A method and system for processing gear scuffing fault data through multi-sensor fusion

Abstract

This invention discloses a multi-sensor fusion method and system for processing gear scuffing fault data. The system includes: a multi-sensor synchronous acquisition module for synchronously acquiring signals from multiple sources; a chaotic resonance enhancement module for enhancing the signal-to-noise ratio of weak fault signals using a nonlinear bistable physical model; a decoupling representation learning module for orthogonally decoupling fault features from operating condition information through adversarial training; a meta-learning adaptive module for rapidly adapting to new operating conditions with small samples through model-independent meta-learning; a physical semantic field mapping module for constructing a three-dimensional multi-physical domain semantic field tensor; a dynamic causal graph reasoning module for learning the causal path graph of fault propagation; a multi-modal fusion decision module for fusing diagnostic results with causal path information; and an output module for visual output. This invention solves the problems of high noise interference, difficulty in separating feature coupling, and high cost of adapting to new operating conditions, improving diagnostic accuracy and model versatility.

Description

Technical Field

[0001] This invention relates to the field of gearbox fault diagnosis and data processing technology, specifically to a multi-sensor fusion method and system for processing gear scuffing fault data. Background Technology

[0002] As a core transmission component, gears are prone to tooth surface scuffing failure under high-speed, heavy-load, and poorly lubricated conditions, which can easily lead to equipment downtime and safety accidents. Traditional inspection methods rely on periodic disassembly and inspection, which cannot provide real-time dynamic monitoring and suffer from problems such as fragmented data, low efficiency in fault feature extraction and diagnosis, making it difficult to meet the precise early warning requirements under complex operating conditions. With the upgrading of intelligent operation and maintenance of high-end equipment, there is an urgent need to build a dedicated data processing system to achieve multi-source signal integration, feature mining, and intelligent diagnosis, ensuring the safe and stable operation of gear transmission systems.

[0003] In existing technologies, gear fault diagnosis often employs single signal processing and conventional machine learning schemes, failing to integrate chaotic resonance enhancement, orthogonal decoupling, and meta-learning. The lack of chaotic resonance enhancement makes weak fault signals easily submerged by strong noise, resulting in a low signal-to-noise ratio. The absence of orthogonal decoupling leads to deep coupling between operating condition fluctuations and fault features in multi-source signals, making accurate separation and extraction difficult. Furthermore, the lack of meta-learning means the model relies on massive amounts of labeled samples for training, resulting in poor generalization ability under varying operating conditions and new scenarios, necessitating retraining. The failure to form a collaborative technical loop among these three elements directly leads to significant noise interference from multi-source signals, difficulty in decoupling feature coupling, and high costs of adapting to new operating conditions. Ultimately, this results in low diagnostic accuracy and poor model versatility, failing to meet the demand for efficient and accurate diagnosis of gear scuffing faults under complex operating conditions. Summary of the Invention

[0004] In view of the above problems, the present invention provides a multi-sensor fusion method and system for processing gear scuffing fault data, which solves the problems of low diagnostic accuracy and poor versatility in existing gear scuffing fault diagnosis technology, such as large noise interference from multiple sources, deep coupling between fault features and working condition information, poor model generalization ability under new working conditions requiring a lot of retraining.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a multi-sensor fusion-based gear scuffing fault data processing system, comprising: The multi-sensor synchronous acquisition module is used to synchronously acquire multi-source signals, including gearbox vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals. The chaotic resonance enhancement module is connected to the multi-sensor synchronous acquisition module. It is used to input the acquired multi-source signals into a nonlinear bistable physical model. By adjusting the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model, the physical model is made to enter a random resonance state, thereby outputting a resonance enhancement signal with enhanced signal-to-noise ratio. The decoupled representation learning module, connected to the chaotic resonance enhancement module, is used to extract a first feature vector and a second feature vector from the resonance enhancement signal, and to make the first feature vector orthogonal to the operating condition information through adversarial training, thereby generating a decoupled fault feature vector; wherein, the first feature vector is used to characterize the scuffing fault state of the gear, and the second feature vector is used to characterize the operating condition of the gearbox. The meta-learning adaptive module, connected to the decoupled representation learning module, is used to perform a model-independent meta-learning process, specifically including: constructing multiple meta-tasks based on historical fault data for meta-training to learn a set of model initialization parameters that are universal across working conditions; and when facing a new working condition, using the universal model initialization parameters and a preset number K of labeled samples collected under the new working condition to perform a finite number L gradient updates on the working condition adaptive diagnostic model, and output a working condition adaptive diagnostic model adapted to the new working condition. The physical semantic field mapping module, connected to the chaotic resonance enhancement module, is used to map the sensor readings corresponding to each sampling time in the resonance enhancement signal to the three-dimensional spatial grid of the gearbox according to the sensor installation position. It also converts the vibration signal, temperature field distribution signal, and acoustic emission signal into physical quantities characterizing the dynamic state, thermodynamic state, and tribological state, respectively, to obtain multi-physical domain state information and fill it into the corresponding position of the three-dimensional spatial grid to construct a three-dimensional physical semantic field tensor containing spatial position information and multi-physical domain state information. The dynamic causal graph reasoning module, connected to the physical semantic field mapping module, is used to construct a dynamic Bayesian causal graph based on the three-dimensional physical semantic field tensor. The graph uses the physical state changes at different positions and times in the three-dimensional physical semantic field tensor as nodes and the driving relationship between nodes that conforms to physical laws as directed edges. Through graph structure learning and counterfactual reasoning based on attention mechanism, the module outputs a causal path graph representing the fault propagation path. The multimodal fusion decision module, which connects the meta-learning adaptive module and the dynamic causal graph reasoning module, is used to perform weighted fusion of the diagnostic results, which include fault type and fault severity, output by the working condition adaptive diagnostic model, with the fault causal propagation path information represented by the causal path graph, to generate the final fusion diagnostic result. The output module, connected to the multimodal fusion decision module, is used to visualize and output the fusion diagnostic results.

[0006] Furthermore, the chaotic resonance enhancement module includes: The signal input unit is used to receive multi-source signals output by the multi-sensor synchronous acquisition module, and to weight and superimpose the vibration signal, acoustic emission signal, temperature field distribution signal and operating condition parameter signal according to a preset weighting coefficient to form a fused input signal; The parameter adjustment unit is used to dynamically adjust the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model. The nonlinear system evolution unit is used to execute the state evolution equations of the nonlinear bistable physical model, and the state evolution equations are in the following form: ; in, Let be the potential function of the nonlinear bistable physical model. Let a and b represent the state variables at time t, and a and b be the barrier height parameters. This represents the signal collected by the i-th type of sensor at time t. This represents the corresponding signal weight, where N is the total number of sensor categories. The background noise at time t is represented by the noise intensity parameter D. The resonant output unit is used to output the resonant enhancement signal when the signal-to-noise ratio gain of the output signal reaches a preset gain threshold.

[0007] Furthermore, the parameter adjustment unit employs a particle swarm optimization algorithm to dynamically adjust the barrier height parameters a and b, as well as the noise intensity parameter D, with the goal of maximizing the signal-to-noise ratio gain of the output signal.

[0008] Among them, the adjustment range of barrier height parameter a is 0.5-2.0, the adjustment range of barrier height parameter b is 0.1-0.8, the adjustment range of noise intensity parameter D is 0.01-0.1, and the adjustment step size is 0.01.

[0009] Furthermore, the parameter adjustment unit includes: The oil film state identification subunit is used to analyze the acoustic emission signal and temperature field distribution signal in the multi-source signal in real time, extract the high-frequency components of the acoustic emission signal and the gradient change of the temperature field distribution signal, and identify the lubricating oil film state of the current gear meshing area online. The lubricating oil film state includes elastohydrodynamic lubrication state, mixed lubrication state, boundary lubrication state and oil film rupture state. The parameter mapping subunit is connected to the oil film state identification subunit. It is used to store the pre-calibrated oil film state-random resonance parameter mapping table, and according to the current lubricating oil film state output by the oil film state identification subunit, it queries and outputs the adjustment range and initial value of the corresponding barrier height parameters a, b and noise intensity parameter D from the oil film state-random resonance parameter mapping table. A dynamic adjustment execution subunit, connected to the parameter mapping subunit, is used to dynamically adjust the barrier height parameters a and b and the noise intensity parameter D within the adjustment range based on the adjustment range and initial value output by the parameter mapping subunit, using a particle swarm optimization algorithm with the goal of maximizing the signal-to-noise ratio gain of the output signal. The oil film state-random resonance parameter mapping table is pre-calibrated through multi-condition calibration experiments on a gearbox test bench. When the oil film state identification subunit identifies that the current lubricating oil film state has switched from elastohydrodynamic lubrication state to boundary lubrication state or oil film rupture state, the lower limit of the adjustment range of the barrier height parameter a output by the parameter mapping subunit is reduced to 50%-70% of the original lower limit of the adjustment range. This enhances the resonance amplification capability of the chaotic resonance enhancement module for the weak impact components in the acoustic emission signal that characterize the initiation stage of bonding.

[0010] Furthermore, the decoupled representation learning module includes: A fault feature extractor is used to extract features from the resonance enhancement signal; A condition discriminator is used to predict the condition category based on the features output by the fault feature extractor. And a gradient inversion layer connecting the fault feature extractor and the operating condition discriminator; During the training phase, the gradient reversal layer reverses the gradient flowing from the condition discriminator to the fault feature extractor with a fixed negative coefficient during backpropagation. The negative coefficient ranges from -0.2 to -0.05. By inverting the gradient through the gradient inversion layer, the fault feature extractor is trained to generate the decoupled fault feature vector that is orthogonal to the operating condition information.

[0011] Furthermore, the meta-learning adaptive module includes: Meta-task construction unit, used to divide historical fault data into multiple meta-tasks according to operating condition categories. Each meta-task It contains a support set and a query set; The inner loop optimization unit, connected to the meta-task construction unit, is used in each meta-task Internally, based on its support set, the current model parameters Perform inner gradient updates to obtain task-specific parameters. The updated formula is as follows: ; in, The inner learning rate, For meta-task loss function on, Represents the loss function Regarding model parameters The gradient vector; The outer loop meta-update unit, connected to the inner loop optimization unit, is used to update the current model parameters based on the query set loss of all meta-tasks. Perform a meta-update to learn a general model initialization parameter. Its optimization objective is: ; The rapid adaptation unit, connected to the outer loop element update unit, is used to initialize the parameters based on the general model when fault data of a new operating condition is received. Fine-tuning is performed where the number of gradient update steps required for fine-tuning is L≤10, and the number of samples used in each step is K≤5.

[0012] Furthermore, it also includes a physical constraint regularization module, connected to the meta-learning adaptive module, wherein the physical constraint regularization module includes: The theoretical calculation unit is used to calculate the theoretical temperature rise rate Y during the development of adhesive failure based on the input current operating parameters. The formula for calculating Y is as follows: Y = ,in, The coefficient of friction, For rotational speed, For torque, The thermal conversion coefficient; The deviation detection unit is used to calculate the temperature rise rate predicted by the meta-learning adaptive module. The deviation between the theoretical temperature rise rate output by the theoretical calculation unit and the theoretical temperature rise rate. The regularization penalty unit is used to construct a physical consistency regularization term based on the deviation output by the deviation detection unit. This is then incorporated into the training loss function of the meta-learning adaptive module; wherein the physical consistency regularization term is expressed as follows: ; Where λ is the regularization weight coefficient. Indicates the calculation of L2 norm; The regularization penalty unit is configured to increase the regularization weight coefficient λ to strengthen the physical constraint when the deviation between the predicted temperature rise rate and the theoretical temperature rise rate increases.

[0013] Furthermore, the dynamic causal graph reasoning module includes: The causal structure learning unit is used to learn the graph structure of the dynamic Bayesian causal graph by using a neural causal model based on the attention mechanism. The graph structure consists of nodes with physical state changes at different positions and times in the three-dimensional physical semantic field tensor and directed edges with driving relationships between nodes that conform to physical laws. The counterfactual reasoning unit, connected to the causal structure learning unit, is used to generate counterfactual samples. By comparing the differences between the real observation results and the counterfactual inference results, the confidence of the causal edges in the dynamic Bayesian causal graph is strengthened, so as to output the causal path graph representing the fault propagation path. The number of nodes in the dynamic Bayesian causal graph is set to M, which corresponds to at least one spatiotemporal event among abnormal gear meshing, bearing wear, excessive temperature, and excessive vibration. The time window of the node is set to T, and the initial confidence of the directed edge is set to C0. The number of learning iterations for the causal structure learning unit is set to I, and the learning rate for the causal structure learning unit is set to β. The counterfactual inference unit generates P counterfactual samples each time. The counterfactual sample generation uses a Monte Carlo simulation method. When the deviation between the actual observation and the counterfactual inference is less than a first deviation threshold δ1, the confidence of the corresponding causal edge increases by a first increment Δ1. When the deviation between the actual observation and the counterfactual inference is greater than a second deviation threshold δ2, the confidence decreases by a second decrement Δ2. Here, δ1 < δ2, Δ1 and Δ2 are both positive numbers, and Δ1 < Δ2. Ultimately, the confidence range of the causal edge is controlled within the lower limit C. min Up to the upper limit C max Between, of which, C min <C max And C min and C max All are positive numbers between 0 and 1.

[0014] Furthermore, the output module includes: The 3D model rendering unit is used to load the 3D model of the gearbox and perform visualization rendering based on the 3D model of the gearbox. A causal path drawing unit, connected to the three-dimensional model rendering unit, is used to draw the fault causal propagation path represented by the causal path diagram on the three-dimensional model of the gearbox in the form of directed edges with arrows. The confidence level identification unit is connected to the causal path drawing unit and is used to identify causal edges with different confidence levels using differentiated colors based on the confidence level of each causal edge in the causal path graph output by the dynamic causal graph inference module. Wherein, the confidence level is the causal edge confidence level enhanced by the dynamic causal graph reasoning module through counterfactual reasoning; Causal edges with a confidence level higher than a preset first confidence level threshold are classified as high-confidence causal edges, and causal edges with a confidence level lower than a preset second confidence level threshold are classified as low-confidence causal edges, wherein the preset first confidence level threshold is greater than the preset second confidence level threshold.

[0015] In a second aspect, the present invention provides a multi-sensor fusion method for processing gear scuffing fault data, the method being applicable to the system described in the first aspect of the present invention, the method comprising the following steps: S1: Synchronously acquire multi-source signals, including gearbox vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals; S2: Input the collected multi-source signals into a nonlinear bistable physical model. By adjusting the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model, the physical model is made to enter a stochastic resonance state, thereby outputting a resonance enhancement signal with enhanced signal-to-noise ratio. S3: Extract the first feature vector and the second feature vector from the resonance enhancement signal, and make the first feature vector orthogonal to the working condition information through adversarial training to generate a decoupled fault feature vector; wherein, the first feature vector is used to characterize the scuffing fault state of the gear, and the second feature vector is used to characterize the operating condition of the gearbox. S4: Perform a model-independent meta-learning process, specifically including: constructing multiple meta-tasks based on historical fault data for meta-training to learn a set of model initialization parameters that are universal across working conditions; when facing a new working condition, based on the universal model initialization parameters, using a preset number K of labeled samples collected under the new working condition to perform a finite number L gradient updates on the working condition adaptive diagnostic model, and output a working condition adaptive diagnostic model adapted to the new working condition. S5: The sensor readings corresponding to each sampling time in the resonance enhancement signal are mapped to the three-dimensional spatial grid of the gearbox according to the sensor installation position, and the vibration signal, temperature field distribution signal, and acoustic emission signal are converted into physical quantities that characterize the dynamic state, thermodynamic state, and tribological state, respectively, to obtain multi-physical domain state information and fill it into the corresponding position of the three-dimensional spatial grid to construct a three-dimensional physical semantic field tensor containing spatial position information and multi-physical domain state information; S6: Based on the three-dimensional physical semantic field tensor, construct a dynamic Bayesian causal graph with the physical state changes at different positions and times in the three-dimensional physical semantic field tensor as nodes and the driving relationship between nodes that conforms to physical laws as directed edges. Then, through graph structure learning and counterfactual reasoning based on attention mechanism, output a causal path graph representing the fault propagation path. S7: The diagnostic results output by the adaptive diagnostic model of the working condition, which include the fault type and fault severity, are weighted and fused with the fault causal propagation path information represented by the causal path diagram to generate the final fused diagnostic result. S8: Visualize the fusion diagnostic results.

[0016] Unlike existing technologies, the above solution discloses a multi-sensor fusion method and system for processing gear scuffing fault data. The system includes: a multi-sensor synchronous acquisition module for synchronously acquiring multi-source signals such as vibration, acoustic emission, temperature field, and operating parameters; a chaotic resonance enhancement module that utilizes the stochastic resonance effect of a nonlinear bistable physical model to enhance the signal-to-noise ratio of weak fault signals; a decoupled representation learning module that orthogonally decouples fault features from operating condition information through adversarial training to generate pure fault feature vectors; a meta-learning adaptive module that obtains model initialization parameters applicable across operating conditions through model-independent meta-learning, enabling rapid adaptation to new operating conditions with small samples; a physical semantic field mapping module that maps sensor readings to a three-dimensional spatial grid to construct a multi-physical domain semantic field tensor; a dynamic causal graph reasoning module that learns the causal path graph of fault propagation based on this tensor; a multi-modal fusion decision module that weightedly fuses diagnostic results with causal path information; and an output module that visualizes the fused diagnostic results. This invention solves the technical problems of large noise interference, difficulty in separating feature coupling, and high cost of adapting to new working conditions by synergistically integrating chaotic resonance enhancement, orthogonal decoupling, and meta-learning, thereby improving diagnostic accuracy and model versatility.

[0017] The above description of the invention is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical solution of the present invention and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of the present invention easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of the present invention. Attached Figure Description

[0018] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on the present invention.

[0019] In the accompanying drawings of the instruction manual: Figure 1 This is a schematic diagram of a multi-sensor fusion gear scuffing fault data processing system according to an exemplary embodiment of the present invention; Figure 2 This is a flowchart of a multi-sensor fusion method for processing gear scuffing fault data according to an exemplary embodiment of the present invention; The reference numerals used in the above figures are explained as follows: 101. Power supply module; 102. Multi-sensor synchronous acquisition module; 103. Chaotic resonance enhancement module; 104. Physical semantic field mapping module; 105. Decoupled representation learning module; 106. Physical constraint regularization module; 107. Meta-learning adaptive module; 108. Dynamic causal graph reasoning module; 109. Multimodal fusion decision module; 110. Output module. Detailed Implementation

[0020] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that can be achieved by this invention, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this invention, and are therefore only examples, and should not be used to limit the scope of protection of this invention.

[0021] In the first aspect, such as Figure 1 As shown, the present invention provides a multi-sensor fusion-based gear scuffing fault data processing system, comprising: Power supply module 101 is used to provide power to other modules in the multi-sensor fusion gear seizure fault data processing system; The multi-sensor synchronous acquisition module 102 is used to synchronously acquire multi-source signals, including gearbox vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals. The chaotic resonance enhancement module 103 is connected to the multi-sensor synchronous acquisition module and is used to input the acquired multi-source signals into a nonlinear bistable physical model. By adjusting the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model, the physical model is made to enter a random resonance state, thereby outputting a resonance enhancement signal with enhanced signal-to-noise ratio. The decoupled representation learning module 105 is connected to the chaotic resonance enhancement module. It is used to extract a first feature vector and a second feature vector from the resonance enhancement signal, and to make the first feature vector orthogonal to the working condition information through adversarial training to generate a decoupled fault feature vector. The first feature vector is used to characterize the scuffing fault state of the gear, and the second feature vector is used to characterize the operating condition of the gearbox. The meta-learning adaptive module 107, connected to the decoupled representation learning module, is used to perform a model-independent meta-learning process, specifically including: constructing multiple meta-tasks based on historical fault data for meta-training to learn a set of model initialization parameters that are universal across working conditions; and when facing a new working condition, using the universal model initialization parameters and a preset number K of labeled samples collected under the new working condition to perform a finite number L gradient updates on the working condition adaptive diagnostic model, and outputting a working condition adaptive diagnostic model adapted to the new working condition. The physical semantic field mapping module 104, connected to the chaotic resonance enhancement module, is used to map the sensor readings corresponding to each sampling time in the resonance enhancement signal to the three-dimensional spatial grid of the gearbox according to the sensor installation position, and convert the vibration signal, temperature field distribution signal, and acoustic emission signal into physical quantities characterizing the dynamic state, thermodynamic state, and tribological state respectively, to obtain multi-physical domain state information and fill it into the corresponding position of the three-dimensional spatial grid to construct a three-dimensional physical semantic field tensor containing spatial position information and multi-physical domain state information; The dynamic causal graph reasoning module 108 is connected to the physical semantic field mapping module. It is used to construct a dynamic Bayesian causal graph based on the three-dimensional physical semantic field tensor. The graph uses the physical state changes at different positions and times in the three-dimensional physical semantic field tensor as nodes and the driving relationship between nodes that conforms to physical laws as directed edges. It outputs a causal path graph representing the fault propagation path through graph structure learning and counterfactual reasoning based on attention mechanism. The multimodal fusion decision module 109 connects the meta-learning adaptive module and the dynamic causal graph reasoning module. It is used to perform weighted fusion of the diagnostic results, which include fault type and fault severity, output by the working condition adaptive diagnostic model and the fault causal propagation path information represented by the causal path graph, to generate the final fusion diagnostic result. Output module 110 is connected to the multimodal fusion decision module and is used to visualize and output the fusion diagnostic results.

[0022] In this embodiment, the multi-sensor synchronous acquisition module uses a unified hardware clock as a reference to simultaneously drive multiple different types of sensors to sample signals, ensuring that signals such as vibration, acoustic emission, temperature field, rotational speed, and torque are strictly aligned on the time axis, with synchronization error controlled at the microsecond level, thus eliminating feature distortion caused by timing offset.

[0023] The chaotic resonance enhancement module is based on the principle of nonlinear bistable random resonance. By adjusting the system potential barrier and noise intensity, the system enters the optimal random resonance state, amplifies the energy of weak fault features that are submerged by strong noise, significantly improves the signal-to-noise ratio, and realizes the effective extraction of weak fault signals.

[0024] Decoupled representation learning module: Adversarial learning mechanism is adopted to orthogonally separate fault-related features and operating condition-related features in a high-dimensional feature space, so that the fault features do not contain operating condition information such as speed, load, and torque, and a pure fault representation is obtained.

[0025] The meta-learning adaptive module adopts the Model Independent Meta-Learning (MAML) framework. Through inner and outer double-layer loop optimization, it learns a set of universal model initialization parameters applicable to multiple working conditions. This enables the diagnostic model to quickly adapt to new working conditions with only a very small number of labeled samples and a limited number of gradient updates, achieving small-sample cross-working-condition generalization.

[0026] The physical semantic field mapping module maps the sensor-acquired signals to spatial coordinates according to the actual three-dimensional structure of the gearbox, and converts the signals into physical quantities in the three physical domains of dynamics, thermodynamics, and tribology. It constructs a three-dimensional tensor with spatial location, time sequence, and physical meaning, thereby realizing the physical semantic expression of the data.

[0027] The dynamic causal graph reasoning module is based on the physical semantic field tensor to construct a dynamic Bayesian causal network. It uses spatiotemporal physical events as nodes and physical driving relationships as directed edges. It learns the causal structure through an attention mechanism and verifies and strengthens the causal confidence through counterfactual reasoning, outputting the causal path of fault occurrence and propagation.

[0028] The multimodal fusion decision module adaptively weights and fuses data-driven diagnostic results (fault type, severity) with physical causal reasoning results (fault location, propagation path, confidence level), balancing diagnostic accuracy and physical interpretability.

[0029] The output module uses the 3D CAD model of the gearbox as a carrier to visualize and render the diagnostic results, fault location, fault level, causal propagation path, and confidence level, thereby achieving intuitive and efficient human-computer interaction.

[0030] In this embodiment, the power supply module uses an AC-DC switching power supply with an input voltage of 220V AC and an output voltage of 12V DC, with a rated output power of not less than 50W. It also integrates a backup lithium battery module, which can maintain normal operation of all modules for at least 30 minutes when the external power supply is interrupted. In the multi-sensor synchronous acquisition module, vibration signals are acquired using a piezoelectric accelerometer with a sampling frequency of 10kHz; acoustic emission signals are acquired using a piezoelectric acoustic emission sensor with a sampling frequency of 50kHz; temperature field distribution signals are acquired using an infrared thermal imaging sensor with a temperature measurement range of -20℃ to 150℃; and operating parameter signals are transmitted via the gearbox's built-in speed sensor. Data is collected by sensors such as torque sensors, with synchronization errors controlled within ±1ms. Based on a model-independent meta-learning framework, historical fault data is divided into multiple meta-tasks containing support sets and query sets according to operating condition categories. Within each meta-task, the inner gradient of the current model parameters is updated based on the support set, and the outer parameter update is completed through the query set loss of all meta-tasks. This allows the learning of universal initialization parameters across operating conditions. When a new operating condition is introduced, relying on these universal model initialization parameters, only a small number of new operating condition samples are used for a limited number of gradient updates, such as using the number of samples K for L gradient updates, where K≤10 and the number of updates L≤5, to quickly complete model adaptation and finally output an adaptive diagnostic model for operating conditions.

[0031] The working principle of the multi-sensor fusion gear scuffing fault data processing system involved in this embodiment is as follows: After the system powers on, the power supply module provides stable and continuous power to all functional units. The multi-sensor synchronous acquisition module is triggered by a unified clock and acquires multiple signals in parallel: a piezoelectric accelerometer acquires vibration signals from the gearbox housing, a piezoelectric acoustic emission sensor acquires high-frequency acoustic emission signals from gear meshing, and an infrared thermal imaging sensor acquires temperature field distribution signals from the gearbox surface. Simultaneously, the gearbox's built-in sensors acquire operating parameter signals such as speed, torque, and load. All signals are sampled synchronously in the same clock domain, with time synchronization errors controlled within ±1ms, ensuring strict alignment of multi-source information in time and space.

[0032] The acquired multi-source raw signals are sent to the chaotic resonance enhancement module. The chaotic resonance enhancement module weights and fuses the multiple signals according to preset weights to form a comprehensive input signal; then, through adaptive parameter adjustment, it drives the nonlinear bistable system into the optimal random resonance state, so that the weak glue fault characteristics are effectively pumped and amplified by noise energy, the noise is suppressed, and finally the output resonance enhancement signal with a significantly improved signal-to-noise ratio gain is achieved.

[0033] The resonant enhancement signal is processed through two parallel links: The first link is data-driven diagnosis: the resonance enhancement signal enters the decoupled representation learning module, where high-dimensional features are learned through the fault feature extractor and the operating condition feature extractor, respectively. Then, adversarial training is implemented through a gradient inversion layer, forcing the fault features and operating condition features to be orthogonal, completely eliminating the influence of the operating condition, and outputting a pure fault feature vector unrelated to the operating condition. This pure fault feature vector is then input into the meta-learning adaptive module. The meta-learning adaptive module first constructs a meta-task on a multi-operating condition historical fault dataset, learning universal model initialization parameters across operating conditions through inner loop single-task updates and outer loop global optimization. When a new operating condition is introduced, based on these universal parameters, a high-precision operating condition adaptive diagnostic model can be quickly generated using only a small number of samples (≤10) and a very small number of gradient updates (≤5 times), outputting the fault type, fault location, and fault severity.

[0034] The second link is oriented towards physically interpretable reasoning: the resonance enhancement signal enters the physical semantic field mapping module, which maps the signal to a 1:1 three-dimensional spatial mesh of the gearbox according to the actual installation coordinates of the sensor; simultaneously, the vibration signal is mapped to physical quantities in the dynamic domain, the temperature field signal to physical quantities in the thermodynamic domain, and the acoustic emission signal to physical quantities in the tribological domain; partial differential equation constraints are used to interpolate and complete the mesh for missing points, ultimately generating a three-dimensional physical semantic field tensor containing spatial location, temporal information, and multi-physical domain information. This tensor is input into the dynamic causal graph reasoning module, which constructs a dynamic Bayesian causal graph with spatiotemporal physical state changes as nodes and physical driving relationships as edges; the causal structure is automatically learned through an attention mechanism, and the confidence of causal edges is verified through Monte Carlo counterfactual reasoning, outputting a causal path graph containing the fault origin, propagation path, and scope of influence.

[0035] The multimodal fusion decision module receives the diagnostic results output by the meta-learning model and the causal path information output by the dynamic causal graph. It dynamically adjusts the fusion weights based on operating parameters and causal confidence levels, weighting and fusing the two types of information to obtain a final fusion diagnostic result that combines high accuracy and high interpretability. Finally, the output module loads a 3D model of the gearbox in STL format, and uses color differentiation, arrow annotations, and numerical displays to perform 3D visualization rendering of the diagnostic conclusions, fault locations, causal propagation paths, and confidence levels, presenting them intuitively to maintenance personnel.

[0036] The above-mentioned scheme, through multi-source information complementarity, makes early fault detection easier. It simultaneously acquires multi-modal data including vibration, acoustic emission, temperature field, and operating parameters, covering the multi-dimensional manifestations of gear scuffing faults in structural vibration, high-frequency radiation, temperature changes, and operating status. Compared to single-signal methods, it has higher fault sensitivity and coverage, capturing weak anomalies in the early stages of scuffing. By utilizing a stochastic resonance mechanism to amplify the energy of weak fault features, the signal-to-noise ratio is significantly improved without increasing hardware costs, solving the problem of early fault signals being submerged in high-noise industrial environments. Through adversarial learning, fault features and operating condition features are completely separated, eliminating interference from speed, load, and torque fluctuations in diagnosis, ensuring the diagnostic model maintains high accuracy under varying operating conditions and loads.

[0037] Meta-learning enables small-sample adaptation, significantly reducing on-site deployment costs. By obtaining universal model initialization parameters through model-independent meta-learning, new working conditions can be quickly adapted with only a few samples, eliminating the need for large-scale data collection and long-term training, thus significantly shortening the deployment cycle and reducing data costs.

[0038] By combining physical semantics and causal reasoning, this approach overcomes the limitations of black-box diagnostics by constructing a three-dimensional physical semantic field and dynamic causal graph. This imbues diagnostic results with clear physical meaning and causal relationships, enabling the identification of fault sources and tracing of propagation paths. This provides crucial support for maintenance decisions, root cause remediation, and lifespan prediction. By integrating data-driven diagnostics with physical causal reasoning, it ensures both diagnostic accuracy and improved result reliability, avoiding misjudgments from single models and meeting the high reliability requirements of industrial environments.

[0039] In some embodiments, the chaotic resonance enhancement module includes: The signal input unit is used to receive multi-source signals output by the multi-sensor synchronous acquisition module, and to weight and superimpose the vibration signal, acoustic emission signal, temperature field distribution signal and operating condition parameter signal according to a preset weighting coefficient to form a fused input signal; The parameter adjustment unit is used to dynamically adjust the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model. The nonlinear system evolution unit is used to execute the state evolution equations of the nonlinear bistable physical model, and the state evolution equations are in the following form: ; in, Let be the potential function of the nonlinear bistable physical model. Let a and b represent the state variables at time t, and a and b be the barrier height parameters. This represents the signal collected by the i-th type of sensor at time t. This represents the corresponding signal weight, where N is the total number of sensor categories. The background noise at time t is represented by the noise intensity parameter D. The resonant output unit is used to output the resonant enhancement signal when the signal-to-noise ratio gain of the output signal reaches a preset gain threshold.

[0040] In this embodiment, the signal input unit is responsible for receiving multiple raw sensor signals, normalizing, aligning and weighting them according to preset rules, and fusing the multi-source heterogeneous signals into a unified input signal, providing a standard input for subsequent nonlinear system processing.

[0041] The parameter adjustment unit adjusts the key parameters of the nonlinear bistable system in real time and adaptively based on the characteristics of the input signal, changes in operating conditions, and oil film status, so that the system always maintains the optimal stochastic resonance state.

[0042] The nonlinear system evolution unit is used to execute the bistable stochastic resonance state equation, complete the interaction between signal, noise and system potential field, and realize the resonance enhancement of weak fault characteristics.

[0043] The resonant output unit monitors the signal-to-noise ratio gain of the output signal in real time, determines whether the enhancement threshold has been reached, and outputs a stable and reliable resonant enhancement signal.

[0044] The nonlinear bistable physical model is a typical nonlinear dynamic model driven by a potential function and has two stable equilibrium states. Under the combined action of a signal and appropriate noise, it can produce a random resonance phenomenon, thereby amplifying weak signals.

[0045] The potential function is a function that describes the energy distribution of a bistable system and determines the shape of the potential well, the height of the potential barrier, and the steady state of the system.

[0046] Signal-to-noise ratio (SNR) gain is the ratio of the output signal SNR to the input signal SNR. It is used to quantify the signal enhancement effect. A gain greater than 1 indicates that the signal has been effectively enhanced.

[0047] D=12 This is the noise intensity parameter for background noise. The ensemble average notation is used to characterize the statistical averaging operation on random variables. This state evolution equation is the core state evolution model of a multi-source signal-driven nonlinear bistable stochastic resonance system, which dynamically adjusts the barrier height parameter. With noise intensity parameters This allows the system to enter the optimal random resonance state, enabling enhanced extraction of multi-source weak fault feature signals and providing high signal-to-noise ratio feature inputs for subsequent cross-condition fault diagnosis.

[0048] In this embodiment, the specific working principle of the chaotic resonance enhancement module is as follows: The signal input unit first receives multiple signals from the multi-sensor synchronous acquisition module, including vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals. The unit first performs mean reduction, normalization, and filtering preprocessing on each signal to eliminate dimensional differences and abnormal interference. Then, it weights and superimposes the signals according to preset weighting coefficients. The weights are allocated based on the sensor's sensitivity to bonding faults; for example, vibration signals have a weight of 0.4, acoustic emission signals 0.35, and temperature field signals 0.25, with a total weight of 1. This fuses the multi-source heterogeneous signals into a single fused input signal containing comprehensive fault information, providing a unified input for subsequent random resonance processing.

[0049] The fused input signal is fed into the nonlinear system evolution unit, which then executes the preset nonlinear bistable system state evolution equation: .

[0050] Under the combined excitation of signal and noise, the system enters a random resonance state: noise energy is coupled into the weak fault signal, significantly increasing the amplitude of the fault characteristic and suppressing the noise. The parameter adjustment unit monitors the system output in real time and dynamically adjusts the barrier height parameters a and b, and the noise intensity parameter D, to maintain the system at the optimal resonance point. The resonance output unit continuously calculates the signal-to-noise ratio (SNR) of the input and output signals. When the SNR gain reaches a preset gain threshold (e.g., 1.2 or higher), the signal enhancement is deemed effective, and a resonant enhanced signal is output. If the preset gain threshold is not reached, the parameter adjustment unit is triggered to re-optimize until the enhancement requirements are met.

[0051] The above scheme fully utilizes the advantages of different sensors by fusing signals such as vibration, acoustic emission, and temperature field according to sensitivity weights, avoiding the loss of information from a single signal and improving the completeness of fault characterization. It leverages the stochastic resonance effect of a nonlinear bistable system to amplify weak fault features in a strong noise background, making it particularly suitable for detecting weak signals in the early stages of bonding, overcoming the bottleneck of traditional filtering methods that "filter noise while weakening the fault signal." By adjusting the potential barrier height and noise intensity in real time, the system maintains optimal resonance under different operating conditions and noise levels, ensuring stable and reliable enhancement. By setting a clearly defined gain threshold for controllable output quality, it ensures a sufficiently high signal-to-noise ratio for the output signal, providing high-quality data input for subsequent feature decoupling, fault diagnosis, and physical modeling.

[0052] In some embodiments, the parameter adjustment unit employs a particle swarm optimization algorithm to dynamically adjust the barrier height parameters a and b and the noise intensity parameter D with the goal of maximizing the signal-to-noise ratio gain of the output signal.

[0053] Among them, the adjustment range of barrier height parameter a is 0.5-2.0, the adjustment range of barrier height parameter b is 0.1-0.8, the adjustment range of noise intensity parameter D is 0.01-0.1, and the adjustment step size is 0.01.

[0054] Particle Swarm Optimization (PSO) is a global optimization method based on swarm intelligence. It iteratively searches for the optimal solution in the solution space by simulating the foraging behavior of bird flocks. It does not require gradient calculation, has a fast convergence speed, and is robust, making it suitable for real-time parameter optimization.

[0055] The fitness function is an objective function used to evaluate the quality of parameter combinations. This module uses the signal-to-noise ratio gain as the fitness function, and the larger the value, the better the signal enhancement effect.

[0056] The barrier height parameter 'a' is used to control the depth of the left and right potential wells in a bistable system. The larger 'a' is, the deeper the potential well, the more stable the system, and the weaker the response to weak signals. The smaller 'a' is, the shallower the potential well, the more sensitive the system, and the stronger the response to weak signals.

[0057] The barrier height parameter b is used to control the system barrier height and nonlinear strength, affecting the difficulty of resonance and the signal amplification amplitude.

[0058] The noise intensity parameter D is used to characterize the average energy of the background noise. If D is too small, resonance cannot be excited, and if D is too large, the signal will be overwhelmed. There is an optimal value.

[0059] The parameter adjustment range and step size are determined based on the characteristics of gear scuffing fault signals, system dynamics, and engineering tests to ensure an effective search space and high optimization efficiency.

[0060] The specific working principle of the parameter adjustment unit for adjusting parameters is as follows: The parameter adjustment unit uses particle swarm optimization as its core, constructing a search space by using barrier height parameters a and b and noise intensity parameter D as three-dimensional optimization variables. The parameter ranges are set according to experimental calibration: a∈[0.5,2.0], b∈[0.1,0.8], D∈[0.01,0.1], and the adjustment step size is uniformly 0.01 to ensure a balance between parameter adjustment accuracy and search efficiency.

[0061] The algorithm initializes a swarm of particles, each representing a set of parameters (a, b, D). Particles move at a certain speed in the search space, continuously updating their position and velocity based on their historical best position and the swarm's global best position. In each iteration, the particle substitutes its current parameters into the nonlinear bistable system, calculates the signal-to-noise ratio gain of the output signal, and uses this gain as its fitness value. A higher fitness indicates a better signal enhancement effect corresponding to that set of parameters.

[0062] During the iteration process, the particle swarm continuously converges towards a high-fitness region, ultimately rapidly searching for the optimal parameter combination that maximizes the signal-to-noise ratio gain. The parameter adjustment unit sends the optimal parameters to the nonlinear system evolution unit in real time, ensuring the system always operates in the optimal stochastic resonance state, thereby maximizing the signal enhancement effect. The entire optimization process is completed in milliseconds, meeting real-time processing requirements.

[0063] Particle swarm optimization (PSO) possesses excellent global search capabilities, effectively avoiding local optima and finding the parameter combination that truly maximizes signal enhancement. PSO converges rapidly with short optimization times per iteration, enabling real-time online signal enhancement. It covers most industrial operating conditions and fault types, ensuring effective enhancement across various scenarios.

[0064] In some embodiments, the parameter adjustment unit includes: The oil film state identification subunit is used to analyze the acoustic emission signal and temperature field distribution signal in the multi-source signal in real time, extract the high-frequency components of the acoustic emission signal and the gradient change of the temperature field distribution signal, and identify the lubricating oil film state of the current gear meshing area online. The lubricating oil film state includes elastohydrodynamic lubrication state, mixed lubrication state, boundary lubrication state and oil film rupture state. The parameter mapping subunit is connected to the oil film state identification subunit. It is used to store the pre-calibrated oil film state-random resonance parameter mapping table, and according to the current lubricating oil film state output by the oil film state identification subunit, it queries and outputs the adjustment range and initial value of the corresponding barrier height parameters a, b and noise intensity parameter D from the oil film state-random resonance parameter mapping table. A dynamic adjustment execution subunit, connected to the parameter mapping subunit, is used to dynamically adjust the barrier height parameters a and b and the noise intensity parameter D within the adjustment range based on the adjustment range and initial value output by the parameter mapping subunit, using a particle swarm optimization algorithm with the goal of maximizing the signal-to-noise ratio gain of the output signal. The oil film state-random resonance parameter mapping table is pre-calibrated through multi-condition calibration experiments on a gearbox test bench. When the oil film state identification subunit identifies that the current lubricating oil film state has switched from elastohydrodynamic lubrication state to boundary lubrication state or oil film rupture state, the lower limit of the adjustment range of the barrier height parameter a output by the parameter mapping subunit is reduced to 50%-70% of the original lower limit of the adjustment range. This enhances the resonance amplification capability of the chaotic resonance enhancement module for the weak impact components in the acoustic emission signal that characterize the initiation stage of bonding.

[0065] In this embodiment, the oil film state identification subunit analyzes the high-frequency components of the acoustic emission signal, the temperature field gradient change, and the vibration and impact characteristics to identify the lubrication state of the lubricating oil film in the gear meshing area online and determine the integrity of the oil film.

[0066] Oil film condition includes elastohydrodynamic lubrication, mixed lubrication, boundary lubrication, and oil film rupture, and is the core physical indicator for judging the risk of adhesion.

[0067] The oil film state-random resonance parameter mapping table refers to a data table establishing the correspondence between oil film state and optimal random resonance parameters through numerous calibration tests conducted on a gearbox test bench under different speeds, loads, and lubrication conditions.

[0068] The parameter mapping subunit is used to query and output the initial values ​​and adjustment ranges of the appropriate parameters a, b, and D from the mapping table based on the real-time identified oil film state, providing physical prior guidance for particle swarm optimization.

[0069] The dynamic adjustment execution subunit is used to continue using the particle swarm optimization algorithm for fine optimization within the range given by the parameter mapping subunit, realizing a two-layer adjustment of "physical prior + intelligent optimization".

[0070] The weak impact component refers to the high-frequency, low-amplitude impact signal generated by the rupture of the oil film and the contact of the micro-protrusions on the tooth surface in the early stage of the bonding failure. It is the most critical characteristic of early bonding.

[0071] The working principle of this embodiment is as follows: First, the oil film state identification subunit extracts key features from multi-source signals, specifically including: performing wavelet decomposition on the acoustic emission signal to extract high-frequency component energy and kurtosis index; calculating the spatial gradient and temperature rise rate of the temperature field distribution signal; and combining the impact characteristics of the vibration signal with the operating parameters to comprehensively determine whether the current gear meshing area lubricating oil film is in a state of elastohydrodynamic lubrication, mixed lubrication, boundary lubrication, or oil film rupture.

[0072] The identification results are sent to the parameter mapping subunit.

[0073] The parameter mapping subunit has a built-in pre-calibrated oil film state-random resonance parameter mapping table. Based on the current oil film state, it directly outputs the initial values ​​and recommended adjustment ranges of the corresponding parameters a, b, and D. When the oil film state switches from good elastohydrodynamic lubrication to boundary lubrication or oil film rupture, it indicates a sharp increase in the risk of galling. At this time, the parameter mapping subunit automatically reduces the lower limit of the adjustment range of the barrier height parameter a to 50%-70% of the original lower limit, reducing the system barrier height and making the system more sensitive to high-frequency weak impact signals, thereby enhancing the amplification capability of early galling characteristics.

[0074] The dynamic adjustment execution subunit receives the initial values ​​and adjustment range mentioned above. Under this physical prior constraint, it continues to run the particle swarm optimization algorithm to perform fine-tuning with the goal of maximizing the signal-to-noise ratio gain, and finally outputs the optimal parameter combination. Through this two-layer adjustment mechanism of "physical prediction first, then intelligent optimization", the physical rationality of parameter adjustment is ensured, while optimization efficiency and accuracy are improved.

[0075] The above scheme introduces the oil film state as a physical prior, avoiding blind particle swarm searching, significantly shortening the convergence time, and improving the physical consistency of parameter adjustment. When the oil film ruptures and the risk of adhesion increases, the potential barrier height is actively reduced to specifically amplify the weak impact signal in the early stage of adhesion, achieving earlier and more sensitive fault detection.

[0076] In some embodiments, the decoupled representation learning module includes: A fault feature extractor is used to extract features from the resonance enhancement signal; A condition discriminator is used to predict the condition category based on the features output by the fault feature extractor. And a gradient inversion layer connecting the fault feature extractor and the operating condition discriminator; During the training phase, the gradient reversal layer reverses the gradient flowing from the condition discriminator to the fault feature extractor with a fixed negative coefficient during backpropagation. The negative coefficient ranges from -0.2 to -0.05. By inverting the gradient through the gradient inversion layer, the fault feature extractor is trained to generate the decoupled fault feature vector that is orthogonal to the operating condition information.

[0077] In this embodiment, the fault feature extractor is composed of a multi-layer convolutional neural network, which is used to learn deep high-dimensional features directly related to gear scuffing faults from the resonance enhancement signal, including time-domain impact features, frequency-domain resonance features, and time-frequency joint features.

[0078] The structure of the operating condition feature extractor is the same as that of the fault feature extractor, and it is used to learn operating condition-related features such as speed, load, and torque from the same enhanced signal.

[0079] The working condition discriminator is composed of a fully connected neural network and is used to predict the working condition category and parameters from the feature vector. It is a discriminator in adversarial learning.

[0080] Gradient Reversal Layer (GRL) is a special network layer that directly transmits features during forward propagation without any transformation; during backward propagation, the gradient is multiplied by a fixed negative coefficient to reverse the gradient direction.

[0081] Orthogonal decoupling refers to the fact that in a high-dimensional feature space, the inner product of the fault feature vector and the operating condition feature vector approaches 0 and is orthogonal to each other, indicating that the fault features do not contain operating condition information.

[0082] The working principle of the decoupled representation learning module is as follows: First, the resonance enhancement signal is simultaneously input into both the fault feature extractor and the operating condition feature extractor. Both extractors employ a 3-layer convolutional neural network with kernel sizes of 3×3, 5×3, and 3×3, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and finally, a 128-dimensional feature vector is output through global average pooling.

[0083] The feature vector output by the fault feature extractor is fed into the operating condition discriminator. The operating condition discriminator is a 2-layer fully connected network with 128-dimensional input and 3-dimensional output, corresponding to three types of operating condition parameters: speed, torque, and load, respectively, used to predict operating condition information.

[0084] The gradient inversion layer is inserted between the fault feature extractor and the operating condition discriminator. During forward propagation, the gradient inversion layer directly transmits the fault features to the operating condition discriminator without changing the feature values. During backpropagation to update the parameters, the gradient inversion layer multiplies the gradient returned from the operating condition discriminator by a fixed negative coefficient in the range of [-0.2, -0.05], thus completely reversing the gradient direction.

[0085] Under this mechanism, the optimization goal of the operating condition discriminator is to identify the operating condition from the fault features as accurately as possible; while the optimization goal of the fault feature extractor, due to gradient inversion, becomes to prevent the operating condition discriminator from identifying the operating condition. The two form an adversarial game. After multiple iterations of training, the fault feature extractor eventually learns feature representations completely unrelated to the operating condition information, making the fault features and operating condition features strictly orthogonal in high-dimensional space, with an orthogonality error of less than 0.01, thus achieving complete decoupling and outputting a pure fault feature vector.

[0086] The above scheme completely eliminates the influence of operating conditions such as speed, load, and torque at the feature level through orthogonal adversarial decoupling, solving the pain point of traditional models where "accuracy drops sharply when operating conditions change." The model automatically learns the optimal decoupling representation, eliminating the need for manual design of filtering, normalization, and feature selection rules, resulting in stronger adaptability and better generalization. Using orthogonality error as the evaluation metric, the error is strictly controlled to be less than 0.01, ensuring stable, consistent, and reproducible decoupling effects.

[0087] In some embodiments, the meta-learning adaptive module includes: Meta-task construction unit, used to divide historical fault data into multiple meta-tasks according to operating condition categories. Each meta-task It contains a support set and a query set; The inner loop optimization unit, connected to the meta-task construction unit, is used in each meta-task Internally, based on its support set, the current model parameters Perform inner gradient updates to obtain task-specific parameters. The updated formula is as follows: ; in, The inner learning rate, For meta-task loss function on, Represents the loss function Regarding model parameters The gradient vector; The outer loop meta-update unit, connected to the inner loop optimization unit, is used to update the current model parameters based on the query set loss of all meta-tasks. Perform a meta-update to learn a general model initialization parameter. Its optimization objective is: ; The rapid adaptation unit, connected to the outer loop element update unit, is used to initialize the parameters based on the general model when fault data of a new operating condition is received. Fine-tuning is performed where the number of gradient update steps required for fine-tuning is L≤10, and the number of samples used in each step is K≤5.

[0088] In this embodiment, the meta-task refers to encapsulating the fault dataset under each working condition into an independent learning task, including a support set (for rapid model adaptation) and a query set (for evaluating the adaptation effect), simulating the scenario of "rapid learning under new working conditions".

[0089] The support set refers to a small set of labeled samples used for gradient updates within a single task within the meta-task, corresponding to a small amount of data that can be obtained under the new working condition.

[0090] A query set refers to the set of samples within a meta-task used to evaluate the model's generalization ability on that task, and is used to guide the optimization of outer parameters.

[0091] Inner loop optimization refers to using a support set to update model parameters with a small number of gradients within a single meta-task, simulating the process of "rapid fine-tuning under new operating conditions".

[0092] Outer loop meta-update refers to uniformly optimizing the initial parameters of the model on the query set of all meta-tasks, so that the parameters can achieve high accuracy with only a small number of updates on any new task.

[0093] Common model initialization parameters It refers to a set of "optimal starting point" parameters obtained through multi-condition training, which can significantly reduce the difficulty of adapting to new conditions and the sample requirements.

[0094] A fast adaptation unit refers to a model that adapts itself to operating conditions by using no more than 10 samples and no more than 5 gradient updates based on common model initialization parameters, and outputs an adaptive diagnostic model.

[0095] In this embodiment, the meta-learning adaptive module uses a two-layer optimization structure to achieve cross-condition small-sample adaptation. The overall process is divided into a meta-training stage and a rapid adaptation stage. The specific working principle is as follows: First, the meta-task construction unit divides historical fault data into operating condition categories such as speed, torque, and load, and each category is constructed as an independent meta-task. Based on typical gearbox operating conditions, the tasks can be categorized into 9 types according to combinations of rotational speed (500 r / min, 1000 r / min, 1500 r / min) and torque (50 N•m, 100 N•m, 150 N•m). Each task is further divided into a support set and a query set at a 1:4 ratio. The support set is used for internal loop updates, and the query set is used for external loop evaluation.

[0096] Then the inner loop optimization unit optimizes each metatask. Internally, based on the current model parameters Starting with the support set samples, calculate the fault diagnosis loss and perform gradient updates according to the following formula: ,in, This is the inner learning rate, typically set to 0.01. Cross-entropy loss function, used for classification tasks. Represents the loss function Regarding model parameters The gradient vector; only a small number of updates are performed within each meta-task (usually 3 times), simulating the constraint of "only a limited number of fine-tunings" in real-world scenarios.

[0097] Then, the outer loop meta-update unit optimizes the current model parameters by using the sum of the query set losses of all meta-tasks as the optimization objective. Update and learn model initialization parameters that are applicable across different working conditions. The formula is as follows: The core objective of this process is not to make the model optimal under a single working condition, but to optimize the current model parameters. It has the ability to "quickly improve under any new operating conditions". The outer loop learning rate is set to 0.001, and iterates 300 times to ensure full convergence.

[0098] When the system encounters new operating conditions it has never seen before, it quickly adapts to the unit loading model initialization parameters. Using only K≤10 labeled samples and performing L≤5 gradient updates, a high-precision and robust adaptive diagnostic model for new working conditions can be quickly obtained, with a diagnostic accuracy of no less than 95% on new working conditions.

[0099] The above solution utilizes meta-learning to learn common fault patterns across operating conditions. New operating conditions require only a handful of samples for adaptation, completely resolving the pain point of traditional deep learning's reliance on massive amounts of labeled data. Model fine-tuning takes ≤5 steps, completing adaptation in seconds, eliminating the need for extensive data re-collection and lengthy training, significantly shortening the on-site deployment cycle. Through rapid adaptation via inner-loop simulation and initial point optimization via outer-loop, the model maintains balanced performance across multiple operating conditions, avoiding overfitting to a single condition and improving overall robustness. Pure fault features eliminate operating condition interference, making meta-learning more adept at learning common fault patterns, further improving the accuracy and stability of small-sample adaptation. It is suitable for variable industrial scenarios, quickly adapting to new equipment, loads, speeds, and environments, meeting the changing operating condition maintenance needs of equipment in wind power, rail transit, and construction machinery.

[0100] In some embodiments, a physical constraint regularization module 106 is further included, connected to the meta-learning adaptive module, the physical constraint regularization module comprising: The theoretical calculation unit is used to calculate the theoretical temperature rise rate Y during the development of adhesive failure based on the input current operating parameters. The formula for calculating Y is as follows: Y = ,in, The coefficient of friction, For rotational speed, For torque, The thermal conversion coefficient; The deviation detection unit is used to calculate the temperature rise rate predicted by the meta-learning adaptive module. The deviation between the theoretical temperature rise rate output by the theoretical calculation unit and the theoretical temperature rise rate. The regularization penalty unit is used to construct a physical consistency regularization term based on the deviation output by the deviation detection unit. This is then incorporated into the training loss function of the meta-learning adaptive module; wherein the physical consistency regularization term is expressed as follows: ; Where λ is the regularization weight coefficient. Indicates the calculation of L2 norm; The regularization penalty unit is configured to increase the regularization weight coefficient λ to strengthen the physical constraint when the deviation between the predicted temperature rise rate and the theoretical temperature rise rate increases.

[0101] In this embodiment, the physical constraint regularization module incorporates the physical mechanism of gear scuffing failure into the data-driven model training, constructs a physical consistency regularization term, and constrains the model prediction results to conform to the laws of thermodynamics and tribology.

[0102] The theoretical temperature rise rate refers to the ideal temperature rise rate calculated based on the physical formula for heat generation from gear friction, reflecting the true physical trend of adhesive development.

[0103] coefficient of friction This refers to the interfacial friction parameters determined by the gear material and lubrication conditions. For 40Cr steel, the parameters are 0.008-0.012 under lubricating oil conditions.

[0104] thermal conversion coefficient It refers to the proportionality coefficient of frictional work converted into heat energy, reflecting the energy conversion efficiency. The preferred value is 0.75.

[0105] The deviation detection unit is used to calculate the relative error between the model's predicted temperature rise and the theoretical temperature rise, and to determine whether the model deviates from the physical laws.

[0106] The physical consistency regularization term refers to adding the temperature rise deviation in the form of the L2 norm to the total loss, thus penalizing predictions that deviate from physical laws.

[0107] The adaptive regularization weight λ is dynamically adjusted according to the magnitude of the deviation. The larger the deviation, the stronger the penalty, and the smaller the deviation, the weaker the penalty, thus achieving a dynamic balance between data-driven and physical-driven approaches.

[0108] The physical constraint regularization module and the meta-learning adaptive module are jointly trained to embed prior physical knowledge into the model learning process, correcting model prediction biases from a mechanistic perspective. The working principle is as follows: First, the theoretical calculation unit calculates based on the operating parameters (speed). Torque Material parameters (coefficient of friction) ) and energy conversion parameters (thermal conversion coefficient) The theoretical temperature rise rate Y during the gear scuffing process is calculated using the following formula: Y = Among them, rotational speed Converted to rad / s, torque M is in N•m, and all parameters are from real-time acquired operating condition signals.

[0109] Then, the deviation detection unit reads the predicted temperature rise rate output by the meta-learning model in real time. The relative deviation between the predicted and theoretical temperature rise rate Y is calculated. Preferably, the preset deviation threshold can be set to 5%: when the deviation between the predicted and theoretical temperature rise rate is less than 5%, the model prediction is considered to conform to physical laws; when the deviation between the predicted and theoretical temperature rise rate is greater than 5%, the model is judged to deviate from the physical mechanism, and regularization penalty is activated.

[0110] Regularized penalty units construct physical constraint regularization terms And add it to the total meta-learning loss function: Among them, the regularization weight coefficient The following adaptive adjustments can be made: The initial value is set to 0.1. For every 1% increase in the predicted temperature rise rate compared to the theoretical temperature rise rate, Increase by 0.05, with a maximum of 1.0. The greater the deviation between the predicted and theoretical temperature rise rates, the stronger the penalty, forcing the model predictions to conform to physical laws.

[0111] In the meta-learning inner and outer loop optimization, the total loss includes both diagnostic classification loss and physical regularization loss, so that the model can strictly follow the physical laws of gear friction heat generation and temperature rise evolution while ensuring diagnostic accuracy, and avoid erroneous predictions that violate the mechanism.

[0112] The above scheme embeds physical mechanisms into model training, enabling diagnostic results to possess both statistical accuracy and physical plausibility. Temperature rise constraints prevent unreasonable model outputs under sparse data and extreme conditions, thus improving reliability under extreme conditions. Regularization weight coefficients... The model is dynamically adjusted according to the deviation between the predicted and theoretical temperature rise rates. It does not restrict the normal learning of the model, but strengthens the constraints only when it deviates from the physical laws, thus balancing accuracy and constraint strength.

[0113] In some embodiments, the dynamic causal graph reasoning module includes: The causal structure learning unit is used to learn the graph structure of the dynamic Bayesian causal graph by using a neural causal model based on the attention mechanism. The graph structure consists of nodes with physical state changes at different positions and times in the three-dimensional physical semantic field tensor and directed edges with driving relationships between nodes that conform to physical laws. The counterfactual reasoning unit, connected to the causal structure learning unit, is used to generate counterfactual samples. By comparing the differences between the real observation results and the counterfactual inference results, the confidence of the causal edges in the dynamic Bayesian causal graph is strengthened, so as to output the causal path graph representing the fault propagation path. The number of nodes in the dynamic Bayesian causal graph is set to M, which corresponds to at least one spatiotemporal event among abnormal gear meshing, bearing wear, excessive temperature, and excessive vibration. The time window of the node is set to T, and the initial confidence of the directed edge is set to C0. The number of learning iterations for the causal structure learning unit is set to I, and the learning rate for the causal structure learning unit is set to β. The counterfactual inference unit generates P counterfactual samples each time. The counterfactual sample generation uses a Monte Carlo simulation method. When the deviation between the actual observation and the counterfactual inference is less than a first deviation threshold δ1, the confidence of the corresponding causal edge increases by a first increment Δ1. When the deviation between the actual observation and the counterfactual inference is greater than a second deviation threshold δ2, the confidence decreases by a second decrement Δ2. Here, δ1 < δ2, Δ1 and Δ2 are both positive numbers, and Δ1 < Δ2. Ultimately, the confidence range of the causal edge is controlled within the lower limit C. min Up to the upper limit C max Between, of which, C min <C max And C min and C max All are positive numbers between 0 and 1.

[0114] In this embodiment, the dynamic Bayesian causal graph refers to a probabilistic graphical model that uses spatiotemporal physical events as nodes and physical driving relationships as directed edges, and updates the state and confidence over time.

[0115] Causal structure learning refers to the use of attention mechanism neural causal models to automatically mine causal dependencies between nodes from data without relying on human priors.

[0116] Counterfactual reasoning refers to generating "hypothetical scenario" samples through Monte Carlo simulation to verify the reliability of causal edges and distinguish between true causality and false correlation.

[0117] The confidence level of the causal edge is a value between 0 and 1, representing the degree of credibility of the causal relationship. It is updated iteratively through counterfactual comparison.

[0118] The fault propagation path refers to a chain structure composed of high-confidence causal edges, which shows the complete transmission link of a fault from its source to its affected area.

[0119] In this embodiment, the dynamic causal graph reasoning module, based on the physical semantic field tensor, completes fully automated learning and verification from data to causal structure. The specific working principle is as follows: Using spatiotemporal physical events in a three-dimensional physical semantic field as nodes, a total of 20 nodes are set, including: abnormal gear meshing, bearing wear, abnormal temperature rise, excessive vibration, oil film rupture, excessively rapid local temperature rise, surge in impact energy, and abnormal friction intensity. The node time window is set to 0.1s, and the initial confidence of directed edges is C0=0.5.

[0120] The causal structure learning unit employs an attention-based neural causal model, using a physical semantic field tensor as input to learn the causal dependencies between nodes. The model automatically captures physical causal chains such as "temperature increase → oil film rupture → adhesion," outputting an initial causal graph structure. The learning iterations are 200 times with a learning rate of 0.002 to ensure stable convergence of the causal structure.

[0121] The counterfactual reasoning unit then generates counterfactual samples through Monte Carlo simulation, as follows: Assuming the state of a certain node changes, infer the changing trends of other nodes and compare them with actual observations. The rules are as follows: If the deviation between the actual observation and the counterfactual inference is less than the first deviation threshold δ1 (which can be set to 10%), the causal edge confidence is increased by the first increment Δ1 (which can be set to 0.1). If the deviation between the actual observation result and the counterfactual inference result is greater than the second deviation threshold δ2 (e.g., set to 30%), the causal edge confidence is reduced by a second reduction Δ2 (which can be set to 0.2). The final confidence level was constrained to be between 0.1 and 0.9 to avoid extreme values.

[0122] Through multiple iterative verifications, genuine causal relationships are reinforced while spurious correlations are suppressed, resulting in a high-confidence dynamic Bayesian causal graph. The dynamic causal graph inference module extracts the primary and secondary paths with the highest confidence from the dynamic Bayesian causal graph, forming a fault causal propagation path diagram. This clarifies the fault origin, propagation sequence, scope of impact, and confidence level, providing direct evidence for fault tracing and maintenance.

[0123] The above-mentioned solution upgrades from "correlation" to "causality," automatically learning the physical causal structure of faults. It filters out spurious associations through hypothesis testing mechanisms, making causal paths more credible and closely aligned with physical mechanisms, thus reducing the false positive rate. By adaptively mining causal relationships from data, it is applicable to different gearbox models. Dynamically updated in 0.1-second time windows, it can track the complete development process of faults from their early stages to severity in real time. By outputting high-confidence causal paths and integrating them with data-driven diagnostic results, it significantly improves the reliability and interpretability of the final decision.

[0124] In some embodiments, the dynamic causal graph reasoning module includes: The causal path encoding unit is used to encode the causal path graph output by the dynamic causal graph reasoning module into a causal path feature vector.

[0125] The diagnostic result fusion unit is used to weight and fuse the causal path feature vector with the output of the working condition adaptive diagnostic model to generate a fused diagnostic result.

[0126] The weight adaptive unit is used to dynamically adjust the fusion weights based on the current operating parameters and the confidence level of the causal path graph.

[0127] In some embodiments, the causal path encoding unit uses a graph neural network (GNN) to encode the causal path graph, outputting a 64-dimensional causal path feature vector. During the encoding process, causal edges with a high confidence level ≥ 0.8 are assigned higher encoding weights. The diagnostic result fusion unit uses a weighted summation method for fusion, with the fusion formula being: Fusion diagnostic result = ω1 × causal path feature vector prediction value + ω2 × operating condition adaptive diagnostic model output value, where ω1 + ω2 = 1. In the weight adaptive unit, when the average confidence level of the causal path graph is ≥ 0.7, ω1 = 0.6, ω2 = 0.4; when the average confidence level is < 0.7, ω1 = 0.4, ω2 = 0.6. Simultaneously, the weights are fine-tuned by ±0.05 based on changes in rotational speed and torque.

[0128] In some embodiments, the output module includes: The 3D model rendering unit is used to load the 3D model of the gearbox and perform visualization rendering based on the 3D model of the gearbox. A causal path drawing unit, connected to the three-dimensional model rendering unit, is used to draw the fault causal propagation path represented by the causal path diagram on the three-dimensional model of the gearbox in the form of directed edges with arrows. The confidence level identification unit is connected to the causal path drawing unit and is used to identify causal edges with different confidence levels using differentiated colors based on the confidence level of each causal edge in the causal path graph output by the dynamic causal graph inference module. Wherein, the confidence level is the causal edge confidence level enhanced by the dynamic causal graph reasoning module through counterfactual reasoning; Causal edges with a confidence level higher than a preset first confidence level threshold are classified as high-confidence causal edges, and causal edges with a confidence level lower than a preset second confidence level threshold are classified as low-confidence causal edges, wherein the preset first confidence level threshold is greater than the preset second confidence level threshold.

[0129] In some embodiments, a 3D model rendering unit refers to loading a gearbox STL format 3D model and providing a 1:1 real physical space rendering base.

[0130] The causal path drawing unit draws the fault propagation path on the 3D model with directed edges bearing arrows, intuitively showing the direction of transmission.

[0131] The confidence level labeling unit uses different saturation colors to mark the confidence level of causal edges, quickly distinguishing between high / medium / low confidence paths.

[0132] The preset first confidence threshold can be set to 0.8, and the preset second confidence threshold can be set to 0.5. High-confidence causal edges are those with a confidence value greater than or equal to 0.8, and can be marked with high-saturation red to represent a definite fault propagation relationship; high-confidence causal edges are those with a confidence value in the range of [0.5, 0.8], and are marked with medium-saturation yellow to represent a high-probability causal relationship; low-confidence causal edges are those with a confidence value less than 0.5, and are marked with low-saturation blue to represent an uncertain association.

[0133] In this embodiment, the output module focuses on 3D visualization and confidence level grading to transform abstract diagnostic results into intuitive images. The specific working principle is as follows: The 3D model rendering unit loads the gearbox STL format 3D model, renders it at a 1:1 realistic scale, with a frame rate of ≥30fps, supports interactive operations such as rotation, scaling, and translation, and provides a complete spatial perspective. The causal path drawing unit maps the path output from the dynamic causal graph to the corresponding physical location on the 3D model, drawing directed edges with 2mm wide lines, 5mm long arrows, and 0.8 opacity, clearly showing the complete path of the fault from the starting point to the propagation area.

[0134] The confidence level indicator automatically assigns colors based on the confidence level of the causal edge. Specifically, it assigns a high-saturation red to causal edges with high confidence, a medium-saturation yellow to causal edges with medium confidence, and a low-saturation blue to causal edges with low confidence. Additionally, a two-decimal-digit confidence level value is displayed next to each causal edge to visually represent its reliability.

[0135] Then, text information such as fault type, severity, occurrence time, and total confidence score are overlaid on the 3D view to form a complete visual diagnostic report.

[0136] The above solution presents the fault location and propagation path using a 3D model, which can be quickly understood without a professional background, significantly improving on-site maintenance efficiency. Through confidence level color grading, high-confidence paths are highlighted, helping maintenance personnel prioritize the most likely fault sources and transmission chains. It supports observation of fault distribution from any angle, meeting the inspection needs of complex gearbox internal structures. Simultaneously, it displays diagnostic results, location, severity level, causal path, and confidence level, providing all maintenance decision-making information in a one-stop shop.

[0137] In the second aspect, such as Figure 2 As shown, the present invention provides a multi-sensor fusion method for processing gear scuffing fault data. The method is applicable to the system described in the first aspect of the present invention, and includes the following steps: S1: Synchronously acquire multi-source signals, including gearbox vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals; S2: Input the collected multi-source signals into a nonlinear bistable physical model. By adjusting the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model, the physical model is made to enter a stochastic resonance state, thereby outputting a resonance enhancement signal with enhanced signal-to-noise ratio. S3: Extract the first feature vector and the second feature vector from the resonance enhancement signal, and make the first feature vector orthogonal to the working condition information through adversarial training to generate a decoupled fault feature vector; wherein, the first feature vector is used to characterize the scuffing fault state of the gear, and the second feature vector is used to characterize the operating condition of the gearbox. S4: Perform a model-independent meta-learning process, specifically including: constructing multiple meta-tasks based on historical fault data for meta-training to learn a set of model initialization parameters that are universal across working conditions; when facing a new working condition, based on the universal model initialization parameters, using a preset number K of labeled samples collected under the new working condition to perform a finite number L gradient updates on the working condition adaptive diagnostic model, and output a working condition adaptive diagnostic model adapted to the new working condition. S5: The sensor readings corresponding to each sampling time in the resonance enhancement signal are mapped to the three-dimensional spatial grid of the gearbox according to the sensor installation position, and the vibration signal, temperature field distribution signal, and acoustic emission signal are converted into physical quantities that characterize the dynamic state, thermodynamic state, and tribological state, respectively, to obtain multi-physical domain state information and fill it into the corresponding position of the three-dimensional spatial grid to construct a three-dimensional physical semantic field tensor containing spatial position information and multi-physical domain state information; S6: Based on the three-dimensional physical semantic field tensor, construct a dynamic Bayesian causal graph with the physical state changes at different positions and times in the three-dimensional physical semantic field tensor as nodes and the driving relationship between nodes that conforms to physical laws as directed edges. Then, through graph structure learning and counterfactual reasoning based on attention mechanism, output a causal path graph representing the fault propagation path. S7: The diagnostic results output by the adaptive diagnostic model of the working condition, which include the fault type and fault severity, are weighted and fused with the fault causal propagation path information represented by the causal path diagram to generate the final fused diagnostic result. S8: Visualize the fusion diagnostic results.

[0138] Finally, it should be noted that although the above embodiments have been described in the description and drawings of this invention, this should not limit the scope of patent protection of this invention. Any technical solutions that are based on the essential concept of this invention, utilize the content described in the description and drawings of this invention to make equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this invention.

Claims

1. A multi-sensor fusion system for processing gear scuffing fault data, characterized in that, include: The multi-sensor synchronous acquisition module is used to synchronously acquire multi-source signals, including gearbox vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals. The chaotic resonance enhancement module is connected to the multi-sensor synchronous acquisition module. It is used to input the acquired multi-source signals into a nonlinear bistable physical model. By adjusting the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model, the physical model is made to enter a random resonance state, thereby outputting a resonance enhancement signal with enhanced signal-to-noise ratio. The decoupled representation learning module, connected to the chaotic resonance enhancement module, is used to extract a first feature vector and a second feature vector from the resonance enhancement signal, and to make the first feature vector orthogonal to the operating condition information through adversarial training, thereby generating a decoupled fault feature vector; wherein, the first feature vector is used to characterize the scuffing fault state of the gear, and the second feature vector is used to characterize the operating condition of the gearbox. The meta-learning adaptive module, connected to the decoupled representation learning module, is used to perform a model-independent meta-learning process, specifically including: constructing multiple meta-tasks based on historical fault data for meta-training to learn a set of model initialization parameters that are universal across working conditions; and when facing a new working condition, using the universal model initialization parameters and a preset number K of labeled samples collected under the new working condition to perform a finite number L gradient updates on the working condition adaptive diagnostic model, and output a working condition adaptive diagnostic model adapted to the new working condition. The physical semantic field mapping module, connected to the chaotic resonance enhancement module, is used to map the sensor readings corresponding to each sampling time in the resonance enhancement signal to the three-dimensional spatial grid of the gearbox according to the sensor installation position. It also converts the vibration signal, temperature field distribution signal, and acoustic emission signal into physical quantities characterizing the dynamic state, thermodynamic state, and tribological state, respectively, to obtain multi-physical domain state information and fill it into the corresponding position of the three-dimensional spatial grid to construct a three-dimensional physical semantic field tensor containing spatial position information and multi-physical domain state information. The dynamic causal graph reasoning module, connected to the physical semantic field mapping module, is used to construct a dynamic Bayesian causal graph based on the three-dimensional physical semantic field tensor. The graph uses the physical state changes at different positions and times in the three-dimensional physical semantic field tensor as nodes and the driving relationship between nodes that conforms to physical laws as directed edges. Through graph structure learning and counterfactual reasoning based on attention mechanism, the module outputs a causal path graph representing the fault propagation path. The multimodal fusion decision module, which connects the meta-learning adaptive module and the dynamic causal graph reasoning module, is used to perform weighted fusion of the diagnostic results, which include fault type and fault severity, output by the working condition adaptive diagnostic model, with the fault causal propagation path information represented by the causal path graph, to generate the final fusion diagnostic result. The output module, connected to the multimodal fusion decision module, is used to visualize and output the fusion diagnostic results.

2. The multi-sensor fusion gear scuffing fault data processing system as described in claim 1, characterized in that, The chaotic resonance enhancement module includes: The signal input unit is used to receive multi-source signals output by the multi-sensor synchronous acquisition module, and to weight and superimpose the vibration signal, acoustic emission signal, temperature field distribution signal and operating condition parameter signal according to a preset weighting coefficient to form a fused input signal; The parameter adjustment unit is used to dynamically adjust the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model. The nonlinear system evolution unit is used to execute the state evolution equations of the nonlinear bistable physical model, and the state evolution equations are in the following form: ； in, Let be the potential function of the nonlinear bistable physical model. Let a and b represent the state variables at time t, and a and b be the barrier height parameters. This represents the signal collected by the i-th type of sensor at time t. This represents the corresponding signal weight, where N is the total number of sensor categories. The background noise at time t is represented by the noise intensity parameter D. The resonant output unit is used to output the resonant enhancement signal when the signal-to-noise ratio gain of the output signal reaches a preset gain threshold.

3. The multi-sensor fusion gear scuffing fault data processing system as described in claim 2, characterized in that, The parameter adjustment unit adopts a particle swarm optimization algorithm to dynamically adjust the barrier height parameters a and b and the noise intensity parameter D with the goal of maximizing the signal-to-noise ratio gain of the output signal. Among them, the adjustment range of barrier height parameter a is 0.5-2.0, the adjustment range of barrier height parameter b is 0.1-0.8, the adjustment range of noise intensity parameter D is 0.01-0.1, and the adjustment step size is 0.

01.

4. The multi-sensor fusion gear scuffing fault data processing system as described in claim 3, characterized in that, The parameter adjustment unit includes: The oil film state identification subunit is used to analyze the acoustic emission signal and temperature field distribution signal in the multi-source signal in real time, extract the high-frequency components of the acoustic emission signal and the gradient change of the temperature field distribution signal, and identify the lubricating oil film state of the current gear meshing area online. The lubricating oil film state includes elastohydrodynamic lubrication state, mixed lubrication state, boundary lubrication state and oil film rupture state. The parameter mapping subunit is connected to the oil film state identification subunit. It is used to store the pre-calibrated oil film state-random resonance parameter mapping table, and according to the current lubricating oil film state output by the oil film state identification subunit, it queries and outputs the adjustment range and initial value of the corresponding barrier height parameters a, b and noise intensity parameter D from the oil film state-random resonance parameter mapping table. A dynamic adjustment execution subunit, connected to the parameter mapping subunit, is used to dynamically adjust the barrier height parameters a and b and the noise intensity parameter D within the adjustment range based on the adjustment range and initial value output by the parameter mapping subunit, using a particle swarm optimization algorithm with the goal of maximizing the signal-to-noise ratio gain of the output signal. The oil film state-random resonance parameter mapping table is pre-calibrated through multi-condition calibration experiments on a gearbox test bench. When the oil film state identification subunit identifies that the current lubricating oil film state has switched from elastohydrodynamic lubrication state to boundary lubrication state or oil film rupture state, the lower limit of the adjustment range of the barrier height parameter a output by the parameter mapping subunit is reduced to 50%-70% of the original lower limit of the adjustment range. This enhances the resonance amplification capability of the chaotic resonance enhancement module for the weak impact components in the acoustic emission signal that characterize the initiation stage of bonding.

5. The multi-sensor fusion gear scuffing fault data processing system as described in claim 1, characterized in that, The decoupled representation learning module includes: A fault feature extractor is used to extract features from the resonance enhancement signal; A condition discriminator is used to predict the condition category based on the features output by the fault feature extractor. And a gradient inversion layer connecting the fault feature extractor and the operating condition discriminator; During the training phase, the gradient reversal layer reverses the gradient flowing from the condition discriminator to the fault feature extractor with a fixed negative coefficient during backpropagation. The negative coefficient ranges from -0.2 to -0.

05. By inverting the gradient through the gradient inversion layer, the fault feature extractor is trained to generate the decoupled fault feature vector that is orthogonal to the operating condition information.

6. The multi-sensor fusion gear scuffing fault data processing system as described in claim 1, characterized in that, The meta-learning adaptive module includes: Meta-task construction unit, used to divide historical fault data into multiple meta-tasks according to operating condition categories. Each meta-task It contains a support set and a query set; The inner loop optimization unit, connected to the meta-task construction unit, is used in each meta-task Internally, based on its support set, the current model parameters Perform inner gradient updates to obtain task-specific parameters. The updated formula is as follows: ； in, The inner learning rate, For meta-task loss function on, Represents the loss function Regarding model parameters The gradient vector; The outer loop meta-update unit, connected to the inner loop optimization unit, is used to update the current model parameters based on the query set loss of all meta-tasks. Perform a meta-update to learn a general model initialization parameter. Its optimization objective is: ; The rapid adaptation unit, connected to the outer loop element update unit, is used to initialize the parameters based on the general model when fault data of a new operating condition is received. Fine-tuning is performed where the number of gradient update steps required for fine-tuning is L≤10, and the number of samples used in each step is K≤5.

7. The multi-sensor fusion gear scuffing fault data processing system as described in claim 1, characterized in that, It also includes a physical constraint regularization module, connected to the meta-learning adaptive module, the physical constraint regularization module comprising: The theoretical calculation unit is used to calculate the theoretical temperature rise rate Y during the development of adhesive failure based on the input current operating parameters. The formula for calculating Y is as follows: Y = ,in, The coefficient of friction, For rotational speed, For torque, The thermal conversion coefficient; The deviation detection unit is used to calculate the temperature rise rate predicted by the meta-learning adaptive module. The deviation between the theoretical temperature rise rate output by the theoretical calculation unit and the theoretical temperature rise rate. The regularization penalty unit is used to construct a physical consistency regularization term based on the deviation output by the deviation detection unit. This is then incorporated into the training loss function of the meta-learning adaptive module; wherein the physical consistency regularization term is expressed as follows: ； Where λ is the regularization weight coefficient. Indicates the calculation of L2 norm; The regularization penalty unit is configured to increase the regularization weight coefficient λ to strengthen the physical constraint when the deviation between the predicted temperature rise rate and the theoretical temperature rise rate increases.

8. The multi-sensor fusion gear scuffing fault data processing system as described in claim 1, characterized in that, The dynamic cause-effect graph reasoning module includes: The causal structure learning unit is used to learn the graph structure of the dynamic Bayesian causal graph by using a neural causal model based on the attention mechanism. The graph structure consists of nodes with physical state changes at different positions and times in the three-dimensional physical semantic field tensor and directed edges with driving relationships between nodes that conform to physical laws. The counterfactual reasoning unit, connected to the causal structure learning unit, is used to generate counterfactual samples. By comparing the differences between the real observation results and the counterfactual inference results, the confidence of the causal edges in the dynamic Bayesian causal graph is strengthened, so as to output the causal path graph representing the fault propagation path. The number of nodes in the dynamic Bayesian causal graph is set to M, which corresponds to at least one spatiotemporal event among abnormal gear meshing, bearing wear, excessive temperature, and excessive vibration. The time window of the node is set to T, and the initial confidence of the directed edge is set to C0. The number of learning iterations for the causal structure learning unit is set to I, and the learning rate for the causal structure learning unit is set to β. The counterfactual inference unit generates P counterfactual samples each time. The counterfactual sample generation uses a Monte Carlo simulation method. When the deviation between the actual observation and the counterfactual inference is less than a first deviation threshold δ1, the confidence of the corresponding causal edge increases by a first increment Δ1. When the deviation between the actual observation and the counterfactual inference is greater than a second deviation threshold δ2, the confidence decreases by a second decrement Δ2. Here, δ1 < δ2, Δ1 and Δ2 are both positive numbers, and Δ1 < Δ2. Ultimately, the confidence range of the causal edge is controlled within the lower limit C. min Up to the upper limit C max Between, of which, C min <C max And C min and C max All are positive numbers between 0 and 1.

9. The multi-sensor fusion gear scuffing fault data processing system as described in claim 1, characterized in that, The output module includes: The 3D model rendering unit is used to load the 3D model of the gearbox and perform visualization rendering based on the 3D model of the gearbox. A causal path drawing unit, connected to the three-dimensional model rendering unit, is used to draw the fault causal propagation path represented by the causal path diagram on the three-dimensional model of the gearbox in the form of directed edges with arrows. The confidence level identification unit is connected to the causal path drawing unit and is used to identify causal edges with different confidence levels using differentiated colors based on the confidence level of each causal edge in the causal path graph output by the dynamic causal graph inference module. Wherein, the confidence level is the causal edge confidence level enhanced by the dynamic causal graph reasoning module through counterfactual reasoning; Causal edges with a confidence level higher than a preset first confidence level threshold are classified as high-confidence causal edges, and causal edges with a confidence level lower than a preset second confidence level threshold are classified as low-confidence causal edges, wherein the preset first confidence level threshold is greater than the preset second confidence level threshold.

10. A method for processing gear scuffing fault data using multi-sensor fusion, characterized in that, The method is applicable to the system as described in any one of claims 1 to 9, and the method includes the following steps: S1: Synchronously acquire multi-source signals, including gearbox vibration signals, acoustic emission signals, temperature field distribution signals, and operating condition parameter signals; S2: Input the collected multi-source signals into a nonlinear bistable physical model. By adjusting the barrier height parameter and noise intensity parameter of the nonlinear bistable physical model, the physical model is made to enter a stochastic resonance state, thereby outputting a resonance enhancement signal with enhanced signal-to-noise ratio. S3: Extract the first feature vector and the second feature vector from the resonance enhancement signal, and make the first feature vector orthogonal to the working condition information through adversarial training to generate a decoupled fault feature vector; wherein, the first feature vector is used to characterize the scuffing fault state of the gear, and the second feature vector is used to characterize the operating condition of the gearbox. S4: Execute the model-independent meta-learning process, which specifically includes: constructing multiple meta-tasks based on historical fault data for meta-training to learn a set of model initialization parameters that are universal across working conditions; when facing a new working condition, based on the universal model initialization parameters, using a preset number K of labeled samples collected under the new working condition to perform a finite number L gradient updates on the working condition adaptive diagnostic model, and outputting a working condition adaptive diagnostic model adapted to the new working condition. S5: The sensor readings corresponding to each sampling time in the resonance enhancement signal are mapped to the three-dimensional spatial grid of the gearbox according to the sensor installation position, and the vibration signal, temperature field distribution signal, and acoustic emission signal are converted into physical quantities that characterize the dynamic state, thermodynamic state, and tribological state, respectively, to obtain multi-physical domain state information and fill it into the corresponding position of the three-dimensional spatial grid to construct a three-dimensional physical semantic field tensor containing spatial position information and multi-physical domain state information; S6: Based on the three-dimensional physical semantic field tensor, construct a dynamic Bayesian causal graph with the physical state changes at different positions and times in the three-dimensional physical semantic field tensor as nodes and the driving relationship between nodes that conforms to physical laws as directed edges. Then, through graph structure learning and counterfactual reasoning based on attention mechanism, output a causal path graph representing the fault propagation path. S7: The diagnostic results output by the adaptive diagnostic model of the working condition, which include the fault type and fault severity, are weighted and fused with the fault causal propagation path information represented by the causal path diagram to generate the final fused diagnostic result. S8: Visualize the fusion diagnostic results.

Images