Mine personnel trajectory prediction method, system and equipment based on artificial intelligence

Abstract

This application relates to the field of mine safety analysis technology and discloses an artificial intelligence-based method, system, and device for predicting the trajectory of mine personnel. The method includes: collecting and standardizing mine personnel location data; calculating trajectory quality using a neural network; converting spatiotemporal feature vectors for image recognition; applying activation functions for multi-time-scale prediction; matching the spatial structure of the mine to cluster and identify group behavior patterns; calculating safety indicators, and generating a safety scheduling strategy. This application accurately predicts the future trajectory of mine personnel based on the specific environmental constraints of the mine and the working behavior patterns of miners, and converts the prediction results into a safety scheduling strategy.

Description

Technical Field

[0001] The present application relates to the technical field of mine safety analysis, and in particular to a method, system and device for predicting mine personnel trajectories based on artificial intelligence. Background Art

[0002] Mine operating environments are complex and potentially dangerous, making personnel safety management a crucial task for mining companies. Traditional mine personnel management methods rely primarily on manual check-ins, telephone reporting, and simple positioning devices. These only provide real-time location information for miners, but are unable to predict their future movements. With the advancement of IoT and positioning technologies, mine personnel positioning systems have gained widespread adoption. These systems deploy sensor networks at key locations within the mine, receiving signals from smart positioning devices worn by miners, enabling real-time monitoring of their locations. These systems typically utilize technologies such as RFID, Bluetooth, and UWB to record miners' location coordinates and corresponding timestamps, generating historical trajectory data. Furthermore, environmental monitoring systems can collect environmental parameters such as temperature, humidity, gas concentrations, and ventilation conditions across various areas of the mine, providing essential data support for safety management.

[0003] However, existing technologies have significant shortcomings. First, these systems can only provide historical and real-time location information for miners and lack the ability to predict future trajectories, making them unable to support forward-looking safety management and resource allocation. Second, some existing trajectory prediction methods are mostly designed for open environments such as urban roads, without considering the physical constraints and safety requirements of the special underground environment of mines. Direct application will result in prediction results that are inconsistent with reality. Third, existing methods often ignore the quality heterogeneity of miner trajectory data. The quality of data collected at different times and in different regions varies significantly. Indiscriminate use will reduce prediction accuracy. Fourth, existing methods usually treat miners as independent individuals for prediction, ignoring the group coordination relationship in mine operations, resulting in prediction results that are inconsistent with actual operating rules. Finally, there is a lack of a mechanism to directly link prediction results with safety management decisions, making it difficult to translate prediction results into specific safety scheduling measures. Summary of the Invention

[0004] This application provides an artificial intelligence-based mine personnel trajectory prediction method, system and equipment, which are used to accurately predict the future trajectories of mine personnel through the special environmental constraints of the mine and the work behavior patterns of miners, and convert the prediction results into a safe scheduling strategy.

[0005] In the first aspect, the present application provides an artificial intelligence-based mine personnel trajectory prediction method, which includes: collecting mine personnel positioning data and environmental parameter data, standardizing the mine personnel positioning data, and obtaining structured trajectory data; performing quality assessment on the structured trajectory data, and calculating the trajectory quality by constructing an artificial neural network with two hidden layers to obtain trajectory data; converting the trajectory data and the environmental parameter data into spatiotemporal feature vectors, performing image pattern recognition processing through the artificial neural network to obtain a trajectory vector; applying a rectified linear unit activation function and a Sigmoid function to the trajectory vector for multi-time scale processing, calculating the gradient through a quality-weighted loss function, and obtaining trajectory prediction data; constraining the trajectory prediction data with the mine spatial structure, identifying the miner group behavior pattern through a density clustering algorithm based on computer vision, and obtaining target trajectory prediction data; calculating the spatial safety distance, time conflict probability, and regional load index based on the target trajectory prediction data to generate a mine personnel safety scheduling strategy.

[0006] In a second aspect, the present application provides an artificial intelligence-based mine personnel trajectory prediction system, the artificial intelligence-based mine personnel trajectory prediction system comprising:

[0007] An acquisition module is used to collect mine personnel positioning data and environmental parameter data, and perform standardization processing on the mine personnel positioning data to obtain structured trajectory data;

[0008] a calculation module, configured to perform quality assessment on the structured trajectory data, and calculate the trajectory quality by constructing an artificial neural network with two hidden layers to obtain trajectory data;

[0009] a conversion module, configured to convert the trajectory data and the environmental parameter data into a spatiotemporal feature vector, and perform image pattern recognition processing through the artificial neural network to obtain a trajectory vector;

[0010] a processing module, configured to apply a rectified linear unit activation function and a sigmoid function to the trajectory vector for multi-time scale processing, calculate a gradient using a loss function weighted by a quality weight, and obtain trajectory prediction data;

[0011] A matching module is used to perform constraint matching between the trajectory prediction data and the spatial structure of the mine, identify the behavior pattern of the miners' group through a density clustering algorithm based on computer vision, and obtain the target trajectory prediction data;

[0012] A generation module is used to calculate the spatial safety distance, time conflict probability and regional load index based on the target trajectory prediction data, and generate a mine personnel safety scheduling strategy.

[0013] In a third aspect, an artificial intelligence-based mine personnel trajectory prediction device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the artificial intelligence-based mine personnel trajectory prediction device executes the above-mentioned artificial intelligence-based mine personnel trajectory prediction method.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned artificial intelligence-based mine personnel trajectory prediction method.

[0015] The technical solution provided in this application achieves a structured representation of miner trajectories by collecting and standardizing mine personnel location data and environmental parameter data. This structured trajectory data is then quality-assessed and trajectory quality is calculated using an artificial neural network with two hidden layers. This effectively addresses the problem of data quality heterogeneity in mine environments and improves the reliability of prediction model training data. Trajectory and environmental parameter data are converted into spatiotemporal feature vectors and processed using an artificial neural network for image pattern recognition, leveraging the advantages of deep learning in spatial pattern recognition to effectively extract key features of miner trajectories. Trajectory vectors are processed using a rectified linear unit activation function and a sigmoid function for multi-time scale processing, enabling trajectory prediction for three different time spans: short-term, medium-term, and long-term, meeting the needs of various application scenarios. The design of a quality-weighted loss function for calculating the gradient effectively suppresses the interference of low-quality data on model training. The trajectory prediction data is constrained and matched to the mine's spatial structure, and a density clustering algorithm based on computer vision is used to identify the behavior patterns of miner groups. This ensures that the prediction results conform to both the physical constraints of the mine and the actual working patterns of miners, significantly improving prediction accuracy. Finally, based on the target trajectory prediction data, the spatial safety distance, temporal conflict probability, and regional load index are calculated, and a mine personnel safety scheduling strategy is generated, realizing the direct transformation of the prediction results into safety management decisions. In the specific application field of mines, this solution fully considers the contribution of artificial intelligence algorithm features to the solution, especially the multi-level feature extraction capabilities of artificial neural networks, the nonlinear mapping capabilities of rectified linear units and sigmoid activation functions, the pattern recognition capabilities of computer vision, and the group behavior modeling capabilities of density clustering. Compared with traditional methods, the present invention can not only accurately predict the future trajectory of miners, but also identify potential safety risks and automatically generate scheduling strategies, significantly improving the accuracy of mine safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 This is a schematic diagram of an embodiment of a method for predicting mine personnel trajectories based on artificial intelligence in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an embodiment of a mine personnel trajectory prediction system based on artificial intelligence in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of a mine personnel trajectory prediction device based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a method, system and device for predicting the trajectory of personnel in a mine based on artificial intelligence. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of a method for predicting mine personnel trajectories based on artificial intelligence includes:

[0022] Step S101: Collecting mine personnel positioning data and environmental parameter data, and performing standardization processing on the mine personnel positioning data to obtain structured trajectory data;

[0023] Step S102: performing quality assessment on the structured trajectory data by constructing an artificial neural network with two hidden layers to calculate the trajectory quality and obtain trajectory data;

[0024] Step S103: converting the trajectory data and environmental parameter data into a spatiotemporal feature vector, performing image pattern recognition processing through an artificial neural network, and obtaining a trajectory vector;

[0025] Step S104: Apply the rectified linear unit activation function and the Sigmoid function to the trajectory vector for multi-time scale processing, calculate the gradient through the loss function weighted by the quality weight, and obtain the trajectory prediction data;

[0026] Step S105: performing constraint matching on the trajectory prediction data and the spatial structure of the mine, identifying the behavior pattern of the miner group through a density clustering algorithm based on computer vision, and obtaining target trajectory prediction data;

[0027] Step S106: Calculate the spatial safety distance, time conflict probability, and regional load index based on the target trajectory prediction data to generate a mine personnel safety scheduling strategy.

[0028] It is understandable that the execution subject of this application can be an artificial intelligence-based mine personnel trajectory prediction system, or a terminal or server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0029] Specifically, this is accomplished through a network of sensors deployed at pre-determined locations within the mine. These sensors receive location coordinate data and timestamps uploaded by smart positioning devices worn by miners. Environmental parameter data, including temperature, humidity, gas concentration, and ventilation conditions, is also collected. The collected raw trajectory data is cleaned, primarily targeting three types of outliers: coordinate points outside the mine's physical boundaries, points with unusual timestamps, and points with coordinate jumps. Missing points are often found in the cleaned trajectory data. These points are then supplemented using a spatiotemporal interpolation algorithm. This algorithm considers the continuity of miners' movements and calculates reasonable coordinates for missing points based on the spatiotemporal relationships between previous and subsequent points. The complete trajectory sequence is segmented into multiple segments according to pre-determined time windows. Spatial and temporal features are extracted from these segments to generate feature-rich trajectory data. This data is then associated with the functional attributes of the mine area, and the trajectory points are semantically annotated to generate structured trajectory data.

[0030] To assess the quality of structured trajectory data, initial quality parameters are first obtained through multi-dimensional scoring, including completeness, accuracy, regularity, and semantic consistency. These initial quality parameters are weighted and combined using a trajectory quality scoring function to generate a normalized quality score. Next, an artificial neural network is constructed with two hidden layers: an input layer with two nodes receiving the basic features of the trajectory data, two hidden layers with eight nodes each processing feature transformation, and an output layer with one node outputting the quality score. The structured trajectory data is converted into a two-dimensional spatiotemporal image and then fed into the neural network. Local features are extracted through convolution, followed by global features extracted through pooling and fully connected operations to generate a trajectory quality score. To optimize the scoring accuracy, a trajectory contrast learning task is constructed. Trajectories of the same miner with similar time periods are assigned similar scores, while trajectories with significant differences in regularity and periodicity are assigned different scores. The network parameters are optimized by minimizing the contrastive loss function to obtain a calibrated quality score. Based on this score, quality weights are assigned to the structured trajectory data to generate trajectory data.

[0031] To convert trajectory data and environmental parameter data into spatiotemporal feature vectors, the trajectory data is converted into a time-series position matrix in the mine's three-dimensional coordinate system, while the environmental parameter data is mapped to attribute values ​​at corresponding locations. A multi-channel mine trajectory activity heat map is generated through spatial rasterization, forming a trajectory image matrix. This matrix is ​​contrast-enhanced to highlight trajectory features in key areas such as tunnels, working faces, and refuges. The enhanced trajectory image is decomposed into four channels: spatial, velocity, acceleration, and environmental parameter. After compression and dimensionality reduction, a spatiotemporal feature vector is generated. This vector is fed into the first hidden layer of an artificial neural network. A 3×3 convolution kernel is used across eight nodes to extract local features, identifying behavioral characteristics such as miners' turning patterns, dwell point distribution, and speed changes, generating a behavioral feature map. This map undergoes a nonlinear transformation using a rectified linear unit activation function and is then fed into the second hidden layer of the neural network. A 5×5 convolution kernel is used across eight nodes to extract contextual information with a larger receptive field, identifying miners' work patterns and path selection patterns in different mine areas, generating contextual feature representation data. Finally, the mine topology constraint is applied to the data, and the three-dimensional information of space, time and environment is integrated through the fully connected layer to generate the trajectory vector.

[0032] When processing trajectory vectors at multiple time scales, the prediction task is divided into three scales: short-term, medium-term, and long-term. Trajectory vectors are partitioned along the time dimension to produce three sets of prediction input data. The short-term prediction input data is processed using a rectified linear unit activation function to capture the local continuity and short-term variation patterns of the trajectory. The medium-term prediction input data models long-range dependencies by calculating the correlation weight matrix for each time point in the trajectory sequence. The long-term prediction input data uses a sigmoid function to process the hidden state output to model the long-term uncertainty of the trajectory. Based on the quality weights of the trajectory data, a weighted loss function is constructed to calculate the prediction error. The backpropagation algorithm is used to calculate the gradient and update the network parameters to optimize the prediction results. Finally, the prediction results from the three time scales are fused to produce the trajectory prediction data.

[0033] When constraining the trajectory prediction data against the mine's spatial structure, the predicted data is spatially compared with the mine's three-dimensional structural map. Predicted points that are physically inaccessible are eliminated, including those located within rock formations, crossing safety isolation zones, and exceeding the working range. This results in spatially constrained trajectory data. The historical trajectories of multiple miners are then grouped by work team, job type, and activity area to form a grouped trajectory set. A computer vision-based density clustering process is applied to this set, clustering temporally and spatially similar trajectory points. The center point and activity range of each group are extracted to generate a behavioral density distribution map. Based on this map, miners' activity patterns in different areas are identified, including work stop patterns, inspection route patterns, material transportation patterns, and emergency evacuation patterns, generating behavioral pattern data. The spatially constrained trajectory data is then matched with the behavioral pattern data. The direction and tempo of the predicted trajectory are adjusted based on the miner's current location and task type to align with the corresponding behavioral pattern characteristics, generating a behaviorally optimized trajectory. Based on the collaborative relationship between miners in the same work group, the behavior optimization trajectory is adjusted for group consistency to ensure the spatiotemporal coordination of the predicted trajectories of the collaborative miners and obtain the target trajectory prediction data.

[0034] To generate a safe scheduling strategy based on target trajectory prediction data, the data is grouped by miner ID and time series. The position coordinate sequence of each miner in the future time period is calculated to form a predicted personnel distribution map. Based on this map, the minimum spatial distance between any two miners is calculated using the Euclidean distance formula. If it is less than a preset safety threshold, it is marked as a potential dangerous contact point, resulting in a spatial safety distance index. A temporal probability analysis is performed on these potential dangerous contact points to calculate the probability of two miners appearing in the same area and time period. The spatial overlap is calculated using a Gaussian distribution model for the miners' location coordinates, resulting in a temporal conflict probability index. The predicted target trajectory data is mapped onto a mine zoning map. The miner density in each area during different time periods is calculated. The load factor is calculated by comparing the regional capacity to the actual number of people, resulting in a regional load index. A multi-objective optimization model is constructed based on these three indicators. These indicators are combined into a safety risk score using a linear weighting method. Each area and time period is then assigned a risk level, generating a risk distribution map. Based on the risk distribution map, a mine personnel safety scheduling strategy is generated, including personnel diversion recommendations, work schedule adjustment plans, and emergency evacuation routes.

[0035] In an embodiment of the present application, by collecting and standardizing mine personnel positioning data and environmental parameter data, a structured representation of miner trajectories is achieved, the quality of the structured trajectory data is evaluated, and an artificial neural network with two hidden layers is applied to calculate the trajectory quality, effectively solving the problem of data quality heterogeneity in the mine environment and improving the reliability of the prediction model training data. The trajectory data and environmental parameter data are converted into spatiotemporal feature vectors, and image pattern recognition processing is performed through an artificial neural network, making full use of the advantages of deep learning in spatial pattern recognition and effectively extracting the key features of the miner's trajectory. The trajectory vector is processed by applying a rectified linear unit activation function and a sigmoid function for multi-time scale processing, achieving trajectory prediction for three different time spans: short-term, medium-term, and long-term, meeting the needs of different application scenarios. The design of the loss function calculation gradient weighted by quality weight effectively suppresses the interference of low-quality data on model training. The trajectory prediction data is constrained and matched with the spatial structure of the mine, and the behavior pattern of the miner group is identified through a density clustering algorithm based on computer vision, so that the prediction results are consistent with both the physical constraints of the mine and the actual working rules of the miners, greatly improving the prediction accuracy. Finally, based on the target trajectory prediction data, the spatial safety distance, temporal conflict probability, and regional load index are calculated, and a mine personnel safety scheduling strategy is generated, realizing the direct transformation of the prediction results into safety management decisions. In the specific application field of mines, this solution fully considers the contribution of artificial intelligence algorithm features to the solution, especially the multi-level feature extraction capabilities of artificial neural networks, the nonlinear mapping capabilities of rectified linear units and sigmoid activation functions, the pattern recognition capabilities of computer vision, and the group behavior modeling capabilities of density clustering. Compared with traditional methods, the present invention can not only accurately predict the future trajectory of miners, but also identify potential safety risks and automatically generate scheduling strategies, significantly improving the accuracy of mine safety management.

[0036] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0037] A sensor network deployed at preset locations in the mine collects location coordinate data and timestamp data uploaded by the smart positioning devices worn by miners to obtain original trajectory data. It also collects temperature, humidity, gas concentration, and ventilation data in the mine area to obtain environmental parameter data.

[0038] The original trajectory data is cleaned to remove coordinate points beyond the physical boundary of the mine, abnormal time stamp points, and coordinate jump points to obtain cleaned trajectory data;

[0039] The missing data points in the cleaned trajectory data are supplemented by a spatiotemporal interpolation algorithm to obtain a trajectory sequence;

[0040] The trajectory sequence is segmented according to the preset time window to obtain multiple trajectory segments;

[0041] Extract spatial and temporal features from multiple trajectory segments, including displacement vector, velocity, acceleration, steering angle, dwell time, and movement frequency, to obtain characterized trajectory data;

[0042] The characterized trajectory data are associated with the functional attributes of the mine area, and the trajectory points are semantically annotated to obtain structured trajectory data.

[0043] Specifically, data is collected through a network of sensors deployed at pre-set locations within the mine. These sensor networks primarily consist of positioning signal receivers located at key locations, such as tunnel intersections, work surfaces, and refuge rooms. These receivers receive signals from smart positioning devices worn by miners. These devices are small, portable devices carried by miners and typically incorporate RFID tags, Bluetooth, or UWB (ultra-wideband) positioning modules, capable of broadcasting the miner's location information in real time. As miners carry these devices within the mine, receivers at pre-set locations capture the signals and record the location coordinates (including X, Y, and Z coordinates) along with the corresponding timestamps, generating raw trajectory data. Simultaneously, environmental monitoring sensors deployed throughout the mine collect data on temperature, humidity, gas concentrations, and ventilation conditions in various areas.

[0044] The cleaning process detects and removes coordinate points that extend beyond the physical boundaries of the mine. These points often result in positioning errors due to signal reflection or interference. Specifically, each coordinate point is compared with the mine's 3D structural map. If a point's coordinates fall within the solid rock layer or outside the mine's mining area, it is considered an outlier and deleted. Secondly, the cleaning process identifies points with unusual timestamps, including those with time reversals (a later point's timestamp is earlier than the previous one) and time jumps (an unusually large time interval between two adjacent points). Finally, the cleaning process detects coordinate jumps, which occur when miners move unreasonably long distances in a very short period of time. These points are often caused by transient interference in the positioning signal or equipment failure. The process calculates the movement speed between adjacent points. If the speed exceeds the normal human walking speed threshold (set at 2-3 meters per second), the point is identified as an outlier and deleted. While the trajectory data obtained after the cleaning process eliminates outliers, it may still contain missing data.

[0045] Missing data points in the cleaned trajectory data are supplemented using a spatiotemporal interpolation algorithm. This algorithm comprehensively considers spatial and temporal continuity, inferring the location of missing points based on known points before and after the missing point. Common interpolation methods include linear interpolation and spline interpolation. In a mine environment, given that miners' movement paths are restricted by the mine structure, spatiotemporal interpolation must also be combined with the physical constraints of the mine passages. Specifically, the temporal location of the missing point is determined, and then the coordinates and timestamps of the known points before and after are calculated to determine the miner's most likely location at that point in time. If the missing time period is long, complex interpolation must be performed based on the miner's historical movement patterns and mine path planning. After interpolation is completed, a continuous and complete trajectory sequence is formed, ensuring the temporal continuity of the trajectory data.

[0046] The completed trajectory sequence is segmented according to preset time windows. The size of the time window is typically determined by the miner's work characteristics and analysis requirements. Common time windows include 10 minutes, 30 minutes, or 1 hour. The segmentation process divides the complete trajectory into fixed-length segments in chronological order. Each segment contains information about all the miner's locations within that time window. For example, if a 30-minute time window is selected, a miner's trajectory over 8 hours of work will be segmented into 16 segments. Window boundaries must be addressed during segmentation. Spatial and temporal features are extracted from the resulting trajectory segments. Spatial features include displacement vectors (the direction and distance between two adjacent points), velocity (displacement divided by the time interval), acceleration (rate of change of velocity), and steering angle (angle of change in direction of movement). Temporal features include dwell time (the duration of near-zero velocity within a specific area) and movement frequency (the number of moves per unit time). The feature extraction process essentially transforms the raw position-time data into a high-level feature set that describes the miner's behavioral patterns. For example, for a trajectory segment, the displacement vector between each two adjacent points is calculated, followed by the average velocity and acceleration per unit time. Time periods with near-zero velocity are detected to identify stops, and the number of direction changes greater than a specific angle is counted to identify turns. These features together constitute the characterized trajectory data.

[0047] The purpose of associating the characterized trajectory data with the functional attributes of mine areas and semantically annotating trajectory points is to enhance the semantic understanding of the trajectory data. Mine area functional attributes refer to the purpose and characteristics of different areas, such as the mining working face, haul tunnels, safe havens, and equipment maintenance areas. The association process first matches the spatial coordinates of each trajectory point with a map of the mine's functional areas to determine the functional area in which the point is located. Semantic labels, such as "working at the working face," "moving along the haul tunnel," and "resting in the safe haven," are then added to the trajectory points based on the functional attributes of the area and the miners' behavioral characteristics in that area. This semantic annotation transforms trajectory data into a simple sequence of locations, instead providing a record of activities with business meaning, thus forming structured trajectory data.

[0048] For example, after a worker wearing a smart positioning device entered the mine, the sensor network recorded his movement trajectory from the locker room to the work face, including location coordinates and corresponding timestamps every 5 seconds. Environmental sensors also recorded environmental parameters along the way. During the data cleaning phase, the system detected some anomalies, such as wall penetrations and speed fluctuations (e.g., moving instantly from the main tunnel to the refuge room), and removed them. After cleaning, it was discovered that the worker had lost approximately one minute of data while passing through an area with weak signal. Using a spatiotemporal interpolation algorithm, the system calculated possible path points during the missing period based on the last location before the missing point and the first location after the missing point, combined with the mine tunnel structure, and supplemented the data. The completed trajectory was segmented into 30-minute segments. Within each segment, features were extracted, such as the worker's average speed in different areas (0.8 m / s in the main tunnel, 0.3 m / s at the work face) and his rest behavior (25 consecutive minutes at the work face). Finally, based on the mine functional area map, semantic annotations are added to the trajectory points to clearly identify workers' movements in the main tunnels, coal mining operations at the working face, short stops in the rest area, and other activities, forming structured trajectory data containing rich semantic information.

[0049] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0050] Perform multi-dimensional scoring on structured trajectory data to obtain initial quality parameters. Then, a preset trajectory quality scoring function is used to weight the initial quality parameters according to the importance of each indicator to obtain a normalized quality score.

[0051] An artificial neural network with two hidden layers is constructed and trained, wherein the input layer has two nodes, the two hidden layers each have eight nodes, and the output layer has one node, thereby obtaining an artificial neural network;

[0052] The structured trajectory data is converted into a two-dimensional spatiotemporal image and input into an artificial neural network. Local features are extracted through convolution operations, and global features are extracted through pooling operations and full connection operations to obtain trajectory quality scores.

[0053] A trajectory contrast learning task is constructed to generate similar scores for trajectories of the same miner with similar time periods, and different scores for trajectories with significant differences in regularity and periodicity. The network parameters are optimized by minimizing the contrastive loss function to obtain a calibration quality score.

[0054] Based on the calibration quality score, quality weights are assigned to the structured trajectory data to obtain trajectory data.

[0055] Specifically, structured trajectory data is scored across multiple dimensions, including completeness, accuracy, regularity, and semantic consistency. The completeness metric measures the temporal coverage of trajectory data, calculated by calculating the proportion of valid data points to total working time. Complete trajectory data should cover the entire work process of a miner without significant missing segments. The accuracy metric measures the positional accuracy of trajectory data, assessed by analyzing the degree of deviation between the positioning point and known reference points in the mine. Accurate trajectory data should accurately reflect the miner's actual location. The regularity metric measures the intensity of periodicity reflected in the trajectory data, assessed by analyzing the repetitive patterns and temporal regularity of the trajectory. Highly regular trajectories generally reflect stable working behavior patterns of miners. The semantic consistency metric measures the degree of match between trajectory data and miners' work tasks, assessed by comparing the degree of conformity between the trajectory activity area and the predetermined work plan. High semantic consistency indicates that the trajectory data accurately reflects the miner's actual work content.

[0056] The measurement results of these four metrics form initial quality parameters, which are then weighted and combined using a pre-defined trajectory quality scoring function. This weighting process considers the importance of each metric. Generally, completeness and accuracy are given higher weights, as they are fundamental quality characteristics. Regularity and semantic consistency are given relatively lower weights, but their weighting may be increased for specific analysis tasks. The weighted combination results are then normalized to map the quality score to a range of 0 to 1, with 1 representing the highest quality and 0 representing the lowest. This normalization process ensures that the quality scores of different trajectories are comparable.

[0057] The artificial neural network architecture consists of an input layer, two hidden layers, and an output layer. The input layer has two nodes, each receiving basic trajectory features such as the number of trajectory points and the duration of coverage. The two hidden layers, each with eight nodes, are responsible for feature transformation and nonlinear mapping. The output layer has one node, outputting the final quality score. The network is trained using supervised learning, using existing high-quality trajectory samples as positive examples and low-quality trajectory samples as negative examples. The network parameters are optimized through a backpropagation algorithm, enabling the network to accurately distinguish between trajectory data of varying quality levels. After training, the artificial neural network can assess the quality of new trajectory data, eliminating the need for manual calculation of various metrics.

[0058] The structured trajectory data is converted into a two-dimensional spatiotemporal image to facilitate neural network processing. This conversion process maps the three-dimensional spatial trajectory and one-dimensional temporal information onto a two-dimensional plane, forming a spatiotemporal image. Specifically, trajectory points are mapped to image pixels, with time as the horizontal axis and spatial position as the vertical axis. Continuous trajectories form lines or curves within the image. Different colors or brightness levels can represent different motion states or environmental parameter values. This image representation allows for simultaneous visualization of spatial and temporal patterns, facilitating feature extraction by convolutional neural networks. The converted two-dimensional spatiotemporal image is input into an artificial neural network. The convolutional layer first extracts local features, such as short-term movement patterns and turning characteristics. The convolution operation slides multiple different filters across the image to identify pattern features such as edges and textures. The pooling layer then reduces the dimensionality of the feature map, preserving key information while reducing computational effort. A fully connected layer extracts global features, integrating relationships between local features to form a comprehensive assessment of overall trajectory quality, outputting a trajectory quality score.

[0059] Constructing a trajectory contrastive learning task is an unsupervised learning method for evaluating trajectory quality in the absence of explicit quality labels. The core idea of ​​contrastive learning is that trajectories from the same miner under similar conditions should have similar quality characteristics, while trajectories from different miners or under significantly different conditions may have different quality characteristics. In its implementation, we first define a trajectory similarity criterion: Trajectory pairs from the same miner and occurring at similar times (e.g., different times of the day) are considered positive pairs and should receive similar quality scores. Trajectory pairs with significant differences in regularity and periodicity (e.g., trajectories from a normal working day versus a day with equipment failure) are considered negative pairs and should receive different quality scores. For each pair of trajectories, feature representations are computed and their differences are quantified using a contrastive loss function. The contrastive loss function is designed to minimize the distance between positive pairs and maximize the distance between negative pairs. By minimizing this loss function, the neural network parameters are continuously optimized, enabling the network to learn a more accurate trajectory quality evaluation criterion and obtain a calibrated quality score. Assigning quality weights to structured trajectory data based on the calibrated quality score is a key step in applying the quality assessment results to real-world data processing. This quality weight assignment uses a nonlinear mapping function to convert the calibrated quality score into a weighting coefficient. A commonly used mapping function is an exponential function, which gives high-quality trajectories exponentially higher weights and significantly reduces the influence of low-quality trajectories. This weighting strategy ensures that subsequent trajectory analysis and prediction are primarily based on high-quality data, mitigating interference from low-quality data. After assigning quality weights, the structured trajectory data forms the trajectory data.

[0060] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0061] The trajectory data is converted into a time-series position matrix in the mine's three-dimensional coordinate system, and the environmental parameter data is mapped into the attribute values ​​of the corresponding position points. A multi-channel mine trajectory activity heat map is generated through spatial rasterization processing to obtain a trajectory image matrix.

[0062] Performing contrast enhancement on the trajectory image matrix to obtain an enhanced trajectory image. The contrast enhancement is used to enhance the trajectory features of the tunnel, working surface and refuge room.

[0063] Decompose the enhanced trajectory image into spatial channel, velocity channel, acceleration channel and environmental parameter channel, perform compression and dimension reduction processing to generate spatiotemporal feature vectors;

[0064] The 8 nodes in the first hidden layer of the artificial neural network are used to apply a 3×3 convolution kernel to the spatiotemporal feature vector to extract local features, identify the miner's behavioral characteristics including turning patterns, stop point distribution, and speed changes, and obtain a behavioral feature map.

[0065] The behavioral feature map undergoes a nonlinear transformation using a rectified linear unit activation function and is input into the second hidden layer of the artificial neural network. A 5×5 convolution kernel is used across eight nodes to extract contextual information with a larger receptive field. This identifies the miners' work patterns and path selection patterns in different mine areas, generating contextual feature representation data.

[0066] The mine topology constraint is applied to the contextual feature representation data, and the three-dimensional spatial, temporal and environmental information is integrated through a fully connected layer to obtain the trajectory vector.

[0067] Specifically, trajectory data contains the miner's location information at different points in time, typically represented as a (x, y, z, t) tuple, where (x, y, z) represents the miner's coordinates in the mine's three-dimensional space, and t represents the corresponding timestamp. A time-series location matrix arranges these discrete location points in chronological order, forming a set of points with temporal and spatial correlations. Furthermore, environmental parameter data (including temperature, humidity, gas concentration, and ventilation conditions) must be mapped to corresponding location points as their attribute values. This mapping establishes a correlation between location and environmental conditions, ensuring that each location point carries not only location information but also environmental parameter information. Spatial rasterization discretizes the continuous three-dimensional space into a regular three-dimensional grid, with each grid cell corresponding to a small area within the mine space. This process is similar to dividing the mine space into three-dimensional "pixels." When a location point falls into a grid cell, the cell's value increases. If multiple location points fall into the same cell, the cell values ​​accumulate to form a heat value. Different types of data (such as position, velocity, and various environmental parameters) are rasterized on different "channels," ultimately forming a multi-channel mine trajectory activity heat map, or trajectory image matrix. This matrix is ​​a three-dimensional array, with the first two dimensions representing spatial coordinates and the third dimension representing the different data channels.

[0068] Contrast enhancement of the trajectory image matrix is ​​performed to highlight trajectory features in key areas. Contrast enhancement is an image processing technique that adjusts the distribution of pixel values ​​to make important features in the image more distinct. In mine trajectory analysis, particular attention is paid to trajectory features in key areas such as tunnels, working faces, and refuge chambers, as these areas are the primary locations for miners' activities and contain important behavioral information. Specific contrast enhancement methods include histogram equalization, contrast stretching, and adaptive contrast enhancement. Histogram equalization redistributes the grayscale levels of image pixels to create a uniformly distributed grayscale histogram; contrast stretching expands the range of pixel values ​​through linear or nonlinear mapping; and adaptive contrast enhancement uses different enhancement strategies based on the characteristics of different image regions. Through contrast enhancement, trajectory features in tunnels, working faces, and refuge chambers become more distinct, forming an enhanced trajectory image. Decomposing the enhanced trajectory image into multiple channels allows for processing different types of information separately. The spatial channel records the miners' location distribution, demonstrating their activity range and hotspots within the mine. The velocity channel records miners' movement speed, reflecting the intensity of activity in different areas. The acceleration channel records changes in speed, capturing key behaviors such as starting, stopping, and turning. The environmental parameter channel records the environmental conditions at each location, correlating the relationship between activity and the environment. This high-dimensional data often contains a large amount of redundant information, requiring compression and dimensionality reduction to reduce its dimensionality. Common dimensionality reduction methods include principal component analysis (PCA), linear discriminant analysis (LDA), and autoencoders. PCA projects high-dimensional data into a low-dimensional space by identifying the main directions of variation in the data (principal components). LDA considers categorical information and finds the projection directions that best distinguish between different categories. Autoencoders use neural networks to learn a compressed representation of the data. The resulting low-dimensional representation, called a spatiotemporal feature vector, retains the key information of the original data while significantly reducing the data size.

[0069] Local features are extracted from the spatiotemporal feature vectors using the first hidden layer of the artificial neural network. The first hidden layer contains eight nodes, each corresponding to a 3×3 convolution kernel. A convolution kernel is a small 3×3 matrix that detects local patterns by sliding it over the input data and performing a convolution operation. The 3×3 size enables the convolution kernel to capture feature variations within a small area, such as local structures like corners and intersections. The convolution operation involves element-wise multiplication of the convolution kernel with a subregion of the input data and summing the results to produce the corresponding value in the output feature map. The eight different convolution kernels can detect eight different local patterns, such as vertical edges, horizontal edges, and diagonal lines. In this way, the neural network can identify behavioral characteristics such as miners' turning patterns (such as turning at tunnel intersections), dwell point distribution (such as extended stops at the working face), and speed changes (such as deceleration near the refuge). These local features collectively form a behavioral feature map, a new multi-channel feature map with each channel corresponding to a detected local behavioral pattern.

[0070] The behavioral feature map undergoes a nonlinear transformation using the rectified linear unit activation function to enhance the network's expressive power. The rectified linear unit (ReLU) is an activation function defined as f(x) = max(0, x), meaning it returns zero for negative inputs and itself for positive inputs. The primary function of the ReLU is to introduce nonlinearity, enabling the network to learn more complex patterns. Furthermore, its computational simplicity and stable gradients facilitate the training of deep networks. The ReLU-processed feature map retains positive features (representing detected patterns) while suppressing negative features (representing irrelevant patterns). These processed features are input to the second hidden layer of the artificial neural network, which also contains 8 nodes but uses a larger 5×5 convolution kernel. The 5×5 convolution kernel has a larger receptive field, enabling it to capture a wider range of contextual information. The receptive field refers to the area in the input feature map corresponding to a point in the output feature map. A larger receptive field allows it to consider more information from surrounding pixels. By using these larger convolutional kernels, the second hidden layer can identify higher-level patterns, such as miners' work patterns in different mine areas (e.g., cyclical work patterns at the working face) and path selection patterns (e.g., preferred choices among multiple paths). These high-level features form contextual feature representations of the data, encompassing not only local behavior information but also the association of these behaviors with the broader spatial and temporal context.

[0071] Applying mine topology constraints to the contextual feature representation data ensures that the generated trajectory vectors conform to the physical constraints of the mine. Mine topology refers to the spatial layout of the mine, including the connectivity of tunnels, impassable areas, and the distribution of functional areas. Using this topology as a constraint eliminates physically impossible trajectory predictions (such as those that pass through walls or over obstacles). In practice, the mine structure is typically represented as a navigation graph, where nodes represent key locations (such as intersections and functional areas) and edges represent navigable paths. The contextual features are combined with the navigation graph via fully connected layers to further integrate and refine the features. A fully connected layer is a type of layer in a neural network where every input is connected to every output, enabling the learning of global combinations of input features. This approach allows the system to integrate spatial information (where the miner is), temporal information (when they arrive), and environmental information (what the surrounding environment is like) to generate a trajectory vector that comprehensively considers these three dimensions. This trajectory vector is a highly abstract and compressed representation that captures the key features and behavioral patterns of the miner's trajectory.

[0072] For example, a worker's movement from the main shaft to various equipment maintenance points in a single day generates raw trajectory data. This trajectory data is converted into a time-series position matrix in the mine's three-dimensional coordinate system. Parameters such as temperature and methane concentration measured by environmental sensors along the way are mapped to corresponding locations. Then, spatial rasterization is used to divide the mine space into 10 cm × 10 cm × 10 cm cubes. Each cube is assigned a heat value based on the worker's frequency of movement, forming a multi-channel heat map matrix. This heat map matrix is ​​then contrast-enhanced to highlight trajectory features in key areas such as equipment maintenance points and main passages. The enhanced heat map is decomposed into a spatial channel (reflecting location distribution), a velocity channel (reflecting movement speed), an acceleration channel (reflecting speed changes), and an environmental parameter channel (reflecting environmental conditions at each point). Principal component analysis is used for dimensionality reduction, resulting in a 100-dimensional spatiotemporal feature vector. This vector is then fed into the first hidden layer of an artificial neural network. Eight 3×3 convolution kernels detect different local features, such as worker pauses at equipment points and turns at roadway intersections, to form a behavioral feature map. This mapping is processed through a ReLU activation function and then fed into the second hidden layer. Eight 5×5 convolutional kernels further extract higher-level features, such as the worker's inspection route pattern and the patterns of work performed in different areas at different times, to form a contextual feature representation. Finally, these features are combined with the mine navigation map, and a fully connected layer integrates all information to generate a trajectory vector that accurately describes the maintenance worker's behavior patterns.

[0073] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0074] The prediction time is divided into three scales: short-term prediction, medium-term prediction and long-term prediction. The trajectory vector is divided into three time dimensions to obtain three sets of prediction input data.

[0075] The short-term prediction input data is fed into the temporal convolutional network structure. The output of each convolutional layer is processed by applying the rectified linear unit activation function to capture the local continuity and short-term change pattern of the trajectory, thus obtaining the short-term trajectory prediction result.

[0076] The mid-term prediction input data is fed into the self-attention mechanism network structure, and the long-distance dependency relationship is modeled by calculating the correlation weight matrix of each time point in the trajectory sequence to obtain the mid-term trajectory prediction result;

[0077] The long-term prediction input data is fed into the recurrent neural network structure, and the output of the hidden state is processed by the Sigmoid function to model the long-term uncertainty of the trajectory and obtain the long-term trajectory prediction result;

[0078] Based on the quality weight of the trajectory data, a weighted loss function is constructed to calculate the prediction error for the short-term trajectory prediction results, the medium-term trajectory prediction results, and the long-term trajectory prediction results. The target prediction result is obtained by calculating the gradient and updating the network parameters.

[0079] The target prediction results are fused and the prediction points that do not meet the physical reachability are filtered out using the mine space constraints to obtain the trajectory prediction data.

[0080] Specifically, categorizing the prediction time scale into short-term, medium-term, and long-term is a key strategy in mine personnel trajectory prediction methods. This multi-scale prediction framework provides specialized mechanisms for forecasting requirements across different time spans. Specifically, short-term prediction typically refers to trajectory predictions for the next 5-15 minutes, focusing on the miner's upcoming location; medium-term predictions refer to trajectory predictions for the next 15-60 minutes, focusing on the miner's movement path within a single shift; and long-term predictions refer to trajectory predictions for the next 1-4 hours, focusing on the miner's activity patterns throughout the entire shift. Temporal segmentation of trajectory vectors involves dividing them into three different sets of prediction input data based on the length of historical data and the time span of the prediction target. For example, for short-term predictions, the last 10 minutes of trajectory data might be used as input; for medium-term predictions, the last 30 minutes of data might be used; and for long-term predictions, the last two hours of data might be used. This segmentation takes into account the varying degree of historical information dependence of prediction tasks at different time scales: short-term predictions rely more on recent behavior, while long-term predictions require more historical information to identify patterns.

[0081] Feeding short-term prediction input data into a temporal convolutional network (TCN) is a core step in short-term trajectory prediction. A temporal convolutional network (TCN) is a variant of a convolutional neural network specifically designed for processing time series data. Its hallmarks are its use of one-dimensional convolution operations to process time series data, while causal and dilated convolution techniques expand the receptive field. Causal convolution ensures that predictions depend only on current and past inputs and are unaffected by future data. Dilated convolution, by inserting holes in the convolution kernel, effectively increases the receptive field and captures longer-range temporal dependencies. In mine personnel trajectory prediction, a TCN receives short-term prediction input data and extracts temporal features of the trajectory through multiple layers of convolution operations. The output of each convolutional layer is processed using a rectified linear unit (RLU) activation function, which retains useful positive features while zeroing out irrelevant negative features, enhancing the network's expressive power. In this way, the TCN effectively captures both local continuity (such as continuous movement along the same roadway) and short-term variation patterns (such as back-and-forth movement at a working face) in the trajectory. Ultimately, it outputs a short-term trajectory prediction, representing a sequence of possible miner positions within the next 5-15 minutes. Feeding mid-term forecast input data into a self-attention network structure is a method for processing forecasts over longer time horizons. The self-attention mechanism is a technique that computes relationships between elements within a sequence. It learns the strength of the association between each element and all other elements in the sequence and generates a weight matrix that expresses these associations. In mine personnel trajectory prediction, the self-attention mechanism allows the model to account for long-range dependencies between different time points in the trajectory sequence, which is crucial for capturing miners' movement patterns over longer periods of time. Specifically, mid-term forecast input data (such as trajectories from the past 30 minutes) is fed into the self-attention network. The network first calculates the query vector, key vector, and value vector for each time point in the trajectory sequence. It then calculates a correlation score by taking the dot product of the query vector and the key vector. The scores are then converted to weights using a softmax function. Finally, the value vector is weighted summed using these weights to produce a representation that accounts for long-range dependencies. This mechanism can identify trajectory points that are distant in time but strongly correlated, such as the behavior of miners regularly returning to an area to inspect equipment, thereby generating more accurate mid-term trajectory forecasts.

[0082] Feeding long-term prediction input data into a recurrent neural network (RNN) is a method for processing predictions over the longest timeframe. A recurrent neural network (RNN) is a specialized neural network for processing sequential data. Its characteristic is its ability to transfer information through its internal state (memory), enabling it to process sequences of arbitrary length. In mine personnel trajectory prediction, long-term predictions require considering the patterns of miners' activities throughout an entire work shift, or even across multiple shifts. Traditional RNNs struggle to handle such long-term dependencies, so improved RNN variants such as long-short-term memory (LSTM) networks or gated recurrent units (GRU) are often employed. These variants incorporate gating mechanisms to selectively remember important information and forget irrelevant information, effectively processing long sequences. Long-term prediction input data (e.g., trajectories from the past two hours) is fed into a recurrent neural network. The network continuously updates its internal state with each time step, accumulating historical information. Finally, the network's hidden state is processed using a sigmoid function, compressing the output value to a range between 0 and 1. This processing helps model long-term trajectory uncertainty, as long-term predictions often have higher uncertainty and need to be expressed probabilistically. In this way, recurrent neural networks can generate trajectory predictions that account for long-term behavioral patterns.

[0083] Each trajectory is assigned a quality weight, reflecting the reliability of its data quality. These quality weights are used to construct a weighted loss function, assigning different learning importance to trajectory samples of different qualities. Specifically, for the short-term, medium-term, and long-term prediction results, the error between the predicted value and the true value is calculated. Common error metrics include mean absolute error (MAE), mean squared error (MSE), or mean absolute percentage error (MAPE). These errors are then weighted by the quality weight, so that the error from high-quality trajectories contributes more to the total loss, thereby guiding the model to learn more patterns from high-quality samples. The weighted loss function calculates the gradient and updates the network parameters using optimization algorithms such as gradient descent, continuously adjusting the model to reduce the weighted loss, ultimately obtaining the target prediction result optimized based on high-quality data. Short-term, medium-term, and long-term predictions each have their own advantages: short-term predictions are more accurate in the near term, medium-term predictions can capture longer-term dependencies, and long-term predictions are better at grasping overall patterns. Fusion processing combines these three prediction results to generate the final predicted trajectory. Common fusion methods include weighted averaging, ensemble learning, or multi-task learning. In the weighted averaging method, the three prediction results are combined according to pre-set weights or dynamically calculated weights based on the reliability of the current prediction. In the ensemble learning method, the three predictions are treated as different "expert opinions" and integrated through voting or averaging. In the multi-task learning method, the three prediction tasks share underlying features but have their own prediction heads, achieving complementarity through joint optimization. The fused prediction results are then filtered using the mine's spatial constraints to eliminate predicted points that are not physically accessible. These constraints are derived from the mine's three-dimensional structural diagram and ensure that the predicted trajectory does not involve physically impossible situations such as penetrating walls, overcoming obstacles, or traversing unreasonable speeds. After constraint filtering, the final trajectory prediction data is obtained, which conforms to physical laws.

[0084] For example, a worker performs cyclical work at the coal mining face every day, with a relatively fixed work pattern. First, trajectory vectors are extracted from historical trajectory data and then divided into three groups based on the prediction time scale: short-term prediction uses trajectory vectors from the last 15 minutes, medium-term prediction uses trajectory vectors from the last 45 minutes, and long-term prediction uses trajectory vectors from the last two hours. For short-term prediction, the trajectory vectors from the last 15 minutes are input into a temporal convolutional network. The network processes this through multiple layers of convolution and ReLU activation functions to capture the worker's local movement patterns on the coal mining face, such as round-trip movement from one end to the other, and predicts the worker's possible position changes over the next 10 minutes. For medium-term prediction, the trajectory vectors from the last 45 minutes are input into a self-attention network. By calculating the correlation weight matrix between different time points, the network identifies a pattern in which the worker returns to the support control area for adjustments approximately every 30 minutes, and predicts the worker's activity trajectory for the next 30 minutes. For long-term prediction, the trajectory vectors from the last two hours are fed into an LSTM network. The network accumulates historical information through memory cells, identifying changes in a worker's work rhythm throughout a shift, such as a pattern of intense activity in the first half and a slight slowdown in the second half. This allows the network to predict the approximate range of activity for the remaining working time. These three predictions are then trained and optimized using a loss function based on trajectory quality weights. For example, a trajectory at a coal face with good signal quality and high data quality receives a higher quality weight and a greater weight in the loss calculation. Finally, the three optimized predictions are fused. The short-term prediction contributes most to the nearest point location, the medium-term prediction contributes most to the location within approximately 30 minutes, and the long-term prediction provides guidance on overall trends. The fused trajectory also passes mine structural constraints to ensure that the predicted path follows only existing roadways and working faces, avoiding crossing walls or obstacles, thereby generating trajectory prediction data that conforms to physical laws.

[0085] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0086] Compare the spatial coordinates of the trajectory prediction data with the three-dimensional structure diagram of the mine, and eliminate the predicted points that do not match the mine passage, including points inside the rock layer, points that cross the safety isolation area, and points that exceed the working range, to obtain the spatially constrained trajectory data;

[0087] The historical trajectories of multiple miners in the spatially constrained trajectory data are classified and sorted, and grouped according to work teams, work nature, and activity areas to obtain a grouped trajectory set;

[0088] Apply density clustering based on computer vision to the grouped trajectory set, clustering trajectory points that are similar in time and space into a group, and extracting the center point and activity range of each group to obtain a behavior density distribution map;

[0089] Based on the behavior density distribution map, the activity patterns of miners in different areas are identified, including the stay patterns at the work site, inspection route patterns, material transportation patterns, and emergency evacuation patterns, to obtain behavior pattern data;

[0090] The spatial constraint trajectory data is matched with the behavior pattern data. According to the current miner position and task type, the direction and rhythm of the predicted trajectory are adjusted to make it conform to the characteristics of the corresponding behavior pattern, and the behavior optimization trajectory is obtained.

[0091] Based on the collaborative relationship between miners in the same work group, the behavior optimization trajectory is adjusted for group consistency to ensure the spatiotemporal coordination of the predicted trajectories of the collaborative miners and obtain the target trajectory prediction data.

[0092] Specifically, a 3D mine structure diagram is a digital representation of the mine space, encompassing the spatial layout and geometry of various functional areas, such as tunnels, working faces, and refuges, as well as the location of inaccessible areas, such as rock formations and safety isolation zones. The spatial coordinate comparison process first compares the 3D coordinates (x, y, z) of each predicted point with the passageway location in the mine structure diagram to check whether the point falls within the legal traversable area. Predicted points located within the rock formation—that is, points whose coordinates fall within the solid rock area—clearly violate the laws of physics, as miners cannot pass through walls and must instead move along the excavated tunnels and working faces. Predicted points that cross safety isolation zones—that is, points whose coordinates pass through restricted areas established for safety reasons (e.g., high gas concentrations, unstable roofs)—are considered safety violations and should not be entered by miners. Predicted points that fall outside the working range—that is, points whose coordinates fall outside the normal working range of miners—may be unreasonable predictions due to model extrapolation. By eliminating these three types of predicted points that do not conform to the mine's physical constraints, the remaining predicted points form spatially constrained trajectory data, representing physically feasible miner movement paths. The historical trajectories of multiple miners in the spatially constrained trajectory data are categorized to identify the behavioral patterns of different miner types. First, trajectories of miners working on the same shift (e.g., morning, afternoon, and night) are grouped together, as miners in the same shift typically have similar work schedules and collaborative relationships. Second, trajectories of miners with the same job type (e.g., coal miners, support workers, and mechanical and electrical maintenance workers) are grouped together, as miners of the same job type perform similar tasks and exhibit similar movement patterns. Finally, trajectories of miners primarily located in the same area (e.g., a working face or a section of a haulage tunnel) are grouped together, as regional characteristics can influence miners' movement behavior. This multi-dimensional categorization results in a set of grouped trajectories.

[0093] Applying computer vision-based density clustering to grouped trajectory sets is a method that uses visual analysis techniques to identify trajectory patterns. The core of density clustering is the DBSCAN (Density-Based Spatial Clustering Applications with Noise) algorithm, which groups data points based on their density distribution. It can detect clusters of arbitrary shapes and identify noise points. In trajectory analysis, spatiotemporal trajectory points are first mapped into two-dimensional or three-dimensional space, where points with similar temporal and spatial proximity are represented as clustered areas. The DBSCAN algorithm is based on two parameters: ε (neighborhood radius) and MinPts (minimum number of points required to form a cluster). The algorithm iterates over each point. If a point's ε-neighborhood contains at least MinPts points, it is considered a core point. If a point is within the ε-neighborhood of a core point but is not itself a core point, it is considered a boundary point. Points that are neither core points nor boundary points are considered noise points. Clusters are formed by connecting density-reachable core points. For each cluster, its center (the average coordinate value of all points) and activity range (the spatial distribution of the points) are calculated to form a behavioral density distribution map. This distribution map visually demonstrates the concentration and boundaries of miners' activities in different areas of the mine. By analyzing the spatial location, density variation, and time series characteristics of the cluster distribution, typical behavioral patterns can be identified. Work point dwell patterns are characterized by high-density clusters formed at specific locations (such as shearer operating positions and support work areas). These areas have high density and long time spans of trajectory points, indicating that miners perform fixed tasks at these locations for extended periods of time. Inspection route patterns are characterized by continuous linear trajectories with moderate density, covering multiple pieces of equipment or areas, reflecting miners' behavior of inspecting and monitoring equipment along predetermined routes. Material transportation patterns are characterized by round-trip trajectories between material storage and use points, often exhibiting regular cycles, reflecting material handling and distribution. Emergency evacuation patterns are characterized by rapid movement from the work area to the safety exit or refuge room, with high speed and relatively direct paths, reflecting evacuation behavior in emergency situations. By identifying and classifying these patterns, behavioral pattern data is generated, containing characteristic descriptions and parameter settings for different types of behavior.

[0094] Based on the miner's current location and task information (obtained from the scheduling system or work order), the most likely behavior pattern for the miner is determined. For example, if a miner is located at a coal face and tasked with operating a shearer, their behavior pattern might be a "coal mining point dwell mode." The predicted trajectory is then compared with the characteristics of this behavior pattern. If the predicted trajectory deviates from the typical characteristics of the behavior pattern (for example, the predicted trajectory does not reflect the expected dwell time at the work point), adjustments are made. These adjustments include correcting the predicted path's direction to align with the typical path selection of the behavior pattern; adjusting the movement speed and rhythm to align with the typical activity characteristics of the behavior pattern; and increasing or decreasing dwell time at key points to align with operational patterns. Through this matching and adjustment process, the predicted trajectory is made to better align with the miner's actual work behavior characteristics, forming an optimized behavior trajectory. Mining operations are typically team-based, with collaborative relationships between miners in the same workgroup, such as the need for coordination between support workers and shearers, and the need for time synchronization between transporters and loaders. Group consistency adjustment first constructs a network of relationships among miners within the workgroup, defining the types of collaboration between members (such as sequential connection, simultaneous coordination, and mutual support) and constraints (such as time difference and distance requirements). The optimized behavior trajectory is then checked to ensure that these collaborative relationships are met. For example, if the predicted trajectories of two collaborating miners indicate that they will arrive at locations where they should be working simultaneously at different times, their schedules need to be adjusted to synchronize them. If the predicted trajectories indicate that the distance between the collaborating miners exceeds the effective collaborative range, their positions need to be adjusted to maintain an appropriate working distance. This group consistency adjustment ensures that the predicted trajectory of each miner in the team is coordinated in time and space with the trajectories of other collaborating members, ultimately generating target trajectory prediction data that conforms to actual operational patterns.

[0095] For example, a work group consists of four miners: shearer operator A, support workers B and C, and safety monitor D. First, the preliminary trajectory prediction data generated by the model was compared with the three-dimensional structure diagram of the mine. Issues were identified in some prediction points: three prediction points for shearer A fell within the coal wall adjacent to the roadway, two prediction points for support worker B passed through the temporarily closed return air lane, and several prediction points for safety monitor D exceeded their assigned area. After eliminating these unreasonable points, spatially constrained trajectory data was generated that met physical constraints. Next, these trajectory data were categorized by work group (all belonging to the morning shift), work type (mining, support, and monitoring), and activity area (all working on the same working face but responsible for different sections), forming a grouped trajectory set. Density clustering was then performed using the DBSCAN algorithm, with ε = 5 meters (considering the width of the mine passages) and MinPts = 10 (based on the data sampling frequency). The algorithm identified multiple clusters: a high-density area near the shearer, a strip-like distribution in the support operation area, and a regular, point-like distribution near the ventilation equipment. These clusters together formed a behavioral density distribution map. Based on this distribution map, several typical behavioral patterns were further identified: shearer A exhibited a "reciprocating operation mode," manipulating the shearer back and forth along the working face; support workers B and C exhibited a "follow-up support mode," performing support operations at a distance behind the shearer; and safety monitor D exhibited a "patrol monitoring mode," regularly checking each monitoring point. Matching the spatially constrained trajectory data with these behavioral patterns revealed that some of shearer A's predictions did not reflect the expected reciprocating rhythm, and support worker B's predicted trajectory did not match the mining progress, requiring adjustments. After these adjustments, the coordination within the work group was considered: shearer A was required to maintain a safe distance from support workers B and C, support work was required to track mining progress, and safety monitor D was required to communicate regularly with all three. Through group consistency adjustment, the predicted trajectories of the four people are ensured to be coordinated in time and space to form the target trajectory prediction data.

[0096] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0097] The target trajectory prediction data is grouped and sorted according to the miner ID and time series, and the position coordinate sequence of each miner in the future time period is calculated to obtain the personnel distribution prediction map;

[0098] Based on the personnel distribution prediction map, the minimum spatial distance between any two miners is calculated using the Euclidean distance formula. If the minimum spatial distance is less than the preset safety threshold, it is marked as a potential dangerous contact point, and the spatial safety distance index is obtained;

[0099] Conduct a temporal probability analysis of potential dangerous contact points, calculate the probability of two miners appearing in the same area and at the same time, and obtain a temporal conflict probability index;

[0100] Map the target trajectory prediction data onto the mine area division map, count the density of miners in each area in different time periods, calculate the load rate by the ratio of regional capacity to actual number of people, and obtain the regional load index;

[0101] Based on the spatial safety distance index, the temporal conflict probability index, and the regional load index, a multi-objective optimization model was constructed. The three indicators were combined into a safety risk score using a linear weighting method. The risk level of each area and time period was divided into different levels to obtain a risk distribution map.

[0102] Based on the risk distribution map, a mine personnel safety scheduling strategy is generated, which includes personnel diversion suggestions, work time adjustment plans and emergency evacuation routes. The diversion suggestions are calculated by minimizing the regional load variance, the time adjustment plan is generated by the peak-shifting algorithm, and the emergency evacuation route is determined by the shortest path algorithm.

[0103] Specifically, the target trajectory prediction data is grouped and organized by miner ID and time series to organize the prediction results in an orderly manner by individual and time dimensions. This involves grouping all miners' predicted trajectory data by unique identifier (miner ID) and then sorting each group in ascending timestamp order. This generates a continuous sequence of position coordinates for each miner over a future time period (e.g., the next 2-4 hours). These sequences are aggregated to form a personnel distribution prediction map, a spatiotemporal data structure that records the projected locations of all miners at various points in the future.

[0104] Calculating the minimum spatial distance between any two miners based on the personnel distribution prediction map is an important part of safety management. This process uses the Euclidean distance formula, which is calculated as follows:

[0105]

[0106] in, Represents miners and miners The minimum spatial distance between them; T represents the prediction time set; 、 、 Representing miners The three-dimensional coordinate values ​​at time t; 、 、 Representing miners The three-dimensional coordinate values ​​at time t are calculated. The Euclidean distance between the two miners at each point in the predicted time range is calculated, and the minimum value is taken as the minimum spatial distance between the two miners. If this minimum distance is less than a preset safety threshold (such as 1.5 meters, set based on epidemic prevention or safe operation requirements), the time-space point is marked as a potential dangerous contact point, forming a spatial safety distance indicator. This indicator not only includes the distance value, but also the time and spatial location information of the minimum distance.

[0107] A further step in assessing risk is to perform a temporal probability analysis of potential hazardous contact points. This process calculates the probability of two miners being in the same area at the same time. The calculation method is:

[0108]

[0109] in, Represents miners and miners Time period within region R The probability of conflict in represents the distance between two people at time t; Indicates the safety distance threshold; and They represent the positions of the two people at time t respectively; It is an indicator function, which takes the value 1 when the condition is true, and 0 otherwise; Represents the length of the time period. This calculation considers two factors: whether the distance between two miners is less than the safety threshold, and whether they are in the same area at the same time. By counting the time points that meet the conditions within the predicted time period and dividing it by the total time length, we obtain the temporal collision probability index. This index reflects the probability of close contact between two miners within a specific area and time period.

[0110] Mapping the predicted target trajectory data onto a mine zone map is the starting point for analyzing regional load. A mine zone map divides the mine space into multiple zones based on functional or physical boundaries, such as the working face, haulage tunnels, and refuge rooms. The mapping process involves assigning each miner's predicted position at each point in time to the zone to which they belong. The density of miners in each zone (that is, the number of people per unit area) is then calculated over different time periods. The load factor is then calculated by comparing the actual number of miners to the zone's capacity (the maximum number of people allowed based on the area and safety regulations). For example, if the working face has a capacity of 10 people and a forecast of 7 people will be present during a certain time period, the load factor is 70%. This calculation requires considering the granularity of the time period, typically hourly or half-hourly. The load factor for each zone in each time period is then calculated to form a regional load index. Constructing a multi-objective optimization model based on the spatial safety distance index, the temporal conflict probability index, and the regional load index is the core of safety risk assessment. A multi-objective optimization model is an optimization problem that simultaneously considers multiple objective functions. In this context, each of the three indicators represents a safety risk factor. The linear weighting method linearly combines multiple objective functions into a single objective function. Specifically, the three indicators are assigned different weights based on their importance, and the weighted sum is used to generate a comprehensive safety risk score. For example, the spatial safety distance indicator has a weight of 0.4, the temporal conflict probability indicator has a weight of 0.3, and the regional load indicator has a weight of 0.3, with the total weights summing to 1. This results in a comprehensive safety risk score for each area and time period. Based on these scores, each area and time period is classified according to pre-defined risk classification criteria (such as low risk, medium risk, and high risk). This generates a visual risk distribution map that displays the risk level of different areas at different times.

[0111] The final application step is generating a safe mine personnel scheduling strategy based on the risk distribution map. This strategy consists of three main components: personnel diversion recommendations, work schedule adjustment plans, and emergency evacuation routes. Personnel diversion recommendations address overcrowding by redistributing personnel to areas with high loads. These recommendations are calculated by minimizing the regional load variance, maximizing load balance across areas to avoid overcrowding in some areas and underutilization in others. Work schedule adjustment plans address high time conflicts by staggering work schedules. These plans are generated using a peak-shifting algorithm, which adjusts the work hours of different types of work or teams to reduce the concentration of personnel during the same time period. Emergency evacuation routes are prepared for emergencies and guide miners to evacuate quickly and safely. These routes are determined using a shortest path algorithm (such as the Dijkstra algorithm), which finds the shortest path from any location to the nearest safe exit, taking into account the mine's physical structure and the distribution of potential hazardous areas.

[0112] The above describes the mine personnel trajectory prediction method based on artificial intelligence in the embodiment of the present application. The following describes the mine personnel trajectory prediction system based on artificial intelligence in the embodiment of the present application. Figure 2 In the embodiments of the present application, an embodiment of a mine personnel trajectory prediction system based on artificial intelligence includes:

[0113] The acquisition module 201 is used to collect mine personnel positioning data and environmental parameter data, and perform standardization processing on the mine personnel positioning data to obtain structured trajectory data;

[0114] A calculation module 202 is configured to perform a quality assessment on the structured trajectory data by constructing an artificial neural network with two hidden layers to calculate the trajectory quality and obtain trajectory data;

[0115] A conversion module 203 is used to convert the trajectory data and the environmental parameter data into a spatiotemporal feature vector, and perform image pattern recognition processing through the artificial neural network to obtain a trajectory vector;

[0116] a processing module 204 for applying a rectified linear unit activation function and a sigmoid function to the trajectory vector to perform multi-time scale processing, and calculating a gradient using a loss function weighted by a quality weight to obtain trajectory prediction data;

[0117] Matching module 205, for performing constraint matching between the trajectory prediction data and the spatial structure of the mine, identifying the behavior pattern of the miner group through a density clustering algorithm based on computer vision, and obtaining target trajectory prediction data;

[0118] The generation module 206 is used to calculate the spatial safety distance, time conflict probability and regional load index based on the target trajectory prediction data, and generate a mine personnel safety scheduling strategy.

[0119] Through the collaborative efforts of the aforementioned components, a structured representation of miner trajectories is achieved by collecting and standardizing mine personnel location data and environmental parameter data. This structured trajectory data is then assessed for quality, and trajectory quality is calculated using an artificial neural network with two hidden layers. This effectively addresses the heterogeneity of data quality in the mine environment and improves the reliability of the prediction model training data. Trajectory and environmental parameter data are converted into spatiotemporal feature vectors and processed using an artificial neural network for image pattern recognition, leveraging the advantages of deep learning in spatial pattern recognition to effectively extract key features of miner trajectories. The trajectory vectors are processed using a rectified linear unit activation function and a sigmoid function for multi-time scale processing, enabling trajectory prediction for short, medium, and long time spans, meeting the needs of diverse application scenarios. The design of a quality-weighted loss function for calculating the gradient effectively suppresses the interference of low-quality data on model training. The predicted trajectory data is constrained and matched to the mine's spatial structure, and a computer vision-based density clustering algorithm is used to identify the behavioral patterns of miner groups. This ensures that the prediction results are consistent with both the mine's physical constraints and the actual working patterns of miners, significantly improving prediction accuracy. Finally, based on the target trajectory prediction data, the spatial safety distance, temporal conflict probability, and regional load index are calculated, and a mine personnel safety scheduling strategy is generated, realizing the direct transformation of the prediction results into safety management decisions. In the specific application field of mines, this solution fully considers the contribution of artificial intelligence algorithm features to the solution, especially the multi-level feature extraction capabilities of artificial neural networks, the nonlinear mapping capabilities of rectified linear units and sigmoid activation functions, the pattern recognition capabilities of computer vision, and the group behavior modeling capabilities of density clustering. Compared with traditional methods, the present invention can not only accurately predict the future trajectory of miners, but also identify potential safety risks and automatically generate scheduling strategies, significantly improving the accuracy of mine safety management.

[0120] above Figure 2 The artificial intelligence-based mine personnel trajectory prediction system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The artificial intelligence-based mine personnel trajectory prediction device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0121] Figure 3This is a schematic diagram of the structure of an AI-based mine personnel trajectory prediction device provided by an embodiment of the present invention. The AI-based mine personnel trajectory prediction device 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage medium 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instruction operations within the AI-based mine personnel trajectory prediction device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, allowing the AI-based mine personnel trajectory prediction device 300 to execute the series of instruction operations stored in the storage medium 330 to implement the steps of the aforementioned AI-based mine personnel trajectory prediction method.

[0122] The artificial intelligence-based mine personnel trajectory prediction device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the artificial intelligence-based mine personnel trajectory prediction device shown does not constitute a limitation on the artificial intelligence-based mine personnel trajectory prediction device provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0123] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the artificial intelligence-based mine personnel trajectory prediction method.

[0124] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling an artificial intelligence-based mine personnel trajectory prediction device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A mine personnel trajectory prediction method based on artificial intelligence, characterized in that: The method comprises: Collecting mine personnel positioning data and environmental parameter data, and performing standardization processing on the mine personnel positioning data to obtain structured trajectory data; Performing a quality assessment on the structured trajectory data, calculating the trajectory quality by constructing an artificial neural network with two hidden layers to obtain trajectory data; Converting the trajectory data and the environmental parameter data into a spatiotemporal feature vector, performing image pattern recognition processing through the artificial neural network to obtain a trajectory vector; Applying a rectified linear unit activation function and a sigmoid function to the trajectory vector for multi-time scale processing, calculating the gradient through a loss function weighted by a quality weight, and obtaining trajectory prediction data; The trajectory prediction data is constrained and matched with the spatial structure of the mine, and the behavior pattern of the miner group is identified by a density clustering algorithm based on computer vision to obtain target trajectory prediction data, including: performing spatial coordinate comparison between the trajectory prediction data and the three-dimensional structure diagram of the mine, eliminating prediction points that do not match the mine passage, including points located inside the rock layer, points crossing the safety isolation area, and points beyond the working range, to obtain spatial constraint trajectory data; classifying and arranging the historical trajectories of multiple miners in the spatial constraint trajectory data, grouping them according to work teams, work nature and activity areas, and obtaining a group trajectory set; applying density clustering processing based on computer vision to the group trajectory set, clustering trajectory points that are similar in time and space The miners are grouped into a group, and the center point and activity range of each group are extracted to obtain a behavior density distribution map; based on the behavior density distribution map, the activity patterns of miners in different areas are identified, including the stay pattern at the work point, the inspection route pattern, the material transportation pattern, and the emergency evacuation pattern, to obtain behavior pattern data; the spatial constraint trajectory data is matched with the behavior pattern data, and the direction and rhythm of the predicted trajectory are adjusted according to the current miner position and task type to make it conform to the characteristics of the corresponding behavior pattern, to obtain the behavior optimization trajectory; based on the collaborative relationship between miners in the same work group, the behavior optimization trajectory is adjusted for group consistency to ensure the spatiotemporal coordination of the predicted trajectory of the collaborative miners, to obtain the target trajectory prediction data; Based on the target trajectory prediction data, the spatial safety distance, time conflict probability and regional load index are calculated to generate a mine personnel safety scheduling strategy.

2. The method for predicting mine personnel trajectories based on artificial intelligence according to claim 1, characterized in that: The collecting of mine personnel positioning data and environmental parameter data, and the standardization of the mine personnel positioning data to obtain structured trajectory data include: A sensor network deployed at preset locations in the mine collects location coordinate data and timestamp data uploaded by the smart positioning devices worn by miners to obtain original trajectory data. It also collects temperature, humidity, gas concentration, and ventilation data in the mine area to obtain environmental parameter data. Cleaning the original trajectory data to remove coordinate points beyond the physical boundary of the mine, abnormal time stamp points, and coordinate jump points to obtain cleaned trajectory data; The missing data points in the cleaned trajectory data are supplemented by a spatiotemporal interpolation algorithm to obtain a trajectory sequence; Segmenting the trajectory sequence according to a preset time window to obtain multiple trajectory segments; Extracting spatial features and temporal features of the plurality of trajectory segments, including displacement vector, velocity, acceleration, steering angle, dwell time, and movement frequency, to obtain characterized trajectory data; The characterized trajectory data is associated with the functional attributes of the mine area, and the trajectory points are semantically annotated to obtain the structured trajectory data.

3. The method for predicting mine personnel trajectory based on artificial intelligence according to claim 1, characterized in that: The quality assessment of the structured trajectory data is performed by constructing an artificial neural network with two hidden layers to calculate the trajectory quality to obtain the trajectory data, including: Performing a multi-dimensional scoring on the structured trajectory data to obtain initial quality parameters, and performing a weighted combination of the initial quality parameters according to the importance of each indicator using a preset trajectory quality scoring function to obtain a normalized quality score; An artificial neural network with two hidden layers is constructed and trained, wherein the input layer has two nodes, the two hidden layers each have eight nodes, and the output layer has one node, thereby obtaining an artificial neural network; Converting the structured trajectory data into a two-dimensional spatiotemporal image, inputting the image into the artificial neural network, extracting local features through convolution operations, and extracting global features through pooling operations and full connection operations to obtain a trajectory quality score; A trajectory contrast learning task is constructed to generate similar scores for trajectories of the same miner with similar time periods, and different scores for trajectories with significant differences in regularity and periodicity. The network parameters are optimized by minimizing the contrastive loss function to obtain a calibration quality score. Based on the calibration quality score, a quality weight is assigned to the structured trajectory data to obtain the trajectory data.

4. The method for predicting mine personnel trajectory based on artificial intelligence according to claim 1, characterized in that: The step of converting the trajectory data and the environmental parameter data into a spatiotemporal feature vector and performing image pattern recognition processing through the artificial neural network to obtain a trajectory vector includes: The trajectory data is converted into a time-series position matrix in a three-dimensional coordinate system of a mine, the environmental parameter data is mapped into attribute values ​​of corresponding position points, and a multi-channel mine trajectory activity heat map is generated through spatial rasterization processing to obtain a trajectory image matrix; Performing contrast enhancement on the trajectory image matrix to obtain an enhanced trajectory image, wherein the contrast enhancement is used to enhance trajectory features of the tunnel, working surface, and refuge chamber; Decomposing the enhanced trajectory image into a spatial channel, a velocity channel, an acceleration channel, and an environmental parameter channel, performing compression and dimensionality reduction processing to generate a spatiotemporal feature vector; Applying a 3×3 convolution kernel to the spatiotemporal feature vector through the eight nodes of the first hidden layer of the artificial neural network to perform local feature extraction, identify the miner's behavioral features including turning patterns, stop point distribution, and speed changes, and obtain a behavioral feature map; The behavioral feature map is nonlinearly transformed using a modified linear unit activation function and input into the second hidden layer of the artificial neural network. A 5×5 convolution kernel is used through 8 nodes to extract contextual information with a larger receptive field, identify the working patterns and path selection patterns of miners in different mine areas, and obtain contextual feature representation data; A mine topology constraint is applied to the context feature representation data, and the three-dimensional information of space, time and environment is integrated through a fully connected layer to obtain the trajectory vector.

5. The method for predicting mine personnel trajectory based on artificial intelligence according to claim 1, characterized in that: The step of applying a rectified linear unit activation function and a sigmoid function to the trajectory vector for multi-time scale processing, calculating a gradient through a quality-weighted loss function, and obtaining trajectory prediction data includes: The prediction time is divided into three scales: short-term prediction, medium-term prediction and long-term prediction, and the trajectory vector is divided into three time dimensions to obtain three sets of prediction input data; The short-term prediction input data is fed into the temporal convolutional network structure. The output of each convolutional layer is processed by applying the rectified linear unit activation function to capture the local continuity and short-term change pattern of the trajectory, thus obtaining the short-term trajectory prediction result. The mid-term prediction input data is fed into the self-attention mechanism network structure, and the long-distance dependency relationship is modeled by calculating the correlation weight matrix of each time point in the trajectory sequence to obtain the mid-term trajectory prediction result; The long-term prediction input data is fed into the recurrent neural network structure, and the output of the hidden state is processed by the Sigmoid function to model the long-term uncertainty of the trajectory and obtain the long-term trajectory prediction result; Based on the quality weight of the trajectory data, a weighted loss function is constructed, prediction errors are calculated for the short-term trajectory prediction results, the medium-term trajectory prediction results, and the long-term trajectory prediction results, and a target prediction result is obtained by calculating the gradient and updating the network parameters; The target prediction results are fused and the prediction points that do not meet the physical reachability are filtered out through the mine space constraint to obtain the trajectory prediction data.

6. The method for predicting mine personnel trajectory based on artificial intelligence according to claim 1, characterized in that: The method of calculating the spatial safety distance, time conflict probability and regional load index based on the target trajectory prediction data and generating a mine personnel safety scheduling strategy includes: The target trajectory prediction data is grouped and sorted according to the miner ID and time series, and the position coordinate sequence of each miner in the future time period is calculated to obtain a personnel distribution prediction map; Based on the personnel distribution prediction map, the minimum spatial distance between any two miners is calculated using the Euclidean distance formula. If the minimum spatial distance is less than a preset safety threshold, it is marked as a potential dangerous contact point, and a spatial safety distance index is obtained; Performing a time probability analysis on the potential dangerous contact points, calculating the probability of two miners appearing in the same area and at the same time period, and obtaining a time conflict probability index; Mapping the target trajectory prediction data onto a mine area division map, counting the density of miners in each area in different time periods, and calculating the load rate by the ratio of regional capacity to actual number of people to obtain a regional load index; Based on the spatial safety distance index, time conflict probability index and regional load index, a multi-objective optimization model is constructed. The three indicators are combined into a safety risk score through a linear weighting method, and each area and time period is divided into risk levels to obtain a risk distribution map; Based on the risk distribution map, a mine personnel safety scheduling strategy is generated, which includes personnel diversion suggestions, work time adjustment plans and emergency evacuation routes. The diversion suggestions are calculated by minimizing the regional load variance, the time adjustment plans are generated by the peak-shifting algorithm, and the emergency evacuation routes are determined by the shortest path algorithm.

7. A mine personnel trajectory prediction system based on artificial intelligence, characterized in that: For implementing the method for predicting mine personnel trajectory based on artificial intelligence according to any one of claims 1 to 6, the system for predicting mine personnel trajectory based on artificial intelligence comprises: An acquisition module is used to collect mine personnel positioning data and environmental parameter data, and perform standardization processing on the mine personnel positioning data to obtain structured trajectory data; a calculation module, configured to perform quality assessment on the structured trajectory data, and calculate the trajectory quality by constructing an artificial neural network with two hidden layers to obtain trajectory data; a conversion module, configured to convert the trajectory data and the environmental parameter data into a spatiotemporal feature vector, and perform image pattern recognition processing through the artificial neural network to obtain a trajectory vector; a processing module, configured to apply a rectified linear unit activation function and a sigmoid function to the trajectory vector for multi-time scale processing, calculate a gradient using a loss function weighted by a quality weight, and obtain trajectory prediction data; The matching module is used to perform constraint matching between the trajectory prediction data and the spatial structure of the mine, identify the behavior pattern of the miner group through the density clustering algorithm based on computer vision, and obtain the target trajectory prediction data, including: performing spatial coordinate comparison between the trajectory prediction data and the three-dimensional structure diagram of the mine, eliminating the predicted points that do not match the mine passage, including points located inside the rock layer, points crossing the safety isolation area, and points beyond the working range, to obtain spatial constraint trajectory data; classifying and arranging the historical trajectories of multiple miners in the spatial constraint trajectory data, grouping them according to work teams, work nature and activity areas, and obtaining a group trajectory set; applying density clustering processing based on computer vision to the group trajectory set to separate trajectories that are similar in time and space The trace points are clustered into a group, and the center point and activity range of each group are extracted to obtain a behavior density distribution map. Based on the behavior density distribution map, the activity patterns of miners in different areas are identified, including the stay pattern at the work point, the inspection route pattern, the material transportation pattern, and the emergency evacuation pattern, to obtain behavior pattern data. The spatial constraint trajectory data is matched with the behavior pattern data, and the direction and rhythm of the predicted trajectory are adjusted according to the current miner position and task type to make it conform to the characteristics of the corresponding behavior pattern, to obtain the behavior optimization trajectory. Based on the collaborative relationship between miners in the same work group, the behavior optimization trajectory is adjusted for group consistency to ensure the spatiotemporal coordination of the predicted trajectory of the collaborative miners, to obtain the target trajectory prediction data. A generation module is used to calculate the spatial safety distance, time conflict probability and regional load index based on the target trajectory prediction data, and generate a mine personnel safety scheduling strategy.

8. An artificial intelligence-based mine personnel trajectory prediction device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the mine personnel trajectory prediction method based on artificial intelligence as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to execute the mine personnel trajectory prediction method based on artificial intelligence according to any one of claims 1 to 6.

Images