSTM) to automatically learn features and patterns from raw or simply preprocessed data. For example, one-dimensional CNN is used to extract features directly from vibration signals for fault classification, or LSTM is used to model the time-dependent relationship of sensor signals. Deep learning models have shown great ability in complex pattern recognition, avoiding the limitations of manual feature extraction. However, existing methods still have many shortcomings: most studies focus on single type of sensor data (such as vibration), and fail to fully integrate multi-source heterogeneous sensor information (such as temperature field, oil particle, displacement); the model is mostly single network structure, which is difficult to adapt to the unique representation needs of different fault modes; in addition, existing methods mostly focus on the accuracy of fault classification, but lack of dynamic decision mechanism for risk warning, which is difficult to meet the comprehensive needs of real-time, robustness and explainability in industrial field.

[0007] Although existing methods have made progress in gearbox working condition monitoring, there are still the following main defects: (1) insufficient multi-source heterogeneous data fusion capability, unable to systematically integrate vibration, temperature, displacement, oil and other multi-modal sensor information, and lack of effective alignment method for multi-sensor asynchronous sampling data, which restricts the full representation of fault features; (2) limited time series dynamic modeling capability, most methods fail to fully exploit the long-range dependence and dynamic evolution patterns in sensor data, with low sensitivity to early faults and progressive faults; (3) poor model generalization and explainability, the black box nature of deep learning makes the diagnostic decision process difficult to understand, and the model is often designed for specific equipment or working conditions, which is difficult to adapt to different operating conditions and equipment variants; (4) imperfect warning mechanism, existing methods mostly focus on identification after fault occurs or in severe stage, lack of multi-level warning strategy based on confidence evaluation and trend prediction, unable to realize real advance warning and predictive maintenance. SUMMARY

[0008] To solve the above problems, the present application provides a high-power gearbox operating condition recognition and risk warning method, comprising the following steps: S1, based on multi-class sensor, collecting real-time running data of high-power gearbox, including bearing vibration signal data, axial displacement data, gear vibration signal data, bearing temperature field data, gear temperature field data and metal particle number data; S2, input the data collected in S1 into the trained multi-branch feature extraction model, first extract the gearbox working condition multi-source data time-dependent deep features based on the gearbox working condition feature extraction backbone network, then share the multi-source data time-dependent deep features based on the multi-branch extraction network, and output the timing enhancement features of the broken tooth working condition, the deep features of the bearing fault working condition, the deep features of the axial displacement working condition and the deep features of the tooth side clearance change working condition; S3, input the four kinds of features obtained in S2 into the trained gearbox running condition recognition model, the model adopts a multi-feature fusion mechanism and a multi-label classification task modeling method to predict the current running condition of the gearbox, and obtains four prediction results and their corresponding confidence degrees, i.e., broken tooth, bearing fault, axial displacement and tooth side clearance; S4, based on the prediction results and confidence degrees obtained in S3, corresponding early warning is performed through a preset risk early warning strategy.

[0009] Preferably, an acceleration sensor is arranged at the bearing of the gearbox to collect vibration signal data of the bearing running process as bearing vibration signal data; a displacement sensor is arranged at the input shaft and the output shaft of the gearbox to collect axial displacement data of the bearing running process; an acceleration sensor is arranged at the gear meshing position of the gearbox to collect vibration signal data of the gear running process as gear vibration signal data; a thermistor sensor array is arranged at the bearing of the gearbox to collect temperature data of the bearing running process as bearing temperature field data; a thermocouple sensor array is arranged at the gear meshing position of the gearbox to collect temperature data of the gear running process as gear temperature field data; and an oil particle counter is installed at the oil outlet of the gearbox to collect metal particle quantity data of the gearbox running process.

[0010] Preferably, the multi-branch feature extraction model comprises a gearbox working condition feature extraction backbone network, a broken tooth working condition feature extraction branch network, a bearing fault working condition feature extraction branch network, an axial displacement working condition feature extraction branch network and a tooth side clearance change working condition feature extraction branch network; The gearbox working condition feature extraction backbone network takes the gearbox working condition multi-source data composed of the bearing vibration signal data, the axial displacement data, the gear vibration signal data, the bearing temperature field data, the gear temperature field data and the metal particle quantity data as input, and obtains gearbox working condition multi-source data time-dependent deep features, including gear vibration signal data time-dependent deep features, gear temperature field data time-dependent deep features, bearing vibration signal data time-dependent deep features, bearing temperature field data time-dependent deep features, metal particle quantity data time-dependent deep features and axial displacement data time-dependent deep features; The broken tooth working condition feature extraction branch network takes the gear vibration signal data time-dependent deep features and the gear temperature field data time-dependent deep features as input, and outputs the timing enhancement features of the broken tooth working condition; The bearing fault working condition feature extraction branch network inputs bearing vibration signal data time-dependent deep features, bearing temperature field data time-dependent deep features, and metal particle quantity data time-dependent deep features, and outputs bearing fault working condition deep features. The axial displacement working condition feature extraction branch network inputs bearing vibration signal data time-dependent deep features, and outputs axial displacement working condition deep features. The tooth side clearance change working condition feature extraction branch network inputs gear vibration signal data time-dependent deep features, and outputs tooth side clearance change working condition deep features.

[0011] Preferably, the gearbox working condition feature extraction main network comprises a feature alignment layer and two feature modeling networks connected in series, and the feature modeling network comprises a fully connected layer, a long short-term memory network layer, and a ReLU activation function layer. First, the gearbox working condition multi-source data composed 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 are input into the feature alignment layer for time scale feature alignment, and the feature alignment layer uses an autoregressive difference moving average model to align the time dimension of the gearbox working condition multi-source data. Second, the aligned gearbox working condition multi-source data are input into the first feature modeling network for feature extraction, and the feature values of the feature points with a feature value of 0 are discarded after the feature extraction by the fully connected layer and the feature truncation by the ReLU activation function layer. Then, the output of the ReLU activation function layer is input into the long short-term memory network layer to model the time-dependent relationship of the features and capture the potential forward time-dependent relationship of each working condition data in the gearbox working condition multi-source data, thereby obtaining gearbox working condition multi-source data time-dependent features. The time-dependent features output by the first feature modeling network are input into the second feature modeling network to obtain gearbox working condition multi-source data time-dependent deep features.

[0012] Preferably, the gearbox working condition recognition model comprises a feature weighted fusion network, a working condition perception enhancement network, and a multi-label working condition recognition network. The feature weighted fusion network serves as a preposition layer of the gearbox working condition recognition model, receives the tooth break timing enhancement features, bearing fault working condition deep features, axial displacement working condition deep features, and tooth side clearance change working condition deep features from the multi-branch feature extraction model, and obtains tooth break working condition attention features, bearing fault working condition attention features, axial displacement working condition attention features, and tooth side clearance change working condition attention features based on an attention mechanism and feature weighting. The working condition attention features output by the feature weighted fusion network are input into the working condition perception enhancement network, and the working condition attention features are enhanced by the working condition perception enhancement network. The stacked Transformer encoder performs layer-by-layer feature enhancement on the working condition attention features, and outputs the gear tooth missing working condition perception enhanced feature, the bearing fault working condition perception enhanced feature, the axial displacement working condition perception enhanced feature and the tooth side gap change working condition perception enhanced feature; the four enhanced features are input into a multi-label working condition recognition network to output the working condition prediction result and the confidence thereof, wherein the multi-label working condition recognition network first constructs a working condition correlation graph, and then adopts a graph attention network to recognize and predict the working condition.

[0013] Preferably, the feature weighted fusion network comprises a feature splicing layer, a self-attention mechanism layer and a Softmax activation function layer; the obtained gear tooth missing working condition time series enhanced feature, bearing fault working condition deep feature, axial displacement working condition deep feature and tooth side gap change working condition deep feature are first input into the feature splicing layer for feature splicing to form an aggregated feature matrix; the aggregated feature matrix is then input into the self-attention mechanism layer, which maps the input aggregated feature matrix into three different representation spaces of query Q, key K and value V through three learnable weight matrices; the dot product of Q and K is calculated, and a scaling operation is performed to obtain the attention scores between the gear tooth missing time series enhanced feature, the bearing fault working condition deep feature, the axial displacement working condition deep feature and the tooth side gap change working condition deep feature; the attention scores are then input into the Softmax activation function layer for normalization to generate attention weights ranging from 0 to 1, including the gear tooth missing time series enhanced feature attention weight, the bearing fault working condition deep feature weight, the axial displacement working condition deep feature weight and the tooth side gap change working condition deep feature weight; finally, the four working condition features are dot multiplied with the corresponding weights to obtain the gear tooth missing working condition attention feature, the bearing fault working condition attention feature, the axial displacement working condition attention feature and the tooth side gap change working condition attention feature.

[0014] Preferably, the working condition perception enhanced network comprises The stacked Transformer encoder performs layer-by-layer feature enhancement on the working condition attention features, and outputs the gear tooth missing working condition perception enhanced feature, the bearing fault working condition perception enhanced feature, the axial displacement working condition perception enhanced feature and the tooth side gap change working condition perception enhanced feature. The stacked Transformer encoder performs layer-by-layer feature enhancement on the working condition attention features, and outputs the gear tooth missing working condition perception enhanced feature, the bearing fault working condition perception enhanced feature, the axial displacement working condition perception enhanced feature and the tooth side gap change working condition perception enhanced feature.

[0015] Preferably, the multi-label working condition recognition network first constructs a working condition correlation graph, each node in the working condition graph represents a working condition label, and the initial edge length between the nodes is the attention score obtained between the working conditions. The working condition correlation graph is input into a graph attention network for reasoning, the graph attention network continuously updates the edge length in the working condition graph and the feature of each node in the reasoning process, thereby explicitly quantifying the instantaneous mutual influence strength between the working conditions, after information propagation and aggregation of the multi-layer graph attention network, each node obtains a final feature representation fused with related working condition information; finally, the updated node features are output through an independent linear layer and a Sigmoid activation function to output the working condition prediction result and the confidence thereof.

[0016] Preferably, the working condition prediction result is a four-dimensional vector, each dimension of the vector stores the prediction result and the confidence of the tooth breakage, bearing fault, axial displacement and tooth side gap working conditions; the working condition prediction result in a certain dimension is 1, which indicates that the gearbox is currently in the working condition corresponding to the vector dimension, if it is 0, it indicates that the gearbox is not in the working condition corresponding to the vector dimension; the confidence is a value between 0 and 1, which reflects the possibility of the model for a certain working condition, the larger the value, the higher the possibility; if the working condition prediction result in all dimensions is 0, the gearbox is currently in a normal working condition.

[0017] Preferably, S4 is specifically: A threshold value is set for the confidence of different working conditions, when the confidence of a certain working condition output by S3 is greater than the threshold value for a certain time , the working condition is confirmed as a to-be-observed event that needs to start an early warning process; If a working condition is determined as a to-be-observed event, firstly, the gradient of the working condition confidence of the working condition in time is calculated, and the confidence change speed of the working condition in time is determined according to the gradient; secondly, the working condition confidence of the working condition in time is input into a long short-term memory network to output working condition confidence prediction values of future to multiple time points, if the working condition confidence prediction value of a certain time point exceeds the threshold value and the gradient of the working condition confidence of the working condition in time exceeds a preset threshold value, an alarm is triggered.

[0018] The innovation points of the application include: (1) Collect gear box operation data by arranging vibration, displacement, temperature, oil particle and other types of sensors to build high-quality time series data sets covering five typical working conditions; This method ensures the comprehensiveness and time sequence consistency of the data, providing a basis for subsequent deep feature extraction; (2) Multi-branch time series feature extraction architecture: According to the physical prior knowledge of different fault modes of the gear box, a special branch network is designed to extract deep time series features of heterogeneous multi-source sensor data, realize the targeted enhancement and representation of fault features, and overcome the limitation that traditional single model is difficult to adapt to multiple fault features; (3) Multi-label working condition recognition mechanism based on attention and graph reasoning: The multi-branch features are integrated by using a feature weighting fusion network, cross-dimensional feature interaction is realized through an improved Transformer encoder, and the correlation between working condition labels is explicitly modeled by introducing a graph attention network, which improves the recognition accuracy and interpretability of multiple concurrent faults; (4) Multi-level warning strategy combining instantaneous confidence and evolution trend: A dual confirmation mechanism based on confidence duration and gradient change is proposed, combined with a long short-term memory network (LSTM) for confidence prediction, realizing the escalation of warning from state recognition to trend judgment, effectively suppressing false positives and improving warning reliability.

[0019] Compared with the prior art, the present application has the following advantages: (1) Multi-branch time series feature extraction architecture: According to the physical prior knowledge of different fault modes of the gear box, a special branch network is designed to extract deep time series features of heterogeneous multi-source sensor data, realize the targeted enhancement and representation of fault features, and overcome the limitation that traditional single model is difficult to adapt to multiple fault features; (2) Multi-label working condition recognition mechanism based on attention and graph reasoning: The multi-branch features are integrated by using a feature weighting fusion network, cross-dimensional feature interaction is realized through an improved Transformer encoder, and the correlation between working condition labels is explicitly modeled by introducing a graph attention network, which improves the recognition accuracy and interpretability of multiple concurrent faults; (3) Multi-level warning strategy combining instantaneous confidence and evolution trend: A dual confirmation mechanism based on confidence duration and gradient change is proposed, combined with a long short-term memory network (LSTM) for confidence prediction, realizing the escalation of warning from state recognition to trend judgment, effectively suppressing false positives and improving warning reliability. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The overall technical route flowchart of the present application.

[0021] Figure 2 The overall framework diagram of the multi-branch feature extraction model.

[0022] Figure 3 Schematic diagram of the gearbox operating condition identification model structure.

[0023] Figure 4 This is a schematic diagram of the gearbox multi-level dynamic warning process.

[0024] Figure 5 1 is a comparison chart of the recognition delay of early faults by different models in the embodiment.

[0025] Figure 6 This is a performance comparison chart of different models in the embodiment on the gearbox working condition identification task.

[0026] Figure 7 It is a radar chart of early warning performance in the embodiment.

[0027] Figure 8 This is a confusion matrix diagram for a multi-fault concurrent scenario in an embodiment. DETAILED DESCRIPTION

[0028] This paper proposes a method for identifying the operating conditions and risk warning of high-power gearboxes based on a time series deep network, aiming to achieve accurate perception of the operating status of high-power gearboxes and advance warning of fault risks. Figure 1 As shown in the paper: First, the operating data of high-power gearboxes under multiple working conditions are collected based on multiple types of sensors, and the missing values ​​are supplemented by interpolation methods to form a high-quality time series dataset; then, a multi-branch feature extraction model is designed, and the deep features of specific working conditions such as broken teeth, bearing faults, axial displacement and tooth side clearance changes are extracted respectively based on prior knowledge; then, a gearbox operating condition recognition model is constructed, and multi-label working condition recognition and confidence assessment are realized through feature fusion based on attention mechanism, Transformer enhanced modeling and graph reasoning network; finally, a multi-level dynamic warning module is designed to trigger graded warnings based on the instantaneous state, evolution trend and prognostic evaluation to ensure the safe and stable operation of high-power gearboxes.

[0029] The specific implementation process of the present invention is described in detail below with reference to specific embodiments.

[0030] 1. Data collection for identifying operating conditions of high-power gearboxes Gearbox operating conditions are categorized into five types: normal, broken tooth, bearing failure, axial displacement, and side clearance change. A broken tooth manifests as sudden impact vibration and abnormal temperature rise, while bearing failure is characterized by high-frequency vibration, localized temperature changes, and wear particles in the oil. Axial displacement is monitored and identified through vibration analysis and shaft displacement sensors. Side clearance change affects the gear meshing process, leading to increased vibration and noise. Guided by this prior knowledge, the present invention designed the following data collection method to cover these five gearbox operating conditions.

[0031] The acceleration sensor is arranged at the bearing of the gearbox to collect vibration signal data of the bearing during operation, which is recorded as bearing vibration signal data; the displacement sensor is arranged at the input shaft and the output shaft of the gearbox to collect axial displacement data of the bearing during operation; the acceleration sensor is arranged at the gear meshing of the gearbox to collect vibration signal data of the gear during operation, which is recorded as gear vibration signal data; the thermistor sensor array is arranged at the bearing of the gearbox to collect temperature data of the bearing during operation, which is recorded as bearing temperature field data; the thermocouple sensor array is arranged at the gear meshing of the gearbox to collect temperature data of the gear during operation, which is recorded as gear temperature field data; the oil particle counter is installed at the oil outlet of the gearbox to collect metal particle quantity data of the gearbox during operation. The collected data is subjected to missing value completion, specifically, the missing values of the working condition data are completed by using a cubic spline interpolation method.

[0032] II. Design of multi-branch feature extraction model A multi-branch feature extraction model is designed under the guidance of prior knowledge that "gear tooth breakage uses gear vibration signal data and gear temperature field data; bearing failure uses bearing vibration signal data, bearing temperature field data and metal particle quantity data; axial displacement uses bearing vibration signal data and axial displacement data; gear side clearance change uses gear vibration signal data". The multi-branch feature extraction model comprises a gearbox working condition feature extraction backbone network, a gear tooth breakage working condition feature extraction branch network, a bearing failure working condition feature extraction branch network, an axial displacement working condition feature extraction branch network and a gear side clearance change working condition feature extraction branch network. All branch networks share working condition features extracted by the backbone network, and the overall structure is as shown in Figure 2 .

[0033] 1. A gear box working condition feature extraction main network: in order to effectively extract the common features of gear box working condition multi-source data, the application constructs a gear box working condition feature extraction main network, which includes a feature alignment layer and two feature modeling networks connected in series, the feature modeling network includes a fully connected layer, a long short-term memory network layer and a ReLU activation function layer; first, the gear box working condition multi-source data composed 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 are input into the feature alignment layer for time scale feature alignment, the feature alignment layer uses an autoregressive difference moving average model to align the time dimension of the gear box working condition multi-source data, solving the problem that the gear box working condition multi-source data cannot be aligned due to the difference in sensor sampling frequency; second, the aligned gear box working condition multi-source data is input into the first feature modeling network for feature extraction, the fully connected layer is used for feature extraction, and the ReLU activation function layer is used for feature truncation after the feature extraction, and the feature points with a feature value of 0 are discarded, 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 dependence relationship of the features, and the potential forward time dependence relationship of each working condition data in the gear box working condition multi-source data is captured to obtain the time dependence features of the gear box working condition multi-source data; the time dependence features output by the first feature modeling network are input into the second feature modeling network to obtain the time dependence deep features of the gear box working condition multi-source data; the time dependence deep features of the gear box working condition multi-source data are composed of gear vibration signal data time dependence deep features, gear temperature field data time dependence deep features, bearing vibration signal data time dependence deep features, bearing temperature field data time dependence deep features, metal particle quantity data time dependence deep features and axial displacement data time dependence deep features.

[0034] 2. A broken tooth working condition feature extraction branch network: in order to effectively fuse the multi-source deep features related to the broken tooth working condition, the application constructs a broken tooth working condition feature extraction branch network, which inputs the gear vibration signal data time dependence deep features and the gear temperature field data time dependence deep features output by the gear box working condition feature extraction main network. The branch network is composed of a feature fusion layer and a time sequence feature enhancement module. First, the two deep features are input into the feature fusion layer, the features are spliced through the fully connected layer, and the ReLU activation function is used for nonlinear transformation to obtain the fused preliminary features of the broken tooth; then the fused features are input into the time sequence 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 high-order features, the ReLU function eliminates redundant information, and the Bi-LSTM layer simultaneously captures the forward and backward time sequence dependence relationship to output the time sequence enhanced features of the broken tooth working condition.

[0035] 3. Bearing fault working condition feature extraction branch network: In view of the characteristics of bearing fault multi-source heterogeneous deep features, the bearing fault working condition feature extraction branch network is designed, and the input is bearing vibration signal data time-dependent deep feature, bearing temperature field data time-dependent deep feature and metal particle number data time-dependent deep feature. The branch network includes a feature aggregation layer, a gated recurrent unit layer and a convolution enhancement module. First, the three kinds of deep features are input into the feature aggregation layer, and the aggregated bearing fault preliminary features are output by the feature weighting fusion through the full connection layer; then the preliminary features are input into the gated recurrent unit layer, the time sequence dynamic change is captured, and the bearing fault time sequence features are output; finally, the time sequence features are input into the convolution enhancement module, the local time sequence mode is further extracted and the feature robustness is enhanced, and the bearing fault working condition deep features are output; the convolution enhancement module is composed of a one-dimensional convolution layer, a ReLU activation function and a maximum pooling layer in series.

[0036] 4. Axial displacement working condition feature extraction branch network: in order to efficiently model the time sequence characteristics of axial displacement working condition, the axial displacement working condition feature extraction branch network is constructed, and the input is bearing vibration signal data time-dependent deep feature. The branch network adopts a lightweight structure, which is composed of two series of time sequence modeling units and a skip connection. Each time sequence modeling unit includes a full connection layer, a long short-term memory network layer and a ReLU activation function layer: the full connection layer performs feature transformation, the long short-term memory network layer models the time dependence, and the ReLU function enhances the nonlinearity. The first time sequence modeling unit outputs the primary time sequence features, which are added to the output of the second time sequence modeling unit through the skip connection, and then adjusted in dimension through the full connection layer, and finally the axial displacement working condition deep features are output. The design alleviates the gradient vanishing problem of deep network and improves the feature reuse ability.

[0037] 5. Tooth side gap change working condition feature extraction branch network: in view of the nonlinear time sequence characteristics of tooth side gap change, the tooth side gap change working condition feature extraction branch network is designed, and the input is gear vibration signal data time-dependent deep feature. The branch network is composed of an attention mechanism module and a time sequence convolution network module in series. First, the input features are input into the attention mechanism module, which includes a self-attention layer and a full connection layer, focusing on the key time sequence information by calculating the feature weight, and outputting the weighted attention features; then the weighted attention features are input into the time sequence convolution network module, which is composed of a causal convolution layer, a dilated convolution layer and a ReLU activation function layer, which captures long time sequence dependence by using causal convolution and dilated convolution, and outputs the tooth side gap change working condition deep features. This structure avoids recursive calculation, improves parallel efficiency and long sequence modeling ability.

[0038] III. Wheel box running working condition recognition model design After the multi-branch feature extraction model generates the deep features of each working condition, the gear box operation condition recognition model designed by the application integrates the features from different branch networks, adopts a multi-feature fusion mechanism and a multi-label classification task modeling method to recognize and predict the current operation condition of the gear box, and provides confidence evaluation for each prediction result. The gear box operation condition recognition module adopts a feature weighted fusion network based on an attention mechanism as a pre-layer, receives the outputs from each branch network of the multi-branch feature extraction model, including the tooth break timing enhancement feature, the bearing fault working condition deep feature, the axial displacement working condition deep feature and the tooth side gap change working condition deep feature, and obtains the tooth break working condition attention feature, the bearing fault working condition attention feature, the axial displacement working condition attention feature and the tooth side gap change working condition attention feature based on the attention mechanism and feature weighting. The working condition attention features output by the feature weighted fusion network are input into the working condition perception enhancement network; the working condition attention features are subjected to layer-by-layer feature enhancement by a stacked Transformer encoder, and the tooth break working condition perception enhancement feature, the bearing fault working condition perception enhancement feature, the axial displacement working condition perception enhancement feature and the tooth side gap change working condition perception enhancement feature are output; the four enhanced features are input into the multi-label working condition recognition network, and the working condition prediction result and its confidence are output; the overall structure is shown in Figure 3

[0039] 1. The feature weighted fusion network based on the attention mechanism: the feature weighted fusion network based on the attention mechanism includes a feature splicing layer, a self-attention mechanism layer and a Softmax activation function layer; first, the obtained tooth break timing enhancement feature, bearing fault working condition deep feature, axial displacement working condition deep feature and tooth side gap change working condition deep feature 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, and the self-attention mechanism maps the input aggregated feature matrix into three different representation spaces of query Q, key K and value V through three learnable weight matrices. The attention scores between the tooth break timing enhancement feature, the bearing fault working condition deep feature, the axial displacement working condition deep feature and the tooth side gap change working condition deep feature are obtained by calculating the dot product of Q and K and performing scaling operation; third, the attention scores are input into the Softmax activation function layer for normalization to generate attention weights ranging from 0 to 1, including tooth break timing enhancement feature attention weight, bearing fault working condition deep feature weight, axial displacement working condition deep feature weight and tooth side gap change working condition deep feature weight, which reveal the attention degree of the model to different working condition features; finally, the four working condition features and the corresponding weights are subjected to dot product to obtain the tooth break working condition attention feature, the bearing fault working condition attention feature, the axial displacement working condition attention feature and the tooth side gap change working condition attention feature.

[0040] ​​2. Working condition awareness enhancement network: after the feature weighting fusion network outputs the tooth breakage working condition attention feature, bearing fault working condition attention feature, axial displacement working condition attention feature and tooth side gap change working condition attention feature, in order to mine the complex nonlinear relationship between the dimension feature values of each working condition attention feature itself, the working condition awareness enhancement network is designed by improving the core component encoder layer in the Transformer architecture, realizing the deep interaction between different dimensions of features, generating feature representation with higher information density and stronger context awareness ability. The working condition awareness enhancement network comprises a stack of Transformer encoders, each encoder layer containing a multi-head attention mechanism layer and a feedforward neural network layer; the multi-head attention mechanism layer performs information interaction between each dimension of each working condition attention feature, learns the dependency relationship between features from different representation subspaces through multiple attention heads in parallel, enhances the working condition attention feature expression ability, and obtains working condition context awareness features; the feedforward neural network layer further nonlinearly transforms the context awareness features output by the multi-head attention mechanism layer, enhancing the expression ability of the model. After a stack of Transformer encoders, the working condition attention features are enhanced layer by layer, and the working condition awareness enhancement network finally outputs the tooth breakage working condition awareness enhancement feature, bearing fault working condition awareness enhancement feature, axial displacement working condition awareness enhancement feature and tooth side gap change working condition awareness enhancement feature.

[0041] 3. Multi-label working condition recognition network based on graph reasoning: In order to explicitly model the internal correlation of each working condition feature of the gearbox, the application 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 a 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 gear tooth breakage working condition perception enhancement feature, the bearing fault working condition perception enhancement feature, the axial displacement working condition perception enhancement feature and the tooth side clearance change working condition perception enhancement feature output by the working condition perception enhancement network are taken as query information. An initial feature representation is given to each working condition node in the graph through linear transformation. 2) Input the working condition correlation graph into the graph attention network for reasoning. The graph attention network updates the edge length in the working condition graph and the feature of each node during the reasoning process, thereby explicitly quantifying the instantaneous mutual influence strength between working conditions. After information propagation and aggregation through multiple layers of graph attention network, each node obtains a final feature representation that integrates related working condition information. Finally, the updated node features are output through an independent linear layer and a Sigmoid activation function to obtain the working condition prediction result and its confidence. The working condition prediction result is a four-dimensional vector, and each dimension of the vector stores the prediction result and confidence of the gear tooth breakage, bearing fault, axial displacement and tooth side clearance working conditions. If the working condition prediction result in a certain dimension is 1, it indicates that the gearbox is currently in the working condition corresponding to the vector dimension. If it is 0, it indicates that the gearbox is not currently in the working condition corresponding to the vector dimension. The confidence is a value between 0 and 1, reflecting the possibility of the model for a certain working condition. The larger the value, the higher the possibility. If the working condition prediction result in all dimensions is 0, the gearbox is currently in normal working condition.

[0042] IV. Designing a gearbox early warning module After achieving accurate identification and confidence evaluation of the running working condition of the gearbox, a multi-level dynamic early warning system integrating instant state, evolution trend and prognosis evaluation is designed to convert these instantaneous diagnostic conclusions into early warning strategies. The overall process is shown in Figure 4

[0043] 1) Determining the working condition event to be observed: To avoid false positives caused by sensor transient noise or accidental fluctuations in the model, the preliminary trigger of the early warning adopts a dual confirmation mechanism of instantaneous confidence and duration. Specifically, the system sets a threshold for the confidence of different working conditions. When the confidence of a certain working condition output by S3 is greater than the threshold for a certain time , the working condition is confirmed as a working condition event to be observed that needs to start the early warning process. 2) Fault evolution trend analysis and early warning method: If a working condition is determined to be a working condition event to be observed, it will enter 2). First, calculate the working condition​ The confidence gradient of the working condition within the range is used to determine the working condition time. The confidence change rate within the time The confidence level of the working condition in the long short-term memory network is input and the future to The confidence prediction values ​​of the working condition at multiple moments, if the value of the confidence prediction value of the working condition at a certain moment exceeds the attention threshold described in S4-1 and the working condition at time If the gradient of the confidence level of the working condition exceeds the preset threshold, an alarm is triggered.

[0044] 5. Analysis of experimental results The experiment took a 1.5MW wind turbine gearbox as the object, arranged vibration, displacement, temperature, and oil sensors according to the patented scheme, collected 100 hours of data, including early and concurrent faults, and after ARIMA alignment and interpolation completion, compared the performance of traditional schemes, including support vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), long short-term memory network (LSTM), convolutional neural network-long short-term memory network combination model (CNN-LSTM), and the present invention.

[0045] like Figure 5 As shown in the figure, in the gearbox early fault identification scenario, the recognition delay of the SVM is 15.8 hours, while the recognition delays of the 1D-CNN, LSTM, and CNN-LSTM are 12.5 hours, 8.3 hours, and 5.8 hours, respectively. The present invention achieves an average accuracy of 0.85 within 8 hours and a recognition delay of only 2.1 hours, a 64% reduction compared to the CNN-LSTM, effectively addressing the "low sensitivity to early faults" issue raised in the patent background.

[0046] like Figure 6 As shown in the figure, the accuracy, precision, recall rate and F1 score of the key indicators for evaluating the gearbox working condition recognition performance of the present invention are higher than those of SVM, 1D-CNN, LSTM and CNN-LSTM, which fully demonstrates its excellent performance in the gearbox working condition recognition task.

[0047] like Figure 7 As shown in the figure, from the perspective of the warning lead time and average accuracy of multiple types of faults such as bearing fault, broken tooth, axial displacement, and tooth side clearance, the performance of the present invention is better than existing methods such as SVM, 1D-CNN, LSTM, and CNN-LSTM in all dimensions, demonstrating better gearbox fault warning performance in all aspects.

[0048] Figure 8Compared with the confusion matrix of CNN-LSTM and the application in the multi-fault concurrent scene, the accuracy rate of various fault identification of the application is higher, the probability of error identification is lower, and the complex situation of gear box multi-fault concurrency can be accurately responded.

[0049] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0050] Although the specific embodiments of the present application are described above, they are not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A high-power gearbox operating condition identification and risk early warning method, characterized in that, The method comprises the following steps: S1, collecting real-time operation data of a large-power gearbox based on multiple 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, inputting the data collected in S1 into a trained multi-branch feature extraction model, first extracting gearbox working condition multi-source data time-dependent deep features based on a gearbox working condition feature extraction backbone network, then sharing multi-source data time-dependent deep features based on a multi-branch extraction network, and outputting tooth breakage working condition time series enhancement features, bearing fault working condition deep features, axial displacement working condition deep features and gear side clearance change working condition deep features; S3, inputting the four features obtained in S2 into a trained gearbox working condition recognition model, the model adopts a multi-feature fusion mechanism and a multi-label classification task modeling method to predict the current working condition of the gearbox, and obtains four prediction results and their corresponding confidence degrees, i.e., tooth breakage, bearing fault, axial displacement and gear side clearance; S4, based on the prediction results and confidence degrees obtained in S3, corresponding early warning is performed through a preset risk early warning strategy.

2. A method for operating condition identification and risk early warning of a high-power gearbox according to claim 1, characterized in that: An acceleration sensor is arranged at the bearing of the gearbox to collect vibration signal data of the bearing during operation as bearing vibration signal data; a displacement sensor is arranged on the input shaft and the output shaft of the gearbox to collect axial displacement data of the bearing during operation; an acceleration sensor is arranged at the gear meshing position of the gearbox to collect vibration signal data of the gear during operation as gear vibration signal data; a thermistor sensor array is arranged at the bearing of the gearbox to collect temperature data of the bearing during operation as bearing temperature field data; a thermocouple sensor array is arranged at the gear meshing position of the gearbox to collect temperature data of the gear during operation as gear temperature field data; and an oil particle counter is installed at the oil outlet of the gearbox to collect metal particle quantity data during operation of the gearbox.

3. The operating condition identification and risk early warning method for a high-power gearbox according to claim 1, characterized in that: The multi-branch feature extraction model comprises a gearbox working condition feature extraction backbone network, a tooth breakage working condition feature extraction branch network, a bearing fault working condition feature extraction branch network, an axial displacement working condition feature extraction branch network and a gear side clearance change working condition feature extraction branch network; The gearbox working condition feature extraction backbone network takes gearbox working condition multi-source data composed 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, and obtains gearbox working condition multi-source data time-dependent deep features, including gear vibration signal data time-dependent deep features, gear temperature field data time-dependent deep features, bearing vibration signal data time-dependent deep features, bearing temperature field data time-dependent deep features, metal particle quantity data time-dependent deep features and axial displacement data time-dependent deep features; The tooth breakage working condition feature extraction branch network takes gear vibration signal data time-dependent deep features and gear temperature field data time-dependent deep features as input, and outputs tooth breakage working condition time series enhancement features; The bearing fault condition feature extraction branch network inputs bearing vibration signal data time-dependent deep features, bearing temperature field data time-dependent deep features and metal particle quantity data time-dependent deep features, and outputs bearing fault condition deep features; The axial displacement condition feature extraction branch network inputs bearing vibration signal data time-dependent deep features, and outputs axial displacement condition deep features; The tooth side clearance change condition feature extraction branch network inputs gear vibration signal data time-dependent deep features, and outputs tooth side clearance change condition deep features.

4. A method for operating condition identification and risk early warning of a high-power gearbox according to claim 3, characterized in that: The gearbox condition feature extraction main network includes a feature alignment layer and two feature modeling networks connected in series, and the feature modeling network includes a fully connected layer, a long short-term memory network layer and a ReLU activation function layer. First, the gearbox condition multi-source data composed 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 are input into the feature alignment layer for time scale feature alignment, and the feature alignment layer uses an autoregressive difference moving average model to align the time dimension of the gearbox condition multi-source data. Second, the aligned gearbox condition multi-source data are input into the first feature modeling network for feature extraction, and the feature extraction is performed by the fully connected layer, and the feature points with a feature value of 0 are discarded after the feature is truncated by the ReLU activation function layer. Then, the output of the ReLU activation function layer is input into the long short-term memory network layer to model the time-dependent relationship of the features and capture the potential forward time-dependent relationship of each condition data in the gearbox condition multi-source data, thereby obtaining the time-dependent features of the gearbox condition multi-source data. The time-dependent features output by the first feature modeling network are input into the second feature modeling network to obtain the time-dependent deep features of the gearbox condition multi-source data.

5. A method for operating condition identification and risk early warning of a high-power gearbox according to claim 1, characterized in that: The gearbox operating condition recognition model includes a feature weighted fusion network, a condition perception enhancement network and a multi-label condition recognition network. The feature weighted fusion network is used as the front layer of the gearbox operating condition recognition model, receives the tooth break timing enhancement features, bearing fault condition deep features, axial displacement condition deep features and tooth side clearance change condition deep features from the multi-branch feature extraction model, and obtains tooth break condition attention features, bearing fault condition attention features, axial displacement condition attention features and tooth side clearance change condition attention features based on the attention mechanism and feature weighting. The condition attention features output by the feature weighted fusion network are input into a condition perception enhancement network, and the condition attention features are subjected to layer-by-layer feature enhancement through a stacked Transformer encoder The condition attention features are output as the tooth wear condition perception enhancement features, the bearing fault condition perception enhancement features, the axial displacement condition perception enhancement features and the tooth side clearance change condition perception enhancement features. The four enhanced features are input into the multi-label condition recognition network to output condition prediction results and their confidence, and the multi-label condition recognition network first constructs a condition correlation graph, and then uses a graph attention network to recognize and predict the conditions.

6. A method for operating condition identification and risk early warning of a high-power gearbox according to claim 5, characterized in that: The feature weighting fusion network comprises a feature splicing layer, a self-attention mechanism layer and a Softmax activation function layer; first, the obtained timing enhanced feature of tooth breakage, the bearing fault condition deep feature, the axial displacement condition deep feature and the tooth side gap change condition deep feature are input into the feature splicing layer for feature splicing to form an aggregated feature matrix; secondly, the aggregated feature matrix is input into the self-attention mechanism layer, and the self-attention mechanism maps the input aggregated feature matrix into three different representation spaces of query Q, key K and value V through three learnable weight matrices; the dot product of Q and K is calculated, and the tooth breakage timing enhanced feature, the bearing fault condition deep feature, the axial displacement condition deep feature and the tooth side gap change condition deep feature are obtained through scaling operation; Thirdly, the attention score is input into the Softmax activation function layer for normalization to generate attention weights ranging from 0 to 1, including tooth breakage timing enhanced feature attention weight, bearing fault condition deep feature weight, axial displacement condition deep feature weight and tooth side gap change condition deep feature weight; Finally, the four kinds of condition features are dot multiplied with the corresponding weights to obtain tooth breakage condition attention feature, bearing fault condition attention feature, axial displacement condition attention feature and tooth side gap change condition attention feature.

7. A method for operating condition identification and risk early warning of a high-power gearbox according to claim 6, characterized in that: The working condition perception enhancement network comprises a plurality of stacked Transformer encoders, each encoder layer comprising a multi-head attention mechanism layer and a feedforward neural network layer; the multi-head attention mechanism layer performs information interaction between each dimension inside each working condition attention feature, learns the dependency relationship between features from different representation subspaces through a plurality of attention heads in parallel, enhances the expression capability of the working condition attention feature, and obtains working condition context perception features; The feedforward neural network layer performs nonlinear transformation on the context perception features output by the multi-head attention mechanism layer, and enhances the expression ability of the model. The stacked Transformer encoders perform layer-by-layer feature enhancement on the working condition attention features, and the working condition perception enhancement network finally outputs the working condition perception enhancement features of the gear tooth breakage, bearing fault, axial displacement and tooth side gap change.

8. A method of operating condition identification and risk early warning for a high-power gearbox according to claim 7, characterized in that: The multi-label condition recognition network first constructs a condition correlation graph, each node in the condition graph represents a condition label, and the initial edge length between nodes is the attention score between the obtained conditions; the output tooth breakage condition perception enhanced feature, bearing fault condition perception enhanced feature, axial displacement condition perception enhanced feature and tooth side gap change condition perception enhanced feature are taken as query information, and each condition node in the graph is given an initial feature representation through linear transformation; The condition correlation graph is input into the graph attention network for reasoning, and the edge length in the condition graph and the feature of each node are updated during the reasoning process, so as to quantitatively measure the instantaneous mutual influence strength between conditions; after information propagation and aggregation of multiple layers of graph attention network, each node obtains a final feature representation which integrates related condition information; finally, the updated node features are output through an independent linear layer and a Sigmoid activation function to output the condition prediction result and its confidence.

9. A method of operating condition identification and risk early warning for a high-power gearbox according to claim 8, characterized in that: The condition prediction result is a four-dimensional vector, and each dimension of the vector stores the prediction result and confidence of tooth breakage, bearing fault, axial displacement and tooth side gap; if the condition prediction result in a certain dimension is 1, it indicates that the gearbox is currently in the condition corresponding to the vector dimension, and if it is 0, it indicates that the gearbox is not currently in the condition corresponding to the vector dimension; the confidence is a value between 0 and 1, which reflects the possibility of the model for a certain condition, and the larger the value, the higher the possibility; if the condition prediction result in all dimensions is 0, the gearbox is currently in normal condition.

10. A method for operating condition identification and risk early warning of a high-power gearbox according to claim 1, characterized in that: The S4 is specifically A threshold of attention is set for the confidence of different working conditions. When the confidence of a working condition output by S3 is greater than the threshold for a certain time , the working condition is confirmed as a to-be-observed event requiring to start a pre-warning process. If a certain working condition is determined as an event to be observed, firstly, the gradient of the working condition confidence of the working condition within time is calculated, the confidence change speed of the working condition within time is determined according to the gradient; secondly, the working condition confidence of the working condition within time is input into a long short-term memory network to output the working condition confidence prediction value of future to multiple time points, and if the value of a certain time point in the working condition confidence prediction value exceeds the attention threshold value and the gradient of the working condition confidence of the working condition within time exceeds the preset threshold value, an alarm is triggered.

Images