Safety anti-collision early warning method and device for industrial vehicle

Abstract

The invention belongs to the technical field of industrial vehicle active safety control, and particularly discloses an industrial vehicle safety anti-collision early warning method and device, and the method comprises the steps: building a unified space-time perception reference through fusing ultrasonic wave, inertial navigation, laser radar and camera data; extended Kalman filtering is fused with inertial navigation prediction and laser radar poses, and high-precision positioning in the tunnel is achieved. Constructing a high-dimensional time sequence feature vector, and predicting a future collision risk in combination with a Bi-LSTM model; the early warning threshold value is dynamically adjusted according to the load, the gradient and the sensor state; the method has the following advantages that the early warning accuracy and response speed under the complex working condition are remarkably improved, and the operation safety of the industrial vehicle is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of active safety control technology for industrial vehicles, and more specifically, to a method and device for early warning and collision avoidance for industrial vehicles. Background Technology

[0002] In mining or tunnel-type infrastructure, heavy-duty automated guided vehicles (AGVs) are widely used for fixed-track transportation of materials such as ores and building materials. Existing automated transportation systems mostly rely on ultrasonic sensors for obstacle detection and use inertial navigation units (IMUs) in conjunction with wheel speed odometers to estimate position. However, in underground tunnels where satellite signals are unavailable, the structure is enclosed, and the environment is complex, ultrasonic sensors are susceptible to interference from multiple reflections from walls and vehicle surfaces, resulting in significant multipath echo interference. Simultaneously, inertial navigation systems inevitably suffer from cumulative drift errors, leading to a decrease in the accuracy of vehicle distance calculations after a certain distance of operation, which can easily cause misjudgments of following distance.

[0003] Due to the distortion of perception data and the uncontrollable positioning error, traditional collision avoidance and early warning methods cannot accurately assess the risk of collisions between vehicles. They generally suffer from problems such as high false alarm rate, delayed response, and intervention failure, making it difficult to meet the safety and reliability requirements under high load and high density queuing operation in tunnels.

[0004] Therefore, a method and device for safety collision avoidance early warning of industrial vehicles are proposed to solve the above-mentioned problems. Summary of the Invention

[0005] The present invention aims to provide a safety collision avoidance and early warning method for industrial vehicles, in order to solve or improve the problem mentioned above where heavy-duty automated guided vehicles operating in underground tunnels suffer from the superposition of ultrasonic sensor multipath echo interference and inertial navigation cumulative error, resulting in unreliable sensing data and positioning results.

[0006] In view of this, a first aspect of the present invention is to provide a method for safety collision avoidance warning of industrial vehicles.

[0007] A second aspect of the present invention is to provide an apparatus.

[0008] The first aspect of this invention provides a method for collision avoidance warning of industrial vehicles, comprising the following steps: real-time synchronous acquisition of ultrasonic sensor data, inertial measurement unit data, lidar data, and visual camera data on the vehicle, and obtaining multi-source sensing data under a unified spatiotemporal reference through preprocessing; calculating the vehicle's predicted position state by using strapdown inertial navigation system (SINS) position; scanning and matching the lidar data using lidar synchronous positioning and mapping methods to obtain the vehicle's absolute pose observation value; fusing the vehicle's predicted position state and the absolute pose observation value using an extended Kalman filter algorithm to obtain the vehicle's pose estimate; constructing a high-dimensional temporal feature vector including the vehicle's current speed, acceleration, relative distance to obstacles, relative speed, braking performance parameters, and environmental interference indicators based on the pose estimate and the multi-source sensing data; performing temporal feature trend analysis on the high-dimensional temporal feature vector to obtain a collision risk index at multiple subsequent time points; adjusting the warning threshold at the current time point according to the vehicle's current load state, real-time slope, and sensor quality indicators; determining the warning level based on the collision risk index and the real-time adjusted warning threshold, and determining whether to generate a warning message based on the warning level.

[0009] A second aspect of the present invention provides an apparatus comprising: a data acquisition module for preprocessing data under a unified spatiotemporal reference to obtain multi-source sensing data; a pose fusion module for performing position prediction based on the inertial measurement unit data and performing scan matching based on the lidar data, combined with an extended Kalman filter algorithm to obtain a vehicle pose estimate; a feature analysis module for constructing the high-dimensional temporal feature vector and performing temporal trend analysis to obtain a collision risk index; and a warning decision module for adjusting a warning threshold according to the vehicle load status, real-time slope, and sensor quality indicators, and determining the warning level based on the collision risk index to generate warning information.

[0010] The beneficial effects of this invention compared to the prior art are as follows:

[0011] By synchronously acquiring and preprocessing multi-sensor data, especially by performing energy envelope and multipath echo removal processing on ultrasonic signals, and filtering out scattering clutter and outlier removal on lidar data, multi-source sensing data achieves precise spatiotemporal unification, significantly reducing the false alarm rate of sensing data under complex tunnel conditions, with the overall false alarm rate controllable below 5%.

[0012] By using an extended Kalman filter fusion positioning strategy that combines inertial navigation prediction with lidar scanning, the problem of large cumulative drift error in traditional inertial navigation systems during long-term operation in underground tunnel environments is effectively solved. This achieves high-precision positioning in GPS-free environments, compresses pose estimation error to the centimeter level, and significantly improves the positioning stability of vehicles in tunnels.

[0013] By constructing a high-dimensional temporal feature vector and employing a bidirectional long short-term memory (Bi-LSTM) network for temporal trend prediction, the dynamic trend of collision risk can be accurately identified and predicted in advance, enabling early warning response of collision risk 0.3 to 0.5 seconds in advance. This reduces the response speed of collision warning from the traditional 500ms to less than 120ms, effectively improving the timeliness of collision warning.

[0014] By dynamically adjusting the warning threshold based on a combination of factors including load status, real-time slope, and sensor data quality, the warning system can automatically increase its sensitivity and provide early warning intervention in extreme scenarios such as heavy loads, downhill slopes, and severe dust interference. In scenarios with light loads, uphill slopes, and favorable environments, it can effectively avoid unnecessary false alarms, thereby further improving the reliability and operational efficiency of the warning system.

[0015] Additional aspects and advantages of embodiments of the invention will become apparent in the following description or may be learned by practice of embodiments of the invention. Attached Figure Description

[0016] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0017] Figure 1 This is a flowchart of the method steps of the present invention;

[0018] Figure 2 This is a flowchart of the multi-source data preprocessing process of the present invention;

[0019] Figure 3 This is a flowchart of the inertial navigation-laser fusion positioning process of the present invention;

[0020] Figure 4 This is a flowchart of the time-series risk prediction process of the present invention;

[0021] Figure 5 This is a flowchart of the dynamic threshold early warning control of the present invention;

[0022] Figure 6 This is a complete closed-loop flowchart of the present invention;

[0023] Figure 7 This is a block diagram of the device logic structure of the present invention. Detailed Implementation

[0024] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0026] Please see Figures 1-7 The following describes some embodiments of the industrial vehicle safety collision avoidance warning method and device of the present invention.

[0027] An embodiment of the first aspect of the present invention provides a method for safety collision avoidance and early warning of industrial vehicles. In some embodiments of the present invention, such as... Figures 1-6 As shown, the method includes:

[0028] S101 collects data from ultrasonic sensors, inertial measurement units, lidar, and vision cameras on the vehicle in real time and synchronously, and obtains multi-source perception data under a unified spatiotemporal reference through preprocessing.

[0029] Here, ultrasonic sensors are used to detect the relative positional changes of nearby obstacles around the vehicle, especially in tunnel scenarios to capture the proximity signals of static objects in front or to the side. However, because tunnels are enclosed structures, there is severe multipath reflection between the sound wave signal and the tunnel wall and vehicle structure, so the raw acquired signal may contain multiple interfering echoes. Therefore, the raw ultrasonic data needs to be bandpass filtered first to suppress background noise, and then the Hilbert-Huang Transform (HHT) method is introduced to obtain the energy envelope curve of each echo signal. By setting an energy threshold and jointly judging the density of the main echo within a limited time window, the first main echo moment in the ultrasonic echo signal is identified and extracted, ultimately achieving effective removal of multipath interference components generated by wall reflection, thereby obtaining reliable obstacle location information.

[0030] Meanwhile, the inertial measurement unit (IMU) collects the vehicle's three-axis acceleration and angular velocity data in real time. Based on this high-frequency kinematic information, a strapdown inertial navigation algorithm is used to calculate the vehicle's predicted position in the current coordinate system, continuously providing trajectory information in environments without external positioning sources. However, considering the cumulative zero-bias drift error inherent in inertial navigation, which can easily lead to position estimation distortion, especially after prolonged operation, this process is not feasible.

[0031] For the lidar sensor, a high-precision 3D lidar is used to acquire geometric information about the tunnel's internal structure and surrounding environment. Considering the unstable point cloud reflection intensity caused by complex media such as dust and water vapor in tunnels, a filtering threshold is set using the point cloud intensity value. Combined with the local point cloud density distribution, a statistical outlier removal algorithm is used to remove false or discrete points, improving the realism of the environmental model. The complete frame of point cloud output by the lidar is then used to perform a simultaneous localization and mapping process. High-precision registration between the point cloud and the local map is performed through feature extraction and iterative nearest-point method to obtain the absolute pose observation value of the vehicle in the tunnel map.

[0032] Visual cameras are deployed in front of or around the vehicle to capture images of the tunnel environment, personnel activity, and other moving targets. In the preprocessing layer, visual data is primarily used for target detection and scene semantic segmentation, combined with deep learning algorithms such as YOLOv5 to identify important semantic objects such as pedestrians, obstacles, and traffic signs. To address the drastic changes in lighting inside the tunnel, enhancement algorithms such as adaptive histogram equalization and image gamma correction are employed to improve image quality in low-light or reflective conditions, ensuring detection accuracy.

[0033] As described above, the four types of sensors collect information from multiple perspectives, including geometric structure, dynamic state, environmental reflection, and visual cues. To achieve consistent representation of multi-source information, time synchronization and spatial coordinate transformation are performed during the preprocessing stage: a high-precision clock synchronization module is used to uniformly align the sampling times of different sensors, ensuring that the outputs of different sensors correspond to the physical state at the same moment; and the local coordinate system data of each sensor are uniformly transformed to the vehicle's central reference coordinate system using the calibrated extrinsic parameter matrix. Finally, a fusion data format under a unified spatiotemporal reference is generated, providing rigorous data support for subsequent pose fusion, feature vector construction, and risk calculation.

[0034] Specifically, the steps for obtaining multi-source sensing data under a unified spatiotemporal reference through preprocessing include:

[0035] The energy envelope of the echo signal is obtained by bandpass filtering the ultrasonic sensor data and then performing Hilbert-Huang transform.

[0036] The energy envelope is used as the input signal for energy threshold determination, and the position of the first main echo is determined in conjunction with the time window density.

[0037] Based on the location of the first main echo, multipath interference echoes caused by reflections from the tunnel walls are eliminated from the ultrasonic sensor data.

[0038] Regarding the specific description above, in the unified spatiotemporal reference preprocessing link, for the original echo signal of the ultrasonic sensor, taking 40kHz as an example, a fifth-order Butterworth bandpass filter is first designed in the digital domain based on the sensor's center frequency, limiting the cutoff frequency window to 38–42kHz to reduce low-frequency drift caused by vehicle vibration and high-frequency spikes generated by tunnel mechanical noise. After filtering, multiple echo components caused by wall reflection may still be retained. Therefore, Hilbert-Huang transform is further introduced to perform empirical mode decomposition on the signal, adaptively separating several intrinsic mode functions. Hilbert spectral analysis is then performed on the first two IMFs with the most concentrated energy to obtain the instantaneous energy envelope curve E(t) that varies with time.

[0039] Next, a dual decision rule is constructed for E(t): First, a baseline energy threshold E_th = α·E, α = 1.8, and E is the average energy value of the entire segment. The time point when the energy first jumps above the threshold is recorded as a candidate main echo. Second, a sliding time window of length Δt = 0.6ms is established with this candidate point as the center. The number ρ of effective sampling points exceeding 0.8·E_th within the window is counted. Only when ρ ≥ ρ_min, where ρ_min is empirically taken as 5, is the candidate determined to be the first main echo time t0, thus avoiding misjudging occasional spikes as real echoes. After determining t0, the corresponding distance d0 = c·t0 / 2 is written into the distance buffer table using the sound speed c and sampling frequency f_s. Amplitude suppression or direct shielding is applied to all waveforms in the signal after t0+τ, where τ is approximately the round-trip time difference of one reflection from the tunnel wall, completely removing multipath interference echoes formed by one or more wall reflections.

[0040] As can be seen above, after three layers of processing including bandpass noise reduction, HHT energy spectrum resolution, and threshold-density dual decision, the ultrasonic ranging error was reduced from ±0.35m in the original working condition to ±0.07m, and the false alarm rate was reduced from 32% to less than 5%. This laid a highly reliable foundation for the spatiotemporal alignment of multi-source data and laser point cloud and risk assessment of close-range obstacle distance input.

[0041] Specifically, the steps for obtaining multi-source sensing data under a unified spatiotemporal reference through preprocessing also include:

[0042] Based on the intensity characteristics of the laser point cloud, intensity threshold filtering is applied to the lidar data for scattering clutter points caused by dust and water vapor. Random noise points are then removed by calculating and employing statistical outlier filtering based on the local density of the point cloud.

[0043] For the above specific description, in the multi-source perception preprocessing process, in order to ensure that the lidar point cloud can still maintain high fidelity in the underground tunnel environment with dust, high humidity and severe beam scattering, a clutter removal algorithm is designed based on the dual constraints of echo intensity characteristics and spatial distribution characteristics: First, sort and count the original point cloud according to the single-point intensity value I, calculate the average intensity ī and standard deviation σ_I of the whole frame, and set the low-intensity threshold as I_th = ī - 1.2σ_I based on the measured dust scattering attenuation model; for the point cloud with I < I_th, it is considered that its reflected energy mainly comes from floating dust or water vapor particles, rather than the tunnel wall or physical obstacles, so such points are marked as scattered clutter candidates.

[0044] To avoid misdeletion due to weak reflection on the surface of large-sized obstacles, local spatial density verification is introduced based on the marked candidate points: a spherical neighborhood with a radius r = 0.25m is constructed centered on each candidate point, and the number of valid points k in the neighborhood is counted; if k ≥ k_min, where k_min depends on the number of radar lines and the tunnel cross-section and is usually set to 10, it means that there are still continuous physical reflections near this area, and this point is retained; otherwise, it is considered that this point comes from isolated scattering and is deleted.

[0045] After completing the intensity threshold filtering, there may still be a small amount of random noise or hardware glitch points remaining. Therefore, statistical outlier removal is performed again on the filtering result: for each point, calculate its average distance d_i to 20 nearest neighbor points, establish the global average distance μ_d and standard deviation σ_d; if a certain point d_i > μ_d + 2.5σ_d, it is marked as an outlier and deleted.

[0046] As can be seen from the above, this two-stage processing process takes into account both the lidar echo intensity characteristics and the spatial structure consistency. After field tests in the tunnel, when the dust concentration (TSP) is about 150mg / m 3 it can retain 98% of the effective wall point cloud while reducing the proportion of scattered clutter points from 23% to less than 5%; at the same time, the density of random noise points is reduced by two orders of magnitude, providing a high-integrity and low-noise three-dimensional geometric information benchmark for subsequent point cloud-IMU time alignment, feature extraction, and vehicle pose optimization.

[0047] S102,推算惯性测量单元数据以获得车辆的位置预测状态;通过激光雷达同步定位与建图方法对激光雷达数据进行扫描匹配,获得车辆的绝对位姿观测值;通过扩展卡尔曼滤波算法融合车辆位置预测状态与绝对位姿观测值,以获得车辆的位姿估计。 Estimate the data of the inertial measurement unit through the strapdown inertial navigation position to obtain the predicted position state of the vehicle; perform scan matching on the lidar data through the lidar simultaneous localization and mapping method to obtain the absolute pose observation value of the vehicle; fuse the predicted position state of the vehicle and the absolute pose observation value through the extended Kalman filter algorithm to obtain the pose estimation of the vehicle.

[0048] Here, the process of extrapolating inertial measurement unit (IMU) data using strapdown inertial navigation (SINS) position mainly involves using the three-axis acceleration and three-axis angular velocity data output by the vehicle's IMU to perform continuous integration and obtain the vehicle's displacement and attitude information in inertial space. In specific implementation, a strapdown inertial navigation (SINS) algorithm structure is adopted. This involves decoupling and transforming the IMU output data to convert the original data in the body coordinate system to the navigation coordinate system, and then calculating state variables such as velocity, displacement, and attitude angles through integration. It should be noted that in this technical scenario, due to the lack of external signal sources such as GPS or geomagnetic references for calibration in the underground tunnel environment, inertial navigation errors tend to accumulate over time. Therefore, when a vehicle enters a short-term stationary state, such as a waiting area in a tunnel or a parking area at a loading / unloading point, a Zero-Speed ​​Update (ZUPT) mechanism is triggered. This mechanism determines whether the vehicle is stationary by detecting whether the acceleration amplitude and angular velocity change rate of the IMU are within the stationary threshold range. Once the stationary state is confirmed, the current speed estimate is forced to converge to zero, and the gyroscope offset term and acceleration deviation are updated accordingly, thereby effectively suppressing long-term drift of the inertial system.

[0049] LiDAR data is scanned and matched using a simultaneous localization and mapping (LiDAR) method. This process typically involves front-end point cloud feature extraction and back-end pose optimization working in tandem. In the front-end processing, the real-time 3D point cloud data generated by the LiDAR is converted to polar coordinates, and key feature points such as edge points and planar points are extracted. Subsequently, the feature points of the current frame are iteratively registered with the historically constructed local map. Matching algorithms such as ICP (Iterative Closest Point) or NDT (Normal Distribution Transform) can be used to solve for the optimal pose transformation of the current LiDAR frame relative to the historical map by minimizing the Euclidean residuals or probability distribution differences between point clouds, thereby obtaining the vehicle's current absolute pose observations, including position and orientation. In complex tunnel environments, this method can further incorporate loop closure detection and loop optimization mechanisms to improve the robustness of localization in structurally repetitive areas such as long straight sections.

[0050] By fusing the two complementary pose information sources, the Extended Kalman Filter (EKF) algorithm is employed for state estimation. As a minimum mean square error estimator suitable for nonlinear systems, the EKF's core principle is to construct two stages: prediction and correction, utilizing the system's nonlinear state transition equation and observation model. In the prediction stage, the vehicle's position state, calculated using strapdown inertial navigation (SINS), is used as a priori estimate, and its covariance matrix is ​​calculated. In the correction stage, the absolute pose obtained from lidar scanning is used as the observation, and the observation residuals and covariance weights of both are combined. The overall pose estimation state and error covariance are updated via Kalman gain, thus preserving the high-frequency response capability of SINS while introducing lidar drift correction capability, achieving high-precision estimation of the vehicle's current spatial position and attitude. The fused pose results not only provide a reliable spatial reference for subsequent risk assessment and path control but also significantly improve the practicality and stability of the entire early warning system under extreme tunnel conditions.

[0051] As described above, to achieve high-precision positioning of industrial vehicles in enclosed spaces such as underground tunnels, the following steps are taken: First, the three-axis acceleration and angular velocity data output by the inertial measurement unit (IMU) are combined with a strapdown inertial navigation algorithm to continuously calculate the vehicle's trajectory and obtain a preliminary position prediction. To overcome the problem of error accumulation during long-term operation of inertial navigation, a zero-speed update mechanism is triggered when the vehicle is stationary or in a low-speed stable state to correct the predicted state with zero-bias drift. Subsequently, the point cloud features of the current frame are extracted using a simultaneous localization and mapping method with lidar, and iteratively matched with the tunnel environment map to calculate the absolute pose observation value of the vehicle in the global coordinate system. Finally, an extended Kalman filter algorithm is used to dynamically fuse the inertial navigation prediction state and lidar observation results. This retains the advantage of the rapid response of the inertial navigation system while introducing the global correction capability of lidar observation, thereby obtaining a stable, continuous, and accurate pose estimate, providing a reliable spatiotemporal reference basis for subsequent collision risk analysis and control command generation.

[0052] Specifically, the steps for extrapolating inertial measurement unit data based on the strapdown inertial navigation system position include:

[0053] Based on the triaxial acceleration and angular velocity data output by the inertial measurement unit, the position prediction state is calculated using strapdown inertial navigation.

[0054] Zero-speed updates are initiated when the vehicle is stationary at preset time intervals to correct the zero-bias drift error accumulated by inertial navigation in the position prediction state in real time.

[0055] Regarding the specific description above, in the inertial navigation prediction stage of the pose estimation process, the linear acceleration vector a output by the three-axis inertial measurement unit (IMU) installed near the vehicle's center of gravity is first sampled at a frequency of 100Hz or higher. b =[ax a y a z ] T With angular velocity vector ω b =[ω x ω y ω z ] T As the initial input, the attitude, velocity, and position are continuously integrally calculated using the Strap-down Inertial Navigation System (SINS) algorithm. Specifically, the equations are first updated using quaternions.

[0056]

[0057] in, This represents the time derivative of the quaternion q, i.e., the rate of change of the quaternion;

[0058] Real-time calculation of the attitude rotation matrix from the vehicle's mechanical system to the navigation coordinate system Then, the acceleration of the machine system is transformed into the navigation system and the gravity vector g is compensated to obtain the relative force of the navigation system. The velocity v is then updated in the form of a first- or second-order integral. k =v k-1 +f n Δt and cumulative displacement p k =p k-1 +v k Δt, thus generating the vehicle's predicted position state at the current moment. However, due to error sources such as zero bias, scaling factor, and random walk, this pure inertial navigation integral will exhibit exponential drift over long-term operation. To suppress this phenomenon, a zero-rate update (ZUPT) trigger criterion is set at the task scheduling layer: when the linear velocity detected by the wheel encoder is below 0.05 m / s, and the acceleration magnitude measured by the IMU is ||a b ||and angular velocity modulus||ω b || All of them fall within the rest threshold range for 0.5 seconds consecutively; specifically, ||a b -g||<0.03g and ||w b If || < 0.01 rad / s, the vehicle is determined to be stationary or briefly stopped at the top of a hill. In this case, the speed is forcibly set to zero, and the observed residuals are processed using an amplified state Kalman filter. Inverse mapping to gyroscope zero bias b ω and acceleration zero bias b aReal-time error correction prevents drift gain from accumulating over time. After a 1.5km continuous test in a mining tunnel, the zero-speed update mechanism automatically triggers every 300m on average, reducing pure inertial navigation position drift from ±12m to ±0.9m. This lays a highly reliable prior prediction foundation for subsequent extended Kalman filter fusion with lidar-matched observations.

[0059] As described above, by generating the predicted position state through continuous integration using high-frequency strapdown inertial navigation in the absence of GPS in underground tunnels, and then triggering zero-speed update and zero-bias correction in real time at the top of a low-speed incline during brief vehicle stops or low-speed climbs, this invention retains the advantage of inertial navigation in millisecond-level response to instantaneous attitude and velocity changes, while effectively suppressing the cumulative drift that grows exponentially over time. This reduces the pure inertial navigation position error over long distances from the order of ten meters to less than one meter. The highly reliable and continuous predictive trajectory not only provides a robust prior for subsequent extended Kalman fusion with lidar observations, enabling faster and more accurate fusion positioning, but also maintains uninterrupted vehicle positioning output even when the quality of the laser echo is instantaneously deteriorated by dust or strong reflections. This significantly improves the reliability, robustness, and safety redundancy of the entire collision avoidance warning system under extreme tunnel conditions.

[0060] Specifically, the steps for scanning and matching lidar data using the lidar synchronous positioning and mapping method include:

[0061] Real-time feature extraction is performed on the lidar point cloud data, and a scanning matching method based on the iterative nearest point method is used to register the current frame lidar point cloud with the local map of the tunnel to obtain the absolute pose observation value of the vehicle.

[0062] Regarding the specific description above, in the simultaneous localization and mapping stage of the LiDAR, the 3D point cloud acquired by the LiDAR at 10Hz sampling is first preprocessed in real time: the original millions of points are downsampled to approximately 120k points through voxel mesh filtering to reduce the computational load and homogenize the spatial distribution; then, curvature analysis is performed on each scan line in polar coordinates, selecting the first 2% of the steepest points as edge features and the last 2% of the gentlest points as planar features, and using KD-Tree clustering in the local neighborhood to remove isolated noise, resulting in a set of approximately 8000 structured feature points. Next, the predicted pose output from the previous fusion cycle is... As an initial transformation, the feature points of the current frame are projected onto the local map formed by the previous keyframe. Based on this, an improved Iterative Nearest Point (ICP) algorithm is used to perform fine registration: first, corresponding sets are established for edge-to-edge and plane-to-plane features, and a residual function is constructed through nearest neighbor search matching.

[0063]

[0064] in, Let i be the reference point for the i-th planar feature. Let be the coordinates of the i-th planar feature point in the current frame or the source frame, and n be the normal vector. To iterate through the set of indexes for matching pairs of all points ∑ i∈ε The algorithm iterates through the set of matching pairs ε for all edge or planar feature points. Then, it solves for the pose increment δξ using Lie algebra linearization and a Lewenberg-Marquardt (LM)-based Gauss-Newton iteration. During the iteration, Huber loss is introduced to suppress outlier matches, and pairs with residuals greater than 0.3m are dynamically removed. To prevent registration divergence caused by repetition of long straight segments in the tunnel environment, the algorithm adds ground constraints between every two frames: Random Sample Consensus (RANSAC) plane fitting is performed on the ground point cloud, and the vehicle's z-axis ground clearance residual is added to the cost function to lock the vertical drift. After 5–8 iterations, if the pose increment norm ||δξ|| < 1 × 10⁻⁶, the algorithm is considered successful. -4 If the mean square error is less than 0.01m, convergence is determined and the relative pose of the current frame is output.

[0065] Then By overlaying the data onto the global coordinate system of the local map, we obtain the absolute pose observations of the vehicle in the world coordinate system. A sliding window approach is used to retain a sub-map with a length of 20m and a span of 50 frames to limit storage growth. This pose is not only used for real-time updates of the local map but also triggers loop closure constraints when loop closure conditions are detected. The difference between the decoder mileage and the laser pose is <0.5m, and the yaw difference is <5°. Factor graph optimization of the global topology map is used to further reduce cumulative errors. Based on field measurements in a 1.5km tunnel, this scanning matching process works well with a dust concentration of 150mg / m³. 3 Even under severe tunnel wall reflections, it can maintain inter-frame registration time of <30ms and pose drift rate of <0.08% (8cm / 100m), providing centimeter-level absolute observation input for the extended Kalman filter, significantly improving overall positioning accuracy and robustness.

[0066] As described above, with the help of feature-constrained iterative nearest point registration and real-time local map updates, this lidar scanning and matching process can still complete the precise registration of a point cloud and local map within 30ms in a dusty, highly reflective, and structurally repetitive tunnel environment, outputting centimeter-level absolute pose observation values. By introducing ground plane constraints to suppress vertical drift and incorporating closure factors to optimize and limit divergence in long straight segments, the overall pose drift rate is controlled within 0.08%, making lidar observation a robust measurement source for correcting inertial navigation drift. This significantly improves the accuracy and robustness of fusion positioning and provides a high-confidence spatial reference for subsequent collision risk assessment and safety control.

[0067] S103, based on pose estimation and multi-source perception data, constructs a high-dimensional temporal feature vector including the vehicle's current speed, acceleration, relative distance to obstacles, relative speed, braking performance parameters, and environmental interference indicators; performs temporal feature trend analysis on the high-dimensional temporal feature vector to obtain the collision risk index at multiple subsequent time points.

[0068] Here, based on the fused vehicle pose estimation results and multi-source perception data calibrated with a unified spatiotemporal reference, a high-dimensional temporal feature vector reflecting the relationship between the current vehicle state and the surrounding environment is constructed. Specifically, the vehicle's current velocity and acceleration information are extracted by differentially calculating continuous displacement and attitude changes in the pose estimation results. Furthermore, ultrasonic sensors and lidar are used together to perceive obstacles ahead, and the relative distance and relative velocity between the vehicle and the obstacle are calculated by combining the vehicle's current pose coordinates with the obstacle's location. Among these, lidar provides high-resolution distance measurement to enhance accuracy, while ultrasonic data provides redundancy for low-speed, close-range conditions.

[0069] Building upon this, the impact of the vehicle's current braking performance on the warning strategy also needs to be considered. Therefore, based on the vehicle's load status and real-time slope data, its effective braking distance or braking hysteresis response time is calculated as an important parameter for dynamically adjusting the warning response timing. To further enhance the model's ability to perceive interference from complex external environments, environmental interference indicators reflecting perception quality are introduced, such as the sensor's current signal-to-noise ratio, the completeness of sampled data, the proportion of effective data, and the marking of interference periods, to help improve the overall robustness of the model.

[0070] The data from the aforementioned multiple dimensions constitute a set of feature vectors at each time point. Sliding sampling is performed within a fixed time window to form a temporal feature vector sequence containing multiple consecutive time points. This temporal sequence is then input into a pre-trained offline neural network model, preferably a Bidirectional Long Short-Term Memory (Bi-LSTM) network with memory and prediction capabilities. Through a comprehensive analysis of the current and historical states, the model predicts the collision risk index at multiple subsequent time points. This collision risk index is a quantitative risk assessment result represented by a continuous variable between 0 and 1; a higher value indicates a greater likelihood of a collision at a future time. Through this mechanism, not only can the existence of a collision risk be statically determined, but the trend of risk changes can also be dynamically perceived, enabling a highly reliable risk management strategy of early warning and gradual intervention.

[0071] As described above, by integrating vehicle pose estimation and multi-source perception data, a high-dimensional temporal feature vector reflecting the vehicle's dynamic state and environmental characteristics is constructed, covering multiple key dimensions such as speed, acceleration, relative information of obstacles, braking performance parameters, and environmental interference indicators. Based on this, a bidirectional long short-term memory network is used to predict the trend of the feature vector sequence within a continuous time window, outputting collision risk indices for multiple future moments. This enables dynamic assessment and early identification of potential vehicle collision risks, providing a reliable basis for subsequent graded early warning and active control.

[0072] Specifically, the constructed high-dimensional time-series feature vectors include:

[0073] The vehicle's current speed and acceleration are extracted based on pose estimation and multi-source perception data;

[0074] The relative distance and relative speed of obstacles in front of the vehicle are determined based on data from ultrasonic sensors and lidar.

[0075] Based on the vehicle's current load mass and the tunnel's real-time slope, determine the vehicle's current braking performance parameters;

[0076] Environmental interference indicators are determined based on the real-time signal-to-noise ratio or effective data ratio of the sensors.

[0077] Regarding the specific description above, in the risk assessment stage, to fully capture the temporal evolution characteristics of the interaction between vehicle dynamics and the tunnel environment, a high-dimensional temporal feature vector was constructed around four dimensions: motion, obstacle, braking, and environment. This vector was then continuously sampled using a fixed window method to form a sequence input. First, the translation and pose difference between adjacent time points were analyzed from the fused pose results.

[0078] The longitudinal axis serves as the fundamental dynamic characteristic for measuring the current driving conditions (acceleration, constant speed, deceleration). Secondly, using denoised ultrasonic ranging and lidar point cloud depth, the nearest obstacle is searched within a fan-shaped area in front of the vehicle, and its distance d is recorded. t The relative distance dimension is used, and the radial relative velocity Δv of the vehicle relative to the obstacle is obtained by either differentially analyzing the distance between two consecutive frames or by directly tracking the obstacle's velocity vector. t This quantifies the urgency of a potential collision. Third, based on the real-time load mass m provided by the onboard weighing sensor... t and the tunnel longitudinal slope angle α obtained by inertial navigation / laser fusion t Call the braking performance model a max =f(m t α t Dynamically calculate the maximum available braking deceleration and further derive the safe braking distance. This value, as a braking performance parameter, reflects the current vehicle's safe braking capability boundary. Finally, to improve the model's robustness in scenarios with abnormal perception, the signal-to-noise ratio, effective point ratio, and frame drop rate of the output frames from each core sensor (LiDAR, ultrasonic, and vision) are statistically analyzed in real time, and the comprehensive score Q is calculated. t The values ​​∈ [0, 1] are recorded as environmental interference indicators, where increased dust concentration, weakened echo, or blurred image will affect Q. t The 'decline' occurs. Thus, the high-dimensional feature vector of a complete time step can be expressed as:

[0079] F t =[v t α t d t , △v t D brake，t Q t ] T

[0080] Collect the past n frames {F} using a sliding window t-n+1 F t The data is input into a bidirectional LSTM model for time series trend analysis, enabling the network to both review the past and capture short-term future dependencies, thereby accurately predicting the collision risk index at multiple subsequent moments and driving a graded early warning strategy.

[0081] As mentioned above, by integrating four dimensions—vehicle longitudinal dynamics (vehicle speed, acceleration), obstacle proximity (relative distance, relative speed), adaptive braking capability (real-time maximum deceleration and safe braking distance under load-slope coupling), and perception credibility (multi-sensor signal-to-noise ratio and data integrity score)—into a continuously sampled high-dimensional temporal feature vector, the system not only comprehensively depicts the multi-factor safety situation of vehicle-obstacle-road-ring at the single-frame level, but also accurately captures the micro-trend changes of risk quantification indicators over time by using bidirectional LSTM to model the historical-future dependency of the feature sequence. This upgrades collision prediction from simple threshold triggering to factor coupling judgment for complex working conditions. It can maintain high sensitivity in extreme scenarios such as heavy-load downhill and dust obstruction, while reducing false alarms when lightly loaded downhill and with good perception signals. This achieves risk prediction at the millisecond level in advance and stable early warning with a false alarm rate of less than 5%, thereby significantly improving the operational safety margin and overall traffic efficiency of heavy-duty AGVs in underground tunnels.

[0082] Specifically, the steps for performing time series feature trend analysis on high-dimensional time series feature vectors include:

[0083] High-dimensional temporal feature vectors are input into a trained bidirectional long short-term memory network model to predict the collision risk index at multiple subsequent time points.

[0084] By analyzing the trend of the collision risk index over multiple consecutive time periods, changes in the risk level can be determined, enabling early identification and warning response to collision risks.

[0085] Regarding the specific description above, in the time-series feature trend analysis stage, the high-dimensional time-series feature vector sequence {F}, composed of four dimensions—velocity, acceleration, relative distance, relative velocity, braking performance, and environmental disturbance—is first continuously sampled at a 50ms period. t-(L-1) F t The input is fed into a pre-trained bidirectional long short-term memory (Bi-LSTM) network model. This model recursively operates simultaneously in both forward and backward time flows: the forward LSTM sequentially reads features from the past L frames, capturing the cumulative effect of risk factors from the past to the present; the backward LSTM backtracks along the reverse time axis, enhancing its perception of sudden acceleration spikes or sharp drops in signal-to-noise ratios, extending from the present to the future. The two hidden state vectors are concatenated at each time step and fed into a fully connected layer with sigmoid activation, outputting a collision risk index sequence for the next k prediction steps, such as 0.1s, 0.2s, and 0.3s.

[0086]

[0087] r t+i ∈[0,1]

[0088] During the model training phase, ground truth data covering various scenarios such as normal driving, emergency stop, rear-end collision, and dust obstruction are used. The loss function is a binary cross-entropy weighted with hard samples. Combined with Adam optimization, the model is iterated on 1 million frames of samples until convergence. This enables the network to identify both long-term risks caused by the increased braking distance under heavy loads on long slopes and short-term risk pulses caused by sudden acceleration or obstacles.

[0089] When running online, first... Take the maximum value r max Assuming peak risk, calculate the risk gradient:

[0090]

[0091] This is used to determine risk trends: when r max A value ≥0.9 or Δr>0.25 indicates a sharp increase in risk, requiring immediate emergency braking; if 0.6≤r max If r < 0.9 and Δr shows a positive increase, then early deceleration is triggered; if r max If the value is less than 0.3 and Δr ≤ 0, then a safe state is maintained. Taking a 1.5km continuous test in a mining tunnel as an example, under the condition of detecting a stationary obstacle 3.2m ahead and the vehicle speed 1.8m / s, the output r is generated 0.8s (approximately 1.4m) away from the collision point. max=0.93, 0.35s earlier than the traditional TTC single threshold method; while with increasing dust concentration and sensor quality score Q t During the descent, the network can also automatically improve the output risk through environmental interference dimensions to compensate for missed detections caused by distance measurement jitter. With the help of this bidirectional memory-multi-step prediction-gradient trend joint mechanism, millisecond-level early identification and graded warning response of collision risks of heavy-load AGVs in underground tunnels are achieved, while the false alarm rate is controlled below 4.7%.

[0092] As described above, by inputting the high-dimensional state sequence within a continuous time window into a bidirectional LSTM and outputting multi-step risk predictions, and then combining the risk peak and gradient changes to determine the warning level in real time, this step not only fully inherits the cumulative effects of historical inertia, speed, and environmental noise using forward memory, but also responds sensitively to short-term sudden conditions through reverse reasoning. Therefore, it can provide a quantitative risk index based on centimeter-level pose and multi-source perception even hundreds of milliseconds before a collision, or even before the vehicle has significantly decelerated, achieving unified advanced identification of both long-term gradual and near-term sudden dangers. At the same time, through adaptive trade-offs in environmental interference dimensions, it effectively suppresses missed detections and false alarms under extreme conditions such as dust, water vapor, and echo attenuation, ultimately advancing the emergency braking trigger by about 0.3 seconds and reducing the false alarm rate to less than 5%. This allows the entire warning system to maintain both high sensitivity and stability, significantly improving the safety margin and passage efficiency of heavy-duty AGVs in underground tunnels.

[0093] S104: Adjust the warning threshold at the current moment based on the vehicle's current load status, real-time slope, and sensor quality indicators; determine the warning level based on the collision risk index and the real-time adjusted warning threshold, and determine whether to generate a warning message based on the warning level.

[0094] Here, the collision risk assessment criteria are dynamically adjusted based on the vehicle's current operating status and environmental information. Specifically, the vehicle's current load status is read in real time, and its inertial parameters, such as braking distance and acceleration limits, are calculated to reflect the vehicle's current deceleration and emergency braking capabilities. Simultaneously, the terrain slope information of the tunnel section the vehicle is currently in is used to further assess the changes in the vehicle's motion inertia caused by variations in the gravitational component, thereby revising the original collision risk assessment threshold. For example, on uphill sections, where vehicle deceleration is easier, the risk threshold can be appropriately increased; while on downhill sections, the threshold should be decreased to enhance sensitivity.

[0095] In addition, the data quality metrics of various sensors, such as signal-to-noise ratio, effective frame rate, and echo recognition rate, are evaluated to dynamically adjust the data reliability weights. When sensor signal quality deteriorates, for example, due to severe interference from walls in ultrasonic echoes or strong scattering from lidar, an early warning mechanism will be triggered by lowering the warning threshold to avoid missed detections caused by information distortion.

[0096] After completing the aforementioned multi-factor adaptive adjustment, the collision risk index at the current moment is compared with the real-time adjusted warning threshold to determine the current risk level range, and a corresponding warning level is generated accordingly, such as safe, alert, and dangerous levels. If the assessment result meets the warning trigger conditions, the corresponding warning information will be output immediately, prompting the driver through the human-machine interface, or the signal will be connected to the automatic control system to perform deceleration, stopping, or other operations to ensure the safe operation of heavy-load vehicles and the stability of the convoy within the tunnel. This mechanism, while ensuring timely response, can significantly reduce the problem of safety intervention failures caused by misjudgments or omissions.

[0097] As described above, by comprehensively analyzing multiple factors such as vehicle load, tunnel slope, and sensor mass, the collision risk judgment threshold is dynamically adjusted, enabling the early warning system to adapt to changes in complex tunnel environments. Based on this, the real-time calculated collision risk index is compared with the adjusted threshold to accurately classify risk levels and trigger corresponding early warning information. This achieves early identification and proactive intervention of industrial vehicle collision risks, significantly improving reliability and safety in high-density, low-visibility environments.

[0098] Specifically, the steps for adjusting the warning threshold include:

[0099] Based on the vehicle's current load status, the vehicle's inertia factor is calculated in real time, and the warning threshold is adjusted according to the magnitude of the inertia factor.

[0100] The warning threshold is dynamically adjusted based on the real-time slope changes in the tunnel area where the vehicle is located.

[0101] Adjust the corresponding warning thresholds based on real-time sensor data quality indicators.

[0102] Regarding the specific description above, at the risk decision-making level, in order to enable the early warning threshold to adaptively adjust to real-time changes in vehicle dynamics and external perception reliability, the current load mass m of the vehicle is first obtained by combining the on-board weighing sensor and the job scheduling database. t The inertia factor η is obtained by comparing it with the unloaded mass m0. m =m t / m0; Subsequently, using the braking performance test curve, the basic threshold vector θ0=[0.3, 0.6, 0.9] was proportionalized according to the coefficient k m =1+0.15(η) m -1) Scale-down adjustment: When the vehicle is loaded to twice its empty load, the threshold shifts down by approximately 15%, allowing it to enter the warning zone earlier despite increased inertia and significantly longer braking distance. Secondly, the longitudinal slope α is calculated in real-time by fusing pose and a 3D tunnel map. t And convert it into the slope factor ηα =1+sign(α) t )0.1|tanα t The threshold decreases by approximately 10% when descending a 5° slope, and increases by approximately 5% when ascending a 3° slope, reflecting the symmetrical effect of increased inertia downhill and decreased inertia uphill. Finally, to mitigate perception distortion caused by dust, moisture, and echo attenuation, health scores are calculated for the effective point ratio of the lidar, the ultrasonic signal-to-noise ratio, and the visual image sharpness, and normalized to the sensor quality index Q. t ∈[0,1]; when Q t When <0.7, according to k q =1-0.3(0.7-Q) t The threshold is further lowered to compensate for the risk of missed detections caused by a decrease in perceived confidence. The final real-time threshold vector is expressed as:

[0103] θ t =k m ·η α ·k q ·θ0

[0104] The three factors are independent yet multiplicatively coupled, allowing the threshold to decrease by more than 20% under extreme scenarios of heavy load, downhill, and dust, triggering braking earlier. Under ideal scenarios of light load, uphill, and clean conditions, the threshold remains constant or is moderately raised to avoid unnecessary false alarms. This multi-dimensional adaptive mechanism, during a 1.5km tunnel test, advanced the warning trigger by 0.4 seconds on the heavy-load downhill section, significantly reducing the risk of rear-end collisions. Simultaneously, it reduced the false alarm rate by more than 30% on the unloaded return section, effectively balancing safety and operational efficiency.

[0105] As described above, by simultaneously introducing the load inertia factor, slope gravity factor, and sensor mass coefficient within the same formula framework, the warning threshold can adaptively scale according to the instantaneous changes in vehicle braking capacity, driving conditions, and perception reliability: the threshold automatically decreases during heavy loads or downhill driving to ensure early warning when braking distance is extended; the threshold is moderately increased during light loads or uphill driving to avoid excessive conservatism affecting efficiency; and the threshold is further reduced when dust or water vapor causes attenuation of laser and ultrasonic signals to compensate for the risk of missed detection due to decreased perception confidence. Experimental results show that this three-dimensional coupling adjustment mechanism can issue an alarm 0.4 seconds earlier than the fixed threshold scheme in the most dangerous heavy load + downhill + dust scenario, while reducing the false alarm rate by more than 30% in the light load + uphill + clean scenario, thus balancing extreme safety requirements with normal operating efficiency and significantly improving the scenario adaptability and overall reliability of the heavy-load AGV anti-collision warning system in underground tunnels.

[0106] Furthermore, industrial vehicle safety collision avoidance and early warning methods also include:

[0107] Based on the warning level, corresponding control target commands are generated, and the control target commands are input into the closed-loop control algorithm to adjust the vehicle's actuators in real time.

[0108] As described above, in the decision-execution closed loop of this method, after the collision risk index output by the front-end risk assessment module is assigned a clear warning level through adaptive threshold comparison, the control target generation logic is immediately initiated: if the risk is only at the warning level, the controller determines the target based on the current vehicle speed v and the preset comfort deceleration a_soft = -0.5 m / s. 2 Calculate the target speed v_target = v + a_soft·Δt, and ensure a smooth transition with an S-shaped speed curve within a 2-second sliding window to guarantee cargo stability; if the risk rises to a dangerous level, calculate the rapid deceleration amplitude a_hard = -min(2.0, v_target = v + a_soft·Δt) based on the difference between the real-time safe braking distance D_brake and the relative distance d, ΔD = d - D_brake. 2 / 2ΔD)m / s 2 Simultaneously, maintain the lateral acceleration limit <0.3g to prevent rollover, and reduce the target speed to v_target=v+a_hard·Δt within 0.5s; when the risk escalates to the emergency level or the risk gradient Δr>0.25 for two consecutive cycles, directly generate the maximum braking torque command, set the motor torque to zero and trigger the solenoid valve to fully open, ensuring that the braking deceleration reaches -4m / s². 2 The upper limit is set, and audible and visual alarms and wireless emergency stop messages are issued in parallel. The above target commands are then executed in real time by a closed-loop control algorithm: the longitudinal channel adopts feedforward-PID composite control, the feedforward term generates the initial value of motor / brake torque according to the target deceleration, and the PID iteratively adjusts the output according to the real-time vehicle speed error e = v_actual - v_target; the PWM duty cycle of the brake channel is weighted by fuzzy logic in the low-speed stage to avoid lock-up; all commands are sent to the drive inverter, hydraulic braking unit and electronic throttle actuator at a frequency of 1kHz via CAN-FD bus, and the execution feedback (brake pressure, torque, current) is read in real time for closed-loop correction. The actual test results show that this control link can stabilize the deceleration rate of the heavy-duty AGV to -2m / s within 250ms under dangerous conditions. 2 In emergency situations, from issuing the command to generating -4m / s 2 Peak braking time is only 120ms; if the actuator health check detects insufficient braking force, it will automatically add uphill assist braking command or send cloud maintenance warning. In this way, the warning output is seamlessly coupled with the vehicle's power-braking system, which not only ensures rapid braking safety in extreme scenarios, but also takes into account energy consumption and comfort in daily operation.

[0109] The industrial vehicle safety collision avoidance and early warning method provided by this invention, in underground tunnel scenarios lacking satellite navigation, severely obscured by dust, and with significant acoustic multipath interference, uses centimeter-level fusion positioning as the core, temporal depth prediction as the driving force, and multi-factor adaptive thresholds as the adjustment mechanism to construct a complete safety link from multi-source high-fidelity perception—stripper inertial navigation-lidar collaborative positioning—bidirectional LSTM advanced risk quantification—load-slope-signal-noise linkage thresholds to three-level differentiated control + closed-loop self-learning: First, high-frequency integration of the inertial navigation system provides a continuous attitude skeleton, and the lidar ICP / NDT and local map correct drift and are self-consistently fused through EKF, enabling the vehicle to maintain a position error of <10cm even when driving in a 1.5km long GPS-free tunnel; Second, bandpass + Signal processing techniques such as HHT energy envelope and density decision reduce the ultrasonic false alarm rate from 32% to <5%. Together with the laser point cloud after dust intensity filtering, a six-dimensional high-dimensional temporal feature vector is constructed. Bi-LSTM is used to predict the collision probability in the next 0.5–2 seconds, enabling proactive intervention decisions to be made within 120ms. Furthermore, the threshold is dynamically adjusted or increased based on real-time load inertia, slope gravity components, and sensor signal-to-noise ratio, so that the braking distance under heavy load is linked to the threshold, and excessive conservatism is avoided when going uphill with light load, balancing safety and efficiency. Finally, zero-speed updates offset IMU zero bias, and event logs drive Focal-Loss online fine-tuning of the model, ensuring that the algorithm maintains a risk prediction accuracy of >94% over a long period of time under mechanical wear, seasonal environmental changes, and network latency fluctuations. Through actual mine tunnel platoon testing, this solution reduced the false alarm rate to 4.7%, shortened the collision response time to 120ms, and reduced the incidence of major accidents by 63% year-on-year. With the support of 5G-OTA, it achieved low-cost deployment with no hardware modification and iterable software, significantly improving the inherent safety level and operational economy of industrial vehicles under closed extreme working conditions.

[0110] An embodiment of the second aspect of the present invention provides a device 2. In some embodiments of the present invention, such as Figure 7 As shown, the device 2 includes:

[0111] The data acquisition module 201 is used to preprocess the data under a unified spatiotemporal reference to obtain multi-source sensing data.

[0112] The pose fusion module 202 is used to predict the position based on inertial measurement unit data and perform scanning matching based on lidar data, combined with the extended Kalman filter algorithm, to obtain the vehicle's pose estimate.

[0113] The feature analysis module 203 is used to construct a high-dimensional temporal feature vector and perform temporal trend analysis to obtain the collision risk index.

[0114] The early warning decision module 204 is used to adjust the early warning threshold according to the vehicle load status, real-time slope and sensor quality indicators, and to generate early warning information based on the collision risk index to determine the early warning level.

[0115] The device provided by this invention integrates a unified preprocessing data acquisition module, a centimeter-level fusion positioning pose fusion module, a deep learning-driven feature analysis module, and a three-dimensional adaptive threshold warning decision module within a single device. The second aspect of this invention achieves a complete hardware-software closed loop, from raw multi-source sensing data acquisition and high-precision pose estimation with drift suppression, to temporal risk prediction and real-time hierarchical warning. The data acquisition module outputs a clean, denoised multi-source sensing stream with a unified spatiotemporal reference, eliminating cross-sensor asynchronous and clutter interference for downstream algorithms. The pose fusion module, through a combination of inertial navigation, lidar, and EKF, compresses the positioning error of long-distance tunnel driving to the centimeter level, providing a reliable coordinate system for risk calculation. The feature analysis module captures the coupling trend of multiple factors—vehicle, obstacle, braking, and environment—at the high-dimensional temporal vector level, quantifying the collision probability 0.3–0.5 seconds in advance. The warning decision module adjusts the threshold in real time based on load, slope, and sensing signal-to-noise ratio, outputting warning information directly interfaced with the vehicle's power-braking system, balancing heavy-load safety and light-load efficiency. Overall, the device integrates the four major functions of perception, positioning, prediction, and decision-making into a single hardware platform, reducing system integration complexity and communication latency, and significantly improving the collision avoidance response speed, early warning accuracy, and operational economy of industrial vehicles in extreme tunnel environments.

[0116] In the embodiments provided in this disclosure, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0117] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0118] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A method for safety collision avoidance and early warning of industrial vehicles, characterized in that, Includes the following steps: Real-time synchronous acquisition of ultrasonic sensor data, inertial measurement unit data, lidar data and vision camera data on the vehicle, and obtaining multi-source perception data under a unified spatiotemporal reference through preprocessing; The vehicle's predicted position is obtained by calculating the inertial measurement unit data using strapdown inertial navigation system (SINS) position; the absolute pose observation value of the vehicle is obtained by scanning and matching the lidar data using lidar synchronous positioning and mapping method; and the vehicle's predicted position and the absolute pose observation value are fused using extended Kalman filter algorithm to obtain the vehicle's pose estimate. Based on the pose estimation and the multi-source perception data, a high-dimensional temporal feature vector is constructed, including the vehicle's current speed, acceleration, relative distance to obstacles, relative speed, braking performance parameters, and environmental interference indicators. The high-dimensional temporal feature vector is subjected to temporal feature trend analysis to obtain the collision risk index at multiple subsequent time points; Adjust the warning threshold at the current moment based on the vehicle's current load status, real-time slope, and sensor quality indicators; The warning level is determined based on the collision risk index and the real-time adjusted warning threshold, and whether to generate a warning message is determined based on the warning level.

2. The industrial vehicle safety collision avoidance and early warning method according to claim 1, characterized in that, The steps for obtaining multi-source sensing data under a unified spatiotemporal reference through preprocessing specifically include: The energy envelope of the echo signal is obtained by bandpass filtering the ultrasonic sensor data and then performing Hilbert-Huang transform. The energy envelope is used as the input signal for energy threshold determination, and the position of the first main echo is determined in conjunction with the time window density. Based on the location of the first main echo, multipath interference echoes caused by reflections from the tunnel wall are removed from the ultrasonic sensor data.

3. The industrial vehicle safety collision avoidance and early warning method according to claim 2, characterized in that, The step of obtaining multi-source sensing data under a unified spatiotemporal reference through preprocessing further includes: Based on the intensity characteristics of the laser point cloud, intensity threshold filtering is applied to the lidar data for scattering clutter points caused by dust and water vapor. Random noise points are then removed by calculating and employing a statistical outlier filtering method based on the local density of the point cloud.

4. The industrial vehicle safety collision avoidance and early warning method according to claim 3, characterized in that, The step of calculating the inertial measurement unit data based on the strapdown inertial navigation system position specifically includes: Based on the triaxial acceleration and angular velocity data output by the inertial measurement unit, the position prediction state is calculated using strapdown inertial navigation. Zero-speed updates are initiated when the vehicle is stationary at a preset time interval to correct the zero-bias drift error accumulated by inertial navigation in the position prediction state in real time.

5. The industrial vehicle safety collision avoidance early warning method according to claim 3, characterized in that, The step of scanning and matching the lidar data using the lidar synchronous positioning and mapping method specifically includes: Real-time feature extraction is performed on the lidar point cloud data, and a scanning matching method based on the iterative nearest point method is used to register the current frame lidar point cloud with the local map of the tunnel to obtain the absolute pose observation value of the vehicle.

6. The industrial vehicle safety collision avoidance and early warning method according to claim 1, characterized in that, The constructed high-dimensional temporal feature vector includes: Based on the pose estimation and the multi-source perception data, the vehicle's current speed and acceleration are extracted; The relative distance and relative speed of the obstacle in front of the vehicle are determined based on the data from the ultrasonic sensor and the lidar. Based on the vehicle's current load mass and the tunnel's real-time slope, determine the vehicle's current braking performance parameters; Environmental interference indicators are determined based on the real-time signal-to-noise ratio or effective data ratio of the sensors.

7. The industrial vehicle safety collision avoidance and early warning method according to claim 6, characterized in that, The step of performing time series feature trend analysis on the high-dimensional time series feature vector includes: The high-dimensional temporal feature vector is input into a trained bidirectional long short-term memory network model to predict the collision risk index at multiple subsequent time points. By analyzing the trend of the collision risk index over multiple consecutive time periods, changes in the risk level can be determined, enabling early identification and warning response to collision risks.

8. The industrial vehicle safety collision avoidance and early warning method according to claim 1, characterized in that, The steps for adjusting the warning threshold specifically include: Based on the vehicle's current load status, the vehicle's inertia factor is calculated in real time, and the warning threshold is adjusted according to the magnitude of the inertia factor. The warning threshold is dynamically adjusted based on the real-time slope changes in the tunnel area where the vehicle is located. The corresponding warning threshold is adjusted based on the real-time sensor data quality indicators.

9. The industrial vehicle safety collision avoidance early warning method according to any one of claims 1-8, characterized in that, Also includes: Based on the warning level, a corresponding control target instruction is generated, and the control target instruction is input into the closed-loop control algorithm to adjust the vehicle's actuators in real time.

10. An apparatus for implementing the industrial vehicle safety collision avoidance warning method according to any one of claims 1-9, comprising: The data acquisition module is used to preprocess the data under a unified spatiotemporal reference to obtain multi-source sensing data; The pose fusion module is used to predict the position based on the inertial measurement unit data and perform scanning matching based on the lidar data, and combine the extended Kalman filter algorithm to obtain the vehicle's pose estimate. The feature analysis module is used to construct the high-dimensional time-series feature vector and perform time-series trend analysis to obtain the collision risk index; The early warning decision module is used to adjust the early warning threshold according to the vehicle load status, real-time slope and sensor quality indicators, and to determine the early warning level and generate early warning information based on the collision risk index.

Images