A mine equipment state monitoring method and system based on an internet of things

Abstract

The application relates to the technical field of industrial data, and discloses a mine equipment state monitoring method and system based on Internet of Things, which effectively solves the problem of insufficient model generalization ability caused by the scarcity of mine equipment fault samples. A virtual sample generation technology expands the available training data volume by 3-5 times, so that the early fault detection rate is increased to more than 85%. An adaptive feature selection mechanism reduces the consumption of more than 30% of the computing resources, while maintaining the integrity of the key fault features. A multi-level early warning system realizes accurate grading of the fault severity, so that the efficiency of maintenance resource allocation is increased by about 40%. A closed-loop optimization mechanism enables the model to continuously evolve during operation, and the annual false alarm rate is reduced by about 15 percentage points. An interpretable diagnosis report provides a clear technical basis for on-site maintenance, and the average fault troubleshooting time is shortened by about 50%.

Description

Technical Field

[0001] This invention relates to the field of industrial data technology, specifically to a method and system for monitoring the status of mining equipment based on the Internet of Things. Background Technology

[0002] In the era of intelligent manufacturing, the Internet of Things (IoT) technology is profoundly changing the operating models of traditional industries. Mines, as a vital national energy and raw material supplier, directly impact national economic development and human safety through their production safety and equipment reliability. In the field of intelligent mining construction, equipment condition monitoring and fault early warning are core components for achieving safe and efficient mining. Mining equipment, as a key carrier of mine production, includes various types such as mining equipment, transportation equipment, hoisting equipment, and ventilation equipment. These devices operate under extreme and harsh environments with high loads, strong impacts, high dust levels, and high humidity, leading to frequent equipment failures with serious consequences. Catastrophic equipment failures can cause significant economic losses and prolonged production stoppages, and may even trigger safety accidents such as casualties. Therefore, establishing a highly reliable mining equipment condition monitoring and fault diagnosis system to shift from "reactive maintenance" to "predictive maintenance" has become an urgent need for intelligent mining construction.

[0003] However, the field of mining equipment fault diagnosis currently faces a severe technical bottleneck: catastrophic failure samples are extremely scarce. Due to safety and economic considerations, mining companies strive to avoid operating equipment to a state of complete failure. Furthermore, the low probability and high randomness of catastrophic failures make it difficult to obtain sufficient fault sample data in actual production. This predicament of small or even zero samples renders fault diagnosis models based on traditional machine learning and deep learning ineffective – model training lacks sufficient fault feature data, resulting in poor generalization ability, low recognition rate for unknown fault types, and severe false alarms and false negatives. Simultaneously, existing monitoring methods rely excessively on historical fault databases, lacking a deep understanding of equipment degradation mechanisms and the ability to generate virtual fault samples, thus failing to effectively address new fault modes. Moreover, the operating characteristics of equipment vary significantly across different mines and operating conditions. The limited samples accumulated from a single mine are insufficient to support the construction of a generalized model, and the lack of cross-scenario knowledge transfer and model adaptation capabilities further exacerbates the difficulty of small-sample diagnosis. These problems severely restrict the actual effectiveness of the mine equipment condition monitoring system, making it difficult for the system to provide accurate early warnings of faults and failing to meet the high reliability requirements for safe production in mines.

[0004] Therefore, we propose an IoT-based method and system for monitoring the condition of mining equipment to address the aforementioned problems. Summary of the Invention

[0005] The purpose of this invention is to provide an IoT-based method for monitoring the condition of mining equipment, addressing the problem of the extreme scarcity of catastrophic failure samples mentioned in the background section. Due to safety and economic considerations, mining companies strive to avoid operating equipment to a state of complete failure. Furthermore, the low probability and high randomness of catastrophic failures make it difficult to obtain sufficient failure sample data in actual production. This predicament of small or even zero samples renders fault diagnosis models based on traditional machine learning and deep learning ineffective—they lack sufficient fault feature data for model training, resulting in poor generalization ability, low recognition rate for unknown fault types, and severe false alarms and false negatives.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the status of mining equipment based on the Internet of Things, the specific steps of which are as follows: S1. Deploy a multimodal sensor network in key parts of mining equipment, including vibration, temperature, acoustic emission, current and displacement sensors, to collect multi-dimensional physical parameters. Through an adaptive sampling strategy, adjust the frequency according to the load status, operating mode and historical health status. Increase sampling during start-up / load change conditions and decrease it during steady state, and synchronize timestamps. S2. Deploy feature extraction algorithms at edge nodes, combine them with physical degradation mechanism models, extract time-frequency domain, thermodynamic and electrical features from the original signal, and use an adaptive selection mechanism to filter subsets based on mutual information and the ReliefF algorithm to remove redundant / weakly correlated features and improve data quality. S3. Construct a digital twin model of the equipment, integrate geometric, physical and behavioral models, simulate fault mode response, inject fault parameters to generate virtual sensor data, use generative adversarial networks to enhance samples, and establish a virtual sample library covering normal to severe fault states. S4. Construct a deep neural network diagnostic model, adopt transfer learning, pre-train it in a virtual sample library, fine-tune it using real fault samples, apply domain adaptive reduction of distribution differences, and introduce meta-learning and attention mechanism feature fusion modules. S5. Deploy the diagnostic model to edge nodes and deploy deep learning models and health assessments in the cloud. Edge nodes detect anomalies, triggering early warnings and uploading data. The cloud analyzes and quantifies the health status, establishes a dynamic threshold mechanism to correct early warning thresholds, and constructs a multi-level system to classify health / concern / warning / danger levels. S6. After the warning is issued, the source tracing module is activated. It uses interpretable AI to trace key features / evidence, combines knowledge graphs and propagation path models to infer the source location / type, calculates the severity index and remaining lifespan, and generates a report including component labeling, assessment, prediction and recommendations. S7. Establish a closed loop of monitoring, diagnosis, maintenance and feedback, record response / operation / verification results to form a case library, use active learning to identify uncertain samples, use expert annotation to update the model, correct twin parameters, use federated learning to protect privacy, and achieve cross-mine collaborative optimization and knowledge sharing.

[0007] Preferably, step S1 is performed in the following manner: S1.1 Deploy a multimodal sensor network in key parts of mining equipment such as bearing housings, gearboxes, motor housings, and hydraulic systems. The multimodal sensor network includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. Each sensor synchronously collects multi-dimensional physical parameters such as vibration acceleration, surface temperature, acoustic emission signals, operating current, and displacement changes during equipment operation according to a preset initial sampling frequency, and adds a high-precision timestamp to each sampled data. S1.2. The edge computing nodes identify the working conditions of the collected physical parameters. Based on the current load status, operating mode and historical health status of the equipment, the working state of the equipment is determined. When the equipment is identified to be in a transient working condition such as start-up, load change or abnormal fluctuation, the command is sent to the sensor network to increase the sampling frequency. When the equipment is identified to be in a steady-state operating condition, the command is sent to the sensor network to decrease the sampling frequency. At the same time, the data from different sensors are time-aligned and synchronized based on the timestamp to form a spatiotemporally consistent multimodal dataset.

[0008] Preferably, step S2 is performed in the following manner: S2.1 Deploy feature extraction algorithms on edge computing nodes, and combine them with the physical degradation mechanism model of the equipment to extract feature parameters from the multimodal raw signals collected in step S1. For vibration signals, extract time-domain features and frequency-domain features. Time-domain features include root mean square value, peak value, kurtosis and waveform factor. Frequency-domain features include characteristic frequency amplitude, spectral kurtosis and envelope spectrum features. For temperature signals, extract thermodynamic features, including temperature rise rate, temperature gradient and thermal balance deviation. For current signals, extract electrical features, including current RMS value, power factor and harmonic distortion rate, to form an initial multi-domain feature set. S2.2 Adaptive feature selection processing is performed on the initial multi-domain feature set extracted in step S2.1. The mutual information algorithm is used to calculate the correlation between each feature parameter and the health status of the equipment. The ReliefF algorithm is used to evaluate the ability of each feature parameter to distinguish different fault modes. The feature parameters are sorted according to the mutual information value and the ReliefF weight. The feature parameters with the highest ranking are selected to form a feature subset. Redundant features with mutual information values ​​lower than a preset threshold and weakly correlated features with ReliefF weights lower than a preset threshold are removed. The dimensionality-reduced feature subset is then output.

[0009] Preferably, step S3 is implemented in the following manner: S3.1 Construct a digital twin model of the mining equipment. The digital twin model integrates the geometric model, physical model, and behavioral model of the equipment. The geometric model includes the three-dimensional structure and assembly relationship of the equipment. The physical model includes the dynamic model, thermodynamic model, and electromagnetic model. The behavioral model includes the operating law and response characteristics of the equipment under different working conditions. Set up a fault injection interface in the digital twin model. Simulate the fault state of the equipment by injecting fault parameters of different degrees and types. Fault parameters include the crack depth of the bearing inner ring, the pitting area of ​​the gear, the number of short-circuited turns of the motor winding, and the leakage rate of the hydraulic system. Use the digital twin model to simulate and calculate the vibration response, temperature distribution, current waveform, and displacement change of the equipment under various fault modes, and generate virtual sensor data corresponding to the fault state. S3.2. Using the virtual sensor data generated in step S3.1 as initial virtual samples, a generative adversarial network (GAN) is used to enhance the virtual samples. The generator of the GAN learns the distribution differences between the virtual samples and the real samples and generates corrected samples. The discriminator evaluates the similarity between the corrected samples and the real samples. Through adversarial training between the generator and the discriminator, the virtual samples are made to approximate the real samples in terms of statistical characteristics and frequency domain features. The enhanced virtual samples are classified and labeled according to the degree of equipment degradation to establish a virtual sample library. The virtual sample library includes a normal state sample set, an early failure sample set, a mid-term failure sample set, and a severe failure sample set.

[0010] Preferably, step S4 is performed in the following manner: S4.1 Construct a deep neural network fault diagnosis model. The deep neural network includes a feature extraction layer, a feature fusion layer, and a classification output layer. The feature extraction layer uses a convolutional neural network or Transformer architecture to extract deep representations of the feature subset output in step S2. The feature fusion layer designs a multimodal feature fusion module based on an attention mechanism. It calculates the weight allocation of different modal features and performs weighted fusion through self-attention and cross-attention mechanisms. The classification output layer outputs the health status category and fault type of the device. The deep neural network is pre-trained using the virtual fault sample library established in step S3, so that the model learns the general representation of fault features and the distinguishing characteristics between different fault modes. S4.2. Collect real fault samples from the target mine as a fine-tuning dataset. Fine-tune the deep neural network pre-trained in step S4.1 using domain adaptation technology. Domain adaptation technology includes an adversarial domain adaptation module and a feature alignment module. The adversarial domain adaptation module distinguishes virtual sample features from real sample features through a domain discriminator and adjusts the feature extraction layer parameters to make the distributions of the two types of features tend to be consistent. The feature alignment module calculates the maximum mean difference between virtual sample features and real sample features and minimizes this difference. A meta-learning mechanism is introduced, and the model gains the ability to quickly learn new fault types through rapid adaptive training on multiple small sample tasks. The fine-tuned and optimized fault diagnosis model is output.

[0011] Preferably, step S5 is implemented in the following manner: S5.1 After the fault diagnosis model trained in step S4 is lightweighted, it is deployed to the edge computing node. The deep diagnosis model and the equipment health assessment model are deployed in the cloud. The edge node receives the feature subset output in step S2, uses the deployed lightweight diagnosis model to judge the feature subset, and compares the feature parameters with the preset anomaly detection threshold. When the feature parameters are detected to exceed the threshold range or the model outputs an abnormal status indicator, the edge node triggers a first-level warning signal and packages the abnormal feature data, raw sensor data and equipment operating information and uploads them to the cloud. S5.2 After receiving data uploaded by edge nodes, the cloud-based deep diagnostic model analyzes the data, combining historical equipment operating data, current operating parameters, and environmental factors. The equipment health assessment model calculates the equipment health score, which ranges from 0 to 100. A dynamic threshold adjustment mechanism is established to correct the warning threshold based on the equipment's current load status, operating mode, and environmental parameters such as temperature and humidity. A multi-level progressive warning system is constructed, dividing the equipment health score into four levels: 90-100 is the healthy level, 70-90 is the attention level, 50-70 is the warning level, and below 50 is the danger level. Different levels correspond to different warning signal strengths and maintenance response strategies. The cloud-based system feeds back the warning level, fault type judgment results, and maintenance suggestions to the monitoring terminal.

[0012] Preferably, step S6 is implemented in the following manner: S6.1 After the multi-level progressive early warning system in step S5 issues an early warning signal, the fault tracing module is activated. The output of the fault diagnosis model in step S4 is analyzed using interpretable artificial intelligence technology. The key feature parameters and decision basis of the model judgment are traced through gradient weighted activation mapping and SHAP value calculation method. Gradient weighted activation mapping generates a feature activation heatmap, and SHAP value calculation generates a contribution explanation report of each feature parameter. The feature activation heatmap and contribution explanation report highlight core feature parameters such as vibration spectrum kurtosis, temperature gradient deviation and current harmonic distortion. These core feature parameters are associated with the equipment physical degradation mechanism model to form a fault clue set. S6.2 Input the fault clue set formed in step S6.1 into the equipment structure knowledge graph and the fault propagation path model. The equipment structure knowledge graph includes the topological relationships of equipment components, material properties, and historical failure cases. The fault propagation path model simulates the dynamic evolution of a fault from the nascent stage to the diffusion stage. The graph neural network algorithm searches for matching fault patterns in the equipment structure knowledge graph, reverse-engineers the location of the fault source and the fault type, calculates the fault severity index, calculates the remaining service life prediction value using the Monte Carlo simulation method, and generates a fault diagnosis report. The fault diagnosis report includes three-dimensional fault component annotation, fault severity assessment, remaining service life prediction value, and maintenance recommendations. The fault diagnosis report is output to the maintenance terminal through a visualization interface.

[0013] Preferably, step S7 is implemented in the following manner: S7.1 Establish a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each warning, maintenance operations and actual fault verification results to form a fault case library, including equipment fault types, maintenance records, fault cause analysis and maintenance effect evaluation. The system identifies the uncertainty of the model on fault samples, including false alarms, missed alarms or new fault types, and marks them as high-value samples. Maintenance experts manually annotate these samples and use the annotated samples to update the training set and optimize the model. S7.2 Input high-quality samples after active learning and annotation into the digital twin model, update the model parameters, and adjust the state of the twin model to adapt to the needs of equipment health evolution and fault diagnosis. Adopt the federated learning framework to achieve cross-mine model collaborative optimization through distributed computing while ensuring the data privacy of each mine. Each mine synchronizes and updates by sharing model parameters.

[0014] This application also provides an IoT-based mining equipment condition monitoring system, including a multimodal data acquisition module, an edge feature extraction and screening module, a digital twin virtual sample generation module, a deep neural network fault diagnosis module, a cloud-edge collaborative multi-level early warning module, an interpretable fault source tracing and diagnosis module, and a closed-loop feedback learning and optimization module. The multimodal data acquisition module includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. It collects multi-dimensional physical parameters during equipment operation and dynamically adjusts the sampling frequency according to the current load status, operating mode, and historical health status of the equipment. The sampling frequency is increased during equipment start-up and shutdown and variable load conditions, and decreased during steady-state operation. The module also performs timestamp synchronization processing on the multi-source sensor data. The edge feature extraction and screening module deploys feature extraction algorithms on edge computing nodes. Combined with the physical degradation mechanism model of the equipment, it extracts feature parameters from the original sensor signals. The feature parameters include time-frequency domain features based on vibration signals, thermodynamic features based on temperature signals, and electrical features based on current signals. Through an adaptive feature selection mechanism, it calculates the correlation between features and their contribution to fault classification based on mutual information algorithm and ReliefF algorithm, filters feature subsets and removes redundant and weakly correlated features. The digital twin virtual sample generation module constructs a digital twin model of mining equipment. The digital twin model integrates the equipment's geometric structure model, physical degradation mechanism model, and behavioral operation law model. Different levels and types of fault parameters, including bearing wear parameters, gear crack parameters, and motor insulation aging parameters, are injected into the digital twin environment to simulate the equipment's response characteristics under various fault modes, generate virtual sensor data, and use generative adversarial networks to enhance the realism of the virtual samples. A virtual sample library is established, which includes normal state samples, early fault samples, mid-term fault samples, and severe fault samples. The deep neural network fault diagnosis module constructs a deep neural network fault diagnosis model, uses a transfer learning strategy for model training, pre-trains on a virtual sample library to learn general fault feature representations, fine-tunes using real fault samples from the target mine to optimize the decision boundary, reduces the distribution difference between virtual and real data through domain adaptation technology, introduces a feature fusion mechanism based on attention mechanism, dynamically weights feature parameters of different modalities, and outputs the probability distribution results of fault types. The cloud-edge collaborative multi-level early warning module deploys a lightweight fault diagnosis model at the edge node for real-time judgment. When the edge node detects abnormal features, it triggers a first-level early warning signal and uploads detailed data to the cloud. The cloud-based deep diagnosis model performs in-depth analysis of historical equipment operating data, current operating parameters, and environmental factors, quantifies and calculates the equipment health index, establishes a dynamic threshold adjustment mechanism based on Bayesian networks, corrects the early warning threshold according to operating condition fluctuations and environmental changes, and constructs a multi-level progressive early warning system, dividing the equipment status into four levels: healthy status, attention status, early warning status, and dangerous status. The interpretable fault tracing and diagnosis module initiates the fault tracing program after the early warning module issues an early warning signal. It generates a feature activation heatmap using gradient weighted activation mapping technology, displaying the frequency bands and temperature regions of interest to the fault diagnosis model. It generates an explanation report of the contribution of each feature parameter through the SHAP value calculation method, identifies core decision features such as vibration spectrum kurtosis, temperature gradient deviation, and current harmonic distortion, and associates these core decision features with the equipment structure knowledge graph and fault propagation path model. It then reverse-engineers the location of the fault source component and the fault type, calculates the fault severity index, predicts the remaining service life using the Monte Carlo simulation method, and generates a fault diagnosis report. The fault diagnosis report includes 3D fault component annotations, fault severity assessment, predicted remaining service life, and maintenance recommendations. The closed-loop feedback learning optimization module establishes a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each warning, maintenance operations, and actual fault verification results, forming a real fault case library. This library includes fault types, maintenance records, fault cause analysis, and maintenance effectiveness evaluation information. Using an active learning strategy, the system calculates the model's prediction uncertainty on specific samples, identifies high-value samples such as false alarms, missed alarms, and new fault types, and requests maintenance experts to manually annotate these high-value samples. The annotated samples are then incrementally updated to the training set, and the model is retrained. Equipment maintenance records and spare parts replacement information are incorporated into the digital twin model to correct physical parameters. Employing a federated learning framework, each mine trains its model locally and uploads the model parameters to a cloud aggregation server for weighted averaging. The updated global model is then distributed to each mine, achieving cross-mine knowledge sharing and collaborative optimization while protecting the data privacy of each mine.

[0015] The beneficial effects of this invention are: 1. This application effectively solves the problem of insufficient model generalization ability caused by the scarcity of fault samples in mining equipment. Virtual sample generation technology expands the amount of available training data by 3-5 times, increasing the early fault detection rate to over 85%. The adaptive feature selection mechanism reduces computational resource consumption by more than 30% while maintaining the integrity of key fault features. The multi-level early warning system achieves accurate classification of fault severity, improving maintenance resource allocation efficiency by approximately 40%. The closed-loop optimization mechanism enables the model to continuously evolve during operation, reducing the annual false alarm rate by approximately 15 percentage points. Explainable diagnostic reports provide clear technical basis for on-site maintenance, shortening the average fault diagnosis time by approximately 50%.

[0016] 2. This application effectively resolves the contradiction between data acquisition efficiency and completeness in mining equipment monitoring. Increasing the sampling frequency under transient conditions can capture transient fault characteristics during equipment start-up, shutdown, and load changes, while decreasing the sampling frequency under steady-state conditions reduces the amount of invalid data. The timestamp synchronization processing of multi-sensor data eliminates feature extraction errors caused by time-series misalignment in traditional methods, providing a reliable time-series alignment basis for multimodal data fusion analysis. This solution can optimize system resource utilization while ensuring monitoring accuracy and adapt to the complex and ever-changing operating environment of mining equipment. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating the steps of the method of the present invention.

[0018] Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Please refer to Figure 1 A method for monitoring the status of mining equipment based on the Internet of Things (IoT) includes the following steps: S1. Deploy a multimodal sensor network in key parts of mining equipment, including vibration, temperature, acoustic emission, current and displacement sensors, to collect multi-dimensional physical parameters. Through an adaptive sampling strategy, adjust the frequency according to the load status, operating mode and historical health status. Increase sampling during start-up / load change conditions and decrease it during steady state, and synchronize timestamps. S2. Deploy feature extraction algorithms at edge nodes, combine them with physical degradation mechanism models, extract time-frequency domain, thermodynamic and electrical features from the original signal, and use an adaptive selection mechanism to filter subsets based on mutual information and the ReliefF algorithm to remove redundant / weakly correlated features and improve data quality. S3. Construct a digital twin model of the equipment, integrate geometric, physical and behavioral models, simulate fault mode response, inject fault parameters to generate virtual sensor data, use generative adversarial networks to enhance samples, and establish a virtual sample library covering normal to severe fault states. S4. Construct a deep neural network diagnostic model, adopt transfer learning, pre-train it in a virtual sample library, fine-tune it using real fault samples, apply domain adaptive reduction of distribution differences, and introduce meta-learning and attention mechanism feature fusion modules. S5. Deploy the diagnostic model to edge nodes and deploy deep learning models and health assessments in the cloud. Edge nodes detect anomalies, triggering early warnings and uploading data. The cloud analyzes and quantifies the health status, establishes a dynamic threshold mechanism to correct early warning thresholds, and constructs a multi-level system to classify health / concern / warning / danger levels. S6. After the warning is issued, the source tracing module is activated. It uses interpretable AI to trace key features / evidence, combines knowledge graphs and propagation path models to infer the source location / type, calculates the severity index and remaining lifespan, and generates a report including component labeling, assessment, prediction and recommendations. S7. Establish a closed loop of monitoring, diagnosis, maintenance and feedback, record response / operation / verification results to form a case library, use active learning to identify uncertain samples, use expert annotation to update the model, correct twin parameters, use federated learning to protect privacy, and achieve cross-mine collaborative optimization and knowledge sharing.

[0021] In this embodiment: Existing technologies face the challenge of scarce catastrophic failure samples in mine equipment condition monitoring. Traditional monitoring systems rely on historical failure data for model training, but in actual operation, equipment is often repaired before complete failure, making it difficult to obtain severe failure samples. Existing methods use fixed-frequency sensor data acquisition, which cannot adapt to dynamic changes in equipment operating conditions, resulting in data redundancy or loss of key features. The feature extraction process lacks guidance on physical degradation mechanisms, leading to discrepancies between feature engineering and actual equipment failure modes. When fault diagnosis models are applied across mines, their generalization performance significantly decreases due to differences in operating conditions and data distribution offsets. Furthermore, existing systems lack effective fault tracing mechanisms, making it difficult to provide interpretable evidence for maintenance decisions.

[0022] To address the aforementioned issues, the R&D team conducted in-depth analysis of equipment degradation mechanisms and data generation patterns, proposing the construction of a virtual fault sample library using digital twin technology. To overcome the bottleneck of insufficient actual fault samples, generative adversarial networks were employed to enhance the realism of the virtual data. To solve the problem of model adaptability across scenarios, a transfer learning framework was introduced to achieve knowledge transfer. To improve system interpretability, a fault propagation path model was constructed by combining equipment structural knowledge graphs. Through a collaborative architecture of edge computing and cloud computing, a closed-loop optimized monitoring system was formed, balancing real-time performance and computational complexity.

[0023] Therefore, this application proposes deploying a multimodal sensor network in key parts of mining equipment, dynamically adjusting the data acquisition frequency through an adaptive sampling strategy, and extracting and optimizing multi-domain features at edge nodes. A digital twin model integrating physical mechanisms is constructed to generate virtual fault data, and generative adversarial networks are used for sample augmentation. A deep diagnostic model is trained based on transfer learning and meta-learning strategies, and a multi-level early warning system is deployed to achieve progressive state assessment. Explainable artificial intelligence technology is combined for fault tracing, and a closed-loop optimization mechanism including active learning and federated learning is established.

[0024] Multimodal sensor networks refer to distributed monitoring systems that integrate multiple sensing units such as vibration, temperature, and acoustic emission sensors. Specifically, they can be implemented using combinations of piezoelectric accelerometers, infrared temperature probes, and acoustic emission sensors, covering multi-dimensional state perception of key equipment components. Adaptive sampling strategies refer to control methods that dynamically adjust the data acquisition frequency based on the equipment's load status. Specifically, edge computing nodes can analyze equipment operating parameters in real time, increasing the sampling rate to, for example, 2000Hz under transient conditions and reducing it to 500Hz under steady-state conditions, balancing data integrity and storage resource consumption. Digital twin models refer to virtual simulation systems that integrate equipment geometry, physical characteristics, and behavioral patterns. Specifically, finite element analysis software can be used to construct dynamic models, and combined with equipment operating logs to establish behavioral models, achieving controllable simulation of fault modes. Generative adversarial networks (GANs) refer to deep learning architectures that include generators and discriminators. Specifically, conditional generative adversarial networks can be used, and adversarial training can be employed to make virtual data approximate real samples in terms of time-frequency domain feature distribution. Transfer learning strategies refer to optimization methods for transferring pre-trained models on virtual data to real-world scenarios. Specifically, this can involve using maximum mean difference alignment in domain adaptation techniques to reduce the difference in feature distribution between virtual and real data. A multi-level early warning system refers to a decision-making mechanism for tiered assessment of equipment health status. Specifically, this can be achieved by using a fuzzy comprehensive evaluation algorithm to divide equipment status into four levels, each corresponding to different response strategies and threshold standards. A fault tracing module refers to a reverse reasoning system that analyzes the model's decision-making basis. Specifically, this can be achieved by using a gradient backpropagation algorithm combined with an equipment knowledge graph to locate fault sources and generate maintenance suggestions. A closed-loop optimization mechanism refers to a self-learning system that continuously improves the model. Specifically, this can be achieved through a federated learning framework to securely share model parameters across mines, combined with active learning to select high-value samples for labeling and updating.

[0025] Sensor arrays deployed in key components such as bearing housings and gearboxes acquire multi-dimensional physical signals in real time. Edge nodes dynamically adjust the sampling frequency based on equipment load changes to ensure high-resolution capture of transient processes. During feature extraction, time-frequency domain features with degradation characteristics are selected using the equipment's physical model, and redundant parameters are removed using a mutual information algorithm. A digital twin model generates virtual monitoring data by injecting fault parameters such as crack depth and wear amount. After enhancement by a generative adversarial network, a sample library containing early fault features is formed. Deep neural networks are pre-trained on the virtual sample library and then fine-tuned using real-world operating data to improve the model's ability to identify unknown faults. Lightweight models are deployed on edge nodes for real-time anomaly detection, while cloud-based deep models comprehensively assess health based on historical data, dynamically adjusting warning thresholds to adapt to environmental changes. Fault diagnosis reports use interpretable algorithms to correlate feature parameters with physical damage mechanisms, combined with Monte Carlo simulations to predict remaining lifespan. System maintenance records are fed back to the digital twin model to correct simulation parameters, and a federated learning framework enables collaborative optimization of multiple mine models.

[0026] This solution overcomes the reliance on historical fault data in traditional methods by solving the small-sample learning challenge through virtual sample generation. Compared to fixed-frequency sampling, the dynamic sampling strategy reduces data storage by approximately 40% while maintaining feature integrity. The physical mechanism-based feature selection method improves fault identification accuracy by approximately 25 percentage points compared to traditional statistical feature extraction. The application of the transfer learning framework shortens the model's adaptation period in new mines by approximately 60%. A multi-level early warning system combined with dynamic threshold adjustment keeps the false alarm rate below 5%. The introduction of interpretability technology reduces maintenance decision response time by approximately 30%, achieving a fault location accuracy of over 92%.

[0027] Through the above technical solutions, this application effectively solves the problem of insufficient model generalization ability caused by the scarcity of fault samples in mining equipment. Virtual sample generation technology expands the amount of available training data by 3-5 times, increasing the early fault detection rate to over 85%. The adaptive feature selection mechanism reduces computational resource consumption by more than 30% while maintaining the integrity of key fault features. The multi-level early warning system achieves accurate classification of fault severity, improving maintenance resource allocation efficiency by approximately 40%. The closed-loop optimization mechanism enables the model to continuously evolve during operation, reducing the annual false alarm rate by approximately 15 percentage points. Explainable diagnostic reports provide clear technical basis for on-site maintenance, shortening the average fault diagnosis time by approximately 50%.

[0028] Example 2: Please refer to Figure 1 The specific method for step S1 is as follows: S1.1 Deploy a multimodal sensor network in key parts of mining equipment such as bearing housings, gearboxes, motor housings, and hydraulic systems. The multimodal sensor network includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. Each sensor synchronously collects multi-dimensional physical parameters such as vibration acceleration, surface temperature, acoustic emission signals, operating current, and displacement changes during equipment operation according to a preset initial sampling frequency, and adds a high-precision timestamp to each sampled data. S1.2. The edge computing nodes identify the working conditions of the collected physical parameters. Based on the current load status, operating mode and historical health status of the equipment, the working state of the equipment is determined. When the equipment is identified to be in a transient working condition such as start-up, load change or abnormal fluctuation, the command is sent to the sensor network to increase the sampling frequency. When the equipment is identified to be in a steady-state operating condition, the command is sent to the sensor network to decrease the sampling frequency. At the same time, the data from different sensors are time-aligned and synchronized based on the timestamp to form a spatiotemporally consistent multimodal dataset.

[0029] In this embodiment: This application further proposes deploying a multimodal sensor network in key parts of mining equipment. The multimodal sensor network includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. Each sensor synchronously collects multi-dimensional physical parameters such as vibration acceleration, surface temperature, acoustic emission signals, operating current, and displacement changes during equipment operation according to a preset initial sampling frequency, and adds a high-precision timestamp to each sampled data. The collected physical parameters are used to identify the operating conditions through edge computing nodes. The operating state of the equipment is determined based on the current load status, operating mode, and historical health status of the equipment. When the equipment is identified as being in a transient operating condition such as start-up, load change, or abnormal fluctuation, a command is sent to the sensor network to increase the sampling frequency. When the equipment is identified as being in a steady-state operating condition, a command is sent to the sensor network to decrease the sampling frequency. At the same time, the data from different sensors are time-aligned and synchronized based on the timestamp to form a spatiotemporally consistent multimodal dataset.

[0030] Multimodal sensor networks refer to heterogeneous sensing systems composed of various types of sensors. Specifically, vibration sensors can monitor mechanical vibrations, temperature sensors can capture thermodynamic changes, acoustic emission sensors can detect internal material defects, current sensors can analyze electrical characteristics, and displacement sensors can measure structural deformation. Through the collaborative acquisition of multi-dimensional physical parameters, key characteristics of equipment operating status are comprehensively covered. High-precision timestamps refer to attaching a time stamp accurate to the microsecond level to each sampled data point. This can be achieved by using GPS-synchronized clocks or network time protocols to align the timing of multi-sensor data. Adaptive sampling strategies refer to control methods that dynamically adjust the sensor sampling frequency according to equipment operating conditions. Specifically, edge computing nodes can analyze load status, operating modes, and historical health parameters in real time. Under transient conditions, the sampling frequency can be increased to capture rapidly changing fault characteristics, while under steady-state conditions, the sampling frequency can be decreased to reduce redundant data.

[0031] Vibration sensors are installed on the surfaces of key mechanical components such as bearing housings, gearboxes, motor housings, and hydraulic systems to collect vibration acceleration signals during equipment operation. Temperature sensors are placed at motor windings and bearing locations to monitor the surface temperature distribution of the equipment in real time. Acoustic emission sensors are deployed in the gear meshing area to capture stress wave signals generated by the propagation of internal cracks in the material. Current sensors are installed in the motor power supply circuit to record waveform changes in the operating current. Displacement sensors are installed at structural connections to detect mechanical deformation during equipment operation. Each sensor samples synchronously according to a preset initial frequency; for example, the initial sampling frequency for the vibration sensor is set to 10kHz, the temperature sensor to 1Hz, and the current sensor to 5kHz. Edge computing nodes analyze the fluctuation amplitude of the equipment load current, the spectral characteristics of the vibration signal, and the temperature change trend to determine whether the equipment is in a start-up, variable load, or steady-state operation. When a sudden drop in current is detected accompanied by a shift of vibration spectrum energy to higher frequencies, the equipment is determined to have entered a transient operating condition, and the sampling frequency of the vibration sensor is increased to 20kHz and the temperature sensor to 10Hz. When the equipment operating parameters tend to stabilize, the sampling frequency is restored to its initial value. All sensor data is aligned using high-precision timestamps, such as by using the IEEE 1588 precise time protocol to achieve microsecond-level synchronization, ensuring strict correspondence between different modal data in the time dimension.

[0032] Traditional mining equipment monitoring systems typically use a fixed sampling frequency, generating a large amount of redundant data under steady-state conditions, while potentially missing key fault characteristics under transient conditions. Existing data synchronization technologies often employ simple time window alignment methods, which struggle to eliminate timing errors caused by sensor response delays. This solution utilizes an adaptive sampling strategy driven by operating condition identification, significantly reducing data storage and transmission load while ensuring data integrity. High-precision timestamp synchronization technology achieves spatiotemporal consistency of multimodal data, providing an accurate data foundation for subsequent feature extraction and fault diagnosis.

[0033] Through the above technical solution, this application effectively resolves the contradiction between data acquisition efficiency and completeness in mining equipment monitoring. Increasing the sampling frequency under transient conditions can capture transient fault characteristics during equipment start-up, shutdown, and load changes, while decreasing the sampling frequency under steady-state conditions reduces the amount of invalid data. The timestamp synchronization processing of multi-sensor data eliminates feature extraction errors caused by time-series misalignment in traditional methods, providing a reliable time-series alignment basis for multi-modal data fusion analysis. This solution can optimize system resource utilization while ensuring monitoring accuracy and adapt to the complex and ever-changing operating environment of mining equipment.

[0034] Example 3: Please refer to Figure 1 The specific method for step S2 is as follows: S2.1 Deploy feature extraction algorithms on edge computing nodes, and combine them with the physical degradation mechanism model of the equipment to extract feature parameters from the multimodal raw signals collected in step S1. For vibration signals, extract time-domain features and frequency-domain features. Time-domain features include root mean square value, peak value, kurtosis and waveform factor. Frequency-domain features include characteristic frequency amplitude, spectral kurtosis and envelope spectrum features. For temperature signals, extract thermodynamic features, including temperature rise rate, temperature gradient and thermal balance deviation. For current signals, extract electrical features, including current RMS value, power factor and harmonic distortion rate, to form an initial multi-domain feature set. S2.2 Adaptive feature selection processing is performed on the initial multi-domain feature set extracted in step S2.1. The mutual information algorithm is used to calculate the correlation between each feature parameter and the health status of the equipment. The ReliefF algorithm is used to evaluate the ability of each feature parameter to distinguish different fault modes. The feature parameters are sorted according to the mutual information value and the ReliefF weight. The feature parameters with the highest ranking are selected to form a feature subset. Redundant features with mutual information values ​​lower than a preset threshold and weakly correlated features with ReliefF weights lower than a preset threshold are removed. The dimensionality-reduced feature subset is then output.

[0035] In this embodiment: This application further proposes deploying a feature extraction algorithm on edge computing nodes, combining it with a physical degradation mechanism model of the equipment, to extract feature parameters from the multimodal raw signals. For vibration signals, time-domain features and frequency-domain features are extracted. Time-domain features include root mean square value, peak value, kurtosis, and waveform factor. Frequency-domain features include characteristic frequency amplitude, spectral kurtosis, and envelope spectrum features. For temperature signals, thermodynamic features are extracted, including temperature rise rate, temperature gradient, and thermal balance deviation. For current signals, electrical features are extracted, including current RMS value, power factor, and harmonic distortion rate, forming an initial multi-domain feature set. Adaptive feature selection processing is performed on the extracted initial multi-domain feature set. The mutual information algorithm is used to calculate the correlation between each feature parameter and the equipment health status. The ReliefF algorithm is used to evaluate the ability of each feature parameter to distinguish different fault modes. The feature parameters are sorted according to the mutual information value and the ReliefF weight. The feature parameters with the highest ranking are selected to form a feature subset. Redundant features with mutual information values ​​below a preset threshold and weakly correlated features with ReliefF weights below a preset threshold are removed. The dimensionality-reduced feature subset is then output.

[0036] Multimodal raw signals refer to various physical quantity data collected by vibration sensors, temperature sensors, current sensors, etc. Specifically, synchronous acquisition and spatiotemporal alignment techniques can be used to integrate the data and comprehensively reflect the equipment's operating status. Equipment physical degradation mechanism models are mathematical models established based on equipment material properties, structural dynamics, and failure modes. These can be implemented using finite element analysis or empirical formulas and are used to guide feature extraction. Time-domain features refer to statistics extracted from the signal's time dimension. Specifically, sliding window calculations of root mean square values ​​or kurtosis can be used to characterize macroscopic changes in the equipment's operating status. Frequency-domain features refer to the spectral characteristics obtained through Fourier transform. These can be implemented using the Fast Fourier Transform algorithm and are used to capture specific frequency components caused by equipment failures. Mutual information algorithms are mathematical methods that measure the correlation between features and equipment health status. These can be implemented using probability density estimation and are used to filter features strongly correlated with failures. The ReliefF algorithm is a feature weighting algorithm that evaluates the ability of features to distinguish failure categories. It can be implemented using nearest neighbor sample distance calculations and is used to identify discriminative features sensitive to multiple failure modes.

[0037] In edge computing nodes, the feature extraction direction is first determined based on the physical degradation mechanism model of the equipment. For example, for gear wear faults, the envelope spectrum features of the vibration signal are extracted. Time-domain statistics are extracted from the vibration signal, such as monitoring vibration energy changes by calculating the root mean square value, and the gear meshing frequency amplitude is extracted through spectrum analysis. For temperature signals, abnormal heating areas are identified by calculating the temperature rise rate, and the status of the heat dissipation system is judged by combining the thermal balance deviation. For current signals, the health of the motor windings is evaluated by harmonic distortion rate analysis. After completing multi-domain feature extraction, the mutual information algorithm is used to quantify the correlation between each feature and the equipment health score, such as screening out the spectral kurtosis index that is highly correlated with bearing faults. At the same time, the ReliefF algorithm is used to evaluate the feature's ability to distinguish different faults such as gear pitting and bearing cracks, such as retaining the temperature gradient feature that can significantly distinguish between early wear and normal state. The features are ranked through a dual evaluation mechanism, such as retaining features with mutual information values ​​higher than 0.3 and ReliefF weights higher than 0.15, finally forming a feature subset containing key discriminative features.

[0038] Traditional methods typically employ fixed feature sets or single evaluation metrics, such as relying solely on vibration spectrum peak values, leading to feature redundancy and insufficient discriminative power. Existing technologies utilize unsupervised dimensionality reduction methods like principal component analysis, which may lose key features directly related to the fault mechanism. Our proposed solution, however, integrates physical degradation mechanisms to guide feature extraction and combines mutual information and ReliefF as dual evaluation mechanisms, effectively preserving discriminative features strongly correlated with equipment health status and sensitive to multiple fault modes.

[0039] Through the above technical solutions, this application solves the problems of low model training efficiency and poor generalization ability in small sample scenarios caused by feature redundancy in traditional methods. Feature extraction guided by a mechanistic model ensures that features are directly related to the physical failure process, avoiding interference from invalid features. The adaptive feature selection mechanism reduces data dimensionality while retaining key discriminative information, thus reducing the number of samples required for subsequent model training, making it particularly suitable for mining scenarios where fault samples are scarce. The dual evaluation strategy effectively balances feature relevance and discriminativeness, improving the adaptability of feature subsets to unknown fault types.

[0040] Example 4: Please refer to Figure 1 The specific method for step S3 is as follows: S3.1 Construct a digital twin model of the mining equipment. The digital twin model integrates the geometric model, physical model, and behavioral model of the equipment. The geometric model includes the three-dimensional structure and assembly relationship of the equipment. The physical model includes the dynamic model, thermodynamic model, and electromagnetic model. The behavioral model includes the operating law and response characteristics of the equipment under different working conditions. Set up a fault injection interface in the digital twin model. Simulate the fault state of the equipment by injecting fault parameters of different degrees and types. Fault parameters include the crack depth of the bearing inner ring, the pitting area of ​​the gear, the number of short-circuited turns of the motor winding, and the leakage rate of the hydraulic system. Use the digital twin model to simulate and calculate the vibration response, temperature distribution, current waveform, and displacement change of the equipment under various fault modes, and generate virtual sensor data corresponding to the fault state. S3.2. Using the virtual sensor data generated in step S3.1 as initial virtual samples, a generative adversarial network (GAN) is used to enhance the virtual samples. The generator of the GAN learns the distribution differences between the virtual samples and the real samples and generates corrected samples. The discriminator evaluates the similarity between the corrected samples and the real samples. Through adversarial training between the generator and the discriminator, the virtual samples are made to approximate the real samples in terms of statistical characteristics and frequency domain features. The enhanced virtual samples are classified and labeled according to the degree of equipment degradation to establish a virtual sample library. The virtual sample library includes a normal state sample set, an early failure sample set, a mid-term failure sample set, and a severe failure sample set.

[0041] In this embodiment: This application further proposes a specific implementation method for constructing a digital twin model of mining equipment. The digital twin model integrates the geometric model, physical model, and behavioral model of the equipment. The geometric model includes the three-dimensional structure and assembly relationship of the equipment; the physical model includes a dynamic model, a thermodynamic model, and an electromagnetic model; and the behavioral model includes the operating laws and response characteristics of the equipment under different working conditions. A fault injection interface is set in the digital twin model to simulate the fault state of the equipment by injecting fault parameters of different degrees and types. Fault parameters include the crack depth of the bearing inner ring, the pitting area of ​​the gear, the number of short-circuited turns of the motor winding, and the leakage rate of the hydraulic system. The vibration of the equipment under various fault modes is simulated and calculated using the digital twin model. The system generates virtual sensor data corresponding to the fault state by analyzing response, temperature distribution, current waveform, and displacement changes. This virtual sensor data is then used as initial virtual samples. A generative adversarial network (GAN) is employed to enhance these virtual samples. The generator in the GAN learns the distribution differences between virtual and real samples and generates corrected samples. The discriminator evaluates the similarity between the corrected samples and real samples. Through adversarial training between the generator and the discriminator, the virtual samples are made to approximate real samples in terms of statistical characteristics and frequency domain features. The enhanced virtual samples are then classified and labeled according to the degree of equipment degradation, establishing a virtual sample library containing normal state sample sets, early fault sample sets, mid-term fault sample sets, and severe fault sample sets.

[0042] A digital twin model is a virtual mapping system that integrates the geometry, physical characteristics, and behavioral patterns of equipment. Specifically, it can be implemented using a multiphysics coupling simulation platform, such as co-simulating a SolidWorks geometric model with an ANSYS physical model. This model establishes a mapping relationship between equipment operating mechanisms and sensor data, providing a source of simulation data to address the problem of scarce actual fault samples.

[0043] A fault injection interface is a programming interface that allows fault parameters to be injected into a virtual model. This can be achieved using parametric modeling tools, such as by modifying the material properties or boundary conditions of the finite element model. By simulating the physical evolution of real faults, this interface makes the generated virtual data physically interpretable.

[0044] Generative Adversarial Networks (GANs) are deep learning frameworks consisting of a generator and a discriminator, specifically implemented using a conditional generative adversarial network architecture. This network enhances the realism of virtual samples through adversarial training mechanisms, addressing the problem of insufficient model generalization ability caused by the distribution difference between simulated and measured data.

[0045] In the implementation process, a three-dimensional geometric model is first constructed based on the equipment's CAD drawings, and a multiphysics model incorporating dynamic and thermodynamic characteristics is established through finite element analysis. Adjustable fault parameter controls are set within the digital twin environment; for example, the bearing crack depth is set to an adjustable range of 0.1-2.0 mm. Batch simulations are performed by changing the combination of fault parameters to obtain virtual sensor data containing information such as vibration spectrum and temperature field distribution. Subsequently, the simulation data is input into a generative adversarial network (GAN). The generator learns the time-frequency characteristics of real fault data through a convolutional neural network, while the discriminator uses a spectrum discriminator to evaluate the frequency domain similarity of the generated data. After iterative training, the enhanced virtual samples achieve feature alignment with real fault data in dimensions such as time-domain waveform and spectral energy distribution, ultimately forming a standardized sample library containing four types of degradation states.

[0046] Traditional methods primarily rely on historical fault data or single-physics simulations to generate training samples, which struggles to cover complex multi-fault coupled scenarios. This solution utilizes a multi-physics digital twin model to generate fault mechanism-driven data. Combined with the distribution correction capabilities of generative adversarial networks, it can construct a high-fidelity virtual sample library even without real fault data. Compared to simple data augmentation techniques, the samples generated by this method not only conform to the physical degradation patterns of equipment but also possess statistical characteristics similar to measured data.

[0047] By employing the aforementioned technical solution, this application effectively addresses the problem of insufficient training of diagnostic models due to the scarcity of catastrophic failure samples in mining equipment. Through dual verification using physical mechanisms and data-driven approaches, the generated virtual samples accurately reflect progressive failure characteristics such as bearing crack propagation and gear pitting development, providing deep neural networks with training data covering the entire lifecycle. This method eliminates excessive reliance on historical failure databases, supports rapid modeling of novel failure modes, and simultaneously eliminates feature bias between the simulation and real domains through adversarial training, significantly improving the generalization performance of the fault diagnosis model across various operating conditions.

[0048] Example 5: Please refer to Figure 1 The specific method for step S4 is as follows: S4.1 Construct a deep neural network fault diagnosis model. The deep neural network includes a feature extraction layer, a feature fusion layer, and a classification output layer. The feature extraction layer uses a convolutional neural network or Transformer architecture to extract deep representations of the feature subset output in step S2. The feature fusion layer designs a multimodal feature fusion module based on an attention mechanism. It calculates the weight allocation of different modal features and performs weighted fusion through self-attention and cross-attention mechanisms. The classification output layer outputs the health status category and fault type of the device. The deep neural network is pre-trained using the virtual fault sample library established in step S3, so that the model learns the general representation of fault features and the distinguishing characteristics between different fault modes. S4.2. Collect real fault samples from the target mine as a fine-tuning dataset. Fine-tune the deep neural network pre-trained in step S4.1 using domain adaptation technology. Domain adaptation technology includes an adversarial domain adaptation module and a feature alignment module. The adversarial domain adaptation module distinguishes virtual sample features from real sample features through a domain discriminator and adjusts the feature extraction layer parameters to make the distributions of the two types of features tend to be consistent. The feature alignment module calculates the maximum mean difference between virtual sample features and real sample features and minimizes this difference. A meta-learning mechanism is introduced, and the model gains the ability to quickly learn new fault types through rapid adaptive training on multiple small sample tasks. The fine-tuned and optimized fault diagnosis model is output.

[0049] In this embodiment: This application further proposes a specific implementation method for constructing a deep neural network fault diagnosis model, including using a transfer learning strategy for model training, pre-training on a virtual fault sample library and then fine-tuning using real fault samples, reducing the distribution difference between virtual data and real data through domain adaptation technology, and introducing a meta-learning mechanism to design a feature fusion module based on an attention mechanism.

[0050] Transfer learning strategies refer to the technique of transferring knowledge learned from a virtual sample database to real-world scenarios. This can be achieved through pre-training and fine-tuning. The model is trained on a virtual sample database to gain basic fault identification capabilities, and then its parameters are adjusted using a small number of real samples to adapt to the target scenario. Domain adaptation techniques aim to eliminate the differences in distribution between virtual and real data. This can be achieved using adversarial domain adaptation modules and feature alignment modules. A domain discriminator guides the feature extraction layer to generate domain-invariant features while minimizing the statistical differences between virtual and real sample features. Meta-learning mechanisms enable models to quickly adapt to new tasks. This is achieved by training the model on multiple small-sample tasks to optimize model parameters, allowing it to quickly adjust decision boundaries when facing new fault types. The feature fusion module of the attention mechanism is a technique for weighted integration of features from different modalities. This can be achieved using self-attention and cross-attention mechanisms to calculate the weight allocation of features across modalities, highlighting key features and suppressing noise interference.

[0051] The deep neural network consists of a feature extraction layer, a feature fusion layer, and a classification output layer. The feature extraction layer uses a convolutional neural network or Transformer architecture to perform deep representation learning of multimodal features, capturing the potential correlations between signals such as vibration, temperature, and current. The feature fusion layer dynamically adjusts the contribution of each modality feature through an attention mechanism; for example, it automatically enhances the weight of vibration spectrum kurtosis features in bearing fault scenarios and increases the saliency of current harmonic distortion features in motor fault scenarios. The classification output layer combines the general fault patterns learned in the pre-training stage with the mine-specific fault features adapted in the fine-tuning stage to output the equipment health status classification result. During domain adaptation, adversarial training gradually approximates the feature distribution of virtual samples to real samples, while feature alignment eliminates the distance between the two types of data in the regenerating kernel Hilbert space through kernel function mapping. The meta-learning module optimizes the model parameter update direction by simulating various small-sample fault diagnosis tasks, enabling it to complete parameter adjustment with only a small number of samples when encountering new faults.

[0052] Traditional methods typically train a single diagnostic model directly on real data, which can easily lead to overfitting when there are insufficient real fault samples. Existing domain adaptation techniques mostly use static feature alignment, which is difficult to handle the dynamic distribution differences of mining equipment under multiple operating conditions. Conventional feature fusion methods use a fixed weight addition method, which cannot dynamically adjust the importance of multimodal features according to the fault type.

[0053] Through the above technical solutions, this application effectively solves the data distribution offset problem between virtual samples and real scenes, improving the fault diagnosis accuracy under small sample conditions. The dynamic feature fusion mechanism enhances the feature representation capability of multi-source heterogeneous data, enabling the model to adaptively focus on key signal features related to the current fault. The introduction of the meta-learning mechanism significantly improves the model's ability to quickly adapt to newly emerging fault modes, reducing the number of samples and data acquisition costs required for diagnosing new faults.

[0054] Example 6: Please refer to Figure 1 The specific method for step S5 is as follows: S5.1 After the fault diagnosis model trained in step S4 is lightweighted, it is deployed to the edge computing node. The deep diagnosis model and the equipment health assessment model are deployed in the cloud. The edge node receives the feature subset output in step S2, uses the deployed lightweight diagnosis model to judge the feature subset, and compares the feature parameters with the preset anomaly detection threshold. When the feature parameters are detected to exceed the threshold range or the model outputs an abnormal status indicator, the edge node triggers a first-level warning signal and packages the abnormal feature data, raw sensor data and equipment operating information and uploads them to the cloud. S5.2 After receiving data uploaded by edge nodes, the cloud-based deep diagnostic model analyzes the data, combining historical equipment operating data, current operating parameters, and environmental factors. The equipment health assessment model calculates the equipment health score, which ranges from 0 to 100. A dynamic threshold adjustment mechanism is established to correct the warning threshold based on the equipment's current load status, operating mode, and environmental parameters such as temperature and humidity. A multi-level progressive warning system is constructed, dividing the equipment health score into four levels: 90-100 is the healthy level, 70-90 is the attention level, 50-70 is the warning level, and below 50 is the danger level. Different levels correspond to different warning signal strengths and maintenance response strategies. The cloud-based system feeds back the warning level, fault type judgment results, and maintenance suggestions to the monitoring terminal.

[0055] In this embodiment: This application further proposes to deploy the fault diagnosis model trained in step S4 to an edge computing node after lightweight processing, and to deploy a deep diagnostic model and a device health assessment model in the cloud. The edge node receives the feature subset output in step S2, uses the deployed lightweight diagnostic model to judge the feature subset, and compares the feature parameters with a preset anomaly detection threshold. When the feature parameters are detected to exceed the threshold range or the model outputs an abnormal state identifier, the edge node triggers a first-level warning signal, and simultaneously packages and uploads the abnormal feature data, raw sensor data, and device operating condition information to the cloud. After receiving the data uploaded by the edge node, the cloud analyzes the data and comprehensively sets the parameters. Based on historical operating data, current operating parameters, and environmental factors, the equipment health assessment model calculates the equipment health score, which ranges from 0 to 100. A dynamic threshold adjustment mechanism is established to correct the warning threshold based on the equipment's current load status, operating mode, and environmental parameters such as temperature and humidity. A multi-level progressive warning system is constructed, dividing the equipment health score into four levels: 90 to 100 points is the healthy level, 70 to 90 points is the attention level, 50 to 70 points is the warning level, and below 50 points is the danger level. Different levels correspond to different warning signal intensities and maintenance response strategies. The cloud platform feeds back the warning level, fault type judgment results, and maintenance suggestions to the monitoring terminal.

[0056] Lightweight processing refers to reducing the computational and storage requirements of deep neural networks through model compression and quantization techniques. Specifically, it can be achieved using knowledge distillation or channel pruning methods, enabling the model to run efficiently on edge computing nodes with limited resources.

[0057] The dynamic threshold adjustment mechanism refers to the adaptive correction of the anomaly detection threshold based on the real-time operating conditions of the equipment and environmental parameters. Specifically, it can be implemented using a sliding window statistical method or an online learning algorithm. By analyzing the historical operating data of the equipment and the current environmental parameters, the threshold range is dynamically adjusted to avoid misjudgment problems caused by fixed thresholds.

[0058] A multi-level progressive early warning system refers to dividing the health status of equipment into different levels and matching them with differentiated early warning strategies. Specifically, it can be implemented using fuzzy comprehensive evaluation methods or hierarchical analysis methods. Through a quantitative scoring mechanism, changes in equipment status are transformed into operable maintenance instructions, achieving a precise correspondence between early warning signals and maintenance responses.

[0059] The trained fault diagnosis model, after being lightweighted, is deployed to edge computing nodes for real-time monitoring of changes in feature subsets. When an edge node detects that a feature parameter exceeds a dynamically adjusted threshold or the model outputs an anomaly flag, it immediately triggers a level-one warning and uploads relevant data to the cloud. The cloud performs secondary analysis on the uploaded data using a deep diagnostic model, combining historical equipment operating data and environmental parameters to calculate a health score. The health score is updated in real-time based on equipment load, operating mode, and environmental conditions such as temperature and humidity. The dynamic threshold adjustment mechanism uses a sliding window to statistically analyze the distribution characteristics of historical equipment data and automatically corrects the triggering conditions for anomaly detection. The multi-level progressive warning system maps the health score to four levels, with different levels corresponding to different colored warning signals and maintenance response priorities. For example, a health level requires no intervention, a monitoring level initiates regular checks, a warning level schedules preventative maintenance, and a danger level requires immediate shutdown for repair.

[0060] Existing methods typically use fixed thresholds for anomaly detection, which cannot adapt to changes in equipment operating conditions and environmental interference, easily leading to false alarms or missed alarms. Traditional early warning systems only provide binary alarm signals, lacking progressive assessment of equipment health status, resulting in untargeted maintenance strategies. This solution solves the problem of poor adaptability of fixed thresholds through a dynamic threshold adjustment mechanism. It achieves refined hierarchical management of equipment status using a multi-level progressive early warning system, and combined with a collaborative computing architecture between the edge and cloud, it improves diagnostic accuracy while ensuring real-time performance.

[0061] Through the above technical solutions, this application can dynamically adjust the detection sensitivity according to the actual operating status of the equipment, reducing false alarms caused by environmental interference or fluctuations in operating conditions. The multi-level early warning system provides differentiated maintenance guidance for different health states, avoiding over-maintenance or under-maintenance. The edge and cloud-based hierarchical processing mechanism reduces network transmission pressure while ensuring that the need for in-depth analysis of complex faults is met, thus comprehensively improving the reliability and maintenance efficiency of the mining equipment condition monitoring system.

[0062] Example 7: Please refer to Figure 1 The specific method for step S6 is as follows: S6.1 After the multi-level progressive early warning system in step S5 issues an early warning signal, the fault tracing module is activated. The output of the fault diagnosis model in step S4 is analyzed using interpretable artificial intelligence technology. The key feature parameters and decision basis of the model judgment are traced through gradient weighted activation mapping and SHAP value calculation method. Gradient weighted activation mapping generates a feature activation heatmap, and SHAP value calculation generates a contribution explanation report of each feature parameter. The feature activation heatmap and contribution explanation report highlight core feature parameters such as vibration spectrum kurtosis, temperature gradient deviation and current harmonic distortion. These core feature parameters are associated with the equipment physical degradation mechanism model to form a fault clue set. S6.2 Input the fault clue set formed in step S6.1 into the equipment structure knowledge graph and the fault propagation path model. The equipment structure knowledge graph includes the topological relationships of equipment components, material properties, and historical failure cases. The fault propagation path model simulates the dynamic evolution of a fault from the nascent stage to the diffusion stage. The graph neural network algorithm searches for matching fault patterns in the equipment structure knowledge graph, reverse-engineers the location of the fault source and the fault type, calculates the fault severity index, calculates the remaining service life prediction value using the Monte Carlo simulation method, and generates a fault diagnosis report. The fault diagnosis report includes three-dimensional fault component annotation, fault severity assessment, remaining service life prediction value, and maintenance recommendations. The fault diagnosis report is output to the maintenance terminal through a visualization interface.

[0063] In this embodiment: This application further proposes to activate the fault tracing module after the multi-level progressive early warning system issues an early warning signal. It utilizes interpretable artificial intelligence technology to analyze the output of the fault diagnosis model, traces the key feature parameters and decision-making basis of the model's judgment through gradient weighted class activation mapping and SHAP value calculation methods, generates a feature activation heatmap and a contribution explanation report, associates these core feature parameters with the equipment physical degradation mechanism model to form a fault clue set, inputs the fault clue set into the equipment structure knowledge graph and fault propagation path model, searches for matching fault patterns in the equipment structure knowledge graph using graph neural network algorithms, infers the fault source location and fault type in reverse, calculates the fault severity index, calculates the remaining service life prediction value using Monte Carlo simulation methods, and generates a fault diagnosis report containing three-dimensional fault component annotations, fault severity assessment, remaining service life prediction value, and maintenance recommendations. The report is then output to the maintenance terminal through a visualization interface.

[0064] Gradient-weighted activation mapping (GEM) is a technique for visualizing the decision-making processes of deep neural networks. Specifically, it can be implemented using the gradient-weighted averaging method of feature maps in convolutional neural networks, generating feature activation heatmaps to locate the sensor data regions that have the greatest impact on model decisions. SHAP value calculation is a technique based on a unified framework of game theory to explain model prediction results. Specifically, it can use additive feature attribution methods to calculate the contribution of each feature parameter to the prediction result. Equipment structure knowledge graphs are structured knowledge bases describing the topological relationships, material properties, and historical failure cases of equipment components. Specifically, they can use graph databases to store the assembly relationships and failure association rules of equipment components. Fault propagation path models are mathematical models that simulate the dynamic evolution of faults from the nascent stage to the diffusion stage. Specifically, they can use differential equations or state transition diagrams to describe the propagation laws of faults among equipment components. Monte Carlo simulation methods are numerical calculation methods based on probability statistics. Specifically, they can use random sampling and repeated experiments to estimate the probability distribution of the remaining service life of equipment.

[0065] Once the warning signal is triggered, interpretable artificial intelligence technology first analyzes the decision logic of the diagnostic model. It generates heatmaps of key features such as vibration spectrum kurtosis and temperature gradient deviation through gradient-weighted class activation mapping, while using SHAP values ​​to quantify the contribution of each feature to fault classification. These interpretive results are combined with a physical degradation mechanism model of the equipment to screen out fault clues that conform to the laws of mechanical failure. Subsequently, the fault clues are mapped into the equipment structure knowledge graph. Combined with a fault propagation path model, a graph neural network traverses the connection relationships of equipment components to identify the most likely location of the fault source. The severity of the fault is determined by calculating the degree to which the current feature deviates from the normal threshold, and the remaining service life is predicted probabilistically by randomly generating equipment degradation paths through Monte Carlo simulation. The final diagnostic report visually displays the faulty components with 3D annotations and recommends targeted maintenance strategies based on maintenance history.

[0066] Traditional fault tracing methods rely on expert experience to manually analyze characteristic parameters, lacking interpretable verification of model decision-making basis and struggling to handle complex correlations between multi-sensor data. Existing knowledge graph applications are mostly limited to static relationship queries, failing to integrate with dynamic fault propagation models to achieve reasoning capabilities. Conventional life prediction methods are based on fixed degradation models, unable to adapt to individual equipment differences and operating condition fluctuations. This solution integrates interpretable artificial intelligence with mechanistic models to achieve cross-validation between model decision-making basis and physical laws, utilizes dynamic knowledge graphs to enhance the accuracy of fault reasoning, and combines probabilistic simulation to improve the robustness of life prediction.

[0067] Through the above technical solutions, this application effectively solves the problem of misjudgment caused by poor interpretability of fault diagnosis models under small sample conditions. It improves the reliability of fault clues by constraining physical mechanisms, and makes up for the lack of data by using dynamic knowledge reasoning. This improves the accuracy of fault location and the efficiency of maintenance decision-making, while providing maintenance personnel with intuitive and reliable diagnostic basis and reducing excessive reliance on historical fault data.

[0068] Example 8: Please refer to Figure 1 The specific method for step S7 is as follows: S7.1 Establish a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each warning, maintenance operations and actual fault verification results to form a fault case library, including equipment fault types, maintenance records, fault cause analysis and maintenance effect evaluation. The system identifies the uncertainty of the model on fault samples, including false alarms, missed alarms or new fault types, and marks them as high-value samples. Maintenance experts manually annotate these samples and use the annotated samples to update the training set and optimize the model. S7.2 Input high-quality samples after active learning and annotation into the digital twin model, update the model parameters, and adjust the state of the twin model to adapt to the needs of equipment health evolution and fault diagnosis. Adopt the federated learning framework to achieve cross-mine model collaborative optimization through distributed computing while ensuring the data privacy of each mine. Each mine synchronizes and updates by sharing model parameters.

[0069] In this embodiment: This application further proposes to establish a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each early warning, maintenance operations, and actual fault verification results to form a fault case library, including equipment fault types, maintenance records, fault cause analysis, and maintenance effect evaluation information. The system identifies the uncertainty of the model on fault samples, including false alarms, missed alarms, or new fault types, and marks them as high-value samples. Maintenance experts manually annotate these samples and use the annotated samples to update the training set and optimize the model. The high-quality samples after active learning and annotation are input into the digital twin model to update the model parameters and adjust the twin model state to adapt to the needs of equipment health evolution and fault diagnosis. A federated learning framework is adopted to achieve cross-mine model collaborative optimization through distributed computing while ensuring the data privacy of each mine. Each mine synchronizes and updates by sharing model parameters.

[0070] A closed-loop mechanism refers to forming a complete data link between equipment monitoring data, diagnostic results, maintenance operations, and verification feedback. This can be achieved by establishing an event-driven database for continuous optimization of model parameters and correction of diagnostic logic. Active learning annotation involves using uncertainty sampling algorithms to filter samples with low model confidence, such as entropy thresholding or committee query strategies. Maintenance personnel then manually verify key samples, improving annotation efficiency. Federated learning frameworks employ distributed machine learning architectures, such as secure aggregation protocols or differential privacy technologies, to enable multiple mining nodes to exchange model parameters without sharing raw data, addressing the data silo problem. Digital twin model parameter updates involve inputting actual maintenance records and failure analysis data back into the simulation model. For example, Bayesian update algorithms can be used to adjust fault evolution parameters in the model, enhancing the model's accuracy in fitting the equipment degradation process.

[0071] After issuing a fault warning, the system continuously tracks the actual effectiveness of maintenance operations, linking and storing maintenance records with the original diagnostic data. When the model exhibits uncertainty in its fault judgment under specific operating conditions, such as when the confidence level for identifying novel complex faults falls below a set threshold, the system automatically marks such samples as unlabeled samples. Maintenance experts then precisely label the samples based on equipment disassembly and inspection results, and the labeled data is used for incremental training of the diagnostic model. The updated diagnostic model parameters are securely aggregated with other mine nodes using a federated averaging algorithm, achieving knowledge sharing while protecting the data privacy of each node. Simultaneously, actual degradation data such as bearing wear and gear clearance changes discovered during equipment maintenance are synchronously input into the digital twin model. The parameter calibration module dynamically adjusts the material fatigue coefficient and fault propagation rate parameters in the model, making the virtual simulation results closer to the actual equipment condition.

[0072] Traditional methods typically employ single-batch model training and lack continuous optimization mechanisms, leading to models that cannot adapt to feature drift caused by equipment aging. Existing technologies require manual full-scale sample labeling, resulting in low efficiency and high costs. This solution, however, accurately identifies high-value samples through an active learning mechanism, reducing invalid labeling workload by over 60%. Regarding cross-mine collaboration, traditional centralized training requires aggregating sensitive data from various mines, posing a data leakage risk. This solution, however, achieves collaborative model optimization through federated learning, where each mine only exchanges encrypted model gradient parameters, keeping the original data locally. In terms of digital twin model maintenance, existing technologies often use static parameter settings, failing to reflect the actual equipment degradation process. This solution, however, achieves dynamic calibration of model parameters through back-injection of maintenance data, reducing simulation errors by approximately 40%.

[0073] Through the above technical solutions, this application effectively solves the problem of insufficient model generalization ability caused by the scarcity of fault samples, and continuously accumulates high-quality fault case data through a closed-loop feedback mechanism. The active learning strategy significantly improves sample annotation efficiency, allowing limited manual annotation resources to focus on key and difficult samples. The federated learning framework breaks down data silos and enhances model robustness by utilizing common degradation problems of equipment across multiple mines. The dynamic calibration mechanism of the digital twin model ensures that virtual simulation data always reflects the real equipment state, providing reliable data support for model training under small sample conditions. This solution improves fault diagnosis accuracy by approximately 25%, reduces the response time for identifying new faults by more than 50%, and simultaneously achieves cross-mine knowledge sharing without leaking sensitive data.

[0074] This application also provides an IoT-based mining equipment condition monitoring system; please refer to [link / reference]. Figure 2It includes a multimodal data acquisition module, an edge feature extraction and screening module, a digital twin virtual sample generation module, a deep neural network fault diagnosis module, a cloud-edge collaborative multi-level early warning module, an interpretable fault source tracing and diagnosis module, and a closed-loop feedback learning and optimization module. The multimodal data acquisition module includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. It collects multi-dimensional physical parameters during equipment operation and dynamically adjusts the sampling frequency according to the current load status, operating mode, and historical health status of the equipment. The sampling frequency is increased during equipment start-up and shutdown and variable load conditions, and decreased during steady-state operation. The module also performs timestamp synchronization processing on the multi-source sensor data. The edge feature extraction and screening module deploys feature extraction algorithms on edge computing nodes. Combined with the physical degradation mechanism model of the equipment, it extracts feature parameters from the original sensor signals. The feature parameters include time-frequency domain features based on vibration signals, thermodynamic features based on temperature signals, and electrical features based on current signals. Through an adaptive feature selection mechanism, it calculates the correlation between features and their contribution to fault classification based on mutual information algorithm and ReliefF algorithm, filters feature subsets and removes redundant and weakly correlated features. The digital twin virtual sample generation module constructs a digital twin model of mining equipment. The digital twin model integrates the equipment's geometric structure model, physical degradation mechanism model, and behavioral operation law model. Different levels and types of fault parameters, including bearing wear parameters, gear crack parameters, and motor insulation aging parameters, are injected into the digital twin environment to simulate the equipment's response characteristics under various fault modes, generate virtual sensor data, and use generative adversarial networks to enhance the realism of the virtual samples. A virtual sample library is established, which includes normal state samples, early fault samples, mid-term fault samples, and severe fault samples. The deep neural network fault diagnosis module constructs a deep neural network fault diagnosis model, uses a transfer learning strategy for model training, pre-trains on a virtual sample library to learn general fault feature representations, fine-tunes using real fault samples from the target mine to optimize the decision boundary, reduces the distribution difference between virtual and real data through domain adaptation technology, introduces a feature fusion mechanism based on attention mechanism, dynamically weights feature parameters of different modalities, and outputs the probability distribution results of fault types. The cloud-edge collaborative multi-level early warning module deploys a lightweight fault diagnosis model at the edge node for real-time judgment. When the edge node detects abnormal features, it triggers a first-level early warning signal and uploads detailed data to the cloud. The cloud-based deep diagnosis model performs in-depth analysis of historical equipment operating data, current operating parameters, and environmental factors, quantifies and calculates the equipment health index, establishes a dynamic threshold adjustment mechanism based on Bayesian networks, corrects the early warning threshold according to operating condition fluctuations and environmental changes, and constructs a multi-level progressive early warning system, dividing the equipment status into four levels: healthy status, attention status, early warning status, and dangerous status. The interpretable fault tracing and diagnosis module initiates the fault tracing program after the early warning module issues an early warning signal. It generates a feature activation heatmap using gradient weighted activation mapping technology, displaying the frequency bands and temperature regions of interest to the fault diagnosis model. It generates an explanation report of the contribution of each feature parameter through the SHAP value calculation method, identifies core decision features such as vibration spectrum kurtosis, temperature gradient deviation, and current harmonic distortion, and associates these core decision features with the equipment structure knowledge graph and fault propagation path model. It then reverse-engineers the location of the fault source component and the fault type, calculates the fault severity index, predicts the remaining service life using the Monte Carlo simulation method, and generates a fault diagnosis report. The fault diagnosis report includes 3D fault component annotations, fault severity assessment, predicted remaining service life, and maintenance recommendations. The closed-loop feedback learning optimization module establishes a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each warning, maintenance operations, and actual fault verification results, forming a real fault case library. This library includes fault types, maintenance records, fault cause analysis, and maintenance effectiveness evaluation information. Using an active learning strategy, the system calculates the model's prediction uncertainty on specific samples, identifies high-value samples such as false alarms, missed alarms, and new fault types, and requests maintenance experts to manually annotate these high-value samples. The annotated samples are then incrementally updated to the training set, and the model is retrained. Equipment maintenance records and spare parts replacement information are incorporated into the digital twin model to correct physical parameters. Employing a federated learning framework, each mine trains its model locally and uploads the model parameters to a cloud aggregation server for weighted averaging. The updated global model is then distributed to each mine, achieving cross-mine knowledge sharing and collaborative optimization while protecting the data privacy of each mine.

[0075] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0076] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the condition of mining equipment based on the Internet of Things, characterized in that: The specific steps are as follows: S1. Deploy a multimodal sensor network in key parts of mining equipment, including vibration, temperature, acoustic emission, current and displacement sensors, to collect multi-dimensional physical parameters. Through an adaptive sampling strategy, adjust the frequency according to the load status, operating mode and historical health status. Increase sampling during start-up / load change conditions and decrease it during steady state, and synchronize timestamps. S2. Deploy feature extraction algorithms at edge nodes, combine them with physical degradation mechanism models, extract time-frequency domain, thermodynamic and electrical features from the original signal, and use an adaptive selection mechanism to filter subsets based on mutual information and the ReliefF algorithm to remove redundant / weakly correlated features and improve data quality. S3. Construct a digital twin model of the equipment, integrate geometric, physical and behavioral models, simulate fault mode response, inject fault parameters to generate virtual sensor data, use generative adversarial networks to enhance samples, and establish a virtual sample library covering normal to severe fault states. S4. Construct a deep neural network diagnostic model, adopt transfer learning, pre-train it in a virtual sample library, fine-tune it using real fault samples, apply domain adaptive reduction of distribution differences, and introduce meta-learning and attention mechanism feature fusion modules. S5. Deploy the diagnostic model to edge nodes, deploy deep models and health assessments in the cloud, trigger early warnings and upload data when the edge judges anomalies, analyze and quantify the health status in the cloud, establish a dynamic threshold mechanism to correct the early warning threshold, and build a multi-level system to classify health / concern / early warning / danger levels. S6. After the warning is issued, the source tracing module is activated. It uses explainable artificial intelligence technology to trace key features / evidence, combines knowledge graphs and propagation path models to infer the source location / type, calculates the severity index and remaining lifespan, and generates a report including component labeling, assessment, prediction and recommendations. S7. Establish a closed loop of monitoring, diagnosis, maintenance and feedback, record response / operation / verification results to form a case library, use active learning to identify uncertain samples, use expert annotation to update the model, correct twin parameters, use federated learning to protect privacy, and achieve cross-mine collaborative optimization and knowledge sharing.

2. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 1, characterized in that: The specific method of step S1 is as follows: S1.1 Deploy a multimodal sensor network in key parts of the bearing housing, gearbox, motor housing, and hydraulic system of the mining equipment. The multimodal sensor network includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. Each sensor synchronously collects multi-dimensional physical parameters such as vibration acceleration, surface temperature, acoustic emission signal, operating current, and displacement changes during equipment operation according to a preset initial sampling frequency, and adds a high-precision timestamp to each sampled data. S1.

2. The edge computing nodes identify the working conditions of the collected physical parameters. Based on the current load status, operating mode and historical health status of the equipment, the working state of the equipment is determined. When the equipment is identified to be in a start-up, stop, load change or abnormal fluctuation transient working condition, the command is sent to the sensor network to increase the sampling frequency. When the equipment is identified to be in a steady state operating condition, the command is sent to the sensor network to decrease the sampling frequency. At the same time, the data from different sensors are time-aligned and synchronized based on the timestamp to form a spatiotemporally consistent multimodal dataset.

3. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 2, characterized in that: The specific method of step S2 is as follows: S2.1 Deploy feature extraction algorithms on edge computing nodes, and combine them with the physical degradation mechanism model of the equipment to extract feature parameters from the multimodal raw signals collected in step S1. For vibration signals, extract time-domain features and frequency-domain features. Time-domain features include root mean square value, peak value, kurtosis and waveform factor. Frequency-domain features include characteristic frequency amplitude, spectral kurtosis and envelope spectrum features. For temperature signals, extract thermodynamic features, including temperature rise rate, temperature gradient and thermal balance deviation. For current signals, extract electrical features, including current RMS value, power factor and harmonic distortion rate, to form an initial multi-domain feature set. S2.2 Adaptive feature selection processing is performed on the initial multi-domain feature set extracted in step S2.

1. The mutual information algorithm is used to calculate the correlation between each feature parameter and the health status of the equipment. The ReliefF algorithm is used to evaluate the ability of each feature parameter to distinguish different fault modes. The feature parameters are sorted according to the mutual information value and the ReliefF weight. The feature parameters with the highest ranking are selected to form a feature subset. Redundant features with mutual information values ​​lower than a preset threshold and weakly correlated features with ReliefF weights lower than a preset threshold are removed. The dimensionality-reduced feature subset is then output.

4. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 3, characterized in that: The specific method of step S3 is as follows: S3.1 Construct a digital twin model of the mining equipment. The digital twin model integrates the geometric model, physical model, and behavioral model of the equipment. The geometric model includes the three-dimensional structure and assembly relationship of the equipment. The physical model includes the dynamic model, thermodynamic model, and electromagnetic model. The behavioral model includes the operating law and response characteristics of the equipment under different working conditions. Set up a fault injection interface in the digital twin model. Simulate the fault state of the equipment by injecting fault parameters of different degrees and types. Fault parameters include the crack depth of the bearing inner ring, the pitting area of ​​the gear, the number of short-circuited turns of the motor winding, and the leakage rate of the hydraulic system. Use the digital twin model to simulate and calculate the vibration response, temperature distribution, current waveform, and displacement change of the equipment under various fault modes, and generate virtual sensor data corresponding to the fault state. S3.

2. Using the virtual sensor data generated in step S3.1 as initial virtual samples, a generative adversarial network (GAN) is used to enhance the virtual samples. The generator of the GAN learns the distribution differences between the virtual samples and the real samples and generates corrected samples. The discriminator evaluates the similarity between the corrected samples and the real samples. Through adversarial training between the generator and the discriminator, the virtual samples are made to approximate the real samples in terms of statistical characteristics and frequency domain features. The enhanced virtual samples are classified and labeled according to the degree of equipment degradation to establish a virtual sample library. The virtual sample library includes a normal state sample set, an early failure sample set, a mid-term failure sample set, and a severe failure sample set.

5. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 4, characterized in that: The specific method of step S4 is as follows: S4.1 Construct a deep neural network fault diagnosis model. The deep neural network includes a feature extraction layer, a feature fusion layer, and a classification output layer. The feature extraction layer uses a convolutional neural network or Transformer architecture to extract deep representations of the feature subset output in step S2. The feature fusion layer designs a multimodal feature fusion module based on an attention mechanism. It calculates the weight allocation of different modal features and performs weighted fusion through self-attention and cross-attention mechanisms. The classification output layer outputs the health status category and fault type of the device. The deep neural network is pre-trained using the virtual fault sample library established in step S3, so that the model learns the general representation of fault features and the distinguishing characteristics between different fault modes. S4.

2. Collect real fault samples from the target mine as a fine-tuning dataset. Fine-tune the deep neural network pre-trained in step S4.1 using domain adaptation technology. Domain adaptation technology includes an adversarial domain adaptation module and a feature alignment module. The adversarial domain adaptation module distinguishes virtual sample features from real sample features through a domain discriminator and adjusts the feature extraction layer parameters to make the distributions of the two types of features tend to be consistent. The feature alignment module calculates the maximum mean difference between virtual sample features and real sample features and minimizes this difference. A meta-learning mechanism is introduced, and the model gains the ability to quickly learn new fault types through rapid adaptive training on multiple small sample tasks. The fine-tuned and optimized fault diagnosis model is output.

6. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 5, characterized in that: The specific method of step S5 is as follows: S5.1 After the fault diagnosis model trained in step S4 is lightweighted, it is deployed to the edge computing node. The deep diagnosis model and the equipment health assessment model are deployed in the cloud. The edge node receives the feature subset output in step S2, uses the deployed lightweight diagnosis model to judge the feature subset, and compares the feature parameters with the preset anomaly detection threshold. When the feature parameters are detected to exceed the threshold range or the model outputs an abnormal status indicator, the edge node triggers a first-level warning signal and packages the abnormal feature data, raw sensor data and equipment operating information and uploads them to the cloud. S5.2 After receiving data uploaded by edge nodes, the cloud-based deep diagnostic model analyzes the data, combining historical equipment operating data, current operating parameters, and environmental factors. The equipment health assessment model calculates the equipment health score, which ranges from 0 to 100. A dynamic threshold adjustment mechanism is established to correct the warning threshold based on the equipment's current load status, operating mode, and environmental parameters such as temperature and humidity. A multi-level progressive warning system is constructed, dividing the equipment health score into four levels: 90 to 100 points is the healthy level, 70 to 90 points is the attention level, 50 to 70 points is the warning level, and below 50 points is the danger level. Different levels correspond to different warning signal strengths and maintenance response strategies. The cloud-based system feeds back the warning level, fault type judgment results, and maintenance suggestions to the monitoring terminal.

7. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 6, characterized in that: The specific method of step S6 is as follows: S6.1 After the multi-level progressive early warning system in step S5 issues an early warning signal, the fault tracing module is activated. The output of the fault diagnosis model in step S4 is analyzed using interpretable artificial intelligence technology. The key feature parameters and decision basis of the model are traced through gradient weighted activation mapping and SHAP value calculation method. Gradient weighted activation mapping generates a feature activation heatmap, and SHAP value calculation generates a contribution explanation report for each feature parameter. The feature activation heatmap and contribution explanation report highlight the core feature parameters of vibration spectrum kurtosis, temperature gradient deviation and current harmonic distortion. These core feature parameters are associated with the equipment physical degradation mechanism model to form a fault clue set. S6.2 Input the fault clue set formed in step S6.1 into the equipment structure knowledge graph and the fault propagation path model. The equipment structure knowledge graph includes the topological relationships of equipment components, material properties, and historical failure cases. The fault propagation path model simulates the dynamic evolution of a fault from the nascent stage to the diffusion stage. The graph neural network algorithm searches for matching fault patterns in the equipment structure knowledge graph, reverse-engineers the location of the fault source and the fault type, calculates the fault severity index, calculates the remaining service life prediction value using the Monte Carlo simulation method, and generates a fault diagnosis report. The fault diagnosis report includes three-dimensional fault component annotation, fault severity assessment, remaining service life prediction value, and maintenance recommendations. The fault diagnosis report is output to the maintenance terminal through a visualization interface.

8. The method for monitoring the status of mining equipment based on the Internet of Things according to claim 7, characterized in that: The specific method of step S7 is as follows: S7.1 Establish a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each warning, maintenance operations and actual fault verification results to form a fault case library, including equipment fault types, maintenance records, fault cause analysis and maintenance effect evaluation information. The system identifies the uncertainty of the model on fault samples, including false alarms, missed alarms or new fault types, and marks them as high-value samples. Maintenance experts manually annotate these samples and use the annotated samples to update the training set and optimize the model. S7.2 Input high-quality samples after active learning and annotation into the digital twin model, update the model parameters, and adjust the state of the twin model to adapt to the needs of equipment health evolution and fault diagnosis. Adopt the federated learning framework to achieve cross-mine model collaborative optimization through distributed computing while ensuring the data privacy of each mine. Each mine synchronizes and updates by sharing model parameters.

9. A mining equipment condition monitoring system based on the Internet of Things, characterized in that: The IoT-based mining equipment condition monitoring system is used to execute the IoT-based mining equipment condition monitoring method described in any one of claims 1 to 8 above, including a multimodal data acquisition module, an edge feature extraction and screening module, a digital twin virtual sample generation module, a deep neural network fault diagnosis module, a cloud-edge collaborative multi-level early warning module, an interpretable fault source tracing and diagnosis module, and a closed-loop feedback learning optimization module. The multimodal data acquisition module includes vibration sensors, temperature sensors, acoustic emission sensors, current sensors, and displacement sensors. It collects multi-dimensional physical parameters during equipment operation and dynamically adjusts the sampling frequency according to the current load status, operating mode, and historical health status of the equipment. The sampling frequency is increased during equipment start-up and shutdown and variable load conditions, and decreased during steady-state operation. The module also performs timestamp synchronization processing on the multi-source sensor data. The edge feature extraction and screening module deploys feature extraction algorithms on edge computing nodes. Combined with the physical degradation mechanism model of the equipment, it extracts feature parameters from the original sensor signals. The feature parameters include time-frequency domain features based on vibration signals, thermodynamic features based on temperature signals, and electrical features based on current signals. Through an adaptive feature selection mechanism, it calculates the correlation between features and their contribution to fault classification based on mutual information algorithm and ReliefF algorithm, filters feature subsets and removes redundant and weakly correlated features. The digital twin virtual sample generation module constructs a digital twin model of mining equipment. The digital twin model integrates the equipment's geometric structure model, physical degradation mechanism model, and behavioral operation law model. Different levels and types of fault parameters, including bearing wear parameters, gear crack parameters, and motor insulation aging parameters, are injected into the digital twin environment to simulate the equipment's response characteristics under various fault modes, generate virtual sensor data, and use generative adversarial networks to enhance the realism of the virtual samples. A virtual sample library is established, which includes normal state samples, early fault samples, mid-term fault samples, and severe fault samples. The deep neural network fault diagnosis module constructs a deep neural network fault diagnosis model, uses a transfer learning strategy for model training, pre-trains on a virtual sample library to learn general fault feature representations, fine-tunes using real fault samples from the target mine to optimize the decision boundary, reduces the distribution difference between virtual and real data through domain adaptation technology, introduces a feature fusion mechanism based on attention mechanism, dynamically weights feature parameters of different modalities, and outputs the probability distribution results of fault types. The cloud-edge collaborative multi-level early warning module deploys a lightweight fault diagnosis model at the edge node for real-time judgment. When the edge node detects abnormal features, it triggers a first-level early warning signal and uploads detailed data to the cloud. The cloud-based deep diagnosis model performs in-depth analysis of historical equipment operating data, current operating parameters, and environmental factors, quantifies and calculates the equipment health index, establishes a dynamic threshold adjustment mechanism based on Bayesian networks, corrects the early warning threshold according to operating condition fluctuations and environmental changes, and constructs a multi-level progressive early warning system, dividing the equipment status into four levels: healthy status, attention status, early warning status, and dangerous status. The interpretable fault tracing and diagnosis module initiates the fault tracing program after the early warning module issues an early warning signal. It generates a feature activation heatmap using gradient weighted activation mapping technology, displaying the frequency bands and temperature regions of interest to the fault diagnosis model. It generates an explanation report of the contribution of each feature parameter through the SHAP value calculation method, identifies the core decision features of vibration spectrum kurtosis, temperature gradient deviation, and current harmonic distortion, and associates these core decision features with the equipment structure knowledge graph and fault propagation path model. It then reverse-engineers the location of the fault source component and the fault type, calculates the fault severity index, predicts the remaining service life through Monte Carlo simulation, and generates a fault diagnosis report. The fault diagnosis report includes 3D fault component annotation, fault severity assessment, predicted remaining service life, and maintenance recommendations. The closed-loop feedback learning optimization module establishes a closed-loop mechanism from monitoring, diagnosis, maintenance to feedback. The system records the response to each warning, maintenance operations, and actual fault verification results, forming a real fault case library. This library includes fault types, maintenance records, fault cause analysis, and maintenance effectiveness evaluation information. It uses an active learning strategy to calculate the model's prediction uncertainty on specific samples, identifies high-value samples such as false alarms, missed alarms, and new fault types, and requests maintenance experts to manually annotate these high-value samples. The annotated samples are then incrementally updated to the training set, and the model is retrained. Equipment maintenance records and spare parts replacement information are incorporated into the digital twin model to correct physical parameters. A federated learning framework is adopted, where each mine trains its model locally and then uploads the model parameters to a cloud aggregation server for weighted averaging. The updated global model is then distributed to each mine, achieving cross-mine knowledge sharing and collaborative optimization while protecting the data privacy of each mine.

Images