Coal mine tunneling sensing test method and system based on multi-sensor combination configuration

Abstract

The application discloses a coal mine tunneling sensing test method and system based on a multi-sensor combination configuration, relates to the technical field of data processing, and comprises the following steps: constructing a test scene parameter, performing real-time sensing on a multi-sensor combination configured for a coal mine equipment, obtaining multi-modal sensing data for multi-source heterogeneous data synchronous denoising analysis, and generating a standard test data set; based on the standard test data set, a coal mine tunneling environment dynamic sensing graph is constructed; multi-directional sensing test of the multi-sensor combination is performed according to a sensing feature tensor, a multi-directional sensing test parameter set is calculated, the multi-directional sensing test parameter set is fed back to the multi-sensor combination for configuration adjustment, and a coal mine tunneling sensing test report is generated. The application solves the technical problem that in the prior art, the multi-sensor combination configuration under the coal mine tunneling environment cannot be adaptively optimized according to the sensing test result, and achieves the technical effect of improving the sensing test accuracy and configuration adaptability of the multi-sensor combination to the coal mine tunneling environment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a coal mine tunneling sensing and testing method and system based on a multi-sensor combination configuration. Background Technology

[0002] Coal mine tunneling operations are characterized by confined spaces, high dust concentrations, large humidity variations, strong vibrations and shocks, and frequent obstructions and interference. A single sensor is insufficient to reliably meet the testing requirements of tunneling equipment for target recognition, pose sensing, motion tracking, and environmental reconstruction. To improve sensing reliability, multiple sensors, such as lidar, millimeter-wave radar, vision sensors, and inertial sensors, are typically combined on coal mine equipment. However, the sensing performance of different sensors varies significantly under different dust, humidity, vibration, and obstruction conditions. Sensor combination parameters often rely on preset experience or static debugging, and there is a lack of effective feedback between test results and sensor configuration adjustments. This makes it difficult for multi-sensor combinations to adaptively optimize themselves in response to changes in the tunneling environment and sensing test results. Summary of the Invention

[0003] This application provides a coal mine tunneling perception testing method and system based on multi-sensor combination configuration, which is used to address the technical problem that in the prior art, it is difficult to adaptively optimize the multi-sensor combination configuration in the coal mine tunneling environment based on the perception test results.

[0004] In view of the above problems, this application provides a coal mine tunneling sensing test method and system based on multi-sensor combination configuration.

[0005] The first aspect of this application provides a coal mine tunneling sensing test method based on a multi-sensor combination configuration, the method comprising: Test scenario parameters are constructed, and multi-sensor combinations are configured on coal mine equipment for real-time sensing. Multimodal sensing data is obtained, and multi-source heterogeneous data synchronous denoising analysis is performed to generate a standard test dataset. Based on the standard test dataset, multi-level feature fusion is performed to construct a dynamic perception map of the coal mine tunneling environment, which includes a perception feature tensor. Multi-directional perception tests of the multi-sensor combination are performed according to the perception feature tensor, and a multi-directional perception test parameter set is calculated. The multi-directional perception test parameter set is fed back to the multi-sensor combination for configuration adjustment, and a coal mine tunneling perception test report is generated.

[0006] A second aspect of this application provides a coal mine tunneling sensing and testing system based on a multi-sensor combination configuration, the system comprising: The real-time sensing module is used to construct test scenario parameters, configure multiple sensor combinations for real-time sensing of coal mine equipment, obtain multi-modal sensing data, perform synchronous denoising analysis of multi-source heterogeneous data, and generate a standard test dataset. The perception map construction module is used to perform multi-level feature fusion based on the standard test dataset to construct a dynamic perception map of the coal mine tunneling environment, which includes a perception feature tensor. The configuration adjustment module is used to perform multi-directional perception tests of the multi-sensor combination based on the perception feature tensor, calculate the multi-directional perception test parameter set, feed the multi-directional perception test parameter set back to the multi-sensor combination for configuration adjustment, and generate a coal mine tunneling perception test report.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application constructs test scenario parameters, performs real-time sensing on a multi-sensor combination of coal mine equipment, obtains multi-modal sensing data, performs synchronous denoising analysis on multi-source heterogeneous data, and generates a standard test dataset. Based on the standard test dataset, multi-level feature fusion is performed to construct a dynamic perception map of the coal mine tunneling environment, which includes a perception feature tensor. Multi-directional perception tests of the multi-sensor combination are conducted based on the perception feature tensor, a multi-directional perception test parameter set is calculated, and the multi-directional perception test parameter set is fed back to the multi-sensor combination for configuration adjustment, generating a coal mine tunneling perception test report. This invention solves the technical problem in the prior art where the configuration of multi-sensor combinations in a coal mine tunneling environment is difficult to adaptively optimize based on perception test results. By constructing a standard test dataset, fusing and generating a dynamic perception map of the coal mine tunneling environment, and performing multi-directional perception tests and feedback configuration adjustments based on the perception feature tensor, the technical effect of improving the accuracy and configuration adaptability of multi-sensor combinations in perceiving and testing the coal mine tunneling environment is achieved. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic diagram of the coal mine tunneling sensing and testing method based on multi-sensor combination configuration provided in an embodiment of this application; Figure 2 Time history curve of tunneling vibration provided in the embodiment of this application for a coal mine tunneling sensing and testing method based on multi-sensor combination configuration; Figure 3This is a schematic diagram of the structure of a coal mine tunneling sensing and testing system based on a multi-sensor combination configuration, provided in an embodiment of this application.

[0010] Figure labeling: Real-time sensing module 11, perception map construction module 12, configuration adjustment module 13. Detailed Implementation

[0011] This application provides a coal mine tunneling perception testing method and system based on multi-sensor combination configuration. It addresses the technical problem in the prior art that it is difficult to adaptively optimize the multi-sensor combination configuration in the coal mine tunneling environment based on perception test results. By constructing a standard test dataset, fusing and generating a dynamic perception map of the coal mine tunneling environment, and performing multi-directional perception testing and feedback configuration adjustment based on perception feature tensors, the technical effect of improving the accuracy and configuration adaptability of multi-sensor combination perception testing of the coal mine tunneling environment is achieved.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides a coal mine tunneling sensing and testing method based on a multi-sensor combination configuration, the method comprising: Step S100: Construct test scenario parameters, configure multi-sensor combinations for real-time sensing of coal mine equipment, obtain multi-modal sensing data, perform multi-source heterogeneous data synchronous denoising analysis, and generate a standard test dataset.

[0015] In this embodiment, when constructing test scenario parameters, the coal mine tunneling task is first analyzed to obtain a set of tunneling analysis parameters including three-dimensional geometric parameters, ventilation parameters, cutting head working condition parameters, and coal and rock property parameters. Then, dust diffusion analysis and settlement simulation analysis are performed based on the three-dimensional geometric parameters, ventilation parameters, and cutting head working condition parameters to construct a dust concentration distribution field and a humidity simulation distribution field. Furthermore, the coal falling impact response analysis is performed by combining the cutting head working condition parameters and coal and rock property parameters to generate a tunneling vibration time history curve. The tunneling vibration time history curve is then dynamically coupled and mapped with the dust concentration distribution field to obtain a dust re-entrainment coupling compensation field. Finally, the dust concentration distribution field, humidity simulation distribution field, tunneling vibration time history curve, and dust re-entrainment coupling compensation field are spatiotemporally aligned and encapsulated to form the test scenario parameters.

[0016] After constructing the test scenario parameters, the coal mine equipment and the multi-sensor combination are configured and associated. Based on the test scenario parameters, the acquisition time, acquisition frequency, data format, and acquisition range of each sensor in the multi-sensor combination are determined. Subsequently, the multi-sensor combination is controlled to perform real-time sensing during the coal mine tunneling test, ensuring that each sensor continuously outputs raw sensing data according to its corresponding acquisition frequency. The acquisition time, sensor source, data format, and data content are recorded in each raw sensing data entry, thus obtaining multimodal sensing data containing different sensor sources, data formats, and physical meanings. After obtaining the multimodal sensing data, multi-source heterogeneous data synchronization and denoising analysis is performed. First, the data from different sensor sources are sorted according to the acquisition time, and a time correspondence between data from different sensor sources is established based on the acquisition time. When the acquisition frequencies of different sensors are inconsistent, the acquisition time of the data to be synchronized is compared with the acquisition times of other sensor data. The data with the smallest acquisition time difference and within the same acquisition period is selected as the corresponding data, thereby completing time synchronization. After time synchronization is completed, the data output from each sensor is uniformly organized according to the data format, converting data of different formats into a unified record format that includes acquisition time, sensor source, data format, and data value. Then, the data in the unified record format is identified and corrected: missing data is identified through acquisition time and sensor source. If a sensor source fails to generate a data record at an acquisition time when data should exist, the data for that acquisition time is determined to be missing data. If valid data from the same sensor source exists at both the preceding and following acquisition times, it is supplemented according to the proportion of time that acquisition time occupies between the preceding and following acquisition times. That is, the valid data value of the preceding acquisition time is used as the starting value, and the valid data value of the following acquisition time is used as the ending value, calculating the data value corresponding to the acquisition time to be supplemented according to the acquisition time interval. Duplicate data is identified by whether there are more than two data records under the same acquisition time, same sensor source, and same data format. When the content of duplicate data is the same, one data record is retained and the rest are deleted. When the content of duplicate data is inconsistent, each data record is compared with the data value range defined by the test scenario parameters and the data change relationship between adjacent acquisition times. Data records that are within the data value range and have continuous data changes with the acquisition time before and after are retained.Abnormal data mutations are identified by the magnitude of change between the data value at the current acquisition time and the data values ​​at the previous and subsequent acquisition times. When the data value at the current acquisition time deviates from the allowable range relative to both the previous and subsequent acquisition times, and the data from the previous acquisition time to the subsequent acquisition time does not show a continuous change in the same direction, the data at the current acquisition time is determined to be abnormal data mutation data. When correcting abnormal data mutations, the valid data values ​​at the previous and subsequent acquisition times are first obtained. Then, the time difference between the current acquisition time and the previous acquisition time, and the time difference between the subsequent acquisition time and the previous acquisition time are calculated. The ratio of these two is used as the correction ratio. Finally, the corrected data value at the current acquisition time is obtained by adding the product of the correction ratio and the difference between the valid data values ​​at the previous acquisition time to the valid data value at the previous acquisition time. When the data content contains multiple data values, the above correction process is performed separately for each data value. Noise data is identified by whether the data value exceeds the data range defined by the test scenario parameters and whether it deviates from the data change relationship between adjacent acquisition times. When the data value exceeds the data range defined by the test scenario parameters and the corrected data value cannot be determined by the valid data values ​​before and after, the data is discarded. When the data value does not exceed the data range defined by the test scenario parameters but deviates from the data change relationship between adjacent acquisition times, it is corrected according to the correction process for abnormal mutation data. After completing time synchronization, format unification, missing data filling, duplicate data deletion, abnormal mutation data correction, and noise data removal or correction, the processed multimodal sensor data is associated and encapsulated according to test scenario parameters, acquisition time, sensor source, data format, and data content to generate a standard test dataset.

[0017] Furthermore, in the method provided in the application embodiments, the process of constructing test scenario parameters further includes: The process involves retrieving coal mine tunneling tasks for analysis to obtain a set of tunneling analysis parameters, including three-dimensional geometric parameters, ventilation parameters, cutting head operating parameters, and coal and rock property parameters. Based on these parameters, dust diffusion analysis is performed to construct a dust concentration distribution field. Settlement simulation analysis is also conducted to construct a humidity simulation distribution field. Furthermore, coal impact response analysis is performed based on the cutting head operating parameters and coal and rock property parameters to construct a tunneling vibration time history curve. This time history curve is then dynamically coupled and mapped with the dust concentration distribution field to generate a dust re-entrainment coupling compensation field. Finally, the dust concentration distribution field, the humidity simulation distribution field, the tunneling vibration time history curve, and the dust re-entrainment coupling compensation field are spatiotemporally aligned and encapsulated to construct test scenario parameters.

[0018] In this embodiment, when retrieving coal mine tunneling tasks for tunneling analysis, the system reads the roadway cross-sectional dimensions, roadway extension direction, tunneling distance, tunneling face location, ventilation layout data, cutting head operation data, and coal and rock occurrence data recorded in the coal mine tunneling task. The parameters are then categorized according to data attributes to obtain a tunneling analysis parameter set. This parameter set includes three-dimensional geometric parameters, ventilation parameters, cutting head operating parameters, and coal and rock property parameters. The three-dimensional geometric parameters include roadway length, roadway width, roadway height, roadway cross-sectional shape, tunneling face coordinates, and coal mine equipment coordinates. The ventilation parameters include airflow direction, wind speed, air volume, and ventilation duct outlet location. The cutting head operating parameters include cutting head rotation speed, cutting trajectory, cutting depth, advance speed, and cutting load. The coal and rock property parameters include coal and rock density, coal and rock hardness, coal and rock particle size, coal and rock moisture content, and coal and rock structural strength. During the tunneling analysis, the tunnel length direction is used as the longitudinal coordinate, the tunnel width direction as the transverse coordinate, and the tunnel height direction as the vertical coordinate. The coordinates of the tunnel face, the coal mine equipment, the ventilation duct outlet position, and the cutting head trajectory coordinates are written into the same three-dimensional coordinate system. Subsequently, the position, advance distance, and cutting status of the cutting head at each calculation time are read according to the calculation time interval, so that the three-dimensional geometric parameters, ventilation parameters, cutting head working condition parameters, and coal and rock property parameters form a corresponding relationship under the same calculation time and the same three-dimensional coordinate system.

[0019] Dust diffusion analysis is conducted based on three-dimensional geometric parameters, ventilation parameters, and cutting head operating parameters. When constructing the dust concentration distribution field, multiple three-dimensional coordinate positions are first determined within the roadway space defined by the three-dimensional geometric parameters according to the longitudinal, transverse, and vertical coordinates, and continuous calculation time is determined according to the calculation time interval. At any calculation time, the dust generation location is determined according to the cutting head trajectory and the coordinates of the tunnel face. The advancing speed is multiplied by the calculation time interval to calculate the advancing distance. The advancing distance, cutting depth, and cutting head effective width are multiplied to calculate the cut coal and rock volume. Finally, the cut coal and rock volume is multiplied by the coal and rock density to calculate the cut coal and rock mass. The dust generation ratio is determined using a benchmark ratio correction method: First, the basic dust generation ratio, benchmark cutting load, and benchmark coal and rock hardness are set. The cutting load is divided by the benchmark cutting load to obtain the load ratio, and the coal and rock hardness is divided by the benchmark coal and rock hardness to obtain the hardness ratio. When the load ratio is greater than 1, the load correction factor is taken as the load ratio; when the load ratio is less than or equal to 1, the load correction factor is taken as 1. When the hardness ratio is greater than 1, the hardness correction factor is taken as the hardness ratio; when the hardness ratio is less than or equal to 1, the hardness correction factor is taken as 1. The basic dust generation ratio, load correction factor, and hardness correction factor are multiplied together to obtain the dust generation ratio. Then, the mass of the cut coal and rock is multiplied by the dust generation ratio to obtain the dust source strength at that calculation moment, and the dust source strength is written into the three-dimensional coordinate position of the cutting head. Next, the dust migration distance is calculated according to the wind direction and wind speed. The dust migration distance is the product of the wind speed and the calculation time interval. For any three-dimensional coordinate position, the dust concentration value at the previous calculation time is read. The dust source strength is divided by the spatial volume at the corresponding position to obtain the dust concentration increment. The dust concentration that migrates into the three-dimensional coordinate position upwind is counted as the influx dust concentration. The dust concentration that migrates out along the wind direction at the current position is counted as the outflux dust concentration. The current dust concentration value, the settling ratio, and the calculation time interval are multiplied to obtain the dust concentration that decreases due to settling. Finally, the dust concentration value at the previous calculation time is added to the dust concentration increment and the influx dust concentration, and then the outflux dust concentration and the dust concentration that decreases due to settling are subtracted to obtain the dust concentration value at the current calculation time. This value is stored according to the calculation time and the corresponding three-dimensional coordinate position to construct a dust concentration distribution field.

[0020] Settlement simulation analysis was conducted based on three-dimensional geometric parameters, ventilation parameters, and cutting head operating parameters. When constructing the humidity simulation distribution field, the tunnel space, three-dimensional coordinate positions, and continuous calculation time were defined by the three-dimensional geometric parameters. The humidity value at each three-dimensional coordinate position at the previous calculation time was read, and the humidity migration process was calculated according to the airflow direction, wind speed, and air volume in the ventilation parameters. Among them, the humidity migration distance is the product of wind speed and calculation time interval. The amount of humidity migrating in is calculated from the humidity value and air volume at the upwind three-dimensional coordinate position, and the amount of humidity migrating out is calculated from the humidity value and air volume at the current three-dimensional coordinate position. Subsequently, the humidity increment in the cutting area is calculated based on the cutting head trajectory, cutting depth, advancing speed, and coal / rock moisture content. Specifically, the advancing speed is multiplied by the calculation time interval to obtain the advancing distance; the advancing distance, cutting depth, and cutting head width are multiplied to obtain the cut coal / rock volume; the cut coal / rock volume is multiplied by the coal / rock density to obtain the cut coal / rock mass; and the cut coal / rock mass is multiplied by the coal / rock moisture content to obtain the cut-released water content. This cut-released water content is then converted into the humidity increment at the corresponding three-dimensional coordinate position of the cutting area. Next, humidity settlement is calculated. Humidity settlement is the product of the humidity value, settlement ratio, and calculation time interval at the previous calculation moment. For any three-dimensional coordinate position, the humidity value at the previous calculation moment is added to the influx of humidity and the humidity increment in the cutting area, and then subtracted from the outflux of humidity and humidity settlement to obtain the humidity value at the current calculation moment. This value is stored according to the calculation moment and three-dimensional coordinate position to construct a humidity simulation distribution field.

[0021] Based on the cutting head operating parameters and coal and rock property parameters, the coal falling impact response analysis is performed. When constructing the tunneling vibration time history curve, the coal falling location and volume within continuous calculation time are determined according to the cutting head rotation speed, cutting trajectory, cutting depth, advance speed, and cutting load. At any calculation time, the advance speed is multiplied by the calculation time interval to obtain the advance distance, the advance distance, cutting depth, and cutting head action width are multiplied to obtain the coal falling volume, and the coal falling volume is multiplied by the coal and rock density to obtain the coal falling mass. Subsequently, the coal falling speed is calculated. First, the basic coal falling speed is calculated based on the height difference between the coal falling location and the coal falling impact location. Then, a coal falling speed correction coefficient is set according to the coal and rock hardness, coal and rock particle size, coal and rock moisture content, and coal and rock structural strength. The coal falling speed is obtained by multiplying the basic coal falling speed by the coal falling speed correction coefficient. Among them, the coal falling speed correction coefficient decreases when the coal and rock hardness, coal and rock moisture content, and coal and rock structural strength increase when the coal and rock particle size increases. After calculating the coal falling mass and velocity, the impact momentum is obtained by multiplying the coal falling mass and velocity, and then the impact force is obtained by dividing the impact momentum by the impact duration. Subsequently, the vibration attenuation is calculated based on the distance between the coal falling impact location and the vibration response location, and the impact force is subtracted from the vibration attenuation to convert it into the vibration response value at that calculation time. If multiple coal falling locations exist at the same calculation time, the vibration response value corresponding to each location is calculated separately, and the multiple vibration response values ​​are summed to form the total vibration response value at that calculation time. If there is residual vibration from a previous calculation time, the vibration response value from the previous calculation time is multiplied by the vibration attenuation coefficient to obtain the residual vibration response value, and this residual vibration response value is added to the total vibration response value at the current calculation time to obtain the vibration response value at the current calculation time. Finally, the vibration response values ​​at each calculation time are arranged in chronological order to construct the tunneling vibration time history curve. Figure 2 As shown, the tunneling vibration time history curve uses the calculation time as the horizontal axis and the vibration response value as the vertical axis. During plotting, the vibration response values ​​corresponding to each calculation time are first read in ascending order. Then, each calculation time is used as the horizontal axis point, and the vibration response value at that calculation time is used as the vertical axis point. Adjacent coordinate points are then connected in the order of the calculation times to form the tunneling vibration time history curve. Figure 2 In the curve, each coordinate point corresponds to a vibration response value at a calculation time. When there is a large coal falling mass, coal falling speed, or multiple coal falling locations at a certain calculation time, the vibration response value corresponding to that calculation time increases and is represented as a peak value in the curve. When the vibration response value of the next calculation time is obtained by adding the residual vibration response value of the previous calculation time and the total vibration response value of the current calculation time, the coordinate point corresponding to the next calculation time is located at the vertical coordinate position determined by the residual vibration response value and the current total vibration response value. Figure 2The horizontal axis scale corresponds to each calculation time, and the vertical axis scale corresponds to the vibration response value. The curve records the vibration response value at each calculation time in the order of calculation time.

[0022] Next, the tunneling vibration time history curve and the dust concentration distribution field are dynamically coupled and mapped. First, the tunneling vibration time history curve is locally calculated to obtain the local floor acceleration excitation parameters characterizing the local vibration excitation state of the floor. Then, the local floor acceleration excitation parameters are correlated with the dust concentration distribution field to set a critical stripping acceleration threshold for judging whether the deposited dust has been stripped and re-erected. Subsequently, the local floor acceleration is calculated and compared with the critical stripping acceleration threshold. When the local floor acceleration is greater than the critical stripping acceleration threshold, the dust re-erecting source term is activated. Then, the local dust deposition concentration value is calculated based on the dust concentration distribution field, and the intensity value of the dust re-erecting source term is made proportional to the vibration acceleration amplitude and the local dust deposition concentration value. In this way, the dust re-erecting effect caused by vibration excitation is compensated into the dust concentration distribution field, and a dust re-erecting coupled compensation field is constructed.

[0023] Finally, when encapsulating the dust concentration distribution field, humidity simulation distribution field, tunneling vibration time history curve, and dust re-entrainment coupling compensation field in a spatiotemporal alignment, a unified spatial coordinate system is first established using the longitudinal, transverse, and vertical coordinates in the three-dimensional geometric parameters, with the calculation time interval of the dust concentration distribution field serving as the unified time interval. Then, according to the unified spatial coordinates and unified time interval, the dust concentration values ​​in the dust concentration distribution field, the humidity values ​​in the humidity simulation distribution field, and the compensated dust concentration values ​​in the dust re-entrainment coupling compensation field are read, along with the vibration response values ​​of the tunneling vibration time history curve at the corresponding calculation time. When the dust concentration distribution field, humidity simulation distribution field, and dust re-entrainment coupling compensation field have the same spatial coordinates and the same calculation time, the corresponding dust concentration values, humidity values, and compensated dust concentration values ​​are directly written into the data entries under the same spatial coordinates and the same calculation time. When the calculation time of any distribution field is not consistent with the unified time interval, the calculation time is adjusted accordingly. When the time intervals are completely consistent, select two adjacent calculated times before and after the current calculation time, and perform linear conversion on the corresponding values ​​according to the time ratio between the current calculation time and the two adjacent calculated times before and after, and then write them into the unified calculation time. For the tunneling vibration time history curve, extract the vibration response value according to the calculation time, and establish a correlation between the vibration response value and the spatial coordinates under the same calculation time. When there is a vibration response position in the tunneling vibration time history curve, write the vibration response value into the spatial coordinates of the corresponding vibration response position. When there is no vibration response position, write the vibration response value into the spatial coordinates corresponding to the tunneling face and the bottom plate according to the same calculation time. After completing the above correspondence, encapsulate the dust concentration value, humidity value, vibration response value and compensated dust concentration value under the same spatial coordinates and the same calculation time together with the three-dimensional geometric parameters, ventilation parameters, cutting head working condition parameters and coal and rock property parameters to construct the test scene parameters.

[0024] Furthermore, in the method provided in the application embodiment, the dynamic coupling mapping of the tunneling vibration time history curve and the dust concentration distribution field to generate a dust re-entrainment coupling compensation field further includes: The vibration time history curve is locally calculated to obtain local base plate acceleration excitation parameters; based on the local base plate acceleration excitation parameters and the dust concentration distribution field, a critical peeling acceleration threshold is set for vibration excitation; the local base plate acceleration is calculated, and when the local base plate acceleration is greater than the critical peeling acceleration threshold, the dust re-eruption source term is activated; based on the dust concentration distribution field, the local dust deposition concentration value is calculated, and when the intensity value of the dust re-eruption source term is proportional to the vibration acceleration amplitude and the local dust deposition concentration value, the dust re-eruption coupling compensation field is constructed.

[0025] In this embodiment, when performing local calculations on the vibration time history curve, the vibration response value in the vibration time history curve is read according to the calculation time, and the vibration response value is correlated with the spatial position of the same calculation time in the dust concentration distribution field. If the vibration response value is the base plate displacement value, the base plate displacement values ​​of the previous calculation time, the current calculation time, and the next calculation time are read. The base plate displacement value of the next calculation time is subtracted by twice the base plate displacement value of the current calculation time, and then added to the base plate displacement value of the previous calculation time. This is then divided by the square of the calculation time interval to obtain the local base plate acceleration excitation parameter at the current calculation time. If the vibration response value is the base plate velocity value, the difference between the base plate velocity value of the current calculation time and the base plate velocity value of the previous calculation time is calculated and divided by the calculation time interval to obtain the local base plate acceleration excitation parameter at the current calculation time. If the vibration response value is the base plate acceleration value, the base plate acceleration value is used as the local base plate acceleration excitation parameter.

[0026] When setting the critical peeling acceleration threshold based on the local base plate acceleration excitation parameters combined with the dust concentration distribution field for vibration excitation, the dust concentration value in the dust concentration distribution field at the same spatial location and the same calculation time is read, and the preset reference peeling acceleration, concentration correction coefficient, and minimum peeling acceleration are called; the dust concentration value is multiplied by the concentration correction coefficient to obtain the threshold correction amount; the reference peeling acceleration is subtracted from the threshold correction amount to obtain the initial peeling threshold; when the initial peeling threshold is greater than or equal to the minimum peeling acceleration, the initial peeling threshold is used as the critical peeling acceleration threshold; when the initial peeling threshold is less than the minimum peeling acceleration, the minimum peeling acceleration is used as the critical peeling acceleration threshold.

[0027] When calculating the local floor acceleration, the local floor acceleration excitation parameters at the current calculation moment are read, and their absolute values ​​are taken as the local floor acceleration. The local floor acceleration is then compared with the critical peeling acceleration threshold at the same spatial location and the same calculation moment. When the local floor acceleration is less than or equal to the critical peeling acceleration threshold, the intensity value of the dust re-emission source term at the corresponding calculation moment is recorded as zero. When the local floor acceleration is greater than the critical peeling acceleration threshold, the dust re-emission source term is activated, and the difference between the local floor acceleration and the critical peeling acceleration threshold is taken as the vibration acceleration amplitude.

[0028] When calculating the local dust deposition concentration value based on the dust concentration distribution field, the dust concentration value, settling ratio, and calculation time interval at the same spatial location at the previous calculation time are read. The dust concentration value, settling ratio, and calculation time interval at the previous calculation time are multiplied together to obtain the deposition concentration increment corresponding to the current calculation time. The deposition concentration increment is added to the dust deposition concentration value retained at the previous calculation time, and the dust concentration value that has been re-erected at the previous calculation time is subtracted to obtain the local dust deposition concentration value at the current calculation time. After the dust re-eruption source term is activated, the vibration acceleration amplitude, the local dust deposition concentration, and the re-eruption ratio coefficient are multiplied to obtain the intensity value of the dust re-eruption source term, making the intensity value of the dust re-eruption source term proportional to the vibration acceleration amplitude and the local dust deposition concentration. Then, the intensity value of the dust re-eruption source term is superimposed on the dust concentration value at the same spatial location and the same calculation time in the dust concentration distribution field to obtain the compensated dust concentration value. The compensated dust concentration values ​​are then arranged according to spatial location and calculation time to construct the dust re-eruption coupled compensation field.

[0029] Step S200: Perform multi-level feature fusion based on the standard test dataset to construct a dynamic perception map of the coal mine tunneling environment, wherein the dynamic perception map of the coal mine tunneling environment contains a perception feature tensor.

[0030] In this embodiment, when performing multi-level feature fusion based on a standard test dataset, the multimodal sensing data in the standard test dataset is first parsed to obtain a denoised point cloud parameter set and a millimeter-wave radar detection vector. Then, local geometric features and intensity distribution features are extracted based on the denoised point cloud parameter set, and electromagnetic scattering features are generated based on the millimeter-wave radar detection vector. Subsequently, the local geometric features, intensity distribution features, and electromagnetic scattering features are fused to form a spatiotemporal perception feature stream, and the spatiotemporal perception feature stream is implicitly decoded through a conditional neural implicit field decoder to construct an implicit dynamic perception map. Finally, the perception feature tensor is calculated based on the conditional neural implicit field decoder, and the perception feature tensor is combined with the implicit dynamic perception map to construct a dynamic perception map of the coal mine tunneling environment containing the perception feature tensor.

[0031] Furthermore, in the method provided in the application embodiments, the construction of a dynamic perception map of the coal mine tunneling environment based on the standard test dataset through multi-level feature fusion also includes: Based on the standard test dataset, multidimensional analysis is performed to obtain a denoised point cloud parameter set and a millimeter-wave radar detection vector. Multi-scale dynamic voxel partitioning is then performed based on the denoised point cloud parameter set to extract local geometric features and intensity distribution features. Electromagnetic scattering analysis is performed based on the millimeter-wave radar detection vector to generate electromagnetic scattering features. Heterogeneous feature fusion is performed based on the local geometric features, intensity distribution features, and electromagnetic scattering features to construct a spatiotemporal perception feature stream. A conditional neural implicit field decoder is constructed to implicitly decode the spatiotemporal perception feature stream, constructing an implicit dynamic perception map. Perceptual analysis is performed on the standard test dataset based on the conditional neural implicit field decoder to calculate the perception feature tensor. The perception feature tensor is combined with the implicit dynamic perception map to construct the dynamic perception map of the coal mine tunneling environment.

[0032] In this embodiment of the application, when performing multidimensional analysis based on the standard test dataset, the data is first classified and read according to the sensor source and data format corresponding to each data record in the standard test dataset. For point cloud data, the acquisition time, spatial coordinates, and point cloud intensity are extracted, and data records that have been identified as abnormal or noisy during the synchronous denoising analysis are deleted. Then, the retained point cloud data are arranged according to the acquisition time and spatial coordinates to obtain a denoised point cloud parameter set. For millimeter-wave radar data, the acquisition time, detection range, azimuth angle, radial velocity, and echo intensity are extracted, and the detection range, azimuth angle, radial velocity, and echo intensity under the same acquisition time are arranged according to the preset field order to obtain the millimeter-wave radar detection vector.

[0033] Next, multi-scale dynamic voxel partitioning is performed based on the denoised point cloud parameter set. When extracting local geometric features and intensity distribution features, the point cloud coordinates, point cloud intensity, and acquisition time in the denoised point cloud parameter set are first read, and the point cloud density is calculated according to the point cloud coordinates at the same acquisition time. The point cloud density is calculated by the number of points in a unit spatial range. Then, the voxel resolution is set according to the point cloud density distribution parameters. When the point cloud density is high, a fine-grained voxel resolution is used to partition the point cloud, and when the point cloud density is low, a coarse-grained voxel resolution is used to partition the point cloud, so that the denoised point cloud parameter set forms voxelized point cloud data according to different spatial scales. After completing the voxel segmentation, a geometric analysis is performed on the point cloud coordinates within each voxel. The center coordinates, coordinate dispersion, normal direction, curvature change, and height difference of the point cloud within the voxel are calculated. The results of the geometric analysis are then arranged according to the voxel position to extract local geometric features. Simultaneously, the point cloud intensity within each voxel is statistically analyzed. The mean, maximum, minimum, and variance of the point cloud intensity, as well as the intensity variation between adjacent voxels, are calculated. The intensity statistical results are then arranged according to the voxel position to extract intensity distribution features.

[0034] Subsequently, when performing electromagnetic scattering analysis based on the millimeter-wave radar detection vector, the detection range, azimuth, radial velocity, and echo intensity corresponding to the same acquisition time in the millimeter-wave radar detection vector are first read. The spatial position of the millimeter-wave radar detection point is then calculated according to the detection range and azimuth. Specifically, the lateral position is obtained by multiplying the detection range by the cosine of the azimuth, and the longitudinal position is obtained by multiplying the detection range by the sine of the azimuth, ensuring that each millimeter-wave radar detection point has a corresponding spatial position, radial velocity, and echo intensity. The millimeter-wave radar detection points are then arranged according to the acquisition time, and within the same acquisition time, they are sorted by detection range from near to far and azimuth from small to large. For any millimeter-wave radar detection point, its echo intensity is multiplied by the square of the detection range to obtain the range-corrected echo intensity, thus reducing the impact of changes in detection range on the echo intensity calculation. Next, calculations are performed on adjacent millimeter-wave radar detection points at the same acquisition time. The difference in echo intensity after distance correction between adjacent millimeter-wave radar detection points is taken as the echo intensity variation value, the difference in azimuth angle between adjacent millimeter-wave radar detection points is taken as the azimuth variation value, and the difference in radial velocity between adjacent millimeter-wave radar detection points is taken as the velocity variation value. Simultaneously, the average value of the distance-corrected echo intensity at the same acquisition time is calculated to obtain the average scattering intensity, the maximum value of the distance-corrected echo intensity is taken to obtain the peak scattering intensity, and the average value of the radial velocity is calculated to obtain the average velocity. Finally, the spatial position, distance-corrected echo intensity, echo intensity variation value, azimuth variation value, velocity variation value, average scattering intensity, peak scattering intensity, and average velocity at each acquisition time are combined in chronological order to generate electromagnetic scattering characteristics.

[0035] Subsequently, heterogeneous feature fusion is performed based on local geometric features, intensity distribution features, and electromagnetic scattering features. In this process, the voxelized point cloud center data and millimeter-wave radar detection point data are first located, and these data are used as graph nodes to construct a dynamic connection graph. Then, query vectors, key vectors, and value vectors are generated based on local geometric features and electromagnetic scattering features, respectively. Cross-modal association weights are calculated through cross-attention to initialize the graph edge weights in the dynamic connection graph. Further, the dynamic connection graph is traversed to perform node motion consistency analysis, and the dynamic connection graph is updated and fused layer by layer based on the phase separation results and graph edge weights. This allows local geometric features, intensity distribution features, and electromagnetic scattering features to be associated and expressed in both time and space dimensions, obtaining a spatiotemporal sensing feature flow.

[0036] Next, a conditional neural implicit field decoder is constructed to implicitly decode the spatiotemporal sensing feature stream. When constructing the implicit dynamic sensing map, the fusion features corresponding to each acquisition time and spatial coordinate in the spatiotemporal sensing feature stream are first read, along with the corresponding spatial coordinates and acquisition time. Then, the spatial coordinates, acquisition time, and fusion features are concatenated according to a fixed field order to form the input data for the conditional neural implicit field decoder. The conditional neural implicit field decoder uses predetermined decoding parameters to perform layer-by-layer calculations on the input data. During layer-by-layer calculations, each feature value in the input data is multiplied by its corresponding decoding parameter and summed to obtain intermediate calculated values. If the intermediate calculated value is greater than zero, it is retained; if it is less than or equal to zero, it is set to zero, and the processed intermediate calculated value is used as the input for the next layer. After continuous layer-by-layer calculations, the implicit decoding results under the same acquisition time and spatial coordinates are output. Finally, the implicit decoding results are arranged according to the acquisition time and spatial coordinates to construct the implicit dynamic sensing map.

[0037] When further performing perceptual analysis on the standard test dataset based on the conditional neural implicit field decoder, the denoised point cloud parameter set, millimeter-wave radar detection vector, and corresponding spatiotemporal sensing feature stream in the standard test dataset are first read according to the acquisition time and spatial coordinates. Then, the data under the same acquisition time and spatial coordinates are input into the conditional neural implicit field decoder. The intermediate calculated values ​​and final decoded output values ​​of the conditional neural implicit field decoder in the implicit decoding process are extracted, and the intermediate calculated values ​​and final decoded output values ​​are arranged in correspondence with the corresponding local geometric features, intensity distribution features, and electromagnetic scattering features. During the arrangement, the acquisition time is used as the time dimension, the spatial coordinates as the spatial dimension, and the corresponding feature values ​​as the feature dimension. The feature values ​​obtained under each acquisition time and each spatial coordinate are written into the corresponding positions. When there are multiple feature values ​​under the same acquisition time and the same spatial coordinate, the average of the multiple feature values ​​is calculated and written into the corresponding position. When a corresponding feature value is missing, the feature values ​​under adjacent acquisition times or adjacent spatial coordinates are selected for linear filling, thereby calculating the perceptual feature tensor.

[0038] Finally, the perceptual feature tensor and the implicit dynamic perceptual map are combined to construct the dynamic perceptual map of the coal mine tunneling environment. During this process, the perceptual feature tensor and the implicit dynamic perceptual map are matched according to the acquisition time and spatial coordinates. When the feature values ​​in the perceptual feature tensor and the implicit decoding results in the implicit dynamic perceptual map have the same acquisition time and spatial coordinates, they are written into the same data record. When the acquisition times are inconsistent, the implicit decoding results from adjacent acquisition times are selected and linearly converted according to the ratio of the current acquisition time to the adjacent acquisition times before being written. When the spatial coordinates are inconsistent, the implicit decoding result corresponding to the nearest spatial coordinate is selected and written. After matching is completed, the perceptual feature tensor and the implicit dynamic perceptual map under each acquisition time and each spatial coordinate are associated and stored to construct the dynamic perceptual map of the coal mine tunneling environment, which includes the perceptual feature tensor.

[0039] Furthermore, in the method provided in the application embodiments, the multi-scale dynamic voxel partitioning based on the denoised point cloud parameter set and the extraction of local geometric features further includes: Density calculation is performed based on the denoised point cloud parameter set. The voxel resolution is set according to the point cloud density distribution parameters, including fine-grained and coarse-grained voxel resolutions. Distance division is performed based on the denoised point cloud parameter set to determine near-distance and far-distance regions. Geometric analysis is performed on the near-distance region according to the fine-grained voxel resolution to obtain a first geometric descriptor. Geometric analysis is also performed on the far-distance region according to the fine-grained voxel resolution to obtain a second geometric descriptor. The first and second geometric descriptors are then aggregated to generate the local geometric features.

[0040] In this embodiment, when calculating density based on a denoised point cloud parameter set, the point cloud coordinates, point cloud intensity, and acquisition time are first read from the denoised point cloud parameter set, and the point cloud coordinates are grouped according to the acquisition time. At the same acquisition time, the maximum and minimum values ​​of the point cloud coordinates in the vertical, horizontal, and vertical coordinates are read to determine the point cloud distribution range, and then the point cloud distribution range is divided according to a preset statistical interval. For each division location, the number of point clouds falling within that location is counted, and the number of point clouds is divided by the spatial volume corresponding to that location to obtain the point cloud density at that location. After calculating the point cloud density at each division location in sequence, the point cloud densities are arranged according to spatial location to form point cloud density distribution parameters. Subsequently, the average point cloud density at each location is calculated at the same acquisition time. Locations with point cloud densities greater than or equal to the average value are set to fine-grained voxel resolution, and locations with point cloud densities less than the average value are set to coarse-grained voxel resolution, so that the voxel resolution includes both fine-grained and coarse-grained voxel resolutions.

[0041] Next, when dividing the distance based on the denoised point cloud parameter set, the distance between each point cloud coordinate and the coal mine equipment coordinate is read. During distance calculation, the squared differences between the point cloud coordinates and the coal mine equipment coordinates in the vertical, horizontal, and longitudinal coordinates are summed, and then the square root of the sum is taken to obtain the distance corresponding to that point cloud coordinate. Subsequently, the point cloud coordinates are divided according to distance. Point cloud coordinates with distances less than or equal to a preset distance threshold are assigned to the near-distance region, and point cloud coordinates with distances greater than the preset distance threshold are assigned to the far-distance region. The point cloud coordinates, point cloud intensity, and acquisition time are retained for both the near-distance and far-distance regions.

[0042] Subsequently, geometric analysis of the near-field region is performed at a fine-grained voxel resolution. First, the point cloud coordinates in the near-field region are divided into voxels at a fine-grained voxel resolution, and the point cloud coordinates within each voxel are read. For any voxel, the average values ​​of all point cloud coordinates within that voxel are calculated in the longitudinal, lateral, and vertical coordinates to obtain the voxel center coordinates. Then, the differences between the maximum and minimum values ​​of the longitudinal, lateral, and vertical coordinates within that voxel are calculated, and these three differences are used as the geometric span of the voxel in the three coordinate directions. Next, the distances from each point cloud coordinate within the voxel to the voxel center coordinates are calculated, and the average of all distances is taken as the coordinate dispersion. Simultaneously, the maximum and minimum values ​​of the vertical coordinates within the voxel are read, and the difference between them is taken as the height difference. Finally, the voxel center coordinates, the geometric spans in the three coordinate directions, the coordinate dispersion, and the height difference are arranged according to the voxel position to obtain the first geometric descriptor.

[0043] Next, geometric analysis is performed on the distant region according to fine-grained voxel resolution. First, the point cloud coordinates in the distant region are divided into voxels according to fine-grained voxel resolution, and the point cloud coordinates in each voxel are read. For any voxel, the voxel center coordinates, geometric span in three coordinate directions, coordinate dispersion and height difference are calculated in the same way as in the near region. The above calculation results are arranged according to the voxel position to obtain the second geometric descriptor.

[0044] Finally, when aggregating the first and second geometric descriptors, the first and second geometric descriptors are first matched according to their acquisition time, and then the spatial arrangement order is determined according to the voxel center coordinates. For the first and second geometric descriptors acquired at the same time, the voxel center coordinates, geometric span in the three coordinate directions, coordinate dispersion, and height difference are read and written in parallel according to the same field order. For geometric descriptors that exist only in the near or far regions, they are written directly according to their acquisition time and voxel center coordinates. After aggregation, the geometric description results corresponding to each acquisition time and voxel position are arranged uniformly to generate local geometric features.

[0045] Furthermore, in the method provided in the application embodiments, the heterogeneous feature fusion based on the local geometric features, the intensity distribution features, and the electromagnetic scattering features to construct a spatiotemporally perceptible feature stream further includes: The voxelized point cloud center data and millimeter-wave radar detection point data are located, and the voxelized point cloud center data and the millimeter-wave radar detection point data are used as graph nodes to construct a dynamic connection graph. Based on the local geometric features and the electromagnetic scattering features, query vectors, key vectors, and value vectors are generated respectively, and cross-attention is performed to calculate cross-modal association weights, and graph edge weight values ​​are initialized. The dynamic connection graph is traversed to perform node motion consistency analysis, and the dynamic connection graph is updated and fused layer by layer according to the phase separation results and graph edge weight values ​​to obtain the spatiotemporal perception feature flow.

[0046] In this embodiment, the voxelized point cloud center data and millimeter-wave radar detection point data are first located. These data are then used as graph nodes to construct a dynamic connection graph. During this process, the voxel center coordinates, acquisition time, and point cloud intensity statistics corresponding to the intensity distribution features are first read. The average vertical, horizontal, and vertical coordinates of the point cloud within the same voxel are then used as the voxelized point cloud center data. Next, the millimeter-wave radar detection point data corresponding to the electromagnetic scattering features is read. This data is calculated from the detection range and azimuth angle, where the horizontal position is the product of the detection range and the cosine of the azimuth angle, and the vertical position is the product of the detection range and the sine of the azimuth angle. Subsequently, the voxelized point cloud center data and the millimeter-wave radar detection point data are used as graph nodes and grouped according to the acquisition time. Under the same acquisition time, the spatial distance between the voxelized point cloud center data and the millimeter-wave radar detection point data is calculated. When the spatial distance is less than or equal to the preset connection distance, graph edges are established between the corresponding graph nodes. When the spatial distance is greater than the preset connection distance, graph edges are not established. The graph nodes and graph edges are established sequentially according to the acquisition time to construct a dynamic connection graph.

[0047] Next, query vectors, key vectors, and value vectors are generated based on local geometric features and electromagnetic scattering features, respectively. Cross-attention calculation is then performed to determine cross-modal association weights, and graph edge weights are initialized. Specifically, local geometric features, electromagnetic scattering features, and intensity distribution features are first normalized. Normalization involves subtracting the minimum value of a similar feature from the current feature value, and then dividing by the difference between the maximum and minimum values ​​of the same feature. When the maximum and minimum values ​​are equal, the normalized feature value is recorded as zero. Subsequently, the voxel center coordinates, geometric span, coordinate dispersion, and height difference from the local geometric features are arranged into query vectors according to a fixed field order. The spatial location, distance-corrected echo intensity, echo intensity variation, azimuth variation, and velocity variation from the electromagnetic scattering features are arranged into key vectors according to a fixed field order. The distance-corrected echo intensity, mean scattering intensity, peak scattering intensity, and mean velocity from the electromagnetic scattering features are arranged into value vectors. If the lengths of the query vector, key vector, and value vector are inconsistent, they are processed according to a preset vector length. Parts exceeding the preset vector length are truncated according to the field order, and parts insufficient for the preset vector length are padded with zeros. For any voxelized point cloud center data corresponding to a graph node, its query vector is multiplied bitwise with the key vector corresponding to the connected millimeter-wave radar detection point data, and then summed to obtain a dot product value. The dot product value is then divided by the square root of the vector length to obtain the attention score. The attention scores corresponding to multiple millimeter-wave radar detection point data connected to the same voxelized point cloud center data are subjected to exponential normalization processing, that is, the exponent of each attention score is taken, and each exponent value is divided by the sum of all exponent values ​​to obtain the cross-modal association weight. The cross-modal association weight is written into the corresponding graph edge, the graph edge weight value is initialized, and the cross-modal association weight is multiplied with the corresponding value vector and summed to obtain the radar weighted feature corresponding to the voxelized point cloud center data. The radar weighted feature and the intensity distribution feature are then written in parallel to form the fusion feature of the graph node.

[0048] Finally, the dynamic connection graph is traversed to perform node motion consistency analysis. Based on the phase separation results and graph edge weights, the dynamic connection graph is updated and fused layer by layer. Graph nodes in the dynamic connection graph are read in the order of acquisition time. For voxelized point cloud center data that are spatially adjacent in adjacent acquisition times, the difference between the spatial position of the later acquisition time and the spatial position of the earlier acquisition time is calculated, and this difference is divided by the time interval to obtain the point cloud side motion velocity. For millimeter-wave radar detection point data, its radial velocity is read, and the radial velocity is decomposed into lateral velocity and longitudinal velocity according to the azimuth angle. Then, the point cloud side motion velocity is compared with the velocity corresponding to the millimeter-wave radar detection point data. If the velocity difference between the two is less than or equal to a preset velocity difference, the corresponding graph edge is assigned to the motion-consistent phase separation result. If the velocity difference is greater than the preset velocity difference, the corresponding graph edge is assigned to the motion-inconsistent phase separation result. During layer-by-layer update fusion, the current edge weight value is retained for graph edges corresponding to motion-consistent phase separation results, while for graph edges corresponding to motion-inconsistent phase separation results, the current edge weight value is multiplied by a preset attenuation coefficient. Then, for each graph node, the fusion features of adjacent graph nodes are read, and the fusion features of adjacent graph nodes are multiplied by their corresponding edge weight values, summed, and then added to the fusion features of the current graph node and averaged to obtain the updated graph node features. The dynamic connection graph is traversed in the above manner, and the updated graph node features of each layer are used as input for the next layer update, until the preset number of update layers is reached. Finally, the features are arranged according to the acquisition time, graph node spatial location, and updated graph node features to obtain a spatiotemporally aware feature stream.

[0049] Step S300: Perform multi-sensor combination multi-directional sensing test based on the sensing feature tensor, calculate the multi-directional sensing test parameter set, feed the multi-directional sensing test parameter set back to the multi-sensor combination for configuration adjustment, and generate a coal mine tunneling sensing test report.

[0050] In this embodiment, when performing multi-sensor combined multi-directional perception testing based on the perception feature tensor, a multi-directional perception decoding head is first constructed, and the perception feature tensor is input into the multi-directional perception decoding head for cross-attention decoding prediction to obtain multi-directional target pose prediction values ​​and multi-directional motion prediction values. Subsequently, a true reference target is set based on the test scene parameters, and the multi-directional target pose prediction values ​​and multi-directional motion prediction values ​​are compared with the true reference target to calculate and form a multi-directional perception test parameter set. Furthermore, the multi-directional perception test parameter set is used as state feedback input to the deep reinforcement learning process of the multi-sensor combination to obtain configuration adjustment action parameters, and the parameters of the multi-sensor combination are adjusted online according to the configuration adjustment action parameters to generate a sensor configuration sequence. Finally, multiple rounds of testing are summarized based on the multi-directional perception test parameter set and the sensor configuration sequence, and the results of each round of multi-directional perception testing, configuration adjustment action parameters, and corresponding sensor configuration sequences are correlated and organized to construct a coal mine tunneling perception test report.

[0051] Furthermore, in the method provided in the application embodiment, the multi-sensor combination multi-directional sensing test is performed based on the sensing feature tensor, the multi-directional sensing test parameter set is calculated, the multi-directional sensing test parameter set is fed back to the multi-sensor combination for configuration adjustment, and the coal mine tunneling sensing test report is generated, which further includes: A multi-directional sensing decoder is constructed, and the sensing feature tensor is input into the multi-directional sensing decoder for cross-attention decoding and prediction to generate multi-directional target pose prediction values ​​and multi-directional motion prediction values. A ground truth reference target is set based on the test scene parameters. The multi-directional target pose prediction values ​​and multi-directional motion prediction values ​​are compared with the ground truth reference target to calculate a multi-directional sensing test parameter set. The multi-directional sensing test parameter set is used as state feedback to perform deep reinforcement learning on a multi-sensor combination to construct configuration adjustment action parameters. The multi-sensor combination is adjusted online according to the configuration adjustment action parameters to generate a sensor configuration sequence. Multiple rounds of testing are conducted based on the multi-directional sensing test parameter set and the sensor configuration sequence to summarize and construct the coal mine tunneling sensing test report.

[0052] In this embodiment, a multi-directional sensing decoding head is first constructed, and the sensing feature tensor is input into the multi-directional sensing decoding head for cross-attention decoding prediction. In this process, feature values ​​corresponding to the same acquisition time and spatial coordinates in the sensing feature tensor are first read, with the coal mine equipment coordinates used as the spatial reference. The sensing feature tensor consists of feature data arranged according to acquisition time, spatial coordinates, and feature dimensions. The multi-directional sensing decoding head is composed of multiple dedicated decoding branches for each direction, including directional sectors, horizontal sectors, and vertical sectors. Specifically, the longitudinal, horizontal, and vertical coordinate differences between each spatial coordinate and the coal mine equipment coordinates are first calculated, and the corresponding spatial coordinates are assigned to the directional, horizontal, or vertical sector based on the coordinate differences. Subsequently, the sensing feature tensors within the corresponding sectors are read by the directional, horizontal, and vertical sectors, and the sensing feature tensors are weighted and read using dot product attention. During dot product attention calculation, the feature values ​​within the corresponding sector are multiplied one by one with the feature values ​​corresponding to each spatial coordinate at the same acquisition time, and the results are summed to obtain the dot product attention score. Then, each dot product attention score is divided by the sum of all dot product attention scores within the same sector to obtain the attention weight for the corresponding spatial coordinate. When the sum of all dot product attention scores is zero, the attention weights for each spatial coordinate within the same sector are set to the same value. After weighted reading, each spatial coordinate is multiplied by its corresponding attention weight and summed to obtain the target position coordinates for the corresponding sector. The target attitude angle is obtained based on the azimuth and pitch angles of the target position coordinates relative to the coal mine equipment coordinates. Finally, the target velocity is calculated based on the change in target position coordinates between adjacent acquisition times and the acquisition time interval, and the direction of change in target position coordinates is taken as the target velocity direction. The first pose prediction value and the first motion prediction value are obtained from the directional sector, the second pose prediction value and the second motion prediction value are obtained from the lateral sector, and the third pose prediction value and the third motion prediction value are obtained from the vertical sector. The first pose prediction value, the second pose prediction value and the third pose prediction value are combined into a multi-directional target pose prediction value, and the first motion prediction value, the second motion prediction value and the third motion prediction value are combined into a multi-directional motion prediction value.

[0053] Next, a true reference target is set based on the test scenario parameters. The multi-directional target pose prediction values ​​and multi-directional motion prediction values ​​are compared with the true reference target to calculate the multi-directional perception test parameter set. Specifically, the 3D geometric parameters, coal mine equipment coordinates, tunnel face coordinates, cutting head working parameters, cutting trajectory, and calculation time from the test scenario parameters are read. The true reference target is then set based on the target's true position, true attitude, true motion speed, and true motion direction at the same calculation time. Subsequently, the first pose prediction value and first motion prediction value of the directional sector, the second pose prediction value and second motion prediction value of the horizontal sector, and the third pose prediction value and third motion prediction value of the vertical sector are compared with the true reference target to obtain multi-directional error values. These multi-directional error values ​​include the position difference between the predicted position and the true position in each sector, the attitude difference between the predicted attitude and the true attitude, the velocity difference between the predicted motion speed and the true motion speed, and the direction difference between the predicted motion direction and the true motion direction. Next, covariance analysis is performed on the multi-directional error values. First, the average error values ​​of the directional sector, lateral sector, and vertical sector are calculated at multiple acquisition times. Then, the average error value of the corresponding sector is subtracted from the error value at each acquisition time to obtain the error deviation value. Subsequently, the error deviation values ​​of any two sectors are multiplied and averaged to obtain the covariance between the two sectors. The covariance is then arranged in the order of directional sector, lateral sector, and vertical sector to construct the directional sensitivity matrix. Then, confidence is calculated based on the first pose prediction value, the first motion prediction value, the second pose prediction value, the second motion prediction value, the third pose prediction value, and the third motion prediction value to obtain the multi-directional detection confidence. Specifically, the dot product attention score and the corresponding prediction value of each sector are read. When the attention weights within the same sector are more concentrated and the predicted position changes more continuously at adjacent acquisition times, the multi-directional detection confidence of the corresponding sector is higher. Subsequently, blind zone extreme value sensing is performed based on multi-directional detection confidence. Continuous angular ranges where the multi-directional detection confidence is below a preset confidence threshold are defined as sensing blind zones. The result with the largest angular range among multiple sensing blind zones is selected to calculate the maximum sensing blind zone angle. Simultaneously, multi-directional target tracking is performed based on multi-directional detection confidence. The time difference between the transition from a low-confidence state to a stable detection state for the same target in adjacent acquisition times is read to calculate the dynamic target tracking delay parameter. Finally, the directional sensitivity matrix, the maximum sensing blind zone angle, and the dynamic target tracking delay parameter are combined to obtain the multi-directional sensing test parameter set.

[0054] Subsequently, the multi-directional sensing test parameter set is used as state feedback to perform deep reinforcement learning on the multi-sensor combination, constructing configuration adjustment action parameters. In this process, the directional sensitivity matrix, maximum sensing blind zone angle, and dynamic target tracking delay parameter are first arranged in a fixed order as the state feedback for the current round, and the current configuration of the multi-sensor combination is read. Then, based on the configuration items allowed for online adjustment in the multi-sensor combination and their allowed ranges, several candidate configuration adjustment action parameters are generated; each candidate configuration adjustment action parameter corresponds to one executable multi-sensor combination configuration change. When numerically processing the state feedback for the current round, the average of the absolute values ​​of each element in the directional sensitivity matrix is ​​first calculated to obtain the directional sensitivity value; then, the maximum sensing blind zone angle is divided by a preset maximum angle to obtain the blind zone angle value; the dynamic target tracking delay parameter is divided by a preset allowed delay to obtain the tracking delay value; finally, the directional sensitivity value, blind zone angle value, and tracking delay value are added together to obtain the comprehensive test value for the current round. When calculating the evaluation value for each candidate configuration adjustment action parameter, the candidate configuration adjustment action parameter is first written into the multi-sensor combination and maintained for a test acquisition period. During this test acquisition period, the multi-directional perception test is re-executed to obtain the directional sensitivity matrix, maximum perception blind zone angle, and dynamic target tracking delay parameter corresponding to the candidate configuration adjustment action parameter. Then, the comprehensive test value corresponding to the candidate configuration adjustment action parameter is calculated using the same calculation method as the current round. Subsequently, the comprehensive test value corresponding to the candidate configuration adjustment action parameter is subtracted from the comprehensive test value of the current round to obtain the feedback increment of the candidate configuration adjustment action parameter. If the feedback increment is positive, it indicates that the candidate configuration adjustment action parameter reduces the comprehensive test value; if the feedback increment is negative or zero, it indicates that the candidate configuration adjustment action parameter does not reduce the comprehensive test value. Next, the evaluation value of the candidate configuration adjustment action parameter saved in the previous round is read. If no evaluation value exists in the previous round, it is recorded as zero. Then, the maximum value among the evaluation values ​​saved in the previous round for each candidate configuration adjustment action parameter after execution is read. The feedback increment, the maximum value multiplied by the preset discount coefficient, and the result are added together to obtain the target evaluation value. Finally, the target evaluation value is subtracted from the previous round evaluation value to obtain the evaluation difference. The evaluation difference is then multiplied by the preset learning step size and added back to the previous round evaluation value to obtain the current evaluation value of the candidate configuration adjustment action parameter. The current evaluation values ​​of all candidate configuration adjustment action parameters are calculated sequentially in the above manner, and the candidate configuration adjustment action parameter with the largest current evaluation value is selected as the configuration adjustment action parameter output in this round.

[0055] Then, the multi-sensor combination is adjusted online according to the configured adjustment action parameters. When generating the sensor configuration sequence, the configuration adjustment action parameters output in this round are read and written into the adjustable configuration items of the multi-sensor combination. Before writing, the allowable range of each adjustable configuration item is read. When the adjustment result corresponding to the configuration adjustment action parameter is within the allowable range, the adjustment result is written into the multi-sensor combination. When the adjustment result is lower than the lower limit of the allowable range, the lower limit of the allowable range is written. When the adjustment result is higher than the upper limit of the allowable range, the upper limit of the allowable range is written. After completing one round of online parameter adjustment, the acquisition time, multi-directional sensing test parameter set, configuration adjustment action parameters, and adjusted multi-sensor combination configuration corresponding to this round are recorded. Then, the next round of multi-directional sensing test is started, and the adjusted multi-sensor combination configuration of each round is recorded in sequence according to the test round to generate the sensor configuration sequence.

[0056] Finally, based on the multi-directional perception test parameter set and sensor configuration sequence, multiple rounds of testing are summarized to construct a coal mine tunneling perception test report. In this process, the sensor configuration sequences generated in each round are first arranged according to the test round to form a sensor configuration sequence. Then, the corresponding test scene parameters, multi-directional target pose prediction values, multi-directional motion prediction values, true reference targets, multi-directional perception test parameter sets, and configuration adjustment action parameters are read round by round. Subsequently, the directional sensitivity matrix, maximum perception blind zone angle, and dynamic target tracking delay parameters in each round are correlated with the sensor configuration sequence of the same round, and the changes in the directional sensitivity matrix, maximum perception blind zone angle, and dynamic target tracking delay parameters are calculated across multiple test rounds. Finally, the test scene parameters, multi-directional target pose prediction values, multi-directional motion prediction values, true reference targets, multi-directional perception test parameter sets, configuration adjustment action parameters, sensor configuration sequences, and the summarized results of multiple test rounds are arranged according to the test round to construct the coal mine tunneling perception test report.

[0057] Furthermore, the method provided in the application embodiments also includes: The multi-directional sensing decoding head consists of multiple dedicated decoding branches for different directions. These dedicated decoding branches include directional sectors, lateral sectors, and vertical sectors. The multiple dedicated decoding branches use dot product attention to perform weighted readings on the sensing feature tensor to obtain the first pose prediction value and the first motion prediction value for the directional sector, the second pose prediction value and the second motion prediction value for the lateral sector, and the third pose prediction value and the third motion prediction value for the vertical sector.

[0058] In this embodiment, the multi-directional sensing decoding head is composed of multiple dedicated decoding branches for different directions. These dedicated decoding branches include directional sectors, horizontal sectors, and vertical sectors. The directional sectors correspond to the spatial coordinates distributed along the front-to-back direction of the coal mining equipment in the sensing feature tensor, the horizontal sectors correspond to the spatial coordinates distributed along the left-to-right direction of the coal mining equipment in the sensing feature tensor, and the vertical sectors correspond to the spatial coordinates distributed along the up-and-down direction of the coal mining equipment in the sensing feature tensor. When constructing the multi-directional sensing decoder, the feature values ​​corresponding to each spatial coordinate at the same acquisition time in the sensing feature tensor are first read, and the coordinates of the coal mine equipment are used as the basis for sector division. Then, the differences between each spatial coordinate and the coal mine equipment coordinates in the longitudinal, lateral, and vertical coordinates are calculated. When the absolute value of the longitudinal coordinate difference is greater than or equal to the absolute values ​​of the lateral and vertical coordinate differences, the spatial coordinate and its corresponding feature value are assigned to the directional sector. When the absolute value of the lateral coordinate difference is greater than the absolute values ​​of the longitudinal and vertical coordinate differences, the spatial coordinate and its corresponding feature value are assigned to the lateral sector. When the absolute value of the vertical coordinate difference is greater than the absolute values ​​of the longitudinal and lateral coordinate differences, the spatial coordinate and its corresponding feature value are assigned to the vertical sector. Through these steps, the sensing feature tensor is input into the corresponding directional dedicated decoding branch according to the directional, lateral, and vertical sectors.

[0059] When multiple directional dedicated decoding branches perform weighted reading of the perceptual feature tensor using dot-product attention, they first read the feature values ​​at corresponding spatial coordinates within the directional sector, horizontal sector, and vertical sector, respectively. Dot-product attention involves multiplying the feature values ​​to be decoded within the same sector with the feature values ​​participating in the reading within the perceptual feature tensor, summing the results, and determining the weight of each spatial coordinate based on the summation. Specifically, for a directional sector, the feature values ​​within the directional sector are first selected as the features to be decoded. Then, the features to be decoded in the directional sector are multiplied with the feature values ​​corresponding to each spatial coordinate within the directional sector, and summed to obtain the dot-product attention score for each spatial coordinate within the directional sector. Subsequently, each dot-product attention score is divided by the sum of all dot-product attention scores within the directional sector to obtain the attention weight for each spatial coordinate within the directional sector. When the sum of all dot-product attention scores is zero, the attention weight for each spatial coordinate within the directional sector is set to the same value. The horizontal and vertical sectors calculate their respective dot-product attention scores and attention weights in the same way. After completing the attention weight calculation, the feature values ​​of each spatial coordinate in each sector are multiplied by the corresponding attention weight and then summed to obtain the weighted reading result of the corresponding sector.

[0060] After obtaining the weighted reading results of the directional sector, the dedicated directional decoding branch corresponding to the directional sector calculates the first pose prediction value and the first motion prediction value based on the weighted reading results. Specifically, each spatial coordinate within the directional sector is multiplied by its corresponding attention weight and then summed to obtain the target position coordinates of the directional sector. Then, the target attitude angle is obtained based on the azimuth and pitch angles of the target position coordinates relative to the coordinates of the coal mine equipment. The target position coordinates and the target attitude angle are used as the first pose prediction value of the directional sector. Subsequently, the target position coordinates of the directional sector at adjacent acquisition times are read. The target position coordinates of the later acquisition time are subtracted from the target position coordinates of the previous acquisition time and divided by the acquisition time interval to obtain the target motion velocity. The direction of change of the target position coordinates is used as the target motion direction, and the target motion velocity and the target motion direction are used as the first motion prediction value of the directional sector. Following the same calculation process, the dedicated directional decoding branch corresponding to the horizontal sector obtains the second pose prediction value and the second motion prediction value based on the weighted reading results of the horizontal sector, and the dedicated directional decoding branch corresponding to the vertical sector obtains the third pose prediction value and the third motion prediction value based on the weighted reading results of the vertical sector. Therefore, by using multiple dedicated decoding branches corresponding to the directional sector, horizontal sector and vertical sector, the perceptual feature tensor is read by dot product attention weighting, and the first pose prediction value and first motion prediction value of the directional sector, the second pose prediction value and second motion prediction value of the horizontal sector, and the third pose prediction value and third motion prediction value of the vertical sector are obtained.

[0061] Furthermore, in the method provided in the application embodiments, the calculation of the multi-directional perception test parameter set by comparing the multi-directional target pose prediction value, the multi-directional motion prediction value, and the true reference target further includes: The first pose prediction value, first motion prediction value of the directional sector, the second pose prediction value, second motion prediction value of the lateral sector, the third pose prediction value, and third motion prediction value of the vertical sector are compared with the true reference target to obtain multi-directional error values. Covariance analysis is performed on the multi-directional error values ​​to construct a directional sensitivity matrix. Confidence calculation is performed based on the first pose prediction value, first motion prediction value of the directional sector, the second pose prediction value, second motion prediction value of the lateral sector, and third pose prediction value of the vertical sector to obtain a multi-directional detection confidence score. Blind zone extreme value perception is performed based on the multi-directional detection confidence score to calculate the maximum perception blind zone angle. Multi-directional target tracking is performed based on the multi-directional detection confidence score to calculate the dynamic target tracking delay parameter. The directional sensitivity matrix, the maximum perception blind zone angle, and the dynamic target tracking delay parameter are combined to obtain the multi-directional perception test parameter set.

[0062] In this embodiment of the application, when comparing the first pose prediction value and the first motion prediction value of the directional sector, the second pose prediction value and the second motion prediction value of the lateral sector, and the third pose prediction value and the third motion prediction value of the vertical sector with the true reference target, the first pose prediction value and the first motion prediction value of the directional sector, the second pose prediction value and the second motion prediction value of the lateral sector, and the third pose prediction value and the third motion prediction value of the vertical sector are read according to the acquisition time, and the true pose and true motion data in the true reference target at the same acquisition time are read simultaneously; wherein, the first pose prediction value, the second pose prediction value and the third pose prediction value all include the target position coordinates and the target attitude angle, and the first motion prediction value, the second motion prediction value and the third motion prediction value all include the target motion speed and the target motion direction. Subsequently, for any sector, the differences between the predicted position coordinates and the true position coordinates in the true reference target are squared in the longitudinal, lateral, and vertical coordinates, summed, and then the square root of the sum is taken to obtain the position difference. Simultaneously, the absolute values ​​of the differences between the predicted and true attitude angles are taken to obtain the attitude difference, and the absolute values ​​of the differences between the predicted and true motion velocities are taken to obtain the velocity difference. The angle between the predicted and true motion directions is used as the direction difference. Next, the position, attitude, velocity, and direction differences for the same acquisition time and the same sector are arranged to obtain multi-directional error values. Thus, the multi-directional error values ​​corresponding to the directional sector, the lateral sector, and the vertical sector together form the error calculation results for each sector.

[0063] Next, covariance analysis is performed on the multi-directional error values. When constructing the directional sensitivity matrix, the multi-directional error values ​​at multiple acquisition times are first organized according to directional sector, lateral sector, and vertical sector. For each sector, the position difference, attitude difference, velocity difference, and direction difference at the same acquisition time are summed to obtain the multi-directional error value corresponding to that sector at that acquisition time. Subsequently, the average value of the multi-directional error values ​​for the directional sector, lateral sector, and vertical sector at all acquisition times is calculated. Based on this, the average value of the corresponding sector is subtracted from the multi-directional error value of each sector at each acquisition time to obtain the error deviation of the corresponding sector at that acquisition time. Further, any two sectors are selected, and their error deviations at the same acquisition time are multiplied and averaged over all acquisition times to obtain the covariance between the two sectors. The covariance between a sector and itself is obtained by averaging the squares of the sector's error deviations. Finally, the covariance between each sector is written into a three-row, three-column matrix in the order of directional sector, lateral sector, and vertical sector to construct the directional sensitivity matrix; the elements in the directional sensitivity matrix record the synchronous change relationship of multi-directional error values ​​between directional sector, lateral sector, and vertical sector.

[0064] Subsequently, confidence calculations are performed based on the first pose prediction value and first motion prediction value of the directional sector, the second pose prediction value and second motion prediction value of the lateral sector, and the third pose prediction value and third motion prediction value of the vertical sector. In this process, the pose prediction values ​​and motion prediction values ​​of the directional sector, lateral sector, and vertical sector at adjacent acquisition times are first read. For any sector, the target position coordinates and target motion velocity of the previous acquisition time are read first, and the target motion velocity of the previous acquisition time is multiplied by the acquisition time interval to obtain the position change. Then, the target position coordinates of the previous acquisition time are added to the position change to obtain the reference position coordinates of the current acquisition time. Next, the target position coordinates in the pose prediction value of the current acquisition time are compared with the reference position coordinates, and the position continuity difference is calculated by taking the square root of the sum of the squares of the differences in the longitudinal, lateral, and vertical coordinates. Simultaneously, the target attitude angle of the current acquisition time is subtracted from the target attitude angle of the previous acquisition time, and the absolute value is taken to obtain the attitude continuity difference; the target motion velocity of the current acquisition time is subtracted from the target motion velocity of the previous acquisition time, and the absolute value is taken to obtain the velocity continuity difference. Finally, the position continuous difference, attitude continuous difference, and velocity continuous difference are added together, and the sum of 1 and 1 is divided to obtain the confidence value of the sector at that acquisition time. The confidence values ​​of the directional sector, lateral sector, and vertical sector at each acquisition time are calculated in the same way, and arranged according to the acquisition time and sector type to obtain the multi-directional detection confidence.

[0065] Next, based on the multi-directional detection confidence level, extreme value perception of blind zones is performed. The maximum perception blind zone angle is calculated, and the directional, lateral, and vertical sectors are mapped to the angle ranges in the test scene parameters. The multi-directional detection confidence levels corresponding to each sector are arranged in order of angle. Subsequently, the multi-directional detection confidence levels within each angle range are compared with a preset confidence threshold. When the multi-directional detection confidence level within a certain angle range is less than the preset confidence threshold, that angle range is determined as a perception blind zone. Furthermore, if the multi-directional detection confidence levels within adjacent angle ranges are all less than the preset confidence threshold, the adjacent angle ranges are merged into a continuous perception blind zone, and the angle span of this continuous perception blind zone is calculated. If the multi-directional detection confidence level within only a single angle range is less than the preset confidence threshold, the angle span of that angle range is taken as the angle span of the corresponding perception blind zone. Finally, the angle spans of each perception blind zone are compared, and the angle value corresponding to the perception blind zone with the largest angle span is taken as the maximum perception blind zone angle.

[0066] Subsequently, multi-directional target tracking is performed based on multi-directional detection confidence. Dynamic target tracking delay parameters are calculated. The start time of target appearance in the ground truth reference target is read according to the acquisition time, and multi-directional detection confidence is read separately for directional, lateral, and vertical sectors. For any sector, starting from the target appearance start time, the multi-directional detection confidence is checked sequentially at subsequent acquisition times. When the multi-directional detection confidence of a sector continuously reaches or exceeds a preset confidence threshold, and the position difference between the corresponding pose prediction value and the ground truth reference target is less than a preset position difference, the acquisition time that first meets the above conditions is taken as the tracking stabilization time of that sector. Next, the tracking stabilization time is subtracted from the target appearance start time to obtain the tracking delay of that sector. If no tracking stabilization time meets the conditions before the end of the test, the test end time is subtracted from the target appearance start time as the tracking delay of that sector. Finally, the tracking delays of the directional, lateral, and vertical sectors are calculated separately in the above manner, and the tracking delays of the three sectors are averaged to calculate the dynamic target tracking delay parameters.

[0067] Finally, the directional sensitivity matrix, maximum blind zone angle, and dynamic target tracking delay parameters are combined, and these parameters are read according to the test round and acquisition time. The directional sensitivity matrix is ​​then used as the covariance analysis result of the multi-directional error values, the maximum blind zone angle as the blind zone extreme value perception result corresponding to the multi-directional detection confidence, and the dynamic target tracking delay parameters as the multi-directional target tracking result, all arranged in a fixed field order. Finally, the directional sensitivity matrix, maximum blind zone angle, and dynamic target tracking delay parameters from the same test round are written together to obtain the multi-directional sensing test parameter set.

[0068] In summary, the embodiments of this application have at least the following technical effects: This application constructs test scenario parameters, performs real-time sensing on a multi-sensor combination of coal mine equipment, obtains multi-modal sensing data, performs synchronous denoising analysis on multi-source heterogeneous data, and generates a standard test dataset. Based on the standard test dataset, multi-level feature fusion is performed to construct a dynamic perception map of the coal mine tunneling environment, which includes a perception feature tensor. Multi-directional perception tests of the multi-sensor combination are conducted based on the perception feature tensor, a multi-directional perception test parameter set is calculated, and the multi-directional perception test parameter set is fed back to the multi-sensor combination for configuration adjustment, generating a coal mine tunneling perception test report. This invention solves the technical problem in the prior art where the configuration of multi-sensor combinations in a coal mine tunneling environment is difficult to adaptively optimize based on perception test results. By constructing a standard test dataset, fusing and generating a dynamic perception map of the coal mine tunneling environment, and performing multi-directional perception tests and feedback configuration adjustments based on the perception feature tensor, the technical effect of improving the accuracy and configuration adaptability of multi-sensor combinations in perceiving and testing the coal mine tunneling environment is achieved.

[0069] Example 2 is based on the same inventive concept as the coal mine tunneling sensing and testing method based on multi-sensor combination configuration in the previous examples, such as... Figure 3 As shown, this application provides a coal mine tunneling sensing and testing system based on a multi-sensor combination configuration. The system and method embodiments in this application are based on the same inventive concept. The system includes: The real-time sensing module 11 is used to construct test scenario parameters, configure multiple sensor combinations for real-time sensing of coal mine equipment, obtain multi-modal sensing data, perform multi-source heterogeneous data synchronous denoising analysis, and generate a standard test dataset. The perception map construction module 12 is used to perform multi-level feature fusion based on the standard test dataset to construct a dynamic perception map of the coal mine tunneling environment, which includes a perception feature tensor. The configuration adjustment module 13 is used to perform multi-directional perception tests of the multi-sensor combination according to the perception feature tensor, calculate the multi-directional perception test parameter set, feed the multi-directional perception test parameter set back to the multi-sensor combination for configuration adjustment, and generate a coal mine tunneling perception test report.

[0070] Furthermore, the system is also used to implement the following functions: The process involves retrieving coal mine tunneling tasks for analysis to obtain a set of tunneling analysis parameters, including three-dimensional geometric parameters, ventilation parameters, cutting head operating parameters, and coal and rock property parameters. Based on these parameters, dust diffusion analysis is performed to construct a dust concentration distribution field. Settlement simulation analysis is also conducted to construct a humidity simulation distribution field. Furthermore, coal impact response analysis is performed based on the cutting head operating parameters and coal and rock property parameters to construct a tunneling vibration time history curve. This time history curve is then dynamically coupled and mapped with the dust concentration distribution field to generate a dust re-entrainment coupling compensation field. Finally, the dust concentration distribution field, the humidity simulation distribution field, the tunneling vibration time history curve, and the dust re-entrainment coupling compensation field are spatiotemporally aligned and encapsulated to construct test scenario parameters.

[0071] Furthermore, the system is also used to implement the following functions: The vibration time history curve is locally calculated to obtain local base plate acceleration excitation parameters; based on the local base plate acceleration excitation parameters and the dust concentration distribution field, a critical peeling acceleration threshold is set for vibration excitation; the local base plate acceleration is calculated, and when the local base plate acceleration is greater than the critical peeling acceleration threshold, the dust re-eruption source term is activated; based on the dust concentration distribution field, the local dust deposition concentration value is calculated, and when the intensity value of the dust re-eruption source term is proportional to the vibration acceleration amplitude and the local dust deposition concentration value, the dust re-eruption coupling compensation field is constructed.

[0072] Furthermore, the system is also used to implement the following functions: Based on the standard test dataset, multidimensional analysis is performed to obtain a denoised point cloud parameter set and a millimeter-wave radar detection vector. Multi-scale dynamic voxel partitioning is then performed based on the denoised point cloud parameter set to extract local geometric features and intensity distribution features. Electromagnetic scattering analysis is performed based on the millimeter-wave radar detection vector to generate electromagnetic scattering features. Heterogeneous feature fusion is performed based on the local geometric features, intensity distribution features, and electromagnetic scattering features to construct a spatiotemporal perception feature stream. A conditional neural implicit field decoder is constructed to implicitly decode the spatiotemporal perception feature stream, constructing an implicit dynamic perception map. Perceptual analysis is performed on the standard test dataset based on the conditional neural implicit field decoder to calculate the perception feature tensor. The perception feature tensor is combined with the implicit dynamic perception map to construct the dynamic perception map of the coal mine tunneling environment.

[0073] Furthermore, the system is also used to implement the following functions: Density calculation is performed based on the denoised point cloud parameter set. The voxel resolution is set according to the point cloud density distribution parameters, including fine-grained and coarse-grained voxel resolutions. Distance division is performed based on the denoised point cloud parameter set to determine near-distance and far-distance regions. Geometric analysis is performed on the near-distance region according to the fine-grained voxel resolution to obtain a first geometric descriptor. Geometric analysis is also performed on the far-distance region according to the fine-grained voxel resolution to obtain a second geometric descriptor. The first and second geometric descriptors are then aggregated to generate the local geometric features.

[0074] Furthermore, the system is also used to implement the following functions: The voxelized point cloud center data and millimeter-wave radar detection point data are located, and the voxelized point cloud center data and the millimeter-wave radar detection point data are used as graph nodes to construct a dynamic connection graph. Based on the local geometric features and the electromagnetic scattering features, query vectors, key vectors, and value vectors are generated respectively, and cross-attention is performed to calculate cross-modal association weights, and graph edge weight values ​​are initialized. The dynamic connection graph is traversed to perform node motion consistency analysis, and the dynamic connection graph is updated and fused layer by layer according to the phase separation results and graph edge weight values ​​to obtain the spatiotemporal perception feature flow.

[0075] Furthermore, the system is also used to implement the following functions: A multi-directional sensing decoder is constructed, and the sensing feature tensor is input into the multi-directional sensing decoder for cross-attention decoding and prediction to generate multi-directional target pose prediction values ​​and multi-directional motion prediction values. A ground truth reference target is set based on the test scene parameters. The multi-directional target pose prediction values ​​and multi-directional motion prediction values ​​are compared with the ground truth reference target to calculate a multi-directional sensing test parameter set. The multi-directional sensing test parameter set is used as state feedback to perform deep reinforcement learning on a multi-sensor combination to construct configuration adjustment action parameters. The multi-sensor combination is adjusted online according to the configuration adjustment action parameters to generate a sensor configuration sequence. Multiple rounds of testing are conducted based on the multi-directional sensing test parameter set and the sensor configuration sequence to summarize and construct the coal mine tunneling sensing test report.

[0076] Furthermore, the system is also used to implement the following functions: The multi-directional sensing decoding head consists of multiple dedicated decoding branches for different directions. These dedicated decoding branches include directional sectors, lateral sectors, and vertical sectors. The multiple dedicated decoding branches use dot product attention to perform weighted readings on the sensing feature tensor to obtain the first pose prediction value and the first motion prediction value for the directional sector, the second pose prediction value and the second motion prediction value for the lateral sector, and the third pose prediction value and the third motion prediction value for the vertical sector.

[0077] Furthermore, the system is also used to implement the following functions: The first pose prediction value, first motion prediction value of the directional sector, the second pose prediction value, second motion prediction value of the lateral sector, the third pose prediction value, and third motion prediction value of the vertical sector are compared with the true reference target to obtain multi-directional error values. Covariance analysis is performed on the multi-directional error values ​​to construct a directional sensitivity matrix. Confidence calculation is performed based on the first pose prediction value, first motion prediction value of the directional sector, the second pose prediction value, second motion prediction value of the lateral sector, and third pose prediction value of the vertical sector to obtain a multi-directional detection confidence score. Blind zone extreme value perception is performed based on the multi-directional detection confidence score to calculate the maximum perception blind zone angle. Multi-directional target tracking is performed based on the multi-directional detection confidence score to calculate the dynamic target tracking delay parameter. The directional sensitivity matrix, the maximum perception blind zone angle, and the dynamic target tracking delay parameter are combined to obtain the multi-directional perception test parameter set.

[0078] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A coal mine tunneling sensing and testing method based on multi-sensor combination configuration, characterized in that, The method includes: Construct test scenario parameters, configure multi-sensor combinations for coal mine equipment to perform real-time sensing, obtain multi-modal sensing data, perform multi-source heterogeneous data synchronous denoising analysis, and generate a standard test dataset; Multi-level feature fusion is performed based on the standard test dataset to construct a dynamic perception map of the coal mine tunneling environment, which includes a perception feature tensor. Based on the perception feature tensor, a multi-sensor combination multi-directional perception test is performed, a multi-directional perception test parameter set is calculated, the multi-directional perception test parameter set is fed back to the multi-sensor combination for configuration adjustment, and a coal mine tunneling perception test report is generated.

2. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 1, characterized in that, The process of constructing test scenario parameters includes the following methods: The coal mine tunneling task is retrieved for tunneling analysis to obtain a tunneling analysis parameter set, which includes three-dimensional geometric parameters, ventilation parameters, cutting head working condition parameters, and coal and rock property parameters. Based on the three-dimensional geometric parameters, the ventilation parameters, and the cutting head operating parameters, dust diffusion analysis is performed to construct a dust concentration distribution field; Settlement simulation analysis was performed based on the three-dimensional geometric parameters, ventilation parameters, and cutting head operating parameters to construct a humidity simulation distribution field; Based on the cutting head working parameters and the coal and rock property parameters, the coal falling impact response analysis is carried out, the tunneling vibration time history curve is constructed, and the tunneling vibration time history curve is dynamically coupled and mapped with the dust concentration distribution field to generate a dust re-raising coupling compensation field. The dust concentration distribution field, the humidity simulation distribution field, the tunneling vibration time history curve, and the dust re-entrainment coupling compensation field are spatiotemporally aligned and encapsulated to construct test scenario parameters.

3. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 2, characterized in that, The method involves dynamically coupling and mapping the tunneling vibration time history curve with the dust concentration distribution field to generate a dust re-entrainment coupling compensation field, including: The vibration time history curve is locally calculated to obtain the local base plate acceleration excitation parameters. Based on the local base plate acceleration excitation parameters and the dust concentration distribution field, a critical peeling acceleration threshold is set for vibration excitation. Calculate the local base plate acceleration. When the local base plate acceleration is greater than the critical peeling acceleration threshold, activate the dust re-generating source term. Based on the dust concentration distribution field, the local dust deposition concentration value is calculated. When the intensity value of the dust re-eruption source term is proportional to the vibration acceleration amplitude and the local dust deposition concentration value, the dust re-eruption coupling compensation field is constructed.

4. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 1, characterized in that, Based on the aforementioned standard test dataset, a multi-level feature fusion is performed to construct a dynamic perception map of the coal mine tunneling environment. The method includes: Based on the standard test dataset, multidimensional analysis is performed to obtain the denoised point cloud parameter set and millimeter-wave radar detection vector; Based on the denoised point cloud parameter set, multi-scale dynamic voxel partitioning is performed to extract local geometric features and intensity distribution features. Electromagnetic scattering analysis is performed based on the millimeter-wave radar detection vector to generate electromagnetic scattering characteristics; Heterogeneous feature fusion is performed based on the local geometric features, the intensity distribution features, and the electromagnetic scattering features to construct a spatiotemporal sensing feature stream; A conditional neural implicit field decoder is constructed to implicitly decode the spatiotemporal sensing feature stream and construct an implicit dynamic sensing map. Based on the conditional neural implicit field decoder, perceptual analysis is performed on the standard test dataset to calculate the perceptual feature tensor; The perceptual feature tensor is combined with the implicit dynamic perceptual map to construct the dynamic perceptual map of the coal mine tunneling environment.

5. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 4, characterized in that, Based on the denoised point cloud parameter set, multi-scale dynamic voxel partitioning is performed to extract local geometric features. The method includes: Density calculation is performed based on the denoised point cloud parameter set, and the voxel resolution is set according to the point cloud density distribution parameters. The voxel resolution includes fine-grained voxel resolution and coarse-grained voxel resolution. Based on the denoised point cloud parameter set, distance division is performed to determine the near-distance region and the far-distance region; Perform geometric analysis on the near-field region according to the fine-grained voxel resolution to obtain a first geometric descriptor; Perform geometric analysis on the distant region according to the fine-grained voxel resolution to obtain a second geometric descriptor; The first geometric descriptor and the second geometric descriptor are aggregated to generate the local geometric features.

6. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 5, characterized in that, Heterogeneous feature fusion based on the local geometric features, the intensity distribution features, and the electromagnetic scattering features to construct a spatiotemporal sensing feature stream includes the following methods: The voxelized point cloud center data and millimeter-wave radar detection point data are located, and the voxelized point cloud center data and millimeter-wave radar detection point data are used as graph nodes to construct a dynamic connection graph; Based on the local geometric features and the electromagnetic scattering features, query vectors, key vectors and value vectors are generated respectively. Cross-attention calculation is performed to calculate cross-modal association weights, and graph edge weights are initialized. The dynamic connection graph is traversed to perform node motion consistency analysis. Based on the phase separation results and graph edge weight values, the dynamic connection graph is updated and fused layer by layer to obtain the spatiotemporal sensing feature flow.

7. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 1, characterized in that, The method includes performing multi-sensor combined multi-directional sensing tests based on the sensing feature tensor, calculating the multi-directional sensing test parameter set, feeding the multi-sensor combined multi-sensor configuration adjustment, and generating the coal mine tunneling sensing test report. Construct a multi-directional sensing decoding head, input the sensing feature tensor into the multi-directional sensing decoding head for cross-attention decoding prediction, and generate multi-directional target pose prediction values ​​and multi-directional motion prediction values; Based on the test scenario parameters, a true reference target is set, and the multi-directional perception test parameter set is calculated by comparing the multi-directional target pose prediction value, the multi-directional motion prediction value, and the true reference target. The multi-directional sensing test parameter set is used as state feedback to perform deep reinforcement learning on the multi-sensor combination to construct configuration adjustment action parameters; Adjust the action parameters according to the configuration to perform online parameter adjustment of the multi-sensor combination and generate a sensor configuration sequence; Based on the multi-directional sensing test parameter set and the sensor configuration sequence, multiple rounds of tests are conducted to summarize the results and construct the coal mine tunneling sensing test report.

8. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 7, characterized in that, The multi-directional sensing decoding head consists of multiple dedicated decoding branches for different directions. These dedicated decoding branches include directional sectors, lateral sectors, and vertical sectors. The multiple dedicated decoding branches use dot product attention to perform weighted readings on the sensing feature tensor to obtain the first pose prediction value and the first motion prediction value for the directional sector, the second pose prediction value and the second motion prediction value for the lateral sector, and the third pose prediction value and the third motion prediction value for the vertical sector.

9. The coal mine tunneling sensing and testing method based on multi-sensor combination configuration as described in claim 8, characterized in that, The multi-directional perception test parameter set is calculated by comparing the predicted multi-directional target pose value, the predicted multi-directional motion value, and the true reference target. The method includes: The first pose prediction value and the first motion prediction value of the directional sector, the second pose prediction value and the second motion prediction value of the horizontal sector, the third pose prediction value and the third motion prediction value of the vertical sector are compared with the true reference target to obtain multi-directional error values. Covariance analysis was performed on the multi-directional error values ​​to construct a directional sensitivity matrix; Confidence calculations are performed based on the first pose prediction value and the first motion prediction value of the directional sector, the second pose prediction value and the second motion prediction value of the horizontal sector, and the third pose prediction value and the third motion prediction value of the vertical sector to obtain the multi-directional detection confidence. Based on the multi-directional detection confidence, blind zone extreme value perception is performed, and the maximum perception blind zone angle is calculated. Multi-directional target tracking is performed based on the multi-directional detection confidence, and dynamic target tracking delay parameters are calculated. The directional sensitivity matrix, the maximum perception blind zone angle, and the dynamic target tracking delay parameter are combined to obtain the multi-directional perception test parameter set.

10. A coal mine tunneling sensing and testing system based on a multi-sensor combination configuration, characterized in that, The system is used to execute the coal mine tunneling sensing test method based on a multi-sensor combination configuration as described in any one of claims 1-9, and the system includes: The real-time sensing module is used to construct test scenario parameters, configure multiple sensor combinations for real-time sensing of coal mine equipment, obtain multi-modal sensing data, perform synchronous noise reduction analysis of multi-source heterogeneous data, and generate a standard test dataset. The perception map construction module is used to perform multi-level feature fusion based on the standard test dataset to construct a dynamic perception map of the coal mine tunneling environment, wherein the dynamic perception map of the coal mine tunneling environment contains a perception feature tensor. The configuration adjustment module is used to perform multi-directional sensing tests of multiple sensors based on the sensing feature tensor, calculate the multi-directional sensing test parameter set, feed the multi-directional sensing test parameter set back to the multi-sensor combination for configuration adjustment, and generate a coal mine tunneling sensing test report.

Images