An intelligent fault management and feedback system based on changes in downhole equipment parameters

Abstract

This invention discloses an intelligent fault management and feedback system based on changes in underground equipment parameters, belonging to the field of coal mine equipment lifecycle management and intelligent fault diagnosis technology. The system is based on the main thread of equipment lifecycle information. It synchronously acquires equipment operating parameters, environmental parameters, and lifecycle management data through a multi-source heterogeneous data acquisition and spatiotemporal alignment module. After joint preprocessing, a dual-stream BERT temporal coding module extracts deep features. A cross-modal interactive modeling module achieves bidirectional fusion of environmental, equipment, and lifecycle data. A gated attention classification module completes fault classification and risk assessment. Finally, an intelligent feedback module links the equipment lifecycle management process to generate contextualized handling solutions, and a model continuous learning module ensures the system's long-term adaptability. This invention achieves collaborative modeling of environment, equipment, and lifecycle management, significantly improving fault diagnosis accuracy and equipment management closed-loop efficiency.

Description

Technical Field

[0001] This invention belongs to the field of coal mine equipment life cycle management and intelligent fault diagnosis technology, specifically involving an intelligent fault management and feedback system based on changes in underground equipment parameters. Background Technology

[0002] The underground working environment in mines is harsh and complex. Environmental factors such as high temperature, high humidity, high dust and unstable surrounding rock make underground equipment (such as belt conveyors, motors, reducers, etc.) prone to failure. Moreover, the causes of failure often involve the coupled influence of multiple dimensions such as equipment operating status, environmental changes and equipment life cycle status.

[0003] Existing technologies mostly collect equipment operating parameters or environmental parameters separately, failing to achieve effective correlation between equipment operating data, downhole environmental data, and equipment lifecycle management data (such as maintenance records, spare parts information, lubrication records, etc.). The data collection process suffers from low spatiotemporal alignment accuracy and unreliable data quality, resulting in a lack of comprehensive and reliable data support for subsequent fault diagnosis, making it difficult to accurately identify faults caused by multi-factor coupling. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent fault management and feedback system based on changes in downhole equipment parameters, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent fault management and feedback system based on changes in underground equipment parameters. This system, with the equipment's entire lifecycle information as its core, is used for comprehensive monitoring and intelligent diagnosis of the operating status of underground mining equipment in complex environments. It includes a multi-source heterogeneous data acquisition and spatiotemporal alignment module, an environment-equipment-lifecycle joint preprocessing and feature engineering module, a BERT-based dual-stream temporal coding and cross-modal interactive modeling module, a gated attention-based risk perception and classification module, an intelligent feedback and lifecycle process linkage module, and a model continuous learning and environment adaptive optimization module. The intelligent fault management and feedback system is deployed between the mining middleware platform and the equipment asset layer, supporting data interoperability with ERP and SCADA systems.

[0006] The multi-source heterogeneous data acquisition and spatiotemporal alignment module is used to simultaneously acquire three types of data and directly output them to the environment-device-lifecycle joint preprocessing and feature engineering module as the raw data source for preprocessing. The environment-device-lifecycle joint preprocessing and feature engineering module is used to perform denoising, missing value imputation, standardization, and multi-dimensional feature extraction on the received data, generating a three-channel input of device parameter sequence, environment parameter sequence, and lifecycle sequence. The BERT-based dual-stream temporal coding and cross-modal interaction modeling module is used to perform deep feature coding on the three-channel sequences to generate a fused multi-dimensional correlation information. The enhanced feature vector of the information is received by the gated attention-based risk perception classification module, which dynamically adjusts the feature weights through a dual-gating mechanism, outputs the probability distribution of fault types, risk levels, and life-cycle associated risks, and pushes the diagnostic results to the intelligent feedback and life-cycle process linkage module. The model continuous learning and environment adaptive optimization module detects data drift through KL divergence, triggers lightweight fine-tuning or federated learning updates, and synchronizes the optimized model parameters to the BERT-based dual-stream temporal coding and cross-modal interaction modeling module and the gated attention-based risk perception classification module.

[0007] Preferably, the three types of data include:

[0008] Equipment operating parameters: motor current, bearing vibration spectrum, hydraulic pressure, speed, oil particle size, and shell temperature;

[0009] Underground environmental parameters: tunnel temperature, relative humidity, methane concentration, CO concentration, wind speed, surrounding rock displacement rate, and ground sound activity intensity;

[0010] Equipment lifecycle management data: equipment planning information, inbound and outbound information, monitoring information, lubrication information, fault information, maintenance information, and scrapping information;

[0011] All data is collected synchronously with a unified timestamp and uploaded to the mining platform via an industrial IoT gateway using the MQTT / OPCUA protocol, achieving data spatiotemporal alignment.

[0012] Preferably, the environment-equipment-full lifecycle joint preprocessing and feature engineering module performs the following operations: denoising, missing value imputation, and standardization processing on the raw data, with the standardization formula being: , where x is the original value, and μ and σ are the historical mean and standard deviation of the corresponding parameters;

[0013] Extracting multi-dimensional features: equipment operation features, environmental correlation features, and full lifecycle features;

[0014] A three-channel input structure is constructed: device parameter sequence, environmental parameter sequence, and full life cycle sequence, which are then fed into the subsequent fusion network.

[0015] Preferably, the BERT-based dual-stream temporal coding and cross-modal interaction modeling module includes a three-branch Transformer encoder structure, which processes device, environment, and full lifecycle sequences respectively. Each branch contains no less than 6 layers of BERT-style coding blocks, uses mask language for pre-training, and introduces a double cross-attention mechanism in the high-level feature space to achieve bidirectional interaction modeling of "environment-device" and "device-full lifecycle". The calculation process is as follows:

[0016] C eqp Let C be the vector of environmental influence on the equipment. eqm H represents the impact vector of full lifecycle data on the device. eqp H is a characteristic of the device branch itself. env H represents the environmental branching characteristics. mgt For managing equipment lifecycle data, through residual connections Generate enhanced features.

[0017] Preferably, the risk perception classification module based on gated attention introduces an environment-lifecycle dual gating mechanism to dynamically adjust feature weights. The gating coefficient calculation formula is as follows: , where h env h represents a single-step vector representing environmental characteristics. mgt h represents a single-step vector representing the features across the entire lifecycle. eqp W represents a single-step vector representing a device feature. g b is a learnable weight matrix. g σ represents the learnable bias term, g represents the activation function, and g is the gating coefficient.

[0018] Preferably, when there are drastic changes in the environment or changes in the state throughout the entire life cycle, g will automatically amplify the weight of the corresponding features, allowing the model to focus on key influencing factors.

[0019] Preferably, the risk perception classification module based on gated attention further includes a main classifier. The main classifier adopts a structure of multi-head self-attention layer and fully connected layer. It takes the fused feature vector after cross-modal interaction modeling as input, generates a weighted context vector through multi-head self-attention layer to capture the global correlation between features, and inputs the weighted context vector into fully connected layer. It outputs the probability distribution of fault type through Softmax activation function, and outputs risk level and full life cycle associated risk through another branch fully connected layer.

[0020] Preferably, the specific calculation process of the main classifier is as follows: ,in The gated weighted feature vector To score attention, This represents the attention weight.

[0021] Preferably, the intelligent feedback and full lifecycle process linkage module is based on a three-element knowledge graph of environment-fault-full lifecycle countermeasures, and automatically generates a structured report containing fault causes, handling suggestions, spare parts scheduling schemes, and maintenance work orders.

[0022] Preferably, the model continuous learning and environment adaptive optimization module detects data drift through KL divergence. When the KL divergence exceeds the threshold, it triggers lightweight fine-tuning or federated learning updates and periodically feeds back the full lifecycle management data to the training set.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. This invention achieves spatiotemporal alignment by synchronously collecting three types of data: equipment operation, environment, and the entire lifecycle, providing a foundation for collaborative analysis. The system employs advanced data fusion algorithms to deeply process the collected multi-source heterogeneous data, effectively eliminating redundancy and conflicts and improving data quality. Based on the processed high-quality data, a multi-dimensional fault characteristic model is constructed to accurately characterize the fault features of equipment under different operating states and environmental conditions, providing a reliable basis for fault diagnosis. Simultaneously, machine learning algorithms are used to mine and analyze historical fault data, automatically learning fault modes and evolution patterns to achieve early warning and accurate prediction of faults, enabling the development of response strategies in advance and reducing the probability and impact of faults.

[0025] 2. This invention constructs a three-channel input structure by extracting multi-dimensional features from equipment, environment, and the entire lifecycle. This three-channel input structure can comprehensively capture various key information during equipment operation. The equipment channel focuses on the equipment's own operating parameters, such as speed, temperature, and pressure, accurately reflecting the equipment's real-time operating status. The environment channel covers downhole environmental factors, such as humidity, gas concentration, and geological conditions, which have a significant impact on equipment operation. The lifecycle channel records all relevant data from the equipment's commissioning to the present moment, including usage time, maintenance records, and replacement parts, providing a long-term basis for equipment health assessment. Through this three-channel input structure, the system can more accurately analyze the root causes of equipment failures, improving the accuracy and reliability of fault diagnosis.

[0026] 3. This invention utilizes a three-branch encoder and dual-cross attention to achieve bidirectional interactive modeling of three types of data, revealing their correlations. The three-branch encoder independently encodes data from the equipment channel, environment channel, and full lifecycle channel, fully exploring the inherent characteristics of each channel's data. The dual-cross attention mechanism cleverly establishes a bidirectional interactive channel between these three types of encoded data. It can automatically focus on key correlation information between data from different channels and dynamically adjust attention weights to deeply reveal the complex correlations between equipment operating parameters, downhole environmental factors, and equipment lifecycle data, thereby providing richer and more accurate feature representations for subsequent precise fault diagnosis and health assessment. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention;

[0028] Figure 2 This is a schematic diagram of the BERT-based multi-parameter temporal model network structure of the present invention. Detailed Implementation

[0029] 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.

[0030] Example 1

[0031] Please see Figure 1 and Figure 2 This invention provides an intelligent fault management and feedback system based on changes in underground equipment parameters. The intelligent fault management and feedback system takes the entire life cycle information of equipment as its core and is used to comprehensively monitor and intelligently diagnose the operating status of underground mining equipment in complex environments. It includes a multi-source heterogeneous data acquisition and spatiotemporal alignment module, an environment-equipment-full life cycle joint preprocessing and feature engineering module, a BERT-based dual-stream temporal coding and cross-modal interactive modeling module, a gated attention-based risk perception and classification module, an intelligent feedback and full life cycle process linkage module, and a model continuous learning and environmental adaptive optimization module. The intelligent fault management and feedback system is deployed between the mining middle platform and the equipment asset layer and supports data interoperability with ERP and SCADA systems.

[0032] The multi-source heterogeneous data acquisition and spatiotemporal alignment module is used to simultaneously acquire three types of data and directly output them to the environment-equipment-lifecycle joint preprocessing and feature engineering module as the original data source for preprocessing. The environment-equipment-lifecycle joint preprocessing and feature engineering module is used to perform denoising, missing value imputation, standardization and multi-dimensional feature extraction on the received data, generating three-channel inputs: equipment parameter sequence, environmental parameter sequence and lifecycle sequence. The BERT-based dual-stream temporal coding and cross-modal interactive modeling module is used to perform deep feature coding on the three-channel sequence to generate enhanced feature vectors that integrate multi-dimensional correlation information. The gated attention-based risk perception classification module is used to receive the enhanced feature vectors and dynamically adjust the feature weights through a dual-gating mechanism, outputting the fault type probability distribution, risk level and lifecycle-related risks, and pushing the diagnostic results to the intelligent feedback and lifecycle process linkage module. The model continuous learning and environment adaptive optimization module detects data drift through KL divergence, triggers lightweight fine-tuning or federated learning updates, and synchronizes the optimized model parameters to the BERT-based dual-stream temporal coding and cross-modal interactive modeling module and the gated attention-based risk perception classification module.

[0033] Furthermore, the relationships between the modules are as follows: The multi-source heterogeneous data acquisition and spatiotemporal alignment module is used to simultaneously acquire three types of data and directly outputs them to the environment-equipment-full lifecycle joint preprocessing and feature engineering module as the raw data source for preprocessing. Its output three-channel sequence is then input into the three branch encoders of the BERT-based dual-stream temporal coding and cross-modal interaction modeling module, providing structured input for feature fusion. The enhanced feature vectors generated by the BERT-based dual-stream temporal coding and cross-modal interaction modeling module are input to the gated attention-based risk perception classification module. After gating and weighting, the main classifier completes the fault and risk determination. The fault type, risk level, and full lifecycle associated risk results output by the classification module are fed into the intelligent feedback and full lifecycle process linkage module as the core basis for knowledge graph matching and report generation. The actual handling results (including manual confirmation information) and full lifecycle management data (maintenance records, scrapping information, etc.) from the intelligent feedback module flow back to the model continuous learning and environment adaptive optimization module, providing data support for model updates. After the optimization module completes model fine-tuning or federated learning updates, it synchronizes the updated model parameters to the cross-modal interactive modeling module and the risk perception classification module to achieve dynamic optimization of model performance.

[0034] The three types of data include:

[0035] Equipment operating parameters: motor current, bearing vibration spectrum, hydraulic pressure, speed, oil particle size, and casing temperature. Downhole environmental parameters: tunnel temperature, relative humidity, methane concentration, CO concentration, wind speed, surrounding rock displacement rate, and ground sound activity intensity. Equipment lifecycle management data: equipment planning information, inventory entry / exit information, monitoring information, lubrication information, fault information, maintenance information, and scrapping information. All data is collected synchronously with a unified timestamp and uploaded to the mining platform via an industrial IoT gateway using the MQTT / OPCUA protocol, achieving data spatiotemporal alignment.

[0036] Furthermore, details of equipment operating parameter acquisition are as follows: Motor current is acquired using a Hall current sensor installed on the motor stator winding (measuring range covering 0-150% of the equipment's rated current), reflecting the motor load status in real time. Bearing vibration spectrum is acquired using a piezoelectric vibration sensor (sampling frequency 2kHz) to collect X / Y / Z triaxial vibration signals, which are then converted into spectral data via Fast Fourier Transform (FFT) to identify bearing wear, imbalance, and other fault characteristics. Oil particle size is acquired using an online oil monitoring sensor, outputting NAS1-12 level cleanliness data, which is correlated with the equipment's lubrication system failure risk.

[0037] The monitoring range of underground environmental parameters is as follows: Tunnel temperature and relative humidity are collected using integrated temperature and humidity sensors (ranges -20~80℃ and 0~100%RH, respectively), deployed at intervals not exceeding 50m to cover the entire tunnel area. Methane and CO concentrations are collected using explosion-proof gas sensors (measurement accuracy ±0.01%). When the methane concentration ≥1.0%, the sampling frequency is automatically increased to 2Hz. The surrounding rock displacement rate is collected using a combination of anchor stress sensors and displacement gauges, and the ground sound activity intensity is collected using a ground sound monitor (sensitivity 10dB). These two sensors work together to reflect the impact of tunnel surrounding rock stability on the equipment.

[0038] Field specifications for equipment lifecycle management data: Equipment planning information includes plan number, equipment number, plan content, and execution cycle. Inbound / outbound information includes inbound / outbound order number, equipment / spare part code, quantity, operation time, and responsible person. Lubrication information includes lubrication location, lubrication medium, lubrication cycle, and last lubrication time. Fault information includes fault order number, equipment number, fault type, occurrence time, and handling result. All fields adopt the unified coding standards of the mining equipment management system to ensure data associativity and traceability.

[0039] The spatiotemporal alignment is achieved by: The industrial IoT gateway has a built-in time synchronization module that calibrates the local clock using the GPS clock source (time synchronization accuracy ±1ms) from the mining platform. Time-series data such as equipment operating parameters and environmental parameters are tagged with a unified timestamp at 1-second intervals. Event-based data, such as equipment lifecycle management data, is automatically mapped to the nearest 1-second time node based on the event trigger time, ensuring that the alignment error of the three types of data in the time dimension does not exceed 500ms. This provides a spatiotemporally consistent foundation for subsequent cross-modal interaction modeling.

[0040] The Environment-Equipment-Full Lifecycle Joint Preprocessing and Feature Engineering module performs the following operations: denoising, missing value imputation, and standardization on the raw data. The standardization formula is: Where x is the original value, and μ and σ are the historical mean and standard deviation of the corresponding parameters. Multi-dimensional features are extracted: equipment operation features, environmental correlation features, and full lifecycle features. A three-channel input structure is constructed: equipment parameter sequence, environmental parameter sequence, and full lifecycle sequence, which are then fed into the subsequent fusion network.

[0041] Furthermore, the denoising strategy is differentiated as follows: Vibration spectrum data in equipment operating parameters are denoised using a db4 wavelet basis (5-level decomposition) to retain the impact signals characteristic of faults. Gas concentration data in downhole environmental parameters is smoothed using Kalman filtering to reduce concentration fluctuations. Equipment lifecycle management data (event-related) does not require denoising; only field integrity is verified.

[0042] Missing value imputation adaptation rules: When the duration of missing time-series data for device operation / environment is ≤30s, linear interpolation is used for imputation. When the duration of missing data is >30s, backup sensor data from the same location is retrieved. For missing fields in the full lifecycle management data, the system automatically initiates a supplementary data entry request to the corresponding management module.

[0043] The specific details of multi-dimensional features:

[0044] Equipment operating characteristics: mean, variance, peak factor, kurtosis, and skewness within a sliding window (10s).

[0045] Environmental correlation characteristics: the correlation coefficient between the rate of change of environmental parameters (such as the change in temperature within 1 minute) and equipment parameters.

[0046] Lifecycle characteristics: equipment service duration, cumulative number of maintenance operations, time interval between the most recent lubrication operations, and spare parts replacement records.

[0047] Three-channel sequence structure specifications: Equipment parameter sequence T is the time step (default 10), D p For device feature dimensions (default 12). Environmental parameter sequence. :D v Environmental feature dimension (default 8). Full lifecycle sequence. :D m The three feature dimensions (default 5) are strictly aligned in time step to ensure the temporal consistency of the subsequent fusion network.

[0048] The BERT-based dual-stream temporal coding and cross-modal interaction modeling module includes a three-branch Transformer encoder structure, which processes device, environment, and full lifecycle sequences respectively. Each branch contains no less than 6 layers of BERT-style coding blocks. It employs masked language modeling pre-training and introduces a double cross-attention mechanism in the high-level feature space to achieve bidirectional interaction modeling of "environment-device" and "device-full lifecycle". The computation process is as follows:

[0049] C eqp Let C be the vector of environmental influence on the equipment. eqm H represents the impact vector of full lifecycle data on the device. eqp H is a characteristic of the device branch itself. env H represents the environmental branching characteristics. mgt For managing equipment lifecycle data, through residual connections Generate enhanced features.

[0050] Furthermore, CrossAttn(A, B) is the cross-attention calculation function. With A as the query vector and B as the key and value vectors, it calculates the influence weight of B on A and generates enhanced features.

[0051] The structural parameters of the three-branch Transformer encoder are as follows: each BERT-style encoding block contains 8 self-attention layers (64 dimensions per layer), layer normalization (ε=1e-5), and a feedforward network (512 hidden layer dimensions, GELU activation function). Masked Language Modeling (MLM) is used during the pre-training phase, with a masking ratio of 15%. The pre-training dataset consists of the mine's equipment operation data (including environmental and full lifecycle information) from the past year.

[0052] The computational logic of the double cross-attention mechanism:

[0053] "Environment-Device" Cross-Attention: Based on Device Branch Features H eqp For the query, the environmental branch feature H env For key / value pairs, calculate the weight of the impact of the environment on the device.

[0054] "Device-Lifecycle" Cross-Attention: Based on Device Branch Features H eqpFor the query, the full lifecycle branch feature H mgt Using key / value pairs, calculate the weights of the impact of lifecycle data on the device. The output vector C of both pairs... eqp and C eqm All need to undergo layer normalization before being compared with the original device feature H. eqp Perform residual connection.

[0055] Validation of the enhanced features: When environmental parameters (such as tunnel temperature) exceed the threshold, C eqp This will significantly increase the weight of equipment temperature-related features. When the cumulative number of equipment repairs reaches a threshold, C eqm It will amplify the weight of features related to equipment performance degradation, thereby achieving a precise correlation between the environment and the entire life cycle of equipment status.

[0056] The risk perception and classification module based on gated attention introduces a dual gating mechanism of environment and life cycle, dynamically adjusting feature weights. The gating coefficient is calculated using the following formula: , where h env h represents a single-step vector representing environmental characteristics. mgt h represents a single-step vector representing the features across the entire lifecycle. eqp W represents a single-step vector representing a device feature. g b is a learnable weight matrix. g σ represents the learnable bias term, g is the gating coefficient, and g is the activation function. When the environment changes drastically or the lifecycle state changes, g automatically amplifies the weights of the corresponding features, allowing the model to focus on key influencing factors. The risk perception classification module based on gated attention also includes a main classifier. The main classifier adopts a structure of multi-head self-attention layers and fully connected layers. It takes the fused feature vector after cross-modal interaction modeling as input, generates a weighted context vector through the multi-head self-attention layer to capture the global correlation between features, and inputs the weighted context vector into the fully connected layer. The probability distribution of fault types is output through the Softmax activation function, while another fully connected layer outputs the risk level and lifecycle-related risks. The specific calculation process of the main classifier is as follows: ,in The gated weighted feature vector To score attention, This represents the attention weight.

[0057] Furthermore, the parameters and functions of the dual-gating mechanism are refined: weight matrix W g The dimensions are 256 × (environmental characteristic dimension, lifecycle characteristic dimension, and device characteristic dimension), with bias term b. g The initial value is set to 0.1. The activation function σ output ranges from 0 to 1. When g approaches 1, the corresponding feature weights are amplified. When g approaches 0, the corresponding feature weights are suppressed.

[0058] The computational process of the main classifier is complete: First, the input feature vector is processed by a gating and weighting module to obtain a gating and weighted feature vector. Next, an attention score is calculated using an attention mechanism, reflecting the importance of different features. Then, attention weights are calculated based on the attention scores, and these weights are used to perform a weighted summation of the feature vectors. During the weighted summation, features with larger weights contribute more to the final result, while features with smaller weights contribute less. Finally, the weighted summation result is input into the classifier to obtain the final classification result.

[0059] The logic behind the output results is as follows: The fault type probability distribution (Softmax output) includes 8 typical faults of downhole equipment (such as bearing failure and motor overheating). The risk level is divided into 4 levels: low, medium, high, and emergency, which are strongly correlated with the fault confidence level (confidence level > 0.95 corresponds to emergency risk). The risks associated with the entire life cycle include insufficient spare parts inventory and tight maintenance resources, which directly link to the process scheduling of the subsequent intelligent feedback module.

[0060] The intelligent feedback and full lifecycle process linkage module is based on a three-element knowledge graph of environment-fault-full lifecycle countermeasures, which automatically generates a structured report containing fault causes, handling suggestions, spare parts scheduling plans, and maintenance work orders.

[0061] Furthermore, the matching logic of the ternary knowledge graph is as follows: using environmental features and fault types as search keys, the corresponding full lifecycle countermeasures are matched in the knowledge graph. For example, for high temperature of 35℃ and motor insulation aging, the corresponding spare parts library is used to retrieve insulating paint, open the tunnel ventilation, and dispatch a maintenance work order. The matching process uses a semantic similarity algorithm (such as cosine similarity), prioritizing historical cases with a confidence score ≥0.85. If no matching case is found, the basic solution is generated by calling the expert rule base.

[0062] Content and linkage methods of structured reports:

[0063] Causes of failure: By combining the characteristics of environment-equipment interaction, identify the causes of failure based on environmental factors (such as high temperature) and the full life cycle status (such as 7 months of service).

[0064] Spare parts scheduling scheme: Automatically associate spare parts inventory information with the equipment's full lifecycle management data, and trigger the procurement plan module if the inventory is insufficient.

[0065] Maintenance work order: includes work order number, equipment number, maintenance content and priority (associated with risk level), and is directly pushed to the maintenance management process module of the mining platform.

[0066] Multi-channel push and feedback mechanism: Reports can be pushed to the mobile app of maintenance personnel and the large screen of the dispatch center. After the maintenance is completed, the manually confirmed handling results will flow back to the knowledge graph and model training set to achieve two-way optimization of knowledge and model.

[0067] The process for obtaining the ternary knowledge graph of environment-fault-full lifecycle countermeasures is as follows:

[0068] Process 1: Initial Construction of Basic Map: By mining the full life cycle data of existing coal mine equipment, the correlation patterns of environmental conditions, fault types, and handling processes are extracted to form an initial map.

[0069] Data collection scope: Historical environmental monitoring records (e.g., "tunnel temperature 35℃, relative humidity 80%", "gas concentration 1.2%"). Equipment failure logs (e.g., "motor insulation aging", "conveyor belt misalignment"). Equipment planning, inbound / outbound, maintenance, and spare parts scheduling records (e.g., "application for insulating varnish spare parts in August 2024", "shutdown for maintenance in September 2024", etc.).

[0070] Association pattern extraction: Analyze the data using association rule algorithms (such as Apriori) to extract frequently occurring triples, for example:

[0071] Three-element group 1: Environment: High temperature 35℃ + high humidity 80%, Fault: Motor insulation aging, Full life cycle countermeasures: Take out insulation paint from the spare parts warehouse, turn on the ventilation in the tunnel and stop the machine to replace the insulation layer.

[0072] Triad 2: Environment: Surrounding rock displacement rate 5mm / d, Fault: Conveyor belt misalignment, Full life cycle countermeasures: Equipment maintenance module dispatches adjustment work orders, spare parts warehouse has conveyor belt correction wheels.

[0073] Phase 2, Knowledge Graph Rule Enhancement: Invite coal mine equipment operation and maintenance and management experts to transform industry standards and practical experience into knowledge graph rule entries, supplementing scenarios that cannot be covered by data.

[0074] Expert input:

[0075] The corresponding handling logic of industry standards: For example, "When the gas concentration is ≥1.5%, the explosion-proof procedure should be initiated first when handling the fault."

[0076] Emergency response measures for high-risk scenarios: such as environment: CO concentration exceeding 0.002%, fault: blockage of ventilation fan duct, full life cycle response: the equipment scheduling module switches to the backup ventilation fan and the safety evaluation module triggers a safety rating update.

[0077] How the rules are implemented:

[0078] Expert experience is broken down into structured triples, embedded into the relationships in the knowledge graph, and labeled with an "expert verification" tag to increase its priority.

[0079] Phase 3, Dynamic Updates of the Map: During system operation, the map is continuously improved through feedback from real-world cases and the model.

[0080] New Case Supplement: After the system diagnoses and handles a new fault, the environmental parameters - fault type - actual executed full lifecycle process (verified manually) will be automatically added as a new triplet, for example:

[0081] New Triad: Environment: Ground noise intensity 90dB, Fault: Broken gear teeth in reducer, Life cycle countermeasures: Equipment scrapping module assesses residual value and procurement module initiates new reducer procurement plan.

[0082] Weight optimization: Statistically analyze the actual application frequency of triples, increase the matching weight for high-frequency effective cases, and reduce the weight or mark invalid cases as pending verification to ensure the practicality of the graph.

[0083] The model continuous learning and environment adaptive optimization module detects data drift through KL divergence. When the KL divergence exceeds the threshold, it triggers lightweight fine-tuning or federated learning updates and periodically feeds back the full lifecycle management data to the training set.

[0084] Furthermore, the specific rules for data drift detection are as follows: the KL divergence calculation targets the "current input data distribution" and the "benchmark data distribution during model training," with a threshold set to 0.5. Adversarial validation (AUC > 0.7) is used as a supplementary criterion to avoid misjudgment based on a single indicator, ensuring an accuracy rate of ≥98% for data drift detection.

[0085] Model update adaptation strategy: Lightweight fine-tuning adopts LoRA technology (rank r=8, learning rate 5e-5), only updating the low-rank matrix of the model's attention layer, with fine-tuning time ≤3 hours. For cross-mine scenarios, federated learning (FedAvg algorithm) is adopted, where each mine node only uploads the model parameter gradients rather than the original data to ensure data privacy, and the parameter update cycle is 7 days / time.

[0086] Standardized data feedback throughout the entire lifecycle: Feedback data must include manually confirmed fault diagnosis results, maintenance and handling records, and equipment scrapping assessment information. After data cleaning (duplicate removal and validity labeling), it should be added to the training set. The training set should be incrementally updated quarterly, and after the update, the model's fault diagnosis F1-score under local operating conditions should improve by ≥5%.

[0087] Example 2

[0088] This embodiment is deployed in transport roadway No. 3, targeting the DSJ1200 belt conveyor, and verifies the effectiveness of the system by linking the entire life cycle process of equipment planning, warehousing, and maintenance.

[0089] Multi-source heterogeneous data acquisition and spatiotemporal alignment. Equipment operation data: motor current (CSM300 sensor), bearing vibration (YD-105 sensor), casing temperature (DS18B20), sampling frequency 2Hz. Environmental data: roadway temperature (SHT35), gas concentration (KGJ16), surrounding rock displacement (YHW150), sampling frequency 1Hz. Full lifecycle data: equipment plan (procurement in 2024), inventory records (warehousing in May 2024), maintenance records (bearing replacement in August 2024), synchronized to the mine management platform.

[0090] Combined preprocessing and feature engineering. Feature extraction: equipment operating characteristics (current variance, vibration peak factor), environmental characteristics (temperature change rate), and life-cycle characteristics (7 months of service, 1 maintenance). Constructing a three-channel input: , , .

[0091] Cross-modal interactive modeling. After three-branch BERT encoding, double cross-attention calculation shows that the influence weight of "tunnel temperature 35℃, service for 7 months" on the aging characteristics of motor insulation reaches 45%, revealing the interactive fault path of "high environmental temperature and equipment aging".

[0092] Fault classification and process linkage. The dual-gating mechanism outputs "Motor insulation aging (confidence level 92%, high risk), spare parts inventory sufficient." The intelligent feedback module links with the equipment maintenance management process, automatically dispatches maintenance work orders, and triggers the insulation varnish release application from the spare parts inventory management module.

[0093] Furthermore, the system traces the equipment's historical maintenance records using a knowledge graph, discovering that the insulation components were not replaced simultaneously when the bearing was replaced in August 2024. Combining this with current service life and temperature data, a causal chain of reasoning is generated: "High temperatures accelerate the deterioration of insulation materials." The maintenance work order automatically links to Article 4.3.2 of the "Coal Mine Electromechanical Equipment Maintenance Standard," recommending the replacement with Nomex insulation paper of temperature resistance grade H, and simultaneously pushes the information to the maintenance team's mobile terminal. The spare parts management module automatically generates a procurement plan based on the inventory threshold (triggers an alert when ≤5 sets), updating the demand for insulating varnish (model JZ-200) to the ERP system. Simultaneously, it coordinates with the transportation scheduling module to plan the optimal retrieval route, ensuring maintenance personnel can obtain spare parts within 2 hours.

[0094] The working principle and usage process of this invention are as follows: S1. Synchronously collect three types of data: equipment operation, environment, and full life cycle data, and upload them to the mining platform after spatiotemporal alignment.

[0095] S2. Preprocess the data and extract multi-dimensional features to construct a three-channel input sequence.

[0096] S3. Deep features of each sequence are extracted through a three-branch BERT encoder and interactive fusion is achieved through double cross attention.

[0097] S4. Use a dual-gating mechanism to adjust feature weights to complete fault classification and risk assessment.

[0098] S5. Generate structured reports based on ternary knowledge graphs and trigger action by linking the entire lifecycle process modules.

[0099] S6. Feed the processing results back to the training set to trigger model fine-tuning and achieve continuous optimization.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent fault management and feedback system based on changes in downhole equipment parameters, characterized in that, The intelligent fault management and feedback system takes the equipment's entire lifecycle information as its core and is used to comprehensively monitor and intelligently diagnose the operating status of underground mining equipment in complex environments. It includes a multi-source heterogeneous data acquisition and spatiotemporal alignment module, an environment-equipment-entire lifecycle joint preprocessing and feature engineering module, a BERT-based dual-stream temporal coding and cross-modal interactive modeling module, a gated attention-based risk perception and classification module, an intelligent feedback and lifecycle process linkage module, and a model continuous learning and environmental adaptive optimization module. The intelligent fault management and feedback system is deployed between the mining middle platform and the equipment asset layer and supports data interoperability with ERP and SCADA systems. The multi-source heterogeneous data acquisition and spatiotemporal alignment module is used to simultaneously acquire three types of data and directly output them to the environment-device-lifecycle joint preprocessing and feature engineering module as the raw data source for preprocessing. The environment-device-lifecycle joint preprocessing and feature engineering module is used to perform denoising, missing value imputation, standardization, and multi-dimensional feature extraction on the received data, generating a three-channel input of device parameter sequence, environment parameter sequence, and lifecycle sequence. The BERT-based dual-stream temporal coding and cross-modal interaction modeling module is used to perform deep feature coding on the three-channel sequences to generate a fused multi-dimensional correlation information. The enhanced feature vector of the information is received by the gated attention-based risk perception classification module, which dynamically adjusts the feature weights through a dual-gating mechanism, outputs the probability distribution of fault types, risk levels, and life-cycle associated risks, and pushes the diagnostic results to the intelligent feedback and life-cycle process linkage module. The model continuous learning and environment adaptive optimization module detects data drift through KL divergence, triggers lightweight fine-tuning or federated learning updates, and synchronizes the optimized model parameters to the BERT-based dual-stream temporal coding and cross-modal interaction modeling module and the gated attention-based risk perception classification module.

2. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 1, characterized in that, The three types of data include: Equipment operating parameters: motor current, bearing vibration spectrum, hydraulic pressure, speed, oil particle size, and shell temperature; Underground environmental parameters: tunnel temperature, relative humidity, methane concentration, CO concentration, wind speed, surrounding rock displacement rate, and ground sound activity intensity; Equipment lifecycle management data: equipment planning information, inbound and outbound information, monitoring information, lubrication information, fault information, maintenance information, and scrapping information; All data is collected synchronously with a unified timestamp and uploaded to the mining platform via an industrial IoT gateway using the MQTT / OPCUA protocol, achieving data spatiotemporal alignment.

3. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 1, characterized in that, The environment-equipment-full lifecycle joint preprocessing and feature engineering module performs the following operations: denoising, missing value imputation, and standardization on the raw data, with the standardization formula as follows: , where x is the original value, and μ and σ are the historical mean and standard deviation of the corresponding parameters; Extracting multi-dimensional features: Equipment operating characteristics, environmental characteristics, and full life cycle characteristics; A three-channel input structure is constructed: device parameter sequence, environmental parameter sequence, and full life cycle sequence, which are then fed into the subsequent fusion network.

4. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 3, characterized in that, The BERT-based dual-stream temporal coding and cross-modal interaction modeling module includes a three-branch Transformer encoder structure, which processes device, environment, and full lifecycle sequences respectively. Each branch contains no less than 6 layers of BERT-style coding blocks. It uses mask language for pre-training and introduces a double cross-attention mechanism in the high-level feature space to achieve bidirectional interaction modeling of "environment-device" and "device-full lifecycle". The calculation process is as follows: C eqp Let C be the vector of environmental influence on the equipment. eqm H represents the impact vector of full lifecycle data on the device. eqp H is a characteristic of the device branch itself. env H represents the environmental branching characteristics. mgt For managing equipment lifecycle data, through residual connections Generate enhanced features.

5. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 1, characterized in that, The risk perception classification module based on gated attention introduces a dual gating mechanism of environment and life cycle, dynamically adjusting feature weights. The gating coefficient is calculated using the following formula: , where h env h represents a single-step vector representing environmental characteristics. mgt h represents a single-step vector representing the features across the entire lifecycle. eqp W represents a single-step vector representing a device feature. g b is a learnable weight matrix. g σ represents the learnable bias term, g represents the activation function, and g is the gating coefficient.

6. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 5, characterized in that, When the environment changes drastically or the state changes throughout the entire life cycle, g will automatically amplify the weight of the corresponding features, allowing the model to focus on key influencing factors.

7. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 1, characterized in that, The risk perception classification module based on gated attention also includes a main classifier. The main classifier adopts a structure of multi-head self-attention layer and fully connected layer. It takes the fused feature vector after cross-modal interaction modeling as input, generates a weighted context vector through multi-head self-attention layer to capture the global correlation between features, and inputs the weighted context vector into fully connected layer. It outputs the probability distribution of fault type through Softmax activation function, and outputs risk level and full life cycle associated risk through another branch fully connected layer.

8. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 7, characterized in that, The specific calculation process of the main classifier is as follows: ,in The gated weighted feature vector To score attention, This represents the attention weight.

9. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 1, characterized in that, The intelligent feedback and full lifecycle process linkage module is based on a ternary knowledge graph of environment-fault-full lifecycle countermeasures, and automatically generates a structured report containing fault causes, handling suggestions, spare parts scheduling plans, and maintenance work orders.

10. The intelligent fault management and feedback system based on changes in downhole equipment parameters according to claim 1, characterized in that, The model continuous learning and environment adaptive optimization module detects data drift through KL divergence. When the KL divergence exceeds the threshold, it triggers lightweight fine-tuning or federated learning updates and periodically feeds back the full lifecycle management data to the training set.

Images