An intelligent early warning method and system for underground mine vehicle collision prevention

Abstract

The application discloses an underground mine vehicle anti-collision intelligent early warning method and system, and relates to the underground mine vehicle field.The method comprises the following steps: deploying a composite detection array comprising a differential pressure sensor and an acoustic vector sensor at the edge of the rock wall of the roadway intersection of the underground mine; collecting the air pressure wave generated by the blind area vehicle expelling air through the differential pressure sensor and extracting the pressure climbing gradient; synchronously collecting the diffraction sound wave formed by the exhaust noise bypassing the edge of the rock wall through the acoustic vector sensor and extracting the continuous change amount of the Doppler frequency shift; inputting the two into a pre-trained motion state analysis model, outputting the absolute approaching distance and relative motion speed of the blind area vehicle, and executing anti-collision early warning intervention when the preset threshold is met.The application solves the problem that the traditional sensor fails in the complex blind area corner of the underground mine, and the multi-path fading leads to extremely unreliable ranging and easily causes sudden collision.

Description

Technical Field

[0001] This invention relates to the field of underground mining vehicles, and more specifically, to an intelligent early warning method and system for collision avoidance in underground mining vehicles. Background Technology

[0002] As the core area for resource extraction, underground mines typically lack open, continuous road environments. Instead, they consist of narrow, semi-enclosed, and multi-branched complex spaces comprised of main haulage roadways, connecting roadways, mining roadways, return air roadways, and multiple intersections. With the continuous improvement of mechanization and intelligence in mining, the density of heavy-duty trackless rubber-tired vehicles, loaders, transport vehicles, and various mining equipment operating underground has significantly increased. Consequently, blind spots between vehicles and between vehicles and personnel at curves, T-junctions, crossroads, and areas where multiple passageways intersect have also increased. If vehicles in underground mines suddenly meet or laterally cut into each other at intersections, limited by roadway width, vehicle weight, braking distance, and driver reaction time, accidents such as scrapes, rear-end collisions, side impacts, and even injuries or fatalities are highly likely.

[0003] Existing mine vehicle safety management technologies have proposed various monitoring, indication, and protection measures for underground vehicle operation. For example, the patent "Monitoring System for Auxiliary Transportation Crossings of Rubber-Tiered Vehicles in Coal Mines" (publication number CN103334792A) discloses a monitoring system installed at curves, T-junctions, crossroads, and complex multi-channel crossings. The system consists of intrinsically safe vehicle detection sensors, explosion-proof and intrinsically safe LED signal indicator screens, and an explosion-proof PLC control box. It manages the passage of rubber-tiered vehicles underground through vehicle detection and signal indication, thereby improving the safety and efficiency of underground auxiliary transportation. Another example is the patent "Stalling Protection Device for Trackless Rubber-Tiered Vehicles" (publication number CN113565562A), which discloses a device that detects the speed of trackless rubber-tiered vehicles using radar speed sensors and sends the data to a controller in real time. When a vehicle stalls, it works in conjunction with arresting cables, arresting blocks, hooks, and alarms to intercept and protect the vehicle, improving the ability to handle stalled vehicles.

[0004] The aforementioned existing technologies can improve underground vehicle traffic order, stall protection, and local safety management to some extent. However, they primarily focus on traffic control after vehicles pass through detection points, interception and protection after abnormal speeds, or traffic management relying on methods such as geomagnetism, radar, communication, and traffic lights. For vehicles in blind spots obscured by rock walls at intersections, especially before a vehicle has left the corner and entered the detection range of conventional line-of-sight sensors, existing technologies still struggle to accurately determine how far the vehicle is from the intersection, whether it is rapidly approaching, and whether a collision risk has been established.

[0005] The unique characteristic of underground mines lies in the fact that their tunnels are not ordinary roads, but rather resemble natural, irregular tubular structures. The tunnel cross-sections are limited, the rock walls extend continuously, and localized areas include ventilation doors, bends, branch tunnels, ramps, and equipment chambers. When large vehicles travel at high speeds within these tunnels, the air in front of the vehicle cannot freely dissipate like on an open road. Instead, it is propelled by the vehicle's windward side and propagates along the tunnel direction, creating a significant aerodynamic piston effect. This phenomenon is typically less pronounced in open surface environments because air has ample space for diffusion; similarly, in typical factory or road scenarios, it is rare to simultaneously present long, enclosed boundaries, narrow cross-sections, and the high duty cycle of heavy vehicles. In contrast, the semi-enclosed structure of underground mine tunnels makes the air pressure disturbances caused by vehicle movement more pronounced. Furthermore, these disturbances are superimposed on factors such as the mine's ventilation system, the opening and closing of ventilation doors, and localized air leaks, creating a complex aerodynamic background.

[0006] Meanwhile, underground mines are environments characterized by high noise, strong reflection, and strong obstruction. Heavy machinery, fans, belt conveyors, drilling rigs, and diesel engines continuously generate high-decibel noise, and the narrow rock walls cause prolonged reverberation, making it difficult for ordinary microphone arrays to stably locate sound sources. Sensors that rely on line-of-sight or near-line-of-sight propagation, such as cameras, lidar, and millimeter-wave radar, are completely blocked at rock wall corners; although conventional radio frequency signals can be diffracted or reflected to some extent, they are prone to severe multipath propagation in the long corridor-like spaces of mine tunnels, causing drift in ranging and orientation determination. In other words, the problem at mine intersections is not simply a matter of not being able to see, but rather a multi-physics coupling problem involving rock wall obstruction, wireless multipath propagation, acoustic reverberation, air pressure fluctuations, and ventilation disturbances.

[0007] Therefore, the key challenge facing existing mine vehicle collision avoidance technology lies in the fact that, before a vehicle is in the blind spot and has moved out of the rock wall's obstruction, the system struggles to reliably distinguish between a truly approaching vehicle and environmental interference such as ventilation, noise, and reflections from complex environmental signals. It also struggles to simultaneously obtain the absolute approach distance and relative speed of the vehicle in the blind spot. If an alarm is only triggered after the vehicle enters the line of sight, it is often already close to the intersection area, leaving insufficient time for the driver or automatic braking system to react. If an alarm is triggered in advance based solely on a single sensor or simple threshold, it is easy to mistake the opening and closing of ventilation doors, fan pressure fluctuations, or mechanical background noise for a vehicle approaching, resulting in frequent false alarms and impacting underground transportation efficiency. How to accurately perceive the distance and speed of vehicles in the blind spot, considering the unique piston effect, sound propagation characteristics at the rock wall edges, and the strong reverberation, multipath, and obstruction environments of confined roadways in underground mines, remains a crucial problem that existing technologies must address. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide an intelligent early warning method and system for preventing collisions of vehicles in underground mines, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent early warning method for collision avoidance in underground mine vehicles includes the following steps: A composite detection array containing micro differential pressure sensors and acoustic vector sensors is deployed at the rock wall edge of the roadway intersection in an underground mine. The differential pressure sensor collects the air pressure wave generated by the vehicle displacing air while driving in a restricted roadway in the blind spot, and extracts the pressure rise gradient of the air pressure wave. The acoustic vector sensor synchronously collects the diffracted sound waves generated by the exhaust noise of the vehicle in the blind spot as it bypasses the edge of the rock wall, and extracts the continuous change in the Doppler frequency shift of the diffracted sound waves. The pressure rise gradient and the continuous change in Doppler frequency shift are input into a pre-trained motion state analysis model, and the absolute approach distance and relative speed of the vehicle in the blind spot are output through the motion state analysis model. When the absolute approach distance is less than a preset first distance threshold and the relative motion speed is greater than a preset first speed threshold, a collision avoidance warning intervention is executed.

[0010] Specifically, the training process of the motion state analysis model includes: The sample pressure rise gradient of the sample air pressure wave, the continuous change of the sample Doppler frequency shift of the sample diffracted sound wave, and the true approach distance and true speed of the sample vehicle corresponding to the sample air pressure wave and the sample diffracted sound wave are obtained. The sample pressure rise gradient and the sample Doppler frequency shift change are used as the input sequence of the motion state analytical model, and the true approximation distance and the true motion velocity are used as the supervision labels of the motion state analytical model. A loss function is constructed based on the first difference between the predicted approximation distance and the actual approximation distance output by the motion state analysis model, and the second difference between the predicted motion speed and the actual motion speed. The network parameters of the motion state analysis model are updated by minimizing the loss function until a preset convergence condition is met.

[0011] Specifically, the step of extracting the pressure rise gradient of the pressure wave includes: The initial air pressure signal acquired by the differential pressure sensor is low-pass filtered to remove signal components with frequencies greater than a preset reference frequency threshold, while retaining the quasi-static air pressure wave signal. The pressure rise gradient of the pressure wave is obtained by taking the derivative of the pressure wave signal in the time dimension.

[0012] Specifically, the step of extracting the continuous change in the Doppler frequency shift of the diffracted sound wave includes: The initial acoustic signal acquired by the acoustic vector sensor is bandpass filtered to retain the principal component frequency band whose frequency is between a preset first frequency band and a preset second frequency band, thereby obtaining the diffracted sound wave; The time-frequency spectrum of the diffracted sound wave is extracted using Fast Fourier Transform, and the trajectory of the main peak frequency in the time-frequency spectrum is traced with time to obtain the continuous change of the Doppler frequency shift.

[0013] Specifically, before the step of inputting the pressure rise gradient and the continuous change in Doppler frequency shift into the pre-trained analytical model of motion state, the method further includes a cross-correlation verification step, specifically including: Obtain the initial change time of the pressure rise gradient and the initial change time of the continuous change in the Doppler frequency shift; Calculate the time difference between the initial change time of the pressure rise gradient and the initial change time of the continuous change in the Doppler frequency shift; If the time difference is less than a preset time matching threshold, it is determined that the air pressure wave and the diffracted sound wave originate from the same vehicle in the blind zone, and the step of inputting the pressure rise gradient and the continuous change in Doppler frequency shift into the pre-trained motion state analytical model is executed.

[0014] Specifically, after the step of extracting the pressure rise gradient of the pressure wave, the method further includes an acoustic-aerophysical field interlocking gating step, specifically including: Determine whether the pressure ramp-up gradient is greater than a preset gradient trigger threshold; When the pressure rise gradient is greater than the gradient trigger threshold, the air pressure wave is used as a gating signal to activate the acoustic feature matching module and open a judgment time window with a preset duration. Within the determination time window, the acoustic feature matching module is used to detect whether the initial acoustic signal collected by the acoustic vector sensor contains a continuous harmonic frequency characterizing mechanical operation. If the continuous harmonic frequency is not detected within the determination time window, the pressure wave is determined to be a ventilation aerodynamic fluctuation interference caused by the opening and closing of the ventilation duct damper or the operation of the fan, and the anti-collision warning intervention is terminated.

[0015] Specifically, the step of using the acoustic feature matching module to detect whether the initial acoustic signal acquired by the acoustic vector sensor contains continuous harmonic frequencies characterizing mechanical operation includes: The initial acoustic signal is converted into a frequency energy distribution sequence; The frequency energy distribution sequence is matched with a pre-constructed mechanical noise frequency template library using cosine similarity. The mechanical noise frequency template library contains multiple standard harmonic frequency distribution templates generated by known diesel engines and electric motors. If the maximum cosine similarity between the calculated frequency energy distribution sequence and each of the standard harmonic frequency distribution templates is less than a preset similarity judgment threshold, the judgment result that the continuous harmonic frequency was not detected is output.

[0016] Specifically, in scenarios involving intersections with multiple blind side tunnels, the acoustic vector sensor is a miniature microphone array containing at least three microphone elements. The method further includes a phase direction finding step based on secondary sound sources at the rock wall edge, specifically including: Obtain the fixed spatial physical coordinates of the rock wall edges of each blind zone branch tunnel as a secondary sound source; The miniature microphone array is used to capture the superimposed diffraction sound waves generated and superimposed by multiple vehicles in the blind spot distributed in different blind spot side lanes; Calculate the signal phase difference sequence between the diffracted sound waves in the superimposed state and each of the microphone array elements; The signal phase difference sequence and the fixed spatial physical coordinates are input into a pre-established spatial sound source decoupling model, and the spatial sound source decoupling model outputs the target blind zone branch alleyway location of each of the blind zone vehicles that generate the pressure wave.

[0017] Specifically, the pre-established spatial sound source decoupling model is a mapping model based on feature distance matching. The step of outputting the target blind zone branch alleyway location of each blind zone vehicle that generates the pressure wave through the spatial sound source decoupling model includes: Obtain the known spatial distance from each of the fixed spatial physical coordinates to each of the microphone array elements; The known spatial distance is input into a preset acoustic phase propagation model to pre-construct a theoretical phase difference template set corresponding to all the secondary sound sources; Multiple discrete peak phase difference features are extracted from the signal phase difference sequence; Each of the peak phase difference features is matched with each template in the theoretical phase difference template set by feature distance, and the matching distance is calculated. The blind zone branch lane where the secondary sound source corresponding to the template with the smallest matching distance is located is selected as the target blind zone branch lane location of each blind zone vehicle.

[0018] This invention also discloses an intelligent early warning system for preventing collisions with underground mine vehicles to implement the above method, comprising: A composite detection array module is deployed at the edge of the rock wall at the intersection of roadways in an underground mine. The composite detection array module includes a micro differential pressure sensor and an acoustic vector sensor. The micro differential pressure sensor is used to collect the air pressure wave generated by the air displaced by vehicles traveling in the blind spot in the restricted roadway. The acoustic vector sensor is used to simultaneously collect the diffracted sound wave formed when the exhaust noise generated by the vehicles in the blind spot bypasses the edge of the rock wall. The signal feature extraction module, connected to the composite detection array module, is used to extract the pressure rise gradient of the pressure wave and the continuous change of the Doppler frequency shift of the diffracted sound wave. The motion state analysis module, connected to the signal feature extraction module, is used to input the pressure rise gradient and the continuous change of the Doppler frequency shift into the pre-trained motion state analysis model, and output the absolute approximation distance and relative speed of the vehicle in the blind spot through the motion state analysis model. The early warning intervention execution module is connected to the motion state analysis module and is used to perform anti-collision early warning intervention when the absolute approach distance is less than a preset first distance threshold and the relative motion speed is greater than a preset first speed threshold.

[0019] The advantages of this invention compared to existing technologies lie in its ability to overcome the limitations of traditional line-of-sight and radio frequency sensors. It creatively utilizes the cross-physical field synergy of aerodynamic piston effect and acoustic edge diffraction to solve the problem of extremely unreliable ranging and speed measurement at complex blind spot corners. By deploying a composite detection array to extract the pressure rise gradient of the air pressure wave generated by vehicle exhaust, the system can accurately pinpoint the absolute approach distance of vehicles in blind spots. Simultaneously, by extracting the Doppler frequency shift characteristics of the diffracted sound waves formed by exhaust noise around rock wall corners, the relative speed of vehicles can be accurately locked. This cross-correlation verification mechanism completely eliminates electromagnetic multipath and acoustic reverberation interference in tunnels, achieving not only advanced and accurate vehicle detection in blind spots but also possessing extremely high environmental adaptability and perception purity, effectively preventing fatal sudden collisions at intersections.

[0020] Furthermore, addressing the issue of false alarms generated by conventional air pressure monitoring due to the complex ventilation network within mines (such as sudden opening and closing of anti-outburst doors or periodic fluctuations in main fan pressure), this invention ingeniously introduces an acoustic-aerodynamic physical field interlocking gating mechanism. Utilizing the physical principle that vehicles simultaneously act as both a source of air pressure thrust and a continuous source of mechanical sound, the system, upon detecting a sharp rise in air pressure, treats it solely as a gating signal. Within a very short time window, it detects whether the acoustic signal contains continuous harmonic frequencies characterizing mechanical operation. This technical feature accurately identifies and filters out aerodynamic fluctuation interference from the ventilation network without relying on complex air pressure waveform filtering algorithms, completely resolving the "ghost vehicle" problem caused by frequent false alarms and ensuring normal transportation efficiency in the mine.

[0021] Furthermore, in intersections with multiple blind-spot side lanes, the piston-pressure waves generated by multiple vehicles approaching simultaneously undergo hydrodynamic superposition, making it easy for the system to lack spatial directionality if relying solely on air pressure. This invention utilizes the phase direction-finding law based on secondary sound sources at the rock wall edge, perfectly entrusting the direction-finding function to the acoustic diffraction field. A miniature microphone array captures the superimposed diffracted sound waves, and combined with the fixed spatial physical coordinates of the rock wall edge as a secondary sound source, the signal phase difference sequence is input into a spatial sound source decoupling model. Even in extreme cases of multiple superimposed air pressure waves, this technique can still clearly isolate and indicate the specific location of vehicles in each blind-spot side lane, effectively solving the problem of the system's inability to provide drivers with accurate avoidance direction guidance. Attached Figure Description

[0022] Figure 1 This is a global working scene diagram of the present invention; Figure 2 This is a schematic diagram of the installation of the composite detection array module of the present invention; Figure 3 This is a schematic diagram of the working principle of the motion state analysis module of the present invention; Figure 4 This is a schematic diagram of the time-frequency and filtering methods for signal feature extraction in this invention; Figure 5 This is a schematic diagram of the anti-interference scenario of the sound and air physical field interlocking gate control of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0024] This invention provides an intelligent early warning method for vehicle collision prevention in underground mines, applicable to main haulage roadways, connecting roadways, mining roadways, return air roadways, and bends, T-junctions, crossroads, or multi-channel intersections formed by the convergence of multiple branch roadways, such as... Figure 1 As shown. This method does not rely on the vehicle having entered the camera, radar, or driver's line of sight. Instead, it utilizes the air displacement effect and mechanical sound edge diffraction effect that inevitably occur when vehicles travel in the confined tunnels of underground mines. While the vehicle is still in the blind spot behind the rock wall, it analyzes the vehicle's approach to the intersection and its relative speed in advance, thereby performing early warning intervention before a collision risk occurs.

[0025] During implementation, a composite detection array is installed at the intersection of underground mine roadways near the edge of the rock wall, such as... Figure 2As shown, the edge of the rock face refers to the corner boundary area that obstructs the view between the main tunnel and the side tunnel. It is also the location where the air pressure waves generated by vehicles in the blind spot propagate to the vicinity of the intersection first. At the same time, it is also the location where vehicle exhaust noise, engine noise, or motor noise is more easily captured after edge diffraction. Placing a composite detection array in this area can simultaneously obtain aerodynamic and acoustic information, avoiding the problem of complete obstruction by the rock face when relying solely on line-of-sight sensors.

[0026] The composite detection array includes a differential pressure sensor and an acoustic vector sensor. The differential pressure sensor is used to detect small, low-frequency, quasi-static pressure changes in the tunnel, with a range selectable from ±100Pa to ±1000Pa, a resolution selectable from 0.05Pa to 0.5Pa, and a sampling frequency selectable from 20Hz to 200Hz. In underground mines, vehicles, especially heavy-duty trackless rubber-tired vehicles, loaders, or transport vehicles, occupy a large cross-sectional area when traveling in narrow tunnels. Unlike on open roads, the air in front of the vehicle cannot diffuse freely in all directions; instead, it is pushed by the vehicle's frontal surface and propagates along the tunnel direction, forming a pressure wave similar to a piston pushing air. This pressure wave can reach the vicinity of intersections before the vehicle leaves the blind spot, thus serving as a physical basis for early detection of approaching vehicles in the blind spot.

[0027] Acoustic vector sensors are used to synchronously acquire acoustic signals from vehicles in blind spots. These sensors can be integrated acoustic sensors with both sound pressure and particle velocity channels, or they can utilize compact multi-microphone arrays to achieve equivalent vector acquisition. The sampling frequency can be selected from 8kHz to 48kHz. Vehicles in blind spots typically generate engine exhaust noise, motor rotation noise, and friction noise between tires and the tunnel floor or rail surface. Exhaust noise and mechanical operating noise have relatively stable dominant frequencies and harmonic structures. When a vehicle is located in a side tunnel behind a rock wall, sound waves cannot reach the intersection along a completely straight path but will diffract at the edge of the rock wall, which can be approximated as a secondary sound source. The acoustic vector sensor acquires precisely this diffracted sound wave that bypasses the rock wall edge. Through this diffracted sound wave, the continuous change in Doppler frequency shift related to the relative motion of the vehicle can be extracted.

[0028] To ensure a unified time reference for both the air pressure and acoustic signals, the differential pressure sensor and acoustic vector sensor in the composite detection array are triggered for sampling by the same acquisition controller. This controller can employ an intrinsically safe edge computing unit and uses a unified clock to timestamp the air pressure and acoustic channels. The time synchronization error should ideally be controlled within 1ms to 20ms. This setup is designed to prevent misinterpretation of ventilation fluctuations and ambient noise as vehicle signals if the time references are inconsistent, in case the air pressure wave and diffracted sound wave originate from the same vehicle in the blind spot.

[0029] This invention first uses a differential pressure sensor to collect the air pressure wave generated by a vehicle moving through a confined roadway in a blind zone, displacing air. The collected initial air pressure signal typically includes background pressure changes in the mine ventilation system, transient disturbances caused by the opening and closing of dampers, periodic fluctuations of the fan, and high-frequency noise from the sensor itself. To highlight the quasi-static air pressure wave caused by vehicle movement, the initial air pressure signal can be low-pass filtered to remove signal components with frequencies higher than a reference frequency threshold, such as... Figure 4 As shown. The reference frequency threshold can be set from 0.5Hz to 10Hz depending on the tunnel length, vehicle speed, and ventilation background, with 2Hz to 5Hz being the most commonly used values. The pressure wave signal retained after low-pass filtering mainly reflects the slow ascent or descent process of the air in the tunnel as a whole is pushed, rather than the high-frequency disturbances caused by local mechanical vibrations.

[0030] After obtaining the pressure wave signal, the rate of pressure change is calculated along the time dimension to obtain the pressure rise gradient. The pressure rise gradient can be understood as the rate at which air pressure rises per unit time. As a vehicle in the blind spot approaches the intersection, the compressed and propelled air in front of the vehicle has a more significant impact on the area near the intersection, and the pressure wave rise is usually steeper. Therefore, compared to simply using the pressure value at a single moment, the pressure rise gradient better reflects the approaching trend of the vehicle and also reduces the influence of differences in basic wind pressure in different roadways on the judgment result. The output unit of the pressure rise gradient can be Pa / s. In practical applications, it can be calculated within a sliding time window, with a sliding time window length ranging from 0.2s to 2s, to balance response speed and noise resistance.

[0031] Simultaneously, initial acoustic signals generated by vehicles in the blind spot are acquired using acoustic vector sensors. These initial acoustic signals simultaneously include vehicle noise, fan noise, conveyor belt noise, drilling rig noise, personnel work noise, and reverberation from rock wall reflections. To extract diffracted sound waves more closely related to vehicle motion, bandpass filtering can be applied to the initial acoustic signals to retain the principal component frequency bands where vehicle mechanical noise is concentrated, such as... Figure 4 As shown. The lower limit of this principal component frequency band can be selected between 20Hz and 100Hz, and the upper limit can be selected between 1000Hz and 4000Hz. For mines with a large number of diesel vehicles, low and mid-frequency exhaust noise and engine harmonics in the range of 30Hz to 2000Hz can be retained first; for mines with a large number of electric vehicles, the upper limit of the frequency band can be appropriately increased to 3000Hz or 4000Hz to retain stable frequency components formed by motor meshing, inverter drive, and tire friction.

[0032] The diffracted sound wave obtained after bandpass filtering is further processed by frame-based fast Fourier transform or short-time fast Fourier transform to generate a time-frequency spectrum, such as... Figure 4As shown, the time-spectrum diagram reflects the locations of frequencies with stronger energy within different time slices. Because of the relative velocity between the vehicle's sound source and the receiving position when the vehicle moves relative to the intersection in the blind spot, the main peak frequency in the diffracted sound wave will continuously shift over time; this frequency shift is called the Doppler shift. The system tracks the trajectory of the main peak frequency change over time in the time-spectrum diagram, and uses the main peak frequency shift between adjacent time slices, the direction of the main peak frequency shift, and the continuity of the shift as the continuous change in Doppler shift. This continuous change does not only take the frequency difference at a single instant, but records the continuous evolution of frequency from high to low or from low to high over a period of time, thus having a stronger ability to suppress occasional noise spikes in the mine reverberation environment.

[0033] After extracting the pressure rise gradient and the continuous change in Doppler frequency shift, these two values ​​are input into a pre-trained motion state analytical model, such as... Figure 3 As shown, this model is used to establish the mapping relationship between aerodynamic characteristics, acoustic motion characteristics, and the actual motion state of vehicles in blind spots. The absolute approximation distance here refers to the distance along the roadway between the front of the vehicle or its equivalent leading edge and the safety boundary of the intersection. It can also be defined as the distance between the vehicle's leading edge and the vertical position of the rock wall edge, depending on mine management needs. The relative motion speed refers to the velocity component of the vehicle approaching the intersection along the branch roadway, with the direction towards the intersection taken as positive, rather than the arbitrary speed displayed on the vehicle's dashboard.

[0034] In one embodiment, the motion state analysis model employs a dual-channel temporal neural network structure, such as... Figure 3 As shown, the first input channel is the pressure rise gradient sequence, and the second input channel is the continuous Doppler frequency shift sequence. Each input sequence covers the same sliding time window, with a window length of 1 to 5 seconds and a step size of 0.1 to 1 second. The two channels are processed by a one-dimensional convolutional layer to extract local temporal variation features, and then by a gated recurrent unit, a long short-term memory network, or a lightweight temporal attention encoder to extract continuous motion trends. The features from the two channels are then fused and fed into a fully connected regression layer, outputting the absolute approximation distance and relative motion velocity, respectively. To adapt to downhole edge computing equipment, the number of model parameters can be controlled between 0.1M and 5M, and the single inference time should be controlled within 10ms to 200ms.

[0035] In another embodiment, the motion state analytical model can also employ gradient boosting trees, random forest regressors, or support vector regressors. In this case, the maximum value, average value, rise duration, and rate of change of the rise slope of the pressure rise gradient, along with the average offset, maximum offset, offset direction stability, and main peak energy continuity of the Doppler frequency shift trajectory, can be used to construct a feature vector. The regression model then outputs the absolute approximation distance and relative motion velocity. This method requires less computation and is suitable for older mine monitoring systems with limited computing power; a dual-channel temporal neural network is more suitable for mines with complex vehicle types, complex roadway structures, and sufficient sample data.

[0036] The motion state analysis model needs to be trained before it can be put into use. During training, sample data is collected at different roadway cross sections, different branch roadway lengths, different vehicle types, and different ventilation conditions. Each set of samples includes the pressure rise gradient of the sample air pressure wave, the continuous change of the Doppler frequency shift of the sample diffracted sound wave, and the actual approach distance and actual speed of the sample vehicle corresponding to that set of air pressure waves and diffracted sound waves. The actual approach distance can be obtained by combining laser ranging calibration devices temporarily deployed in the roadway, UWB positioning tags, vehicle odometers, and manual calibration points; the actual speed can be obtained by time calculation through vehicle wheel speed sensors, inertial measurement units, radar speedometers, or calibration intervals. The training samples should cover an approach distance range of 0m to 80m and a vehicle speed range of 0.5m / s to 12m / s to ensure that the model can cover different scenarios such as low-speed passing, normal transportation, and relatively fast approach in the mine.

[0037] During training, the sample pressure gradient and the continuous change in sample Doppler frequency shift are used as the model input sequence, while the actual approximation distance and actual speed are used as supervision labels. After the model outputs the predicted approximation distance and predicted speed, the first difference between the predicted approximation distance and the actual approximation distance, and the second difference between the predicted speed and the actual speed are calculated respectively. The loss function is composed of the first and second differences, and a network parameter regularization term can be added to avoid the model overfitting to a certain roadway or a certain type of vehicle. The weights of distance error and speed error can be set according to actual safety requirements. For example, the distance error weight can be 0.5 to 0.8, and the speed error weight can be 0.2 to 0.5. When the mine pays more attention to the early braking distance, the distance error weight can be increased; when the risk of vehicles cutting in laterally at mine intersections is high, the speed error weight can be increased.

[0038] Network parameters can be updated using gradient descent-type optimization algorithms, such as the Adam optimizer or stochastic gradient descent optimizer. The learning rate can be between 0.0001 and 0.001, the batch size between 16 and 256, and the number of training epochs between 50 and 500. Preset convergence criteria can be either a validation set loss decreasing by a preset amount for 10 to 30 consecutive training epochs, or a validation set distance error of less than 1m to 3m and a velocity error of less than 0.2m / s to 0.8m / s. After model training, it is deployed to the intersection edge computing unit, which receives the air pressure and acoustic features output by the composite detector array in real time and performs online inference.

[0039] Once the motion state analysis model outputs the absolute approach distance and relative speed of the vehicle in the blind spot, the system executes a collision avoidance warning intervention based on safety thresholds, such as... Figure 3 As shown. The first distance threshold can be set from 5m to 50m based on the intersection width, vehicle braking performance, tunnel gradient, and vehicle speed limit, with commonly used values ​​ranging from 10m to 30m. The first speed threshold can be set from 0.5m / s to 5m / s, with commonly used values ​​ranging from 1m / s to 3m / s. When the absolute approach distance is less than the first distance threshold and the relative speed is greater than the first speed threshold, it indicates that the vehicle in the blind spot has entered the dangerous approach range and is still moving towards the intersection. At this time, collision avoidance warning intervention is executed. Collision avoidance warning intervention can include intersection audible and visual alarms, LED directional indicators, on-board terminal voice reminders, dispatch platform pop-ups, level crossing signal light switching, vehicle speed limit command issuance, and linkage with the automatic braking system. For driverless or remotely driven mining trucks, deceleration, stopping, or yielding commands can also be sent to the vehicle controller.

[0040] To avoid mismatch between air pressure signals and acoustic signals, this invention includes a cross-correlation verification step before inputting the motion state analysis model. The system acquires the initial change time of the pressure rise gradient and the initial change time of the continuous change in Doppler frequency shift. The initial change time of the pressure rise gradient can be defined as the first time point after multiple consecutive sampling points of the pressure rise gradient exceed a certain multiple of the background mean; the initial change time of the continuous change in Doppler frequency shift can be defined as the first time point after the main peak frequency trajectory begins to show a stable directional shift and continues for more than a preset time. The time difference between the two initial change times is then calculated. The time matching threshold can be between 0.1s and 2s, preferably between 0.3s and 1s. If the time difference is less than the time matching threshold, it is determined that the air pressure wave and the diffracted sound wave originate from the same vehicle in the blind zone, and then the pressure rise gradient and the continuous change in Doppler frequency shift are input into the motion state analysis model. If the time difference exceeds the time matching threshold, it is considered that the two types of signals may originate from different physical sources, such as the accidental superposition of ventilation pressure waves and distant mechanical noise. In this case, distance-velocity analysis is not triggered, or the confidence level of the analysis result is reduced.

[0041] In a further embodiment, the present invention includes an acoustic-aerodynamic physical field interlocking gating step to eliminate interference from the ventilation duct network, such as... Figure 5 As shown. In underground mines, sudden opening or closing of anti-outburst ventilation doors, periodic changes in the main fan pressure, starting and stopping of local ventilation fans, and air leakage in the roadway can all cause significant air pressure fluctuations. If the system alarms whenever it detects a pressure rise, it will create ghost vehicles. Therefore, this invention utilizes the physical characteristics of vehicles as both a source of air pressure thrust and a source of continuous mechanical sound. That is, when a real vehicle approaches, it not only pushes air to form an air pressure wave, but also produces continuous harmonic noise from the engine, motor, or transmission system; while simple opening and closing of ventilation doors or fan pressure fluctuations usually lack continuous mechanical harmonics synchronized with vehicle movement.

[0042] In practical implementation, after extracting the pressure rise gradient of the air pressure wave, the system first determines whether the pressure rise gradient is greater than the gradient trigger threshold. The gradient trigger threshold can be set from 0.1 Pa / s to 20 Pa / s, preferably from 0.5 Pa / s to 5 Pa / s, and the specific value can be calibrated according to the ventilation background of the tunnel and the size of the vehicle. If the pressure rise gradient does not exceed the gradient trigger threshold, it means that the current air pressure change is insufficient to indicate that a vehicle is approaching, and the system maintains monitoring status. If the pressure rise gradient exceeds the gradient trigger threshold, the air pressure wave is used as a gating signal to activate the acoustic feature matching module and open a judgment time window, such as... Figure 5 As shown. The judgment time window can be 0.5s to 5s, preferably 1s to 3s. By using the air pressure wave only as a gating signal and not directly as an alarm basis, it is possible to avoid misjudging any pressure disturbance as a vehicle.

[0043] Within the decision time window, the acoustic feature matching module detects whether the initial acoustic signal collected by the acoustic vector sensor contains continuous harmonic frequency components characterizing mechanical operation. Continuous harmonic frequency components refer to frequency components that repeatedly appear in multiple consecutive time slices, have a certain regular frequency interval, and whose energy does not disappear instantaneously with a single impact. For example, a diesel engine may exhibit its fundamental frequency and its harmonics, while an electric motor may exhibit its rotational frequency, electromagnetic noise frequency, and its related harmonics. If no continuous harmonic frequency components are detected within the decision time window, the system determines that the pressure wave belongs to the opening or closing of dampers in the ventilation duct network, the operation of fans, or similar aerodynamic fluctuation interference, and terminates the current collision avoidance warning intervention. This significantly reduces false alarms caused by the ventilation system.

[0044] The acoustic feature matching module operates as follows: First, the initial acoustic signal is converted into a frequency energy distribution sequence. This conversion can be achieved using short-time Fourier transform, Mel spectrum analysis, or piecewise power spectrum estimation. The frequency energy distribution of each time slice reflects the energy intensity of the current acoustic signal at different frequencies. Then, the frequency energy distribution sequence is matched with a pre-built mechanical noise frequency template library using cosine similarity. The mechanical noise frequency template library contains multiple standard harmonic frequency distribution templates, collected from known diesel engines, motors, transmission systems, mining rubber-tired vehicles, loaders, and transport vehicles at different speeds. The template library can be categorized and stored according to vehicle type, power type, speed range, and load state. Cosine similarity measures the degree of closeness between the current frequency energy distribution and the harmonic distribution shape in the template. Its advantage lies in its insensitivity to overall volume, focusing more on the shape of the frequency energy distribution.

[0045] During matching, the system calculates the similarity between the current frequency energy distribution sequence and each standard harmonic frequency distribution template, and selects the maximum similarity as the judgment criterion. The similarity judgment threshold can be between 0.6 and 0.95, preferably between 0.75 and 0.9. If the maximum cosine similarity is less than the similarity judgment threshold, the system outputs a judgment result indicating that no continuous harmonic frequency component was detected, and then marks the corresponding air pressure wave as ventilation aerodynamic fluctuation interference. If the maximum cosine similarity is greater than or equal to the similarity judgment threshold, it indicates that the current air pressure wave is accompanied by vehicle-related mechanical harmonics, and the system continues to perform Doppler frequency shift extraction, cross-correlation verification, and motion state analysis. Through this acoustic-aerophysical field interlocking method, the air pressure signal is responsible for early triggering, and the acoustic signal is responsible for confirming vehicle attributes. The two constrain each other, which can balance lead time and reliability.

[0046] In intersection scenarios with multiple blind-spot side lanes, a single air pressure signal may lack directional discrimination capability. For example, when vehicles in two side lanes approach the intersection simultaneously, the piston air pressure waves generated by each vehicle will superimpose near the intersection, making it difficult for the differential pressure sensor to individually determine which side lane each vehicle comes from. To address this, the present invention further utilizes a phase direction finding step using secondary sound sources at the rock wall edge. Here, the secondary sound source refers to the equivalent sound radiation position formed at the rock wall edge after the vehicle's sound waves are diffracted. Since the rock wall edge positions of each blind-spot side lane are fixed, the fixed spatial physical coordinates of each rock wall edge can be obtained during system installation or tunnel mapping. The fixed spatial physical coordinates can be represented using a local tunnel coordinate system, with the origin set at the intersection center or the installation point of the composite detection array.

[0047] In this scenario, the acoustic vector sensor preferably employs a miniature microphone array containing at least three microphone elements. Three microphone elements constitute a basic planar direction-finding structure; when higher directional resolution is required or when distinguishing between uphill and downhill roadways, four to eight microphone elements can be used to construct a three-dimensional array. The spacing between microphone elements can range from 2cm to 30cm. Too small a spacing will reduce phase difference resolution, while too large a spacing may cause phase ambiguity in the high-frequency band. Since the dominant frequency of mechanical noise from mine vehicles is relatively low, the element spacing is typically 5cm to 15cm to balance structural size and direction-finding stability.

[0048] When multiple vehicles in blind spots are traveling in different side tunnels, the miniature microphone array captures the superimposed state of diffracted sound waves. The system calculates the signal phase difference sequence between the superimposed diffracted sound waves and each microphone element. The phase difference sequence reflects the path difference of the sound waves propagating from different rock wall edges to each element. Since the spatial position of the rock wall edge of each side tunnel is fixed, the distance combination from different secondary sound sources to each microphone element in the array is also different, so the corresponding theoretical phase difference is distinguishable. This invention inputs the signal phase difference sequence and the fixed spatial physical coordinates of each rock wall edge into a spatial sound source decoupling model, and the spatial sound source decoupling model outputs the target blind spot side tunnel location of each vehicle related to the current air pressure triggering event.

[0049] In one embodiment, the spatial sound source decoupling model employs a mapping model based on feature distance matching. This model does not necessarily require a large number of training samples; instead, it utilizes the tunnel geometry and sound wave propagation relationships to establish a template library. The system first obtains the known spatial distances from the fixed spatial physical coordinates of each rock wall edge to each microphone array element. These distances can be obtained from three-dimensional tunnel mapping data, installation calibration data, or manual ranging. Subsequently, these known spatial distances are input into a preset acoustic phase propagation model and, combined with the principal component frequency bands of the diffracted sound waves or multiple selected frequency points, construct a theoretical phase difference template set corresponding to all secondary sound sources. Each theoretical phase difference template corresponds to a rock wall edge, which in turn corresponds to the location of a branch tunnel where vehicles may have blind spots.

[0050] In actual operation, the system extracts multiple discrete peak phase difference features from the signal phase difference sequence corresponding to the superimposed diffracted sound waves. Peak phase difference features can be understood as phase feature points in the phase difference distribution that have strong energy, recur frequently, or are highly correlated with the propagation path of a certain sound source. When multiple vehicles are present simultaneously, multiple peak phase difference features may appear in the superimposed sound waves, each corresponding to a potential secondary sound source. The system performs feature distance matching between each peak phase difference feature and each template in the theoretical phase difference template set, and calculates the matching distance. The matching distance can be Euclidean distance, Manhattan distance, or dynamic time warping distance; the smaller the distance, the closer the actual phase difference is to the theoretical phase difference of a certain rock wall edge.

[0051] The system selects the template with the smallest matching distance and uses the side lane where the secondary sound source corresponding to this template is located as the target blind zone side lane location of the vehicle. To avoid mismatches, a matching distance threshold can be set, which can be 10% to 40% of the average interval of the theoretical phase difference templates. If the minimum matching distance is still greater than the matching distance threshold, the system can output an uncertain location state, only performing a conservative warning without providing a clear directional indication. If multiple peak phase difference features match different side lanes, the system can simultaneously output the locations of multiple target blind zone side lanes and display the risk directions on the intersection indicator screen. For example, when a vehicle is rapidly approaching from the left front side lane, the indicator screen displays a left front direction warning; when a vehicle is also approaching from the right side lane, the system simultaneously displays a right direction warning and can issue yield or deceleration instructions to the corresponding vehicle.

[0052] To further improve system robustness, the composite detection array can undergo background calibration during installation. Background calibration includes acquiring atmospheric pressure background data in a vehicle-free state, pressure fluctuation data of ventilation fans under different operating conditions, noise data of common stationary equipment, and mechanical noise data of different vehicle types. After calibration, the system establishes background atmospheric pressure and background acoustic baselines at roadway intersections. During online operation, the pressure rise gradient can be calculated relative to the background atmospheric pressure baseline, and the continuous change in Doppler frequency shift can be extracted relative to the main peak of the background acoustic signal, thereby reducing deviations caused by different mines, seasons, and ventilation conditions.

[0053] In the actual early warning logic, the system can be configured with multiple risk levels. When the absolute approach distance is less than a second distance threshold and the relative speed is greater than a second speed threshold, a warning at the alert level is output. When the absolute approach distance is less than a first distance threshold and the relative speed is greater than a first speed threshold, a mandatory intervention level warning is output. The second distance threshold can be greater than the first distance threshold; for example, the second distance threshold can be 20m to 60m, and the first distance threshold can be 5m to 30m. The second speed threshold can be less than or equal to the first speed threshold; for example, the second speed threshold can be 0.5m / s to 2m / s, and the first speed threshold can be 1m / s to 5m / s. The alert level warning is used to remind the driver to pay attention to vehicles approaching from the blind spot, while the mandatory intervention level warning is used to trigger audible and visual alarms, traffic lights prohibiting passage, or the vehicle automatically decelerates and stops. This avoids frequent strong alarms from a long distance while ensuring timely intervention when the vehicle is rapidly approaching at close range.

[0054] This invention can further optimize the early warning strategy by combining vehicle identification information. When underground vehicles are equipped with onboard communication terminals, the edge computing unit can cross-validate the blind zone vehicle distance, speed, and branch roadway orientation obtained from the model with the location and speed reported by the onboard terminal. If they match, the confidence level of the early warning is improved; if communication is interrupted or the positioning drifts, the system can still independently complete the early warning by relying on air pressure waves and diffracted sound waves. This design makes the invention compatible with existing mine dispatching systems and does not rely entirely on wireless communication and positioning signals, making it suitable for underground environments with severe multipath interference and severe obstruction.

[0055] Corresponding to the above method, the present invention also provides an intelligent early warning system for collision avoidance of underground mine vehicles, such as... Figure 1 As shown. The system includes a composite detection array module, a signal feature extraction module, a motion state analysis module, and an early warning and intervention execution module. The composite detection array module is installed on the edge of the rock wall at the intersection of underground mine roadways, such as... Figure 2 As shown, the internal components include a differential pressure sensor and an acoustic vector sensor. The differential pressure sensor collects the air pressure wave generated by the vehicle displacing air as it travels in the restricted tunnel, while the acoustic vector sensor simultaneously collects the diffracted sound waves formed by the vehicle's exhaust noise or mechanical noise bypassing the edge of the rock wall.

[0056] The signal feature extraction module is connected to the composite detection array module. It performs low-pass filtering on the initial air pressure signal, extracts the pressure rise gradient, and performs band-pass filtering, time-frequency analysis, and main peak frequency trajectory tracking on the initial acoustic signal, thereby extracting the continuous change in Doppler frequency shift, such as... Figure 4As shown. The signal feature extraction module can also integrate cross-correlation verification and acoustic-aerodynamic field interlocking gating functions. The cross-correlation verification function is used to determine whether the initial change time of the pressure rise gradient matches the continuous change in the Doppler frequency shift; the acoustic-aerodynamic field interlocking gating function is used to confirm whether there are vehicle mechanical harmonics in the acoustic signal after the air pressure wave is triggered, avoiding false alarms caused by ventilation aerodynamic fluctuations, such as... Figure 5 As shown.

[0057] The motion state analysis module is connected to the signal feature extraction module, and internally deploys a pre-trained motion state analysis model, such as... Figure 3 As shown, this module receives the pressure rise gradient and the continuous change in Doppler frequency shift, and outputs the absolute approach distance and relative speed of the vehicle in the blind spot. In scenarios with multiple blind spot side tunnels, the motion state analysis module can also be connected to the spatial sound source decoupling submodule. The spatial sound source decoupling submodule uses the phase difference sequence collected by the miniature microphone array and the fixed spatial physical coordinates of the rock wall edge to determine the orientation of the target blind spot side tunnel corresponding to each vehicle in the blind spot.

[0058] The early warning intervention execution module is connected to the motion state analysis module. When the absolute approach distance is less than a first distance threshold and the relative speed is greater than a first speed threshold, the early warning intervention execution module executes anti-collision early warning intervention. This module can be connected to underground audible and visual alarms, LED signal indicator screens, vehicle-mounted terminals, level crossing PLC control boxes, dispatching platforms, and vehicle automatic control interfaces. For manually driven vehicles, the system can remind the driver to slow down and give way through audible and visual alarms and directional indicators; for unmanned vehicles or mining vehicles with drive-by-wire functions, the system can directly output deceleration, stop, prohibit entry into intersections, or priority passage control commands.

[0059] Through the above implementation methods, this invention utilizes the pressure rise gradient of air pressure waves to obtain vehicle distance information in the blind zone, and utilizes the continuous change in the Doppler frequency shift of diffracted sound waves to obtain vehicle speed information in the blind zone. Then, a motion state analytical model is used to achieve joint output of distance and speed. Since the air pressure waves and acoustic waves originate from different physical fields, an effective early warning is triggered only when both simultaneously satisfy the vehicle approach law. Therefore, it can identify the risk of actual vehicle approach in advance in mine intersection environments where rock walls obstruct the view, acoustic reverberation, wireless multipath propagation, and ventilation disturbances coexist, and effectively reduce false alarms caused by ventilation fluctuations, fixed mechanical noise, and the superposition of multiple branch roadways.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent pre-warning method for collision avoidance of underground mine vehicles, characterized in that, Includes the following steps: A composite detection array containing micro differential pressure sensors and acoustic vector sensors is deployed at the edge of the rock wall at the intersection of roadways in an underground mine. The differential pressure sensor collects the air pressure wave generated by a vehicle driving in a restricted roadway in a blind spot, and extracts the pressure rise gradient of the air pressure wave. The step of extracting the pressure rise gradient includes: performing a low-pass filter on the initial air pressure signal collected by the differential pressure sensor to filter out signal components with frequencies higher than a preset reference frequency threshold, retaining the quasi-static air pressure wave signal; and taking the derivative of the air pressure wave signal in the time dimension to obtain the pressure rise gradient of the air pressure wave. The acoustic vector sensor synchronously acquires the diffracted sound waves formed when the exhaust noise generated by the vehicle in the blind spot bypasses the edge of the rock wall, and extracts the continuous change in the Doppler frequency shift of the diffracted sound waves. The step of extracting the continuous change in the Doppler frequency shift of the diffracted sound waves includes: bandpass filtering the initial acoustic signal acquired by the acoustic vector sensor, retaining the principal component frequency bands with frequencies between a preset first frequency band and a preset second frequency band to obtain the diffracted sound waves; using fast Fourier transform to extract the time-frequency spectrum of the diffracted sound waves, tracing the change trajectory of the main peak frequency in the time-frequency spectrum with time dimension to obtain the continuous change in the Doppler frequency shift. Obtain the initial change time of the pressure rise gradient and the initial change time of the continuous change in the Doppler frequency shift; calculate the time difference between the initial change time of the pressure rise gradient and the initial change time of the continuous change in the Doppler frequency shift; if the time difference is less than a preset time matching threshold, determine that the air pressure wave and the diffracted sound wave originate from the same vehicle in the blind zone. The pressure rise gradient and the continuous change in Doppler frequency shift are input into a pre-trained motion state analysis model, and the absolute approach distance and relative speed of the vehicle in the blind spot are output through the motion state analysis model. When the absolute approach distance is less than a preset first distance threshold and the relative motion speed is greater than a preset first speed threshold, a collision avoidance warning intervention is executed.

2. The intelligent pre-warning method for collision avoidance of underground mine vehicles according to claim 1, characterized in that, The training process of the motion state analysis model includes: The sample pressure rise gradient of the sample air pressure wave, the continuous change of the sample Doppler frequency shift of the sample diffracted sound wave, and the true approach distance and true speed of the sample vehicle corresponding to the sample air pressure wave and the sample diffracted sound wave are obtained. The sample pressure rise gradient and the sample Doppler frequency shift change are used as the input sequence of the motion state analytical model, and the true approximation distance and the true motion velocity are used as the supervision labels of the motion state analytical model. A loss function is constructed based on the first difference between the predicted approximation distance and the actual approximation distance output by the motion state analysis model, and the second difference between the predicted motion speed and the actual motion speed. The network parameters of the motion state analysis model are updated by minimizing the loss function until a preset convergence condition is met.

3. The intelligent early warning method for collision avoidance of underground mine vehicles according to claim 1, characterized in that, After the step of extracting the pressure rise gradient of the pressure wave, the method further includes an acoustic-gas physics field interlocking gating step, specifically including: Determine whether the pressure ramp-up gradient is greater than a preset gradient trigger threshold; When the pressure rise gradient is greater than the gradient trigger threshold, the air pressure wave is used as a gating signal to activate the acoustic feature matching module and open a judgment time window with a preset duration. Within the determination time window, the acoustic feature matching module is used to detect whether the initial acoustic signal collected by the acoustic vector sensor contains a continuous harmonic frequency characterizing mechanical operation. If the continuous harmonic frequency is not detected within the determination time window, the pressure wave is determined to be a ventilation aerodynamic fluctuation interference caused by the opening and closing of the ventilation duct damper or the operation of the fan, and the anti-collision warning intervention is terminated.

4. The intelligent early warning method for collision avoidance of underground mine vehicles according to claim 3, characterized in that, The step of using the acoustic feature matching module to detect whether the initial acoustic signal acquired by the acoustic vector sensor contains continuous harmonic frequencies characterizing mechanical operation includes: The initial acoustic signal is converted into a frequency energy distribution sequence; The frequency energy distribution sequence is matched with a pre-constructed mechanical noise frequency template library using cosine similarity. The mechanical noise frequency template library contains multiple standard harmonic frequency distribution templates generated by known diesel engines and electric motors. If the maximum cosine similarity between the calculated frequency energy distribution sequence and each of the standard harmonic frequency distribution templates is less than a preset similarity judgment threshold, the judgment result that the continuous harmonic frequency was not detected is output.

5. The intelligent early warning method for collision avoidance of underground mine vehicles according to claim 1, characterized in that, In scenarios involving intersections with multiple blind side tunnels, the acoustic vector sensor is a miniature microphone array containing at least three microphone elements. The method further includes a phase direction finding step based on secondary sound sources at the rock wall edge, specifically comprising: Obtain the fixed spatial physical coordinates of the rock wall edges of each blind zone branch tunnel as a secondary sound source; The miniature microphone array is used to capture the superimposed diffraction sound waves generated and superimposed by multiple vehicles in the blind spot distributed in different blind spot side lanes; Calculate the signal phase difference sequence between the diffracted sound waves in the superimposed state and each of the microphone array elements; The signal phase difference sequence and the fixed spatial physical coordinates are input into a pre-established spatial sound source decoupling model, and the spatial sound source decoupling model outputs the target blind zone branch alleyway location of each of the blind zone vehicles that generate the pressure wave.

6. The intelligent early warning method for collision avoidance of underground mine vehicles according to claim 5, characterized in that, The pre-established spatial sound source decoupling model is a mapping model based on feature distance matching. The step of outputting the target blind zone branch alleyway location of each blind zone vehicle that generates the air pressure wave through the spatial sound source decoupling model includes: Obtain the known spatial distance from each of the fixed spatial physical coordinates to each of the microphone array elements; The known spatial distance is input into a preset acoustic phase propagation model to pre-construct a theoretical phase difference template set corresponding to all the secondary sound sources; Multiple discrete peak phase difference features are extracted from the signal phase difference sequence; Each of the peak phase difference features is matched with each template in the theoretical phase difference template set by feature distance, and the matching distance is calculated. The blind zone branch lane where the secondary sound source corresponding to the template with the smallest matching distance is located is selected as the target blind zone branch lane location of each blind zone vehicle.

7. An intelligent early warning system for preventing collisions with underground mining vehicles using the method of claim 1, characterized in that, include: A composite detection array module is deployed at the edge of the rock wall at the intersection of roadways in an underground mine. The composite detection array module includes a micro differential pressure sensor and an acoustic vector sensor. The micro differential pressure sensor is used to collect the air pressure wave generated by the vehicle in the blind spot driving in the restricted roadway and displacing the air. The acoustic vector sensor is used to simultaneously collect the diffracted sound wave formed when the exhaust noise generated by the vehicle in the blind spot bypasses the edge of the rock wall. The signal feature extraction module, connected to the composite detection array module, is used to extract the pressure rise gradient of the pressure wave and the continuous change of the Doppler frequency shift of the diffracted sound wave. The motion state analysis module, connected to the signal feature extraction module, is used to input the pressure rise gradient and the continuous change of the Doppler frequency shift into the pre-trained motion state analysis model, and output the absolute approximation distance and relative speed of the vehicle in the blind spot through the motion state analysis model. The early warning intervention execution module is connected to the motion state analysis module and is used to perform anti-collision early warning intervention when the absolute approach distance is less than a preset first distance threshold and the relative motion speed is greater than a preset first speed threshold.

Images