Intelligent identification and safety intervention method for illegal behaviors of underground workers in non-coal mines

Abstract

This application discloses an intelligent identification and safety intervention method for violations by underground workers in non-coal mines, including the following steps: S1. Multi-source sensing data is collected by various sensors deployed in the underground working area of ​​non-coal mines. This integrates four types of data: infrared and visible light video, UWB personnel positioning, underground environmental parameters, and equipment operating status, significantly improving the completeness and reliability of monitoring complex underground scenarios. A model is trained based on a real underground working dataset, and by combining target tracking, attitude estimation, spatiotemporal graph convolutional networks, and attention mechanisms, personnel trajectories, human skeletons, and spatiotemporal behavioral features are extracted, significantly improving the stability and accuracy of identification. By comprehensively considering multiple factors such as violation type, identification confidence, violation duration, and environmental and equipment status, safety risks are quantified and classified into four risk levels: low, medium, high, and emergency. On-site personnel are quickly assigned to handle the situation, significantly shortening the response time and preventing the risk from escalating.

Description

Technical Field

[0001] This application relates to the technical field of mine monitoring, and in particular to a method for intelligent identification and safety intervention of violations by underground workers in non-coal mines. Background Technology

[0002] The underground working environment in non-coal mines is characterized by enclosed spaces, insufficient lighting, and high dust concentrations, posing extremely high safety risks compared to surface operations. Non-standard operations and violations by underground workers are the core causes of safety accidents such as collapses, mechanical injuries, hazardous gas leaks, and equipment malfunctions. Therefore, accurately identifying violations by underground personnel and promptly intervening in safety matters are crucial for ensuring safe production in non-coal mines. Currently, safety management methods in non-coal mines have significant shortcomings and are ill-suited to the safety control needs of complex underground environments.

[0003] First, current underground violation monitoring relies heavily on single visible light video surveillance, which can only achieve visual image acquisition. Underground environments have problems such as dust obstruction, fluctuating light intensity, and many blind spots. Traditional single video monitoring is prone to problems such as blurry images, missed targets, and false detections. It cannot accurately capture subtle violations by personnel and ignores the cumulative impact of environmental and equipment risks such as gas concentration, oxygen concentration, and equipment failure on violations, resulting in insufficient monitoring comprehensiveness.

[0004] Secondly, traditional monitoring systems only have simple image capture and post-event review functions, lacking autonomous intelligent recognition capabilities. They cannot identify various violations in real time, such as not wearing safety equipment, operating in dangerous areas without authorization, and improper equipment operation. Furthermore, traditional underground violation handling methods primarily rely on manual warnings and post-event accountability. When violations are detected, they cannot trigger differentiated intervention measures based on risk levels, relying solely on manual on-site intervention, resulting in high response delays and untimely handling.

[0005] Application content

[0006] This application aims to address, at least to some extent, the technical problems in the related art.

[0007] To achieve the above objectives, this application proposes a method for intelligent identification and safety intervention of violations by underground workers in non-coal mines, including the following steps:

[0008] S1. Collect multi-source sensing data by deploying various sensors in the underground working area of ​​non-coal mines. The multi-source sensing data includes video surveillance data, UWB personnel positioning data, environmental monitoring data, and equipment operating status data.

[0009] S2. Preprocess the collected multi-source sensing data, including image enhancement, noise filtering and data alignment, and perform spatiotemporal synchronous fusion of the preprocessed multi-source sensing data to generate fused feature data.

[0010] S3. Input the fused feature data into the pre-trained violation recognition model to identify whether the downhole workers have violated regulations, and output the violation type and violation confidence level. The violation type includes at least area intrusion, leaving the post, not wearing safety equipment, violating regulations in dangerous areas, and violating regulations in equipment operation.

[0011] S4. Calculate the safety risk level based on the identified violation type, violation confidence level, and violation duration, combined with the current operating environment parameters and equipment operating status.

[0012] S5. Based on the safety risk level, implement safety intervention measures that match the safety risk level, including at least one of early warning intervention, equipment linkage control intervention, and personnel dispatch intervention;

[0013] S6. Collect feedback data on worker behavior and on-site status after the implementation of safety intervention measures, evaluate the intervention effect, and perform incremental learning and dynamic updates on the violation identification model based on the evaluation results.

[0014] In addition, the application may also include the following additional technical features:

[0015] Specifically, in step S1, the video monitoring data is collected by infrared cameras and visible light cameras installed in key areas underground; the UWB personnel positioning data is collected by the UWB positioning identification cards worn by underground personnel in conjunction with the intrinsically safe card reader substations deployed underground; the environmental monitoring data includes at least gas concentration, oxygen concentration, temperature and humidity data; and the equipment operating status data includes at least equipment start / stop status, operating parameters and fault codes.

[0016] Specifically, the data preprocessing and fusion in step S2 includes:

[0017] Image enhancement processing is performed on video surveillance data, including low-light enhancement and dust interference filtering;

[0018] Kalman filtering is applied to UWB personnel positioning data to smooth the data and eliminate positioning drift.

[0019] The processed video surveillance data is spatiotemporally synchronized with UWB personnel positioning data to establish a one-to-one correspondence between personnel identities and detected targets in video images.

[0020] Environmental monitoring data and equipment operation status data are aligned by timestamp, and together with video data and location data, they form multimodal fusion feature data.

[0021] Specifically, the training method for the violation identification model in step S3 includes:

[0022] Collect historical video data and corresponding violation annotation data of underground operation scenarios in non-coal mines, and construct a violation training dataset;

[0023] A multi-target tracking algorithm is used to extract individual trajectory data from the training data, and a pose estimation algorithm is used to extract the human body key point information of the workers to construct a human skeleton sequence.

[0024] The individual trajectory data and human skeleton sequence are input into a spatiotemporal graph convolutional network and decomposed into a sequence of atomic behaviors with spatiotemporal labels;

[0025] Long Short-Term Memory (LSTM) networks enhanced with attention mechanisms capture the spatiotemporal correlations of behavioral chains and generate behavioral feature vectors.

[0026] Based on the behavioral feature vector and the preset violation judgment rules, the violation recognition result is output.

[0027] Specifically, calculating the security risk level in step S4 includes:

[0028] The basic risk coefficient is determined based on the type of violation, and then the basic risk coefficient is weighted and adjusted based on the confidence level of the violation.

[0029] The cumulative risk increment is calculated based on the duration of the violation, and then the environmental risk correction factor is calculated based on the current operating environment parameters and equipment operating status.

[0030] The comprehensive safety risk value is calculated by combining the basic risk coefficient, weighted correction value, cumulative risk increment, and environmental risk correction factor.

[0031] The comprehensive safety risk value is compared with multiple preset risk level thresholds to determine the safety risk level; the safety risk level includes at least four levels: low risk, medium risk, high risk, and emergency risk.

[0032] Specifically, the equipment linkage control intervention in step S5 includes: sending control commands to relevant production equipment according to the safety risk level and the type of violation, and executing equipment deceleration, equipment shutdown, equipment lockout, or remote emergency stop operations.

[0033] Specifically, the personnel dispatch intervention in step S5 includes: based on the identity and location information of the personnel involved in the violation, sending dispatch instructions to the security personnel or on-duty personnel closest to the location where the violation occurred, and assigning them to go to the scene to handle the situation.

[0034] Specifically, the intervention effect feedback and model update in step S6 include:

[0035] Collect data on worker behavior within a preset time period after safety intervention measures are implemented;

[0036] Determine whether the violation has been corrected. If it has been corrected, record it as a positive sample; if it has not been corrected, record it as a negative sample.

[0037] Add the positive and negative samples to the model training dataset;

[0038] Periodically or when the accumulated sample size reaches a preset threshold, the violation identification model is incrementally trained to update the model parameters.

[0039] In summary, the beneficial effects of the proposed method for intelligent identification and safety intervention of violations by underground workers in non-coal mines are as follows:

[0040] 1. By integrating four types of data—infrared and visible light video, UWB personnel positioning, underground environmental parameters, and equipment operating status—it effectively solves the problems of blurred images, missed targets, and false detections caused by underground lighting fluctuations, dust obstruction, and blind spots. It can accurately capture subtle violations. At the same time, by combining the superimposed risks of environmental hazards and equipment anomalies, it constructs an integrated full-domain perception system for personnel, environment, and equipment, which greatly improves the completeness and reliability of monitoring complex underground scenarios.

[0041] 2. The model is trained based on real underground operation datasets. It combines target tracking, attitude estimation, spatiotemporal graph convolutional networks and attention mechanisms to extract personnel trajectories, human skeletons and spatiotemporal behavioral features. It can accurately identify various types of violations such as area intrusion and dereliction of duty in real time. It can distinguish between temporary actions and continuous violations based on temporal behavioral features, effectively reducing false detection and false detection rates. It has good adaptability to harsh underground environments with low light and high dust, and the recognition stability and accuracy are significantly improved.

[0042] 3. By comprehensively considering multiple factors such as violation type, identification confidence level, violation duration, and environmental and equipment status, safety risks are quantitatively calculated and classified into four levels: low, medium, high, and emergency. Based on the risk level, differentiated intervention measures such as early warning prompts, equipment linkage control, and nearby personnel dispatch are matched to achieve proactive risk control such as equipment deceleration, locking, and emergency stop, and to quickly assign on-site personnel to handle the situation, significantly shortening response time and preventing the risk from escalating. Attached Figure Description

[0043] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0044] Figure 1 This is a flowchart of the method for intelligent identification and safety intervention of violations by underground workers in non-coal mines, as described in this application.

[0045] Figure 2 This is a flowchart illustrating the model training process for the intelligent identification and safety intervention method for violations by underground workers in non-coal mines, as described in this application. Detailed Implementation

[0046] To make the technical means, inventive features, objectives, and effects of this application easier to understand, the application is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0047] The present application will now be described in further detail with reference to the accompanying drawings.

[0048] like Figure 1 As shown in the embodiment of this application, the method for intelligent identification and safety intervention of violations by underground workers in non-coal mines includes the following steps:

[0049] S1. Collect multi-source sensing data through various sensors deployed in the underground working area of ​​non-coal mines. The multi-source sensing data includes video surveillance data, UWB personnel positioning data, environmental monitoring data, and equipment operation status data.

[0050] It should be noted that the UWB (Ultra-Wideband) personnel positioning system achieves centimeter-level high-precision three-dimensional coordinate positioning of underground personnel by deploying a base station matrix and utilizing the time difference of arrival, and supports real-time determination of electronic fences and geofences; environmental monitoring data integrates wind speed sensors, toxic and harmful gas sensors, and dust concentration monitors to capture changes in the micro-environment of the work in real time; equipment operation status data is collected by vibration, temperature, and current transformers installed on core equipment such as hoists, local fans, and rock drilling rigs, and transmitted back with low latency through industrial Ethernet or 5G private networks to ensure the integrity, synchronization, and authenticity of the original data.

[0051] S2. Preprocess the collected multi-source sensing data, including image enhancement, noise filtering and data alignment, and then perform spatiotemporal synchronous fusion of the preprocessed multi-source sensing data to generate fused feature data.

[0052] It should be noted that, in terms of image enhancement, an adaptive histogram equalization algorithm based on Retinex theory or a lightweight deep learning dehazing network is used to effectively restore visual details in high water fog / low brightness scenes; noise filtering targets the positioning and sensing time-series data, using Kalman filtering or wavelet thresholding to remove pulse interference caused by the start-up and shutdown of underground equipment. The core lies in the spatiotemporal synchronization fusion mechanism: in the time dimension, data from each sensor is unified to millisecond-level timestamps based on PTP (Precise Network Time Protocol); in the spatial dimension, the UWS positioning coordinates are mapped to the field of view of the monitoring camera through coordinate transformation, establishing a mapping matrix of pixel coordinates-underground physical coordinates, and concatenating and aligning video target detection features, position trajectory features, and sensing physical quantity features at the feature layer to generate a structured, highly representative fusion feature tensor, providing high-quality input for subsequent model inference.

[0053] S3. Input the fused feature data into the pre-trained violation recognition model to identify whether there are violations by downhole workers, and output the violation type and violation confidence level. The violation types include at least area intrusion, leaving the post, not wearing safety equipment, violation of operating in dangerous areas, and violation of equipment operation.

[0054] It should be noted that the model is based on fused feature data, performing pose estimation and target detection in real time within the video stream, and combining location coordinates with electronic fences for logical cross-validation. Specific identification logic includes: determining area intrusion when personnel coordinates exceed the designated work area and stay beyond a threshold; determining absenteeism when the location tag remains stationary outside the rest area for an extended period and there are no moving targets in the video stream; and using target detection algorithms to check the wearing of safety helmets, reflective vests, and miner's lamps to determine if safety equipment is not being worn.

[0055] The model determines violations in hazardous areas by combining sudden changes in equipment operating current with the coordinate distance of personnel from hazardous sources (such as manhole openings); it also determines equipment violations by comparing the equipment operation timing logic with standard operating procedures. In addition to outputting the specific violation type, the model also generates a probability confidence level between 0 and 1 to characterize the reliability of the current judgment result, effectively filtering false alarms caused by occlusion or data jitter.

[0056] S4. Calculate the safety risk level based on the identified violation type, violation confidence level, and violation duration, combined with the current operating environment parameters and equipment operating status.

[0057] It should be noted that the identified violation types are assigned basic hazard weights (e.g., equipment operation violations have a higher weight than general area intrusions), and then the weights are discounted or strengthened using confidence coefficients. A time accumulation factor is introduced, and as the duration of the violation increases exponentially (e.g., using an inverse half-life function), the risk value rises non-linearly and rapidly, reflecting the safety logic that prolonged problems lead to disasters. At the same time, real-time environmental parameters (e.g., whether the current gas concentration is close to the critical value, whether the wind speed is abnormal) are used as external correction coefficients. If the environment is already severe, the risk level of the same level of violation is automatically increased. Equipment operating status is also used as a working condition weighting item. Finally, through fuzzy comprehensive evaluation or weighted summation normalization, a four-level safety risk level is output, providing a precise quantitative basis for intervention strategies.

[0058] S5. Based on the safety risk level, implement safety intervention measures that match the safety risk level. Safety intervention measures include at least one of the following: early warning intervention, equipment linkage control intervention, and personnel dispatch intervention.

[0059] It should be noted that targeted rectification instructions, including screenshots of violations and location coordinates, are sent to violators and nearby shift leaders via underground sound and light alarms, intelligent mine lamp voice broadcasts, and handheld terminal vibration message pushes; equipment linkage control intervention: control instructions are issued, such as remotely locking the illegally operated mining equipment, automatically increasing the local fan speed in dangerous areas, or triggering the emergency stop protection of the conveyor belt, to achieve technical error prevention; personnel dispatch intervention: linkage with the wellhead safety monitoring center generates dispatch work orders and searches for the nearest safety inspectors or emergency rescue teams through the UWB positioning system, forcibly dispatching them to the site for physical restraint and evacuation command. All intervention instructions are accompanied by timestamps and electronic signatures, forming a closed-loop handling work order.

[0060] S6. Collect feedback data on worker behavior and on-site status after the implementation of safety intervention measures, evaluate the intervention effect, and perform incremental learning and dynamic updates on the violation identification model based on the evaluation results.

[0061] It should be noted that when constructing intervention effectiveness evaluation indicators, if the evaluation results show that the model has missed or false alarms (e.g., misjudging normal operations as violations), the system automatically extracts the relevant multi-source data and sends it to the edge or cloud for automatic pseudo-labeling. Subsequently, using an incremental learning algorithm combined with a memory replay mechanism, new samples are integrated into the training set to fine-tune and dynamically update the parameters of the violation behavior recognition model without forgetting old violation patterns. This process forms a continuous evolutionary closed loop of perception-recognition-intervention-feedback-optimization, ensuring that the model can adapt to long-term data distribution drift caused by seasonal changes in the downhole working environment, the introduction of new equipment, and changes in personnel attire.

[0062] In one embodiment of this application, the video monitoring data in step S1 is collected by infrared cameras and visible light cameras installed in key areas underground; the UWB personnel positioning data is collected by the UWB positioning identification cards worn by underground personnel in conjunction with the intrinsically safe card reader substations deployed underground; the environmental monitoring data includes at least gas concentration, oxygen concentration, temperature and humidity data; and the equipment operation status data includes at least equipment start / stop status, operating parameters and fault codes.

[0063] It should be noted that infrared cameras can effectively cope with insufficient lighting or completely dark scenes underground, ensuring clear imaging of heat source targets; visible light cameras are used to provide high-resolution color and texture information, facilitating subsequent identification of personnel and details of their movements. The UWB personnel positioning system uses ultra-wideband pulse signal technology, consisting of positioning identification cards worn by underground personnel (usually integrated into miners' lamps or safety helmets) and pre-deployed intrinsically safe card reader substations forming a ranging network. It achieves real-time positioning with centimeter-level accuracy through time difference of arrival or two-way ranging algorithms.

[0064] Environmental monitoring data relies on a multi-parameter sensor array, covering at least four basic indicators: methane concentration, oxygen volume fraction, dry-bulb temperature, and relative humidity. Parameters such as wind speed and dust concentration can be added as needed. Equipment operating status data is collected via PLC or fieldbus, including the equipment's start / stop logic status, operating parameters, and fault codes. The collection frequency of these data types can be customized according to actual needs.

[0065] In one embodiment of this application, the data preprocessing and fusion in step S2 includes:

[0066] Image enhancement processing is performed on video surveillance data, including low-light enhancement and dust interference filtering;

[0067] Kalman filtering is applied to UWB personnel positioning data to smooth the data and eliminate positioning drift.

[0068] The processed video surveillance data is spatiotemporally synchronized with UWB personnel positioning data to establish a one-to-one correspondence between personnel identities and detected targets in video images.

[0069] Environmental monitoring data and equipment operation status data are aligned by timestamp, and together with video data and location data, they form multimodal fusion feature data.

[0070] It should be noted that a two-stage image enhancement pipeline was designed for video surveillance data: First, an adaptive low-light enhancement algorithm based on Retinex theory was adopted to decompose the luminance channel into illuminance and reflectance components. The illuminance component underwent nonlinear stretching while suppressing noise amplification in the reflectance component. Second, to address the light scattering effect caused by dust particles in the well, a dehazing strategy combining dark channel priors and guided filtering was introduced to effectively restore target edges and contrast. For UWB positioning data, since non-line-of-sight propagation and multipath effects often lead to location point jumps, a Kalman filter with a constant acceleration motion model was used for state estimation, smoothing the position sequence, removing outlier drift points, and maintaining the tracking response speed for the actual movement trajectory of personnel.

[0071] After single-modal preprocessing, spatiotemporal synchronous fusion is performed: temporally, using a unified system timestamp as a benchmark, linear interpolation is used to align the data of each modality to the same sampling time; spatially, through pre-calibrated camera intrinsic and extrinsic parameters and the UWB positioning base station coordinate system, the UWB three-dimensional coordinates are projected onto the image pixel plane, thereby establishing a correspondence between personnel ID, pixel bounding box, and skeleton key points. Finally, environmental monitoring values ​​and equipment status codes are stitched together with the same timestamp into the above image-positioning joint features, forming multimodal fusion feature data covering four dimensions: visual, spatial, environmental, and equipment.

[0072] In one embodiment of this application, such as Figure 2 As shown, the training method for the violation recognition model in step S3 includes:

[0073] Collect historical video data and corresponding violation annotation data of underground operation scenarios in non-coal mines, and construct a violation training dataset;

[0074] A multi-target tracking algorithm is used to extract individual trajectory data from the training data, and a pose estimation algorithm is used to extract the human body key point information of the workers to construct a human skeleton sequence.

[0075] Individual trajectory data and human skeleton sequences are input into a spatiotemporal graph convolutional network and decomposed into spatiotemporally labeled atomic behavior sequences;

[0076] Long Short-Term Memory (LSTM) networks enhanced with attention mechanisms capture the spatiotemporal correlations of behavioral chains and generate behavioral feature vectors.

[0077] Based on the behavioral feature vector and the preset rules for judging violations, the results of violation recognition are output.

[0078] It should be noted that a large amount of historical monitoring video was collected from actual production scenarios in non-coal mines, covering the operation process under different lighting, dust, and occlusion levels. Safety experts labeled the types of violations and their start and end times frame by frame or segment by segment to construct a training dataset with sufficient scale and controllable annotation quality. Secondly, a multi-object tracking algorithm was used to extract the continuous trajectory coordinate sequence of each worker from the training video; at the same time, a high-resolution pose estimation network was used to extract the two-dimensional / three-dimensional coordinates of 17 or 25 key joints of the human body to construct a temporal skeleton sequence.

[0079] Subsequently, trajectory displacement information, along with kinematic features such as skeleton angles and velocities, are input into a spatiotemporal graph convolutional network (LSTM). This network uses the natural skeletal connections of the human body as a graph structure, performing convolution operations in both the temporal and spatial dimensions to decompose continuous actions into atomic behavioral units with clear spatiotemporal labels. To capture the logical connections and long-range dependencies between multiple atomic behaviors, an attention-enhanced LSTM network is introduced. Self-attention weights are used to calculate the correlation of behavioral features at different times, outputting a fixed-dimensional behavioral feature vector. During training, the model not only relies on data-driven classification loss but also incorporates pre-defined violation judgment rules for posterior probability correction, ensuring that the recognition results conform to both the data distribution and the semantic constraints of safety procedures, thereby significantly reducing the false positive rate.

[0080] In one embodiment of this application, calculating the security risk level in step S4 includes:

[0081] The basic risk coefficient is determined based on the type of violation, and then the basic risk coefficient is weighted and adjusted based on the confidence level of the violation.

[0082] The cumulative risk increment is calculated based on the duration of the violation, and then the environmental risk correction factor is calculated based on the current operating environment parameters and equipment operating status.

[0083] The comprehensive safety risk value is calculated by combining the basic risk coefficient, weighted correction value, cumulative risk increment, and environmental risk correction factor.

[0084] The overall safety risk value is compared with multiple preset risk level thresholds to determine the safety risk level; the safety risk level includes at least four levels: low risk, medium risk, high risk, and emergency risk.

[0085] It should be noted that, firstly, the basic risk coefficient is retrieved from the preset risk knowledge base based on the type of violation. Then, the confidence score output by the violation identification model is used to multiply or weight the coefficient using a S-shaped function. The higher the confidence score, the closer the correction value is to the basic coefficient; conversely, it is appropriately reduced to reflect the uncertainty of the identification result.

[0086] Based on the duration of the violation, additional risk points are assigned according to a non-linear growth function. Short-term accidental triggering does not significantly increase the risk, while prolonged continuous violations accumulate rapidly, reflecting the objective law that the longer the duration, the higher the probability of an accident. Environmental and equipment conditions are introduced as correction factors: for example, when the gas concentration exceeds the limit, the environmental risk correction factor is >1.0, and the same upward adjustment is made when the oxygen concentration is insufficient; if the equipment is in a faulty state or abnormally started or stopped, the correction coefficient is also increased.

[0087] The final comprehensive risk value is obtained by summing the basic risk coefficient, the weighted correction value, and the cumulative risk increment, and then multiplying it by the environmental risk correction factor. This comprehensive value is compared with four preset threshold ranges: below threshold T1 is low risk, T1-T2 is medium risk, T2-T3 is high risk, and above T3 is emergency risk. Each threshold can be dynamically configured according to the actual safety standards of the mine to ensure the rationality of the risk classification and its adaptability to the site.

[0088] In one embodiment of this application, the equipment linkage control intervention in step S5 includes: sending control commands to relevant production equipment according to the safety risk level and the type of violation, and executing equipment deceleration, equipment shutdown, equipment lockout, or remote emergency stop operations.

[0089] It should be noted that for low-risk or initial violations, the system will implement equipment deceleration or issue audible and visual alarms; for medium-risk situations, the system will implement standby shutdown, allowing the equipment to stop smoothly after completing the current cycle; for high-risk situations, the system will implement equipment interlocking, prohibiting local or remote restart; and for emergency risks, the system will directly trigger remote emergency stop of the equipment to cut off the power source and dangerous energy as quickly as possible to ensure personnel safety. Simultaneously, all control commands are accompanied by timestamps and operation logs for easy post-event traceability. After intervention is executed, the system continuously monitors equipment feedback signals. If the equipment does not respond to the command, it will automatically escalate to a higher level of emergency measures and notify the ground dispatch center.

[0090] In one embodiment of this application, the personnel dispatch intervention in step S5 includes: based on the identity information and location information of the personnel involved in the violation, sending a dispatch instruction to the security personnel or on-duty personnel closest to the location where the violation occurred, and assigning them to go to the scene to handle the situation.

[0091] It should be noted that when a violation is confirmed and involves a specific person, the system first obtains the violator's ID and real-time location through the aforementioned UWB positioning and video correlation results. Simultaneously, it queries the personnel scheduling database for the real-time locations and contact information of currently on-duty security personnel, team leaders, or duty dispatchers. Subsequently, based on a path planning algorithm, the system calculates the closest suitable personnel to the location of the violation and pushes structured dispatch instructions to them via mobile terminals (explosion-proof mobile phones or smart mining lamps).

[0092] The content includes: the name of the person committing the violation, the type of violation, precise location coordinates (and nearby landmarks), a snapshot or short video clip of the violation, and suggested actions (such as warning, stopping, removing, or resetting). The dispatch instruction includes a confirmation button and an estimated arrival time input box. If the person does not confirm within the set time, the system automatically forwards the instruction to the nearest secondary personnel and simultaneously notifies the ground dispatch center for manual intervention. The entire dispatch process generates a complete record of the handling chain for use in safety assessments and process optimization.

[0093] In one embodiment of this application, step S6, which involves feedback on the intervention effect and model updating, includes:

[0094] Collect data on worker behavior within a preset time period after safety intervention measures are implemented;

[0095] Determine whether the violation has been corrected. If it has been corrected, record it as a positive sample; if it has not been corrected, record it as a negative sample.

[0096] Add positive and negative samples to the model training dataset;

[0097] Periodically or when the accumulated sample size reaches a preset threshold, the violation identification model is incrementally trained and the model parameters are updated.

[0098] It should be noted that after each safety intervention is implemented, the system automatically starts an effectiveness evaluation timer, continuously collecting video and location data of the same area, the same person, or similar operational behaviors within a preset time window (e.g., 5 minutes, 15 minutes, or 30 minutes, depending on the type of violation). By comparing the behavioral characteristic sequences before and after the intervention, it determines whether the original violation has been corrected: if the target person resumes compliant operation or leaves the danger zone, and the same type of violation does not recur, the intervention event, along with the identification input data before the intervention, is marked as a positive sample; if the violation persists after the time window ends, or recurs repeatedly within a short period, it is marked as a negative sample.

[0099] All positive and negative samples undergo quality screening (removing samples with missing data or ambiguous labels) before being added to the model's incremental training pool. Model updates employ two trigger strategies: one is periodic updates (e.g., weekly), and the other is immediate incremental training initiated when the accumulated sample size reaches a preset threshold (e.g., adding 1000 positive samples and 500 negative samples). During incremental training, the weights are fine-tuned based on the original model, and an experience replay mechanism is used to mix in some old samples to prevent catastrophic forgetting. The updated model is evaluated on an offline validation set. If its performance metrics are not lower than the original model, it is deployed online; otherwise, the original model is retained with an alert, and the quality of the newly added samples is manually reviewed to ensure continuous optimization and stable reliability of the model.

[0100] The present application and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present application. The actual structure is not limited to this. In conclusion, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the protection scope of the present application.

Claims

1. A method for intelligent identification and safety intervention of violations by underground workers in non-coal mines, characterized in that, Includes the following steps: S1. Collect multi-source sensing data by deploying various sensors in the underground working area of ​​non-coal mines. The multi-source sensing data includes video surveillance data, UWB personnel positioning data, environmental monitoring data, and equipment operating status data. S2. Preprocess the collected multi-source sensing data, including image enhancement, noise filtering and data alignment, and perform spatiotemporal synchronous fusion of the preprocessed multi-source sensing data to generate fused feature data. S3. Input the fused feature data into the pre-trained violation recognition model to identify whether the downhole workers have violated regulations, and output the violation type and violation confidence level. The violation type includes at least area intrusion, leaving the post, not wearing safety equipment, violating regulations in dangerous areas, and violating regulations in equipment operation. S4. Calculate the safety risk level based on the identified violation type, violation confidence level, and violation duration, combined with the current operating environment parameters and equipment operating status. S5. Based on the safety risk level, implement safety intervention measures that match the safety risk level, including at least one of early warning intervention, equipment linkage control intervention, and personnel dispatch intervention; S6. Collect feedback data on worker behavior and on-site status after the implementation of safety intervention measures, evaluate the intervention effect, and perform incremental learning and dynamic updates on the violation identification model based on the evaluation results.

2. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, In step S1, the video monitoring data is collected by infrared cameras and visible light cameras installed in key areas underground; the UWB personnel positioning data is collected by the UWB positioning identification cards worn by underground personnel in conjunction with the intrinsically safe card reader substations deployed underground; the environmental monitoring data includes at least gas concentration, oxygen concentration, temperature and humidity data; and the equipment operation status data includes at least equipment start / stop status, operating parameters and fault codes.

3. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, The data preprocessing and fusion in step S2 includes: Image enhancement processing is performed on video surveillance data, including low-light enhancement and dust interference filtering; Kalman filtering is applied to UWB personnel positioning data to smooth the data and eliminate positioning drift. The processed video surveillance data is spatiotemporally synchronized with UWB personnel positioning data to establish a one-to-one correspondence between personnel identities and detected targets in video images. Environmental monitoring data and equipment operation status data are aligned by timestamp, and together with video data and location data, they form multimodal fusion feature data.

4. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, The training method for the violation detection model in step S3 includes: Collect historical video data and corresponding violation annotation data of underground operation scenarios in non-coal mines, and construct a violation training dataset; A multi-target tracking algorithm is used to extract individual trajectory data from the training data, and a pose estimation algorithm is used to extract the human body key point information of the workers to construct a human skeleton sequence. The individual trajectory data and human skeleton sequence are input into a spatiotemporal graph convolutional network and decomposed into a sequence of atomic behaviors with spatiotemporal labels; Long Short-Term Memory (LSTM) networks enhanced with attention mechanisms capture the spatiotemporal correlations of behavioral chains and generate behavioral feature vectors. Based on the behavioral feature vector and the preset violation judgment rules, the violation recognition result is output.

5. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, The calculation of the security risk level in step S4 includes: The basic risk coefficient is determined based on the type of violation, and then the basic risk coefficient is weighted and adjusted based on the confidence level of the violation. The cumulative risk increment is calculated based on the duration of the violation, and then the environmental risk correction factor is calculated based on the current operating environment parameters and equipment operating status. The comprehensive safety risk value is calculated by combining the basic risk coefficient, weighted correction value, cumulative risk increment, and environmental risk correction factor. The comprehensive safety risk value is compared with multiple preset risk level thresholds to determine the safety risk level; the safety risk level includes at least four levels: low risk, medium risk, high risk, and emergency risk.

6. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, The equipment linkage control intervention in step S5 includes: sending control commands to relevant production equipment according to the safety risk level and the type of violation, and executing equipment deceleration, equipment shutdown, equipment lockout, or remote emergency stop operations.

7. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, The personnel dispatch intervention in step S5 includes: based on the identity and location information of the personnel involved in the violation, sending dispatch instructions to the security personnel or on-duty personnel closest to the location where the violation occurred, and assigning them to go to the scene to handle the situation.

8. The method for intelligent identification and safety intervention of violations by underground workers in non-coal mines according to claim 1, characterized in that, The intervention effect feedback and model update in step S6 include: Collect data on worker behavior within a preset time period after safety intervention measures are implemented; Determine whether the violation has been corrected. If it has been corrected, record it as a positive sample; if it has not been corrected, record it as a negative sample. Add the positive and negative samples to the model training dataset; Periodically or when the accumulated sample size reaches a preset threshold, the violation identification model is incrementally trained to update the model parameters.

Images