AGV navigation quality real-time evaluation and predictive maintenance method

Abstract

The application relates to the technical field of mobile robot navigation control, and discloses an AGV navigation quality real-time evaluation and predictive maintenance method and system, which comprises the following steps: step one, multi-source data of laser point cloud data, acceleration data / angle velocity data, wheeled mileage data, visual image data and environment temperature and humidity data; step two, integrating comprehensive health degree H by dynamically weighting and fusing position confidence C1, scanning matching freshness C2, dynamic interference degree C3 and sensor effectiveness C4; step three, taking a historical feature window as a guide, forming a feature vector by jointly splicing humidity change rate dHr / dt, temperature change rate dT / dt, health degree score sequence H(t) and task complexity C_task; and step four, outputting navigation failure probability P_fail(Δt) of multiple future time points Δt to the outside through a lightweight time sequence prediction model. The application improves the reliability, continuity and overall system resilience of AGV work.

Description

Technical Field

[0001] This invention relates to the technical field of mobile robot navigation control, and specifically to a method for real-time assessment and predictive maintenance of AGV navigation quality. Background Technology

[0002] AGVs are widely used in warehousing and logistics, intelligent manufacturing, and other scenarios. Their automation and intelligence greatly improve the efficiency of material handling and production processes. Existing AGV navigation methods mainly include laser SLAM, QR codes, magnetic strips, and visual navigation, each relying on different environmental features or preset paths to achieve positioning and movement. However, in complex actual operating environments, various dynamic interference factors frequently occur, such as changes in light intensity, alterations in ground friction coefficients, shifts in reference object features, and fluctuations in temperature and humidity. These factors can easily cause deviations or even complete failure of the AGV's navigation and positioning system. Current mainstream solutions typically take recovery measures such as restarting the system, repositioning, and resetting after navigation failure. This approach is a typical "post-event recovery" strategy, which cannot prevent problems before they occur. Therefore, it is difficult to avoid serious accidents such as production line shutdowns and equipment collisions caused by navigation interruptions, posing significant risks to production continuity and safety.

[0003] To address the aforementioned technical issues, current methods employ anomaly detection based on Kalman filtering or setting fixed thresholds. However, these methods still exhibit significant limitations and shortcomings in their application. Firstly, when AGV navigation and positioning systems rely on Kalman filtering or fixed threshold-based anomaly detection methods, they can typically only detect anomalies that have already occurred and lack the ability to predict and assess potential future risks. Secondly, positioning detection methods based on Kalman filtering or setting fixed thresholds for anomaly detection fail to fully integrate multi-dimensional information from multiple sensors (such as the degree of dynamic environmental interference and the effectiveness indicators of each sensor itself), thus failing to construct a comprehensive and integrated health evaluation index for the navigation system. Third, the prediction model structure of the positioning detection method based on Kalman filtering or setting a fixed threshold for anomaly judgment is complex, computationally intensive, and highly dependent on the centralized computing power of the cloud, making it difficult to run in real time and efficiently on the edge computing device of the AGV. Finally, most existing solutions lack proactive early warning mechanisms and strategies for graded intervention based on risk level, making it impossible to achieve forward-looking maintenance operations.

[0004] Therefore, the industry urgently needs to develop a new method and supporting system that can operate in real time at the edge of the AGV, comprehensively assess the real-time health status of the navigation system, accurately predict the probability of future failures, and proactively implement preventive maintenance measures on this basis, so as to fundamentally improve the reliability, continuity and overall system resilience of AGV operation. Summary of the Invention

[0005] The present invention aims to provide a method and system for real-time assessment and predictive maintenance of AGV navigation quality, so as to realize a new method and supporting system that can operate in real time at the edge of the AGV, comprehensively assess the real-time health status of the navigation system, accurately predict the probability of future failures, and proactively implement preventive maintenance measures on this basis, so as to fundamentally improve the reliability, continuity and overall system resilience of AGV operation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: 1) A method for real-time assessment and predictive maintenance of AGV navigation quality, comprising the following steps: Step 1: The LiDAR, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor deployed in the AGV layout collect multi-source data from the perspective of the AGV, including laser point cloud data, acceleration / angular velocity data, wheel mileage data, visual image data and environmental temperature and humidity data. Step 2: Perform a multi-dimensional health assessment on the multi-source data from Step 1. Calculate the positional reliability C1 using the AGV's pose status via a laser SLAM node; calculate the scan matching freshness C2 using the perfect match score s_raw between the AGV's current laser point cloud frame and the pre-stored grid map; calculate the dynamic interference C3 using the AGV's self-motion speed provided by the AGV's laser point cloud frame, IMU inertial measurement unit, and odometer; calculate the sensor effectiveness C4 using the status flags or raw data streams of the sensors (LiDAR, IMU, odometer, camera, and temperature / humidity sensor) used in the AGV layout; and then dynamically weight and fuse the positional reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 to obtain the comprehensive health score H. Step 3: Guided by the historical feature window, calculate the humidity change rate dHr / dt and temperature change rate dT / dt from the environmental temperature and humidity data collected by the AGV in Step 1 using the first-order difference mean; obtain the health score sequence H(t) at this time point using the comprehensive health score H in Step 2 guided by the time axis; quantify the task complexity C_task from the visual image data collected by the AGV, and concatenate the humidity change rate dHr / dt, temperature change rate dT / dt, health score sequence H(t), and task complexity C_task together to form a feature vector; Step 4: Input the feature vector from Step 3 into the pre-trained lightweight time series prediction model. The lightweight time series prediction model outputs the navigation failure probability P_fail(Δt) for multiple future time points Δt. Guided by P_fail(Δt), compare P_fail(Δt) with a preset risk threshold. When P_fail(Δt) is low risk, log the AGV actions. When P_fail(Δt) is medium risk, reduce the AGV speed and send a notification to the host for maintenance. When P_fail(Δt) is high risk, reduce the AGV speed and sequentially request relocation, switch navigation modes, replan the route, and send an emergency notification to the host for maintenance.

[0007] This invention is deployed and run in real time at the edge: it adopts a lightweight time-series prediction model, and the entire data acquisition, multi-dimensional evaluation, probabilistic reasoning and maintenance decision-making process revolves around the data collected locally by the AGV. The computational logic is lightweight and adapted to the computing power conditions of the AGV edge, and the entire process can be completed locally by the AGV without relying on a large amount of computing power transmission and calculation in the cloud, thus meeting the real-time requirements of AGV operation.

[0008] Comprehensive health assessment of navigation systems: The assessment indicators are constructed from four dimensions: core navigation performance, map matching status, environmental interference, and hardware reliability. This not only covers the core output quality of navigation and positioning, but also incorporates the impact of sensor hardware status and external environmental interference. The comprehensive health score is obtained through dynamic weighted fusion, which can completely and accurately reflect the current real-time health status of the navigation system and avoid the one-sidedness of single-dimensional assessment.

[0009] Accurately predict navigation failures to support proactive preventive maintenance: By constructing feature vectors based on historical characteristics, environmental information, and task information, and outputting the probability of navigation failure at multiple future time points through a time-series prediction model, potential navigation failures can be predicted in advance; at the same time, based on different risk levels, a graded maintenance strategy is matched, from low-risk log recording and medium-risk early warning and speed reduction to high-risk sequential execution of relocation, switching navigation modes, and replanning routes, etc., to realize the whole process from state awareness to failure prediction and proactive maintenance.

[0010] Comprehensive improvement of AGV overall operating performance: By replacing traditional reactive maintenance after a failure with predictive proactive maintenance, the probability of sudden navigation failures in AGVs is fundamentally reduced, which can effectively improve the reliability of AGV operations, ensure the continuity of work processes, and enhance the overall resilience of the AGV system to cope with internal and external anomalies, adapting to the needs of long-term stable operation in industrial scenarios.

[0011] 2) According to the method for real-time assessment and predictive maintenance of AGV navigation quality described in 1), wherein the positioning confidence C1 is calculated and output by laser SLAM nodes based on the AGV pose state, the method includes the following steps: Step S11: The point cloud data of the lidar with AGV layout, the acceleration / angular velocity data of the IMU inertial measurement unit, and the wheel odometer data are fused by extended Kalman filtering or graph optimization and input into the lidar SLAM model. Step S12: Extract the covariance trace by using the pose estimate with covariance output from the SLAM algorithm within the laser SLAM model: trace = Σ xx + Σ yy + Σ θθ , where: Σ xx Σ represents the location variance, in m². xx Σ represents the positional uncertainty of the AGV along the x-axis in its vehicle coordinate system; yy (where Σ is the location variance in m²) yy Σ represents the positional uncertainty of the AGV in its vehicle coordinate system along the y-axis; θθ The variance of the yaw angle is expressed in rad², Σ. θθ This indicates the uncertainty of the AGV's orientation angle; Step S13: Normalize the mapping by extracting the covariance trace. The normalization mapping uses logarithmic compression + linear mapping. Preset parameters: When the ideal positioning score is ≈100, trace_min = 0.001; when the positioning is about to fail and the score is ≈0, trace_max = 0.5. When the laser SLAM model does not output with covariance, the matching residual is used instead. The normalization formula is similar, with residual min = 0.02m and max = 0.3m. Output: C1 ∈ [0,100]. The higher the positioning confidence C1, the more reliable the positioning.

[0012] This invention enhances the reliability of positioning by fusing multi-source sensor data. It integrates LiDAR point cloud, IMU inertial navigation data, and wheeled odometer data, and completes the fusion input through extended Kalman filtering or graph optimization. This fully leverages the complementary characteristics of different sensors to provide more comprehensive and accurate raw information for LiDAR SLAM positioning calculations, reducing positioning uncertainty caused by errors in single sensor data from the source. Simultaneously, this invention achieves a quantifiable and measurable output of positioning reliability. By extracting the pose estimation covariance trace, it integrates three core positioning error information: AGV vehicle x-axis position uncertainty, y-axis position uncertainty, and yaw angle uncertainty. This transforms abstract positioning uncertainty into a calculable quantitative indicator, fully covering positioning accuracy errors in both position and orientation dimensions, and more comprehensively reflecting the reliability of AGV pose positioning.

[0013] This invention obtains standardized positioning confidence through normalization mapping. It employs a normalization method combining logarithmic compression and linear mapping to map the covariance trace (or matching residual when there is no covariance) to a standardized interval of [0, 100]. Simultaneously, by presetting parameter thresholds for two critical scenarios—ideal positioning and positioning failure—the semantic consistency of the confidence values ​​is ensured: higher values ​​indicate higher positioning confidence, aligning with intuitive judgment logic in engineering scenarios and facilitating direct reading and use in subsequent AGV navigation decisions. When the laser SLAM model cannot output pose estimation with covariance, it can switch to a normalization calculation logic based on matching residuals, still outputting valid positioning confidence results, avoiding algorithm failure, and improving the robustness and adaptability of the entire positioning confidence calculation process. The output positioning confidence C1 clearly quantifies the current reliability of laser SLAM positioning, providing a clear basis for AGV positioning anomaly judgment and navigation re-trigger logic, helping the AGV system to promptly identify positioning failure risks and ensure the safety of automated driving.

[0014] 3) According to the method for real-time evaluation and predictive maintenance of AGV navigation quality described in 1), the scan matching freshness C2 is calculated and output using the perfect matching score s_raw between the current laser point cloud frame of the AGV and the pre-stored raster map, including the following steps: Step S21: Extract obstacle grids from the original grid map based on the occupancy probability > 0.5 among the occupied grids; calculate the Euclidean distance d(x,y) from each grid to the nearest obstacle using the distance transformation algorithm, and construct the Gaussian likelihood field score function: Where: σ=0.15m is the smoothing parameter; the Gaussian likelihood field score function has a value of 1 at the obstacle and decays exponentially with increasing distance; the score is pre-stored in a floating-point two-dimensional array with a grid resolution of 0.05m, and this pre-calculation only needs to be performed once when the map remains unchanged; Then, the matching score of the current scan is calculated online. Assume that the current laser scan frame, after motion distortion correction, yields the point cloud {(x... i ,y i )} i =1 M The AGV's current estimated pose is (x, y, θ). Transform each scan point to the map coordinate system: Then, the score is obtained by looking up the table using the raster index. i Points outside the map boundaries are not counted as valid points. The matching score s_raw is obtained by aggregating the scores.

[0015] Step S22: In the pre-stored grid map coordinate system, sample and generate N poses at certain intervals to cover the AGV working area, and generate a perfectly matching simulation scan for each sampled pose. For each frame of simulation scan, directly calculate the projection score s_raw by using the same correlation matching method as the real scan with the pose known, and take the average score of all frames as s_max. Step S23: Calculate using the normalized formula: C2=min(100,max(0,(s_raw / s_max)×100)), output: C2 ∈ [0,100], the higher the scan matching freshness C2, the fresher the scan and map match.

[0016] This invention significantly reduces the resource overhead of online computation through offline pre-computation. It pre-extracts obstacle grids with an occupancy probability greater than 0.5 from the original grid map, calculates the Euclidean distance from each grid to the nearest obstacle using a distance transformation algorithm, and constructs a Gaussian likelihood field score function that conforms to the obstacle distance decay law. The pre-computed score is stored as a floating-point two-dimensional array with a fixed grid resolution. Pre-computation only needs to be re-executed when the map changes. One pre-computation can be reused, eliminating the need to recalculate the score field for each match, thus avoiding redundant computation and effectively reducing the computing power consumption of online matching at the AGV vehicle end. The matching process only needs to look up the table through the grid index to obtain the score of the corresponding point, which greatly improves the computational efficiency.

[0017] Meanwhile, a Gaussian likelihood field of σ=0.15m is used as the smoothing parameter to ensure that the score of the obstacle grid position is 1. The score decreases exponentially with the increase of the distance from the point to the obstacle. This can reflect the degree of fit between the scan point and the obstacle on the map, and also avoid the problem of sudden score changes caused by grid misalignment in discrete grid matching, thus improving the stability and discriminativeness of the matching score. In the online matching stage, for the current laser point cloud after motion distortion correction, all scan points are first transformed to the pre-stored grid map coordinate system based on the current estimated pose of the AGV. Then, the corresponding score of each point is obtained by looking up the grid index table. At the same time, invalid points that exceed the map boundary are automatically filtered out. Finally, by aggregating the scores of valid points, the original matching score s_raw that can reflect the true matching degree is stably obtained. The calculation logic is clear and the results are accurate and reliable.

[0018] Here, by uniformly sampling N poses at fixed intervals throughout the AGV's entire working area, a perfectly matching simulated scan frame is generated for each sampled pose. Then, a correlation matching process completely consistent with the real scan is used to calculate the projection score s_raw of the simulated scan. Finally, the average score of all simulated frames is taken as the normalization benchmark s_max. This benchmark covers the matching scenarios of the entire AGV working area, eliminating the influence of different scenarios and different numbers of scan points on the absolute value of the score, and ensuring the consistency of freshness calculation. The scan matching freshness C2 is constrained within a fixed range of [0,100]. The numerical meaning is clear: the higher the C2, the better the matching degree between the current laser scan and the pre-stored grid map, and the higher the freshness of the scene. This can provide an intuitive quantitative indicator for AGV positioning stability assessment, and also provide a clear judgment basis for positioning anomaly detection and map update triggering.

[0019] 4) According to the method for real-time assessment and predictive maintenance of AGV navigation quality described in 1), the dynamic interference degree C3 is calculated and output based on the AGV self-movement speed provided by the AGV laser point cloud frame, IMU inertial measurement unit, and odometer, including the following steps: Step S31: Obtain the total number of valid point clouds N_total in the current frame of the original cloud point frame map: The original point clouds in the original cloud point frame map are cropped sequentially with the ground point removal as the height threshold of 0.1m and the distance point retention ≤15m. At the same time, invalid points with an intensity of 0 or coordinates containing NaN / Inf are removed to obtain the total number of valid point clouds N_total in the current frame. Step S32: Obtain the number of dynamic points N_dynamic using the two-frame difference method for dynamic point detection: Transform both the current frame point cloud P_t and the previous frame point cloud P_{t-1} to the world coordinate system, and perform relative pose motion compensation on P_{t-1} using IMU / odometry or SLAM. For each point in the current frame, use nearest neighbor matching to find the point with the closest Euclidean distance in the compensated point cloud of the previous frame, calculate the distance d_i. If d_i > 0.2m, mark it as a dynamic point; otherwise, the nearest neighbor is missing, and it is not counted as a dynamic point until the next frame confirms. The number of dynamic points N_dynamic is obtained when all the points marked as dynamic points are counted. Step S33: Input the total number of valid point clouds N_total in the current frame from step S31 and the number of dynamic points N_dynamic from step S32 into the formula for calculating the proportion of dynamic points: r_dyn=N_dynamic / N_total Calculate C3: C3 = 100 × (1 - r_dyn). Saturation treatment: if r_dyn > 0.8, then C3 = 0; if r_dyn < 0.05, then C3 = 100. Output: C3 ∈ [0, 100]. The higher the dynamic interference C3, the smaller the dynamic interference.

[0020] In this invention, the preprocessing stage efficiently cleans up invalid point clouds through multiple filtering rules. Ground points are removed with a height threshold of 0.1m, point cloud clipping ranges with a distance ≤15m are retained, and invalid points with an intensity of 0 or containing NaN / Inf abnormal coordinates are eliminated. This can accurately filter out redundant and interfering data in the original point cloud frame, quickly and accurately obtain the total number of valid point clouds N_total in the current frame, and provide a clean and reliable analytical basis for subsequent dynamic interference assessment. It avoids invalid data interfering with the dynamic point statistics results and improves the accuracy of the assessment.

[0021] High-precision dynamic point detection is achieved by combining a two-frame difference method with pose compensation. The point clouds of the current frame and the previous frame are uniformly transformed to the world coordinate system. Motion compensation is performed on the previous frame point cloud by using the relative pose provided by IMU / odometry or SLAM. Then, based on Euclidean distance nearest neighbor matching, dynamic points are determined with a threshold of 0.2m. For points with missing nearest neighbors, a rule is set to temporarily store them for confirmation in the next frame. This not only eliminates the inter-frame position deviation caused by the sensor's own motion and avoids misclassifying static points as dynamic points, but also reduces the undercount of dynamic points. It can accurately count the number of dynamic points N_dynamic in the scene. The accuracy of dynamic point detection is significantly improved compared with the uncompensated frame difference method.

[0022] The dynamic point ratio is calculated based on the total number of effective point clouds and the number of dynamic points. A dynamic interference score C3 in the range of 0-100 is obtained through a transformation formula. At the same time, a saturation rule is set: when the dynamic point ratio exceeds 0.8, the score is directly 0; when the dynamic point ratio is below 0.05, the score is directly 100. The final scoring rule is clear and the result is intuitive. The higher the C3 value, the smaller the scene dynamic interference. This not only achieves the standardized quantification of the degree of dynamic interference, but also adapts to the needs of downstream tasks such as SLAM localization and point cloud mapping for judging the confidence of dynamic scenes. It can directly provide a judgment basis for sensor pose estimation and mapping accuracy optimization in dynamic scenes.

[0023] 5) According to the AGV navigation quality real-time assessment and predictive maintenance method described in 1), the sensor validity C4 is calculated and output by collecting the status flags or raw data streams of each sensor in the AGV layout, including the lidar, IMU inertial measurement unit, odometer, camera, and temperature and humidity sensor, and includes the following steps: Step S41: Set the weight w_i of the sensor status identifier or raw data stream of the LiDAR to 0.35; the weight w_i of the sensor status identifier or raw data stream of the IMU inertial measurement unit to 0.20; the weight w_i of the sensor status identifier or raw data stream of the odometer to 0.15; the weight w_i of the sensor status identifier or raw data stream of the camera to 0.15; and the weight w_i of the sensor status identifier or raw data stream of the temperature and humidity sensor to 0.15. Step S42: Assume the individual validity v_i∈{0,1} of each sensor (LiDAR, IMU, odometer, camera, and temperature / humidity sensor) is 1 if the condition is met, and 0 otherwise. Then, apply the weighted average formula: C4=Σw_i·v_i×100 Output: C4∈[0,100], the higher the sensor effectiveness C4, the more robust the sensor system.

[0024] This invention assigns differentiated weights to different types of sensors based on their importance in the perception system. LiDAR, as the core environmental perception sensor, has the highest weight at 0.35, matching its core role in environmental perception, localization, and mapping. The IMU (Inertial Measurement Unit), as the core pose perception device, has a weight of 0.20, matching its key functions in assisting localization and pose measurement. The remaining odometry, camera, and temperature and humidity sensors each have a weight of 0.15, consistent with the secondary functions of these three types of sensors in assisting perception and environmental monitoring. This weight allocation aligns with the actual role division in typical multi-sensor systems such as mobile robots and autonomous driving, accurately reflecting the impact of different sensors on the overall system availability.

[0025] The working status of each sensor is simplified and transformed into a binary single-item validity identifier. A value of 1 is assigned when the sensor is working normally and meets the status requirements, and a value of 0 is assigned when it is malfunctioning or abnormal, which reduces the complexity of single-sensor status judgment. Based on this, a weighted summation and amplification to a percentage range are used to obtain the overall validity value C4 of the sensor system in the range of 0-100. The calculation logic is simple and clear, with low computational load, which is suitable for the real-time computing needs of embedded platforms. The final output C4 value is directly positively correlated with the soundness of the sensor system. The higher the C4 value, the more high-weight core sensors are in normal working condition, and the higher the soundness of the sensor system. This quantitative result is intuitive and easy to understand. It can provide a clear judgment indicator for real-time status monitoring of the system, and also provide a reliable quantitative input for subsequent sensor fault diagnosis and system fault-tolerant control. It can help developers or users quickly grasp the overall health status of multi-sensor systems.

[0026] 6) According to the AGV navigation quality real-time assessment and predictive maintenance method described in 1), the comprehensive health score H is obtained by dynamically weighting and fusing the positional reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4, including the following steps: Step S51: Obtain σ²_Ci by collecting the variance of the location reliability C1 over a set M seconds, obtain σ²_Ci by collecting the variance of the scan matching freshness C2 over a set M seconds, obtain σ²_Ci by collecting the variance of the dynamic interference C3 over a set M seconds, obtain σ²_Ci by collecting the variance of the sensor effectiveness C4 over a set M seconds, and obtain σ²_Ci by collecting the variance of the sensor effectiveness C4 over a set M seconds. The initial variance is 1. Step S52: Input the variance of the location reliability C1, σ²_Ci, into the basic weight formula w_i'=1 / (σ²_Ci+ε), and calculate the basic weight w_i' of the location reliability C1 within ε=0.01; The variance of the scan matching freshness C2 is obtained as σ²_Ci, which is then input into the basic weight formula w_i'=1 / (σ²_Ci+ε). The basic weight w_i' of the scan matching freshness C2 is calculated within ε=0.01. The variance of the dynamic disturbance degree C3 is obtained as σ²_Ci, which is then input into the basic weight formula w_i'=1 / (σ²_Ci+ε), and the basic weight w_i' of the dynamic disturbance degree C3 is calculated within ε=0.01. The variance of sensor effectiveness C4, σ²_Ci, is input into the basic weight formula w_i'=1 / (σ²_Ci+ε), and the basic weight w_i' of sensor effectiveness C4 is calculated within ε=0.01. Step S53: Add the base weights w_i' of location confidence C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 from step S52 to obtain Σw_j'. Then calculate w_i of location confidence C1, w_i of scan matching freshness C2, w_i of dynamic interference C3, and w_i of sensor effectiveness C4 using normalized weights w_i=w_i' / Σw_j'. Step S54: Add the w_i of the location confidence C1, the w_i of the scan matching freshness C2, the w_i of the dynamic interference C3, and the w_i of the sensor effectiveness C4 obtained by normalizing the weights in step S53 to obtain Σw_i, and output the fusion score formula H=Σw_i·C_i corresponding to the location confidence C1, the scan matching freshness C2, the dynamic interference C3, and the sensor effectiveness C4 respectively: H∈[0,100].

[0027] This invention can automatically generate dynamic fusion weights for each of four indicators—positioning reliability (C1), scan matching freshness (C2), dynamic interference (C3), and sensor effectiveness (C4)—based on their real-time fluctuation characteristics. This effectively avoids the shortcomings of fixed weights, which cannot adapt to different operating conditions of autonomous driving positioning systems. It significantly improves the accuracy of the comprehensive health calculation results in representing the actual health status of the system. By statistically analyzing the variance of each indicator within an M-second sliding window, the stability of the indicator's fluctuation is characterized. The smaller the fluctuation (smaller the variance), the higher the reliability of the indicator, and the larger the basic weight obtained after calculation using the basic weight formula. After normalization, the indicator has a higher proportion in the comprehensive health calculation. Conversely, the larger the fluctuation and the lower the reliability of the indicator, the lower the fusion weight is obtained. It can automatically achieve dynamic adaptation with "a higher proportion of reliable indicators and a lower proportion of unreliable indicators," and the weight allocation logic is more in line with the actual operating rules of the positioning system.

[0028] Meanwhile, a minimum value ε=0.01 is introduced to avoid abnormal calculations when the variance is 0. The initial variance is set to 1 to ensure that the initial calculation can start normally. The entire calculation process has no logical breakpoints and can continuously output stable comprehensive health results. The final comprehensive health score H is standardized and mapped to the [0,100] interval. The value is positively correlated with the system health level and can intuitively reflect the overall health status of the positioning system. This not only meets the needs of real-time status monitoring of the autonomous driving system, but also provides a clear and accurate quantitative basis for subsequent system fault warning and maintenance decisions. By updating the variance and weight of each indicator window by window, the dynamic changes in the operating status of the positioning system can be tracked in real time. Under abnormal conditions such as sensor interference and positioning matching failure, the weight of each indicator can be quickly adjusted so that the comprehensive health score can reflect the abnormal status of the system in a timely and accurate manner, ensuring the accuracy of health assessment under dynamic interference scenarios.

[0029] 7) According to the method for real-time assessment and predictive maintenance of AGV navigation quality described in 1), the task complexity C_task is quantified using visual image data collected by the AGV, including the following steps: With path length factor f len Intersection density factor f cross and turning complexity factor f turn The task complexity C_task is obtained by performing computational fusion, where the total length L of the current task path is calculated, accumulated along the planned path from the starting point to the ending point, and then normalized using the weight formula f. len= min(1,L / L ref ) Obtain the path length factor f len If L>2L ref Take 1; In the map topology, nodes with ≥3 connected edges are considered intersections, and the number N of intersections traversed in the task path is counted.cross Through the normalized weight formula f cross =min(1,N cross / N ref The intersection density factor f is obtained. cross ; For each pair of adjacent segments (p) i−1 →p i ) and (p i →p i+1 ), calculate the steering angle θ i θ i The absolute value is 0~π, and all steering angles Θ are accumulated by summing them up. i θ i The calculation involves the cumulative change in steering angle between all adjacent road segments along the path, based on a set reference cumulative steering angle Θ. ref =π, then pass through the normalized weights f turn =min(1,Θ / Θ ref The turning complexity factor f is obtained. turn ; Using weighted summation, C_task = clip(w1*f len +w2*f cross +w3*f turn The task complexity C_task is obtained by calculating the weights w1, w2, and w3, where w1 = 0.3, w2 = 0.4, and w3 = 0.3.

[0030] This invention extracts the core features affecting task difficulty from three independent dimensions: path scale, road topology, and path turning characteristics. Path length directly reflects the overall execution scale of the task, intersection density reflects the complexity of road connectivity in the area traversed by the path, and turning complexity reflects the frequency of operational adjustments during the path journey. These three factors comprehensively cover the core influencing elements of path task complexity from different levels, avoiding the bias of single-dimensional evaluation.

[0031] Meanwhile, all factors are mapped to the [0,1] interval using the normalization rule min(1, actual value / reference value). Factors exceeding the reference value are directly set to the upper limit of 1 to avoid excessive influence of a single factor value on the overall result, ensuring that factors of different dimensions are on the same quantitative scale, providing a basis for subsequent weighted fusion. The three normalized factors are weighted and summed by preset weights. The weight setting highlights the impact of intersection density on task complexity, matching the pattern that intersections have a greater impact on traffic efficiency and operational difficulty in actual scenarios. Then, the clip function is used to constrain the final task complexity within the [0,1] interval, and the output results are standardized and uniform, which can directly provide intuitive and reliable quantitative basis for downstream applications such as path planning, task difficulty classification, and autonomous driving task allocation.

[0032] 8) According to the AGV navigation quality real-time assessment and predictive maintenance method described in 1), a feature vector is formed by concatenating the humidity change rate dHr / dt, temperature change rate dT / dt, health score sequence H(t), and task complexity C_task, including the following steps: Aligning with the time window within the historical feature window, and using the current time as the right endpoint, feature values ​​from the past 60 time points are taken, with zero padding for values ​​less than 60 seconds. Sliding window normalization is applied to the humidity change rate dHr / dt, temperature change rate dT / dt, health score sequence H(t), and task complexity C_task features. Statistics are pre-calculated from historical data: health score sequence H(t) μ=75, σ=15; temperature change rate dT / dt: μ=0, σ=0.5; humidity change rate dHr / dt: μ=0, σ=0.1; task complexity C_task is in the range [0,1] after normalization.

[0033] This invention integrates and splices four heterogeneous features: humidity change rate, temperature change rate, changes in the health status of the equipment itself (health score sequence), and current workload attributes (task complexity). This integration can fully reflect the environmental fluctuation patterns, health evolution trends, and task load background of the target object over a historical period, providing comprehensive information support for subsequent downstream tasks such as status identification, health prediction, and fault warning, and avoiding judgment bias caused by missing single-dimensional feature information. By using a processing method that extracts a fixed 60 historical time-point features with the current time as the right endpoint and zero-padding when there are fewer than 60 time points, historical observation information of different durations is uniformly converted into a fixed-length feature sequence. This perfectly adapts to the requirement of uniform input tensor dimensions for deep learning and machine learning models, solves the problem of inconsistent historical data lengths in different scenarios, and ensures the stability of model inference and training.

[0034] Sliding window normalization was performed on different types of features based on pre-calculated historical statistics. This eliminated the dimensional differences between different features, avoided gradient interference caused by features with large numerical ranges in model training, and preserved the relative change patterns of features over time. Specifically, the health score sequence was standardized with a mean of 75 and a standard deviation of 15, which conforms to the general normal distribution of health scores and can effectively amplify abnormal fluctuations in health status. The temperature and humidity change rates were centered with a mean of 0 to highlight the degree of deviation of environmental changes. The task complexity was normalized to the [0,1] interval, which unified the dimensional differences of different task rating standards. After the overall normalization process, the convergence speed and prediction accuracy of the subsequent model can be effectively improved.

[0035] 9) According to the AGV navigation quality real-time evaluation and predictive maintenance method described in 1), the pre-trained lightweight time-series prediction model includes the following steps: Step S1: Construct the input layer of the lightweight time series prediction model. The input layer includes a first LSTM layer for extracting local time series and a second LSTM layer for capturing long-term dependencies at a higher level of abstraction. The first LSTM layer is constructed using 32 independent LSTM memory blocks, each responsible for extracting different dimensions of time series patterns from the input sequence, resulting in an output dimension of 32. Simultaneously, each LSTM memory block is passed through a forgetting gate f. t Input gate i t Output gate o t and cell state C t Formula output is stored in long-term memory; The second LSTM layer uses 32 independent LSTM memory blocks, each responsible for extracting different dimensions of temporal patterns from the input sequence, constructing an output dimension of 32. Simultaneously, each LSTM memory block is passed through a forget gate f. t Input gate i t Output gate o t and cell state C t The formula outputs long-term memory and is used to process higher-level temporal features of the first LSTM layer; Long-term memory formulas in the first and second LSTM layers: Step S2: Construct a fully connected layer in the lightweight time series prediction model. The fully connected layer has 16 neurons, and the number of neurons is used to fit the nonlinear features in the health sequence using the ReLU activation function. Step S3: Construct an output layer in the lightweight time series prediction model. The output layer has 3 neurons, corresponding to P30s, P60s, and P120s respectively. The number of neurons is used to learn the reasonable relationship of P30≤P60≤P120 using the Sigmoid activation function.

[0036] This invention constructs a lightweight temporal prediction module using an input layer, a fully connected layer, and an output layer. The input layer employs a dual-layer structure with 32 independent LSTM memory blocks in each layer (first and second LSTM layers), enabling hierarchical temporal feature extraction. The first LSTM layer accurately captures local temporal patterns of different dimensions in the input sequence, while the second LSTM layer further extracts higher-level, long-term dependency features based on the output of the first LSTM layer. Simultaneously, the long-term memory output of the LSTM gating mechanism is retained, effectively remembering key long-term information in the temporal sequence and avoiding the gradient vanishing and long-term dependency loss problems encountered in traditional temporal models during long-sequence training. This approach addresses the capture requirements of temporal features at different scales.

[0037] In the fully connected layer, only 16 neurons are set, along with the ReLU activation function. This controls the overall number of model parameters, avoiding the problems of high inference latency and excessive storage resources caused by an overly complex model. At the same time, it can effectively fit the nonlinear features in the health time series data, retaining the ability to model the nonlinear changes of time series data, and achieving a balance between lightweightness and fitting ability. In the output layer, three neurons are set to correspond to prediction tasks of three different durations, P30s, P60s, and P120s, respectively. The Sigmoid activation function is used to guide the model to learn the business logic relationship of P30≤P60≤P120. It can output multi-step time series prediction results that conform to actual business rules, ensuring the rationality of the prediction results. It can directly output health time series predictions of different duration dimensions, meeting the application requirements of multi-step time series prediction.

[0038] The overall model structure has controllable parameters, which ensures the accuracy of time series prediction while achieving model lightweighting, reducing the computational resource consumption of model inference, and making it more suitable for deployment and application in resource-constrained scenarios such as edge devices and end devices, thus balancing prediction performance and deployment practicality.

[0039] 10) According to the AGV navigation quality real-time assessment and predictive maintenance method described in 1), the navigation failure probability P_fail(Δt) for multiple future time points Δt is output outward through a lightweight time series prediction model, including the following steps: Step S4-1, Training Phase: Guided by the constructed lightweight time-series prediction model, a 60-second window is extracted from historical data of the AGV. Based on whether a navigation failure event will occur within the next Δt seconds, the navigation failure time is labeled as y=1 or y=0. Weighted binary cross-entropy loss is used for training. Where: α = 3-5 is the weight of the positive sample; Step S4-2, Inference Phase: A new feature window generated every 1 second is input into the model, and forward propagation yields three Sigmoid outputs: The learned temporal patterns are: a continuous decline in health, a sharp drop in location confidence, a continuously high dynamic interference, and sudden environmental changes all lead to an increase in P_fail; Step S4-3, Probability Calibration: After training, apply temperature scaling to make the output probability consistent with the actual frequency. When P_fail(60s)≥0.7, the actual failure probability is about 85%-90%, and when P_fail(60s)≤0.4, the actual failure probability is less than 10%.

[0040] The lightweight time-series prediction model of this invention can effectively output the navigation failure probability P_fail(Δt) of AGV at multiple future time points Δt. The difference from the traditional threshold judgment is that it comprehensively considers the trend of the past 60 seconds, and only increases P_fail when the trend points to deterioration and the current state is poor, thus avoiding false alarms caused by instantaneous jitter.

[0041] Meanwhile, this invention extracts a 60-second historical data window and labels positive and negative samples based on whether navigation failures occur within the next Δt seconds. It employs a weighted binary cross-entropy loss and matches the actual scenario with a low proportion of navigation failure samples using a positive sample weight of 3-5. This solves the problem of imbalanced positive and negative samples and can specifically drive the model to learn the pre-navigation failure feature patterns, ensuring the effectiveness of model training.

[0042] During the inference phase, the updated feature window only needs to be input into the model every second. The model can quickly output the navigation failure probabilities at multiple time points through a single forward propagation, resulting in high inference efficiency that is suitable for the real-time operation requirements of AGVs. Furthermore, the model has learned the core associated temporal patterns of navigation failures: it can identify several risk scenarios, such as a continuous decline in health, a sharp drop in positioning confidence, a continuous increase in dynamic interference, and sudden environmental changes. The corresponding output increases the failure probability, accurately reflecting the actual risk level and conforming to the actual evolution of AGV navigation failures.

[0043] After probability calibration is completed through temperature scaling during the probability calibration phase, the model output probability and the actual navigation failure frequency have a high degree of matching and strong probability quantification accuracy: when the model outputs a navigation failure probability P_fail(60s) ≥ 0.7 in the next 60 seconds, the corresponding actual navigation failure probability can reach 85%-90%; when the model outputs P_fail(60s) ≤ 0.4, the actual navigation failure probability is less than 10%, and the confidence level of the output probability is high, which can provide a reliable decision-making basis for early risk warning and operation and maintenance intervention of AGV.

[0044] 11) An AGV navigation quality real-time assessment and predictive maintenance system, comprising the AGV navigation quality real-time assessment and predictive maintenance method described in 1)-10) above, including: The data acquisition module is configured to connect to the lidar, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor of the AGV layout to collect multi-source data from the perspective of the AGV, including lidar point cloud data, acceleration data / angular velocity data, wheel mileage data, visual image data and environmental temperature and humidity data. The health assessment module is configured to perform multi-dimensional health assessments on multi-source data. It calculates and outputs the localization reliability C1 based on the AGV's pose status using laser SLAM nodes; it calculates and outputs the scan matching freshness C2 based on the perfect match score s_raw between the AGV's current laser point cloud frame and a pre-stored grid map; it calculates and outputs the dynamic interference C3 based on the AGV's self-movement speed provided by the AGV's laser point cloud frame, IMU inertial measurement unit, and odometer; it calculates and outputs the sensor effectiveness C4 based on the status flags or raw data streams of the sensors (LiDAR, IMU inertial measurement unit, odometer, camera, and temperature / humidity sensor) used in the AGV layout; and it dynamically weights and fuses the localization reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 to obtain the comprehensive health score H. The time series prediction module is configured to run a pre-trained lightweight time series prediction model with feature vectors as input, and output the navigation failure probability P_fail(Δt) for multiple future time points Δt. The early warning decision module is configured to compare P_fail(Δt) with a preset risk threshold, guided by P_fail(Δt). When P_fail(Δt) is low risk, the AGV action is logged; when P_fail(Δt) is medium risk, the AGV speed is reduced and a notification is sent to the upper level for maintenance; when P_fail(Δt) is high risk, the AGV speed is reduced and the upper level is requested to reposition, switch navigation modes, replan the route, and issue an emergency notification in sequence for maintenance. The communication interface is configured to interact with the AGV master controller and scheduling system via Modbus TCP, MQTT, or shared memory.

[0045] The system can be integrated into the existing industrial control computer of the AGV as a software service, or it can be used as an independent edge computing box to communicate with the AGV main controller via Ethernet.

[0046] Compared with the prior art, the present invention also has the following technical effects: From passive recovery to proactive prevention: Early warning and intervention before navigation failure occurs, significantly reducing the number of actual failures and improving AGV operation efficiency (experiments show that the accuracy of high-risk warning is ≥85%, and the failure avoidance rate after intervention is ≥70%); Lightweight edge computing: Model size ≤15MB, inference ≤20ms, CPU usage increment ≤5%, can run in real time on AGV embedded platform, without relying on the cloud.

[0047] More comprehensive multi-dimensional fusion assessment: comprehensively considers positioning, scanning, dynamic interference, and sensor status to avoid misjudgment based on a single threshold; graded early warning and differentiated intervention: take different actions according to the risk level to avoid overreaction and cope with emergencies; model self-evolution: continuously adapt to changes in the field environment through incremental learning; standardized early warning code: facilitates multi-system parsing and automated processing. Attached Figure Description

[0048] Figure 1 This is a flowchart of a method for real-time assessment and predictive maintenance of AGV navigation quality according to the present invention.

[0049] Figure 2 This is a framework diagram of an AGV navigation quality real-time assessment and predictive maintenance system according to the present invention. Detailed Implementation

[0050] The following detailed description illustrates the specific implementation method: See the example. Figure 1 As shown in this embodiment, a method for real-time assessment and predictive maintenance of AGV navigation quality includes the following steps: Step 1: The LiDAR, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor deployed in the AGV layout collect multi-source data from the perspective of the AGV, including laser point cloud data, acceleration / angular velocity data, wheel mileage data, visual image data and environmental temperature and humidity data. Step 2: Perform a multi-dimensional health assessment on the multi-source data from Step 1. Calculate the positional reliability C1 using the AGV's pose status via a laser SLAM node; calculate the scan matching freshness C2 using the perfect match score s_raw between the AGV's current laser point cloud frame and the pre-stored grid map; calculate the dynamic interference C3 using the AGV's self-motion speed provided by the AGV's laser point cloud frame, IMU inertial measurement unit, and odometer; calculate the sensor effectiveness C4 using the status flags or raw data streams of the sensors (LiDAR, IMU, odometer, camera, and temperature / humidity sensor) used in the AGV layout; and then dynamically weight and fuse the positional reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 to obtain the comprehensive health score H. Step 3: Guided by the historical feature window, calculate the humidity change rate dHr / dt and temperature change rate dT / dt from the environmental temperature and humidity data collected by the AGV in Step 1 using the first-order difference mean; obtain the health score sequence H(t) at this time point using the comprehensive health score H in Step 2 guided by the time axis; quantify the task complexity C_task from the visual image data collected by the AGV, and concatenate the humidity change rate dHr / dt, temperature change rate dT / dt, health score sequence H(t), and task complexity C_task together to form a feature vector; Step 4: Input the feature vector from Step 3 into the pre-trained lightweight time series prediction model. The lightweight time series prediction model outputs the navigation failure probability P_fail(Δt) for multiple future time points Δt. Guided by P_fail(Δt), compare P_fail(Δt) with a preset risk threshold. When P_fail(Δt) is low risk, log the AGV actions. When P_fail(Δt) is medium risk, reduce the AGV speed and send a notification to the host for maintenance. When P_fail(Δt) is high risk, reduce the AGV speed and sequentially request relocation, switch navigation modes, replan the route, and send an emergency notification to the host for maintenance.

[0051] In step two, the positioning confidence C1 is calculated and output using the laser SLAM node based on the AGV's pose state, and includes the following steps: Step S11: The point cloud data of the lidar with AGV layout, the acceleration / angular velocity data of the IMU inertial measurement unit, and the wheel odometer data are fused by extended Kalman filtering or graph optimization and input into the lidar SLAM model. Step S12: Output the model with covariance using the SLAM algorithm within the laser SLAM model. (Or 6×6, taking position and yaw angle components) pose estimation. In this embodiment, the AGV's pose (position x, y and orientation angle θ) is estimated in real time by the laser SLAM module, fusing the following sensor data:

[0052] The SLAM algorithm fuses the above data through extended Kalman filtering or graph optimization, and outputs a pose estimate with covariance, where the covariance matrix is ​​the direct input for C1 calculation.

[0053] Recommended principles regarding sensor installation location: LiDAR: Installed at the front or top center of the AGV body, 0.2m-0.5m above the ground (to avoid ground clutter); horizontal scanning range ≥270° (or 360°), and not obstructed by the vehicle body itself. No absolute coordinate limitations are required, but precise calibration of the extrinsic parameters of the radar and the vehicle body is necessary. If installed eccentrically, the covariance trace remains unchanged under coordinate system rotation; therefore, C1 is insensitive to eccentricity.

[0054] IMU: It should be placed as close as possible to the center of mass of the AGV (usually the center of the vehicle body) to reduce the additional acceleration caused by centrifugal force; the shaft system should be aligned with the vehicle body shaft system (deviation angle <5°), otherwise it needs to be calibrated online.

[0055] Odometer: Integrated into the drive wheel motor or axle, it is necessary to ensure accurate wheel diameter parameters and no severe tire slippage.

[0056] If the installation location deviates significantly from the above principles, the extrinsic parameters must be clearly calibrated in the SLAM configuration, and the covariance trace must be verified to be <0.01 through static testing; otherwise, the accuracy of C1 may be affected.

[0057] To extract the covariance trace: trace = Σ xx + Σ yy + Σ θθ In SLAM pose estimation for AGVs, the covariance matrix Σ reflects the uncertainty of the localization result.

[0058] Wherein: Σ xx Σ represents the location variance, in m². xx This represents the positional uncertainty of the AGV along its x-axis in its vehicle coordinate system (usually the forward direction); the larger the value, the less certain the AGV is about "how much it has moved along the forward direction." For example, in a corridor environment, insufficient features of the LiDAR along the forward direction (such as parallel side walls) can lead to Σ xx Increase.

[0059] Σ yy (where Σ is the location variance in m²) yy This represents the positional uncertainty of the AGV in its vehicle coordinate system along the y-axis (usually lateral, i.e., the horizontal direction perpendicular to the direction of travel). This value increases when the AGV passes through narrow passages, where the constraints in the left and right directions are strong. However, if it encounters open areas or dynamic obstacles, the lateral positioning may drift.

[0060] Σ θθ The variance of the yaw angle is expressed in rad², Σ. θθThis represents the uncertainty of the AGV's orientation angle (yaw angle); the larger the value, the more uncertain the AGV is about "which direction it is facing". For example, when the environment scanned by the lidar is symmetrical (such as a long corridor) or rotationally symmetrical (such as a circular shelf), the angle estimation is prone to degradation, and Σθθ will increase significantly.

[0061] Therefore, the covariance trace (trace = Σ) xx + Σ yy + Σ θθ This integrates uncertainties from three directions and is a single scalar measure of the overall positioning reliability of an AGV. The smaller the trace size, the more accurate the positioning; the larger the trace size, the less reliable the positioning.

[0062] Step S13: Normalize the mapping by extracting the covariance trace. The normalization mapping uses logarithmic compression + linear mapping. Preset parameters: When the ideal positioning score is ≈100, trace_min = 0.001; when the positioning is about to fail and the score is ≈0, trace_max = 0.5. When the laser SLAM model does not output with covariance, the matching residual is used instead. The normalization formula is similar, with residual min = 0.02m and max = 0.3m. Output: C1 ∈ [0,100]. The higher the positioning confidence C1, the more reliable the positioning.

[0063] In step two, the scan matching freshness C2 is calculated and output using the perfect match score s_raw between the current laser point cloud frame of the AGV and the pre-stored raster map. The scan matching freshness C2 reflects the degree of matching between the current laser scan and the map, indicating whether the environmental features are still consistent with the terrain. Figure 1 The method employs correlation-based matching, also known as likelihood field matching, which calculates the average response value of the projected points as the matching score by projecting the scanned point cloud onto a pre-built likelihood field map.

[0064] Therefore, scanning to match freshness C2 includes the following steps: Step S21: Extract obstacle grids from the original grid map based on the occupancy probability > 0.5 among the occupied grids; calculate the Euclidean distance d(x,y) from each grid to the nearest obstacle using the distance transformation algorithm, and construct the Gaussian likelihood field score function: Where: σ=0.15m, is the smoothing parameter; the Gaussian likelihood field score function has a value of 1 at obstacles and decays exponentially with increasing distance; the score is pre-stored in a floating-point two-dimensional array with a grid resolution of 0.05m, and this pre-calculation only needs to be performed once when the map remains unchanged; then, the matching score of the current scan is calculated online, assuming that the current laser scan frame, after motion distortion correction, yields the point cloud {(x i ,y i )}i =1 M The AGV's current estimated pose is (x, y, θ). Transform each scan point to the map coordinate system: Then, the score is obtained by looking up the table using the raster index. i Points outside the map boundaries are not counted as valid points. The matching score s_raw is obtained by aggregating the scores. Optimization and acceleration: Downsampling (taking one sample every 3 points) and parallel acceleration using the ARM NEON instruction set.

[0065] Step S22: In the pre-stored grid map coordinate system, N poses are sampled at certain intervals to cover the AGV working area, and a perfectly matching simulation scan is generated for each sampled pose. For each frame of simulation scan, the projection score s_raw is directly calculated by using the same correlation matching method as the real scan with the pose known, and the average score of all frames is taken as s_max. The present invention uses the average value to avoid local underestimation caused by map boundaries or sparse areas.

[0066] In step S22, the perfectly matched simulation scan refers to a set of virtual laser scan data generated in a computer simulation environment based on a known accurate map and an ideal sensor model. The geometric relationship between these scan data and the map is completely known and error-free, meaning that the scan points fall exactly on the obstacle boundaries in the map, and there is no measurement noise, motion distortion, or environmental changes.

[0067] Generation method: Input: A pre-stored high-precision grid map (the same map used by the AGV); a series of known, precise AGV poses (e.g., discrete poses uniformly sampled in the map coordinate system, covering all possible locations and orientations of the AGV).

[0068] Simulated Scan Generation: For each given pose (x, y, θ), a ray casting algorithm is used to simulate the LiDAR measurement process on a map. Rays are emitted from the LiDAR origin according to the LiDAR's angular resolution (e.g., 0.2°). The intersection of each ray with the nearest obstacle grid in the map is calculated, and the distance from this intersection to the origin is the simulated distance value. Because the map is accurate, the pose is precise, and no noise is added, the generated scan point cloud theoretically matches the map perfectly.

[0069] Here, it's necessary to explain why we don't use the maximum value of the actual scan: In real-world scanning, sensor noise and quantization errors exist, and the raw value (s_raw) often fails to reach the theoretical value from simulation. Using simulated scanning to calibrate s_max provides a fixed and reproducible reference benchmark, making the C2 normalized value comparable across different maps and AGVs.

[0070] Step S23: Calculate using the normalized formula: C2=min(100,max(0,(s_raw / s_max)×100)), output: C2 ∈ [0,100], the higher the scan matching freshness C2, the fresher the scan and map match.

[0071] Optional correction: When the AGV speed is <0.1m / s for more than 30 seconds, C2 will decay at a rate of 5% per minute until it resumes movement.

[0072] In step two, the dynamic interference degree C3 is calculated and output using the AGV self-movement speed provided by the AGV laser point cloud frame, IMU inertial measurement unit, and odometer. The dynamic interference degree C3 evaluates the degree of interference of dynamic obstacles on navigation. C3 = 100 × (1 – dynamic point ratio). The dynamic interference degree C3 includes the following steps: Step S31: Obtain the total number of valid point clouds N_total in the current frame of the original cloud point frame map: The original point clouds in the original cloud point frame map are cropped sequentially with the ground point removal as the height threshold of 0.1m and the distance point retention ≤15m. At the same time, invalid points with an intensity of 0 or coordinates containing NaN / Inf are removed to obtain the total number of valid point clouds N_total in the current frame. Step S32: Detect dynamic points using the two-frame difference method to obtain the number of dynamic points N_dynamic: Transform both the current frame point cloud P_t and the previous frame point cloud P_{t-1} to the world coordinate system, and perform relative pose motion compensation on P_{t-1} using IMU / odometry or SLAM. For each point in the current frame, use nearest neighbor matching to find the point with the closest Euclidean distance in the compensated point cloud of the previous frame, calculate the distance d_i. If d_i > 0.2m, mark it as a dynamic point; otherwise, the nearest neighbor is missing, and it is not counted as a dynamic point until the next frame confirms. The number of dynamic points N_dynamic is obtained when all points are marked as dynamic. Step S33: Input the total number of valid point clouds N_total in the current frame from step S31 and the number of dynamic points N_dynamic from step S32 into the formula for calculating the proportion of dynamic points: r_dyn=N_dynamic / N_total Calculate C3: C3 = 100 × (1 - r_dyn). Saturation treatment: if r_dyn > 0.8, then C3 = 0; if r_dyn < 0.05, then C3 = 100. Output: C3 ∈ [0, 100]. The higher the dynamic interference C3, the smaller the dynamic interference.

[0073] Among them, dynamic point detection (two-frame difference method): transform the current point cloud P_t to the world coordinate system according to the AGV pose, transform the previous point cloud P_{t-1} to the current coordinate system, calculate the Euclidean distance to the nearest neighbor point in the previous frame for each point, and mark it as a dynamic point if the distance is greater than the threshold d_dyn (0.2m).

[0074] The basis for determining the dynamic point detection threshold d_dyn = 0.2m is as follows: This threshold was determined through a combination of theoretical derivation and experimental calibration, as follows: Theoretical Derivation: Considering the displacement of stationary points caused by the AGV's own motion (displacement between two frames ≤ 0.1m when speed ≤ 1.0m / s), LiDAR ranging noise (±0.02m), and pose estimation error (±0.01m), the theoretical upper limit of the stationary point distance difference is approximately 0.15m. Dynamic objects (pedestrians, forklifts, etc.) typically move at speeds of 0.5~1.5m / s, with displacement between two frames reaching 0.05~0.15m, which approaches 0.2m after adding noise. Setting the threshold to 0.2m allows for the differentiation between static and dynamic objects while preserving a margin.

[0075] Experimental calibration: In a static scene, 100 frames of point cloud data were collected, and the distance difference of 99% of static points was ≤0.18m. In a dynamic scene, when pedestrians crossed at different speeds, the minimum distance difference of dynamic points was approximately 0.22m. ROC curve analysis showed that the optimal balance was achieved at 0.2m (92% accuracy and 88% recall).

[0076] Adjustability: This threshold can be adjusted based on sensor accuracy and AGV maximum speed (e.g., it can be reduced to 0.15m for high-precision LiDAR and increased to 0.25m for high-speed AGV). Therefore, this embodiment uses 0.2m as the default value, which is suitable for most warehouse AGV scenarios.

[0077] In step two, the sensor validity C4 is calculated and output based on the status flags or raw data streams of the LiDAR, IMU, odometer, camera, and temperature and humidity sensor in the AGV layout. Sensor validity C4 evaluates whether each key sensor is working properly and calculates the overall validity using a weighted average. Sensor validity C4 includes the following steps: Step S41: Set the weight w_i of the sensor status identifier or raw data stream of the LiDAR to 0.35; the weight w_i of the sensor status identifier or raw data stream of the IMU inertial measurement unit to 0.20; the weight w_i of the sensor status identifier or raw data stream of the odometer to 0.15; the weight w_i of the sensor status identifier or raw data stream of the camera to 0.15; and the weight w_i of the sensor status identifier or raw data stream of the temperature and humidity sensor to 0.15. Define the sensor set and weights (adjustable):

[0078] Step S42: Assume the individual validity v_i∈{0,1} of each sensor (LiDAR, IMU, odometer, camera, and temperature / humidity sensor) is 1 if the condition is met, and 0 otherwise. Then, apply the weighted average formula: C4=Σw_i·v_i×100 Output: C4∈[0,100], the higher the sensor effectiveness C4, the more robust the sensor system.

[0079] Therefore, in step two, the location reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 are dynamically weighted and fused to obtain the comprehensive health score H, which includes the following steps: Step S51: Obtain σ²_Ci by collecting the variance of the location reliability C1 over a set M seconds, obtain σ²_Ci by collecting the variance of the scan matching freshness C2 over a set M seconds, obtain σ²_Ci by collecting the variance of the dynamic interference C3 over a set M seconds, obtain σ²_Ci by collecting the variance of the sensor effectiveness C4 over a set M seconds, and obtain σ²_Ci by collecting the variance of the sensor effectiveness C4 over a set M seconds. The initial variance is 1. Step S52: Input the variance of the location reliability C1, σ²_Ci, into the basic weight formula w_i'=1 / (σ²_Ci+ε), and calculate the basic weight w_i' of the location reliability C1 within ε=0.01; The variance of the scan matching freshness C2 is obtained as σ²_Ci, which is then input into the basic weight formula w_i'=1 / (σ²_Ci+ε). The basic weight w_i' of the scan matching freshness C2 is calculated within ε=0.01. The variance of the dynamic disturbance degree C3 is obtained as σ²_Ci, which is input into the basic weight formula w_i'=1 / (σ²_Ci+ε), and the basic weight w_i' of the dynamic disturbance degree C3 is calculated within ε=0.01. The variance of sensor effectiveness C4, σ²_Ci, is input into the basic weight formula w_i'=1 / (σ²_Ci+ε), and the basic weight w_i' of sensor effectiveness C4 is calculated within ε=0.01. Step S53: Add the base weights w_i' of location confidence C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 from step S52 to obtain Σw_j'. Then calculate w_i of location confidence C1, w_i of scan matching freshness C2, w_i of dynamic interference C3, and w_i of sensor effectiveness C4 using normalized weights w_i=w_i' / Σw_j'. Step S54: Add the w_i of the location confidence C1, the w_i of the scan matching freshness C2, the w_i of the dynamic interference C3, and the w_i of the sensor effectiveness C4 obtained by normalizing the weights in step S53 to obtain Σw_i, and output the fusion score formula H=Σw_i·C_i corresponding to the location confidence C1, the scan matching freshness C2, the dynamic interference C3, and the sensor effectiveness C4 respectively: H∈[0,100].

[0080] In step three, the task complexity C_task is obtained by quantizing the visual image data collected by the AGV, including the following steps: With path length factor f len Intersection density factor f cross and turning complexity factor f turn The task complexity C_task is obtained by performing computational fusion, where the total length L of the current task path is calculated, accumulated along the planned path from the starting point to the ending point, and then normalized using the weight formula f. len= min(1,L / L ref ) Obtain the path length factor f len If L>2L ref Take 1; In the map topology, nodes with ≥3 connected edges are considered intersections, and the number N intersections traversed in the task path is counted. cross By using the normalized weight formula f cross =min(1,N cross / N ref The intersection density factor f is obtained. cross ; For each pair of adjacent segments (p) i−1 →p i ) and (p i →p i+1 ), calculate the steering angle θ i θ i The absolute value is 0~π, and all steering angles Θ are accumulated by summing them up. i θ iThe calculation involves the cumulative change in steering angle between all adjacent road segments along the path, based on a set reference cumulative steering angle Θ. ref =π, then pass through the normalized weights f turn =min(1,Θ / Θ ref The turning complexity factor f is obtained. turn ; Using weighted summation, C_task = clip(w1*f len +w2*f cross +w3*f turn The task complexity C_task is obtained by calculating the weights w1, w2, and w3, where w1 = 0.3, w2 = 0.4, and w3 = 0.3.

[0081] Therefore, in step three, a feature vector is constructed by concatenating the health score sequence H(t) over the past N seconds (N=60), the environmental change rate (temperature change slope, humidity change slope), and the task complexity (path length, number of intersections). The specific method is as follows: Feature composition: Each time step (sampling interval 1 second) contains 4 components: H(t), temperature change rate dT / dt (mean of first difference over the past 10 seconds), humidity change rate dHr / dt, and task complexity C_task (normalized [0,1]). Time window alignment: Using the current time as the right endpoint, take the feature values ​​of the past 60 time points, and fill with zero if the time is less than 60 seconds.

[0082] Feature preprocessing and normalization: Sliding window normalization is used, and the statistics are pre-calculated from historical data: H: μ=75, σ=15; dT / dt: μ=0, σ=0.5; dHr / d: μ=0, σ=0.1; C_task is already in [0,1].

[0083] Missing value handling: For a single time step missing value, linear interpolation is used; for consecutive missing values ​​≤ 3 seconds, forward padding is used; for consecutive missing values ​​> 3 seconds, the missing feature is set to 0, and a mask feature vector (length 4, 1 indicates valid, 0 indicates missing) is added, and the final tensor shape is [60, 8] (4 original features + 4 mask features).

[0084] Efficiency optimization: A circular buffer is used to store the features of the most recent 60 seconds, and the buffer is updated incrementally every time new data is received (1Hz).

[0085] Meanwhile, in step four, a pre-trained lightweight time series prediction model is used, including the following steps: Step S1: Construct the input layer of the lightweight time series prediction model. The input layer has a shape of [batch, 60, 8] (60 time steps, each step containing 4 original features + 4 masked features). The input layer includes a first LSTM layer for extracting local temporal patterns (such as short-term fluctuations) and a second LSTM layer for capturing long-term dependencies (such as the overall downward trend of health within 60 seconds) at a higher level of abstraction. The first LSTM layer is constructed using 32 independent LSTM memory blocks, each responsible for extracting different dimensions of temporal patterns from the input sequence, with an output dimension of 32. Each LSTM memory block is then passed through a forgetting gate f. t Input gate i t Output gate o t and cell state C t Formula output is stored in long-term memory; return_sequences=True: Output the hidden state at each time step, with an output shape of [batch, 60, 32].

[0086] dropout=0.2: Randomly drop 20% of neurons at each time step's input and recurrent connections to prevent overfitting.

[0087] The second LSTM layer uses 32 independent LSTM memory blocks, each responsible for extracting different dimensions of temporal patterns from the input sequence, constructing an output dimension of 32. Simultaneously, each LSTM memory block is passed through a forget gate f. t Input gate i t Output gate o t and cell state C t The formula outputs long-term memory and is used to process higher-level temporal features of the first LSTM layer; return_sequences=False: Only output the hidden state of the last time step, with an output shape of [batch, 32].

[0088] dropout=0.2: Same regularization strategy as the first layer.

[0089] Long-term memory formulas in the first and second LSTM layers: The selection criteria for 32 units were based on balancing expressive power and computational efficiency. Accuracy decreased by approximately 8% with 16 units, while 64 units doubled the file size while only improving accuracy by 1%. 32 units represented the optimal compromise.

[0090] Step S2: Construct a fully connected layer in a lightweight time-series prediction model. The fully connected layer has 16 neurons, and the number of neurons is fitted with the nonlinear features in the health sequence using the ReLU activation function. Activation function: ReLU. Definition: ReLU(z) = max(0,z). Function: Introduces nonlinearity, enabling the network to fit nonlinear features in the health sequence (such as sudden drops or oscillatory recoveries); generates sparse activation (approximately 50% of neurons output zero), reducing the risk of overfitting; computationally simple, with no exponential operations, suitable for real-time inference at edge computing environments.

[0091] Step S3: Construct the output layer of the lightweight time series prediction model. The output layer has 3 neurons, corresponding to P30s, P60s, and P120s respectively. The number of neurons learns the reasonable relationship of P30≤P60≤P120 using the activation function Sigmoid. Activation function: Sigmoid. Definition: σ(z) = 1 / (1 + e−z), output range (0,1), can be interpreted as a probability. Reason for selection: The three output probabilities are independent (not mutually exclusive), allowing the model to learn a reasonable relationship of P30≤P60≤P120; it is smooth and differentiable, suitable for gradient descent optimization.

[0092] Therefore, a lightweight time series prediction module (LSTM network or TinyTimeMixer) is constructed by using an input layer, a fully connected layer, and an output layer. The model size is ≤15MB and the single inference time is ≤20ms.

[0093] In step four, a lightweight temporal prediction model outputs the navigation failure probability P_fail(Δt) for multiple future time points Δt, where P_fail(Δt) represents the probability of a navigation failure event occurring within the next Δt seconds. This probability is derived in real-time through a temporal prediction model trained under supervised learning, and includes the following steps: Step S4-1, Training Phase: Guided by the constructed lightweight time-series prediction model, a 60-second window is extracted from historical data of the AGV. Based on whether a navigation failure event will occur within the next Δt seconds, the navigation failure time is labeled as y=1 or y=0. Weighted binary cross-entropy loss is used for training. Where: α = 3-5 is the weight of the positive sample; Step S4-2, Inference Phase: A new feature window generated every 1 second is input into the model, and forward propagation yields three Sigmoid outputs: The learned temporal patterns are: a continuous decline in health, a sharp drop in location confidence, a continuously high dynamic interference, and sudden environmental changes all lead to an increase in P_fail; Step S4-3, Probability Calibration: After training, apply temperature scaling to make the output probability consistent with the actual frequency. When P_fail(60s)≥0.7, the actual failure probability is about 85%-90%, and when P_fail(60s)≤0.4, the actual failure probability is less than 10%.

[0094] The lightweight time-series prediction model can effectively output the navigation failure probability P_fail(Δt) of AGV at multiple future time points Δt. The difference from the traditional threshold judgment is that it comprehensively considers the trend of the past 60 seconds, and only increases P_fail when the trend points to deterioration and the current state is poor, thus avoiding false alarms caused by instantaneous jitter.

[0095] The P_fail(Δt) is compared with a preset risk threshold to implement tiered early warning and preventative actions. Specifically, the preset thresholds are: low risk <40%, medium risk 40%-70%, and high risk ≥70%. Risk Levels and Action Overview

[0096] Independent startup rule: Only one or two actions most relevant to the fault type are triggered at any given time to avoid overreaction.

[0097] Meanwhile, during the joint startup (full execution) of high-risk levels, when P_fail ≥ 70%, all preventative actions will be attempted regardless of the fault type. Among these, forced speed reduction, local alarm, and emergency notification are executed immediately in parallel; relocation, mode switching, and route replanning are executed sequentially, and the decision to continue with subsequent actions is based on the result of the previous action.

[0098] (1) Actions that are executed immediately in parallel (without waiting for results)

[0099] (3) Complete execution flow (logical description) Forced speed reduction (0.2 m / s) / / Effective immediately Red light flashing + buzzer sound to activate / / Effective immediately Send urgent notification (MQTT) / / Asynchronous, no waiting result_reloc = Request relocation (timeout 500ms) if result_reloc == success: Skip the mode switching and wait for health to recover or continue monitoring. else: Perform step 6 result_switch = Switch navigation mode (timeout 300ms) if result_switch == success: Skip to replan route else: Perform step 7 result_replan = Request to replan the route (timeout 1000ms) if result_replan == failure: Park the car and send a "human intervention required" alert. else: Continue driving along the new route and monitor health status. (4) The principle and advantages of sequential execution Relocation priority: Relocation has the lowest cost, does not require changes to the navigation basis, and can recover from most positioning failures caused by transient interference.

[0100] Switching to a backup navigation mode (QR code / magnetic strip) is the second-best option if relocation fails, indicating that the current SLAM environment is unreliable.

[0101] Finally, if all of the above fail, it indicates that the current route itself has systemic risks and a complete change of the driving route is required.

[0102] Forced deceleration and alarms are always executed: ensuring basic physical safety, unaffected by the success or failure of subsequent actions.

[0103] This embodiment involves two speed settings: a recommended speed reduction of 50% (medium risk) and a forced speed reduction to 0.2 m / s (high risk). The choice is based on the following engineering and physical considerations: The recommended speed reduction of 50% is based on the fact that positioning error is positively correlated with speed: the AGV's positioning drift rate (m / s) is approximately linearly related to speed. Reducing the speed can decrease the odometry integration error per unit time and the inter-frame matching error of laser SLAM.

[0104] Preserving Task Efficiency: A 50% speed reduction strikes a balance between safety and operational efficiency, preventing complete shutdowns that could lead to production line congestion. Experimental Data: In actual testing, a 50% speed reduction slows the descent rate of the positional confidence level C1 by approximately 40%, and 90% of medium-risk scenarios can automatically recover to a low-risk state after speed reduction. Industry Experience: Most AGV manufacturers default to limiting the speed to 50%~60% of the rated speed in "cautious mode." This invention uses 50% to provide sufficient safety margin.

[0105] The basis for forcibly reducing the speed to 0.2 m / s is: safe stopping distance: AGV emergency braking distance s = v² / (2a). Assuming the maximum deceleration a = 1.0 m / s² (typical AGV value), the braking distance is only 0.02 m when v = 0.2 m / s, which can avoid collision; if v = 0.5 m / s, the braking distance is 0.125 m, and a slight collision may still occur in narrow alleys.

[0106] Navigation robustness at low speeds: At 0.2 m / s, laser SLAM exhibits inter-frame displacement ≤ 0.2 m (scanning frequency 10 Hz), and matching residuals can be controlled within 0.03 m. Even with a health score H < 40, basic positioning can still be maintained. Facilitates relocation: Relocation algorithms (such as particle filtering) converge faster at low speeds; 0.2 m / s is the tested "maximum speed for reliable relocation".

[0107] Industry Reference: ISO 3691-4, "Safety Requirements for Industrial Vehicles," typically limits the recovery speed of AGVs after a "protective stop" to below 0.3 m / s. This invention uses 0.2 m / s to increase safety redundancy. Experimental Verification: In tests at the Haogongpin warehouse, after a forced deceleration to 0.2 m / s, the navigation failure rate decreased from 72% to 9%, with no collisions occurring.

[0108] The synergistic logic of the two speeds: when the risk is medium, the speed is reduced by 50% to preserve efficiency and expect self-recovery; if the risk continues to deteriorate to high risk, the speed is forcibly reduced to 0.2 m / s, sacrificing efficiency for absolute safety and triggering other intervention actions.

[0109] Conditions for the continuation and exit of an action

[0110] Meanwhile, anti-shake and false trigger protection are implemented, with delayed confirmation: high-risk joint actions are only triggered if P_fail ≥ 70% continuously for more than 2 seconds. Minimum interval: at least 10 seconds between two high-risk warnings for the same AGV. Priority for manual intervention: the operation and maintenance platform can send a "cancel warning" command to forcibly clear the current warning status.

[0111] This embodiment defines a structured warning code format, which is: WARN-NAV-LLL-TTT-CCCC-SSSS

[0112] The recommended transmission format is JSON, as shown in the example below: { "warn_code":"WARN-NAV-HGH-LCF-3A7B-1F02", "timestamp": 1702044800, "P_fail_30s": 0.72, "P_fail_60s": 0.68, "P_fail_120s": 0.55, "health_score": 45 } Publish to the topic / agv / {id} / nav_warning via MQTT, or write to the holding register via Modbus TCP.

[0113] An AGV navigation quality real-time assessment and predictive maintenance system includes: The data acquisition module is configured to connect to the lidar, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor of the AGV layout to collect multi-source data from the perspective of the AGV, including lidar point cloud data, acceleration data / angular velocity data, wheel mileage data, visual image data and environmental temperature and humidity data. The health assessment module is configured to perform multi-dimensional health assessments on multi-source data. It calculates and outputs the localization reliability C1 based on the AGV's pose status using laser SLAM nodes; it calculates and outputs the scan matching freshness C2 based on the perfect match score s_raw between the AGV's current laser point cloud frame and a pre-stored grid map; it calculates and outputs the dynamic interference C3 based on the AGV's self-movement speed provided by the AGV's laser point cloud frame, IMU inertial measurement unit, and odometer; it calculates and outputs the sensor effectiveness C4 based on the status flags or raw data streams of the sensors (LiDAR, IMU inertial measurement unit, odometer, camera, and temperature / humidity sensor) used in the AGV layout; and it dynamically weights and fuses the localization reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 to obtain the comprehensive health score H. The time series prediction module is configured to run a pre-trained lightweight time series prediction model with feature vectors as input, and output the navigation failure probability P_fail(Δt) for multiple future time points Δt. The early warning decision module is configured to compare P_fail(Δt) with a preset risk threshold, guided by P_fail(Δt). When P_fail(Δt) is low risk, the AGV action is logged; when P_fail(Δt) is medium risk, the AGV speed is reduced and a notification is sent to the upper level for maintenance; when P_fail(Δt) is high risk, the AGV speed is reduced and the upper level is requested to reposition, switch navigation modes, replan the route, and issue an emergency notification in sequence for maintenance. The communication interface is configured to interact with the AGV main controller and scheduling system via Modbus TCP, MQTT or shared memory. The system can be integrated into the AGV's original industrial control computer as a software service, or it can be used as an independent edge computing box to communicate with the AGV main controller via Ethernet.

[0114] This embodiment sets up two application scenarios. The first scenario is a software implementation embedded in the XOA controller, and the second scenario is a standalone XOR agent hardware implementation. In the first scenario, the XOR agent is implemented as an independent service process on the XOA industrial control computer (computing power ≥ 6 TOPS, running Linux). Data acquisition subscribes to LiDAR, IMU, and other devices via the ROS2 interface; the health assessment module is written in C++ and runs at a frequency of 10Hz; the time series prediction model uses the TensorFlow Lite framework to load a pre-trained LSTM model (12MB after quantization). The early warning decision module sends deceleration commands to the AGV motion control node via ModbusTCP and simultaneously sends JSON format early warning information to the scheduling system.

[0115] On-site testing at the Haogongpin warehouse: A 50-meter-long aisle was selected, and the AGV was operated during a period of strong afternoon sunlight (illuminance > 8000 lux). The XOR agent detected a drop in positioning confidence from 92 points to 65 points, predicting a 72% failure probability after 60 seconds, triggering a high-risk warning. The AGV automatically reduced its speed to 0.2 m / s and requested to switch to the QR code navigation alternative path. The AGV successfully traversed the area without navigation failure. After 20 consecutive tests, the high-risk warning accuracy rate was 90%, and the failure avoidance rate after intervention was 85%.

[0116] In the second scenario, for older AGVs (with insufficient main control computing power), a dedicated edge computing box (NVIDIA Jetson Orin Nano, 20 TOPS computing power) is used to communicate with the AGV main controller via Ethernet. The XOR agent obtains sensor data from the AGV main controller (converted into UDP packets), runs the same algorithm as in the first scenario, and sends the warning results and deceleration commands back to the AGV main controller via Modbus TCP. Tests show that the network latency is <10ms and the overall response time is <80ms, meeting the real-time requirements.

[0117] Incremental learning implementation method: Collect FP / FN samples offline every week, retrain with the original training set, and after verifying that the accuracy has improved by ≥1%, distribute and hot-reboot the data via OTA.

[0118] The above are merely embodiments of this solution. Commonly known technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solution of this solution. These modifications and improvements should also be considered within the scope of protection of this solution, and will not affect the effectiveness of the implementation of this solution or the practicality of the patent. The scope of protection claimed in this application shall be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for real-time assessment and predictive maintenance of AGV navigation quality, characterized in that, Includes the following steps: Step 1: The LiDAR, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor in the AGV layout collect multi-source data from the perspective of the AGV, including laser point cloud data, acceleration data / angular velocity data, wheel mileage data, visual image data and environmental temperature and humidity data. Step 2: Perform multi-dimensional health assessment on the multi-source data from Step 1. Calculate and output the positioning confidence C1 based on the AGV's pose status using laser SLAM nodes; calculate and output the scan matching freshness C2 based on the perfect match score s_raw between the AGV's current laser point cloud frame and the pre-stored grid map. The dynamic interference degree C3 is calculated and output based on the AGV self-movement speed provided by the AGV laser point cloud frame, IMU inertial measurement unit, and odometer; the sensor effectiveness C4 is calculated and output based on the status flags or raw data streams of each sensor in the AGV layout, including the lidar, IMU inertial measurement unit, odometer, camera, and temperature and humidity sensor; the positioning reliability C1, scan matching freshness C2, dynamic interference degree C3, and sensor effectiveness C4 are dynamically weighted and fused to obtain the comprehensive health degree H. Step 3: Guided by the historical feature window, calculate the humidity change rate dHr / dt and temperature change rate dT / dt from the environmental temperature and humidity data collected by AGV in Step 1 using the first-order difference mean; and obtain the health score sequence H(t) at this time point using the comprehensive health score H in Step 2 guided by the time axis. The task complexity C_task is obtained by quantifying the visual image data collected by AGV. The feature vector is formed by concatenating the humidity change rate dHr / dt, the temperature change rate dT / dt, the health score sequence H(t), and the task complexity C_task. Step 4: Input the feature vector from Step 3 into the pre-trained lightweight time series prediction model. The lightweight time series prediction model outputs the navigation failure probability P_fail(Δt) for multiple future time points Δt. Guided by P_fail(Δt), compare P_fail(Δt) with a preset risk threshold. When P_fail(Δt) is low risk, log the AGV actions. When P_fail(Δt) is medium risk, reduce the AGV speed and send a notification to the host for maintenance. When P_fail(Δt) is high risk, reduce the AGV speed and sequentially request relocation, switch navigation modes, replan the route, and send an emergency notification to the host for maintenance.

2. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The positioning confidence C1, calculated and output from the AGV's pose state via laser SLAM nodes, includes the following steps: Step S11: The point cloud data of the lidar with AGV layout, the acceleration / angular velocity data of the IMU inertial measurement unit, and the wheel odometer data are fused by extended Kalman filtering or graph optimization and input into the lidar SLAM model. Step S12: Extract the covariance trace by using the pose estimate with covariance output from the SLAM algorithm within the laser SLAM model: trace = Σ xx + Σ yy + Σ θθ , where: Σ xx Σ represents the location variance, in m². xx Σ represents the positional uncertainty of the AGV along the x-axis in its vehicle coordinate system; yy (where Σ is the location variance in m²) yy Σ represents the positional uncertainty of the AGV in its vehicle coordinate system along the y-axis; θθ The variance of the yaw angle is expressed in rad², Σ. θθ This indicates the uncertainty of the AGV's orientation angle; Step S13: Normalize the mapping by extracting the covariance trace. The normalization mapping uses logarithmic compression + linear mapping. Preset parameters: When the ideal positioning score is ≈100, trace_min = 0.001; when the positioning is about to fail and the score is ≈0, trace_max = 0.

5. When the laser SLAM model does not output with covariance, the matching residual is used instead. The normalization formula is similar, with residual min = 0.02m and max = 0.3m. Output: C1 ∈ [0, 100]. The higher the positioning confidence C1, the more reliable the positioning.

3. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The scan matching freshness C2 is calculated by using the perfect match score s_raw between the current laser point cloud frame of the AGV and the pre-stored raster map, and includes the following steps: Step S21: Extract obstacle grids from the original grid map based on the occupancy probability > 0.5 among the occupied grids; calculate the Euclidean distance d(x,y) from each grid to the nearest obstacle using the distance transformation algorithm, and construct the Gaussian likelihood field score function: Where: σ=0.15m is the smoothing parameter; the Gaussian likelihood field score function has a value of 1 at the obstacle and decays exponentially with increasing distance; the score is pre-stored in a floating-point two-dimensional array with a grid resolution of 0.05m, and this pre-calculation only needs to be performed once when the map remains unchanged; Then, the matching score of the current scan is calculated online. Assume that the current laser scan frame, after motion distortion correction, yields the point cloud {(x... i ,y i )} i =1 M The AGV's current estimated pose is (x, y, θ). Transform each scan point to the map coordinate system: Then, the score is obtained by looking up the table using the raster index. i Points outside the map boundaries are not counted as valid points. The matching score s_raw is obtained by aggregating the scores. Step S22: In the pre-stored grid map coordinate system, sample and generate N poses at certain intervals to cover the AGV working area, and generate a perfectly matching simulation scan for each sampled pose. For each frame of simulation scan, directly calculate the projection score s_raw by using the same correlation matching method as the real scan with the pose known, and take the average score of all frames as s_max. Step S23: Calculate using the normalized formula: C2=min(100,max(0,(s_raw / s_max)×100)), output: C2 ∈ [0,100], the higher the scan matching freshness C2, the fresher the scan and map match.

4. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The dynamic disturbance degree C3 is calculated and output using the AGV self-motion speed provided by the AGV laser point cloud frame, IMU inertial measurement unit, and odometer, including the following steps: Step S31: Obtain the total number of valid point clouds N_total in the current frame of the original cloud point frame map: The original point clouds in the original cloud point frame map are cropped sequentially with the ground point removal as the height threshold of 0.1m and the distance point retention ≤15m. At the same time, invalid points with an intensity of 0 or coordinates containing NaN / Inf are removed to obtain the total number of valid point clouds N_total in the current frame. Step S32: Obtain the number of dynamic points N_dynamic using two-frame difference dynamic point detection: Transform both the current frame point cloud P_t and the previous frame point cloud P_{t-1} to the world coordinate system, and perform relative pose motion compensation on P_{t-1} using IMU / odometry or SLAM. For each point in the current frame, use nearest neighbor matching to find the point with the closest Euclidean distance in the compensated point cloud of the previous frame, calculate the distance d_i. If d_i > 0.2m, mark it as a dynamic point; otherwise, the nearest neighbor is missing, and it is not counted as a dynamic point until the next frame confirms. The number of dynamic points N_dynamic is obtained when all the points marked as dynamic points are counted. Step S33: Input the total number of valid point clouds N_total in the current frame from step S31 and the number of dynamic points N_dynamic from step S32 into the formula for calculating the proportion of dynamic points: r_dyn=N_dynamic / N_total Calculate C3: C3 = 100 × (1 - r_dyn). Saturation treatment: if r_dyn > 0.8, then C3 = 0; if r_dyn < 0.05, then C3 = 100. Output: C3 ∈ [0, 100]. The higher the dynamic interference C3, the smaller the dynamic interference.

5. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The calculation and output of sensor validity C4, which involves collecting status flags or raw data streams from the LiDAR, IMU, odometer, camera, and temperature and humidity sensors in the AGV layout, includes the following steps: Step S41: Set the weight w_i of the sensor status identifier or raw data stream of the LiDAR to 0.35; the weight w_i of the sensor status identifier or raw data stream of the IMU inertial measurement unit to 0.20; the weight w_i of the sensor status identifier or raw data stream of the odometer to 0.15; the weight w_i of the sensor status identifier or raw data stream of the camera to 0.15; and the weight w_i of the sensor status identifier or raw data stream of the temperature and humidity sensor to 0.

15. Step S42: Assume the individual validity v_i∈{0,1} of each sensor (LiDAR, IMU, odometer, camera, and temperature / humidity sensor) is 1 if the condition is met, and 0 otherwise. Then, apply the weighted average formula: C4=Σw_i·v_i×100 Output: C4∈[0,100], the higher the sensor effectiveness C4, the more robust the sensor system.

6. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The overall health score H is obtained by dynamically weighting and fusing the location reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4, including the following steps: Step S51: Obtain σ²_Ci by collecting the variance of the location reliability C1 over a set M seconds, obtain σ²_Ci by collecting the variance of the scan matching freshness C2 over a set M seconds, obtain σ²_Ci by collecting the variance of the dynamic interference C3 over a set M seconds, obtain σ²_Ci by collecting the variance of the sensor effectiveness C4 over a set M seconds, and obtain σ²_Ci by collecting the variance of the sensor effectiveness C4 over a set M seconds. The initial variance is 1. Step S52: Input the variance of the location reliability C1, σ²_Ci, into the basic weight formula w_i'=1 / (σ²_Ci+ε), and calculate the basic weight w_i' of the location reliability C1 within ε=0.01; The variance of the scan matching freshness C2 is obtained as σ²_Ci, which is then input into the basic weight formula w_i'=1 / (σ²_Ci+ε). The basic weight w_i' of the scan matching freshness C2 is calculated within ε=0.

01. The variance of the dynamic disturbance degree C3 is obtained as σ²_Ci, which is then input into the basic weight formula w_i'=1 / (σ²_Ci+ε), and the basic weight w_i' of the dynamic disturbance degree C3 is calculated within ε=0.

01. The variance of sensor effectiveness C4, σ²_Ci, is input into the basic weight formula w_i'=1 / (σ²_Ci+ε), and the basic weight w_i' of sensor effectiveness C4 is calculated within ε=0.

01. Step S53: Add the base weights w_i' of location confidence C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4 from step S52 to obtain Σw_j'. Then calculate w_i of location confidence C1, w_i of scan matching freshness C2, w_i of dynamic interference C3, and w_i of sensor effectiveness C4 using normalized weights w_i=w_i' / Σw_j'. Step S54: Add the w_i of the location confidence C1, the w_i of the scan matching freshness C2, the w_i of the dynamic interference C3, and the w_i of the sensor effectiveness C4 obtained by normalizing the weights in step S53 to obtain Σw_i, and output the fusion score formula H=Σw_i·C_i corresponding to the location confidence C1, the scan matching freshness C2, the dynamic interference C3, and the sensor effectiveness C4 respectively: H∈[0,100].

7. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The feature vector is formed by concatenating the humidity change rate dHr / dt, the temperature change rate dT / dt, the health score sequence H(t), and the task complexity C_task, and includes the following steps: Aligning with the time window within the historical feature window, and using the current time as the right endpoint, feature values ​​from the past 60 time points are taken, with zero padding for values ​​less than 60 seconds. Sliding window normalization is applied to the humidity change rate dHr / dt, temperature change rate dT / dt, health score sequence H(t), and task complexity C_task features. Statistics are pre-calculated from historical data: health score sequence H(t) μ=75, σ=15; temperature change rate dT / dt: μ=0, σ=0.5; humidity change rate dHr / dt: μ=0, σ=0.1; task complexity C_task is in the range [0,1] after normalization.

8. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The pre-trained lightweight time series prediction model includes the following steps: Step S1: Construct the input layer of the lightweight time series prediction model. The input layer includes a first LSTM layer for extracting local time series and a second LSTM layer for capturing long-term dependencies at a higher level of abstraction. The first LSTM layer is constructed using 32 independent LSTM memory blocks, each responsible for extracting different dimensions of time series patterns from the input sequence, resulting in an output dimension of 32. Simultaneously, each LSTM memory block is passed through a forgetting gate f. t Input gate i t Output gate o t and cell state C t Formula output is stored in long-term memory; The second LSTM layer uses 32 independent LSTM memory blocks, each responsible for extracting different dimensions of temporal patterns from the input sequence, constructing an output dimension of 32. Simultaneously, each LSTM memory block is passed through a forget gate f. t Input gate i t Output gate o t and cell state C t The formula outputs long-term memory and is used to process higher-level temporal features of the first LSTM layer; Long-term memory formulas in the first and second LSTM layers: Step S2: Construct a fully connected layer in the lightweight time series prediction model. The fully connected layer has 16 neurons, and the number of neurons is used to fit the nonlinear features in the health sequence using the ReLU activation function. Step S3: Construct an output layer in the lightweight time series prediction model. The output layer has 3 neurons, corresponding to P30s, P60s, and P120s respectively. The number of neurons is used to learn the reasonable relationship of P30≤P60≤P120 using the Sigmoid activation function.

9. The method for real-time assessment and predictive maintenance of AGV navigation quality according to claim 1, characterized in that: The lightweight time-series prediction model outputs the navigation failure probability P_fail(Δt) for multiple future time points Δt, including the following steps: Step S4-1, Training Phase: Guided by the constructed lightweight time-series prediction model, a 60-second window is extracted from historical data of the AGV. Based on whether a navigation failure event will occur within the next Δt seconds, the navigation failure time is labeled as y=1 or y=0. Weighted binary cross-entropy loss is used for training. The learned temporal patterns are: a continuous decline in health, a sharp drop in location reliability, a continuously high dynamic interference, and sudden environmental changes all lead to an increase in P_fail; Step S4-3, Probability Calibration: After training, apply temperature scaling to make the output probability consistent with the actual frequency. When P_fail(60s)≥0.7, the actual failure probability is about 85%-90%, and when P_fail(60s)≤0.4, the actual failure probability is less than 10%.

10. A real-time assessment and predictive maintenance system for AGV navigation quality, and a method for real-time assessment and predictive maintenance of AGV navigation quality according to any one of claims 1-9, characterized in that: include: The data acquisition module is configured to connect to the lidar, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor of the AGV layout to collect multi-source data from the perspective of the AGV, including lidar point cloud data, acceleration data / angular velocity data, wheel mileage data, visual image data and environmental temperature and humidity data. The health assessment module is configured to perform multi-dimensional health assessments on multi-source data, outputting the positioning reliability C1 calculated by the laser SLAM node based on the AGV's pose status; and outputting the scan matching freshness C2 calculated by the perfect matching score s_raw between the AGV's current laser point cloud frame and the pre-stored grid map. The dynamic interference degree C3 is calculated and output based on the AGV self-movement speed provided by the AGV laser point cloud frame, IMU inertial measurement unit and odometer; the sensor validity C4 is calculated and output based on the status flags or raw data streams of each sensor in the AGV layout, including lidar, IMU inertial measurement unit, odometer, camera and temperature and humidity sensor. The overall health score H is obtained by dynamically weighting and fusing the location reliability C1, scan matching freshness C2, dynamic interference C3, and sensor effectiveness C4. The time series prediction module is configured to run a pre-trained lightweight time series prediction model with feature vectors as input, and output the navigation failure probability P_fail(Δt) for multiple future time points Δt. The early warning decision module is configured to compare P_fail(Δt) with a preset risk threshold, guided by P_fail(Δt). When P_fail(Δt) is low risk, the AGV action is logged; when P_fail(Δt) is medium risk, the AGV speed is reduced and a notification is sent to the upper level for maintenance; when P_fail(Δt) is high risk, the AGV speed is reduced and the upper level is requested to reposition, switch navigation modes, replan the route, and issue an emergency notification in sequence for maintenance. The communication interface is configured to interact with the AGV master controller and scheduling system via Modbus TCP, MQTT, or shared memory.

Images