A worm and gear speed reducer meshing wear state identification method and system

Abstract

The application discloses a kind of worm gear reducer meshing wear state identification method and system, it is related to the technical field of speed reducer condition monitoring and fault diagnosis, method includes multi-source heterogeneous signal synchronous acquisition and preprocessing, working condition perception and feature self-adaptive correction, multi-domain sensitive feature extraction and feature level fusion, layered progressive wear recognition and fault decoupling, offline-online closed-loop model self-update, system includes multi-source synchronous perception, working condition self-adaptive correction, multi-source feature fusion diagnosis, digital twin wear evolution, offline-online closed-loop iteration, hierarchical early warning and operation management, the present application solves the problem that early wear signal is weak in low-speed condition by sound-vibration-oil-electricity multi-source fusion, eliminate variable working condition interference by working condition self-adaptive correction, realize coupling fault decoupling by multi-label deep network, solve long-term model drift problem by offline-online closed loop, can significantly improve the accuracy of worm gear reducer full-cycle wear recognition.

Description

Technical Field

[0001] This invention relates to a broader range of applications, specifically a method and system for identifying the meshing wear state of a worm gear reducer. Background Technology

[0002] In the existing industrial gear reducer operation and maintenance system, the identification of worm gear meshing wear status mainly relies on periodic disassembly inspection and a single vibration threshold alarm mode. Maintenance personnel usually determine a fixed set of status thresholds (such as vibration intensity and oil element concentration) through offline calibration. Subsequently, the online monitoring system executes alarms according to the fixed thresholds. However, under actual complex working conditions, the factors affecting the accuracy of wear status identification are highly time-varying and nonlinear.

[0003] First, worm gears mostly operate under low-speed, heavy-load conditions, with the tooth surfaces primarily engaged in sliding meshing. The impact energy generated by early micro-wear is weak and easily masked by equipment vibration, environmental background noise, and external interference transmitted from the foundation structure. Single vibration sensors have insufficient signal-to-noise ratios, making it difficult to capture micron-level early wear characteristics. While oil monitoring can quantify wear products, there is a significant lag between the abrasive particles detaching from the tooth surface and being detected through lubricant circulation, making it impossible to detect sudden failures such as tooth surface scuffing, or accurately locate the wear position and differentiate failure types. Second, dynamic fluctuations in equipment speed and load can lead to… The vibration amplitude and meshing frequency change synchronously. The existing fixed threshold system cannot distinguish the characteristic changes caused by the fluctuation of the working condition from those caused by wear and degradation. In scenarios such as variable frequency speed regulation and variable load conveying, false alarms and missed alarms occur frequently. In addition, in actual operation, the worm gear tooth surface often exhibits coupled failure modes such as uniform wear and pitting, and pitting and mild scuffing. Traditional single-label fault classification models can only output a single fault type and cannot achieve decoupled identification of multiple failure modes. Moreover, conventional time-frequency analysis methods have limited ability to extract weak wear features under strong noise, resulting in insufficient accuracy in early fault identification. Traditional periodic maintenance can only be carried out by disassembling the machine after a fault occurs, which cannot achieve early warning and is prone to unplanned downtime and production losses. Although some existing multi-sensor monitoring solutions have introduced multiple types of signals, they are mostly simple threshold parallel judgments. They have not achieved deep fusion of multi-source features, nor have they established an adaptive correction mechanism for operating conditions. They are still difficult to adapt to the stable identification requirements under changing operating conditions, and cannot solve the core pain points of decoupling coupled faults and long-term accuracy drift.

[0004] Therefore, those skilled in the art provide a method and system for identifying the meshing wear state of a worm gear reducer to solve the problems mentioned in the background art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying the meshing wear state of a worm gear reducer, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for identifying the meshing wear state of a worm gear reducer, the method comprising the following steps: S1. Synchronous acquisition and preprocessing of multi-source heterogeneous signals: Build a multi-dimensional sensing system of acoustic vibration, oil, electrical and working conditions, complete high-precision synchronous acquisition and targeted noise reduction preprocessing of multi-physical quantity signals, and build a multi-source data foundation for wear identification;

[0007] S2. Working condition perception and feature adaptive correction: Based on real-time speed and load data, working condition clustering and matching are performed, and the original features are normalized and corrected for working conditions and torque compensation is performed. In transient working conditions, order tracking processing is introduced to eliminate the interference of working condition fluctuations on wear characteristics.

[0008] S3. Multi-domain sensitive feature extraction and multi-source feature-level fusion: Extract time-domain, frequency-domain, and time-frequency-domain multi-dimensional features of each sensing source respectively, obtain a wear-sensitive feature subset through feature selection algorithm, and generate a multi-dimensional fusion feature vector by using a weighted fusion algorithm that couples subjective and objective factors.

[0009] S4. Layered progressive wear status identification and fault decoupling: A three-layer progressive diagnostic architecture of early warning layer - accurate classification layer - quantitative prediction layer is adopted. By improving the deep residual shrinkage network, the wear type and degree classification of multi-label wear is realized. The wear quantification and remaining life prediction are completed by combining the attention-enhanced long short-term memory network.

[0010] S5. Offline-Online Closed-Loop Model Self-Update: Regularly integrate offline high-precision detection data to construct labeled samples, and dynamically update the diagnostic model and health benchmark through incremental transfer learning to correct model accuracy drift after long-term operation; In step S5, when the machine needs to be shut down for maintenance, wear true value labels are obtained through visual inspection of the tooth surface and oil ferrography analysis. The online monitoring steady-state data of the corresponding time period are automatically matched according to the timestamp to generate labeled incremental training samples. The bottom feature extraction layer of the network is frozen, and the top classification and prediction layers are fine-tuned through incremental transfer learning. Combined with the domain adaptive algorithm, the differences in data distribution across devices are reduced, the model is efficiently iterated, the health feature benchmarks and early warning thresholds of each working condition cluster are corrected simultaneously, and the feature benchmark drift after long-term operation is corrected.

[0011] As a further embodiment of the present invention, step S1 specifically includes the following steps: S101. Multi-source sensing point arrangement: A high-frequency piezoelectric accelerometer and a broadband acoustic emission sensor are coaxially installed on the bearing housing end face of the worm input end, and a single-axis accelerometer is installed on the bearing housing of the worm output end. A three-phase current transformer, a speed and torque integrated sensor and an online oil sensor are also provided. S102, Hardware-level time synchronization: The PTP precise time protocol is used to achieve clock synchronization of all acquisition channels with a time synchronization error of 1ms; S103. Preprocessing of branch signals: Perform wavelet threshold noise reduction, trend term removal and bandpass filtering on vibration and acoustic emission signals, perform power frequency notch filtering and low-pass filtering on current signals, and perform moving average filtering on oil signals.

[0012] As a further embodiment of the present invention, step S2 specifically includes the following steps: S201. Offline working condition clustering and benchmark library construction: Using speed and load torque as two-dimensional working condition features, the improved K-means++ algorithm is used to divide 5 to 8 typical working condition clusters, and a corresponding health feature benchmark library is established for each working condition cluster; S202. Real-time working condition matching and determination: The Euclidean distance between the current working condition point and the center of each working condition cluster is calculated in real time, and the current working condition cluster is matched to obtain the current working condition cluster. S203. Normalization correction of characteristic working conditions: Normalize the original characteristics according to the formula xi=xi / i, where xi is the original characteristic value and i is the health mean value of the characteristic under the corresponding working condition. S204, Load Torque Compensation Correction: Introducing a torque compensation factor to perform secondary correction on the corrected characteristics, reducing the interference of load fluctuations on the characteristics; S205, Transient Condition Order Normalization: When the condition is determined to be a start-stop or acceleration / deceleration transient condition, the time-domain signal is converted into a stationary angular-domain signal by calculating the order tracking algorithm.

[0013] As a further embodiment of the present invention, step S3 specifically includes the following steps: S301. Feature extraction of each sensing source: Extract 18-dimensional features in time domain, frequency domain, and time-frequency domain from vibration signals; extract 5-dimensional impact features from acoustic emission signals; extract 4-dimensional electrical features from current signals; and extract 3-dimensional abrasive features from oil signals. S302. Sensitive Feature Filtering: Using the maximum correlation and minimum redundancy algorithm, 10-12 core sensitive features are selected based on wear level to form the initial feature vector; S303, Subjective-Objective Coupled Weighted Fusion: The subjective weight of each sensor source is determined by the analytic hierarchy process (AHP), and the objective weight is calculated by the entropy weight method. The subjective and objective weights are coupled in a 6:4 ratio to generate a multi-dimensional fusion feature vector.

[0014] As a further aspect of the present invention: the three-layer progressive diagnostic architecture in step S4 includes an early warning layer, a multi-label classification layer, and a quantitative prediction layer. The early warning layer uses acoustic emission impact count, vibration kurtosis, and abrasive hourly growth rate as input features, and adopts an adaptive thresholding method based on three criteria to achieve millisecond-level early warning. The warning threshold is dynamically updated according to the current working condition benchmark. The multi-label classification layer uses fused feature vectors as input, and performs feature extraction and classification through an improved deep residual shrinkage network with a built-in soft thresholding module. It also outputs the wear degree level and wear failure type, supporting the decoupled identification of coupled wear faults. The quantitative prediction layer adopts an attention-enhanced long short-term memory network, and combines the classification results, abrasive cumulative amount, and historical degradation trend to output the quantitative tooth surface wear depth, remaining service life, and confidence interval.

[0015] As a further aspect of the present invention: the improved deep residual shrinking network in the multi-label classification layer uses a binary cross-entropy loss function for multi-label classification training, and the loss function formula is: Where N is the number of samples, and M is the number of label categories. For real labels, This corresponds to the predicted probability.

[0016] Another aspect of the present invention provides a worm gear reducer meshing wear state identification system, including a multi-source synchronous sensing module, an operating condition adaptive correction module, a multi-source feature fusion diagnosis module, a digital twin wear evolution module, an offline and online closed-loop iteration module, and a hierarchical early warning and operation and maintenance management module; The multi-source synchronous sensing module is used to synchronously collect the vibration, acoustic emission, current, oil and operating parameters of the reducer, and complete signal conditioning and analog-to-digital conversion. The adaptive correction module is used to match the current operating condition category and complete the characteristic operating condition normalization correction, torque compensation and transient order normalization. The multi-source feature fusion diagnostic module is used to extract multi-domain features and filter and fuse them, and completes early warning, multi-label classification and wear quantification prediction based on a hierarchical and progressive diagnostic architecture. The digital twin wear evolution module is used to construct a digital twin of the wear degradation of the reducer, simulate the wear evolution process, and output the remaining service life and maintenance window period; The offline / online closed-loop iterative module is used to access offline high-precision detection data to generate labeled samples and dynamically update the diagnostic model and health benchmark. The tiered early warning and operation and maintenance management module is used to trigger three-level tiered early warnings, manage the full lifecycle health records, and generate operation and maintenance decision suggestions.

[0017] As a further aspect of the present invention: the multi-source synchronous sensing module integrates a high-precision synchronous acquisition unit, which achieves multi-channel microsecond-level time synchronization through the PTP clock synchronization protocol to ensure the timing consistency of multi-source signals; The adaptive correction module for operating conditions integrates an order tracking processing unit, which is used to convert time-domain vibration signals into angular-domain stationary signals under transient operating conditions such as start-stop, acceleration and deceleration, thereby eliminating spectral interference from speed fluctuations. The multi-source feature fusion diagnostic module integrates a soft threshold denoising unit and a built-in adaptive soft threshold module with a deep residual shrinkage network. It can automatically learn the denoising threshold based on the signal noise level and extract weak wear features.

[0018] As a further aspect of the present invention: the digital twin wear evolution module integrates a material wear characteristic library, has built-in wear rate curves and load mapping relationships for various worm gear materials, and corrects the life prediction results by combining load conditions; The offline / online closed-loop iterative module has a built-in incremental learning and domain adaptation toolchain, which supports rapid fine-tuning of the model and remote silent updates without the need for downtime or production stoppage. The graded early warning and operation and maintenance management module integrates a dual-criteria early warning verification unit, which combines vibration-acoustic emission and oil fluid signals to jointly determine anomalies, reducing the risk of unplanned downtime caused by false alarms from a single sensor.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes a high- and low-frequency complementary sensing system through coaxial acoustic and vibration arrays. It captures mid- and low-frequency macroscopic meshing characteristics with vibration signals and high-frequency microscopic wear stress waves with acoustic emission signals. By superimposing direct quantitative characterization of oil wear particles and load-assisted judgment of current signals, a multi-dimensional complementary wear sensing network is constructed. This fundamentally solves the core defects of traditional single-vibration schemes, such as weak early wear signal energy, insufficient signal-to-noise ratio, and high false alarm rate under low-speed and heavy-load conditions. Cross-verification of multi-source information also effectively reduces false judgments caused by interference from single sensors. This scheme can provide early warning in the early stage of wear, effectively preventing wear from developing into malignant failures such as scuffing and tooth breakage. This invention establishes a multi-condition health benchmark library through offline operating condition clustering. It combines real-time operating condition matching, feature normalization correction, torque compensation, and transient order tracking to transform the original features that fluctuate with speed and load into operating condition-independent features that only reflect wear and degradation. This solves the pain point of traditional fixed threshold schemes that cannot distinguish between operating condition fluctuations and wear and degradation, and that cause frequent false alarms and missed alarms under variable frequency speed regulation and variable load operation. In scenarios with large-scale dynamic changes in speed and load, this scheme has a more stable diagnostic accuracy and a lower false alarm rate. It does not require separate calibration of thresholds for different operating conditions and has significantly improved adaptability to complex operating conditions. This invention adopts a three-layer progressive diagnostic architecture of early warning, accurate classification, and quantitative prediction. It achieves millisecond-level anomaly response through the early warning layer and ensures diagnostic accuracy through the classification and prediction layers, thus balancing the real-time performance of early warning and the precision of diagnosis. The deep residual shrinkage network with built-in adaptive soft threshold can automatically learn the noise reduction threshold according to the noise level, and can still effectively extract weak wear features in strong background noise environments. The multi-label classification output mode supports the decoupled identification of multiple failure modes such as uniform wear, pitting, and adhesion, which breaks through the limitation of traditional single-label models that cannot identify coupled faults, and is more in line with the actual scenario of multiple failures superimposed in industrial sites. Attached Figure Description

[0020] Figure 1 A flowchart of an overall method for identifying the meshing wear state of a worm gear reducer; Figure 2 A detailed flowchart of step S1 of a method for identifying meshing wear state of a worm gear reducer; Figure 3 A detailed flowchart of step S2 in a method for identifying meshing wear conditions of a worm gear reducer; Figure 4 A detailed flowchart of step S3 in a method for identifying meshing wear conditions of a worm gear reducer; Figure 5 This is an overall flowchart of a worm gear reducer meshing wear condition identification system. Detailed Implementation

[0021] Please see Figures 1-5 In this embodiment of the invention, a method for identifying the meshing wear state of a worm gear reducer includes the following steps: S1. Synchronous acquisition and preprocessing of multi-source heterogeneous signals: Build a multi-dimensional sensing system of acoustic vibration, oil, electrical and working conditions, complete high-precision synchronous acquisition and targeted noise reduction preprocessing of multi-physical quantity signals, and build a multi-source data foundation for wear identification; Step S1 specifically includes the following steps: S101. Multi-source sensing point arrangement: A high-frequency piezoelectric accelerometer and a broadband acoustic emission sensor are coaxially installed on the bearing housing end face of the worm input end, and a single-axis accelerometer is installed on the bearing housing of the worm output end. A three-phase current transformer, a speed and torque integrated sensor and an online oil sensor are also provided. S102, Hardware-level time synchronization: The PTP precise time protocol is used to achieve clock synchronization of all acquisition channels with a time synchronization error of 1ms; S103, Signal preprocessing: Wavelet threshold denoising, trend term removal and bandpass filtering are performed on vibration and acoustic emission signals, power frequency notch filtering and low-pass filtering are performed on current signals, and moving average filtering is performed on oil signals. It should be noted that this step adopts a sensor layout approach that combines high and low frequencies and uses multi-dimensional verification. The core purpose is to address the technical shortcomings of existing single vibration sensors, such as weak energy, low signal-to-noise ratio, and high false negative rate in early wear signals under low-speed, heavy-load conditions. Vibration and acoustic emission sensors are coaxially arranged at the worm input end, allowing simultaneous acquisition of mid-to-low frequency macroscopic meshing vibrations and high-frequency microscopic stress waves (100-400kHz) via the same transmission path. Vibration signals effectively reflect mid-to-late-stage macroscopic wear and impact on the tooth surface, while acoustic emission signals are far more sensitive than vibration to early damage such as microcrack initiation, micro-stripping, and pre-scuffing precursors. Together, they form a full-frequency coverage sensing system, significantly improving the detection capability of early wear. The vibration sensor at the worm gear output end helps distinguish whether wear occurs on the worm side or the worm gear side, enabling axial fault location. Current, speed, and torque sensors are used to obtain operating condition benchmarks, while the oil sensor directly quantifies the degree of wear from the perspective of material loss. This multi-source information verification effectively reduces misjudgments caused by interference from a single sensor. PTP sensors are used. Precise time protocols achieve microsecond-level synchronization to ensure strict alignment of multi-source signals on the time axis, avoiding phase deviations that could lead to feature fusion failure, especially ensuring the temporal correspondence of impact features. Differentiated preprocessing strategies are adopted for different signals to maximize the removal of various noise interferences while retaining effective wear features, thus providing a high-quality data foundation for subsequent feature extraction.

[0022] S2. Working condition perception and feature adaptive correction: Based on real-time speed and load data, working condition clustering and matching are performed, and the original features are normalized and corrected for working conditions and torque compensation is performed. In transient working conditions, order tracking processing is introduced to eliminate the interference of working condition fluctuations on wear characteristics. Step S2 specifically includes the following steps: S201. Offline working condition clustering and benchmark library construction: Using speed and load torque as two-dimensional working condition features, the improved K-means++ algorithm is used to divide 5 to 8 typical working condition clusters, and a corresponding health feature benchmark library is established for each working condition cluster; S202. Real-time working condition matching and determination: The Euclidean distance between the current working condition point and the center of each working condition cluster is calculated in real time, and the current working condition cluster is matched to obtain the current working condition cluster. S203. Normalization correction of characteristic working conditions: Normalize the original characteristics according to the formula xi=xi / i, where xi is the original characteristic value and i is the health mean value of the characteristic under the corresponding working condition. S204, Load Torque Compensation Correction: Introducing a torque compensation factor to perform secondary correction on the corrected characteristics, reducing the interference of load fluctuations on the characteristics; S205, Transient condition order normalization: When the condition is determined to be a start-stop or acceleration / deceleration transient condition, the time domain signal is converted into a stationary angular domain signal by calculating the order tracking algorithm. It should be noted that the core objective of this step is to eliminate the interference of speed and load fluctuations on wear characteristics, transforming the original characteristics that dynamically change with operating conditions into operating condition-independent characteristics that only reflect the degree of wear degradation. This addresses the core pain point of existing technologies with fixed threshold systems, which frequently produce false alarms and false negatives under varying operating conditions, through offline K-means++. Clustering typical operating condition clusters and establishing corresponding health benchmarks enables an adaptive mechanism that adapts to changes in operating conditions and benchmark changes, avoiding the distortion of judgments under non-rated operating conditions by a single steady-state benchmark. Feature normalization correction eliminates the overall rise and fall of feature amplitude caused by changes in operating conditions by comparing the original feature values ​​with the mean of the health benchmark, making feature values ​​under different operating conditions horizontally comparable. Torque compensation correction can further isolate the influence of load size on vibration and acoustic emission amplitude, further improving the sensitivity of features to wear. Order normalization processing under transient operating conditions can transform non-stationary time-domain signals during start-up, acceleration, and deceleration into stationary signals in the angular domain, eliminating the problems of spectrum broadening and feature ambiguity caused by speed fluctuations, ensuring the stability and accuracy of wear feature extraction during dynamic operation. The entire correction mechanism enables the equipment to maintain a stable diagnostic accuracy even under scenarios with large fluctuations in speed and load.

[0023] S3. Multi-domain sensitive feature extraction and multi-source feature-level fusion: Extract time-domain, frequency-domain, and time-frequency-domain multi-dimensional features of each sensing source respectively, obtain a wear-sensitive feature subset through feature selection algorithm, and generate a multi-dimensional fusion feature vector by using a weighted fusion algorithm that couples subjective and objective factors. Step S3 specifically includes the following steps: S301. Feature extraction of each sensing source: Extract 18-dimensional features in time domain, frequency domain, and time-frequency domain from vibration signals; extract 5-dimensional impact features from acoustic emission signals; extract 4-dimensional electrical features from current signals; and extract 3-dimensional abrasive features from oil signals. S302. Sensitive Feature Filtering: Using the maximum correlation and minimum redundancy algorithm, 10-12 core sensitive features are selected based on wear level to form the initial feature vector; S303, Subjective-Objective Coupled Weighted Fusion: The subjective weight of each sensor source is determined by the analytic hierarchy process (AHP), and the objective weight is calculated by the entropy weight method. The subjective and objective weights are coupled in a 6:4 ratio to generate a multi-dimensional fusion feature vector. It should be noted that this step employs a three-level processing approach—multi-domain feature extraction, redundant feature screening, and multi-source weighted fusion—to achieve complementary advantages of multi-source information and streamlined optimization of feature dimensions. This addresses the issues of one-sided single-source feature information and high redundancy of full-quantity features. Vibration features are extracted from three dimensions: time domain, frequency domain, and time-frequency domain. These features can characterize wear states from multiple perspectives, including amplitude intensity, spectral distribution, and energy migration, comprehensively covering signal features of different failure modes such as uniform wear, pitting, and adhesion. Impact features from acoustic emission focus on early microscopic damage, oil features directly quantify the total wear, and current features assist in distinguishing load disturbances. These multi-source features form a comprehensive signal... The algorithm employs a complementary approach, using a maximum correlation and minimum redundancy algorithm to select core sensitive features. This retains the information most strongly correlated with wear while removing redundant components between features, reducing the computational complexity of subsequent models and avoiding interference from redundant features with diagnostic results. A combined subjective and objective weighting method, integrating domain expert experience into subjective weights to ensure reasonable weight allocation for each sensor source at different wear stages, and adapting objective weights to the distribution characteristics of real-time data, enhances the adaptability of the fused features to different devices and operating conditions. The resulting fused feature vector exhibits higher robustness and discriminative power.

[0024] S4. Layered progressive wear status identification and fault decoupling: A three-layer progressive diagnostic architecture of early warning layer - accurate classification layer - quantitative prediction layer is adopted. By improving the deep residual shrinkage network, the wear type and degree classification of multi-label wear is realized. The wear quantification and remaining life prediction are completed by combining the attention-enhanced long short-term memory network. The aforementioned three-layer progressive diagnostic architecture includes: an early warning layer, a multi-label classification layer, and a quantitative prediction layer. The early warning layer uses acoustic emission impact count, vibration kurtosis, and abrasive hourly growth rate as input features, and adopts an adaptive thresholding method based on three criteria to achieve millisecond-level early warning. The warning threshold is dynamically updated according to the current operating condition benchmark. The multi-label classification layer uses fused feature vectors as input, and performs feature extraction and classification through an improved deep residual shrinkage network with a built-in soft thresholding module. It also outputs the wear degree level and wear failure type, and supports decoupling identification of coupled wear faults. The quantitative prediction layer adopts an attention-enhanced long short-term memory network, and combines the classification results, cumulative abrasive amount, and historical degradation trend to output the quantitative tooth surface wear depth, remaining service life, and confidence interval. Furthermore, the improved deep residual shrinking network in the multi-label classification layer uses a binary cross-entropy loss function for multi-label classification training. The loss function formula is as follows: Where N is the number of samples, and M is the number of label categories. For real labels, For the corresponding predicted probability; It should be noted that this step adopts a three-layer progressive diagnostic architecture. The core is to balance early warning response speed and diagnostic accuracy, resolving the contradiction that existing single diagnostic models cannot simultaneously satisfy rapid early warning and accurate fault classification. The early warning layer selects three of the most sensitive early features and employs a simple and efficient adaptive thresholding method, achieving millisecond-level response at the edge, capturing anomalies and triggering warnings immediately. This is particularly suitable for early warning of sudden faults such as glue adhesion. The accurate classification layer uses a deep residual shrinkage network with a built-in soft thresholding module, which can automatically learn the noise reduction threshold based on the noise level of the input signal. Even in environments with strong background noise, it can effectively extract weak wear features, overcoming the limitation of insufficient recognition accuracy in low signal-to-noise ratio scenarios using traditional algorithms. A multi-label classification mode is used instead of the traditional single-label classification because worm gear wear in industrial settings often presents a state of multiple coupled failure modes. Multi-label output can simultaneously identify the wear degree and failure type, supporting uniform wear + pitting corrosion + The decoupling identification of coupled faults such as adhesive bonding is more in line with actual working conditions. The quantization prediction layer introduces an attention mechanism to enhance the long short-term memory network, which can dynamically assign higher weights to recent degradation trends, more accurately fit the wear and degradation trajectory, and improve the prediction accuracy of wear amount and remaining service life. Among them, the binary cross-entropy loss function is used to train the multi-label classification network because the binary cross-entropy can independently calculate the classification loss for each label, which is suitable for the characteristic that there is no mutual exclusion relationship between multiple labels, ensuring the classification accuracy of each failure type and wear level. Compared with traditional multi-class cross-entropy, it is more suitable for the identification scenario of coupled faults.

[0025] S5. Offline-Online Closed-Loop Model Self-Update: Regularly integrate offline high-precision detection data to construct labeled samples, and dynamically update the diagnostic model and health benchmark through incremental transfer learning to correct model accuracy drift after long-term operation; In step S5, when the machine needs to be shut down for maintenance, wear true value labels are obtained through visual inspection of tooth surfaces and oil ferrography analysis. The online monitoring steady-state data of the corresponding time period are automatically matched according to the timestamp to generate labeled incremental training samples. The bottom feature extraction layer of the network is frozen, and the top classification and prediction layers are fine-tuned through incremental transfer learning. Combined with the domain adaptive algorithm, the data distribution difference across devices is reduced, the model is efficiently iterated, the health feature benchmarks and warning thresholds of each working condition cluster are corrected simultaneously, and the feature benchmark drift after long-term operation is corrected. It should be noted that this step constructs a closed-loop iterative mechanism for offline detection and online monitoring. The core of this mechanism is to solve the common industry problems of sensor drift, feature benchmark shift, and continuous decline in model accuracy after long-term operation of traditional online monitoring systems. By using tooth surface visual inspection and oil ferrography analysis data during downtime maintenance as ground truth labels, the detection accuracy is far higher than that of online sensor data, which can serve as high-quality labeled samples for model iteration. Incremental samples are generated by automatically matching online steady-state data according to timestamps, eliminating the need for additional test benches to collect data. This allows for full utilization of daily maintenance data to accumulate samples. An incremental transfer learning approach of freezing the bottom layer and fine-tuning the top layer is adopted, eliminating the need for full retraining of the model. The efficiency of a single update is improved by more than 80%, and remote silent updates can be completed during equipment operation intervals without downtime or production stoppage, thus not affecting normal production. The introduction of a domain-adaptive algorithm can reduce the differences in data distribution between different equipment and operating conditions, enabling rapid migration of the model to new equipment and reducing the sample requirements and calibration costs for new equipment deployment. Synchronous dynamic calibration of health characteristic benchmarks can correct benchmark offsets caused by slow-changing factors such as long-term sensor drift, changes in equipment foundation, and deterioration of lubricating oil, thereby avoiding false alarms and missed alarms after long-term operation and ensuring the diagnostic reliability of the system throughout its entire life cycle.

[0026] Another aspect of this invention provides a worm gear reducer meshing wear state identification system, comprising a multi-source synchronous sensing module, a working condition adaptive correction module, a multi-source feature fusion diagnosis module, a digital twin wear evolution module, an offline / online closed-loop iteration module, and a hierarchical early warning and operation and maintenance management module. The multi-source synchronous sensing module is used to synchronously collect reducer vibration, acoustic emission, current, oil, and working condition parameters, completing signal conditioning and analog-to-digital conversion. The working condition adaptive correction module is used to match the current working condition category, completing feature working condition normalization correction, torque compensation, and transient order normalization. The multi-source feature fusion diagnosis module is used to extract and filter multi-domain features, completing early warning, multi-label classification, and wear quantification prediction based on a hierarchical progressive diagnosis architecture. The digital twin wear evolution module is used to construct a digital twin of reducer wear degradation, simulate the wear evolution process, and output the remaining service life and maintenance window. The offline / online closed-loop iteration module is used to access offline high-precision detection data to generate labeled samples, dynamically updating the diagnostic model and health benchmark. The hierarchical early warning and operation and maintenance management module is used to trigger three-level hierarchical early warnings, manage the entire lifecycle health record, and generate operation and maintenance decision suggestions. It should be noted that this system adopts a four-level collaborative physical architecture of edge-aware local computing cloud-based iterative application interaction. The six functional modules correspond to the entire process on the method side, forming a complete business closed loop of data acquisition, state correction, intelligent diagnosis, evolution, simulation model iteration, operation and maintenance management. Among them, the multi-source synchronous perception module provides underlying data support for the system, the working condition adaptive correction module is responsible for eliminating working condition interference, the multi-source feature fusion diagnosis module is the core diagnosis execution unit, the digital twin wear evolution module realizes the virtual simulation of the wear process and life prediction, the offline and online closed-loop iteration module ensures the long-term operational accuracy of the system, and the hierarchical early warning and operation and maintenance management module realizes the practical application of diagnostic results. The modules interact with each other through standardized interfaces. The edge terminal is responsible for real-time acquisition, rapid early warning, and lightweight diagnosis, ensuring the real-time performance and availability of on-site monitoring even when the network is down. The cloud terminal is responsible for full data storage, model training iteration, and global device management, leveraging the advantages of computing power and data scale. The collaboration between the two not only meets the real-time requirements of industrial sites but also ensures the long-term intelligent iteration capability of the system, making it adaptable to the deployment needs of industrial sites of different scales.

[0027] The multi-source synchronous sensing module integrates a high-precision synchronous acquisition unit, which achieves multi-channel microsecond-level time synchronization through the PTP clock synchronization protocol to ensure the consistency of multi-source signal timing; the working condition adaptive correction module integrates an order tracking processing unit, which is used to convert time-domain vibration signals into angular-domain stationary signals under transient working conditions such as start-stop, acceleration and deceleration, and eliminate spectral interference from speed fluctuations; the multi-source feature fusion diagnosis module integrates a soft threshold noise reduction unit, which has an adaptive soft threshold module with a built-in deep residual shrinkage network, and can automatically learn the noise reduction threshold according to the signal noise level to extract weak wear features; The digital twin wear evolution module integrates a material wear characteristic library, with built-in wear rate curves and load mapping relationships for various worm gear materials, and corrects life prediction results based on load conditions; the offline and online closed-loop iteration module has built-in incremental learning and domain-adaptive toolchains, supporting rapid model fine-tuning and remote silent updates without downtime; the hierarchical early warning and operation and maintenance management module integrates a dual-criteria early warning verification unit, which integrates vibration-acoustic emission and oil fluid signals to jointly determine anomalies, reducing the risk of unplanned downtime caused by false alarms from a single sensor; It should be noted that the three types of built-in units correspond to the three core technical capabilities of the system, providing hardware and algorithmic support for achieving highly reliable diagnosis. The high-precision synchronous acquisition unit achieves multi-channel microsecond-level time synchronization through the PTP protocol, ensuring the time reference for multi-source feature fusion and guaranteeing strict timing alignment of signals from different physical quantities such as vibration, acoustic emission, and current, avoiding distortion of fused features due to phase deviation. The order tracking processing unit is the core algorithm unit for variable operating conditions. For non-steady-state processes such as start-stop, acceleration, and deceleration, it transforms non-stationary signals into stationary signals through angular domain resampling, solving the problem of feature ambiguity and judgment failure in traditional spectrum analysis under speed fluctuations, and broadening the system's applicable operating conditions. The soft threshold noise reduction unit is the core for extracting weak fault features in strong noise environments. Unlike traditional fixed threshold noise reduction, its adaptive learning threshold characteristic can adapt to different noise levels in different environments. It can effectively suppress background noise and retain weak wear features without manual parameter adjustment, greatly improving the system's adaptability to complex industrial environments, especially enhancing the detection capability of early micro-wear under low-speed conditions.

[0028] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying the meshing wear state of a worm gear reducer, characterized in that, The method includes the following steps: S1. Synchronous acquisition and preprocessing of multi-source heterogeneous signals: Build a multi-dimensional sensing system of acoustic vibration, oil, electrical and working conditions, complete high-precision synchronous acquisition and targeted noise reduction preprocessing of multi-physical quantity signals, and build a multi-source data foundation for wear identification; S2. Working condition perception and feature adaptive correction: Based on real-time speed and load data, working condition clustering and matching are performed, and the original features are normalized and corrected for working conditions and torque compensation is performed. In transient working conditions, order tracking processing is introduced to eliminate the interference of working condition fluctuations on wear characteristics. S3. Multi-domain sensitive feature extraction and multi-source feature-level fusion: Extract time-domain, frequency-domain, and time-frequency-domain multi-dimensional features of each sensing source respectively, obtain a wear-sensitive feature subset through feature selection algorithm, and generate a multi-dimensional fusion feature vector by using a weighted fusion algorithm that couples subjective and objective factors. S4. Layered progressive wear status identification and fault decoupling: A three-layer progressive diagnostic architecture of early warning layer - accurate classification layer - quantitative prediction layer is adopted. By improving the deep residual shrinkage network, the wear type and degree classification of multi-label wear is realized. The wear quantification and remaining life prediction are completed by combining the attention-enhanced long short-term memory network. S5. Offline-Online Closed-Loop Model Self-Update: Regularly integrate offline high-precision detection data to construct labeled samples, and dynamically update the diagnostic model and health benchmark through incremental transfer learning to correct model accuracy drift after long-term operation.

2. The method for identifying the meshing wear state of a worm gear reducer according to claim 1, characterized in that, Step S1 specifically includes the following steps: S101. Multi-source sensing point arrangement: A high-frequency piezoelectric accelerometer and a broadband acoustic emission sensor are coaxially installed on the bearing housing end face of the worm input end, and a single-axis accelerometer is installed on the bearing housing of the worm output end. A three-phase current transformer, a speed and torque integrated sensor and an online oil sensor are also provided. S102, Hardware-level time synchronization: The PTP precise time protocol is used to achieve clock synchronization of all acquisition channels with a time synchronization error of 1ms; S103. Preprocessing of branch signals: Perform wavelet threshold noise reduction, trend term removal and bandpass filtering on vibration and acoustic emission signals, perform power frequency notch filtering and low-pass filtering on current signals, and perform moving average filtering on oil signals.

3. The method for identifying the meshing wear state of a worm gear reducer according to claim 1, characterized in that, Step S2 specifically includes the following steps: S201. Offline working condition clustering and benchmark library construction: Using speed and load torque as two-dimensional working condition features, the improved K-means++ algorithm is used to divide 5 to 8 typical working condition clusters, and a corresponding health feature benchmark library is established for each working condition cluster; S202. Real-time working condition matching and determination: The Euclidean distance between the current working condition point and the center of each working condition cluster is calculated in real time, and the current working condition cluster is matched to obtain the current working condition cluster. S203. Normalization correction of characteristic working conditions: Normalize the original characteristics according to the formula xi=xi / i, where xi is the original characteristic value and i is the health mean value of the characteristic under the corresponding working condition. S204, Load Torque Compensation Correction: Introducing a torque compensation factor to perform secondary correction on the corrected characteristics, reducing the interference of load fluctuations on the characteristics; S205, Transient Condition Order Normalization: When the condition is determined to be a start-stop or acceleration / deceleration transient condition, the time-domain signal is converted into a stationary angular-domain signal by calculating the order tracking algorithm.

4. The method for identifying the meshing wear state of a worm gear reducer according to claim 1, characterized in that, Step S3 specifically includes the following steps: S301. Feature extraction of each sensing source: Extract 18-dimensional features in time domain, frequency domain, and time-frequency domain from vibration signals; extract 5-dimensional impact features from acoustic emission signals; extract 4-dimensional electrical features from current signals; and extract 3-dimensional abrasive features from oil signals. S302. Sensitive Feature Filtering: Using the maximum correlation and minimum redundancy algorithm, 10-12 core sensitive features are selected based on wear level to form an initial feature vector. S303, Subjective-Objective Coupled Weighted Fusion: The subjective weight of each sensor source is determined by the analytic hierarchy process (AHP), and the objective weight is calculated by the entropy weight method. The subjective and objective weights are coupled in a 6:4 ratio to generate a multi-dimensional fusion feature vector.

5. The method for identifying the meshing wear state of a worm gear reducer according to claim 1, characterized in that, The three-layer progressive diagnostic architecture in step S4 includes an early warning layer, a multi-label classification layer, and a quantization prediction layer. The early warning layer uses acoustic emission impact count, vibration kurtosis, and abrasive hourly growth rate as input features, and adopts an adaptive thresholding method based on three criteria to achieve millisecond-level early warning. The warning threshold is dynamically updated according to the current operating condition benchmark. The multi-label classification layer uses fused feature vectors as input, and performs feature extraction and classification through an improved deep residual shrinkage network with a built-in soft thresholding module. It also outputs the wear degree level and wear failure type, supporting the decoupled identification of coupled wear faults. The quantization prediction layer uses an attention-enhanced long short-term memory network, and combines the classification results, cumulative abrasive amount, and historical degradation trend to output the quantified tooth surface wear depth, remaining service life, and confidence interval.

6. A system for identifying the meshing wear state of a worm gear reducer according to any one of claims 1-5, characterized in that, The system includes a multi-source synchronous sensing module, an operating condition adaptive correction module, a multi-source feature fusion diagnosis module, a digital twin wear evolution module, an offline and online closed-loop iteration module, and a hierarchical early warning and operation and maintenance management module. The multi-source synchronous sensing module is used to synchronously collect the vibration, acoustic emission, current, oil and operating parameters of the reducer, and complete signal conditioning and analog-to-digital conversion. The adaptive correction module is used to match the current operating condition category and complete the characteristic operating condition normalization correction, torque compensation and transient order normalization. The multi-source feature fusion diagnostic module is used to extract multi-domain features and filter and fuse them, and completes early warning, multi-label classification and wear quantification prediction based on a hierarchical and progressive diagnostic architecture. The digital twin wear evolution module is used to construct a digital twin of the wear degradation of the reducer, simulate the wear evolution process, and output the remaining service life and maintenance window period; The offline / online closed-loop iterative module is used to access offline high-precision detection data to generate labeled samples and dynamically update the diagnostic model and health benchmark. The tiered early warning and operation and maintenance management module is used to trigger three-level tiered early warnings, manage the full lifecycle health records, and generate operation and maintenance decision suggestions.

7. The method for identifying the meshing wear state of a worm gear reducer according to claim 6, characterized in that, The multi-source synchronous sensing module integrates a high-precision synchronous acquisition unit, which achieves multi-channel microsecond-level time synchronization through the PTP clock synchronization protocol; The adaptive correction module for operating conditions integrates an order tracking processing unit, which is used to convert time-domain vibration signals into angular-domain stationary signals under transient operating conditions such as start-stop, acceleration and deceleration. The multi-source feature fusion diagnostic module integrates a soft threshold denoising unit and a built-in adaptive soft threshold module with a deep residual shrinkage network.

8. The method for identifying the meshing wear state of a worm gear reducer according to claim 6, characterized in that, The digital twin wear evolution module integrates a material wear characteristic library, has built-in wear rate curves and load mapping relationships for various worm gear materials, and corrects the life prediction results by combining load conditions. The offline / online closed-loop iterative module incorporates an incremental learning and domain-adaptive toolchain. The graded early warning and operation and maintenance management module integrates a dual-criteria early warning verification unit, which combines vibration-acoustic emission and oil fluid sensor signals to jointly determine abnormalities.

Images