Unmanned truck travel time adaptive prediction system and method with road condition drift perception

Abstract

The application discloses an unmanned truck passing time adaptive prediction system and method for road condition drift perception, physical derived features are extracted to decouple absolute space coordinates; potential road features are extracted by using feature abstraction and reconstruction network, and abnormal road conditions are triggered for real-time fuse prewarning based on the mapping relationship between instantaneous reconstruction error and adaptive dynamic threshold; road conditions are stratified by using density clustering, soft membership weight is calculated by combining dynamic temperature coefficients derived from vehicle acceleration, and an asymmetric loss function is introduced to apply congestion prevention safety margin to complete global weighted fusion prediction; road condition concept drift is perceived in real time by calculating feature distribution divergence, and adaptive incremental updating is performed on the affected model by combining information entropy access mechanism to extract effective samples. The application effectively overcomes model aging caused by physical environment evolution, reduces global scheduling imbalance and local resource queuing risk, and provides a long-term and high-precision time prediction benchmark for complex dynamic mine environments.

Description

Technical Field

[0001] This invention relates to a method for predicting the travel time of unmanned trucks, specifically to an adaptive prediction system and method for unmanned truck travel time based on road condition drift perception. Background Technology

[0002] In open-pit mine unmanned driving and intelligent scheduling systems, accurate prediction of truck travel time is the cornerstone for achieving efficient collaborative scheduling across the entire "mining-loading-transportation-discharge" process. Currently, most traditional travel time prediction models rely on offline training using static road network topology and absolute spatial coordinates. However, the operating environment in open-pit mines is extremely harsh and undergoes frequent dynamic evolution. With the continuous progress of drilling and blasting, face advancement, and spoil disposal, the physical geometry and road surface characteristics of mine roads undergo substantial changes frequently. This precipitous evolution of the physical environment leads to a severe "road condition concept drift" between historical distribution and real-time status. Consequently, pre-trained models built on fixed coordinate systems experience a significant decline in prediction accuracy after short-term deployment. The continuous decline in travel time prediction accuracy directly disrupts the stability of the overall scheduling system, resulting in imbalances in truck-shovel matching and reduced overall transportation efficiency. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides an adaptive prediction system and method for unmanned truck travel time based on road condition drift perception. This system can decouple absolute coordinates, perceive the intensity of road condition evolution in real time, and has the ability to update the model autonomously and incrementally, thereby ensuring the prediction robustness and long-term viability in complex and dynamic mining environments.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an adaptive prediction system for unmanned truck travel time based on road drift perception, comprising an on-board edge computing device deployed on the unmanned truck for performing trajectory cleaning and completion, physical derived feature extraction, dimensionality reduction of potential road features, and real-time anomaly circuit breaking based on reconstruction error; and a cloud scheduling server, which is communicatively connected to the on-board edge computing device for receiving the dimensionality-reduced potential road features and performing hierarchical soft clustering based on dynamic temperature adjustment, asymmetric global fusion prediction, and background silent incremental updates based on Wasserstein distance and information entropy admission criteria.

[0005] An adaptive prediction method for the travel time of unmanned trucks based on road condition drift perception includes the following steps: S100: Obtain the historical operating trajectory of the unmanned truck in the open mine, including multi-source heterogeneous sensor data and vehicle load status. Use a filtering algorithm to complete the blind zone state space to obtain the filtered trajectory sequence. Based on the vehicle load status, split the filtered trajectory sequence into independent prediction streams to extract the physical derived feature sequence. S200: In the vehicle-mounted edge computing device, the physical derived feature sequence is input into the pre-trained feature abstraction and reconstruction network for dimensionality reduction. The sparse constraints during the network training stage are used to filter out mechanical redundant noise and output potential road features. The mapping relationship between the instantaneous reconstruction error and the adaptive dynamic threshold of the feature abstraction and reconstruction network is used to perform real-time circuit breaking and early warning of unstructured abnormal road conditions to prevent severe sudden road conditions from polluting the subsequent prediction stream. S300 uses a cloud-based scheduling server to perform road condition stratification based on the potential features of the road using a density clustering algorithm. It also generates a dynamic temperature adjustment coefficient based on the real-time acceleration of the vehicle to smooth the decision boundary when switching road conditions. In addition, it combines the dynamic temperature adjustment coefficient to calculate the soft membership weight of the current sample to each road condition cluster. S400. For each road condition cluster, train an intra-cluster prediction model based on an asymmetric loss function, and use soft membership weights to globally weight and fuse the outputs of each intra-cluster prediction model. This eliminates cross-road condition prediction jumps and applies a congestion prevention scheduling safety margin through an asymmetric negative gradient amplification mechanism, outputting a basic travel time prediction value. S500 extracts the real-time probability density distribution of potential road features and calculates its distribution divergence from the baseline feature distribution to overcome the failure of conventional spatial distance metrics on high-dimensional feature sets, thereby enabling real-time perception of road condition concept drift. When road condition concept drift is detected, invalid physical noise is filtered out based on the information entropy admission judgment mechanism, and valid drift samples are extracted to adaptively and incrementally update the affected intra-cluster prediction model.

[0006] Furthermore, the method for splitting the filtered trajectory sequence into independent prediction streams based on the vehicle load status to extract the physical derived feature sequence is as follows: Extracting Physically Derived Feature Sequences Using a Sliding Window Mechanism Discretization calculations include the following dimensions: Rate of change of heading angle at time t (HR) t : ; In the formula: L is the sliding window length, k is the window radius, wrap is the angle normalization function, and θ is the heading angle. The sampling time interval; The road surface slope is calculated using the ratio of the elevation difference between adjacent sampling points to the horizontal displacement. ; Calculate road curvature using the ratio of angular displacement to distance traveled The dynamic attitude disturbance characteristic α is obtained by calculating the variance of the attitude angle change output by the inertial measurement unit within the sliding window. t ; Meanwhile, the three-axis composite acceleration a and real-time vehicle speed v of the unmanned truck are directly extracted from the filtered trajectory sequence and output as bypasses to the downstream S300 and S400.

[0007] Furthermore, the joint optimization objective function J(W,b) during the network training phase is: ; In the formula: b is the bias vector of the hidden layer neurons; the first term is the mean square reconstruction error term, where X i Let i be the i-th physical derived feature of the input. The first term is the i-th physical reconstruction feature, and N is the number of samples; the second term is the regularization term, used to constrain the norm of the weight matrix W through the weight coefficient λ1 to improve the model's robustness to noise; the third term is the sparsity penalty term, which guides the average activation of the hidden layer neurons through the coefficient λ2. Approaching the desired sparsity Among them, average activation Let be the mean activation output of the j-th hidden layer neuron under N input samples; thus, KL divergence is used to achieve automatic dimensionality reduction of redundant features in the mine.

[0008] Furthermore, the specific method for real-time circuit breaking and early warning of unstructured abnormal road conditions by utilizing the mapping relationship between the instantaneous reconstruction error of the feature abstraction and reconstruction network and the adaptive dynamic threshold is as follows: Define the instantaneous reconstruction error as Where X is the input feature, For reconstructing features; The historical error mean is updated in real time using an exponentially weighted moving average algorithm. with the mean square of the error : ; ; Calculate the dynamic real-time threshold based on the historical mean error and the mean squared error. : ; Where: γ is the preset forgetting factor, and k is the sensitivity coefficient; when determining the instantaneous reconstruction error Greater than the dynamic real-time threshold When this occurs, a real-time circuit breaker warning is triggered.

[0009] Furthermore, the specific method for calculating the soft membership weights of the current sample belonging to each road condition cluster is as follows: Obtain the vehicle's current three-axis composite acceleration 'a'; Calculate the dynamic temperature regulation coefficient: In the formula, As the reference temperature coefficient, Sensitivity factor; Calculate soft membership weights : ; In the formula: c is the membership weight of the current road potential feature h to the k-th road condition cluster; h is the current road potential feature vector extracted by the denoising autoencoder; k is the center vector of the k-th road condition cluster; dist is the preset distance metric function; T is the dynamic temperature adjustment coefficient calculated by the above steps; K is the total number of road condition clusters; p is the index variable of the cluster.

[0010] Furthermore, the loss function of the intra-cluster prediction model based on the asymmetric loss function... for: ; In the formula: y is the actual value, For predicted values, The penalty bias coefficient is set to address the risk of underprediction, which exponentially amplifies the error penalty when underprediction occurs.

[0011] Furthermore, the method for calculating the distribution divergence between the real-time probability density distribution of the potential road features and the baseline feature distribution is as follows: The potential feature sequences of roads within a short sliding window are collected, and the real-time probability density function is fitted using kernel density estimation. The real-time probability density function is then integrated to obtain the corresponding real-time cumulative distribution function. ; Calculate the benchmark cumulative distribution function corresponding to the real-time cumulative distribution function and the benchmark characteristic distribution. The first-order Wasserstein distance between them: ; In the formula: P ref and P det These represent the baseline probability distribution and the real-time observation probability distribution, respectively. and These are the cumulative distribution functions corresponding to the baseline characteristic distribution and the real-time observation distribution, respectively; The trend smoothing value of the Wasserstein distance is calculated using an exponentially weighted moving average algorithm. When the trend smoothing value is greater than the adaptive threshold, it is determined that road condition concept drift has occurred.

[0012] Furthermore, the method for filtering invalid physical noise and extracting valid drift samples based on the information entropy admission judgment mechanism to adaptively incrementally update the affected intra-cluster prediction model is as follows: Trajectory samples experiencing road condition concept drift are pushed into a short-term memory (STM) with a first-in, first-out (FIFO) mechanism; the feature information entropy H of the sample set in the STM is calculated; and the method updates the STM only if H is greater than a preset richness threshold H. threshold At that time, the approved sample set is entered into the long-term memory, and the incremental learning interface is called to update the intra-cluster prediction model parameters affected by road condition drift.

[0013] Compared with existing technologies, this invention fundamentally overcomes the problem of road network spatial topology failure caused by high-frequency stripping and blasting operations in open-pit mines by extracting physically derived features independent of absolute geographic coordinates, thus endowing the prediction model with extremely strong cross-regional generalization ability. Simultaneously, the dynamic temperature adjustment and asymmetric loss mechanism introduced at the prediction execution layer not only smoothly eliminates the abrupt jumps in estimated arrival times when trucks travel across road conditions, but also applies scheduling safety margins using the anti-congestion penalty weights at the algorithm's underlying layer, effectively reducing local queuing congestion rates in the mining area. More importantly, this invention constructs a concept drift perception and adaptive evolution architecture based on dual verification of distribution divergence and information entropy, enabling the system to accurately identify short-term physical noise and substantial road network rerouting, achieving silent incremental updates of affected model branches without global retraining. Combining the above algorithmic mechanisms, this invention greatly extends the full lifecycle accuracy of the prediction model, providing long-term and stable underlying time benchmark support for truck-shovel collaborative scheduling in complex and dynamic mining environments. Attached Figure Description

[0014] Figure 1 This is a flowchart of the control method of the present invention; Figure 2 This is a schematic diagram illustrating the multi-model weighted fusion prediction principle and dynamic temperature regulation mechanism based on soft membership degree of the present invention. Figure 3 This is a flowchart of the road condition drift perception and incremental adaptive update mechanism of the present invention; Figure 4 This is a schematic diagram illustrating the physical derived feature space reconstruction and calculation principle of the present invention; Figure 5 This is a diagram illustrating the cloud-edge-device collaborative computing deployment architecture and data flow topology of this invention. Figure 6 This is a comparison chart of the prediction errors of the adaptive prediction model of this invention and the traditional static model under the road condition concept drift condition. Detailed Implementation

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention provides an adaptive prediction system for the travel time of unmanned trucks based on road condition drift perception, such as... Figure 5 The system described adopts a "cloud-edge-device" collaborative architecture, including an onboard edge computing device and a cloud scheduling server. The unmanned truck is equipped with onboard edge computing devices and onboard sensors such as RTK-GPS, IMU, and load sensors. Taking a 220T unmanned dump truck in an open-pit mine as an example, the preferred sampling frequency for the onboard RTK-GPS terminal is 10Hz, and the preferred sampling frequency for the attitude and three-axis composite acceleration of the inertial measurement unit (IMU) is 100Hz to ensure high-frequency dynamic capture of road bumps. The onboard sensors collect multi-source heterogeneous raw trajectory stream P in real time. t and vehicle load status m t The data is then input into the vehicle-mounted edge computing device in real time; the vehicle-mounted edge computing device communicates with the cloud scheduling server. The vehicle-mounted edge computing device is used to perform trajectory cleaning and completion, physical derived feature extraction, dimensionality reduction of potential road features, and real-time anomaly circuit breaking based on reconstruction errors. The cloud scheduling server is used to receive the dimensionality-reduced potential road features and perform hierarchical soft clustering based on dynamic temperature adjustment, asymmetric global fusion prediction, and background silent incremental updates based on Wasserstein distance and information entropy admission criteria.

[0018] This invention also provides an adaptive prediction method for the travel time of unmanned trucks based on road condition drift perception, such as... Figure 1 The steps are as follows: S100. Multi-source heterogeneous data cleaning and physical feature reconstruction: This step aims to physically and semantically decompose the original sensor sequences, decoupling them into a "topological feature flow" for characterizing the spatial road network structure and a "kinematic feature flow" for characterizing the transient dynamics of vehicles. This is specifically divided into the following sub-steps.

[0019] S101. Trajectory Washout and Blind Spot Compensation: The system acquires the raw vehicle sensor sequences. , of which lat t lon t alt t θ t With v tThese include the vehicle's real-time latitude, longitude, elevation, heading angle, and real-time speed. For GPS blind spots in deep mining areas, the system monitors positioning variance in real time. When a decrease in positioning accuracy is detected, the system activates blind spot compensation logic, calls the inertial measurement unit (IMU) to perform dead reckoning, and combines this with the state prediction equation using Kalman filtering. Where A is the state transition matrix, B is the control input matrix, and U... t To control input variables, W t To address system process noise, millisecond-level state completion is performed on the latitude, longitude, and elevation at time t to ensure the physical continuity of the input sequence on the time axis, thereby outputting the filtered trajectory sequence.

[0020] S102. Load-based multi-stream data distribution: The system synchronously retrieves truck load sensor data m t To eliminate the significant differences in braking distance and climbing dynamics between heavily loaded and unloaded trucks, this embodiment sets a load determination threshold M. th Considering the truck's rated load capacity, this threshold is preferably 150 tons. When m t >M th When the data stream is loaded, it enters the "heavy load prediction stream model"; otherwise, it enters the "idle load prediction stream model".

[0021] S103, Physically derived feature sequence X t Discretization reconstruction: like Figure 4 As shown, the system uses a sliding spatiotemporal window of length L and radius k to extract four-dimensional physical feature vectors from the preprocessed sequence. Corresponding to the aforementioned 10Hz positioning sampling frequency, this embodiment preferably sets the sliding window length L=50 (i.e., covering a 5-second time receptive field) and the window radius k=5. The specific calculation operators and variable definitions are as follows: Rate of change of heading angle at time t (HR) t : ; In the formula, L is the sliding window length, k is the window radius, wrap is the angle normalization function, and θ is the heading angle. The sampling time interval; Road slope used to sense the longitudinal resistance characteristics of unmanned trucks on transport routes and mining ramps. : ; In the formula, alt represents the vehicle elevation. The horizontal Euclidean distance between adjacent sampling points; Used to abstractly describe road geometry and topology, achieving road curvature that decouples features from absolute geographic coordinates. : ; By calculating the variance of the roll and pitch angles output by the inertial measurement unit (IMU) within the sliding window, dynamic attitude disturbance characteristics α of unstructured road bumps and mud levels are used to perceive the road surface in real time. t : ; In the formula, For vehicle roll angle, The vehicle's pitch angle, and This is the mean within the window. This feature quantifies the bumpiness and mud resistance of the road surface using the second moment (variance).

[0022] S104, Feature Spatiotemporal Assembly and Multidimensional Sequence Reconstruction: The rate of change of heading angle was calculated separately. Road surface slope Road curvature and dynamic attitude perturbation characteristics Subsequently, the vehicle-mounted edge computing device performs feature space aggregation. Specifically, the four scalar features that are physically and semantically independent are instantaneously concatenated according to a preset topological-kinematic mapping order to reconstruct the physical derived feature vector at time t: ; As the truck continues to travel within the mining area and the sliding spatiotemporal window shifts forward along the time axis, the system reconstructs the feature vectors obtained from each instant. The matrices are concatenated according to their timestamps, and finally aggregated to generate a continuous sequence of physical derived features. This sequence transforms the original fragmented trajectories and sensor signal streams into a unified physical geometry and dynamic response description. While removing absolute latitude, longitude, and elevation coordinates, it achieves complete decoupling of absolute geographic coordinates and serves as the standard input stream to the downstream denoising autoencoder (SDAE) dimensionality reduction module.

[0023] Simultaneously, the system directly extracts the triaxial composite acceleration 'a' and real-time vehicle speed 'v' of the unmanned truck from the filtered trajectory sequence, and outputs them as a "kinematic feature stream" to the downstream steps S300 and S400. This dual-stream decoupling mechanism ensures the purity of the topological features during dimensionality reduction in step S200, while preserving lossless transient dynamic prior information for downstream dynamic temperature regulation and time consumption prediction.

[0024] S200, Feature Abstraction and Adaptive Threshold Circuit Breaker: Deployed on-board edge computing devices in unmanned trucks, this system aims to perform nonlinear dimensionality reduction on high-dimensional, noisy physically derived feature sequences and construct a safety warning envelope that evolves with the conditions of the mine road surface. The specific execution process is as follows.

[0025] S201, Sparse Noise Reduction Learning: On the vehicle side, due to X t Dynamic attitude perturbation α t This includes a significant amount of redundant mechanical noise from the truck's suspension system; without strong suppression, subsequent road condition clustering will become completely ineffective. Therefore, X... t After injecting Gaussian noise, a denoising autoencoder with sparse constraints (SDAE) is input, and the objective function is jointly optimized during the network training phase. The sparse penalty term (KL divergence) filters out mechanically redundant noise, outputting a clean road potential feature h; the objective function is jointly optimized. : ; In the formula: b is the bias vector of the hidden layer neurons; the first term is the mean square reconstruction error term, where Let i be the i-th physical derived feature of the input. The first term is the i-th physical reconstruction feature, where N is the number of samples; the second term is the regularization term, used to apply weight coefficients. (In this embodiment, 1×10 is preferred) -4 The norm of the weight matrix W is constrained to improve the model's noise resistance; preferably, the hidden layer neurons of the denoising autoencoder are set to 32 dimensions to achieve effective compression and abstraction of high-dimensional physically derived feature sequences; the third term is a sparsity penalty term, determined by the coefficient... (In this embodiment, 0.05 is preferred) to guide the average activation level of neurons in the hidden layer. Approaching the desired sparsity (In this embodiment, 0.1 is preferred), wherein the average activation level is... Let be the mean activation output of the j-th hidden layer neuron across N input samples; thus, KL divergence is used to achieve automatic dimensionality reduction of redundant features in the mine. During the model training phase, the Adam optimization algorithm is preferably used to iterate the parameters of the feature abstraction and reconstruction network, with an initial learning rate set to 0.001.

[0026] S202, Dynamic Adaptive Fuse: The system monitors the reconstructed residual at time t in real time. ,in For input features, To reconstruct features; considering that the surface smoothness of open-pit mine roads will slowly change due to weather or regular wear and tear, the system constructs a dynamically evolving real-time threshold for circuit breaking. Instead of using a fixed constant; Statistics are updated dynamically: The historical error mean is updated in real time using an exponentially weighted moving average algorithm. with the mean square of the error : ; ; in, As a forgetting factor, it determines the length of time the threshold is remembered for historical road conditions; Dynamic real-time threshold Calculation: Calculate the dynamic real-time threshold at time t based on second-order statistics: ; Wherein, γ is a preset forgetting factor (which determines the length of the threshold's memory of historical road conditions, preferably 0.95). The sensitivity coefficient (preferably configured as 2.5) is typically taken as a value between 2 and 3; when determining the instantaneous reconstruction error... Greater than the dynamic real-time threshold When this occurs, a real-time circuit breaker warning is triggered; Abnormal circuit breaker triggering: The instantaneous reconstruction error is compared with the dynamic real-time threshold in real time on the edge computing device. If determined... If this occurs, it indicates that the current driving characteristics have undergone an "unstructured mutation" beyond the normal evolution range, such as encountering a local road collapse, a deep pothole, or a major sensor malfunction. The system immediately triggers the circuit breaker protection logic, suspends the current travel time prediction, and forcibly sends a "physical environment anomaly" alarm to the cloud scheduling server, at which point manual driving takes over.

[0027] S300, road condition soft clustering based on dynamic temperature coefficient.

[0028] In the preferred embodiment, the cloud-based scheduling server uses the HDBSCAN algorithm for feature clustering, such as loading platform clusters and mining steep slope clusters. In specific implementations, to effectively filter transient noise points on the road surface and identify physically meaningful continuous road segments, the core parameters of the HDBSCAN algorithm—minimum cluster size and minimum sample size—are preferably set to 120 and 15, respectively. Figure 2 As shown, to prevent abrupt changes in predicted travel time when crossing different road conditions, a dynamic temperature adjustment coefficient T is introduced. This coefficient is used to calculate the weight w of a sample belonging to the k-th road condition cluster, such as an incline or a flat section. k At that time, the system calculates the dynamic temperature regulation coefficient T based on the vehicle's real-time synthetic acceleration a: ; In the formula: T is the dynamic temperature adjustment coefficient; T0 is the preset reference temperature coefficient (preferably 1.0); exp is an exponential function with the natural constant e as the base; A preset sensitivity factor (preferably 0.65) is used to adjust the weight of the contribution of acceleration to temperature; This represents the vehicle's current three-axis composite acceleration. After obtaining the real-time dynamic temperature regulation coefficient T, the system calculates the weights of the current road potential feature h belonging to each road condition cluster. : ; In the formula: c is the membership weight of the current road potential feature h to the k-th road condition cluster; h is the current road potential feature vector extracted by the denoising autoencoder; k is the center vector of the k-th road condition cluster; dist is a preset distance metric function, preferably Euclidean distance in this embodiment; T is the dynamic temperature adjustment coefficient calculated by the above steps; K is the total number of road condition clusters; p is the index variable of the cluster. Based on this coefficient, the soft membership weights of the current sample belonging to each road condition cluster are calculated. If the weight shift is too drastic, it can cause the model to switch abruptly, resulting in a sharp drop in prediction time and potentially leading to an abnormal negative gradient explosion in subsequent asymmetric losses. Therefore, a dynamic temperature adjustment coefficient T is introduced to smooth the decision boundary during road condition changes, balancing smoothness and agile response.

[0029] Step S400: Congestion Prevention Integrated Prediction Based on Gradient Boosting Framework and Asymmetric Loss. For each road condition cluster identified in Step S300, the system constructs an independent intra-cluster travel time prediction model F for each cluster. k Considering the numerous nonlinear characteristics and the ease with which mining area environmental data can be lost, in this embodiment, the prediction model F within each cluster... k Specifically, a heterogeneous ensemble model based on the gradient boosting decision tree (GBDT) architecture is adopted.

[0030] S401, the basic architecture of the intra-cluster prediction model: For the k-th road condition cluster, its travel time prediction model F k It is composed of M regression trees stacked together. In this embodiment, it is preferable to set the number of basic regression trees M=100, and the maximum depth of a single regression tree is limited to 5 to prevent the model from overfitting on sparse road condition features. Let the comprehensive state vector be a combination of vehicle load m, real-time vehicle speed v, and potential road features h. Prediction model Fk The deduction output is: ; In the formula, This represents the space where a single regression tree resides. Each tree splits the input latent topological features, such as implicit elevation differences or curvature, and outputs the predicted travel time increment for that slice of road. The sum of the outputs of all trees is the base predicted travel time for that road condition cluster.

[0031] S402, Gradient optimization of coupled asymmetric loss: In training the prediction model F k At that time, the conventional mean squared error loss was abandoned in favor of an asymmetric loss function. : ; In the formula, y is the actual value. For predicted values, The penalty bias coefficient set for underprediction risk (preferred configuration in this embodiment) =2.0), when underprediction occurs, the error penalty is amplified exponentially (i.e., a 3x error penalty is applied to ensure the safety margin of the scheduling system). During the gradient boosting iteration of the model, the system calculates the negative gradient of the asymmetric loss function with respect to the current predicted value, i.e., the pseudo residual. When "under-prediction" occurs When the bias coefficient ω (ω> 0) is used, the pseudo residual will be amplified: ; In the formula: This refers to the pseudo-residual, or negative gradient, calculated for the i-th training sample during the m-th iteration. The actual vehicle travel time recorded in the historical data for the i-th training sample; The comprehensive state feature vector input to the i-th training sample at time t includes potential road features inherited from step S200, real-time vehicle speed, and load, etc. The predicted travel time for this sample is derived by the current ensemble model consisting of the first m-1 regression trees; Force the next regression tree f m The focus is on fitting this amplified error, thereby creating a "safe prediction margin" at the underlying physical logic of the model, making the final output travel time more conservative.

[0032] This asymmetric loss mechanism, combined with the concept drift perception in step S500, constitutes a deep algorithmic synergy: In the initial stage of substantial deterioration in the open-pit mine road network, such as sudden mudslides or local collapses, before the S500 module collects sufficient samples and triggers incremental updates, the system inevitably experiences a "drift perception blind spot." During this period, traditional models are prone to outputting underestimated prediction times, causing the scheduling system to prematurely dispatch subsequent transportation tasks, thereby leading to resource queuing and scheduling imbalances at loading / unloading points or bottleneck sections. The asymmetric negative gradient amplification mechanism introduced in step S400 provides a strong safety net for this "blind spot," ensuring that before the model completes incremental evolution, the output value can still be protected by an anti-queuing safety margin through the underlying physical logic.

[0033] S403, Global Dynamic Weighted Fusion: After obtaining the independent prediction results for each road condition cluster, the soft membership weights w generated from the dynamic temperature coefficient in step S300 are combined. k Output the global predicted travel time Y for vehicles traversing complex and variable road sections ahead: ; This formula ensures that the estimated travel time is continuous and smooth when vehicles cross different physical road conditions, such as entering a sharp bend from a long, gentle slope.

[0034] S500, adaptive evolution based on Wasserstein distance and information entropy, such as Figure 3 As shown; S501, Drift Sensing and Divergence Measurement: The potential road features h within a short sliding window (preferably with a window size of 1000 sampling sequences in this embodiment) are collected. The h is then estimated using kernel density (preferably a Gaussian kernel function with a bandwidth parameter set to 0.5) and fitted to a continuous real-time probability density function. This function is then integrated to obtain the corresponding real-time cumulative distribution function. ,calculate Compared with the baseline cumulative distribution function The first-order Wasserstein distance between them: ; In the formula, and These are the cumulative distribution functions corresponding to the baseline feature distribution and the real-time observation distribution, respectively. The trend smoothing value of the Wasserstein distance is calculated using the exponentially weighted moving average algorithm. When the trend smoothing value is greater than the adaptive threshold, it is determined that a road condition concept drift has occurred. This distance can overcome the failure problem of conventional spatial distance metrics on high-dimensional sparse feature sets and accurately measure the distribution deviation caused by road precipitous rerouting.

[0035] S502, Access and Incremental Updates: When divergence is abnormal, trajectory samples exhibiting road condition concept drift are pushed into a short-term memory (SSM) with a first-in, first-out (FIFO) mechanism (in this embodiment, the maximum capacity of the SSM is preferably set to 5000 consecutive trajectory samples). The system calculates the feature information entropy H of the SSM sample set, and the entropy is determined if and only if H is greater than a preset richness threshold H0. threshold (In this embodiment, version 1.2 is preferred) When it is determined that the batch contains a sufficient set of novel road condition information samples to enter the long-term memory, the incremental learning interface is called to update the intra-cluster prediction model parameters affected by road condition drift, and for the affected F k Adjust the decision tree weights to complete the adaptive update.

[0036] Taking a real-world scenario as an example, in a mining area where continuous rainfall causes muddy roads and leads to road condition drift, the mean absolute error of traditional prediction models based on static road networks rises sharply. However, with the system configured in this embodiment, relying on distribution divergence measurement and an adaptive incremental update mechanism, the mean absolute error of travel time prediction converges rapidly after accumulating effective samples from approximately 15 transport batches. Compared to traditional models, prediction accuracy is improved by approximately 38%, effectively eliminating the risk of localized vehicle-shovel imbalance and queuing congestion in the cloud-based scheduling system. Figure 6 As shown, it intuitively demonstrates the comparison of prediction errors between the adaptive prediction model and the traditional static model in this embodiment under the condition of sudden road condition concept drift.

[0037] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any minor modifications, equivalent substitutions, and improvements made to the above embodiments based on the technical essence of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive prediction method for the travel time of unmanned trucks based on road condition drift perception, characterized in that, Includes the following steps: S100: Acquire the sensor data and historical operating trajectory of the unmanned truck and its load status. Use a filtering algorithm to complete the blind zone state space to obtain the filtered trajectory sequence. Based on the vehicle load status, split the filtered trajectory sequence into independent prediction streams to extract the physical derived feature sequence. S200: In the vehicle-mounted edge computing device, the physical derived feature sequence is input into the pre-trained feature abstraction and reconstruction network for dimensionality reduction. The sparse constraints of the network training stage are used to filter out mechanical redundant noise and output the potential road features. It also utilizes the mapping relationship between the instantaneous reconstruction error of the feature abstraction and reconstruction network and the adaptive dynamic threshold to perform real-time circuit breaking and early warning for unstructured abnormal road conditions; S300 uses a cloud-based scheduling server to perform road condition stratification based on the potential features of the road using a density clustering algorithm. It also generates a dynamic temperature adjustment coefficient based on the real-time acceleration of the vehicle to smooth the decision boundary when switching road conditions. In addition, it combines the dynamic temperature adjustment coefficient to calculate the soft membership weight of the current sample to each road condition cluster. S400. For each road condition cluster, train an intra-cluster prediction model based on an asymmetric loss function, and use soft membership weights to globally weight and fuse the outputs of each intra-cluster prediction model. Apply a congestion prevention scheduling safety margin through an asymmetric negative gradient amplification mechanism, and output the basic travel time prediction value. S500 extracts the real-time probability density distribution of potential road features, calculates its divergence from the baseline feature distribution, and thus perceives road condition concept drift in real time; when road condition concept drift is determined to occur, it filters invalid physical noise based on the information entropy admission judgment mechanism, extracts valid drift samples, and performs adaptive incremental updates to the affected intra-cluster prediction model. The specific method for calculating the soft membership weights of the current sample belonging to each road condition cluster is as follows: Obtain the vehicle's current three-axis composite acceleration 'a'; Calculate the dynamic temperature regulation coefficient: In the formula, T0 is the reference temperature coefficient. Sensitivity factor; Calculate soft membership weights : ； In the formula: c is the membership weight of the current road potential feature h to the k-th road condition cluster; h is the current road potential feature vector extracted by the denoising autoencoder; k is the center vector of the k-th road condition cluster; dist is the preset distance metric function; T is the dynamic temperature adjustment coefficient calculated by the above steps; K is the total number of road condition clusters; p is the index variable of the cluster; The loss function of the intra-cluster prediction model based on the asymmetric loss function for: ； In the formula: y is the actual value, For predicted values, The penalty bias coefficient is set to address the risk of underprediction, which exponentially amplifies the error penalty when underprediction occurs.

2. The adaptive prediction method for unmanned truck travel time based on road condition drift perception according to claim 1, characterized in that, The method for splitting the filtered trajectory sequence into independent prediction streams based on vehicle load status to extract physically derived feature sequences is as follows: Extracting Physically Derived Feature Sequences Using a Sliding Window Mechanism Discretization calculations include the following dimensions: Rate of change of heading angle at time t (HR) t : ； In the formula: L is the sliding window length, k is the window radius, wrap is the angle normalization function, and θ is the heading angle. The sampling time interval; The road surface slope Grad is calculated using the ratio of the elevation difference between adjacent sampling points to the horizontal displacement. t ; Calculate road curvature using the ratio of angular displacement to distance traveled. t The dynamic attitude disturbance characteristic α is obtained by calculating the variance of the attitude angle change output by the inertial measurement unit within the sliding window. t ; Meanwhile, the three-axis composite acceleration a and real-time vehicle speed v of the unmanned truck are directly extracted from the filtered trajectory sequence and output as bypasses to the downstream S300 and S400.

3. The adaptive prediction method for unmanned truck travel time based on road condition drift perception according to claim 1, characterized in that, The joint optimization objective function J(W,b) during the network training phase is: ； In the formula: b is the bias vector of the hidden layer neurons; the first term is the mean square reconstruction error term, where X i Let i be the i-th physical derived feature of the input. The first term is the i-th physical reconstruction feature, and N is the number of samples; the second term is the regularization term, which is used to constrain the norm of the weight matrix W through the weight coefficient λ1 to improve the model's noise resistance. The third term is the sparsity penalty term, which guides the average activation of neurons in the hidden layer through the coefficient λ2. Approaching the desired sparsity Among them, average activation Let be the mean activation output of the j-th hidden layer neuron under N input samples; thus, KL divergence is used to achieve automatic dimensionality reduction of redundant features in the mine.

4. The adaptive prediction method for unmanned truck travel time based on road condition drift perception according to claim 1, characterized in that, The specific method for real-time circuit breaking and early warning of unstructured abnormal road conditions by utilizing the mapping relationship between the instantaneous reconstruction error of the feature abstraction and reconstruction network and the adaptive dynamic threshold is as follows: Define the instantaneous reconstruction error as Where X is the input feature, For reconstructing features; The historical error mean is updated in real time using an exponentially weighted moving average algorithm. with the mean square of the error : ； ； Calculate the dynamic real-time threshold based on the historical mean error and the mean squared error. : ； Where: γ is the preset forgetting factor, and k is the sensitivity coefficient; when determining the instantaneous reconstruction error Greater than the dynamic real-time threshold When this occurs, a real-time circuit breaker warning is triggered.

5. The adaptive prediction method for unmanned truck travel time based on road condition drift perception according to claim 1, characterized in that, The method for calculating the divergence between the real-time probability density distribution and the baseline feature distribution of the road potential features is as follows: The potential feature sequences of roads within a short sliding window are collected, and the real-time probability density function is fitted using kernel density estimation. The real-time probability density function is then integrated to obtain the corresponding real-time cumulative distribution function. ; Calculate the benchmark cumulative distribution function corresponding to the real-time cumulative distribution function and the benchmark characteristic distribution. The first-order Wasserstein distance between them: ； In the formula: P ref and P det These represent the baseline probability distribution and the real-time observation probability distribution, respectively. and These are the cumulative distribution functions corresponding to the baseline characteristic distribution and the real-time observation distribution, respectively; The trend smoothing value of the Wasserstein distance is calculated using an exponentially weighted moving average algorithm. When the trend smoothing value is greater than the adaptive threshold, it is determined that road condition concept drift has occurred.

6. The adaptive prediction method for unmanned truck travel time based on road condition drift perception according to claim 1, characterized in that, The method for filtering invalid physical noise and extracting valid drift samples based on the information entropy admission judgment mechanism to adaptively incrementally update the affected intra-cluster prediction model is as follows: Trajectory samples experiencing road condition concept drift are pushed into a short-term memory (STM) with a first-in, first-out (FIFO) mechanism; the feature information entropy H of the sample set in the STM is calculated; and the method is applied if and only if H is greater than a preset richness threshold H0. threshold At that time, the approved sample set is entered into the long-term memory, and the incremental learning interface is called to update the intra-cluster prediction model parameters affected by road condition drift.

7. An adaptive prediction system for unmanned truck travel time based on road condition drift perception according to claim 1, characterized in that, This includes in-vehicle edge computing devices deployed on unmanned trucks; The cloud-based scheduling server communicates with the vehicle-mounted edge computing device.

Images