A gear life prediction method based on a multi-modal spatio-temporal coupling graph neural network

Abstract

The application relates to the technical field of gear detection, and discloses a gear life prediction method based on a multi-modal space-time coupling graph neural network, which comprises the following steps: collecting multi-modal data for pretreatment, and extracting multi-modal feature sequences according to sliding time windows. In each sliding time window, a collection node and a measured gear node are taken as a node set, physical coupling edges and data coupling edges are determined, edge weights are fused, and a space-time coupling graph is constructed. Time correlation information is extracted, feature propagation is carried out on the space-time coupling graph, cross-modal attention weight distribution is carried out on different modes, and a degradation state representation of the measured gear is obtained. A parameter set is obtained based on data with a remaining life label. In the running process, newly formed multi-modal feature sequences and the space-time coupling graph are input, and the remaining life of the measured gear is output. When the remaining life is less than a maintenance threshold, a maintenance suggestion is generated. The application realizes stable and reliable prediction of the remaining life of the measured gear.

Description

Technical Field

[0001] This invention relates to the field of gear inspection technology, and more specifically, to a gear life prediction method based on a multimodal spatiotemporal coupled graph neural network. Background Technology

[0002] Gear transmission systems are widely used in wind power, metallurgy, and transportation equipment, operating under complex conditions and experiencing significant load fluctuations. To improve equipment availability and reduce maintenance costs, it is necessary to monitor the condition and assess lifespan without disassembling the equipment, which falls under the category of equipment condition monitoring and lifespan assessment technology. With the development of multi-source sensing and time-series analysis, the industry is gradually moving from single-signal diagnosis to multimodal fusion and spatiotemporal correlation modeling. However, how to stably determine the remaining lifespan under conditions of real structural coupling, data correlation, and operating disturbances remains a key challenge.

[0003] Existing methods often model graph structures or time series separately, making it difficult to simultaneously characterize physical and data coupling edges and perform edge weight fusion within the same process. The allocation of multimodal signal weights is often fixed or empirically set, lacking adaptive attention weights based on propagation results. The training phase rarely incorporates joint constraints on energy budget consistency and edge weight sparsity, resulting in insufficient robustness. Furthermore, the linkage between the online phase and maintenance thresholds is often missing. For example, CN117743964A – a method for predicting remaining equipment lifespan based on graph data – while using graph neural networks for lifespan prediction, fails to determine physical and data coupling edges in parallel within each sliding time window and perform edge weight fusion. It also lacks cross-modal attention weight allocation based on propagation results, and fails to address the joint constraints on energy budget consistency and edge weight sparsity, nor does it generate maintenance recommendations in conjunction with maintenance thresholds.

[0004] Therefore, it is necessary to design a gear life prediction method based on a multimodal spatiotemporally coupled graph neural network to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a gear life prediction method based on a multimodal spatiotemporally coupled graph neural network, which aims to solve the problem of how to stably output the remaining life of the tested gear in a gear transmission scenario where real structural coupling, data correlation and operating condition disturbance coexist without disassembly detection.

[0006] This invention proposes a gear life prediction method based on a multimodal spatiotemporally coupled graph neural network, comprising:

[0007] Multimodal data is collected and preprocessed, and multimodal feature sequences are extracted according to a sliding time window. The multimodal data includes vibration data, acoustic data, temperature data, rotational speed data, torque data, and oil state data.

[0008] Within each sliding time window, the node set is composed of the acquisition node and the measured gear node. The physical coupling edge is determined based on the structural connection, contact relationship and vibration transmission path. The data coupling edge is determined based on the statistical correlation and causal relationship of historical characteristics. The edge weights of the physical coupling edge and the data coupling edge are fused to construct a spatiotemporal coupling graph.

[0009] Temporal correlation information is extracted from the multimodal feature sequence and feature propagation is performed on the spatiotemporal coupling graph. Based on the propagation results, cross-modal attention weights are assigned to different modes to obtain a representation of the degradation state of the gear under test.

[0010] A parameter set for inference is obtained based on data labeled with remaining lifetime. The determination of the parameter set simultaneously constrains the remaining lifetime prediction error, the degradation state fitting error, the energy budget consistency, and the edge weight sparsity.

[0011] During operation, the newly formed multimodal feature sequence and spatiotemporal coupling diagram are input, and the remaining life of the tested gear is output. When the remaining life is less than the maintenance threshold, maintenance suggestions are generated.

[0012] Furthermore, when collecting and preprocessing multimodal data and extracting multimodal feature sequences using a sliding time window, the process includes:

[0013] The timestamps of each channel are synchronized using a unified time base, and the relative positions of the acquisition nodes and the gears under test are recorded and spatially aligned using unified spatial coordinates.

[0014] The multimodal data is subjected to background noise estimation, denoising, zero drift correction, outlier suppression, missing data completion, and unit unification in a preset order. The outlier suppression is determined based on the amplitude range and rate of change, and the missing data completion is constrained based on the consistency between adjacent time periods and related modes.

[0015] The data is segmented according to the sliding time window, with the starting point of the window aligned with the unified time reference. Within each sliding time window, the time domain features, frequency domain features, envelope correlation features, and operating condition features of the multimodal data are extracted. Low-stability features are removed based on feature stability and feature correlation to obtain the multimodal feature sequence.

[0016] Furthermore, when determining the physical coupling edge based on structural connections, contact relationships, and vibration transmission paths, the following is included:

[0017] The connection relationship is established based on gear meshing pairs, coaxial connections, bearing supports and housing constraints. The direction of the edge is determined based on the contact position and vibration transmission path. The weight of the edge is set in stages according to the contact type, assembly preload state, support stiffness and vibration transmission path level.

[0018] The edge weights are tested for stability within multiple consecutive sliding time windows, and the physical coupling edge is retained only when the edge weights fluctuate within a preset fluctuation range.

[0019] Furthermore, when determining data coupling edges based on statistical correlations and causal relationships of historical characteristics, and fusing the edge weights of the physical coupling edges and data coupling edges, the process includes:

[0020] The correlation stability is obtained by judging whether the linkage direction and linkage strength of the same feature pair remain within a preset fluctuation range in the continuous sliding time window based on the multimodal feature sequence; the consistency of the delayed response is obtained by comparing whether the time sequence of the first changing feature and the later responding feature remains within a preset range in the continuous sliding time window; and the consistency of the event linkage is obtained by statistically analyzing the common occurrence and sequence of events in the continuous sliding time window based on the event time points triggered by the multimodal features and in combination with the start / stop, alarm and load changes in the operation log and maintenance record.

[0021] Data coupling edges are determined based on the correlation stability, hysteresis response consistency, and event linkage consistency, and edge weights are assigned to the data coupling edges.

[0022] The weights of the physical coupling edges and the data coupling edges are unified in terms of scale and range, and weighted and fused in order of physical priority. When the conclusions of the two types of edges conflict, the physical coupling edges are retained and adjusted according to the data coupling edges.

[0023] The edge weights after fusion are sparsified, and edges that do not meet the requirements of edge weight stability and window consistency verification are removed in ascending order of weight value. Before entering the next sliding time window, the retained edges are checked based on the window consistency verification result, and finally the spatiotemporal coupling graph is obtained.

[0024] Furthermore, when extracting time-related information from the multimodal feature sequence and performing feature propagation on the spatiotemporal coupling graph, the process includes:

[0025] Within each sliding time window, time-related information is extracted for each modal feature, including trend changes, periodic changes, and abrupt events.

[0026] After the extraction of the temporal correlation information is completed, the edge weights fused in the spatiotemporal coupling graph are used as the basis for transmission. One to several rounds of feature propagation are carried out between adjacent nodes in descending order of edge weights. The number of propagation rounds is within the preset number of propagation rounds. After each round of propagation, when the temporal correlation conflict caused by the propagation exceeds the preset conflict threshold, the contribution of the edge is suppressed. After all propagation rounds are completed, the node temporal correlation aggregation result is obtained.

[0027] Furthermore, when performing cross-modal attention weight allocation for different modes based on the propagation results to obtain the degradation state representation of the tested gear, this includes:

[0028] Non-negative attention weights are set based on the node time correlation aggregation results, correlation stability, lag response consistency and event linkage consistency, and the noise level of each modality in historical samples, and the attention weights are normalized within the same node.

[0029] When the correlation stability of a certain mode is insufficient or conflicts with the working condition indication within a continuous sliding time window, its attention weight is reduced. When the degradation is aggravated by the joint indication of multiple modes, the corresponding attention weight is increased. The results of each mode after the attention weight allocation are weighted and converged within the node to obtain the node-level degradation state result. The results are then weighted and converged between nodes according to the meshing pair and vibration path level to obtain the degradation state representation of the gear under test.

[0030] Furthermore, when determining the parameter set for inference based on data labeled with remaining lifetime, the process includes:

[0031] The historical samples are divided into training windows and validation windows according to the sliding time window.

[0032] The initial values ​​of the parameters are determined based on the degradation state representation and remaining lifetime label within the training window;

[0033] The parameter set is obtained iteratively in batch updates. After each update, the remaining life prediction error, degradation state fitting error, energy balance consistency, and edge weight sparsity are evaluated simultaneously. The energy balance consistency is determined by comparing whether the common trend of speed change, torque change, temperature change, and oil state change is consistent with the energy transmission direction of the gear under test. The edge weight sparsity is constrained by removing low contribution edges according to edge weight from small to large and limiting the number of retained edges of each node to within a preset range.

[0034] When the four evaluation metrics in the verification window meet the preset thresholds for several consecutive rounds, the updates stop and a parameter set for inference is obtained. The parameter set includes attention weights, edge weights, parameters related to remaining lifetime output, and parameters related to degradation state representation.

[0035] Furthermore, when determining the parameter set for inference based on data labeled with remaining lifetime, a phased weighting strategy is employed, including:

[0036] In the first stage, the optimization objectives are only the remaining lifetime prediction error and the degradation state fitting error, and the window consistency check is used as the stopping condition.

[0037] In the second stage, after the conditions of the first stage are met, the constraint weights for energy balance consistency and edge weight sparsity are increased. After each round of updates, edge weight stability test and window consistency check are performed. If the verification window shows performance degradation or consistency does not meet the preset consistency threshold, the parameter set of the previous round is rolled back and the growth rate of constraint weights is reduced.

[0038] After completing the second phase, the parameter set used for inference is finalized, and the threshold, number of windows, and backoff conditions used for retraining are recorded.

[0039] Furthermore, during operation, the newly formed multimodal feature sequence and spatiotemporal coupling diagram are input, and the remaining life of the tested gear is output. When the remaining life is less than the maintenance threshold, maintenance suggestions are generated, including:

[0040] Within each sliding time window, the newly formed multimodal feature sequence is aligned with a unified time reference and a unified spatial coordinate, and the spatiotemporal coupling graph is reconstructed. The aligned multimodal feature sequence and the reconstructed spatiotemporal coupling graph are input into the parameter set to obtain the degradation state representation and remaining lifetime.

[0041] A window consistency check is performed on a consecutive preset number of sliding time windows. Maintenance suggestions are generated only when the remaining lifespan in the consecutive windows is not higher than the maintenance threshold and the degradation status indicates no improvement trend.

[0042] A maintenance preparation recommendation is generated when the remaining lifetime is within the maintenance threshold buffer range and the window consistency check passes.

[0043] A shutdown recommendation is generated when the remaining lifetime continuously decreases and falls below the maintenance threshold within a consecutive window, accompanied by a continuous deterioration in the degradation status.

[0044] After any suggestion is generated, an alarm silence period is entered. During the alarm silence period, the suggestion is updated only if the new window consistency check result is inconsistent with the previous conclusion.

[0045] Furthermore, it also includes:

[0046] When the operating condition distribution changes or new data with remaining lifetime labels appears, the edge weights, cross-modal attention weights, and parameter set are updated on a rolling basis based on the latest data. The remaining lifetime output of historical samples before and after the update is compared with the consistency check threshold. If the comparison fails, the system reverts to the state before the update.

[0047] Compared with existing technologies, the beneficial effects of this invention are as follows: by acquiring and preprocessing multimodal data under a unified benchmark, a spatiotemporal coupled graph is constructed by fusing physical coupling edges and data coupling edges within a sliding time window, and cross-modal attention weight allocation is driven by time-related information extraction and graph feature propagation to obtain an interpretable representation of the degradation state. The invention is then trained and applied online using a parameter set that simultaneously constrains the remaining lifetime prediction error, degradation state fitting error, energy budget consistency, and edge weight sparsity. Finally, maintenance suggestions are generated in conjunction with maintenance thresholds, forming a closed loop from data acquisition, relationship characterization, information fusion to decision output. This application integrates the edge weights of physical coupling edges and data coupling edges, enabling structural connections, contact relationships, vibration transmission paths, statistical correlations, and causal relationships to be simultaneously characterized in the same process, thus improving stability under operating condition disturbances and load fluctuations. Cross-modal attention weights adaptively adjust the influence of different modes based on propagation results, suppressing the interference of low-stability features on lifetime assessment and improving the discriminative power of degradation state representation. Constraints on energy budget consistency and edge weight sparsity ensure that the parameter set used for inference follows the energy transfer direction and eliminates low-contribution connections, reducing overfitting and improving transferability. This achieves highly stable and reliable prediction outputs for the remaining lifetime of the tested gear. Attached Figure Description

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0049] Figure 1 A flowchart of a gear life prediction method based on a multimodal spatiotemporally coupled graph neural network provided in an embodiment of the present invention;

[0050] Figure 2 This is a prediction diagram of the remaining life of gears based on the method of this invention;

[0051] Figure 3 This is a comparison chart of the mean and true values ​​of gears based on the method of this invention;

[0052] Figure 4 This is a schematic diagram of a helical gearbox test bench;

[0053] Figure 5 This is a multi-index degradation curve graph for GB1 and GB2.

[0054] The components include: 1. Electrical measuring device; 2. DAS; 3. Drive motor; 4. Gearbox 1; 5. Gearbox 2; 6. Load motor; 7. Computer; 8. Control cabinet; and 9. Touch screen. Detailed Implementation

[0055] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] In traditional gear life prediction technologies, gear transmission systems are widely used in wind power, metallurgy, and transportation equipment, with complex operating conditions and significant load fluctuations. Existing methods struggle to stably predict remaining life under conditions of real structural coupling, data correlation, and operating disturbances. Specifically, physical and data coupling edges are not simultaneously characterized and their weights are not fused within the same process, resulting in graph models failing to accurately represent system dynamics. Cross-modal attention weight allocation is fixed or empirically set, lacking an adaptive adjustment mechanism based on propagation results. The training phase lacks joint constraints on energy budget consistency and edge weight sparsity, leading to insufficient model robustness. Furthermore, the lack of linkage between online prediction and maintenance thresholds affects the timeliness of maintenance decisions. Consequently, the instability of prediction results directly reduces equipment availability and increases maintenance costs.

[0057] For example, in the monitoring of wind turbine gearboxes, when wind speed changes abruptly, the correlation between vibration and temperature signals changes due to operational disturbances. However, existing methods still employ a fixed weight allocation strategy, and the contribution of each mode is not dynamically adjusted. Furthermore, conflicts arise between physical coupling edges (such as gear meshing pairs) and data coupling edges (such as statistical correlations based on historical features) during edge weight fusion. Due to the lack of a data correction mechanism, the constructed graph structure is distorted. During the feature propagation stage, fixed attention weights cause the degenerate state representation to fail to focus on key information, accumulating remaining lifetime prediction biases. Under continuous operational fluctuations, this bias may lead to false alarms or missed alarms, forcing maintenance personnel to perform unnecessary downtime inspections or delaying the handling of actual faults.

[0058] If the above problems are not addressed, the unreliability of gear life prediction will increase the risk of unplanned equipment downtime and the likelihood of equipment damage, thereby affecting the safe operation and economic benefits of the entire system. The instability of prediction results makes maintenance decisions lack a reliable basis, potentially leading to wasted maintenance resources or failure of critical components, severely limiting the practical application value of equipment condition monitoring technology.

[0059] For this, please refer to Figure 1As shown, this application proposes a gear life prediction method based on a multimodal spatiotemporally coupled graph neural network, including:

[0060] S100: Collects multimodal data for preprocessing and extracts multimodal feature sequences according to a sliding time window. The multimodal data includes vibration data, acoustic data, temperature data, rotational speed data, torque data, and oil state data.

[0061] S200: Within each sliding time window, the data acquisition node and the measured gear node are used as the node set. The physical coupling edge is determined based on the structural connection, contact relationship and vibration transmission path. The data coupling edge is determined based on the statistical correlation and causal relationship of historical characteristics. The edge weights of the physical coupling edge and the data coupling edge are fused to construct a spatiotemporal coupling graph.

[0062] S300: Extract time-related information from multimodal feature sequences and propagate features on a spatiotemporal coupling graph. Based on the propagation results, assign cross-modal attention weights to different modes to obtain a representation of the degradation state of the gear under test.

[0063] S400: The parameter set used for inference is obtained based on data with remaining lifetime annotations. The parameter set is obtained while simultaneously constraining the remaining lifetime prediction error, degradation state fitting error, energy budget consistency, and edge weight sparsity.

[0064] S500: During operation, input the newly formed multimodal feature sequence and spatiotemporal coupling diagram, output the remaining life of the tested gear, and generate maintenance suggestions when the remaining life is less than the maintenance threshold.

[0065] Specifically, multimodal data refers to the measured values ​​of various physical quantities obtained from the gear system. This can be achieved through real-time acquisition using sensor networks or extraction from historical databases. Examples include acquiring vibration data via accelerometers, acoustic data via microphones, temperature data via thermocouples, rotational speed data via encoders, torque data via torque sensors, and oil state data via oil analyzers. This ensures that the input data covers multi-dimensional information about the gear's operating state. Furthermore, physical coupling edges refer to the connection relationships established based on the physical structure of the gear system. These can be identified using mechanical design drawings or 3D modeling software. For example, structural connections can be determined based on mechanical principles, contact relationships can be derived through contact mechanics models, or vibration transmission paths can be set based on vibration transmission theory. Specifically, this approach directly reflects the inherent connections between physical entities without relying on historical data. Data coupling edges refer to statistical correlations mined from historical data. These can be achieved using Pearson correlation coefficient calculations or Granger causality tests. For example, linear correlations of feature pairs can be identified in time series analysis, nonlinear dependencies can be assessed through mutual information, or causal directions can be determined using event sequence matching. As a preferred implementation method, this approach focuses on the intrinsic laws of the data and is independent of the physical model. Edge weight fusion refers to the process of combining the weights of physically coupled edges and data coupled edges. This can be achieved using weighted averaging or maximum value selection methods. For example, a fixed ratio can be set based on domain expert experience, dynamically adjusted using fuzzy logic, or the fusion coefficient can be solved through optimization algorithms. This ensures that the fusion result retains both physical reliability and compatibility with data-driven characteristics. Cross-modal attention weight allocation refers to a mechanism for dynamically adjusting the contribution of features from different modalities. This can be achieved using the softmax function or based on modal reliability indicators. For example, attention scores can be generated based on feature similarity, fixed weights can be allocated using historical performance indicators, or the weight distribution can be automatically learned through neural network layers. Specifically, this allocation process relies on aggregated information after feature propagation rather than pre-defined rules. Energy budget consistency refers to the constraint that verifies whether the energy changes in the system conform to physical laws. This can be achieved using energy conservation equation verification or power balance comparison. For example, by monitoring the difference between input and output power, analyzing the trend of thermodynamic parameter changes, or verifying the energy transfer direction using a kinetic model, this ensures that the prediction results meet the requirements of physical interpretability. Edge weight sparsity refers to the process of reducing the number of low-contribution edges in a graph. This can be achieved using threshold filtering or L1 regularization constraints. For example, setting an absolute weight threshold to eliminate weak connections, applying sparse coding techniques to compress the edge set, or limiting the number of connections based on node degree distribution are preferred implementation methods. This process reduces model complexity during the training phase, thereby improving generalization ability. This application solves the problem in existing technologies of difficulty in characterizing dual coupling relationships within the same process by simultaneously constructing physically coupled edges and data coupled edges and fusing edge weights.By employing cross-modal attention weight allocation based on feature propagation results, the model overcomes the lack of adaptability caused by fixed weight allocation. Joint constraints on energy budget consistency and edge weight sparsity enhance the model's robustness under operating condition disturbances. By linking the remaining lifespan output with maintenance thresholds to generate maintenance suggestions, the model directly translates prediction results into operational decisions, thereby providing stable and reliable gear lifespan prediction capabilities without disassembly.

[0066] Specifically, in the health monitoring of gear transmission systems, multimodal data is collected and preprocessed, covering vibration, acoustics, temperature, rotational speed, torque, and oil state data. Feature sequences are then extracted according to sliding time windows. Within each sliding time window, the data acquisition node and the tested gear node form a node set. Physical coupling edges are determined based on structural connections, contact relationships, and vibration transmission paths. Simultaneously, data coupling edges are determined based on the statistical correlation and causal relationships of historical features. The edge weights of the two types of edges are fused to construct a spatiotemporal coupling graph. Further, temporal correlation information is extracted from the multimodal feature sequences, and feature propagation is performed on the constructed spatiotemporal coupling graph. Based on the propagation results, cross-modal attention weights are assigned to different modes to obtain a representation of the degradation state of the tested gear. Based on historical data labeled with remaining lifetime, a parameter set for inference is obtained. This process simultaneously constrains remaining lifetime prediction error, degradation state fitting error, energy budget consistency, and edge weight sparsity. During operation, the newly formed multimodal feature sequence and spatiotemporal coupling diagram are input, and the remaining life of the tested gear is output. When the remaining life is less than the maintenance threshold, maintenance suggestions are generated.

[0067] Specifically, in the monitoring example of a wind turbine gearbox, vibration data was collected by a piezoelectric accelerometer, acoustic data was acquired through an electret microphone, and temperature data was measured using a resistance temperature detector (RTD). The sliding time window was set to 5 seconds. Physical coupling edges were established based on the relationship between the gear meshing pairs and the bearing support, while data coupling edges were determined based on the cross-correlation coefficient between vibration signals and temperature data. After edge weight fusion, features were propagated multiple times on the spatiotemporal coupling graph. Attention weights were dynamically adjusted based on the stability and consistency of each mode during propagation, and degenerate state representations were obtained through node-level and hierarchical weighted convergence.

[0068] Therefore, by simultaneously characterizing both physical and data coupling edges and fusing edge weights, the problem of stably predicting the remaining life of gears under conditions of real structural coupling, data correlation, and operational disturbances is solved, improving the adaptability of the graphical model to noisy environments. The adaptive cross-modal attention weight allocation mechanism based on propagation results ensures that the degenerate state representation focuses on key information, avoiding bias caused by fixed weight allocation. The joint constraints introduced during parameter set acquisition enable the model to balance prediction accuracy and physical rationality, enhancing generalization performance. Intelligent linkage with maintenance thresholds enables accurate generation of maintenance suggestions, and window consistency checks reduce false alarm rates, ensuring the reliability of maintenance decisions. Taking a monitoring period from 710 to 830 as an example, the proposed method predicts the remaining life of gears as follows: Figure 2 As shown. The mean remaining lifetime of the method in this paper is compared with the true value as follows. Figure 3 As shown.

[0069] Understandably, for the gear dataset, from Figure 2 It can be seen that as the monitoring time increases, its probability density function becomes narrower, the uncertainty decreases, the variance becomes smaller, and the prediction results are closer to the actual lifespan. Figure 3 It can be seen that as the running time increases, the error between the prediction and the actual remaining useful life of all methods decreases with the increase of monitoring data, and the remaining useful life predicted by the method in this paper is more accurate.

[0070] In some of the embodiments described above in this application, it is proposed to collect multimodal data, preprocess it, and extract multimodal feature sequences according to a sliding time window to support the construction of a gear life prediction model. However, in this process, the asynchronous timestamps of the multimodal data lead to time misalignment of the feature sequences, the misalignment of spatial coordinates causes distortion of physical position relationships, background noise and outliers interfere with the accuracy of feature extraction, the lack of constraints in missing data completion causes data distortion, and the failure to screen low-stability features during feature extraction leads to noise interference analysis, making the feature sequences unreliable. This, in turn, affects the accuracy of the physical coupling edges and data coupling edges in the spatiotemporal coupling graph construction, and ultimately reduces the robustness of the remaining life prediction.

[0071] In response, this application further proposes that when collecting and preprocessing multimodal data and extracting multimodal feature sequences according to a sliding time window, the process includes:

[0072] The timestamps of each channel are synchronized using a unified time base, and the relative positions of the acquisition nodes and the gears under test are recorded and spatially aligned using unified spatial coordinates.

[0073] The multimodal data is subjected to background noise estimation, denoising, zero drift correction, outlier suppression, missing data completion, and unit unification in a preset order. Outlier suppression is determined based on amplitude range and rate of change, while missing data completion is constrained based on the consistency between adjacent time periods and related modes.

[0074] The data is segmented according to a sliding time window, with the starting point of each window aligned with a unified time reference. Within each sliding time window, time-domain features, frequency-domain features, envelope correlation features, and operating condition features of the multimodal data are extracted. Low-stability features are then removed based on feature stability and feature correlation to obtain a multimodal feature sequence.

[0075] Specifically, a unified time reference refers to a reference time source used to align the acquisition times of different sensors. This can be achieved using GPS time synchronization or a precise time protocol. Its purpose is to ensure precise alignment of all modal data along the time axis, avoiding feature distortion caused by differences in sampling time. A unified spatial coordinate system defines the coordinate system relative to the sensor and the gear being measured. This can be achieved using a global coordinate system based on a computer-aided design model or a local coordinate system established by laser ranging. Its purpose is to clarify the physical positional relationships, making data from different sensors comparable in spatial dimensions.

[0076] Background noise estimation refers to the quantification process of environmental interference characteristics, which can be achieved based on historical noise databases or real-time environmental monitoring. Its purpose is to separate effective signals from background interference. Outlier suppression refers to the identification and handling of abrupt changes in data. It can be dynamically determined based on amplitude thresholds and rates of change to avoid misjudging normal fluctuations. Missing data completion refers to the intelligent filling of missing values. It can be achieved using time series interpolation or multimodal data association constraints. Its purpose is to ensure data continuity without introducing distortion.

[0077] In practical applications, time-domain characteristics refer to statistical quantities reflecting the instantaneous properties of a signal, which may include root mean square (RMS) values ​​and kurtosis, used to capture degradation trends. Frequency-domain characteristics refer to the signal's performance in the frequency domain, which may include spectral kurtosis and sideband energy, used to reveal periodic fault modes. Envelope correlation characteristics refer to features specific to modulated signals, which may include envelope spectral entropy, used to identify early damage. Operating condition characteristics refer to parameters related to operating conditions, which may include load coefficients and speed fluctuation rates, used to correlate operating states.

[0078] Specifically, this application's solution eliminates time misalignment by synchronizing the timestamps of each channel using a unified time base. Spatial alignment is achieved through unified spatial coordinates to ensure accurate physical location relationships. Preprocessing steps are executed in a preset order to ensure the continuity of processing logic. These steps include background noise estimation and denoising to improve the signal-to-noise ratio, zero-drift correction to eliminate sensor drift, outlier suppression to mitigate dynamic recognition interference, and missing data completion using multimodal correlation for intelligent filling. Unit unification ensures comparability of data with different dimensions. A sliding time window is used to segment and align with the time base, guaranteeing time consistency. Multi-dimensional feature extraction comprehensively captures state information, and feature filtering removes unstable features while retaining features sensitive to degradation. These steps are executed sequentially and mutually supportive, standardizing the preprocessing flow and addressing the issue of insufficient reliability of feature sequences.

[0079] As a specific implementation method, this application employs a network time protocol to synchronize the timestamps of each sensor in a wind turbine gearbox monitoring scenario, and uses a unified spatial coordinate system based on the gearbox's three-dimensional model to record sensor positions. During preprocessing, background noise estimation, wavelet threshold denoising, zero-drift dynamic compensation, outlier detection based on amplitude and rate of change, missing value completion using adjacent time windows and related modal data constraints, and unit standardization are performed sequentially. The system is segmented according to a preset duration sliding window, with the window start point aligned with the synchronized time reference. Within each window, the time-domain root mean square value, frequency-domain spectral kurtosis, envelope demodulation characteristics, and load condition characteristics of the vibration signal are extracted. Low-stability features are removed based on historical stability analysis to obtain a reliable multimodal feature sequence.

[0080] Through the above technical solutions, the reliability of multimodal feature sequences is improved, and the accuracy of the physical coupling edges and data coupling edges constructed by the spatiotemporal coupling graph is guaranteed, thereby enhancing the robustness of remaining lifetime prediction.

[0081] Specifically, in some of the embodiments described above in this application, a method is proposed to determine physical coupling edges based on structural connections, contact relationships and vibration transmission paths to construct a spatiotemporal coupled graph. However, in its implementation, the edge weights may experience non-physical fluctuations due to operating condition disturbances or measurement noise, resulting in instability of the graph structure in the time dimension, which in turn leads to feature propagation distortion and a decrease in the reliability of lifetime prediction.

[0082] In this regard, this application further proposes that when determining the physical coupling edge based on structural connections, contact relationships, and vibration transmission paths, the following should be included:

[0083] The connection relationship is established based on gear meshing pairs, coaxial connections, bearing supports and housing constraints. The direction of the edge is determined based on the contact position and vibration transmission path. The weight of the edge is set in stages according to the contact type, assembly preload state, support stiffness and vibration transmission path level.

[0084] The edge weights are tested for stability within multiple consecutive sliding time windows, and the physical coupling edges are retained only when the edge weights fluctuate within a preset fluctuation range.

[0085] Among them, gear meshing refers to the physical meshing relationship between gear pairs, which can be achieved through spur gear meshing, helical gear meshing, or worm gear meshing, etc. The purpose is to strictly rely on the inherent mechanical structure of the gear system, avoid subjective experience settings, and ensure that the connection relationship truly reflects the physical coupling essence. Contact position and vibration transmission path can be understood as the spatial transmission trajectory of vibration energy in the mechanical structure. Specifically, the contact point can be located using a three-dimensional coordinate system, and the vibration transmission path can be determined by combining finite element analysis. The purpose is to ensure that the edge direction is consistent with the actual vibration transmission direction, ensuring that the characteristic propagation conforms to physical laws rather than random diffusion. Contact type, assembly preload state, support stiffness, and vibration transmission path level are key physical parameters that affect the coupling strength. Contact type can be manifested as point contact or surface contact, assembly preload state reflects the tightness of bolts, support stiffness characterizes the rigidity level of bearings or brackets, and vibration transmission path level describes the structural hierarchy of vibration transmission. The purpose is to quantify the impact of these parameters on vibration transmission efficiency and set edge weights hierarchically, avoiding insufficient adaptability of fixed weights when operating conditions change. Edge weight stability testing refers to a monitoring mechanism for the temporal changes of edge weights. It can use a sliding window statistical method to evaluate fluctuation characteristics. Its purpose is to filter out instantaneous noise interference caused by sudden changes in operating conditions or sensor jitter, and retain only stable edges that conform to physical laws.

[0086] Specifically, this scheme first establishes connection relationships based on gear meshing pairs, coaxial connections, bearing supports, and housing constraints, ensuring that the relationships between nodes strictly correspond to the mechanical structural entities. Then, the direction of the edges is determined based on the contact position and vibration transmission path, ensuring that the feature propagation direction aligns with the directional transmission characteristics of vibration energy. Next, edge weights are hierarchically set based on contact type, assembly preload state, support stiffness, and vibration transmission path hierarchy, allowing the weight values ​​to accurately reflect the differences in vibration transmission efficiency under different physical states. Finally, edge weight stability checks are performed over multiple consecutive sliding time windows. Non-physical fluctuations are eliminated through temporal continuity verification, retaining only physically coupled edges that meet the preset fluctuation range. These steps form a closed-loop mechanism, ensuring that the construction of physically coupled edges conforms to structural constraints and possesses temporal robustness, thus providing a stable and reliable graph structure foundation for feature propagation.

[0087] As a specific implementation method, the solution of this application is implemented as follows: In the wind turbine gearbox monitoring scenario, for the planetary gear transmission system, firstly, a connection relationship is established based on the meshing pair of planetary gears and the sun gear, the coaxial connection between the planet carrier and the planetary gears, the bearing support relationship, and the gearbox constraints. Secondly, the direction of the edge is determined based on the spatial coordinates of the gear meshing point and the vibration transmission path in the gearbox. Next, the edge weight is set hierarchically based on the gear point contact characteristics, bearing preload state, support stiffness, and the vibration transmission level from the gear to the gearbox. Finally, the edge weight change is monitored in a continuous sliding time window sequence, and the physical coupling edge is retained only when the weight fluctuation remains within the preset fluctuation range. In this embodiment, the edge weight stability verification module can be integrated into an industrial edge computing device for real-time temporal fluctuation analysis.

[0088] Through the above solution, this application solves the problem of non-physical fluctuations in edge weights caused by operating condition disturbances or measurement noise, ensures that the spatiotemporal coupling graph maintains temporal consistency during dynamic operation, avoids distortion in the feature propagation process, and thus improves the long-term reliability of gear life prediction.

[0089] In some of the embodiments described above in this application, physical coupling edges are proposed to determine the graph structure based on structural connections, contact relationships, and vibration transmission paths. However, in its implementation, relying solely on the physical structure cannot adapt to fluctuations in operating conditions and dynamic changes in data, leading to inaccuracies in the graph structure under complex operating environments. Furthermore, the introduction of data coupling edges is susceptible to noise interference and causal ambiguity, resulting in unstable edge weights and conflicts between physical and data conclusions. A mechanism is lacking to reliably fuse and dynamically verify the two types of edges.

[0090] In response, this application further proposes a method for determining data coupling edges based on statistical correlations and causal relationships of historical characteristics, and for fusing the edge weights of physical coupling edges and data coupling edges, including:

[0091] The correlation stability is obtained by judging whether the direction and strength of the linkage of the same feature pair within a continuous sliding time window of the multimodal feature sequence remain within a preset fluctuation range. The consistency of delayed response is obtained by comparing the temporal relationship between the features that change first and the features that respond later within the continuous sliding time window. Based on the event time points triggered by the multimodal features and combined with start / stop, alarm, and load changes in the operation logs and maintenance records, the consistency of the co-occurrence and sequence of events is statistically analyzed within the continuous sliding time window to obtain the event linkage consistency.

[0092] Data coupling edges are determined based on correlation stability, lag response consistency, and event linkage consistency, and edge weights are assigned to the data coupling edges.

[0093] The weights of physically coupled edges and data coupled edges are standardized in terms of dimensions and range. They are then weighted and merged in order of physical priority. When the conclusions of the two types of edges conflict, the physically coupled edges are retained and the results are adjusted according to the data coupled edges.

[0094] The edge weights after fusion are sparsified, and edges that do not meet the requirements of edge weight stability and window consistency verification are removed in ascending order of weight value. Before entering the next sliding time window, the retained edges are checked based on the window consistency verification results, and finally the spatiotemporal coupled graph is obtained.

[0095] Among these, correlation stability refers to the stability index of the linkage relationship between features. It can be assessed using the standard deviation of the Pearson correlation coefficient within a rolling window, aiming to filter out single-window noise interference and ensure that the data coupling edges are based on persistent statistical regularities rather than random fluctuations. Lag response consistency can be understood as a stability measure of the temporal causal relationship of feature changes. It can be achieved by determining the fluctuation range of peak delay in cross-correlation analysis, aiming to verify the ability of the data coupling edges to reflect the real physical process and prevent spurious correlations caused by time misalignment. Event linkage consistency is specifically a quantitative representation of the correlation strength between features and operational events. It can be achieved using a dynamic time warping algorithm for event sequences, aiming to bind data features with actual operating events and improve physical interpretability. Dimensional unification and range normalization refer to eliminating dimensional differences in weights from different sources. It can be achieved using min-max normalization combined with logarithmic transformation, aiming to establish comparable scales to support fair fusion. Physically-prioritized weighted fusion can be understood as a mechanism where physical information dominates the fusion strategy. This can be achieved by dynamically adjusting the physical weight coefficient to at least 1.5 times the data weight coefficient, prioritizing the structural reliability while allowing data-driven fine-tuning. Edge weight sparsity specifically focuses on optimizing critical coupling paths. This can be achieved using adaptive threshold truncation based on network topology density, aiming to remove weak connection noise and improve graph structure robustness. Window consistency verification refers to a dynamic verification process over time, which can be achieved by checking the continuity of edge weight correlation coefficients between sliding windows, ensuring the graph structure remains stable during dynamic operation.

[0096] Specifically, this application's solution assesses the multi-dimensional stability of multimodal feature sequences within a continuous sliding time window. First, it determines correlation stability by analyzing the range of fluctuations in linkage direction and strength. Then, it verifies the consistency of delayed responses by combining the temporal relationship between features that change first and those that respond later. Finally, it integrates event timing information from the operation log to form event linkage consistency. These three factors together constitute the basis for determining data coupling edges. Subsequently, it unifies the dimensions and normalizes the range of the edge weights of physical and data coupling edges. After eliminating dimensional differences, it performs weighted fusion according to the physical priority principle. When the conclusions of the two types of edges conflict, the physical coupling edge is retained, and its amplitude is finely adjusted based on the data coupling edge. On this basis, it performs sparsification on the fused edge weights, removing edges that do not meet the stability and window consistency checks by weight values ​​from smallest to largest. Before entering the next sliding time window, the retained edges are dynamically reviewed based on the verification results. This process, through comprehensive judgment of multi-source evidence and continuous verification over time, achieves the organic integration of prior physical structures and dynamic data characteristics, ensuring that the spatiotemporal coupling diagram accurately reflects the true coupling relationship under complex operating conditions.

[0097] As a specific implementation method, this application is implemented as follows: In the wind turbine gearbox monitoring scenario, after the multimodal feature sequences collected by vibration sensors and temperature sensors are processed by a continuous sliding time window, the correlation stability is determined by calculating the fluctuation range of the Pearson correlation coefficient for the vibration amplitude and oil viscosity feature pair through a rolling window. Cross-correlation analysis is used to identify the time delay stability of vibration mutation leading temperature change to obtain the consistency of hysteresis response. Combining the load mutation events recorded by the SCADA system and the alarm time points in the maintenance log, the co-occurrence frequency and sequence consistency of the statistical feature trigger events and operation events form the event linkage consistency. After establishing the data coupling edge based on the comprehensive judgment of the three, the weight of the physical coupling edge (set according to the stiffness classification of the gear meshing pair) and the weight of the data coupling edge are normalized to the [0,1] interval through logarithmic transformation, and weighted fusion is performed according to the principle that the proportion of physical weight is not less than 70%. An adaptive threshold truncation based on network node degree is implemented on the fusion weights to remove weak connection edges with weights lower than the dynamic threshold. Before window switching, it is verified whether the weight correlation coefficient of the retained edges in adjacent windows is higher than 0.85. Edges that fail the verification are temporarily frozen and await review. The final spatiotemporal coupling graph represents the dynamic coupling characteristics of the gear meshing pair under variable load conditions.

[0098] Through the above technical solutions, this application realizes the reliable integration of physical coupling edges and data coupling edges under dynamic working conditions. The multi-dimensional stability assessment avoids the influence of noise interference and causal ambiguity on data coupling edges. The physical priority fusion mechanism solves the graph structure distortion problem caused by the conflict of conclusions between the two types of edges. The edge weight sparsification processing and window consistency verification ensure the stability of the spatiotemporal coupling graph in complex operating environments, thereby improving the accuracy of the assessment of the degradation state of the tested gear.

[0099] In some of the embodiments described above in this application, it is proposed to extract time-related information from multimodal feature sequences and propagate features on a spatiotemporal coupled graph. However, in the implementation process, the feature propagation operation may cause information distortion due to time-related conflicts. For example, the dynamic consistency of trend changes, periodic changes and abrupt events is not considered during the propagation process, which makes the node aggregation results affected by noise interference or logical contradictions, thereby reducing the reliability of the gear degradation state representation.

[0100] In response, this application further proposes steps for extracting temporal correlation information from multimodal feature sequences and propagating features on a spatiotemporal coupling graph, including:

[0101] Within each sliding time window, time-related information is extracted for each modality feature. This time-related information includes trend changes, periodic changes, and abrupt events.

[0102] After extracting the temporal correlation information, the edge weights fused in the spatiotemporal coupled graph are used as the basis for propagation. One to several rounds of feature propagation are performed between adjacent nodes in descending order of edge weight, with the number of propagation rounds falling within a preset range. After each round of propagation, if the temporal correlation conflict caused by the propagation exceeds a preset conflict threshold, the contribution of that edge is suppressed. After completing all propagation rounds, the node temporal correlation aggregation result is obtained.

[0103] Specifically, temporal correlation information refers to the dynamic evolution pattern of multimodal features over time. This can be achieved using sliding window statistical analysis, wavelet transform, or multi-scale decomposition. The aim is to distinguish between long-term degradation trends, periodic load fluctuations, and sudden failure events, avoiding the loss of temporal information due to reliance solely on static features. Feature propagation in descending order of edge weights can be understood as prioritizing the transmission of information from high-confidence connections. This can be implemented using priority queues or sorting algorithms, ensuring that high-weight information from key connections in physical and data coupling is transmitted first, preventing excessive information diffusion and noise accumulation. In practical applications, keeping the number of propagation rounds within a preset range specifically refers to limiting the number of iterations in feature propagation. This can be achieved using a fixed upper limit or dynamic termination conditions, aiming to maintain the stability of the propagation process and ensure computational efficiency. Specifically, suppressing the contribution of an edge when the temporal correlation conflict caused by propagation exceeds a preset conflict threshold refers to dynamically filtering distorted information. This can be achieved using threshold comparators or adaptive suppression mechanisms, aiming to provide real-time feedback on the logical consistency of the time dimension and prevent distorted information such as trend reversals, periodic misalignments, or contradictory abrupt events from contaminating the aggregation results. Among them, the node time correlation aggregation result can be understood as a node representation that integrates time-series dynamic information. It can be implemented by weighted average or max pooling, with the aim of forming an aggregated representation that accurately reflects the degradation and evolution law of gear nodes.

[0104] Specifically, the proposed solution extracts refined temporal correlation information, including trend changes, periodic changes, and abrupt events, within each sliding time window, providing precise temporal guidance for feature propagation. Subsequently, based on the weights of the fusion edges, multiple rounds of feature propagation are performed according to weight priority, ensuring that key connection information is transmitted preferentially while maintaining stability by limiting the range of propagation rounds. After each round of propagation, temporal correlation conflicts are evaluated in real time, and the contribution of highly conflicting edges is suppressed to filter out distorted information. Finally, the conflict-suppressed temporal dynamic information is integrated to form a node temporal correlation aggregation result.

[0105] As a specific implementation method, in the wind turbine gearbox monitoring scenario, the signals collected by vibration sensors are processed using wavelet transform to extract periodic meshing frequency components as periodic variation features. Temperature sensor data is used to obtain trend variation features through linear regression, while sudden changes in oil state are identified as abrupt events through threshold detection. During feature propagation, high-weighted edge information between gear meshing pairs is prioritized. When a propagation causes a contradiction between the temperature trend and the direction of vibration amplitude change, the contribution weight of that edge is automatically reduced. Ultimately, the aggregated results of each node accurately reflect the degradation state of the gear meshing region.

[0106] The above scheme solves the problem of information distortion caused by temporal correlation conflicts during feature propagation, ensures the accuracy and robustness of node aggregation results, thereby improving the reliability of gear degradation state representation and providing a solid foundation for remaining lifetime prediction.

[0107] Specifically, in some of the embodiments described above in this application, cross-modal attention weight allocation is proposed to obtain a representation of the degradation state of the gear under test by means of propagation results. However, in its implementation, the reliability and consistency of different modes are not dynamically considered, which may result in the degradation state representation being affected by noise or conflicting information and thus inaccurate.

[0108] In response, this application further proposes that when allocating cross-modal attention weights to different modalities based on the propagation results to obtain a representation of the degenerate state of the gear under test, the following steps are included:

[0109] Non-negative attention weights are set based on the node time correlation aggregation results, correlation stability, consistency of delayed response and event linkage, and noise level of each modality in historical samples, and the attention weights are normalized within the same node.

[0110] When the correlation stability of a certain mode is insufficient or conflicts with the operating condition indication within a continuous sliding time window, its attention weight is reduced; when multiple modes jointly indicate increased degradation, the corresponding attention weight is increased. The results of each mode after attention weight allocation are weighted and aggregated within the node to obtain the node-level degradation state results. Then, weighted and aggregated between nodes according to the meshing pair and vibration transmission path levels to obtain the degradation state representation of the tested gear.

[0111] Among them, the node temporal correlation aggregation result refers to the aggregated representation of the temporal dynamic characteristics extracted during feature propagation at the node. This can be achieved using statistical features of trend changes, periodic changes, and abrupt events, aiming to reflect the temporal correlation pattern at the node. Correlation stability can be understood as a stability index of the direction and strength of linkage between multimodal feature pairs within a continuous time window. This can be achieved by calculating the fluctuation range of the correlation coefficient, aiming to identify reliable statistical correlations. Lag response consistency specifically measures the consistency of the temporal order between features that change first and those that respond later within a continuous window. For example, it can be achieved based on the time delay stability of the Granger causality test, aiming to verify the reliability of causal relationships. Event linkage consistency refers to the consistency of the co-occurrence and chronological order of events triggered by multimodal features. This can be achieved using the co-occurrence frequency and sequence matching degree of event sequences, aiming to capture coordinated changes at the event level. Noise level can be understood as a data quality index for each modality in historical samples, specifically based on the power ratio of signal to noise or the proportion of outliers, aiming to quantify the reliability of the modality. Attention weight setting and normalization specifically refers to dynamically allocating non-negative weights based on the aforementioned indicators and normalizing them within nodes. This can be achieved, for example, using weighted summation and softmax normalization, aiming to ensure reasonable weight distribution and coordinated contributions from each mode. Dynamic weight adjustment conditions refer to reducing weights when modal reliability decreases or conflicts with operating conditions, and increasing weights when multimodal collaboration indicators degrade. This can be achieved based on threshold comparisons and logical judgments, aiming to optimize weight allocation in real time. Intra-node weighted aggregation can be understood as integrating multimodal results at the same physical location according to weights. This can be achieved using linear weighted averaging, aiming to form a local degradation representation. Inter-node weighted aggregation specifically involves hierarchical aggregation based on the physical structure of the gear system. For example, it can be achieved based on the weights of meshing pair relationships and vibration transmission paths, aiming to output an overall state assessment that conforms to the energy transmission path of the mechanical system.

[0112] Specifically, the proposed scheme sets initial attention weights based on node time correlation aggregation results, correlation stability, hysteresis response consistency, event linkage consistency, and noise level to ensure that the weights reflect the time correlation pattern and modal reliability of the current state. Subsequently, the weights are dynamically adjusted within a continuous sliding time window. When the correlation stability of a certain mode is insufficient or conflicts with the operating condition indication, its weight is reduced; when multiple modes jointly indicate increased degradation, the corresponding weight is increased, thereby suppressing unstable modes and strengthening collaborative degradation signals. Finally, node-level degradation state results are formed through intra-node weighted aggregation, and inter-node weighted aggregation is performed according to the meshing pair and vibration transmission path levels. This ensures that the degradation state representation conforms to both local multimodal integration and the system's physical structure, thus outputting a representation that accurately reflects the actual degradation state of the gear.

[0113] As a preferred embodiment, the solution of this application is implemented as follows: In the wind turbine gearbox monitoring scenario, when the correlation stability between the vibration mode and the temperature mode is lower than a preset threshold within three consecutive sliding time windows, the attention weight of the vibration mode is automatically reduced. When the vibration, acoustic, and oil state modes simultaneously show abnormal amplitude increases and the consistency of hysteresis responses is higher than a preset threshold, the attention weight of these modes is increased. Intra-node weighted convergence uses the sum of the product of each mode weight and its eigenvalue, while inter-node weighted convergence allocates weights according to the contact stiffness of the gear meshing pair and the length of the vibration transmission path, ultimately outputting a representation reflecting the overall degradation state of the gear.

[0114] Through the above scheme, this application can dynamically optimize the integration process of multimodal information, suppress the interference of noise modes and conflicting information, and strengthen the degradation trend of multi-source signal collaborative indication, thereby obtaining a more accurate and reliable representation of gear degradation state and providing high-quality input for remaining lifetime prediction.

[0115] Specifically, in some of the embodiments described above in this application, a gear life prediction method based on a multimodal spatiotemporally coupled graph neural network is proposed to generate a representation of the degradation state. However, in its implementation, the acquisition of the parameter set depends only on the optimization of the remaining life prediction error and the degradation state fitting error. It lacks a joint constraint mechanism for the consistency of energy expenditure and the sparsity of edge weights, which leads to a high risk of the model violating physical laws under operating condition disturbances and a redundant graph structure, thus causing the problem of insufficient robustness of the prediction results.

[0116] In this regard, this application further proposes the following steps:

[0117] The historical samples are divided into training windows and validation windows according to the sliding time window.

[0118] The initial values ​​of the parameters are determined based on the degradation state representation and remaining lifetime label within the training window.

[0119] The parameter set is obtained iteratively in batch updates. After each update, the remaining life prediction error, degradation state fitting error, energy balance consistency, and edge weight sparsity are evaluated simultaneously. The energy balance consistency is determined by comparing whether the common trend of speed change, torque change, temperature change, and oil state change is consistent with the energy transmission direction of the gear under test. The edge weight sparsity is constrained by removing low-contribution edges according to edge weight from small to large and limiting the number of retained edges of each node to within a preset range.

[0120] When the four evaluation metrics in the verification window meet the preset thresholds for several consecutive rounds, the updates stop and a parameter set for inference is obtained. The parameter set includes attention weights, edge weights, parameters related to remaining lifetime output, and parameters related to degradation state representation.

[0121] Among these, energy budget consistency refers to a mechanism that constrains the physical rationality of a model by verifying whether the changing trends of multimodal physical quantities conform to the energy conservation law of a mechanical system. This can be determined by comparing the common trends of changes in rotational speed, torque, temperature, and oil state with the energy transfer direction of the tested gear. Its purpose is to prevent anomalous predictions under operating conditions and ensure that the model output conforms to physical laws. Edge weight sparsity refers to a constraint mechanism that dynamically simplifies coupling relationships based on the contribution of graph structures to avoid redundant edges introducing noise interference. This can be achieved by removing low-contribution edges according to their weights from smallest to largest and limiting the number of retained edges for each node to a preset range. Its purpose is to maintain the integrity of key physical coupling paths while improving the model's generalization ability. Batch update iteration refers to a method of processing training data in batches to gradually optimize parameters. This can employ stochastic gradient descent, aiming to balance training efficiency and parameter convergence stability.

[0122] Specifically, this application's scheme divides historical samples into training and validation windows using a sliding time window, enabling the model to validate the adaptability of parameters to unseen data in real time during the training phase, avoiding overfitting issues caused by relying solely on training data. Initial parameter values ​​are determined based on the degradation state representation and remaining lifetime label within the training window, and the parameters are initialized using the physical correlation between the actual degradation state and lifetime label. The parameter set is iteratively obtained using a batch update method, and after each update, the remaining lifetime prediction error, degradation state fitting error, energy budget consistency, and edge weight sparsity are simultaneously evaluated. Energy budget consistency is used to constrain the model's physical rationality by verifying whether the changing trends of physical quantities such as rotational speed and torque conform to the gear energy transmission direction. Edge weight sparsity is addressed by simplifying the graph structure by removing low-contribution edges according to their weights and limiting the number of nodes. Updates stop when all four indicators within the validation window continuously meet their thresholds, ensuring the stability of the parameter set on actual operating data, thus forming an inference parameter set that balances prediction accuracy, physical rationality, and structural simplicity.

[0123] As a specific implementation method, the scheme of this application is implemented as follows: In the wind turbine gearbox monitoring system, historical operating data is divided into a training window and a validation window according to a sliding time window. The initial parameter values ​​are determined by the least squares method based on the degradation state representation and remaining lifetime label of the training window. A deep learning framework is used for batch update iteration. After each round of update, the remaining lifetime prediction error and degradation state fitting error are calculated, and the consistency of energy expenditure is verified by comparing the changing trends of speed, torque, temperature, and oil state. At the same time, the edge weights are sparsified. Training stops when the four indicators of the validation window meet the preset thresholds for multiple consecutive rounds, resulting in a parameter set for online prediction.

[0124] Through the above solution, this application solves the robustness defects caused by the violation of physical laws and the redundancy of graph structure in the parameter acquisition process, so that the model can maintain physical rationality and reduce redundant interference under operating condition disturbances, thereby improving the accuracy and reliability of gear remaining life prediction.

[0125] Traditional gear life prediction methods simultaneously constrain the remaining life prediction error, degradation state fitting error, energy budget consistency, and edge weight sparsity during parameter set acquisition. However, in this process, the synchronous optimization of multiple constraints is prone to instability in the training process, difficulty in convergence, or insufficient robustness of the parameter set due to potential conflicts between physical laws and data characteristics.

[0126] In response, this application further proposes a phased weighting strategy for obtaining the parameter set used for inference based on data with remaining lifetime annotations, including:

[0127] In the first stage, the optimization objectives are only the remaining lifetime prediction error and the degradation state fitting error, and the window consistency check is used as the stopping condition.

[0128] In the second stage, after the conditions of the first stage are met, the constraint weights for energy expenditure consistency and edge weight sparsity are increased. After each round of updates, edge weight stability and window consistency are checked. If the verification window shows performance degradation or consistency does not meet the preset consistency threshold, the parameter set of the previous round is rolled back and the growth rate of constraint weights is reduced.

[0129] After completing the second phase, the parameter set used for inference is finalized, and the threshold, number of windows, and backoff conditions used for retraining are recorded.

[0130] The phased weighting strategy involves dividing the parameter set acquisition process into optimization stages with clearly defined progressive objectives. This can be achieved by adjusting constraint weight coefficients in stages, aiming to avoid conflicts caused by simultaneous optimization of multiple constraints and prioritizing the establishment of basic predictive capabilities before gradually enhancing physical constraints. Window consistency verification can be understood as a dynamic monitoring mechanism for the stability of parameter performance within a continuous sliding time window. It can be based on the fluctuation range of prediction results in adjacent windows using historical samples, aiming to ensure the reliability of the parameter set in the time series dimension. Edge weight stability verification evaluates the fluctuation characteristics of weights after the fusion of physically coupled and data coupled edges. This can be achieved by comparing the standard deviation within the sliding window with a preset threshold, aiming to identify abnormal edge weight changes caused by operational disturbances. The rollback mechanism is specifically a parameter set state recovery operation triggered when the performance of the verification window does not meet requirements. It can be designed to save historical parameter snapshots and selectively call them based on verification results, aiming to prevent performance crashes caused by excessively rapid growth of constraint weights. The finalization of the parameter set refers to the solidification of parameter configurations that meet multi-dimensional verification standards. It may include locking operations for attention weights, edge weights, and parameters related to remaining lifetime output, with the aim of ensuring stability and reproducibility during the inference phase.

[0131] Specifically, this application's scheme achieves orderly optimization of the parameter set acquisition process through a phased weighting strategy. In the first phase, the model focuses solely on minimizing the remaining lifetime prediction error and the degradation state fitting error, avoiding interference from complex constraints such as energy budget consistency and edge weight sparsity, thereby quickly establishing a reliable initial solution. Window consistency verification serves as a stopping condition to ensure that this phase terminates promptly after the basic error has stabilized and converged, preventing premature introduction of additional constraints that could lead to a deviation in the optimization direction. In the second phase, based on the stable basic model formed in the first phase, the constraint weights for energy budget consistency and edge weight sparsity are gradually increased, making the parameter set more consistent with actual energy transfer patterns and sparse structure characteristics. After each round of edge weight stability verification and window consistency verification, the impact of constraint introduction on model performance is monitored in real time. When the verification window shows performance degradation or consistency fails to meet the preset threshold, a dynamic adjustment mechanism is used to revert to the previous round's parameter set and reduce the growth rate of constraint weights, ensuring a balance between physical consistency and prediction accuracy in the optimization process. After completing the second phase, the parameter set is finally finalized and the optimization results are solidified. Simultaneously, the retraining threshold, window number, and revert conditions are recorded, forming a complete closed loop for parameter set acquisition.

[0132] As a preferred embodiment, the specific implementation of this application's solution is as follows: In the wind turbine gearbox lifetime prediction scenario, historical samples are divided into training windows and validation windows according to a sliding time window. During the first training phase, the optimizer only calculates the weighted sum of the remaining lifetime prediction error and the degradation state fitting error. This phase terminates when the error fluctuation of five consecutive validation windows is less than 5%. After entering the second phase, the energy budget consistency constraint weight gradually increases from 0.1 to 0.5, and the edge weight sparsity constraint weight increases from 0.05 to 0.3. After each round of parameter updates, the standard deviation of the edge weights in the continuous window is automatically checked to see if it exceeds 0.08, and whether the window consistency check passes. If the remaining lifetime prediction error of the validation window increases by more than 8% after a certain round of updates, the system reverts to the previous parameter set and reduces the constraint weight growth rate by 50%. When the second phase is completed, the current parameter set is locked as the final inference parameters, and the window consistency check threshold of 0.05, the minimum number of validation windows of 3, and the revert trigger condition threshold of 8% are stored.

[0133] By employing the aforementioned approach, this application avoids the training instability problem caused by multi-constraint synchronous optimization, enabling an orderly transition between basic predictive capabilities and physical constraints in the parameter set acquisition process, thereby improving the model's convergence reliability under complex operating conditions. Simultaneously, the dynamic backoff mechanism and parameter set finalization strategy ensure the quality and stability of the inference parameters, enhancing adaptability to changes in operating conditions and the accuracy of maintenance decisions.

[0134] Specifically, in some of the embodiments described above in this application, a method is proposed to input newly formed multimodal feature sequences and spatiotemporal coupling diagrams during operation, output the remaining lifetime, and generate maintenance recommendations. However, in its implementation, the remaining lifetime prediction of a single window is easily affected by data noise and instantaneous fluctuations. Directly triggering maintenance recommendations based on thresholds can lead to false alarms (such as unnecessary downtime triggered by short-term fluctuations) or missed alarms (such as degradation trends being masked by noise). At the same time, there is a lack of dynamic control mechanism for the stability of recommendations, making it difficult to balance maintenance timeliness and decision reliability.

[0135] In response, this application further proposes a method for inputting newly formed multimodal feature sequences and spatiotemporal coupling diagrams during operation, outputting the remaining lifespan of the tested gear, and generating maintenance suggestions when the remaining lifespan is less than a maintenance threshold, including:

[0136] Within each sliding time window, the newly formed multimodal feature sequence is aligned with a unified time reference and a unified spatial coordinate, and the spatiotemporal coupling graph is reconstructed. The aligned multimodal feature sequence and the reconstructed spatiotemporal coupling graph are input into the parameter set to obtain the degradation state representation and remaining lifetime.

[0137] A window consistency check is performed on a consecutive preset number of sliding time windows. Maintenance suggestions are generated only when the remaining lifespan within the consecutive windows is not higher than the maintenance threshold and the degradation status indicates no improvement trend.

[0138] A maintenance preparation recommendation is generated when the remaining lifetime is within the maintenance threshold buffer and the window consistency check passes.

[0139] A shutdown recommendation is generated when the remaining lifetime continuously decreases and falls below the maintenance threshold within a consecutive window, accompanied by a continuous deterioration in the degradation status.

[0140] After any suggestion is generated, an alarm silence period is entered. During the alarm silence period, the suggestion is updated only if the new window consistency check result is inconsistent with the previous conclusion.

[0141] The unified time reference and unified spatial coordinate alignment refer to synchronizing multimodal data to a common reference point on the time axis and unifying the spatial coordinate system. Time synchronization can be achieved using a precise time protocol or global positioning system timestamps, while spatial alignment can be achieved using coordinate transformation matrices or laser ranging and positioning technology. The aim is to eliminate time offsets and spatial misalignments between sensors and avoid introducing additional noise. Window consistency verification can be understood as verifying the trend stability of prediction results within a continuous sliding time window. This can be achieved based on the monotonicity test of remaining life or time-series correlation analysis of degradation state representation. Its purpose is to filter out instantaneous fluctuations and ensure that decisions are based on a continuous degradation trend. The buffer range of the maintenance threshold specifically refers to the critical interval set around the maintenance threshold, which can be set to ±5% to ±15% of the maintenance threshold. Its purpose is to identify the critical state of equipment degradation and reserve buffer time for maintenance preparation. The alarm silence period refers to the decision freeze period set after generating a suggestion. It can be dynamically adjusted based on historical fault recovery cycles to suppress frequent suggestion jumps and maintain the executability of maintenance strategies.

[0142] Specifically, the proposed solution ensures baseline consistency of input data by performing spatiotemporal alignment and graph reconstruction operations within each sliding time window. Subsequently, window consistency checks are performed on a predetermined number of consecutive sliding time windows. Maintenance recommendations are triggered only when the remaining lifetime consistently falls below the maintenance threshold and the degradation status remains unresolved. A multi-window continuous verification mechanism filters out single-window noise interference. Simultaneously, based on the relative position of the remaining lifetime and the maintenance threshold, recommendations for routine maintenance, pre-maintenance maintenance, or shutdown are generated, achieving hierarchical decision-making. An alarm silence period is initiated after recommendation generation, updating the recommendations only when new evidence contradicts them, forming a closed-loop process of "data input – trend verification – hierarchical decision-making – silence maintenance." This mechanism eliminates input bias through spatiotemporal alignment, ensures trend reliability through continuous window verification, and achieves dynamic decision-making adjustment by combining buffer ranges and silence periods, thereby stably outputting maintenance recommendations without dismantling the system.

[0143] As a specific implementation method, this application is implemented as follows: In the wind turbine gearbox monitoring system, multimodal data such as vibration and acoustics are collected every 10 minutes to form a new feature sequence. A precise time protocol is used to align the timestamps of each sensor to a unified time reference, and the spatial coordinates are unified to the gear center reference system through a coordinate transformation matrix. Based on the aligned data, a spatiotemporal coupling graph is dynamically reconstructed, which includes the physical coupling edges of the gear meshing pairs and the data coupling edges of the vibration-temperature correlation. The reconstructed graph is input into a parameter set to obtain the current remaining lifetime prediction value of 135 hours. Window consistency verification is performed on the most recent four consecutive sliding time windows, and the remaining lifetimes are found to be 140 hours, 138 hours, 136 hours, and 135 hours, respectively, with the degradation state indicating continuous deterioration. Since the remaining lifetime is below the maintenance threshold of 150 hours and the consecutive window verification is passed, a routine maintenance recommendation is generated. At the same time, a 48-hour alarm silence period is entered. During this period, even if the remaining lifetime of a single window briefly rebounds to 142 hours, the original recommendation is maintained.

[0144] Through the above technical solutions, this application suppresses the interference of single-window data fluctuations on maintenance recommendations, avoiding false alarms and missed alarms caused by instantaneous noise. By employing window consistency verification and buffer range mechanisms, it accurately identifies critical equipment degradation states and generates hierarchical maintenance recommendations, balancing maintenance timeliness and decision reliability. The alarm silence period ensures the stability of the maintenance strategy, reduces unnecessary downtime, and improves the engineering practicality of gear transmission system condition monitoring.

[0145] Specifically, in some of the embodiments described above in this application, a method for predicting gear life based on a parameter set is proposed. However, in its implementation, when the working condition distribution changes or new data with remaining life labels appears, the model cannot be updated adaptively, resulting in decreased prediction accuracy and compromised stability.

[0146] In response, this application further proposes that when the operating condition distribution changes or new data with remaining lifetime labels appears, the edge weights, cross-modal attention weights and the parameter set are updated on a rolling basis based on the latest data, and the remaining lifetime output of historical samples before and after the update is compared with the consistency check threshold. If the consistency check threshold is not met, the system will revert to the state before the update.

[0147] Among these, "changes in operating conditions" refers to deviations in the operating environment or load conditions, which can manifest as an expansion in the range of speed fluctuations, an increase in the frequency of torque mutations, or a change in temperature distribution patterns. Its purpose is to promptly identify changes in the model's applicable boundaries. "Rolling updates" refers to the process of dynamically adjusting model parameters, which can be achieved using sliding window incremental learning or online Bayesian updates. Its purpose is to ensure the model continuously adapts to new operating condition characteristics. "Consistency verification threshold" refers to the criterion for quantitatively verifying the quality of updates. It can be set as the root mean square error rate of change in the remaining lifetime prediction results of historical samples or the trend consistency coefficient. Its purpose is to objectively assess the compatibility of the update operation with existing knowledge. "Rollback mechanism" refers to safety protection measures, which can be configured as parameter snapshot restoration or version control. Its purpose is to ensure service reliability during abnormal updates.

[0148] Specifically, the proposed solution uses changes in operating condition distribution or the addition of data labeled with remaining lifetime as trigger conditions to initiate a rolling update process based on the latest data, dynamically calibrating edge weights, cross-modal attention weights, and parameter sets. Based on this, a consistency verification threshold is used to quantitatively compare the remaining lifetime outputs of historical samples before and after the update. When the output difference exceeds a preset threshold, a rollback mechanism is automatically triggered, restoring the system to the state before the update. This process forms a closed-loop control logic of "trigger-update-verification-decision," enabling the model to both promptly integrate new operating condition features to avoid prediction bias and verify update quality through comparison with historical sample outputs, thereby ensuring the stability of existing prediction models while adapting to new data.

[0149] As a specific implementation method, this application is implemented as follows: In the wind turbine gearbox monitoring scenario, when a sudden change in wind speed is detected causing the rotational speed distribution to deviate from the historical range, newly added vibration and temperature data with remaining lifetime annotations are automatically collected. An online gradient descent algorithm is used to continuously update the edge weights and cross-modal attention weights in the graph neural network. Subsequently, the stored historical sample library is called to calculate the difference in remaining lifetime prediction results for three consecutive months of historical data before and after the update. If the difference exceeds a 5% consistency check threshold, the system immediately reverts to the parameter snapshot state before the update, and the reason for the update failure is recorded. In this embodiment, the rolling update process is executed silently in the background, without affecting the lifetime prediction task of the current window, and the rollback operation achieves millisecond-level state recovery through the parameter version management module.

[0150] Through the above technical solutions, this application solves the problem of inaccurate prediction when operating conditions change suddenly or new data is introduced, ensures that the gear life prediction model maintains prediction accuracy and stability in dynamic operating environments, avoids maintenance decision errors caused by data drift, and ensures the continuous and reliable operation of prediction services through a safety boundary mechanism.

[0151] Optionally, the gearbox run-to-fault dataset provided by the University of Huddersfield can be used for verification, such as... Figure 4 As shown, the system includes an electrical measuring device 1, a DAS (Data Acquisition System) 2, a drive motor 3, a gearbox 1 4, a gearbox 2 5, a load motor 6, a computer 7, a control cabinet 8, and a touch screen 9. Gear 3 in gearbox 1 4 and gear 5 in gearbox 2 5 are connected via a spindle. They exhibit strong physical and data correlation during actual operation and can be used as the measured gear node and key acquisition node for focused modeling. To avoid the unsteady-state effects during the meshing and break-in period, the accelerated degradation section after 300 hours of operation was selected for analysis and verification.

[0152] Under a unified time reference and unified spatial coordinates, the original vibration data is decomposed into multi-scale frequency bands, and characteristic amplitudes that can characterize component degradation are extracted from the frequency band energy and spectral amplitude. Simultaneously with acoustic data, temperature data, rotational speed data, torque data and oil state data, a multi-modal characteristic sequence within the sliding time window is formed. When focusing on gear 3 of gearbox 4 and gear 5 of gearbox 2, they can be included in the measured gear node and adjacent acquisition node of the spatiotemporal coupling diagram according to the meshing pair and vibration transmission path.

[0153] As monitoring time increases, the probability density function of the remaining life prediction after multimodal fusion gradually converges, showing a narrower distribution, reduced uncertainty, and prediction results that are closer to the actual life. Simultaneously, with the increase in available monitoring data, the remaining life errors of all comparative methods show a decreasing trend, with the error reduction of this method being more significant. This method can more stably reflect the degradation state and remaining life of the tested gear (see...). Figure 5 (Including the multi-index degradation curves of gearbox 1-4 and gearbox 2-5) GB1 represents gearbox 1-4, and GB2 represents gearbox 2-5.

[0154] In the above embodiments, multimodal data is acquired and preprocessed under a unified benchmark. A spatiotemporal coupled graph is constructed by fusing physical coupling edges and data coupling edges within a sliding time window. Then, cross-modal attention weight allocation is driven by time-related information extraction and graph feature propagation to obtain an interpretable degradation state representation. A parameter set for inference, simultaneously constraining remaining lifetime prediction error, degradation state fitting error, energy budget consistency, and edge weight sparsity, is used for training and online application. Finally, maintenance recommendations are generated in conjunction with maintenance thresholds. This forms a closed loop from data acquisition, relationship characterization, information fusion to decision output. This application fuses the edge weights of physical coupling edges and data coupling edges, enabling structural connections, contact relationships, vibration transmission paths, statistical correlations, and causal relationships to be characterized simultaneously in the same process, improving stability under operating condition disturbances and load fluctuations. The cross-modal attention weights adaptively adjust the influence of different modes based on propagation results, suppressing the interference of low-stability features on lifetime assessment and improving the discriminative power of the degradation state representation. Constraints such as energy budget consistency and edge weight sparsity are applied. The parameter set used for inference follows the energy transfer direction and eliminates low-contribution connections, reducing overfitting and improving transferability. This achieves highly stable and reliable prediction outputs for the remaining life of the tested gear.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A gear life prediction method based on a multi-modal spatio-temporal coupled graph neural network, characterized in that, The method comprises the following steps: Collecting multi-modal data for preprocessing and extracting multi-modal feature sequences according to sliding time windows, wherein the multi-modal data comprises vibration data, acoustic data, temperature data, rotation speed data, torque data and oil state data; In each sliding time window, taking the collection nodes and the measured gear nodes as a node set, determining physical coupling edges according to structural connections, contact relationships and vibration transmission paths, determining data coupling edges according to statistical correlation relationships and causal relationships of historical features, and fusing edge weights of the physical coupling edges and the data coupling edges to construct a space-time coupling graph; Extracting time correlation information from the multi-modal feature sequences and performing feature propagation on the space-time coupling graph, distributing cross-modal attention weights according to the propagation results, and obtaining a degradation state representation of the measured gear; Based on the data with remaining life annotations, a parameter set for reasoning is obtained, and the parameter set is obtained while constraining the remaining life prediction error, the degradation state fitting error, the energy balance consistency and the edge weight sparsity; During operation, input the newly formed multi-modal feature sequence and the space-time coupling graph, and output the remaining life of the measured gear. When the remaining life is less than the maintenance threshold, a maintenance suggestion is generated; Wherein, when the parameter set for reasoning is obtained based on the data with remaining life annotations, the method comprises the following steps: Divide the historical samples into training windows and validation windows according to the sliding time windows; Determine the initial value of the parameters according to the degradation state representation and the remaining life annotations in the training window; Iteratively obtain the parameter set in batch update mode, and simultaneously evaluate the remaining life prediction error, the degradation state fitting error, the energy balance consistency and the edge weight sparsity after each round of update. The energy balance consistency is determined by comparing whether the common trend of the changes in rotation speed, torque, temperature and oil state is consistent with the energy transmission direction of the measured gear. The edge weight sparsity is constrained by removing low contribution edges according to the edge weight from small to large and limiting the number of retained edges of each node within a predetermined range. Stop updating when the four evaluation indicators in the validation window meet the preset threshold value for a certain number of rounds, and obtain the parameter set for reasoning, which includes attention weights, edge weights, remaining life output related parameters and degradation state representation related parameters.

2. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 1, characterized in that, When collecting multi-modal data for preprocessing and extracting multi-modal feature sequences according to sliding time windows, the method comprises the following steps: Synchronize the timestamps of each channel with a unified time reference, record the relative positions of the collection nodes and the measured gear with a unified spatial coordinate, and perform spatial alignment; Estimate background noise, denoise, zero drift correction, outlier suppression, missing data completion and unit unification on the multi-modal data in a predetermined order, wherein the outlier suppression is determined according to the amplitude range and the change rate, and the missing data completion is constrained according to the consistency of adjacent time periods and related modalities; According to the sliding time window, the window start is aligned with the uniform time reference; the time domain features, the frequency domain features, the envelope related features and the working condition features of the multi-modal data are extracted in each sliding time window, and low stability features are removed according to feature stability and feature correlation to obtain the multi-modal feature sequence.

3. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 1, characterized in that, When the physical coupling edge is determined according to the structural connection, contact relationship and vibration transmission path, it includes: According to the gear engagement pair, coaxial connection, bearing support and shell constraint to establish a connection relationship, according to the contact position and the vibration transmission path to determine the direction of the edge, and according to the contact type, the assembly pre-tightening state, the support stiffness and the vibration transmission path level to grade the edge weight; In the continuous multiple sliding time windows, the edge weight stability of the obtained edge weight is tested, and only when the edge weight fluctuates within a preset fluctuation range, the physical coupling edge is retained.

4. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 3, characterized in that, When the data coupling edge is determined according to the statistical correlation and causality of historical features, and the edge weights of the physical coupling edge and the data coupling edge are fused, it includes: According to the multi-modal feature sequence, whether the linkage direction and linkage strength of the same feature pair remain within a preset fluctuation range in a continuous sliding time window is determined to obtain a correlation stability; whether the time sequence of the first changed feature and the second response feature remains within a preset range in a continuous sliding time window is compared to obtain a lag response consistency; according to the event time point triggered by the multi-modal feature and combined with the start-stop, alarm and load change in the operation log and maintenance record, the consistency of the event common occurrence and sequence is counted in a continuous sliding time window to obtain an event linkage consistency; The data coupling edge is determined according to the correlation stability, lag response consistency and event linkage consistency, and the edge weight of the data coupling edge is given; The edge weights of the physical coupling edge and the data coupling edge are dimensionally unified and range normalized, weighted and fused in the order of physical priority, and when the conclusions of the two types of edges conflict, the physical coupling edge is retained and adjusted according to the data coupling edge; The fused edge weight is sparsified, and the edges that do not meet the edge weight stability and window consistency check requirements are removed in the order of small to large weight values, and the retained edges are reviewed according to the window consistency check results before entering the next sliding time window, and finally the spatio-temporal coupling graph is obtained.

5. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 4, characterized in that, When the multi-modal feature sequence is time-related information extraction and feature propagation on the spatio-temporal coupling graph, it includes: In each sliding time window, time-related information is extracted for each modal feature, including trend change, periodic change and mutation event; After the time-related information extraction is completed, the fused edge weight in the spatio-temporal coupling graph is used as the transmission basis, and one to several rounds of feature propagation are performed between adjacent nodes in the order of large to small edge weight; when the time-related conflict caused by the propagation exceeds the preset conflict threshold, the contribution of the edge is inhibited; after completing all propagation rounds, the node time-related aggregation result is obtained.

6. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 5, characterized in that, According to the propagation result, cross-modal attention weight distribution is performed on different modalities, and a degradation state representation of the measured gear is obtained, including: According to the node time correlation aggregation result, the correlation stability, the lag response consistency and the event linkage consistency, and the noise level of each modality in the historical sample, a non-negative attention weight is set, and the attention weight is normalized in the same node; When a certain modality has insufficient correlation stability or conflicts with the working condition indication within the continuous sliding time window, its attention weight is reduced, and when multiple modalities collectively indicate that the degradation is aggravated, the corresponding attention weight is increased; after the attention weight distribution is completed, the results of each modality are weighted and converged in the node to obtain a node-level degradation state result, and the meshing pairs and the transmission path level are weighted and converged between nodes to obtain a degradation state representation of the measured gear.

7. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 6, characterized in that, When the data with remaining life annotations is used to obtain the parameter set for reasoning, a staged weight strategy is adopted for obtaining, including: In the first stage, only the remaining life prediction error and the degradation state fitting error are used as optimization targets, and window consistency verification is used as the stopping condition; In the second stage, after meeting the conditions of the first stage, the constraint weight of energy balance consistency and edge weight sparsity is increased, and after each round of update, edge weight stability test and window consistency verification are performed, if the performance of the verification window decreases or the consistency does not meet the preset consistency threshold, the parameter set of the last round is rolled back and the growth rate of the constraint weight is reduced; After completing the second stage, the parameter set for reasoning is finally shaped, and the threshold value, window number and rollback condition for retraining are recorded.

8. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 7, characterized in that, In the running process, the newly formed multi-modal feature sequence and the space-time coupled graph are input, and the remaining life of the measured gear is output, when the remaining life is less than the maintenance threshold, a maintenance suggestion is generated, including: In each sliding time window, the newly formed multi-modal feature sequence is aligned and reconstructed according to the unified time reference and the unified spatial coordinates, and the aligned multi-modal feature sequence and the reconstructed space-time coupled graph are input into the degradation state representation and the remaining life obtained by the parameter set; Window consistency verification is performed on a continuous preset number of sliding time windows, and only when the remaining life in the continuous window is not higher than the maintenance threshold and the degradation state representation does not show an improvement trend, a maintenance suggestion is generated; When the remaining life is in the buffer range of the maintenance threshold and the window consistency verification is passed, a preparation maintenance suggestion is generated; When the remaining life continuously decreases and is lower than the maintenance threshold, accompanied by continuous deterioration of the degradation state representation, a shutdown suggestion is generated; After any suggestion is generated, enter the alarm mute period, and only when the new window consistency verification result is inconsistent with the last conclusion, update the suggestion.

9. The gear life prediction method based on the multi-modal spatio-temporal coupled graph neural network according to claim 8, characterized in that, Also includes: When the working condition distribution changes or new data with remaining life annotations appears, the edge weight, cross-modal attention weight and the parameter set are updated based on the latest data, and the remaining life output of the historical sample before and after the update is compared by consistency verification threshold, and if it fails, it is rolled back to the state before the update.

Images