A large-scale efficient reasoning method for a coal mine visual model

Abstract

This application relates to the field of artificial intelligence and data processing technology, and discloses a large-scale and efficient inference method for a coal mine visual model. The method includes: synchronously acquiring real-time video streams from the coal mine, physical environment, and business status parameters; calculating a comprehensive environmental distortion impact factor; constructing a feature weight matrix using the factors to enhance visual features through channel-dimensional weighting; calculating scene priorities and predictive inference load based on environmental factors and business parameters; dynamically allocating batch processing size and graphics processor instances; correcting time decay logic using actual inference frame intervals; and adjusting alarm time windows based on environment and priority to determine risk levels; performing coordinate reverse mapping and sending equipment control commands. This invention, by integrating environmental perception and dynamic scheduling of business priorities, achieves adaptive feature enhancement and on-demand allocation of computing power, helping to solve delays and false alarms caused by resource contention and improving the accuracy of coal mine monitoring.

Claims

1. A large-scale, efficient reasoning method for a visual model of a coal mine, characterized in that, Includes the following steps: S1. Collect coal mine video stream, physical environment parameters and business status parameters to calculate the comprehensive environmental distortion impact factor, use the comprehensive environmental distortion impact factor to enhance the visual model features, and output the scene-based adaptation inference results. S2. Calculate the scenario priority index and predictive inference load by combining the comprehensive environmental distortion impact factor and the business status parameters, and dynamically schedule computing resources and batch processing strategies accordingly, and monitor the actual inference frame interval. S3. Receive the scenario-based adaptation inference result, combine the actual inference frame interval, the comprehensive environmental distortion impact factor and the scenario priority index, dynamically modify the alarm confirmation logic and risk judgment logic, calculate the risk cumulative confidence and determine the risk level. S4. Reconstruct the spatial coordinates of the targets corresponding to the risk level; Based on the risk level, standardized alarm commands are generated, and equipment control signals are sent to the field controller via standard industrial protocols to execute the linkage control of the coal mine field equipment.

2. The large-scale and efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The calculation of the comprehensive environmental distortion impact factor in step S1 specifically includes: The real-time dust concentration, ambient light intensity, and equipment vibration amplitude are obtained from the physical environment parameters. The sharpness index attenuation term caused by the real-time dust concentration is calculated based on the dust optical transmission attenuation coefficient. Based on the optimal inference baseline illumination threshold of the visual model, the image contrast loss term caused by the ambient illumination intensity is calculated; Based on the vibration impact threshold that causes image motion blur, the motion blur effect term caused by the vibration amplitude of the device is calculated; Based on the normalized weight coefficients configured for the current monitoring scenario, the sharpness index attenuation term, the image contrast loss term, and the motion blur effect term are weighted and summed to obtain the dimensionless comprehensive environmental distortion influence factor.

3. The large-scale, efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The enhancement of visual model features using the comprehensive environmental distortion influence factor in step S1 specifically includes: Perform multi-dimensional data synchronous acquisition for coal mine visual models; In the forward inference process of the coal mine visual model, the original feature map output by the backbone network is extracted; the feature weight matrix is ​​constructed using the comprehensive environmental distortion influence factor to enhance the original feature map by channel dimension weighting; specifically, a feature channel importance diagonal matrix is ​​constructed, where the diagonal elements of the feature channel importance diagonal matrix are the sensitivity weights of the feature channels to the preset business objectives; A dynamic gain is generated using the comprehensive environmental distortion impact factor, and the dynamic gain is applied to the feature channel importance diagonal matrix, so that the sensitivity weight increases with the increase of the comprehensive environmental distortion impact factor; The feature channel importance diagonal matrix containing the dynamic gain is multiplied with the original feature map to generate enhanced features.

4. The large-scale, efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The calculation of the scenario priority index in step S2 specifically includes: Extract the scenario risk level quantification value, real-time number of operators, current operation duration, and equipment importance from the business status parameters; Calculate the product of the number of workers in real time and the duration of the current operation, and determine the ratio of the product to the maximum workload of a standard shift as the operation density term; The scenario risk level quantification value, the operation intensity item, and the equipment importance are linearly weighted and summed to generate a normalized scenario priority index.

5. The large-scale, efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, Step S2, which combines the comprehensive environmental distortion impact factor and the business status parameters to determine scenario priority and predicted load, and dynamically schedules computing resources and batch processing strategies, specifically includes: The comprehensive environmental distortion impact factor is introduced as a computing power compensation term in the inference load prediction model; The average load and current concurrent request volume within the historical sliding window are obtained. The average load is adjusted by the comprehensive environmental distortion impact factor to compensate for the computing power consumption of feature enhancement. The average load is adjusted by the scene priority index to reserve high priority computing power. The logarithmic term of the current concurrent request volume is added to obtain the prediction inference load. Based on the scenario priority index and the prediction inference load, the batch size and the number of graphics processor instances for the inference task are dynamically adjusted; specifically: The scene priority index is compared with a preset priority threshold: if the scene priority index is not less than the priority threshold, the batch size is forcibly set to a preset high-priority batch size, and a dedicated graphics processor instance is allocated; if the scene priority index is less than the priority threshold, the batch size is increased according to the predictive inference load, and the number of graphics processor instances occupied is reduced.

6. The large-scale, efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The dynamic modification of the alarm confirmation logic in step S3 specifically includes dynamically adjusting the length of the alarm confirmation time window: Set the base time window length; The length of the alarm confirmation time window is obtained by using the comprehensive environmental distortion influence factor as a positive adjustment parameter and the scene priority index as a negative adjustment parameter to perform a weighted correction calculation on the length of the basic time window.

7. The large-scale, efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The S3 step of dynamically correcting the risk assessment logic and calculating the cumulative risk confidence level specifically includes: Obtain the target risk value from the inference results of multiple consecutive frames; The actual inference frame interval is introduced to correct the exponential term of the time decay function, specifically as follows: The physical time difference is obtained by multiplying the backtracked frame index offset by the actual inference frame interval. The physical time difference is then subjected to exponential decay operation using a preset time decay constant to obtain the time decay weight corresponding to each frame. The cumulative confidence level of the risk is obtained by calculating the weighted sum of the target risk value and the time decay weight.

8. The large-scale and efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The spatial coordinate reconstruction of the target corresponding to the risk level in step S4 specifically includes: Based on the ratio between the original video stream resolution and the model input resolution, reverse mapping reconstruction of the target space coordinates is performed, specifically: obtaining the width and height resolutions of the original video stream, as well as the preset width and height resolutions of the visual model input image; Calculate the scaling ratio of the original video stream and the visual model input image in the horizontal and vertical dimensions; Using the scaling ratio, the normalized target center point coordinates and target box size output by the visual model are linearly mapped back to the original video stream coordinate system, and the mapped pixel coordinates are rounded.

9. The large-scale and efficient reasoning method for a coal mine visual model according to claim 1, characterized in that, The specific steps of executing the linkage control of the field equipment in step S4 include: When the risk level is determined to be high, a high-level signal is written to a designated register of the field programmable logic controller via the industrial communication protocol to trigger the audible and visual alarm. When the risk level is determined to be safe, a low-level signal is written to the designated register to perform a reset. In response to emergency alarms involving equipment safety, a shutdown signal is sent to the equipment control loop.

10. A computer device, characterized in that, It includes a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected via the bus; the memory stores a computer program, and when the processor executes the computer program, it implements a large-scale and efficient reasoning method for a coal mine visual model as described in any one of claims 1-9.

Images