Method and system for predictive maintenance management of equipment based on multi-feature fusion

Abstract

This application relates to a predictive maintenance management method and system for equipment based on multi-feature fusion, belonging to the field of equipment health management technology. It addresses the problems of existing sensor data being easily distorted by interference and lacking a comprehensive modeling framework to reveal the potential causal structure of equipment degradation paths. The method includes: real-time acquisition of multi-source data; extraction of the causal contribution of data to failures to generate causal significance feature values; generation of a three-dimensional health feature vector through a counterfactual neural network model combined with a physical failure model; fusion of the two to obtain an equipment state risk feature matrix, which is then input into the model to output a risk score; and finally, dynamic adjustment of monitoring frequency and maintenance level, and iterative optimization of the strategy. This application can filter out data interference, deeply integrate multiple inference mechanisms, accurately reveal the degradation patterns of equipment, and achieve adaptive maintenance throughout the entire life cycle.

Description

Technical Field

[0001] This invention belongs to the field of equipment health management technology, specifically relating to a predictive maintenance management method and system for equipment based on multi-feature fusion. Background Technology

[0002] As modern equipment becomes increasingly complex, information-based, and intelligent, its operating environment becomes more demanding, and the requirements for mission reliability and integrity continue to rise. Traditional maintenance methods based on scheduled maintenance or simple condition monitoring are no longer sufficient to meet the needs of highly dynamic tasks. There is an urgent need for a predictive maintenance method that can integrate multi-source heterogeneous sensor data, deeply explore the evolution patterns of failures, and achieve risk-driven outcomes. In recent years, condition monitoring technologies based on big data analytics and artificial intelligence have been gradually applied to the field of equipment health management. However, the lack of effective modeling of causal mechanisms, physical failure processes, and the dynamic coupling relationship of tasks limits further improvements in prediction accuracy and strategy adaptability. Existing technologies suffer from the following problems: sensor-collected data is easily distorted by strong electromagnetic interference, rapid changes in operating conditions, and complex environmental coupling conditions; and there is a lack of a comprehensive modeling framework that can combine physical failure mechanisms, causal reasoning mechanisms, and counterfactual reasoning capabilities, making it impossible to reveal the potential causal structure in the equipment degradation path with limited data support.

[0003] In view of this, it is very necessary to provide a predictive maintenance management method and system for equipment based on multi-feature fusion to solve the above-mentioned defects in the prior art. Summary of the Invention

[0004] To address the technical problems of existing technologies, such as the susceptibility of sensor-collected data to distortion due to strong electromagnetic interference, rapid changes in operating conditions, and complex environmental coupling conditions; and the lack of a comprehensive modeling framework that can combine physical failure mechanisms, causal reasoning mechanisms, and counterfactual reasoning capabilities, thus failing to reveal the potential causal structure in equipment degradation paths with limited data support, this invention provides a predictive maintenance management method and system for equipment based on multi-feature fusion to solve the aforementioned technical problems.

[0005] In a first aspect, the present invention provides a predictive maintenance management method for equipment based on multi-feature fusion, comprising:

[0006] Step S1: The step of collecting multi-source data, which involves collecting multi-source data in real time during equipment operation;

[0007] Multi-source data includes electrical parameter data, mechanical vibration data, temperature distribution data, and electromagnetic compatibility parameter data;

[0008] The electrical parameters of the equipment's electrical system are collected using a distributed sensor network.

[0009] Mechanical vibration data of the equipment is acquired using a three-axis MEMS accelerometer array;

[0010] Temperature distribution data were acquired using infrared thermal imaging and embedded thermocouples.

[0011] Electromagnetic compatibility parameter data of the equipment are collected using a broadband electric field probe and a magnetic field sensor.

[0012] Step S2: The step of causal feature analysis, extracting the causal contribution of multi-source data to equipment failure, and generating causal significance feature values;

[0013] Extracting the causal contribution of multi-source data to equipment failure includes pre-acquiring the environmental parameters of the current equipment environment, an environmental correction rule base, and constructing a causal directed graph.

[0014] The environmental parameters currently in which the equipment is located include electromagnetic environment intensity data, atmospheric temperature gradient data, and vibration spectrum characteristic data. Electromagnetic environment intensity data is collected by an electromagnetic field measuring instrument, atmospheric temperature gradient data is collected by a radiosonde or weather tower, and vibration spectrum characteristic data is collected by a triaxial MEMS accelerometer and a dynamic signal analyzer.

[0015] The environmental correction rule base includes the reliability decay coefficients for different equipment models under different combinations of environmental parameters. The reliability decay coefficients range from 0.1 to 1.0 in decimal.

[0016] The multi-source data collected in step S1 is weighted and fused with the credibility attenuation coefficient to obtain the feature sequence of environment adaptation.

[0017] A causal discovery algorithm is used to analyze the conditional independence relationships between feature sequences of environmental adaptation, and a causal directed graph is constructed.

[0018] Among them, conditional independence is whether there is no longer any mutual influence between two feature sequences under specific conditional variables. Conditional independence is used to determine whether there is a real causal relationship between multi-source data and faults.

[0019] The process of generating causal significance feature values ​​includes:

[0020] In a causal directed graph, identify the directed path from the sensor node used to collect multi-source data to the fault node;

[0021] The magnitude of the change in conditional probability between nodes on each path is statistically analyzed and used as the causal strength of the path.

[0022] By combining the causal strength of all paths pointing to the same fault node and superimposing the reliability decay coefficient of the environmental parameters of the current location of the data acquisition equipment, a causal significance feature value is generated.

[0023] Step S3: The step of generating a three-dimensional health feature vector is to fuse multi-source data through a counterfactual neural network model and combine it with the physical failure model of the equipment to generate a three-dimensional health feature vector containing failure probability value, remaining life value and availability value.

[0024] The counterfactual neural network model is a neural network architecture with multimodal input layers, in which multi-source data is processed through independent feature extraction subnetworks;

[0025] A causal attention network is introduced into the output layer of each sub-network, and the feature weights are dynamically adjusted based on the causal significance feature values.

[0026] The multi-source data is nonlinearly coupled through a cross-modal feature interaction layer, outputting a unified feature representation of the fused multi-source data, and an adversarial training strategy is used to optimize the network parameters.

[0027] The process of generating a three-dimensional health feature vector includes:

[0028] The unified feature representation of the fused multi-source data is input into the physical constraint branch. The output of the counterfactual neural network is coordinated with the calculation results of the equipment physical failure model through the physical and data joint optimization layer to generate a three-dimensional health feature vector.

[0029] Among them, the physical constraint branch uses the laws of the equipment physical failure model to transform into a computable constraint branch, and uses physical logic to constrain the output of the data model to ensure that the health status assessment results conform to the data laws and do not violate the physical principles of actual equipment failure.

[0030] The physical constraint branch includes a failure probability calculation unit, a remaining lifetime prediction unit, and an availability assessment unit.

[0031] The failure probability calculation unit constructs the equipment physical failure equation based on the material fatigue cumulative damage theory.

[0032] The remaining lifetime prediction unit is a solver of the differential equations embedded in the component degradation rate;

[0033] Availability assessment unit, which is an environmental dynamics sub-model constructed by combining the environmental parameters of the current environment in which the equipment is located;

[0034] The three-dimensional health feature vector includes the first dimension, the failure probability value; the second dimension, the remaining lifetime value; and the third dimension, the availability value.

[0035] Among them, the first dimension of the failure probability value is a weighted fusion of the counterfactual neural network prediction probability and the calculation result of the equipment physical failure model;

[0036] The second dimension, remaining lifetime, is the intersection of the predicted value of the data-driven branch of the counterfactual neural network and the confidence interval of the solution to the differential equation.

[0037] The third dimension, availability, is the real-time operability coefficient output by the environmental dynamics sub-model.

[0038] Among them, the real-time operability coefficient is a quantitative indicator output by the environmental dynamics sub-model, which comprehensively evaluates whether the equipment can safely and effectively complete the mission under the current environment.

[0039] Step S4: The step of outputting the risk score of the equipment is to fuse the causal significance feature value and the three-dimensional health feature vector into the equipment status risk feature matrix, input the equipment perception optimization model, and output the risk score of the equipment.

[0040] The process of constructing the equipment status risk characteristic matrix includes:

[0041] A spatiotemporally aligned feature fusion framework is established to synchronously match causal significance feature values ​​with three-dimensional health feature vectors according to timestamps;

[0042] The causal significance feature value and the three-dimensional health feature vector are combined by tensor splicing to form an original feature cube with three dimensions. The first dimension corresponds to the equipment subsystem classification, the second dimension corresponds to the time series window, and the third dimension corresponds to the feature type.

[0043] The feature importance filtering layer eliminates redundant feature dimensions based on the feature weight template preset by the equipment model.

[0044] Among them, the feature weight template is a feature importance comparison table pre-set based on design specifications, historical failure data and industry experience for a specific equipment model. The feature weight template provides a clear quantitative standard for feature selection and quickly eliminates redundant features.

[0045] The original feature cube is compressed using a spatiotemporal attention mechanism to output the equipment status risk feature matrix;

[0046] The process of obtaining the risk score for equipment includes:

[0047] The equipment status risk feature matrix is ​​input into the equipment perception optimization model based on ensemble learning, wherein the equipment perception optimization model is constructed using the Stacking ensemble framework.

[0048] The first layer integrates base learners, including gradient boosting decision trees, deep neural networks, and random forests, which respectively handle different dimensions of the equipment state risk feature matrix.

[0049] The Bayesian-optimized support vector regression model is used in the second-layer meta-learner, integrating the output of the base learners;

[0050] The weight coefficients of each base learner are dynamically adjusted through a feature importance feedback mechanism, and a standardized risk score is output.

[0051] Step S5: Dynamically adjust monitoring frequency and maintenance level. Based on the equipment's risk score and task priority, dynamically adjust the monitoring frequency and maintenance response time, and iteratively optimize the maintenance strategy through a closed-loop feedback mechanism to achieve adaptive maintenance management throughout the equipment's life cycle.

[0052] Based on the equipment's risk score and task priority, the monitoring frequency and maintenance level are dynamically adjusted, including pre-constructing a task priority assessment matrix and classifying the equipment status according to the risk score interval.

[0053] The pre-constructed task priority evaluation matrix includes dividing the current task into three levels: Level I, Level II, and Level III. Each level corresponds to a different maintenance response coefficient, the specific value of which is set by experts.

[0054] Based on the risk score range, the equipment status is divided into three categories: emergency status, alert status, and normal status. Specifically, for emergency status, the score is >0.7 and the alarm indicator is set to red; for alert status, the score range is 0.4-0.7 and the alarm indicator is set to yellow; and for normal status, the score range is 0-0.4 and the alarm indicator is set to green.

[0055] Determine the monitoring frequency: Monitoring frequency = base value × maintenance response coefficient × status coefficient, where the base value is a preset value, and the status coefficient is 1.5 for emergency status, 1.2 for alert status, and 1.0 for normal status.

[0056] Determine the maintenance response time: Maintenance response time = base value / (maintenance response coefficient × status coefficient), where the base value is a preset value, and the status coefficient is 1.5 for emergency status, 1.2 for alert status, and 1.0 for normal status.

[0057] Dynamically adjusting monitoring parameters includes adjusting the sensor sampling rate, data transmission interval, and feature analysis depth;

[0058] The closed-loop feedback mechanism iterative optimization and maintenance strategy includes:

[0059] Construct a knowledge graph of equipment health status evolution to record status change data throughout the entire process from initial service to final retirement.

[0060] Establish a maintenance strategy feedback loop to compare and analyze the results of each maintenance with the expected goals;

[0061] The maintenance decision rule base is updated based on the comparison results, the logic for generating subsequent maintenance strategies is optimized, and the optimal maintenance scheme for the current stage is automatically matched.

[0062] Secondly, the technical solution of the present invention also provides an equipment predictive maintenance management system based on multi-feature fusion, including a multi-source data acquisition module, a causal feature analysis module, a three-dimensional health feature vector generation module, an equipment risk score output module, and a dynamic adjustment module for monitoring frequency and maintenance level.

[0063] The multi-source data acquisition module collects multi-source data in real time during equipment operation;

[0064] Multi-source data includes electrical parameter data, mechanical vibration data, temperature distribution data, and electromagnetic compatibility parameter data;

[0065] The electrical parameters of the equipment's electrical system are collected using a distributed sensor network.

[0066] Mechanical vibration data of the equipment is acquired using a three-axis MEMS accelerometer array;

[0067] Temperature distribution data were acquired using infrared thermal imaging and embedded thermocouples.

[0068] Electromagnetic compatibility parameter data of the equipment are collected using a broadband electric field probe and a magnetic field sensor.

[0069] The causal feature analysis module extracts the causal contribution of multi-source data to equipment failure and generates causal significance feature values.

[0070] The module for generating three-dimensional health feature vectors uses a counterfactual neural network model to fuse multi-source data and combine it with the physical failure model of equipment to generate three-dimensional health feature vectors containing failure probability values, remaining life values, and availability values.

[0071] The risk scoring module for output equipment integrates causal significance feature values ​​and three-dimensional health feature vectors into an equipment status risk feature matrix, inputs the equipment perception optimization model, and outputs the equipment risk score.

[0072] The module dynamically adjusts the monitoring frequency and maintenance level based on the equipment's risk score and task priority. It also iteratively optimizes the maintenance strategy through a closed-loop feedback mechanism to achieve adaptive maintenance management throughout the equipment's entire life cycle.

[0073] The beneficial effects of this invention are as follows: The equipment predictive maintenance management method and system based on multi-feature fusion provided by this invention, through the setting of multi-dimensional sensors for real-time and dynamic acquisition of electrical parameters, mechanical vibration, temperature distribution, and electromagnetic compatibility parameters, and combined with advanced signal processing and data analysis algorithms, extracts effective information from massive amounts of raw data, effectively filtering out distortion components caused by various interferences. Simultaneously, relying on a closed-loop feedback mechanism, it dynamically adjusts the monitoring frequency and data acquisition depth according to the latest mission requirements and equipment status, ensuring high-quality monitoring data acquisition under complex environments and different operating conditions, thus solving the problem of sensor-acquired data being easily distorted by interference; by employing a counterfactual neural network model, it closely integrates multi-source monitoring data with the equipment physical failure model. This approach integrates advanced causal feature analysis techniques to accurately quantify the causal contribution of each monitoring parameter to equipment failure, achieving a deep combination of physical failure mechanisms, causal reasoning, and counterfactual reasoning. By simulating different operating scenarios to assess potential risks, and combining failure physics equations based on material fatigue cumulative damage theory and component degradation rate differential equation solvers, it accurately reveals the potential causal structure of equipment degradation paths with limited data. Furthermore, relying on the knowledge graph of equipment health status evolution to associate multi-dimensional historical data, and combining a closed-loop feedback optimization mechanism to incorporate maintenance execution effects into model retraining, it further enhances the ability to recognize and respond to complex failure modes, solving the technical challenges of lacking a comprehensive modeling framework and being unable to reveal the potential causal structure of degradation paths.

[0074] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a flowchart of the equipment predictive maintenance management method based on multi-feature fusion provided by the present invention.

[0077] Figure 2 This is a schematic diagram of the predictive maintenance management system for equipment based on multi-feature fusion provided by the present invention. Detailed Implementation

[0078] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0079] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0080] Example 1:

[0081] like Figure 1 As shown, this embodiment of the invention provides a predictive maintenance management method for equipment based on multi-feature fusion, including the following steps:

[0082] Step S1: The step of collecting multi-source data, which involves collecting multi-source data in real time during equipment operation;

[0083] Multi-source data includes electrical parameter data, mechanical vibration data, temperature distribution data, and electromagnetic compatibility parameter data;

[0084] The electrical parameters of the equipment's electrical system are collected using a distributed sensor network.

[0085] Mechanical vibration data of the equipment is acquired using a three-axis MEMS accelerometer array;

[0086] Temperature distribution data were acquired using infrared thermal imaging and embedded thermocouples.

[0087] Electromagnetic compatibility parameter data of the equipment are collected using a broadband electric field probe and a magnetic field sensor.

[0088] In this embodiment, a Hall effect current sensor (accuracy ±0.5%FS) and an isolated voltage probe (bandwidth DC-10MHz) are installed in the power subsystem of the electrical system to collect electrical parameters of the equipment's electrical system. The sampling frequency is set to 1kHz. For critical circuit nodes, a four-wire measurement method is used to eliminate the influence of lead resistance, and the operating current (range 0-500A), voltage to ground (range 0-1000V), and insulation resistance (range 0-100MΩ) are recorded synchronously. All electrical parameters are transmitted via a CAN bus, and EMI filtering devices are used to eliminate electromagnetic interference. The electrical parameter data packets are timestamped (synchronization accuracy ±1μs) and then stored in a circular buffer.

[0089] A triaxial MEMS accelerometer array (range ±50g, frequency response 0.5-5kHz) is installed in key transmission components of the equipment. Each measurement point collects mechanical vibration data in three dimensions: radial, axial, and tangential. An impact-resistant packaging structure ensures the triaxial MEMS accelerometer array operates normally in harsh environments. The vibration signals are converted by a 24-bit Σ-Δ ADC, and then a dedicated DSP chip calculates time-domain parameters (including RMS value and peak factor) and frequency-domain characteristics (including 1 / 3 octave band spectrum) in real time, generating a set of feature vectors every 100ms. For rotating components, an additional laser speed sensor (accuracy ±0.01%) provides phase reference.

[0090] A fusion temperature measurement system was constructed to collect temperature distribution data, combining infrared thermal imaging (640×512 resolution, thermal sensitivity <50mK) and K-type embedded thermocouples (accuracy ±1℃). A high-temperature resistant thermocouple array (50mm spacing) was deployed in the high-temperature region at a sampling rate of 10Hz; infrared scanning (30Hz frame rate) was used for the non-contact areas of the outer shell. The temperature distribution data was processed using a spatial interpolation algorithm to generate a two-dimensional temperature field distribution map, with hotspot coordinates (positioning accuracy ±2mm) and temperature rise gradient (℃ / s) labeled. All temperature distribution data underwent physical consistency verification using a temperature-stress coupling model.

[0091] The equipment is equipped with broadband electric field probes (10kHz-6GHz) and magnetic field sensors (DC-100kHz) deployed throughout, employing a fiber optic transmission and isolated measurement system. Real-time monitoring includes:

[0092] Conducted interference: Power line noise is captured via LISN network (CE102 standard).

[0093] Radiated field strength: Spatial EMI distribution measured by a three-dimensional isotropic antenna array (RE102 standard);

[0094] Electrostatic accumulation: Electrostatic potential monitoring of rotating parts (range ±50kV);

[0095] Transient pulses: Surge / spike capture compliant with MIL-STD-461G;

[0096] All EMC data are associated with equipment operating modes (such as radar power on / off status) through event-related tagging.

[0097] Step S2: The step of causal feature analysis, extracting the causal contribution of multi-source data to equipment failure, generating causal significance feature values, which are used to quantify the failure early warning value of each monitoring parameter;

[0098] Extracting the causal contribution of multi-source data to equipment failure includes pre-acquiring the environmental parameters of the current equipment environment, an environmental correction rule base, and constructing a causal directed graph.

[0099] The environmental parameters currently in which the equipment is located include electromagnetic environment intensity data, atmospheric temperature gradient data, and vibration spectrum characteristic data. Electromagnetic environment intensity data is collected by an electromagnetic field measuring instrument, atmospheric temperature gradient data is collected by a radiosonde or weather tower, and vibration spectrum characteristic data is collected by a triaxial MEMS accelerometer and a dynamic signal analyzer.

[0100] In this embodiment, a distributed electromagnetic sensor array is installed on the surface of the equipment. Each array node contains a dual-channel detection module for high frequency (1MHz-6GHz) and low frequency (10kHz-1MHz) to collect electromagnetic environment intensity data, which is used to scan the spatial electromagnetic spectrum in real time and identify preset characteristic signals and interference signals of specific frequency bands of radar beams.

[0101] A dual-mode system using a platinum resistance thermometer and an infrared thermal imager is used to collect atmospheric temperature gradient data. 12-16 temperature measurement points are arranged along the main load-bearing structure of the equipment, and a temperature gradient distribution map is generated every 5 seconds.

[0102] An array of 24 triaxial MEMS sensors is used, arranged in an equilateral triangle at key mechanical connection points. The sampling rate is set to 5kHz, and the 1 / 3 octave spectrum features are extracted in real time by an embedded signal processor.

[0103] The collected environmental parameters of the equipment are aligned using a timestamp algorithm to ensure the synchronization of multi-source data.

[0104] The environmental correction rule base includes the reliability attenuation coefficients for different equipment models under different combinations of environmental parameters. The reliability attenuation coefficients range from 0.1 to 1.0 in decimal. The smaller the reliability attenuation coefficient, the greater the interference of the current environment on the sensor data.

[0105] In this embodiment, the reliability attenuation coefficient of data collected by each sensor under different combinations of environmental parameters is obtained by setting up an environmental adaptability experiment. A 6-month environmental adaptability experiment is conducted on each type of equipment, including electromagnetic interference testing, temperature shock testing, and vibration interference testing, establishing a multi-dimensional interference mapping relationship. Electromagnetic interference testing is conducted in a microwave anechoic chamber, applying frequency sweep interference from 10kHz to 6GHz, setting a frequency step of 1% of the current frequency, recording the output deviation curve of the distributed electromagnetic sensor array, and generating a frequency-interference comparison table. Temperature shock testing is conducted in a climate chamber, cycling from -40℃ to 85℃ at a rate of 5℃ / min, establishing a temperature mutation rate-measurement error model. Vibration interference testing uses a six-degree-of-freedom vibration table to simulate random vibrations of different road surface spectra, analyzing the influence of vibrations in each frequency band on the readings of the three-axis MEMS sensor. An environmental correction rule base is formed, comprising a three-layer structure: a base layer storing raw test data, an intermediate layer containing feature extraction rules, and an application layer that is a dynamic reliability calculation engine supporting online updates and maintenance.

[0106] Multi-source data and environmental parameters of the equipment's current location are transmitted to the central processing unit (CPU) via a ruggedized fieldbus. The CPU receives and processes the data, performing a data reliability assessment every 200ms. It queries the interference intensity of the Hall current sensor in the 50–60Hz power frequency band of the current electromagnetic environment, and calculates the electromagnetic reliability factor by combining this with the immunity curve of the Hall current sensor model in the corresponding frequency band. It detects the rate of change of the atmospheric temperature gradient and assesses the temperature influence factor by referring to the resistance-temperature characteristics table of copper conductors. It calculates the mechanical interference factor by combining the number of mechanical shocks reported by the triaxial MEMS sensor and applying a vibration-electrical coupling model. The three factors are weighted geometrically averaged to obtain a reliability decay coefficient, which is used to adjust the confidence weight of the current measurement value. The system maintains a 30-day sliding window historical database, automatically triggering a data repair procedure when the real-time reliability falls below 0.5.

[0107] Among them, the electromagnetic reliability factor, temperature influence factor, and mechanical interference factor together constitute the reliability attenuation coefficient.

[0108] The multi-source data collected in step S1 is weighted and fused with the credibility attenuation coefficient to obtain the feature sequence of environment adaptation.

[0109] A causal discovery algorithm is used to analyze the conditional independence relationships among sensor feature sequences adapted to the environment, and a causal directed graph is constructed.

[0110] In this embodiment, an improved PC algorithm is used to construct a causal directed graph, and a full network reconstruction is performed daily at midnight. During initialization, a fully connected graph is established, containing all sensor nodes used to collect environmental parameters of the equipment and 28 typical fault nodes. Conditional independence testing employs a conditional mutual information test based on mutual information, with the test threshold dynamically adjusted according to the current electromagnetic interference level; the threshold is increased by 30% when interference is severe. During the edge orientation phase, equipment mechanism constraint rules are introduced, such as forcibly retaining the known causal chain from generator speed to lubricating oil temperature. Incremental updates are performed every 2 hours, and newly emerging temporary causal relationships are discovered through a sliding window chi-square test. The completed network is stored in the form of an adjacency matrix, accompanied by confidence scores and timestamp information for each edge.

[0111] Among them, the edge orientation stage is the core step in constructing a causal directed graph.

[0112] The process of generating causal significance feature values ​​includes:

[0113] In a causal directed graph, identify the directed path from the sensor node used to collect multi-source data to the fault node;

[0114] The magnitude of the change in conditional probability between nodes on each path is statistically analyzed and used as the causal strength of the path.

[0115] By combining the causal strength of all paths pointing to the same fault node and superimposing the reliability decay coefficient of the environmental parameters of the current location of the data acquisition equipment, a causal significance feature value is generated.

[0116] Step S3: The step of generating a three-dimensional health feature vector is to fuse multi-source data through a counterfactual neural network model and combine it with the physical failure model of the equipment to generate a three-dimensional health feature vector containing failure probability value, remaining life value and availability value.

[0117] The counterfactual neural network model is a neural network architecture with multimodal input layers, in which multi-source data is processed through independent feature extraction subnetworks;

[0118] In this embodiment, the counterfactual neural network model is a hierarchical neural network architecture. The electrical parameter data channel is configured with a dedicated network layer with current-voltage coupling characteristics, containing eight parallel one-dimensional convolutional branches, each processing electrical signal features in different frequency bands. The mechanical vibration data channel constructs a time-frequency joint analysis network, decomposing the vibration waveform through a wavelet transform layer and then filtering key frequency band components through an attention mechanism. The temperature field processing network processes temperature distribution data, integrating a spatial convolution module and a temporal recursion module to simultaneously capture the spatial gradient changes and temporal evolution trends of the temperature distribution. Electromagnetic compatibility parameter data is processed through a frequency domain attention network, focusing on extracting interference spectrum features related to the equipment's operating mode. A causal attention network is set at the output of each sub-network, dynamically adjusting the feature contribution based on the causal significance feature values ​​generated in step S2, and assigning enhancement coefficients of 0.8-1.2 to features with high causal significance.

[0119] A cross-attention mechanism is deployed in the feature fusion layer to achieve multi-source data association. Electrical and vibration features interact through motion-electromagnetic coupling units to analyze the phase relationship between current harmonics and mechanical vibration. Temperature and electromagnetic features are fused in the dielectric loss calculation layer to evaluate the temperature rise effect of insulating materials in electromagnetic fields. An adversarial training mechanism is introduced into the interaction process: the generator simulates data distortion under extreme conditions, and the discriminator learns to identify and compensate for feature biases. The adversarial network is retrained weekly using the latest data to maintain its adaptability to new interference patterns. The fused unified feature representation is standardized to form a 256-dimensional feature vector for use by downstream modules.

[0120] A causal attention network is introduced into the output layer of each sub-network, and the feature weights are dynamically adjusted based on the causal significance feature values.

[0121] The multi-source data is nonlinearly coupled through a cross-modal feature interaction layer, outputting a unified feature representation of the fused multi-source data, and an adversarial training strategy is used to optimize the network parameters.

[0122] In this embodiment, electrical parameter data and mechanical vibration data interact through a motion-electromagnetic coupling unit to analyze the phase relationship between current harmonics and mechanical vibration; temperature distribution data and electromagnetic compatibility parameter data characteristics are fused in the dielectric loss calculation layer to evaluate the temperature rise effect of insulating materials in electromagnetic fields.

[0123] The process of generating a three-dimensional health feature vector includes:

[0124] The unified feature representation of the fused multi-source data is input into the physical constraint branch. The output of the counterfactual neural network is coordinated with the calculation results of the equipment physical failure model through the physical and data joint optimization layer to generate a three-dimensional health feature vector.

[0125] Among them, the physical constraint branch uses the laws of the equipment physical failure model to transform into a computable constraint branch, and uses physical logic to constrain the output of the data model to ensure that the health status assessment results conform to the data laws and do not violate the physical principles of actual equipment failure.

[0126] The physical constraint branch includes a failure probability calculation unit, a remaining lifetime prediction unit, and an availability assessment unit.

[0127] The failure probability calculation unit constructs the equipment physical failure equation based on the material fatigue cumulative damage theory.

[0128] The remaining lifetime prediction unit is a solver of the differential equations embedded in the component degradation rate;

[0129] Availability assessment unit, which is an environmental dynamics sub-model constructed by combining the environmental parameters of the current environment in which the equipment is located;

[0130] The three-dimensional health feature vector includes the first dimension, the failure probability value; the second dimension, the remaining lifetime value; and the third dimension, the availability value.

[0131] Among them, the first dimension of the failure probability value is a weighted fusion of the counterfactual neural network prediction probability and the calculation result of the equipment physical failure model;

[0132] The second dimension, remaining lifetime, is the intersection of the predicted value of the data-driven branch of the counterfactual neural network and the confidence interval of the solution to the differential equation.

[0133] The third dimension, availability, is the real-time operability coefficient output by the environmental dynamics sub-model.

[0134] Among them, the real-time operability coefficient is a quantitative indicator output by the environmental dynamics sub-model, which comprehensively evaluates whether the equipment can safely and effectively complete the mission under the current environment.

[0135] In this embodiment, the failure probability unit integrates a material micro-damage database, containing SN fatigue curves and crack propagation rate charts for 15 types of metal alloys. During calculation, the material grade of the equipment components is pre-matched, and the cumulative damage is queried based on stress spectrum data interpolation, then converted into failure probability using a Weibull distribution function. The remaining life unit incorporates degradation rate models for bearings and gear standard parts. After inputting current operating condition parameters, it solves differential equations to obtain the theoretical life curve and simultaneously calculates the 95% confidence interval boundaries. The availability unit accesses real-time terrain data, meteorological information, and enemy situation data. Through an environmental dynamics sub-model, it simulates and calculates the equipment's maximum maneuverability and range under the current environment, normalizing these parameters to an operability coefficient of 0-1.

[0136] In this embodiment, a two-way verification mechanism is set up in the physical and data joint optimization layer. The deviation between the counterfactual neural network's predicted probability and the equipment physical failure model's calculation result is less than 15%; otherwise, a review procedure is triggered. The remaining lifetime value is taken as the intersection range of the predicted value of the counterfactual neural network's data-driven branch and the solution of the equipment physical failure model. When the overlap between the two intervals is less than 60%, manual verification is initiated. Availability assessment incorporates feedback from the command system, backpropagating the real-time operability coefficient to optimize the assessment model. The generated three-dimensional feature vector includes confidence labels: failure probability labeled data - physical consistency level, remaining lifetime labeled interval width, and availability labeled environmental coverage. Feature calibration is performed hourly to ensure the uniformity and comparability of the dimensions.

[0137] Step S4: The step of outputting the risk score of the equipment is to fuse the causal significance feature value and the three-dimensional health feature vector into the equipment status risk feature matrix, input the equipment perception optimization model, and output the risk score of the equipment.

[0138] The process of constructing the equipment status risk characteristic matrix includes:

[0139] A spatiotemporally aligned feature fusion framework is established to synchronously match causal significance feature values ​​with three-dimensional health feature vectors according to timestamps;

[0140] The causal significance feature value and the three-dimensional health feature vector are combined by tensor splicing to form an original feature cube with three dimensions. The first dimension corresponds to the equipment subsystem classification, the second dimension corresponds to the time series window, and the third dimension corresponds to the feature type.

[0141] In this embodiment, a spatiotemporal cube architecture is used to organize feature data. The time dimension employs a sliding time window mechanism, with each window containing the data sequence of the current moment and the previous 7 sampling points. The spatial dimension is divided according to equipment subsystems, including 6 major functional modules. The feature dimension is divided into two main categories: causal features and health features, each with 12 specific indicators. Data alignment ensures time synchronization through a unified clock source for the equipment, with errors controlled within 1 millisecond. A layered filling strategy is used during feature stitching, with missing data filled using the moving average of the three most recent valid measurements.

[0142] The feature importance filtering layer eliminates redundant feature dimensions based on the feature weight template preset by the equipment model.

[0143] Among them, the feature weight template is a feature importance comparison table pre-set based on design specifications, historical failure data and industry experience for a specific equipment model. The feature weight template provides a clear quantitative standard for feature selection and quickly eliminates redundant features.

[0144] In this embodiment, the feature filtering layer loads the feature weight template provided by the equipment manufacturer, automatically reduces the dimensionality of features with importance <0.3, and retains the core 56-dimensional feature space.

[0145] In this embodiment, a triple attention network is set up in the feature cube processing. The spatial attention sub-network analyzes the fault transmission relationship between subsystems and assigns higher weights to systems with physical connections. The temporal attention sub-network identifies the trend of feature changes and assigns increasing attention to indicators that are continuously deteriorating. The feature attention sub-network evaluates the coefficient of variation of each indicator and dynamically adjusts the contribution ratio of different features. Each attention sub-network automatically updates its parameters weekly, using the operation and maintenance data of the most recent 30 days as the basis for adjustment. The compressed feature matrix is ​​subjected to robust normalization to eliminate the influence of different dimensions, and finally generates a 24×8×7 equipment status risk feature matrix for use by the prediction model.

[0146] The original feature cube is compressed using a spatiotemporal attention mechanism to output the equipment status risk feature matrix;

[0147] The process of obtaining the risk score for equipment includes:

[0148] The equipment status risk feature matrix is ​​input into the equipment perception optimization model based on ensemble learning, wherein the equipment perception optimization model is constructed using the Stacking ensemble framework.

[0149] The first layer integrates base learners, including gradient boosting decision trees, deep neural networks, and random forests, which respectively handle different dimensions of the equipment state risk feature matrix.

[0150] The Bayesian-optimized support vector regression model is used in the second-layer meta-learner, integrating the output of the base learners;

[0151] In this embodiment, the first-layer ensemble base learner includes a gradient boosting decision tree, a deep neural network, and a random forest. The gradient boosting decision tree has a maximum depth of 8, focusing on analyzing the combined effects between features. The deep neural network uses a 3-layer LSTM structure to capture temporal dependencies, and the random forest uses 200 decision trees to enhance model robustness. The various results output by the first-layer ensemble base learner are tested for confidence. A review process is triggered when the difference between the first-layer ensemble base learners exceeds 20%. The second-layer meta-learner uses a support vector regression model, automatically adjusting the kernel function parameters through Bayesian optimization. The optimization objective is to minimize the mean absolute error on the validation set. The support vector regression model is retrained monthly, retaining the three best historical versions as rollback backups.

[0152] The weight coefficients of each base learner are dynamically adjusted through a feature importance feedback mechanism, and a standardized risk score is output.

[0153] In this embodiment, a feature importance feedback loop is established to record the prediction accuracy of each learner for different types of faults. Specifically, for power system faults, the weight of the gradient boosting decision tree is increased by 0.1; for electronic system faults, the weight of the deep neural network is increased by 0.15. The weight adjustment range is dynamically calculated based on the rolling accuracy of the most recent 100 prediction results, with each adjustment being less than 10% of the base weight.

[0154] The raw score output undergoes a three-stage transformation: a benchmark curve is selected based on the equipment type to map the raw score to a standard range; a correction coefficient for the urgency of the current mission is added, relaxing the emergency level by 0.1; and an exponential smoothing filter is applied to eliminate the impact of instantaneous fluctuations. The score results are divided into 5 levels: 0-0.3 is normal (green indicator), 0.3-0.5 is of concern (blue indicator), 0.5-0.7 is a warning (yellow indicator), 0.7-0.9 is dangerous (orange indicator), and 0.9-1 is emergency (red indicator). Each level corresponds to a standardized contingency plan library, containing 12 typical response measures.

[0155] Step S5: Dynamically adjust monitoring frequency and maintenance level. Based on the equipment's risk score and task priority, dynamically adjust the monitoring frequency and maintenance response time, and iteratively optimize the maintenance strategy through a closed-loop feedback mechanism to achieve adaptive maintenance management throughout the equipment's life cycle.

[0156] Based on the equipment's risk score and task priority, the monitoring frequency and maintenance level are dynamically adjusted, including pre-constructing a task priority assessment matrix and classifying the equipment status according to the risk score interval.

[0157] The pre-constructed task priority evaluation matrix includes dividing the current task into three levels: Level I, Level II, and Level III. Each level corresponds to a different maintenance response coefficient, the specific value of which is set by experts.

[0158] Based on the risk score range, the equipment status is divided into three categories: emergency status, alert status, and normal status. Specifically, for emergency status, the score is >0.7 and the alarm indicator is set to red; for alert status, the score range is 0.4-0.7 and the alarm indicator is set to yellow; and for normal status, the score range is 0-0.4 and the alarm indicator is set to green.

[0159] Determine the monitoring frequency: Monitoring frequency = base value × maintenance response coefficient × status coefficient, where the base value is a preset value, and the status coefficient is 1.5 for emergency status, 1.2 for alert status, and 1.0 for normal status.

[0160] Determine the maintenance response time: Maintenance response time = base value / (maintenance response coefficient × status coefficient), where the base value is a preset value, and the status coefficient is 1.5 for emergency status, 1.2 for alert status, and 1.0 for normal status.

[0161] Dynamically adjusting monitoring parameters includes adjusting the sensor sampling rate, data transmission interval, and feature analysis depth;

[0162] In this embodiment, when entering alert status, the vibration sensor sampling rate is increased from 1kHz to 5kHz, and the temperature monitoring interval is shortened from 5 minutes to 1 minute. Under Level I mission conditions, electromagnetic compatibility monitoring is changed from timed sampling to continuous monitoring, and the feature analysis depth is increased from the conventional 5-layer wavelet decomposition to 7 layers. Resource availability checks are performed before each adjustment, and data compression mode is automatically activated when communication bandwidth is insufficient to ensure real-time transmission of key indicators. All parameter adjustment records are included in the equipment's electronic log.

[0163] The closed-loop feedback mechanism iterative optimization and maintenance strategy includes:

[0164] Construct a knowledge graph of equipment health status evolution to record status change data throughout the entire process from initial service to final retirement.

[0165] In this embodiment, the knowledge graph nodes comprise three parts: an equipment structure tree, a fault mode library, and a maintenance history. The equipment structure tree is structured at three levels: system, subsystem, and component, with over 200 key nodes labeled. The fault mode library includes 137 verified typical fault cases, each labeled with its preceding characteristics, development patterns, and handling methods. The maintenance history is stored using blockchain technology, recording the decision-making basis, implementation process, and effect verification for each maintenance operation. The knowledge graph is automatically updated weekly; new nodes require engineer review and confirmation, and a visual query interface is provided, supporting the tracing of equipment status evolution along a timeline.

[0166] Establish a maintenance strategy feedback loop to compare and analyze the results of each maintenance with the expected goals;

[0167] The maintenance decision rule base is updated based on the comparison results, the logic for generating subsequent maintenance strategies is optimized, and the optimal maintenance scheme for the current stage is automatically matched.

[0168] In this embodiment, each maintenance operation generates an effectiveness evaluation report, including three core indicators: fault confirmation, timeliness of handling, and resource consumption. The actual results are compared with predicted values ​​to calculate a strategy accuracy score. When three consecutive scores are below a threshold, a rule base optimization process is triggered: analyzing the causes of deviations, distinguishing between data quality issues, model errors, or rule defects; adjusting relevant parameters accordingly, such as modifying status level thresholds or optimizing task coefficients; and validating the effectiveness of the new rules in a digital twin environment. Major rule modifications require approval from the equipment management department before deployment.

[0169] In the early stages of equipment service, the focus is on accumulating baseline data, with monitoring frequency set at 1.5 times the standard value. During the mid-term, the emphasis shifts to failure mode learning, generating monthly degradation trend analysis reports. As the equipment approaches retirement, monitoring of critical components is strengthened, improving the maintenance response rate by 20%.

[0170] Example 2:

[0171] like Figure 2 As shown, this embodiment also provides an equipment predictive maintenance management system based on multi-feature fusion, including a multi-source data acquisition module 1, a causal feature analysis module 2, a three-dimensional health feature vector generation module 3, an equipment risk score output module 4, and a dynamic adjustment module for monitoring frequency and maintenance level 5.

[0172] Multi-source data acquisition module 1 collects multi-source data in real time during equipment operation;

[0173] Multi-source data includes electrical parameter data, mechanical vibration data, temperature distribution data, and electromagnetic compatibility parameter data;

[0174] The electrical parameters of the equipment's electrical system are collected using a distributed sensor network.

[0175] Mechanical vibration data of the equipment is acquired using a three-axis MEMS accelerometer array;

[0176] Temperature distribution data were acquired using infrared thermal imaging and embedded thermocouples.

[0177] Electromagnetic compatibility parameter data of the equipment are collected using a broadband electric field probe and a magnetic field sensor.

[0178] Causal feature analysis module 2 extracts the causal contribution of multi-source data to equipment failure and generates causal significance feature values;

[0179] The module 3 for generating three-dimensional health feature vectors uses a counterfactual neural network model to fuse multi-source data and combine it with the physical failure model of equipment to generate three-dimensional health feature vectors containing failure probability values, remaining life values, and availability values.

[0180] The risk scoring module 4 for output equipment integrates causal significance feature values ​​and three-dimensional health feature vectors into an equipment status risk feature matrix, inputs the equipment perception optimization model, and outputs the risk score of the equipment.

[0181] The module 5 dynamically adjusts the monitoring frequency and maintenance level based on the equipment's risk score and task priority. It also iteratively optimizes the maintenance strategy through a closed-loop feedback mechanism to achieve adaptive maintenance management throughout the equipment's entire life cycle.

[0182] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The methods disclosed in the embodiments are described simply because they correspond to the systems disclosed in the embodiments; relevant details can be found in the method section.

[0183] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0184] In the embodiments provided by this invention, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0185] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0186] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit.

[0187] Similarly, in the various embodiments of the present invention, each processing unit can be integrated into a functional module, or each processing unit can exist physically, or two or more processing units can be integrated into a functional module.

[0188] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0189] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0190] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A predictive maintenance management method for equipment based on multi-feature fusion, characterized in that, Includes the following steps: Step S1: The step of collecting multi-source data, which involves collecting multi-source data in real time during equipment operation; Step S2: The steps of causal feature analysis include extracting the causal contribution of multi-source data to equipment failure, constructing a causal directed graph, and generating causal significance feature values. Generating causal significance feature values ​​includes: identifying the directed paths from the sensor nodes used to collect multi-source data to the fault nodes in the causal directed graph; The magnitude of the change in conditional probability between nodes on each path is statistically analyzed and used as the causal strength of the path. By combining the causal strength of all paths pointing to the same fault node, the reliability decay coefficient of the environmental parameters of the current equipment is collected, and causal significance feature values ​​are generated. Step S3: The step of generating a three-dimensional health feature vector is to fuse multi-source data through a counterfactual neural network model and combine it with the physical failure model of the equipment to generate a three-dimensional health feature vector containing failure probability value, remaining life value and availability value. Among them, the counterfactual neural network model introduces a causal attention network into the output layer of each sub-network, dynamically adjusts the feature weights based on the causal significance feature value, inputs the unified feature representation of the fused multi-source data into the physical constraint branch, and coordinates the output of the counterfactual neural network with the calculation results of the equipment physical failure model through the physical and data joint optimization layer to generate a three-dimensional health feature vector; Step S4: The step of outputting the risk score of the equipment is to fuse the causal significance feature value and the three-dimensional health feature vector into the equipment status risk feature matrix, input the equipment perception optimization model, and output the risk score of the equipment. The process of integrating these features into an equipment status risk characteristic matrix includes: A spatiotemporally aligned feature fusion framework is established to synchronously match causal significance feature values ​​with three-dimensional health feature vectors according to timestamps; The causal significance feature value and the three-dimensional health feature vector are combined by tensor splicing to form an original feature cube with three dimensions. The first dimension corresponds to the equipment subsystem classification, the second dimension corresponds to the time series window, and the third dimension corresponds to the feature type. The feature importance filtering layer eliminates redundant feature dimensions based on the feature weight template preset by the equipment model. The original feature cube is compressed using a spatiotemporal attention mechanism to output the equipment status risk feature matrix; The output equipment risk score includes: inputting the equipment status risk feature matrix into the equipment perception optimization model based on ensemble learning, wherein the equipment perception optimization model is constructed using the Stacking ensemble framework; The first layer integrates base learners, including gradient boosting decision trees, deep neural networks, and random forests, which respectively handle different dimensions of the equipment state risk feature matrix. The Bayesian-optimized support vector regression model is used in the second-layer meta-learner, integrating the output of the base learners; The weight coefficients of each base learner are dynamically adjusted through a feature importance feedback mechanism, and a standardized risk score is output. Step S5: Dynamically adjust monitoring frequency and maintenance level. Based on the equipment's risk score and task priority, dynamically adjust the monitoring frequency and maintenance response time, and iteratively optimize the maintenance strategy through a closed-loop feedback mechanism to conduct adaptive maintenance management for the entire equipment life cycle. The mathematical expression for monitoring frequency is: Monitoring frequency = Base value × Maintenance response coefficient × Status coefficient, where the base value is a preset value, and the status coefficient is 1.5 for emergency status, 1.2 for alert status, and 1.0 for normal status. The mathematical expression for maintenance response time is: Maintenance response time = Base value / (Maintenance response coefficient × State coefficient), where the base value is a preset value, and the state coefficient is 1.5 for emergency state, 1.2 for alert state, and 1.0 for normal state; Dynamically adjusting monitoring parameters includes adjusting the sensor sampling rate, data transmission interval, and feature analysis depth; The closed-loop feedback mechanism iterative optimization and maintenance strategy includes: Construct a knowledge graph of equipment health status evolution to record status change data throughout the entire process from initial service to final retirement. Establish a maintenance strategy feedback loop to compare and analyze the results of each maintenance with the expected goals; The maintenance decision rule base is updated based on the comparison results, the logic for generating subsequent maintenance strategies is optimized, and the optimal maintenance scheme for the current stage is automatically matched.

2. The equipment predictive maintenance management method based on multi-feature fusion according to claim 1, characterized in that, The multi-source data includes electrical parameter data, mechanical vibration data, temperature distribution data, and electromagnetic compatibility parameter data.

3. The equipment predictive maintenance management method based on multi-feature fusion according to claim 1, characterized in that, Extracting the causal contribution of multi-source data to equipment failure includes pre-acquiring the environmental parameters of the current equipment environment, a pre-defined environmental correction rule base, and constructing a causal directed graph. The environmental parameters currently in which the equipment is located include electromagnetic environment intensity data, atmospheric temperature gradient data, and vibration spectrum characteristic data. Electromagnetic environment intensity data is collected by an electromagnetic field measuring instrument, atmospheric temperature gradient data is collected by a radiosonde or weather tower, and vibration spectrum characteristic data is collected by a triaxial MEMS accelerometer and a dynamic signal analyzer. The environmental correction rule base includes the reliability decay coefficients for different equipment models under different combinations of environmental parameters. The reliability decay coefficients range from 0.1 to 1.0 in decimal. The multi-source data collected in step S1 is weighted and fused with the credibility attenuation coefficient to obtain the feature sequence of environment adaptation. A causal discovery algorithm is used to analyze the conditional independence relationships between feature sequences of environment adaptation and to construct a causal directed graph.

4. The equipment predictive maintenance management method based on multi-feature fusion according to claim 1, characterized in that, The counterfactual neural network model is a neural network architecture with multimodal input layers, in which multi-source data is processed through independent feature extraction subnetworks; The multi-source data is nonlinearly coupled through a cross-modal feature interaction layer, outputting a unified feature representation of the fused multi-source data, and an adversarial training strategy is used to optimize the network parameters. The process of generating a three-dimensional health feature vector includes: The unified feature representation of the fused multi-source data is input into the physical constraint branch. The output of the counterfactual neural network is coordinated with the calculation results of the equipment physical failure model through the physical and data joint optimization layer to generate a three-dimensional health feature vector. Among them, the physical constraint branch is transformed into a computable constraint branch by utilizing the laws of the equipment physical failure model. The physical constraint branch includes a failure probability calculation unit, a remaining life prediction unit, and an availability assessment unit. The failure probability calculation unit constructs the equipment physical failure equation based on the material fatigue cumulative damage theory. The remaining lifetime prediction unit is a solver of the differential equations embedded in the component degradation rate; Availability assessment unit, which is an environmental dynamics sub-model constructed by combining the environmental parameters of the current environment in which the equipment is located; The three-dimensional health feature vector includes the first dimension, the failure probability value; the second dimension, the remaining lifetime value; and the third dimension, the availability value. Among them, the first dimension of the failure probability value is a weighted fusion of the counterfactual neural network prediction probability and the calculation result of the equipment physical failure model; The second dimension, remaining lifetime, is the intersection of the predicted value of the data-driven branch of the counterfactual neural network and the confidence interval of the solution to the differential equation. The third dimension, availability, is the real-time operability coefficient output by the environmental dynamics sub-model.

5. The equipment predictive maintenance management method based on multi-feature fusion according to claim 1, characterized in that, Based on the equipment's risk score and task priority, the monitoring frequency and maintenance level are dynamically adjusted, including pre-constructing a task priority assessment matrix and classifying the equipment status according to the risk score interval. The pre-constructed task priority evaluation matrix includes dividing the current task into three levels: Level I, Level II, and Level III. Each level corresponds to a different maintenance response coefficient, the specific value of which is set by experts. Based on the risk score range, the equipment status is divided into three categories: emergency status, alert status, and normal status. In the emergency status, the score is >0.7 and the alarm indicator is set to red. In the alert status, the score range is 0.4-0.7 and the alarm indicator is set to yellow. In the normal status, the score range is 0-0.4 and the alarm indicator is set to green.

6. A system for the equipment predictive maintenance management method based on multi-feature fusion as described in claim 1, characterized in that, It includes a multi-source data acquisition module, a causal feature analysis module, a three-dimensional health feature vector generation module, a risk scoring module for output equipment, and a module for dynamically adjusting monitoring frequency and maintenance level; The multi-source data acquisition module collects multi-source data in real time during equipment operation; The causal feature analysis module extracts the causal contribution of multi-source data to equipment failure and generates causal significance feature values. The module for generating a three-dimensional health feature vector generates a three-dimensional health feature vector containing fault probability value, remaining life value and availability value by fusing multi-source data through a counterfactual neural network model and combining the physical failure model rules of equipment. The risk scoring module of the output equipment integrates the causal significance feature value and the three-dimensional health feature vector into an equipment status risk feature matrix, inputs it into the equipment perception optimization model, and outputs the risk score of the equipment. The module for dynamically adjusting monitoring frequency and maintenance level adjusts the monitoring frequency and maintenance level based on the equipment's risk score and task priority, and iteratively optimizes the maintenance strategy through a closed-loop feedback mechanism to perform adaptive maintenance management throughout the equipment's entire life cycle.

7. The system according to claim 6, characterized in that, The multi-source data includes electrical parameter data, mechanical vibration data, temperature distribution data, and electromagnetic compatibility parameter data; The electrical parameters of the equipment's electrical system are collected using a distributed sensor network. Mechanical vibration data of the equipment is acquired using a three-axis MEMS accelerometer array; Temperature distribution data were acquired using infrared thermal imaging and embedded thermocouples. Electromagnetic compatibility parameter data of the equipment are collected using a broadband electric field probe and a magnetic field sensor.

Images