Intelligent fault diagnosis early warning system of mine electromechanical equipment

Abstract

The invention discloses an intelligent fault diagnosis early warning system for mine electromechanical equipment, which comprises a data sensing module used for collecting multi-mode state data of the mine electromechanical equipment and carrying out anti-interference preprocessing, and an edge intelligent module used for receiving the preprocessed data and executing fault diagnosis and dynamic threshold adjustment. The cloud evolution module is used for receiving a diagnosis result and original data and completing self-evolution through data enhancement and model updating, and the application service module provides graded early warning, residual life prediction and operation and maintenance decisions based on the diagnosis result and evolution information; the method effectively captures early weak fault features through multi-modal data deep fusion in combination with a physical mechanism model and a data-driven algorithm, greatly improves the recognition accuracy of major faults, and improves the recognition accuracy of the major faults through deploying a lightweight CNN-LSTM model at an edge end and focusing on model evolution and data enhancement at a cloud end. And collaboration of edge quick response and cloud continuous optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of mining equipment operation and maintenance technology, and in particular to an intelligent fault diagnosis and early warning system for mining electromechanical equipment. Background Technology

[0002] As the core of mineral resource development, the operational reliability of mining electromechanical equipment is directly related to the safety and efficiency of mine production. However, the harsh mining environment and the long-term heavy-load and variable operating conditions of the equipment make it prone to failures such as bearing wear, gear cracks, aging of motor winding insulation, and hydraulic system leakage.

[0003] Traditional fault diagnosis methods often rely on single sensors, ignoring the correlation information of multi-source heterogeneous data such as temperature, current, and pressure. This makes it difficult to capture the multi-dimensional characteristics of complex faults. Furthermore, the scarcity of historical fault samples for mining equipment leads to insufficient generalization ability of traditional deep learning models. Real-time diagnosis is required for underground mining electromechanical equipment, but edge computing power is limited, while remote computing power is strong but suffers from large transmission delays. This results in a contradiction between real-time performance and computing power, making it difficult to adapt simultaneously. In addition, traditional diagnostic systems have poor dynamic adaptability and are prone to false alarms and missed alarms. Therefore, this invention proposes an intelligent fault diagnosis and early warning system for mining electromechanical equipment to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to propose an intelligent fault diagnosis and early warning system for mining electromechanical equipment. This intelligent fault diagnosis and early warning system for mining electromechanical equipment solves key problems such as efficient utilization of multi-source data in complex environments, accurate identification of small-sample faults, dynamic working condition adaptation, and real-time early warning by integrating multimodal perception, edge intelligent decision-making, and cloud self-evolution.

[0005] To achieve the objectives of this invention, the following technical solution is provided: an intelligent fault diagnosis and early warning system for mining electromechanical equipment, comprising a data sensing module, an edge intelligence module, a cloud evolution module, and an application service module. The data sensing module is used to collect multimodal state data of the mining electromechanical equipment and perform anti-interference preprocessing. The edge intelligence module is deployed on underground edge computing nodes to receive the preprocessed data and perform fault diagnosis and dynamic threshold adjustment. The cloud evolution module is deployed on a remote server to receive diagnostic results and raw data, complete self-evolution through data augmentation and model updates, and send optimized model parameters to the edge intelligence module. The application service module provides graded early warning, remaining life prediction, and operation and maintenance decisions based on diagnostic results and evolution information.

[0006] Further improvements are made in that: the data sensing module includes a multi-source sensing submodule and a preprocessing submodule. The multi-source sensing submodule collects vibration signals, surface temperature signals, drive motor current signals, early microcrack acoustic emission signals, and lubricating oil physicochemical property signals based on a multi-source sensor array. The preprocessing submodule collects signals based on a hybrid wiring method of optical fiber transmission and electromagnetic shielded cable, and combines algorithm filtering to remove noise, while simultaneously verifying sensor data to eliminate abnormal data.

[0007] Further improvements include: the multi-source sensor array includes a vibration acceleration sensor, an infrared temperature sensor, a current transformer, an acoustic emission sensor, and an oil sensor. For rotating mining machinery, the vibration acceleration sensor is symmetrically arranged along the axial, radial, and circumferential directions with a sampling rate ≥20kHz. The infrared temperature sensor monitoring points cover the bearing outer ring, motor housing, and key parts of the gearbox. The filtering and noise reduction specifically combines an adaptive notch filter algorithm to suppress power frequency interference, removes random noise through a wavelet packet decomposition-reconstruction algorithm, and then performs consistency verification on the dual redundant sensor data of key measuring points, eliminating abnormal data with deviations exceeding the threshold.

[0008] Further improvements are made in the following ways: The edge intelligence module includes a feature engineering submodule, a lightweight diagnostic submodule, and a dynamic threshold adjustment submodule. The feature engineering submodule is used to perform spatiotemporal alignment on the preprocessed multimodal data, extract features from vibration signals, temperature signals, and oil sensor data, and form a multidimensional joint feature vector. The lightweight diagnostic submodule establishes a diagnostic model based on a knowledge distillation CNN-LSTM hybrid network and outputs fault type classification results and confidence levels. The dynamic threshold adjustment submodule is based on a fuzzy logic controller that inputs current equipment operating parameters, outputs diagnostic threshold correction coefficients, and dynamically adjusts the alarm threshold of the lightweight diagnostic submodule.

[0009] Further improvements are made in the following aspects: the extracted features include the time-domain features, frequency-domain features, and time-frequency-domain features of the vibration signal, the sliding window variance of the temperature signal, the harmonic distortion rate of the current signal, and the feedforward neural network features of the oil sensor; in the diagnostic model, the teacher model is a full-size Transformer model trained in the cloud, and the student model extracts local features through convolutional layers and captures temporal dependencies through long short-term memory layers.

[0010] A further improvement is that the training of the diagnostic model includes:

[0011] S1. Data labeling: Based on the preprocessed data at the edge, combined with manual calibration and mechanism model simulation results, multi-label labeling is performed on the fault samples;

[0012] S2. Knowledge distillation uses the cloud-based Transformer model as the teacher and the edge-based CNN-LSTM model as the student. Through joint optimization of soft label loss and cross-entropy loss, the student model learns the decision boundary of the teacher, keeping the edge model lightweight.

[0013] S3. Model compression: Pruning techniques are used to remove redundant neurons and connections in the student model, while retaining key feature extraction paths to compress the model size to meet the memory limitations of edge embedded GPUs.

[0014] Further improvements are made in the following ways: The cloud-based evolution module includes a sample library construction submodule, a model evolution submodule, and a fusion submodule. The sample library construction submodule generates high-confidence simulated fault samples based on a conditional generative adversarial network, combined with stored historical fault samples and real-time data uploaded from the edge, and constructs a fault sample database. The model evolution submodule aggregates anonymized data from multiple mining equipment based on a federated learning framework, updates the cloud-based master model periodically, and encrypts and distributes the optimized model parameters to the edge intelligence module. The fusion submodule is used to perform residual analysis on the data-driven diagnostic results of the embedded equipment and the simulation results of the mechanism model to correct the prediction bias of the data-driven model.

[0015] Further improvements are made in that the federated learning framework adopts a horizontal federated learning mode, with participants being edge intelligent layer nodes from different mines. Each node updates the parameters of the cloud master model through encrypted gradient exchange without sharing the original data, thereby protecting the privacy of the mining equipment's operational data.

[0016] Further improvements are made in the following aspects: The application service module includes a graded early warning submodule, a life prediction submodule, and a decision generation submodule. The graded early warning submodule classifies early warning levels according to the degree of fault verification and pushes early warning information synchronously through downhole audible and visual alarms, mobile terminal APP, and ground monitoring screen. The life prediction submodule performs degradation modeling of the health indicators of key components based on LSTM network and predicts the probability of failure and remaining usable time in the next 30 days using multi-dimensional joint feature vectors as input. The decision generation submodule automatically recommends the optimal maintenance strategy by combining knowledge graph.

[0017] Further improvements are made in the following ways: The construction of the knowledge graph involves first extracting entities from historical maintenance records, equipment manuals, and expert knowledge, then defining the semantic relationships between entities, and finally using a graph neural network to intelligently match fault-maintenance strategies and output optimal maintenance recommendations with a confidence level of ≥90%.

[0018] The beneficial effects of this invention are as follows: By deeply fusing multimodal data, this invention breaks through the limitations of traditional single sensors, combines physical mechanism models and data-driven algorithms to effectively capture early and weak fault characteristics, and greatly improves the accuracy of identifying major faults; at the same time, it uses generative adversarial networks in the cloud to perform high-confidence simulation sample enhancement for small sample faults, and combines federated learning to aggregate multi-mine data to optimize the model, thereby improving the ability to identify rare faults; by deploying a lightweight CNN-LSTM model at the edge, it meets the low-latency early warning requirements of underground explosion-proof environments, while focusing on model evolution and data enhancement in the cloud, it achieves synergy between rapid edge response and continuous cloud optimization; and by combining fuzzy logic controllers and equipment operating parameters, it dynamically adjusts diagnostic thresholds and feature weights to adapt to changing operating conditions. Attached Figure Description

[0019] Figure 1 This is a diagram of the present invention. Detailed Implementation

[0020] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0021] As core equipment in mineral resource development, the operational reliability of mining electromechanical equipment directly affects the safety and efficiency of mine production. However, the harsh environment of mines, characterized by high dust levels, strong vibrations, humidity, and electromagnetic interference, forces equipment to operate under heavy loads and variable conditions for extended periods. This makes it prone to faults such as bearing wear, gear cracks, aging of motor winding insulation, and hydraulic system leaks. Traditional fault diagnosis methods have the following shortcomings:

[0022] Insufficient data utilization: Existing systems often rely on single sensors, such as vibration signals, ignoring the correlation information of multi-source heterogeneous data such as temperature, current, and pressure, making it difficult to capture the multi-dimensional characteristics of complex faults.

[0023] Weak diagnostic capability with small sample sizes: The probability of equipment failure in mining is low, especially major failures, and historical failure samples are scarce. Traditional deep learning models are prone to insufficient generalization ability due to "overfitting".

[0024] There is a conflict between real-time performance and computing power: downhole equipment needs real-time diagnosis, but edge computing power is limited and complex models are difficult to deploy; cloud computing is powerful, but data transmission delays may lead to delayed early warnings.

[0025] Poor dynamic adaptability: As the equipment enters different working conditions during the mining cycle, such as changes in the load of the hoist and fluctuations in the particle size of the material on the belt conveyor, fixed thresholds or static models cannot adapt to the evolution of working conditions, resulting in a high false alarm / missed alarm rate.

[0026] Based on the above issues, according to Figure 1As shown in the figure, this embodiment provides an intelligent fault diagnosis and early warning system for mining electromechanical equipment, including a data sensing module, an edge intelligence module, a cloud evolution module, and an application service module.

[0027] The data sensing module is used to collect multimodal status data of mining electromechanical equipment and perform anti-interference preprocessing.

[0028] The data sensing module includes a multi-source sensing submodule and a preprocessing submodule. The multi-source sensing submodule collects vibration signals, surface temperature signals, drive motor current signals, early microcrack acoustic emission signals, and lubricating oil physicochemical property signals from the device based on a multi-source sensor array.

[0029] The multi-source sensor array includes vibration acceleration sensors, infrared temperature sensors, current transformers, acoustic emission sensors, and oil sensors. For rotating mining machinery and equipment, such as hoist main shafts and belt conveyor drums, the vibration acceleration sensors are symmetrically arranged axially, radially, and circumferentially, with a sampling rate ≥20kHz. The infrared temperature sensors cover the outer ring of bearings, motor housings, and key parts of gearboxes, with a measurement accuracy ≤±0.5℃. The acoustic emission sensors have a frequency response range of 50kHz-1MHz and are used to capture high-frequency signals of early crack propagation.

[0030] The preprocessing submodule acquires signals based on a hybrid wiring method of fiber optic transmission and electromagnetic shielded cables, and combines algorithm filtering to remove noise, while also performing sensor data verification to eliminate abnormal data;

[0031] Specifically, the filtering and denoising process involves combining an adaptive notch filter algorithm to suppress 50Hz / 60Hz power frequency interference, removing random noise through a wavelet packet decomposition-reconstruction algorithm, and then performing consistency verification on the dual redundant sensor data at key measurement points to eliminate abnormal data with deviations exceeding the threshold.

[0032] The edge intelligence module is deployed on the downhole edge computing node to receive pre-processed data and perform fault diagnosis and dynamic threshold adjustment.

[0033] The edge intelligence module includes a feature engineering submodule, a lightweight diagnostic submodule, and a dynamic threshold adjustment submodule. The feature engineering submodule is used to perform spatiotemporal alignment on the preprocessed multimodal data based on the vibration signal timestamp, extract the features of the vibration signal, temperature signal, and oil sensor, and form a multidimensional joint feature vector.

[0034] The extracted features include time-domain features, frequency-domain features, and time-frequency-domain features of vibration signals; sliding window variance of temperature signals; harmonic distortion rate of current signals; and feedforward neural network features of oil sensors. Time-domain features include root mean square value, kurtosis, peak factor, etc.; frequency-domain features include envelope spectrum peak value, center frequency, etc.; and time-frequency-domain features include wavelet energy entropy, STFT time-frequency matrix, etc.

[0035] The lightweight diagnostic submodule establishes a diagnostic model based on a CNN-LSTM hybrid network with knowledge distillation and outputs the fault type classification results and confidence scores. The fault classification results include normal, bearing inner ring crack, gear pitting, motor winding short circuit, etc.

[0036] In the diagnostic model, the teacher model is a full-size Transformer model trained in the cloud, while the student model extracts local features through convolutional layers and captures temporal dependencies through long short-term memory layers. The parameter compression rate is ≥70%, and it supports millisecond-level inference at the edge.

[0037] Diagnostic model training includes:

[0038] S1. Data labeling: Based on the preprocessed data at the edge, combined with manual calibration and mechanism model simulation results, multi-label labeling is performed on the fault samples;

[0039] S2. Knowledge distillation uses the cloud-based Transformer model as the teacher and the edge-based CNN-LSTM model as the student. Through joint optimization of soft label loss and cross-entropy loss, the student model learns the decision boundary of the teacher, keeping the edge model lightweight.

[0040] S3. Model compression: Pruning techniques are used to remove redundant neurons and connections in the student model, while retaining key feature extraction paths to compress the model size to meet the memory limitations of embedded GPUs at the edge. The final model size is compressed to less than 50MB.

[0041] The dynamic threshold adjustment submodule is based on the input of current equipment operating parameters, such as hoist lifting speed, load rate, and belt conveyor material flow rate, into the fuzzy logic controller. It outputs diagnostic threshold correction coefficients to dynamically adjust the alarm thresholds of the lightweight diagnostic submodule to adapt to changing operating conditions.

[0042] The cloud-based evolution module is deployed on a remote server to receive diagnostic results and raw data. It completes self-evolution through data augmentation and model updates, and then sends the optimized model parameters to the edge intelligence module.

[0043] The cloud-based evolution module includes a sample library construction submodule, a model evolution submodule, and a fusion submodule. The sample library construction submodule generates high-confidence simulated fault samples based on a conditional generative adversarial network, combined with stored historical fault samples and real-time data uploaded from the edge, and builds a fault sample database. Specifically, for small-sample faults with an occurrence rate of <0.1%, a conditional generative adversarial network is used with normal samples as input to generate high-confidence simulated fault samples, thus solving the problem of sample class imbalance.

[0044] The model evolution submodule aggregates anonymized data from multiple mining equipment based on a federated learning framework, and regularly updates the cloud-based main model. For example, it embeds the temporal features of new fault modes into the global attention mechanism of the Transformer, and encrypts and distributes optimized model parameters such as convolutional kernel weights and recurrent unit gating parameters to the edge intelligence module.

[0045] The federated learning framework adopts a horizontal federated learning model, with participants being edge intelligent layer nodes from different mines. Each node updates the parameters of the cloud master model through encrypted gradient exchange without sharing the original data, in order to protect the privacy of the mining equipment's operating data.

[0046] The fusion submodule is used to perform residual analysis on the data-driven diagnostic results of the embedded device and the simulation results of the mechanism model, so as to correct the prediction bias of the data-driven model.

[0047] The application service module provides tiered early warnings, remaining life predictions, and operation and maintenance decisions based on diagnostic results and evolution information.

[0048] The application service module includes a graded early warning submodule, a life prediction submodule, and a decision generation submodule. The graded early warning submodule divides the early warning level according to the degree of fault verification, including Level 1 (normal), Level 2 (attention), Level 3 (warning), and Level 4 (shutdown). The early warning information is pushed synchronously through underground audible and visual alarms, mobile terminal APP (such as mining explosion-proof mobile phones), and ground monitoring screens. Among them, Level 3 and above early warnings are associated with historical fault case databases to display the fault development trend.

[0049] The life prediction submodule uses an LSTM network to model the health indicators of key components such as main motor bearings and hoist wire ropes, and uses a multi-dimensional joint feature vector as input to predict the failure probability and remaining usable time in the next 30 days, with a confidence interval of ±5%.

[0050] The decision generation submodule combines a knowledge graph that includes stored historical repair cases, spare parts inventory status, and expert repair experience to automatically recommend the optimal repair strategy, including repair time window, spare parts model (compatibility check), repair priority, and required human resources configuration.

[0051] The construction of the knowledge graph involves first extracting entities from historical maintenance records, equipment manuals, and expert knowledge, such as equipment components, fault types, maintenance operations, and spare parts. Then, semantic relationships between entities are defined, such as "bearing inner ring crack" → "required" → "ultrasonic flaw detection", "XX type bearing" → "compatible with" → "hoisting machine spindle". Finally, based on the graph neural network, intelligent matching of fault-maintenance strategies is completed, and the optimal operation and maintenance suggestions with a confidence level of ≥90% are output.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A mine electromechanical equipment intelligent fault diagnosis and early warning system, characterized in that: The application relates to a mine electromechanical equipment fault diagnosis system, which comprises a data sensing module, an edge intelligent module, a cloud evolution module and an application service module, the data sensing module is used for collecting multi-modal state data of mine electromechanical equipment and performing anti-interference pretreatment, the edge intelligent module is arranged on an underground edge computing node and is used for accepting the pretreated data, performing fault diagnosis and dynamic threshold adjustment, the cloud evolution module is arranged on a remote server and is used for accepting diagnosis results and original data, completing self-evolution through data enhancement and model updating, and issuing optimized model parameters to the edge intelligent module, and the application service module provides hierarchical early warning, residual life prediction and operation and maintenance decision based on the diagnosis results and evolution information.

2. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 1, characterized in that: The data sensing module comprises a multi-source sensing sub-module and a pretreatment sub-module, the multi-source sensing sub-module collects vibration signals, surface temperature signals, driving motor current signals, early micro-crack acoustic emission signals and lubricating oil physical and chemical characteristic signals of equipment based on a multi-source sensor array, and the pretreatment sub-module collects signals based on a mixed wiring mode of optical fiber transmission and electromagnetic shielding cable, and simultaneously performs sensor data checking and abnormal data elimination through algorithm filtering and denoising.

3. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 2, characterized in that: The multi-source sensor array comprises a vibration acceleration sensor, an infrared temperature sensor, a current transformer, an acoustic emission sensor and an oil sensor, wherein for a rotating mine electromechanical equipment, the vibration acceleration sensor is arranged symmetrically in the axial direction, the radial direction and the circumferential direction, the sampling rate is greater than or equal to 20 kHz, and the infrared temperature sensor monitors key parts of bearing outer rings, motor housings and gearboxes; the filtering and denoising is specifically combined with an adaptive notch filter algorithm to suppress power frequency interference, random noise is removed through a wavelet packet decomposition-reconstruction algorithm, and then the data of double-redundancy sensors of key measuring points are subjected to consistency checking, and abnormal data deviating from a threshold value by more than a threshold value are eliminated.

4. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 1, characterized in that: The edge intelligent module comprises a feature engineering sub-module, a lightweight diagnosis sub-module and a dynamic threshold adjustment sub-module, the feature engineering sub-module is used for performing time-space alignment on the pretreated multi-modal data, extracting features of vibration signals, temperature signals and oil sensors, and forming a multi-dimensional joint feature vector, the lightweight diagnosis sub-module establishes a diagnosis model based on a knowledge distillation CNN-LSTM hybrid network and outputs fault type classification results and confidence, and the dynamic threshold adjustment sub-module inputs current equipment working condition parameters based on a fuzzy logic controller and outputs a diagnosis threshold correction coefficient to dynamically adjust the alarm threshold of the lightweight diagnosis sub-module.

5. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 4, characterized in that: The extracted features include time domain features, frequency domain features and time-frequency domain features of vibration signals, a sliding window variance of temperature signals, a harmonic distortion rate of a current signal and a feedforward neural network feature of an oil sensor; in the diagnosis model, a teacher model is a full-size Transformer model trained in the cloud, and a student model extracts local features through a convolution layer and captures time sequence dependence through a long short-term memory layer.

6. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 4, characterized in that: The diagnosis model training comprises the following steps: S1, data labeling, based on the pretreated data of the edge end, combining artificial calibration and mechanism model simulation results, and labeling fault samples with multiple labels; S2, knowledge distillation, cloud end transformer model as teacher, edge end CNN-LSTM model as student, through the joint optimization of soft label loss and cross entropy loss, let the student model learn the decision boundary of the teacher, so that the edge end model remains lightweight; S3, model compression, using pruning technology to remove redundant neurons and connections in the student model, and retaining key feature extraction paths to compress the model size to meet the memory limit of the embedded GPU at the edge.

7. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 1, characterized in that: The cloud evolution module includes a sample library construction submodule, a model evolution submodule, and a fusion submodule. The sample library construction submodule generates high-confidence simulation fault samples based on conditional generative adversarial networks combined with stored historical fault samples and real-time data uploaded by the edge, constructs a fault sample database, the model evolution submodule aggregates anonymized data from multiple mine equipment based on a federated learning framework, regularly updates the cloud master model, and encrypts and distributes the optimized model parameters to the edge intelligent module, and the fusion submodule is used to perform residual analysis on the data-driven diagnosis results of the embedded device and the simulation results of the mechanism model, and correct the prediction bias of the data-driven model.

8. The intelligent fault diagnosis and early warning system of mine electromechanical equipment according to claim 7, characterized in that: The federated learning framework adopts a horizontal federated learning mode, and the participants are edge intelligent layer nodes of different mines. Each node updates the parameters of the cloud master model by exchanging encrypted gradients without sharing the original data to protect the privacy of the operation data of the mine equipment.

9. The intelligent fault diagnosis and early warning system for mine electromechanical equipment according to claim 1, characterized in that: The application service module includes a hierarchical early warning submodule, a life prediction submodule, and a decision generation submodule. The hierarchical early warning submodule divides the warning level according to the fault verification degree, and synchronously pushes the warning information through the underground sound and light alarm, mobile terminal APP and ground monitoring screen. The life prediction submodule models the degradation of the health index of the key components based on the LSTM network, and predicts the failure probability and remaining available time within the next 30 days with a multi-dimensional joint feature vector as input. The decision generation submodule automatically recommends the optimal maintenance strategy combined with the knowledge graph.

10. The intelligent fault diagnosis and early warning system of mine electromechanical equipment according to claim 9, characterized in that: The construction of the knowledge graph is specifically to extract entities from historical maintenance records, equipment manuals and expert knowledge, then define the semantic relationship between entities, and finally complete the intelligent matching of fault-maintenance strategy based on graph neural network and output the optimal operation and maintenance suggestion with a recommendation confidence of ≥90%.

Images