Multi-physical quantity analysis model fusion driven roller digital twin fault judgment method

Abstract

The application relates to a multiple-physical-quantity analysis model fusion driven roller digital twin fault judgment method in the field of intelligent mine equipment fault monitoring and digital twin technology, which comprises the following steps: underground roller group multi-source physical quantity distributed sensing deployment and original data acquisition; multi-source sensing data space-time two-dimensional accurate registration preprocessing; self-adaptive fault mechanism decoupling modeling and interference stripping according to physical quantities and working conditions; fault confidence driven multi-model dynamic weight self-adaptive fusion; virtual-real mapping construction of a digital twin and initialization of an initial state; physical-data-twin closed loop iterative optimization and model parameter self-correction; roller hidden fault prediction and composite fault fine disassembly and tracing; fault judgment result storage and working condition database iterative update. The application solves the problems of single monitoring dimension and poor anti-interference ability of traditional single physical quantity, overcomes the problems of space-time misplacement of multi-source monitoring data and low fusion efficiency, and solves the problem of lack of a closed loop iterative optimization system in the prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent mining equipment fault monitoring and digital twin technology, specifically to a digital twin fault determination method for idler rollers driven by the fusion of multi-physical quantity analytical models in coal mines. This invention focuses on the intelligent monitoring scenario of idler roller operation status in underground belt conveyors in coal mines, integrating multi-source sensor physical quantities such as temperature, pressure, vibration, and acoustics. It combines analytical modeling, multi-model fusion algorithms, and digital twin iterative technology, belonging to the cross-disciplinary fields of intelligent mining equipment fault diagnosis, industrial digital twins, and multi-source data fusion. It can be applied to the full life-cycle health management and accurate fault determination of idler roller groups under complex working conditions in coal mines. Background Technology

[0002] Belt conveyors are core equipment for material transportation in underground coal mines, and idlers, as key supporting components, directly determine the safety and continuity of the coal mine transportation system through their stable operation. The complex and harsh working conditions underground—high dust, high humidity, strong interference, and variable loads—easily lead to idler bearing wear, jamming, abnormal noise, overheating, and other faults. If these faults cannot be accurately and promptly identified, they can easily cause secondary accidents such as conveyor belt wear, misalignment, and tearing, resulting in underground production shutdowns and even safety hazards. Currently, most coal mine idler fault monitoring technologies adopt a single physical quantity diagnostic mode, with mainstream solutions relying on single sensor data such as vibration or temperature for fault identification. Single physical quantity monitoring has a limited dimension and is easily affected by external factors such as underground equipment vibration, ambient temperature and humidity, and dust interference, resulting in low signal-to-noise ratios and serious problems of missed and false diagnoses. Furthermore, existing monitoring technologies often use fixed threshold judgment logic, failing to dynamically adjust parameters based on the actual operating conditions of the idlers, lacking physical mechanism model support and a real-time iterative optimization mechanism using digital twins, thus failing to achieve a refined and comprehensive characterization of the idler's operating status. Furthermore, existing technologies have not established a spatiotemporal alignment and dynamic weight fusion system for multi-physical quantity data. The independent analysis and fragmentation of each monitoring data make it difficult to achieve complementary verification of multi-source information. This makes it unsuitable for the large-scale, long-cycle, and highly reliable intelligent monitoring needs of underground idler groups, which seriously restricts the intelligent and precise upgrading of underground equipment operation and maintenance in intelligent mines. Summary of the Invention

[0003] This invention aims to address the core technical deficiencies of existing underground idler roller fault monitoring technologies in coal mines and specifically tackle the industry's technical pain points.

[0004] First, it addresses the limitations of traditional single-physical-quantity monitoring, which suffers from limited dimensionality and poor anti-interference capabilities. Existing technologies rely solely on vibration or temperature as a single parameter for fault diagnosis, failing to adapt to complex downhole operating conditions. They are prone to signal distortion due to environmental noise and fluctuations in operating conditions, leading to high rates of missed and false diagnoses, and making it difficult to accurately identify minor, latent, and complex faults. Second, it addresses the technical issues of fixed parameters and poor adaptability in existing monitoring models. Traditional threshold-based and static diagnostic models cannot dynamically iterate according to changes in roller operation and downhole conditions, lacking both physical mechanism and data-driven support, resulting in insufficient model accuracy and generalization ability.

[0005] Secondly, we need to overcome the problems of spatiotemporal misalignment and low fusion efficiency of multi-source monitoring data. Existing technologies have not achieved spatiotemporal alignment processing of multiple physical quantity data such as temperature, pressure, vibration, and acoustics, and there is no dynamic weight fusion mechanism. Data from different dimensions cannot complement each other for verification and collaborative empowerment, and fragmented data is difficult to support full-dimensional fault judgment.

[0006] Finally, the problem of the lack of a closed-loop iterative optimization system in existing technologies needs to be addressed. Traditional monitoring modes lack a closed-loop linkage mechanism of physical models, data-driven and twin iterations, and cannot use real-time sensor data to correct model parameters. The twin simulation accuracy is low, making it difficult to achieve intelligent monitoring and full life cycle health management of the roller group in a full state, with fine detail and high reliability. This fails to meet the core needs of intelligent operation and maintenance of underground equipment in smart mines.

[0007] To address the aforementioned problems, this invention provides a method for determining the fault of a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models, comprising the following steps: S1 Distributed Sensing Deployment and Raw Data Acquisition of Multi-Source Physical Quantities for Underground Idler Groups: Based on the distribution layout, operating conditions, and environmental characteristics of underground belt conveyor idler groups in coal mines, a distributed fixed-point installation and deployment of multiple types of sensors was completed. These sensors respectively collect raw data on four core physical quantities of idler operation: temperature, pressure, vibration, and acoustics. Simultaneously, underground environmental condition data are collected to construct a comprehensive and multi-dimensional raw sensing data sample set. S2 Multi-Source Sensor Data Spatiotemporal Dual-Dimensional Precise Registration Preprocessing: In the time dimension, the sampling timing benchmark of each sensor is unified, and the data transmission delay and sampling phase deviation are calibrated; in the spatial dimension, the original physical quantity data of the idler roller is combined with the actual installation coordinates of the sensor to establish a spatial mapping relationship and correct the spatial offset error of the sensor data at different points. S3 Physical Quantity Adaptive Fault Mechanism Decoupling Modeling and Interference Removal: Independent fault analysis mechanism models for idler rollers are constructed to address the differences in fault mechanisms and environmental interference characteristics, achieving differentiated interference decoupling; irrelevant interference from underground equipment resonance and mechanical vibrations is removed from vibration data; background noise from roadway airflow disturbances and dust friction is filtered from acoustic data; baseline offset errors caused by belt load variations and environmental temperature and humidity fluctuations are eliminated from temperature and pressure data. S4 Fault Confidence-Driven Multi-Model Dynamic Weight Adaptive Fusion: Based on the preliminary fault diagnosis results of each idler roller fault analysis mechanism model, a dual judgment mechanism of working condition and fault confidence is introduced to identify the working condition parameters of the current underground environment in real time, dynamically calculate the effective confidence of each physical quantity diagnosis result, form a working condition confidence evaluation model based on fuzzy membership, construct a fuzzy evaluation set of interference working conditions, adaptively adjust the fusion weight, complete the adaptive weighted fusion of multi-dimensional fault diagnosis results, and output the preliminary comprehensive fault judgment result of the idler roller; S5 Digital Twin Virtual-Real Mapping Construction and Initial State Initialization: Based on the original physical quantity data of the idler rollers, the operating process parameters and the aforementioned pre-processed standard operating data, a digital twin simulation model of the idler roller group in the coal mine is built to complete the one-to-one virtual-real mapping between the physical entity and the virtual twin. S6 Physical-Data-Twin Closed-Loop Iterative Optimization and Model Parameter Self-Correction: Construct a closed-loop iterative optimization system, utilize real-time collected roller operation sensor data from underground and dynamically fused fault diagnosis results to iteratively correct the core parameters of each physical quantity fault analysis model, and synchronously update the simulation working condition parameters, fault judgment thresholds and state mapping rules of the digital twin. S7 Idler Roller Latent Fault Prediction and Complex Fault Refined Disassembly and Source Tracing: Based on the iteratively optimized high-precision twin judgment model and multi-model fusion results, a full-dimensional refined fault analysis is carried out, weak features of latent faults of idlers are extracted and predicted, complex faults are fuzzily disassembled and graded, and finally, a full-state monitoring report of idler group, graded early warning information and operation and maintenance suggestions are output, and the probability of fault source tracing is assessed. S8 Fault Judgment Result Storage and Working Condition Database Iterative Update: Each working condition data, model weight parameters, fault judgment result, and fault source information are completely stored in the underground idler roller intelligent monitoring database to continuously accumulate complex working condition samples and fault samples, and optimize the dynamic weight allocation logic and fault analysis model accuracy.

[0008] To achieve quantitative standardization of the acquisition process, in step S1, the multi-source physical quantity time-series sampling matrix model is constructed as follows: In the formula: —Multi-source physical quantity timing acquisition matrix; —Temperature time-series acquisition vector, The temperature value of the idler roller collected by the i-th temperature sensor at time t; —Pressure timing acquisition vector, The value of the idler support load pressure collected by the i-th pressure sensor at time t; —Vibration timing acquisition vector, Let t be the vibration acceleration value of the roller body collected by the i-th vibration sensor at time t; —Acoustic timing acquisition vector, Let be the roller running noise value collected by the i-th acoustic sensor at time t; — Collect time-series variables.

[0009] To achieve accurate spatiotemporal registration of heterogeneous multi-source data (temperature, pressure, vibration, acoustics), ensuring data consistency and effectiveness in subsequent data fusion and model analysis, and avoiding diagnostic failures caused by spatiotemporal misalignment, step S2 involves temporal calibration based on cubic spline interpolation in the time dimension. This involves using a cubic spline interpolation algorithm to fill in missing values ​​in the temporal data, correct phase shifts, and construct a continuous and smooth temporal data curve. The core constraint equation for interpolation is as follows: In the formula: —The interpolation fitting function for the i-th sampling interval; —The interpolation coefficients for the i-th interval are the coefficients of the function offset term, the first term, the second term, and the third term, respectively; —Adjacent sampling times; —The first and second derivatives of a function characterize the rate of change and acceleration of a physical quantity, respectively; In the spatial dimension, spatial mapping registration based on coordinate transformation is performed: a mapping relationship is established between the local coordinate system of the idler roller and the global monitoring coordinate system downhole. The spatial offset error of the sensor is corrected through a rigid body transformation matrix. The spatial registration formula is as follows: In the formula: —The sensor's global spatial coordinate vector; —Local installation coordinate vector of the sensor roller; —Three-dimensional rotation transformation matrix; — Spatial translation offset vector.

[0010] To avoid interference and superposition issues during multi-source data fusion and improve the accuracy and effectiveness of single-dimensional fault features, a resonance interference stripping model based on time-domain differential filtering is proposed for vibration data. This model utilizes first- and second-order differentials to capture transient vibration fault features and strips out steady-state equipment resonance interference. The fault feature purification formula is then developed. In the formula: —Characteristic quantities of fault vibration after purification; —The original vibration timing signal; —Resonance interference suppression coefficient; For acoustic signals, a background noise filtering model based on Gaussian probability distribution is proposed: Downhole background noise follows a Gaussian normal distribution; effective fault acoustic features are selected through a probability density function to filter out random environmental noise. In the formula: —Acoustic noise probability density value; —Background noise mean and standard deviation; —Refined acoustic characteristics of the fault; —Indicator function; —Noise threshold probability; For temperature and pressure signals, a baseline drift integral correction model is used: to correct steady-state deviations caused by variable loads and temperature and humidity fluctuations through integral calculations. In the formula: —Corrected true temperature and pressure fault characteristics; —Integral time series variable.

[0011] To replace the traditional fixed-weight fusion mode and achieve optimal fusion of multi-dimensional fault features, thereby significantly improving the accuracy of fault determination under complex working conditions, in step S4, the working condition confidence assessment model... In the formula: —Confidence level of the i-th type of physical quantity; —Real-time operating condition interference parameters; —Standard, non-interference operating parameters; — Operating condition interference tolerance threshold.

[0012] To achieve visualized and digital replication of the idler roller's operating status, and to provide a simulation platform and benchmark support for subsequent fault iteration correction and status comparison analysis, step S5 utilizes a virtual-real mapping matrix of a twin structure. In the formula: —The virtual twin's structural parameter matrix, including size, orientation, and position parameters; —Physical idler roller structure parameter matrix; —The virtual-real mapping correction coefficient matrix; Initialization of initial state parameters based on numerical fitting: The standard operating state threshold of the idler roller is solved by the least squares fitting algorithm to complete the initialization of the twin parameters. In the formula: —Parameters from historical normal operation samples; —Initialize the fitting parameters for the twin.

[0013] To address issues such as equipment aging and gradual changes in operating conditions caused by long-term operation of idler rollers, and to continuously perform dynamic calibration between the twin and the physical entity, thus resolving the problems of fixed parameters and accuracy decay in traditional static models, and enabling the model's judgment capability to continuously and adaptively optimize according to the equipment's operating status, step S6 involves dynamic iterative correction of model parameters: based on the gradient descent iteration criterion, the core parameters of the fault analysis model are updated in real time. In the formula: —The core parameters of the model in the kth and k+1th iterations; —Iterative learning step size; —The diagnostic error function for the k-th iteration; —Error gradient; Twin state virtual-real deviation correction: Based on real-time data statistical deviation, the virtual-real simulation state is corrected to achieve virtual-real synchronization. In the formula: —Mean square error between real and virtual states; —Number of samples collected; —Physical entity and twin state parameters; — Deviation compensation coefficient; — Corrected twin state parameters.

[0014] To achieve refined health management throughout the entire lifecycle of idler rollers, step S7 involves the prediction of subtle latent fault features: subtle fault features are extracted from time-series data using high-order differentiation to achieve early fault prediction. When the third-order derivative value exceeds the normal threshold range, an early latent fault is identified. It is the third-order characteristic gradient; Fuzzy Decomposition and Classification of Composite Faults: Constructing a fuzzy fault evaluation matrix to achieve decomposition and classification of composite fault types. In the formula: —Fault evaluation result vector; — Fault feature weight vector; —Fault feature fuzzy relation matrix; —Fuzzy synthesis operator; Fault Origin Probability Assessment: Based on Bayesian probability formulas, the posterior probability of various fault causes is calculated to achieve accurate source tracing. In the formula: —The posterior probability of the i-th type of fault under observed feature X; —The prior probability of the feature corresponding to the i-th type of fault; —The historical prior probability of the i-th type of fault.

[0015] To establish a self-growing closed-loop system encompassing "data acquisition - preprocessing modeling - fusion judgment - iterative optimization - sample accumulation," and to continuously improve the accuracy and adaptability of idler roller fault determination under complex downhole extreme conditions, step S8 involves sample data statistical optimization: by correcting the sample mean and variance, the sample distribution characteristics of the database are optimized. In the formula: —Updated original sample mean; —Updated original sample variance; —Sample forgetting coefficient; —Newly added measured sample data.

[0016] In summary, compared with the prior art, the present invention has the following technical advantages and beneficial effects: (1) This invention uses a spatiotemporal dual-dimensional registration model to solve the industry pain point of spatiotemporal mismatch of distributed heterogeneous sensors, and provides a consistent and standardized data base for subsequent mechanism decoupling and multi-model fusion, which is a prerequisite for high-precision fault judgment. (2) This invention introduces a gradient iteration parameter correction + virtual-real deviation compensation model. The "perception-diagnosis-twin update" closed-loop iteration system solves the industry problem that the accuracy of the traditional static twin model continues to decline with the aging of equipment, and ensures that the fault judgment of the idler roller is stable and reliable throughout its entire life cycle. (3) After adopting the forgetting coefficient dynamic sample update formula of the present invention, as batches accumulate, the sample mean quickly converges and stabilizes, the variance continues to decrease, and the sample distribution approaches the real downhole fault distribution. The existing technology uses full sample storage, with equal weights for new and old samples, and old failed samples continue to interfere with the model. The dynamic statistical update mechanism of the present invention balances the weights of new and old samples, constructs a self-growing fault database, and continuously improves the extremely rare faults. (4) Traditional schemes are affected by environmental resonance, airflow noise and load baseline drift, which submerge the effective fault features and concentrate the feature amplitude in the noise range. The fault and normal samples highly overlap. After differential filtering, Gaussian noise filtering and integral baseline correction in this invention, the fault feature amplitude is significantly separated and the distribution range of normal samples and fault samples is almost non-overlapping. The multi-physical quantity differential decoupling model can remove environmental noise and purify the real fault features, solving the pain point of low feature recognition and high misjudgment rate caused by interference superposition in the existing technology. (5) The sample database iterative update + gradient parameter self-correction closed loop system of the present invention has a false alarm rate that decreases slightly in the early stage and then remains stable at less than 3% in the long term, with no trend of accuracy decay. It reflects the self-growing working condition database and the whole process closed loop system, realizes stable low-misjudgment operation of equipment throughout the entire life cycle, and solves the operation and maintenance pain points of long-term model failure and frequent manual calibration in the existing technology. Attached Figure Description

[0017] Figure 1 This is a flowchart of the fault determination method for idler roller digital twin driven by the fusion of multi-physical quantity analytical models of the present invention; Figure 2 This is a comparison diagram of the original timing misalignment of existing multi-source sensors and the timing after spatiotemporal registration of this invention; Figure 3 This is a comparison chart of the virtual-real error decay curves of the existing static twin model and the closed-loop iterative twin model of this invention; Figure 4 This is a convergence curve of the sample mean / variance of the database sample iterative update model of the present invention; Figure 5 This is a histogram of fault characteristic amplitude distribution before and after interference decoupling according to the present invention; Figure 6 This is a comparison chart of the false alarm rate curves of the existing technology and the full life cycle model of this invention as a function of runtime. Detailed Implementation

[0018] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0019] like Figure 1 The method for determining faults in a digital twin of idler rollers driven by the fusion of multiple physical quantity analytical models includes the following steps: S1 Distributed sensing deployment and raw data acquisition of multi-source physical quantities of the underground idler roller group; S2 Precise registration and preprocessing of multi-source sensor data in both spatiotemporal dimensions; S3 Decoupling modeling and interference removal of fault mechanisms based on the working conditions of the physical quantities; S4 Adaptive fusion of dynamic weights of multiple models driven by fault confidence; S5 Construction of virtual-real mapping of the digital twin and initial state initialization; S6 Iterative optimization of physical-data-twin closed loop and self-correction of model parameters; S7 Prediction of latent faults of idler rollers and refined decomposition and source tracing of complex faults; S8 Storage of fault determination results and iterative update of the working condition database.

[0020] The following is a detailed explanation of each step.

[0021] S1 Downhole Idler Roller Group Multi-Source Physical Quantity Distributed Sensing Deployment and Raw Data Acquisition The evolution of idler roller failure is a dynamic time-series process involving the coupling of multiple physical quantities. Accurate and comprehensive raw sensor data is the foundation for subsequent mechanism modeling, data fusion, and fault diagnosis. A distributed sensor array is used to achieve full coverage acquisition of multi-dimensional physical quantities. A time-series sampling quantization model is used to standardize the acquisition of operational and environmental condition data, providing standardized and quantifiable data samples to support subsequent multi-source data preprocessing and modeling analysis.

[0022] Based on the distribution, operating conditions, and environmental characteristics of belt conveyor idler groups in coal mines, a distributed, fixed-point installation of multiple types of sensors was completed. These sensors collect raw data on four core physical quantities related to idler operation: temperature, pressure, vibration, and acoustics. The sensors cover key fault-prone locations such as idler bearings, roller bodies, and support bases, continuously collecting data on all operating conditions, including idler start-up, stop, constant speed operation, and variable load operation. Simultaneously, environmental data such as underground noise, temperature, humidity, and belt load are also collected, constructing a comprehensive, multi-dimensional raw sensing data sample set.

[0023] To achieve quantitative standardization of the acquisition process, a multi-source physical quantity time-series sampling matrix model (linear algebra) is constructed to realize structured storage and time-series calibration of all sensed data. In the formula: —Multi-source physical quantity time-series acquisition matrix, dimensionless, is a four-dimensional column vector matrix, representing the full-dimensional operating state of the idler roller at time t; —Temperature time-series acquisition vector, unit: °C The temperature value of the idler roller collected by the i-th temperature sensor at time t; —Pressure time-series acquisition vector, unit: Pa, The value of the idler support load pressure collected by the i-th pressure sensor at time t; —Vibration time sequence acquisition vector, unit: m / s 2 , Let t be the vibration acceleration value of the roller body collected by the i-th vibration sensor at time t; —Acoustic timing acquisition vector, unit: dB Let be the roller running noise value collected by the i-th acoustic sensor at time t; — Data collection time series variable, unit: s, representing the real-time moment of data collection.

[0024] This method can unify the temporal format of various types of heterogeneous sensor data, realize the structured regularization of massive unordered collected data, and provide a standardized data base for subsequent spatiotemporal registration, noise reduction and decoupling modeling.

[0025] S2 Multi-Source Sensor Data Spatiotemporal Dual-Dimensional Pre-registration Preprocessing In underground distributed monitoring scenarios, inherent deviations in the sampling frequency, transmission delay, and installation location of different sensors can lead to spatiotemporal misalignment and feature mismatch in multi-source data. By relying on numerical analysis interpolation algorithms, linear algebraic spatial mapping models, and temporal error correction algorithms, precise spatiotemporal registration of data is achieved, eliminating inherent biases in the acquisition system, ensuring the spatiotemporal consistency of multi-source data, and providing a precise data foundation for subsequent multi-model fusion.

[0026] To address the issues of inconsistent sensor sampling frequencies, transmission delay differences, and installation misalignment in downhole distributed monitoring, a multi-source data spatiotemporal alignment preprocessing was conducted. In the temporal dimension, the sampling timing references of each sensor were unified, and data transmission delay and sampling phase deviations were calibrated to eliminate timing discrepancies. In the spatial dimension, a spatial mapping relationship was established by combining the original physical quantity data (structural parameters) of the idler roller with the actual installation coordinates of the sensors. This corrected the spatial offset errors of sensor data from different locations, achieving precise spatiotemporal registration of heterogeneous multi-source data (temperature, pressure, vibration, acoustics), ensuring data consistency and effectiveness for subsequent data fusion and model analysis, and avoiding diagnostic failures caused by spatiotemporal misalignment.

[0027] In the time dimension, time series calibration based on cubic spline interpolation: To address the issues of uneven sensor sampling frequency and transmission delay deviation, a cubic spline interpolation algorithm is employed to fill in missing values ​​in the time-series data and correct phase shifts, thereby constructing a continuous and smooth time-series data curve. The core constraint equation for interpolation is as follows: In the formula: —The interpolation fitting function for the i-th sampling interval, with dimensions consistent with the corresponding physical quantity; —The interpolation coefficients for the i-th interval are the coefficients of the function offset term, the first term, the second term, and the third term, respectively, and are dimensionless; — Adjacent sampling times, unit: seconds; —The first and second derivatives of a function characterize the rate of change and acceleration of a physical quantity, respectively, with dimensions of the corresponding physical quantity / s and the corresponding physical quantity / s. 2 .

[0028] In the spatial dimension, spatial mapping registration based on coordinate transformation: Establish the mapping relationship between the local coordinate system of the idler roller and the global monitoring coordinate system downhole. Correct the spatial offset error of the sensor through a rigid body transformation matrix. Spatial registration formula: In the formula: — Sensor global spatial coordinate vector, unit: m; —Local installation coordinate vector of the sensor roller, unit: m; —A three-dimensional rotation transformation matrix, a 3×3 square matrix, dimensionless, representing spatial angle offset correction; — Spatial translation offset vector, unit: m, representing the amount of compensation for installation position deviation.

[0029] Through the above two-dimensional mathematical model, the problem of spatiotemporal mismatch of multi-source data is completely eliminated, and a standardized monitoring dataset with time synchronization and spatial alignment is output.

[0030] S3 Physical Quantity-Based Adaptive Fault Mechanism Decoupling Modeling and Interference Stripping In complex downhole conditions, multiple physical quantities exhibit environmental interference coupling characteristics. Traditional direct data fusion methods can lead to interference superposition and feature distortion. This method combines dynamic differential modeling and probabilistic statistical noise distribution models to construct a decoupling and noise reduction model targeting the differentiated interference mechanisms of four types of physical quantities. This model removes environmental clutter, retains true fault characteristics, and achieves accurate purification of fault features for each physical quantity.

[0031] To address the differences in fault mechanisms and environmental interference characteristics among the four types of physical quantities, independent fault analysis mechanism models for idler rollers were constructed to achieve differentiated interference decoupling. Irrelevant interference from downhole equipment resonance and mechanical stray vibrations was removed from vibration data; background noise from roadway airflow disturbances and dust friction was filtered out from acoustic data; and baseline offset errors caused by belt load variations and environmental temperature and humidity fluctuations were eliminated from temperature and pressure data. Through mechanism-driven noise reduction and decoupling processing, the true fault characteristics of each physical quantity were preserved, avoiding the interference superposition problem during multi-source data fusion and improving the accuracy and effectiveness of single-dimensional fault features.

[0032] For vibration signals, a resonance interference stripping model based on time-domain differential filtering is proposed: Using first- and second-order differentials to capture transient vibration fault characteristics and eliminate steady-state equipment resonance interference, the fault characteristic purification formula is as follows: In the formula: —Characteristic quantities of fault vibration after purification, unit: m / s 3 ; —Original vibration time sequence signal, unit: m / s 2 ; — Resonance interference suppression coefficient, dimensionless, with a value range of [0,1], which is adaptively adjusted according to the resonance intensity of the downhole equipment.

[0033] For acoustic signals, a background noise filtering model based on Gaussian probability distribution is proposed: Downhole background noise follows a Gaussian normal distribution. Valid fault acoustic features are selected using a probability density function to filter out random environmental noise. In the formula: —Acoustic noise probability density value, dimensionless; —Background noise mean and standard deviation, in dB; —Refined acoustic characteristics of the fault, unit: dB; —Indicator function, dimensionless, takes the value 1 if the condition is met, otherwise takes the value 0; —Noise threshold probability, dimensionless.

[0034] For temperature and pressure signals, a baseline drift integral correction model is used: For baseline offsets caused by variable loads and temperature / humidity fluctuations, steady-state deviations are corrected through integral calculations while preserving dynamic fault characteristics. In the formula: — Corrected actual temperature and pressure fault characteristics, units: ℃, Pa; —Integral time series variable, unit: s.

[0035] S4 Fault Confidence-Driven Multi-Model Dynamic Weight Adaptive Fusion Traditional fixed-weight fusion models cannot adapt to dynamic and varying working conditions in downhole environments, resulting in poor fault diagnosis tolerance. An innovative working condition-adaptive dynamic weight fusion system is constructed, which adaptively allocates fusion weights based on real-time working conditions and the confidence levels of various physical quantities. This achieves optimal fusion of multi-dimensional fault features, significantly improving the accuracy of fault diagnosis under complex working conditions.

[0036] Based on the preliminary fault diagnosis results of each physical quantity single model, a dual judgment mechanism of operating condition and fault confidence is introduced to replace the traditional fixed-weight fusion mode. The mechanism identifies operating condition parameters such as the current environmental interference intensity, idler roller running speed, and load in real time, dynamically calculates the effective confidence of each physical quantity's diagnostic results, and adaptively adjusts the fusion weights. Under high-noise and strong-interference conditions, the fusion weights of steady-state physical quantities such as temperature and pressure are increased; under conditions of high-speed idler roller operation and significant mechanical characteristics, the fusion weights of transient physical quantities such as vibration and acoustics are amplified. This completes the adaptive weighted fusion of multi-dimensional fault diagnosis results and outputs a preliminary comprehensive fault judgment result for the idler roller.

[0037] A working condition confidence assessment model based on fuzzy membership is developed, a fuzzy evaluation set of interference working conditions is constructed, and the effective confidence of each physical quantity is calculated to characterize the reliability of the single-dimensional diagnostic results. In the formula: —Confidence level of the i-th type of physical quantity, dimensionless, with values ​​[0,1]; —Real-time operating condition disturbance parameters, dimensionless; —Standard, non-interference operating condition parameters, dimensionless; — Operating condition disturbance tolerance threshold, dimensionless.

[0038] With minimizing fault diagnosis error as the optimization objective, a weight-constrained optimization model is constructed, and the optimal fusion weight is adaptively solved. A dynamic fusion method based on operations research to solve for the optimal weight is proposed. Multi-physical quantity fault fusion judgment method: In the formula: —The objective function for diagnosing errors is dimensionless; —Dynamically fused weights for the i-th type of physical quantity, dimensionless; — Fault determination value for a single physical quantity, dimensionless; —Comprehensive fault determination confidence level, dimensionless, with values ​​[0,1]. The larger the value, the higher the probability of fault.

[0039] S5 Digital Twin Virtual-Real Mapping Construction and Initial State Initialization Precise mapping between digital twins and the real world is the core technology for realizing idler roller state simulation and fault iterative optimization, requiring accurate spatial structure modeling and state parameter initialization algorithms. This invention constructs an integrated geometry-mechanism-state mapping model for the idler roller twin, achieving a high-precision one-to-one replication of the physical entity and the virtual twin, providing a standardized simulation benchmark for subsequent closed-loop iterative optimization.

[0040] Based on the physical structural parameters and operational process parameters of the idler rollers, along with the aforementioned pre-processed standard operating data, a digital twin simulation model of the underground idler roller group in a coal mine was constructed, achieving a one-to-one virtual-real mapping between the physical entity and the virtual twin. The standard operating parameters, normal state thresholds, and basic fault judgment criteria of the twin were initialized, replicating the actual installation layout, operating conditions, and environmental conditions of the underground idler rollers. This enabled the visualization and digital replication of the idler roller operating status, providing a simulation platform and benchmark support for subsequent fault iteration correction and state comparison analysis.

[0041] Twin structure virtual-real mapping matrix: Construct a mapping matrix for the structural parameters of the idler roller to achieve precise replication of its geometric dimensions and installation posture: In the formula: —The virtual twin structure parameter matrix includes size, orientation, and position parameters, with dimensions varying depending on the parameter type; —The physical structure parameter matrix of the idler roller; —The virtual-real mapping correction coefficient matrix is ​​dimensionless and is the unit correction matrix.

[0042] Initialization of initial state parameters based on numerical fitting: The standard operating state threshold of the idler roller is solved by the least squares fitting algorithm to complete the initialization of the twin parameters. In the formula, —Parameters from historical normal operation samples; —The initial fitting parameters of the twin are dimensionless, and the baseline operating threshold and state parameters of the twin are determined through optimal fitting.

[0043] S6 Physics-Data-Twin Closed-Loop Iterative Optimization and Model Parameter Self-Correction Idler rollers undergo long-term operation, resulting in equipment aging and gradual changes in operating conditions. Static models are prone to accuracy degradation and decreased adaptability. To address this, a closed-loop iterative system of "physical perception - data diagnosis - twin update" was constructed to achieve dynamic self-correction of model parameters and ensure stable fault determination accuracy throughout the entire lifecycle.

[0044] A closed-loop iterative optimization system is constructed, utilizing real-time downhole idler roller operation sensor data and dynamically fused fault diagnosis results to iteratively correct the core parameters of the fault analysis model for each physical quantity, while simultaneously updating the simulation operating condition parameters, fault judgment thresholds, and state mapping rules of the digital twin. Addressing issues such as equipment aging and gradual changes in operating conditions caused by long-term idler roller operation, continuous dynamic calibration of the twin and the physical entity is performed, resolving the problems of fixed parameters and accuracy decay in traditional static models, and enabling the model's judgment capability to continuously and adaptively optimize according to the equipment's operating status.

[0045] Dynamic iterative correction of model parameters: Based on the gradient descent iteration criterion, the core parameters of the fault analysis model are updated in real time. In the formula: —The core parameters of the model in the kth and k+1th iterations are dimensionless; —Iterative learning step size, dimensionless, takes the value (0,1), controls the correction rate; —The diagnostic error function for the k-th iteration, dimensionless; — Error gradient, characterizing the direction and magnitude of parameter correction, is dimensionless.

[0046] Twin state virtual-to-real deviation correction: Based on real-time data statistical deviations, the state of the twin simulation is corrected to achieve virtual-real synchronization: In the formula: —The mean square error between the virtual and real states has the same dimensions as the corresponding physical quantity; —Number of samples collected, dimensionless; —Physical entity and twin state parameters; — Deviation compensation coefficient, dimensionless; — Corrected twin state parameters.

[0047] S7 idler roller latent fault prediction and refined disassembly and source tracing of complex faults Early-stage latent faults in idler rollers often exhibit subtle characteristics and strong coupling between multiple faults, making accurate identification and dismantling difficult using traditional methods. This new method enables latent fault prediction, dismantling of multiple faults, fault hierarchy classification, and precise source tracing, achieving comprehensive and refined fault assessment.

[0048] Based on the iteratively optimized high-precision twin-model judgment and multi-model fusion results, a comprehensive and refined fault analysis is conducted. On the one hand, early-stage subtle fault characteristics of the idlers are extracted, enabling early prediction and warning of non-manifested faults such as minor bearing wear and latent obstruction. On the other hand, for complex faults caused by multiple coupled factors, fault type decomposition, fault severity classification, and precise fault location tracing are completed, distinguishing different fault causes such as mechanical wear faults, environmentally induced faults, and abnormal load faults. Finally, a full-state monitoring report of the idler group, graded early warning information, and operation and maintenance suggestions are output, achieving refined health management of the idlers throughout their entire life cycle.

[0049] Prediction of subtle characteristics of latent faults: Early fault prediction can be achieved by extracting weak fault features from time-series data using higher-order differentiation. When the third-order derivative value exceeds the normal threshold range, an early latent fault is identified. The third-order characteristic gradient characterizes the weak fault evolution trend, with the dimension being the corresponding physical quantity / s. 3 .

[0050] Fuzzy decomposition and classification of complex faults: Construct a fuzzy fault evaluation matrix to achieve composite fault type decomposition and classification: In the formula: — Fault evaluation result vector, dimensionless, corresponding to membership degrees of different fault levels; — Fault feature weight vector, dimensionless; —Fault feature fuzzy relation matrix, dimensionless; — Fuzzy synthesis operator, dimensionless.

[0051] Fault tracing probability assessment: Based on Bayesian probability formulas, the posterior probability of various fault causes is calculated to achieve accurate source tracing: In the formula: —The posterior probability of the i-th type of fault under observed feature X, dimensionless; —The prior probability of the feature corresponding to the i-th type of fault, dimensionless; —The historical prior probability of the i-th type of fault, dimensionless.

[0052] S8 fault determination result storage and iterative updates of the operating condition database To achieve long-term adaptive optimization of the model and continuously improve its adaptability to extreme working conditions, it is necessary to store and iteratively update the entire process of working condition data, model parameters, and fault results in a closed loop. Based on mathematical statistics sample distribution optimization and database dynamic update algorithms, a self-growing sample database is constructed to continuously iterate and optimize the core parameters and fusion logic of the model, forming a closed loop of intelligent optimization throughout the entire life cycle.

[0053] Each set of operating data, model weight parameters, fault determination results, and fault tracing information is completely stored in the downhole idler intelligent monitoring database, continuously accumulating complex operating condition samples and fault samples. Based on massive measured samples, the dynamic weight allocation logic and fault analysis model accuracy are further optimized, forming a self-growing closed-loop system covering the entire process of "data acquisition - preprocessing modeling - fusion judgment - iterative optimization - sample accumulation," continuously improving the accuracy and adaptability of idler fault determination under complex downhole extreme operating conditions.

[0054] Sample data statistical optimization: By correcting for sample mean and variance, the distribution characteristics of the database samples are optimized, thereby improving sample effectiveness. In the formula: —Updated original sample mean, with units corresponding to physical quantities; —Updated original sample variance, with the dimension being the square unit of the corresponding physical quantity; —Sample forgetting coefficient, dimensionless, takes values ​​(0,1), ensuring balanced iteration of old and new samples; —Newly added measured sample data.

[0055] The following section verifies the fault determination method for idler rollers driven by the fusion of multi-physical quantity analytical models of the present invention.

[0056] like Figure 2As shown, the four sub-figures in the top row are the original time-series curves of the four sensors (temperature, vibration, sound, and pressure) without spatiotemporal registration in existing technologies. The maximum sampling phase offset of the four curves reaches 1.2s, and the time axes are not synchronized. At the same moment, the wear fault characteristics of the idler roller bearing are spatiotemporally misaligned and fragmented. Single fault features are scattered across different time slices, and direct fusion will result in the loss of fault features and a significant increase in the false negative rate. The four sub-figures in the bottom row are the synchronized time-series curves after the cubic spline interpolation time-series calibration and rigid body coordinate space mapping registration of this invention. The fault abrupt peaks of the four types of physical quantities are completely aligned, and the spatiotemporal coupling of fault features is completely preserved. This figure intuitively demonstrates that the spatiotemporal dual-dimensional registration model of this invention solves the industry pain point of spatiotemporal mismatch of distributed heterogeneous sensors, provides a consistent and standardized data foundation for subsequent mechanism decoupling and multi-model fusion, and is a prerequisite for high-precision fault determination.

[0057] like Figure 3 As shown, the horizontal axis represents the number of days the model iterates, and the vertical axis represents the mean square error between the physical entity and the digital twin. Existing static digital twins lack a gradient descent parameter self-correction mechanism. As the equipment ages, the virtual-to-real error increases linearly, exceeding the fault determination threshold after 30 days, leading to diagnostic failure. This invention introduces a gradient iteration parameter correction + virtual-to-real deviation compensation model, achieving rapid convergence in the first 7 days. The virtual-to-real error then continuously decays exponentially and remains stably at an extremely low level. Throughout the equipment's lifecycle, the twin simulation state is synchronized with the physical idler in real time. This figure demonstrates that the "perception-diagnosis-twin update" closed-loop iterative system of this invention solves the industry problem of the continuous decline in accuracy of traditional static twin models with equipment aging, ensuring stable and reliable fault determination throughout the idler's lifecycle.

[0058] like Figure 4 As shown, the horizontal axis represents the iterative batch of newly added fault samples in the database, the left vertical axis represents the sample mean, and the right vertical axis represents the sample variance. In the initial stage, the number of fault samples is small, the mean fluctuates greatly, the variance is high, and the model's generalization ability is poor. After adopting the forgetting coefficient dynamic sample update formula of this invention, as batches accumulate, the sample mean quickly converges and stabilizes, the variance continuously decreases, and the sample distribution approaches the true global distribution of downhole faults. Existing technologies use full sample storage with equal weight for new and old samples, resulting in old, failed samples continuously interfering with the model. This invention's dynamic statistical update mechanism balances the weights of new and old samples, constructs a self-growing fault database, and continuously improves the ability to identify extremely rare faults. This figure supports the technical effects of this invention's long-term iterative optimization and continuous enhancement of model generalization ability.

[0059] like Figure 5As shown in the histogram, the fault feature amplitude distributions output by the traditional inorganic decoupling scheme and the present invention's physical quantity adaptive decoupling model are compared. The traditional scheme is affected by environmental resonance, airflow noise, and load baseline drift, resulting in the effective fault features being submerged and feature amplitudes concentrated in the noise range, with a high degree of overlap between fault and normal samples. In contrast, the present invention, after differential filtering, Gaussian noise filtering, and integral baseline correction, significantly separates the fault feature amplitudes, with almost no overlap between the distribution ranges of normal and fault samples. Step S3, the multi-physical quantity differential decoupling model, verifies that it can remove environmental clutter and purify true fault features, solving the pain points of low feature recognition and high misjudgment rate caused by interference superposition in existing technologies.

[0060] like Figure 6 As shown, the horizontal axis represents the continuous operating time of the idler roller (0–180 days), and the vertical axis represents the monthly false alarm rate. The three existing technologies lack a closed-loop self-correction mechanism, and as equipment ages and operating conditions drift, the false alarm rate continuously increases, exceeding 22% after 180 days. The present invention's sample database iterative update + gradient parameter self-correction closed-loop system, after a slight initial decrease, maintains a stable false alarm rate below 3% in the long term, with no trend of accuracy decay. This demonstrates the self-growing operating condition database and the full-process closed-loop system in step S8, achieving stable, low-misjudgment operation throughout the equipment's lifecycle, and solving the maintenance pain points of existing technologies such as model failure after long-term use and frequent manual calibration.

[0061] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape, principle and application direction of this application should be covered within the scope of protection of this application.

Claims

1. A method for determining faults in a digital twin of idler rollers driven by the fusion of multiple physical quantity analytical models, characterized in that, Includes the following steps: S1 Distributed Sensing Deployment and Raw Data Acquisition of Multi-Source Physical Quantities for Underground Idler Groups: Based on the distribution layout, operating conditions, and environmental characteristics of underground belt conveyor idler groups in coal mines, a distributed fixed-point installation and deployment of multiple types of sensors was completed. These sensors respectively collect raw data on four core physical quantities of idler operation: temperature, pressure, vibration, and acoustics. Simultaneously, underground environmental condition data are collected to construct a comprehensive and multi-dimensional raw sensing data sample set. S2 Multi-Source Sensor Data Spatiotemporal Dual-Dimensional Precise Registration Preprocessing: In the time dimension, the sampling timing benchmark of each sensor is unified, and the data transmission delay and sampling phase deviation are calibrated; in the spatial dimension, the original physical quantity data of the idler roller is combined with the actual installation coordinates of the sensor to establish a spatial mapping relationship and correct the spatial offset error of the sensor data at different points. S3 Physical Quantity Adaptive Fault Mechanism Decoupling Modeling and Interference Removal: Independent fault analysis mechanism models for idler rollers are constructed to address the differences in fault mechanisms and environmental interference characteristics, achieving differentiated interference decoupling; irrelevant interference from underground equipment resonance and mechanical vibrations is removed from vibration data; background noise from roadway airflow disturbances and dust friction is filtered from acoustic data; baseline offset errors caused by belt load variations and environmental temperature and humidity fluctuations are eliminated from temperature and pressure data. S4 Fault Confidence-Driven Multi-Model Dynamic Weight Adaptive Fusion: Based on the preliminary fault diagnosis results of each idler roller fault analysis mechanism model, a dual judgment mechanism of working condition and fault confidence is introduced to identify the working condition parameters of the current underground environment in real time, dynamically calculate the effective confidence of each physical quantity diagnosis result, form a working condition confidence evaluation model based on fuzzy membership, construct a fuzzy evaluation set of interference working conditions, adaptively adjust the fusion weight, complete the adaptive weighted fusion of multi-dimensional fault diagnosis results, and output the preliminary comprehensive fault judgment result of the idler roller; S5 Digital Twin Virtual-Real Mapping Construction and Initial State Initialization: Based on the original physical quantity data of the idler rollers, the operating process parameters and the aforementioned pre-processed standard operating data, a digital twin simulation model of the idler roller group in the coal mine is built to complete the one-to-one virtual-real mapping between the physical entity and the virtual twin. S6 Physical-Data-Twin Closed-Loop Iterative Optimization and Model Parameter Self-Correction: Construct a closed-loop iterative optimization system, utilize real-time collected roller operation sensor data from underground and dynamically fused fault diagnosis results to iteratively correct the core parameters of each physical quantity fault analysis model, and synchronously update the simulation working condition parameters, fault judgment thresholds and state mapping rules of the digital twin. S7 Idler Roller Latent Fault Prediction and Complex Fault Refined Disassembly and Source Tracing: Based on the iteratively optimized high-precision twin judgment model and multi-model fusion results, a full-dimensional refined fault analysis is carried out, weak features of latent faults of idlers are extracted and predicted, complex faults are fuzzily disassembled and graded, and finally, a full-state monitoring report of idler group, graded early warning information and operation and maintenance suggestions are output, and the probability of fault source tracing is assessed. S8 Fault Judgment Result Storage and Working Condition Database Iterative Update: Each working condition data, model weight parameters, fault judgment result, and fault source information are completely stored in the underground idler roller intelligent monitoring database to continuously accumulate complex working condition samples and fault samples, and optimize the dynamic weight allocation logic and fault analysis model accuracy.

2. The method for determining the fault of a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S1, the multi-source physical quantity time-series sampling matrix model is constructed as follows: In the formula: —Multi-source physical quantity timing acquisition matrix; —Temperature time-series acquisition vector, The temperature value of the idler roller collected by the i-th temperature sensor at time t; —Pressure timing acquisition vector, The value of the idler support load pressure collected by the i-th pressure sensor at time t; —Vibration timing acquisition vector, Let t be the vibration acceleration value of the roller body collected by the i-th vibration sensor at time t; —Acoustic timing acquisition vector, Let be the roller running noise value collected by the i-th acoustic sensor at time t; — Collect time-series variables.

3. The method for determining faults in a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S2, time-series calibration based on cubic spline interpolation is performed in the time dimension: The cubic spline interpolation algorithm is used to fill in missing values ​​in the time-series data, correct phase shifts, and construct a continuous and smooth time-series data curve. The core constraint equation for interpolation is as follows: In the formula: —The interpolation fitting function for the i-th sampling interval; —The interpolation coefficients for the i-th interval are the coefficients of the function offset term, the first term, the second term, and the third term, respectively; —Adjacent sampling times; —The first and second derivatives of a function characterize the rate of change and acceleration of a physical quantity, respectively; In the spatial dimension, spatial mapping registration based on coordinate transformation is performed: a mapping relationship is established between the local coordinate system of the idler roller and the global monitoring coordinate system downhole. The spatial offset error of the sensor is corrected through a rigid body transformation matrix. The spatial registration formula is as follows: In the formula: —The sensor's global spatial coordinate vector; —Local installation coordinate vector of the sensor roller; —Three-dimensional rotation transformation matrix; — Spatial translation offset vector.

4. The method for determining faults in a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S3, for the vibration data, a resonance interference stripping model based on time-domain differential filtering is used: first-order and second-order differentials are used to capture the transient fault characteristics of vibration, and steady-state equipment resonance interference is stripped away. The fault feature purification formula is then used. In the formula: —Characteristic quantities of fault vibration after purification; —The original vibration timing signal; —Resonance interference suppression coefficient; For acoustic signals, a background noise filtering model based on Gaussian probability distribution is proposed: Downhole background noise follows a Gaussian normal distribution; effective fault acoustic features are selected through a probability density function to filter out random environmental noise. In the formula: —Acoustic noise probability density value; —Background noise mean and standard deviation; —Refined acoustic characteristics of the fault; —Indicator function; —Noise threshold probability; For temperature and pressure signals, a baseline drift integral correction model is used: to correct steady-state deviations caused by variable loads and temperature and humidity fluctuations through integral calculations. In the formula: —Corrected true temperature and pressure fault characteristics; —Integral time series variable.

5. The method for determining the fault of a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S4, the working condition confidence assessment model In the formula: —Confidence level of the i-th type of physical quantity; —Real-time operating condition interference parameters; —Standard, non-interference operating parameters; — Operating condition interference tolerance threshold.

6. The method for determining faults in a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S5, the virtual-real mapping matrix of the twin structure is used. In the formula: —The virtual twin's structural parameter matrix, including size, orientation, and position parameters; —The physical structure parameter matrix of the idler roller; —The virtual-real mapping correction coefficient matrix; Initialization of initial state parameters based on numerical fitting: The standard operating state threshold of the idler roller is solved by the least squares fitting algorithm to complete the initialization of the twin parameters. In the formula: —Parameters from historical normal operation samples; —Initialize the fitting parameters for the twin.

7. The method for determining the fault of a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S6, the model parameters are dynamically iteratively corrected: based on the gradient descent iteration criterion, the core parameters of the fault analysis model are updated in real time. In the formula: —The core parameters of the model in the kth and k+1th iterations; —Iterative learning step size; —The diagnostic error function for the k-th iteration; —Error gradient; Twin state virtual-real deviation correction: Based on real-time data statistical deviation, the virtual-real simulation state is corrected to achieve virtual-real synchronization. In the formula: —Mean square error between real and virtual states; —Number of samples collected; —Physical entity and twin state parameters; — Deviation compensation coefficient; — Corrected twin state parameters.

8. The method for determining faults in a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S7, the prediction of weak features of latent faults is achieved by extracting weak fault features from time-series data through high-order differentiation, thus realizing early fault prediction. When the third-order derivative value exceeds the normal threshold range, an early latent fault is identified. It is the third-order characteristic gradient; Fuzzy Decomposition and Classification of Composite Faults: Constructing a fuzzy fault evaluation matrix to achieve decomposition and classification of composite fault types. In the formula: —Fault evaluation result vector; — Fault feature weight vector; —Fault feature fuzzy relation matrix; —Fuzzy synthesis operator; Fault Origin Probability Assessment: Based on Bayesian probability formulas, the posterior probability of various fault causes is calculated to achieve accurate source tracing. In the formula: —The posterior probability of the i-th type of fault under observed feature X; —The prior probability of the feature corresponding to the i-th type of fault; —The historical prior probability of the i-th type of fault.

9. The method for determining faults in a digital twin of an idler roller driven by the fusion of multiple physical quantity analytical models according to claim 1, characterized in that, In step S8, sample data statistical optimization is performed: the sample distribution characteristics of the database are optimized by correcting for the sample mean and variance. In the formula: —Updated original sample mean; —Updated original sample variance; —Sample forgetting coefficient; —Newly added measured sample data.

Images