A logistics network multi-source data fusion intelligent scheduling decision optimization method

Abstract

The application discloses a kind of logistics network multi-source data fusion intelligent scheduling decision optimization method.The closed loop system is composed of three modules: adaptive data fusion module: construct space-time dynamic health index, combine graph attention network interpolation compensation low confidence data, and with dynamic weighting improves D-S evidence theory, effectively solves the Zadeh paradox of high conflict data fusion.Multi-objective online scheduling module: utilize XGBoost to build lightweight carbon perception agent model to realize microsecond level reasoning, combined with non-dominated sorting deep reinforcement learning, achieve three-objective Pareto online collaborative optimization of carbon emission, time limit and cost.Continuous incremental learning module: parallel double-domain drift detection, only in real drift triggers flexible weight consolidation mechanism, using Fisher information matrix protects key parameters, prevents catastrophic forgetting.

Description

Technical Field

[0001] This invention belongs to the fields of smart logistics, big data analysis and artificial intelligence technology, and specifically relates to a method for intelligent scheduling and decision optimization of logistics networks through multi-source data fusion. Background Technology

[0002] With the continued growth of global e-commerce, modern logistics networks have evolved into highly complex systems encompassing multimodal transport and integrated warehousing and distribution. Large logistics companies have completed the construction of their IoT infrastructure, equipping vehicles with high-precision GPS modules and onboard OBD interfaces, deploying RFID readers at the warehouse level, and obtaining dynamic information about the external environment through real-time traffic APIs from map service providers and weather data interfaces. This multi-source heterogeneous data provides real-time support for intelligent dispatching systems.

[0003] The closest existing technical solution to this invention is: a multi-objective logistics path planning system based on deep reinforcement learning, which combines traditional DS evidence theory for multi-source data fusion preprocessing. Its data fusion layer pre-defines a fixed basic probability assignment for each data source and uses the Dempster merging rule to merge the basic probability assignments from multiple sources pairwise; the scheduling optimization layer uses deep reinforcement learning with an Actor-Critic framework for path planning, and the reward function contains only two objectives: delivery time cost and vehicle mileage cost, which are linearized into a single-objective optimization using a weighted summation method; the model update layer employs a full retraining strategy when the scheduling model performance degrades.

[0004] The aforementioned existing technology has the following three drawbacks:

[0005] First, traditional Data Sector Evidence Theory (DS) uses a basic probability allocation with fixed static weights, which cannot perceive real-time quality changes in the data source. When equipment failure or signal anomalies cause a data source to output abnormally high conflict data, traditional DS theory triggers the "Zadeh Paradox," concentrating all confidence on propositions that contradict all evidence, resulting in severely distorted fused output conclusions. Furthermore, there is no compensation mechanism when data sources are missing, and the data directly participates in subsequent decisions in a missing state, leading to serious biases in scheduling decisions.

[0006] Secondly, the precise carbon emission accounting model takes up to seconds to calculate, making it impossible to directly embed the objective function of a millisecond-level online scheduling engine, resulting in a tradeoff between real-time performance and accuracy. Even if a simplified carbon model is forcibly incorporated, the weighted summation linearization scheme cannot output a true Pareto optimal solution set, depriving decision-makers of a diverse range of strategy options.

[0007] Third, in extreme scenarios such as major promotional events and sudden road closures, logistics demand and road network characteristics undergo precipitous changes, causing trained scheduling models to quickly become ineffective. Full retraining strategies take hours to days and suffer from catastrophic forgetting problems due to incompatibility between new and old knowledge, failing to meet the real-time adaptive requirements of scheduling systems. Summary of the Invention

[0008] In response to the three shortcomings of existing intelligent scheduling systems for logistics networks, the technical problems to be solved by this invention are: how to ensure the reliability of scheduling decisions in scenarios with high conflict of multi-source data, how to overcome the computational bottleneck of real-time accurate carbon emission accounting and online multi-objective scheduling coordination, and how to maintain the continuous effectiveness of the scheduling model under the drift of logistics characteristic concepts.

[0009] This invention provides a method for intelligent scheduling decision optimization of logistics networks through multi-source data fusion. This method comprises three sequentially executed functional modules: adaptive confidence-based multi-source conflict data fusion, lightweight carbon-sensing proxy model and non-dominated ranking deep reinforcement learning multi-objective online scheduling, and dual-domain drift detection and elastic weight consolidation incremental continuous learning. These three modules form an end-to-end closed-loop scheduling system. Specifically, the method includes the following steps:

[0010] Step 1: Preprocessing of multi-source raw data

[0011] The raw data streams collected from four types of data sources—GPS / OBD, RFID, weather API, and traffic API—are format-unified: all data fields are mapped to a unified data pattern to eliminate heterogeneous format differences; the timestamps of each data source are aligned, and data with inconsistent sampling frequencies are resampled using linear interpolation based on the system's unified clock to unify them to a fixed sampling period; and each feature dimension is normalized to a maximum-minimum range, mapping the numerical range to the [0,1] interval.

[0012] Step Two: Dynamic Assessment of Data Source Health

[0013] Suppose the system connects to N data sources, and the health score of the i-th data source at time t is... It is determined by both the time decay factor and the spatial correlation score.

[0014] Time decay factor This reflects the degree to which the real-time reliability of the data source decreases over time since its last normal reporting.

[0015]

[0016] in, The attenuation rate hyperparameter of the i-th data source is determined by prior statistical analysis of the historical failures of that data source. t is the time interval between the current time t and the last time the i-th data source reported normal data; e is a natural constant.

[0017] Spatial correlation score This reflects the degree of consistency between the reading of the i-th data source at the current moment and the readings of its spatially neighboring data sources.

[0018]

[0019] in, Let i be the set of neighboring data sources in the logistics network topology graph; and These are the normalized readings of data source i and neighboring data source j at time t, respectively. The number of neighboring data sources; To prevent extremely small positive numbers from being divided by zero, take ; The value range is [0, 1], and the higher the value, the more consistent it is with the neighboring data source.

[0020] The formula for calculating overall health is:

[0021]

[0022] in, and The fusion weights for the time decay factor and the spatial correlation score are respectively, satisfying... The default value is 0.5.

[0023] Dynamic health threshold The calculation method is as follows:

[0024]

[0025] in, This is the sensitivity coefficient, with a value range of (0, 1), and a default value of 0.6; N is the total number of currently connected data sources. When When the i-th data source is determined to be in a low confidence state, the graph attention network adaptive interpolation compensation process is triggered.

[0026] Step 3: Adaptive Interpolation Compensation using Graph Attention Network

[0027] Constructing a diagram based on the topology of the logistics network Where V is the set of data source nodes and E is the set of edges. When data source i is detected to be in a low-confidence state, the compensation estimate is calculated through the following steps:

[0028] For the feature vector of each node i Perform a linear transformation: , where W is the learnable shared weight matrix.

[0029] Calculate the raw attention score between node pairs (i, j):

[0030]

[0031] in, This is a learnable attention parameter vector; represents the vector concatenation operation; LeakyReLU is the linear rectified activation function with leakage; T is the transpose sign.

[0032] The original attention scores are then subjected to Softmax normalization to obtain the normalized attention coefficients:

[0033]

[0034] in, This represents the attention weight of node i to its neighbor node j at time t; k is the attention weight of node i to its neighbor node j at time t. The traversal index of all neighboring nodes.

[0035] By using attention-weighted aggregation of neighbor node features, a compensation estimate is generated:

[0036]

[0037] in, The feature vector of node i at time t after graph attention network compensation is used to replace the original low-confidence data in subsequent fusion. This is the ELU nonlinear activation function.

[0038] Step 4: Improve the DS evidence theory fusion

[0039] After completing graph attention network compensation for low-confidence data, improved DS evidence theory fusion is performed on all data sources. The health of each data source is considered. The weights of the basic probability assignments are dynamically adjusted. The weighted basic probability assignment for the health of proposition A from the i-th data source is as follows:

[0040]

[0041] in, Let i be the original basic probability assignment of proposition A for the i-th data source; To identify the frame, which is the set of all possible propositions; The size of the power set is the number of all propositions; the second term on the right-hand side of the equation distributes the confidence of low-healthy data sources evenly across all propositions, reducing their interference with the final fusion result.

[0042] All data sources are weighted by health and their basic probability distributions are substituted into the Dempster merging rules for pairwise fusion. The final fusion confidence distribution is output, and the value corresponding to the proposition with the highest confidence is taken as the high-confidence unified data stream and passed to the multi-target online scheduling module.

[0043] Step 5: Offline Training of the XGBoost Lightweight Carbon-Sensing Proxy Model

[0044] Historical logistics operation data were collected to construct a training sample set for the proxy model. The feature vector of each sample includes road slope g (unit: %), vehicle instantaneous speed v (unit: km / h), vehicle current load w (unit: kg), real-time road congestion index c (normalized to [0, 1]), and the label is the actual carbon emissions (unit: g / km) calculated offline by the MOVES physical model.

[0045] The XGBoost algorithm is used to train regression on the sample set, with a new regression tree added in each iteration. Fit the residuals of the predictions from the first t-1 rounds, with the objective function being:

[0046]

[0047] Where M is the total number of training samples; Let the mean squared error loss function be used. This represents the cumulative forecast value from the previous t-1 rounds; This is the output of the t-th regression tree for the k-th sample; The regularization term includes L2 regularization, which includes penalties for the number of leaf nodes in the tree and leaf node weights, to prevent overfitting.

[0048] After training, verify the accuracy on an independent test set, requiring... The coefficient should be no less than 0.95. After successful verification, the surrogate model parameters will be frozen and deployed to the online scheduling environment for forward inference only. At each scheduling decision step, the real-time feature vector of the current road segment will be input into the surrogate model to obtain the estimated carbon emissions within microseconds. The carbon emission reward signal is input into the non-dominated ranking deep reinforcement learning algorithm.

[0049] Step Six: Non-dominated ranking deep reinforcement learning multi-objective online scheduling

[0050] Define the elements of a Markov decision process: state vector This includes a road network congestion state matrix, vehicle location and load state vectors, a set of orders to be delivered, and a time-deadline vector; actions. This involves a joint action that includes vehicle assignment decision and next-hop path decision; the reward is presented as a three-dimensional vector.

[0051]

[0052] in, The timeliness cost reward is defined as the weighted difference between the number of orders delivered within the current step and the timeliness breach penalty. The operating cost reward is defined as the negative value of the vehicle usage cost corresponding to the current mileage traveled. The carbon emission reward is defined as the negative value of the carbon emission estimated by the proxy model.

[0053] The Actor network uses a graph convolutional network to encode the road network state and vehicle-order state, outputting the probability distribution of scheduling actions. The scheduling actions are sampled from this distribution to obtain a three-dimensional reward vector. and the state at the next moment .

[0054] Maintaining the Pareto Archive It stores all non-dominated solutions found in the historical decision trajectory. It then compares the newly acquired cumulative reward vector with... Perform a dominance test on the existing solutions: if the new solution is superior to the previous solution in all three objectives... If a certain solution is found, then the old solution will be removed from the solution list. Remove from the middle; if the new solution is not... If any solution dominates, then add it to the list. .

[0055] Critic network for calculating the three-dimensional dominance function ,in Discount factor; For the Critic network to state The three-dimensional value estimation vector is obtained. The congestion distances corresponding to actions of each non-dominated solution in the Pareto archive are used as additional selection pressure to guide the Actor network to update parameters towards diverse Pareto optimal directions, following a truncated gradient approach for proximal policy optimization. After a sufficient number of online interaction rounds, a Pareto frontier policy set is output from the current Pareto archive, allowing decision-makers to select the final scheduling scheme based on current business priorities.

[0056] Step 7: Virtual Drift Detection

[0057] During system initialization, previously The data feature distribution within each time step is used as the baseline distribution. During system operation, the data feature distribution within the current sliding window is used as the current distribution. The maximum mean difference (MMD) is used to measure the distance between two distributions:

[0058]

[0059] in, To map the original features to the regenerating kernel Hilbert space The kernel function is a radial basis function kernel; The sample size is the baseline distribution. This represents the number of samples within the current sliding window. and These are the sample feature vectors of the baseline distribution and the current distribution, respectively; denoted as the norm in the regenerated kernel Hilbert space.

[0060] The calculated MMD value is compared with the preset virtual drift threshold. Comparison: If If a virtual drift occurs, the system will trigger a distribution drift alarm, increase the monitoring frequency of model prediction errors, but will not trigger model updates for the time being.

[0061] Step 8: Real Drift Detection

[0062] In each sliding monitoring window (window size is...) Within each scheduling decision cycle, calculate the average reward prediction error of the non-dominated ranking deep reinforcement learning scheduling model:

[0063]

[0064] in, This represents the three-dimensional reward vector actually obtained at the k-th decision step. Let be the Critic network's prediction of the reward at step k; Let L1 be the norm of the vector; To prevent extremely small positive numbers from being divided by zero.

[0065] The calculated Compared with the preset true drift threshold Comparison: If If a true drift occurs, the elastic weight consolidation incremental update process is immediately triggered.

[0066] Step Nine: Reinforce Incremental Continuous Learning with Elastic Weights

[0067] After a real drift is triggered, collect a new set of scene data samples following the drift. (Only a small number of recent scheduling trajectories are needed; full historical data is not required.) Based on the training data of the old task, the Fisher information matrix is ​​used to quantify the importance of each parameter of the non-dominated ranking deep reinforcement learning network to the performance of the old task. The Fisher information value of the j-th parameter is... The calculation method is as follows:

[0068]

[0069] in, This is a typical representative subset of samples from the old tasks; For the model in parameters The following sample The output probability; For the log-likelihood with respect to the parameter The partial derivatives; The larger the value, the more important the parameter is to older tasks.

[0070] Construct a loss function that reinforces incremental updates with elastic weights:

[0071]

[0072] in, For the model in the new scene dataset The non-dominated ranking deep reinforcement learning strategy gradient loss drives the model to adapt to new scenarios; These are fixed snapshot values ​​of each parameter after training for the old task; This is the square of the offset between the current parameters and the old task parameters; As a penalty coefficient, strong constraints are applied to important parameters of the old task, and weak constraints are applied to unimportant parameters of the old task. This is a regularization intensity hyperparameter used to balance the learning speed of new tasks with the degree of knowledge retention of old tasks.

[0073] Use the Adam optimizer Perform gradient descent iterative updates until... Drop to new scenario verification set The following verification also confirms that the scheduling reward decay of the model on representative samples of the old scenario does not exceed 5%. After the incremental update is completed, the updated parameters are used to overwrite the existing parameters of the non-dominated ranking deep reinforcement learning scheduling engine, and the baseline distribution and parameter snapshot are updated synchronously, and the system resumes normal scheduling operation.

[0074] The beneficial effects of this invention are:

[0075] 1. This invention completely solves the Zadeh paradox limitation of traditional DS evidence theory by using a dynamic evaluation mechanism for the health of data sources with spatiotemporal consistency constraints. The health of data sources can be dynamically evaluated within milliseconds. The graph attention network adaptive interpolation compensation mechanism significantly improves the data integrity rate and eliminates the risk of decision failure caused by data conflicts and missing data from the root.

[0076] 2. This invention approximates highly complex physical carbon emission models such as MOVES using the XGBoost surrogate model, compressing carbon emission inference time to the microsecond level. This represents a speed improvement of thousands of times compared to physical models, while also increasing prediction accuracy. The coefficient is no less than 0.95; for the first time in the field of logistics scheduling, Pareto online collaborative optimization of the three objectives of carbon emissions, delivery timeliness and operating costs has been achieved, giving decision-makers the ability to flexibly balance between low carbon and high efficiency.

[0077] 3. The dual-domain drift detection mechanism of this invention reduces the concept drift detection latency to the minute level; the elastic weight consolidation incremental update only requires a small number of new scene samples to complete the parameter correction, and the update time is compressed to the minute level. Under the premise of avoiding catastrophic forgetting, the scheduling performance of the old scene does not degrade by more than 5%, enabling the scheduling model to efficiently adapt to extreme concept drift scenarios such as major promotions and sudden events. Attached Figure Description

[0078] The invention will now be further described with reference to the accompanying drawings.

[0079] Figure 1 This is a diagram of the overall system architecture of the present invention, showing the hierarchical relationship and data flow between the data acquisition layer, the adaptive confidence multi-source conflict data fusion module, the lightweight carbon sensing agent model and non-dominated ranking deep reinforcement learning multi-objective online scheduling module, and the dual-domain drift detection and elastic weight consolidation incremental continuous learning module.

[0080] Figure 2 This is a flowchart for adaptive confidence multi-source conflict data fusion, demonstrating the complete processing flow from multi-source raw data input, data preprocessing, dynamic assessment of data source health, low-confidence data detection and graph attention network adaptive interpolation compensation, fusion with improved DS evidence theory, to high-confidence unified data stream output.

[0081] Figure 3 The flowchart of the non-dominated ranking deep reinforcement learning multi-objective online scheduling algorithm is shown, illustrating the complete online optimization loop structure of state space construction, Actor network forward inference, action sampling and execution, three-objective reward calculation, non-dominated ranking and Pareto profile update, Critic network parameter update, and Pareto front policy set output.

[0082] Figure 4The flowchart for dual-domain drift detection and elastic weight consolidation incremental continuous learning demonstrates the complete closed-loop process of virtual drift detection path, real drift detection path, Fisher information matrix calculation after real drift is triggered, construction of elastic weight consolidation loss function, incremental gradient update, performance verification and system recovery.

[0083] Figure 5 To provide a complete online closed-loop time sequence diagram for the system, the key interaction processes and time nodes of the four entities—multi-source data acquisition layer, data fusion module, non-dominated sorting deep reinforcement learning scheduling module, and drift detection and elastic weight consolidation and update module—are displayed in a time sequence diagram style during the normal operation phase, drift occurrence phase, elastic weight consolidation incremental update phase, and update completion and recovery phase. Detailed Implementation

[0084] 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.

[0085] I. System Overall Architecture

[0086] like Figure 1 As shown, the intelligent scheduling decision optimization method for multi-source data fusion in logistics networks proposed in this invention consists of three functional modules connected in series to form a closed-loop system: Module A is an adaptive confidence multi-source conflict data fusion module; Module B is a lightweight carbon-sensing agent model and non-dominated ranking deep reinforcement learning multi-objective online scheduling module; and Module C is a dual-domain drift detection and elastic weight consolidation incremental continuous learning module.

[0087] The data acquisition layer includes four types of data sources: vehicle GPS / OBD data sources, warehouse RFID data sources, external weather API data sources, and real-time traffic API data sources. Each data source aggregates the raw data stream to module A. Module A outputs high-confidence fused data to module B every 500ms. Module B completes scheduling decisions and outputs scheduling instructions within milliseconds, while simultaneously feeding back the three-dimensional reward vector to module C. Module C continuously monitors data distribution and prediction errors, triggering corresponding model updates when drift is detected. After the update is completed, the new parameters are fed back to module B, forming a system closed loop.

[0088] II. Module A: Adaptive Confidence-Based Multi-Source Conflicting Data Fusion

[0089] like Figure 2 As shown, the complete processing flow of module A is as follows.

[0090] Data preprocessing

[0091] The raw data streams collected from four types of data sources—GPS / OBD, RFID, weather API, and traffic API—are format-unified: all data fields are mapped to a unified JSON-LD data mode to eliminate heterogeneous format differences; timestamps from each data source are aligned, using the system's unified clock UTC millisecond precision as a benchmark, and data with inconsistent sampling frequencies are resampled through linear interpolation to unify to a fixed sampling period, with a default value of 500ms; and maximum-minimum normalization is performed on each feature dimension to map the numerical range to the interval between 0 and 1.

[0092] Dynamic assessment of data source health

[0093] Suppose the system connects to N data sources, and the health score of the i-th data source at time t is... Due to time decay factor Spatial correlation score A joint decision.

[0094] Time decay factor The calculation method is as follows:

[0095]

[0096] in, The decay rate hyperparameter of the i-th data source is determined by the prior statistical analysis of the historical failures of that data source. The larger the value, the more easily the data source decays over time. t is the time interval between the current time t and the last time the i-th data source reported normal data; e is a natural constant.

[0097] Spatial correlation score The calculation method is as follows:

[0098]

[0099] in, Let i be the set of neighboring data sources in the logistics network topology graph; and These are the normalized readings of data source i and neighboring data source j at time t, respectively. The number of neighboring data sources; To prevent extremely small positive numbers from being divided by zero, take ; The value range is [0, 1], and the higher the value, the more consistent it is with the neighboring data source.

[0100] The formula for calculating overall health is:

[0101]

[0102] in, and The fusion weights for the time decay factor and the spatial correlation score are respectively, satisfying... The specific values ​​of both are determined through cross-validation on historical data, with a default value of 0.5.

[0103] Dynamic health threshold The calculation method is a certain percentage of the average health of all current data sources:

[0104]

[0105] in, This is the sensitivity coefficient, with a value range of (0, 1), and a default value of 0.6; N is the total number of currently connected data sources. When When the i-th data source is determined to be in a low confidence state, the graph attention network adaptive interpolation compensation process is triggered.

[0106] Graph Attention Network Adaptive Interpolation Compensation

[0107] A graph is constructed based on the topology of the logistics network (nodes are data sources, and edges represent the road network connectivity or business relationships between nodes). , where V is the set of data source nodes and E is the set of edges.

[0108] When data source i is detected to be in a low-confidence state, the feature vector of each node i is... Perform a linear transformation on the data source, including its current multi-dimensional observation features:

[0109]

[0110] Where W is the learnable shared weight matrix.

[0111] Calculate the raw attention score between node pairs (i, j):

[0112]

[0113] in, This is a learnable attention parameter vector; represents the vector concatenation operation; LeakyReLU is the linear rectified activation function with leakage; T is the transpose sign.

[0114] The original attention scores are then subjected to Softmax normalization to obtain the normalized attention coefficients:

[0115]

[0116] in, This represents the attention weight of node i to its neighbor node j at time t; k is the attention weight of node i to its neighbor node j at time t. The traversal index of all neighboring nodes; It is an exponential function.

[0117] By aggregating neighbor node features using attention weighting, a compensation estimate for low-confidence node i is generated:

[0118]

[0119] in, The graph attention network compensated feature vector of node i at time t is used to replace the original low-confidence data in subsequent DS fusion. This is the ELU nonlinear activation function.

[0120] Improved DS evidence theory fusion

[0121] After performing graph attention network compensation on the low-confidence data, an improved DS evidence theory fusion was performed on all data sources (original data for those with normal confidence, and compensated estimates for those with low confidence). The improvement lies in: using the health of each data source... Instead of using fixed static weights, the weights of its basic probability allocation are dynamically adjusted.

[0122] Specifically, the weighted basic probability assignment of the health status of proposition A for the i-th data source is as follows:

[0123]

[0124] in, Let i be the original basic probability assignment of proposition A for the i-th data source; To identify the frame (the set of all possible propositions); The value is the size of the power set, i.e., the number of all propositions; the second term on the right side of the equation indicates that the confidence of low-healthy data sources is evenly distributed across all propositions to reduce their interference with the final fusion result.

[0125] All data sources are weighted by health and their basic probability distributions are substituted into Dempster's merging rules for pairwise fusion. The final fusion confidence distribution is output, and the value corresponding to the proposition with the highest confidence is taken as the high-confidence unified data stream output and passed to module B.

[0126] Module B: Lightweight Carbon-Sensing Proxy Model and Non-Dominated Ranking Deep Reinforcement Learning Multi-Objective Online Scheduling

[0127] like Figure 3As shown, Module B includes the construction of the XGBoost lightweight carbon-aware agent model in the offline pre-training stage, and the non-dominated ranking deep reinforcement learning multi-objective scheduling in the online running stage.

[0128] XGBoost Lightweight Carbon Sensing Proxy Model Construction

[0129] Offline pre-training phase: Collect historical logistics operation data to build a training sample set for the agent model. , where M is the total number of training samples; The feature vector of the k-th sample includes road gradient g (in %), vehicle instantaneous speed v (in km / h), vehicle current load w (in kg), and real-time road congestion index c (normalized to [0, 1], where 0 indicates completely unobstructed traffic and 1 indicates severe congestion). The actual carbon emissions (in g / km) corresponding to the k-th sample are calculated offline by the MOVES physical model and used as a label for supervised learning.

[0130] The XGBoost algorithm was used on the sample set. Perform regression training to obtain the surrogate model. XGBoost iteratively fits the residuals using an additive model, adding a new regression tree in each iteration. To fit the residuals of the predictions from the first t-1 rounds, the objective function is:

[0131]

[0132] in, Let the mean squared error loss function be used. This represents the cumulative forecast value from the previous t-1 rounds; This is the output of the t-th regression tree for the k-th sample; The regularization term includes L2 regularization, which includes penalties for the number of leaf nodes in the tree and leaf node weights, to prevent overfitting.

[0133] After training, the proxy model Accuracy validation was performed on an independent test set, requiring prediction accuracy (in words). The coefficient (measured) should not be lower than 0.95. If this is not met, adjust the hyperparameters of XGBoost, such as the number of trees and maximum depth, and retrain until the accuracy requirement is met. After successful verification, The parameters are frozen, and the system is deployed to the online scheduling environment for forward inference only.

[0134] Online inference phase: At each scheduling decision step, the real-time feature vector of the current road segment is... Input the proxy model and obtain the estimated carbon emissions for the current driving scenario within microseconds. As a carbon emission incentive signal Input non-dominated ranking deep reinforcement learning algorithm. Among them, The current road slope, The current instantaneous speed of the vehicle. Given the current vehicle load, The real-time congestion index for the current road segment is derived from the high-confidence fused data stream output by module A.

[0135] Non-dominated ranking deep reinforcement learning multi-objective online scheduling

[0136] The elements of a Markov decision process are defined as follows:

[0137] state space : The state vector at each time t This includes: a road network state matrix (an adjacency matrix composed of the congestion index of each road segment); vehicle state vectors (the current location node number of each vehicle, the current load as a percentage of the rated load, and the remaining battery power); and order state vectors (the destination node of the currently uncompleted order, the weight of the goods, and the number of minutes remaining until the deadline).

[0138] Action space Actions at each decision step It involves a combination of two types of decisions: vehicle assignment decision (assigning vehicle u to receive order p); and next-hop path decision (specifying the next road network node number that the vehicle should go to).

[0139] Reward Vector Instead of traditional scalar rewards, a three-dimensional reward vector is adopted.

[0140]

[0141] in, The timeliness cost reward is defined as the weighted difference between the number of orders delivered within the current step and the timeliness penalty. The timeliness penalty is the number of minutes exceeding the deadline multiplied by the unit timeliness penalty coefficient. The operating cost reward is defined as the negative value of the vehicle usage cost corresponding to the current mileage traveled. Carbon emission rewards are defined as carbon emission estimates based on proxy models. The negative value.

[0142] The Actor network uses a graph convolutional network to encode the road network state and vehicle-order state, and outputs the probability distribution of each scheduling action. ,in These are the learnable parameters of the Actor network. The scheduling action for the current step is sampled from this probability distribution. And execute, to obtain a three-dimensional reward vector. and the state at the next moment .

[0143] Maintaining the Pareto Archive It stores all non-dominated solutions found in the historical decision trajectory. It then compares the newly acquired cumulative reward vector with... Perform a dominance test on the existing solutions: if the new solution is superior to the previous solution in all three objectives... If a certain solution is found, then the old solution will be removed from the solution list. Remove from the middle; if the new solution is not... If any solution dominates, then add it to the list. .

[0144] Critic network for calculating the three-dimensional dominance function ,in Discount factor; For the Critic network to state The three-dimensional value estimation vector. Utilizing the Pareto archives. The crowding distance for each non-dominated solution corresponds to an action in the network. This distance reflects the sparsity of the solution on the Pareto front and serves as additional selection pressure, guiding the Actor network to update its parameters towards diverse Pareto optimal directions. The Actor-Critic network parameters are updated using a truncated gradient approach similar to proximal policy optimization, ensuring that the policy update step size is controlled.

[0145] After a sufficient number of online interactive rounds, from the current Pareto archive The system outputs a Pareto frontier strategy set, allowing decision-makers to select the final scheduling plan based on current business priorities, such as prioritizing timeliness during peak sales periods and controlling carbon emissions during routine operations.

[0146] Module C: Dual-Domain Drift Detection and Elastic Weight Consolidation Incremental Continuous Learning

[0147] like Figure 4 As shown, module C includes two parallel paths: virtual drift detection and real drift detection, as well as an elastic weight consolidation incremental update process after a real drift is triggered.

[0148] Virtual drift detection

[0149] During system initialization, previously The data feature distribution within each time step is used as the baseline distribution. During system operation, the data feature distribution within the current sliding window is used as the current distribution. The window size is W time steps.

[0150] The maximum mean difference is used to measure the distance between two distributions:

[0151]

[0152] in, To map the original features to the regenerating kernel Hilbert space The kernel function (using radial basis function kernels); The sample size is the baseline distribution. This represents the number of samples within the current sliding window. and These are the sample feature vectors of the baseline distribution and the current distribution, respectively; denoted as the norm in the regenerated kernel Hilbert space.

[0153] The calculated MMD value is compared with the preset virtual drift threshold. Comparison: If If a virtual drift occurs, the system will trigger a distribution drift alarm, increase the monitoring frequency of model prediction errors, but will not trigger model updates for the time being.

[0154] Real drift detection

[0155] In each sliding monitoring window (window size is...) Within each scheduling decision cycle, the average reward prediction error of the non-dominated ranking deep reinforcement learning scheduling model is calculated as the detection signal for true drift.

[0156]

[0157] in, This represents the three-dimensional reward vector actually obtained at the k-th decision step. Let be the Critic network's prediction of the reward at step k; Let L1 be the norm of the vector; To prevent extremely small positive numbers from being divided by zero.

[0158] The calculated Compared with the preset true drift threshold Comparison: If If a true drift occurs, the elastic weight consolidation incremental update process is immediately triggered.

[0159] Elastic weights reinforce incremental continuous learning

[0160] After a real drift is triggered, collect a new set of scene data samples following the drift. (Only a small number of recent scheduling trajectories are needed; full historical data is not required.)

[0161] Based on training data from previous tasks and normal scenarios, calculate the parameters of the non-dominated ranking deep reinforcement learning network. The importance of legacy task performance is quantified using the Fisher information matrix F. The Fisher information value of the j-th parameter... The calculation method is as follows:

[0162]

[0163] in, This is a typical representative sample subset of the old tasks, extracted from historical data, and its quantity is much smaller than the full historical data. For the model in parameters The following sample The output probability; For the log-likelihood with respect to the parameter The partial derivatives; The larger the value, the more important the parameter is to the old task, and the stronger the constraints need to protect it during incremental updates.

[0164] Construct a loss function that reinforces incremental updates with elastic weights:

[0165]

[0166] in, For the model in the new scene dataset The non-dominated ranking deep reinforcement learning strategy gradient loss drives the model to adapt to new scenarios; These are fixed snapshot values ​​of each parameter after training for the old task; This is the square of the offset between the current parameters and the old task parameters, representing the magnitude of the change to the old task parameters; As a penalty coefficient for this offset, important parameters of the old task (large) Strengthen constraints on parameters that are not important for old tasks (small) Apply weak constraints; The hyperparameter for strengthening the regularization strength of the elastic weights is used to balance the learning speed of new tasks with the knowledge retention of old tasks, and is tuned through joint evaluation of dual-task performance on the validation set.

[0167] Use the Adam optimizer Perform gradient descent iterative updates until... On the validation set of the new scenario, the speed was reduced to The following also verifies that the scheduling reward decay of the model on representative samples of the old scenario does not exceed 5%.

[0168] After the incremental update is complete, use the updated parameters. The existing parameters of the non-dominated ranking deep reinforcement learning scheduling engine are overridden, and the system resumes normal scheduling operation. The baseline distribution is updated synchronously. To determine the data distribution for the new scene after drifting and update it. This is a snapshot of the current parameters, preparing for the next round of drift detection and incremental updates.

[0169] V. Complete online operation closed loop

[0170] like Figure 5 As shown, during normal operation, the data acquisition layer continuously sends raw data streams to module A. Module A outputs high-confidence fused data to module B every 500ms. Module B completes scheduling decisions and outputs scheduling instructions within milliseconds, while simultaneously feeding back the three-dimensional reward vector to module C. Module C continuously monitors data distribution and prediction errors, with no drift alarms. During the drift phase, such as at the start of a major promotion, module C's virtual drift detector first triggers an alarm when the MMD exceeds the threshold. Several time steps later, the real drift detector triggers an elastic weight consolidation incremental update when the MAPE exceeds the threshold. During the elastic weight consolidation incremental update phase, module C notifies module B to enter conservative scheduling mode, running with the old parameters, while asynchronously executing the elastic weight consolidation parameter update. After the update is complete, module C notifies module B to switch to the new parameters and resume normal scheduling. After the update is complete, the baseline distribution and parameter snapshot are synchronously refreshed, and the system returns to normal operation closed loop.

[0171] VI. Alternative Implementation Methods

[0172] In the data fusion module, the graph attention network adaptive interpolation compensation submodule can be replaced with a Bayesian deep learning method: for low-confidence data sources, the Bayesian neural network outputs the predicted mean and predicted variance during multiple random forward inferences and Monte Carlo Dropout sampling. The mean is used as the compensation estimate, and the variance is used as the quantitative indicator of the compensation uncertainty. The uncertainty measure is synchronously passed to the improved DS fusion stage to further adjust the basic probability allocation weights of the data source. This scheme has stronger adaptability in scenarios where the data source topology is unclear or the road network structure changes frequently, making graph construction difficult; the scheduling optimization module and the continuous learning module remain unchanged.

[0173] In the multi-objective online scheduling module, the non-dominated ranking deep reinforcement learning algorithm can be replaced with a hybrid architecture of multi-objective evolutionary algorithm and deep learning: a deep neural network is used as the surrogate evaluator for the scheduling strategy, and a decomposition-based multi-objective evolutionary algorithm is used as the population evolution search framework. A set of weight vectors uniformly distributed in the objective space is maintained, decomposing the three-objective problem into several scalar quantum problems to be solved in parallel. The deep neural network surrogate evaluator is used to quickly estimate the three-objective rewards of each candidate scheduling scheme, thereby achieving Pareto front approximation without real-time interaction with the environment. The carbon-sensing surrogate model embedding method remains unchanged, and the continuous learning module remains unchanged.

[0174] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

[0175] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent scheduling and decision optimization of logistics networks through multi-source data fusion, characterized in that, The method includes the following steps: The raw data streams from GPS / OBD data sources, RFID data sources, weather API data sources, and traffic API data sources are processed to unify the format, align the timestamps, and perform maximum-min normalization preprocessing to form a unified data stream. A comprehensive health score is calculated for each data source, which is obtained by weighted fusion of time decay factor and spatial correlation score. A dynamic health threshold is set, and data sources with a comprehensive health score lower than the dynamic health threshold are marked as low confidence states. A graph attention network based on the logistics network topology is triggered to perform adaptive interpolation compensation on the low confidence data sources to generate compensation estimates. The comprehensive health scores of all data sources are used as dynamic weights to adjust their basic probability allocations, and are fused pairwise according to the Dempster merging rule to output a high-confidence unified data stream. In the offline pre-training phase, the XGBoost lightweight carbon-aware proxy model is trained using road gradient, instantaneous vehicle speed, current vehicle load, and road congestion index as input features, and carbon emissions calculated offline by the MOVES physical model as labels. In each online scheduling decision step, the real-time feature vector of the current road segment is input into the proxy model to obtain the estimated carbon emissions within microseconds. A state space is constructed using the high-confidence unified data stream. A three-dimensional reward vector is used as the reward signal for the non-dominated ranking deep reinforcement learning scheduling algorithm. The three-dimensional reward vector consists of time cost reward, operating cost reward, and carbon emission reward provided by the agent model. Pareto archives are maintained through non-dominated ranking, and crowding distance is used to guide the update of Actor-Critic network parameters, outputting a Pareto frontier policy set.

2. The intelligent scheduling and decision-making optimization method for multi-source data fusion in a logistics network according to claim 1, characterized in that, The method also includes the following steps: Parallel operation of virtual drift detection and real drift detection: The distance between the current data distribution and the benchmark distribution is measured based on the maximum mean difference. When the distance exceeds the virtual drift threshold, a distribution drift alarm is triggered. When the average reward prediction error of the scheduling model within the sliding window exceeds the real drift threshold, an elastic weight consolidation incremental update is triggered. The elastic weight consolidation incremental update quantifies the importance of old task parameters using the Fisher information matrix, constructs a loss function that includes the new task learning loss and the old task parameter protection penalty term, performs incremental gradient updates on the scheduling model parameters, and synchronously refreshes the benchmark distribution and parameter snapshot after the update is completed.

3. The method according to claim 2, characterized in that, The time decay factor The calculation method is as follows: in, The attenuation rate hyperparameter of the i-th data source is determined by prior statistical analysis of the historical failures of that data source. The time interval between the current time t and the last time the i-th data source reported normal data; e is a natural constant; The spatial correlation score The calculation method is as follows: in, Let i be the set of neighboring data sources in the logistics network topology graph; and These are the normalized readings of data source i and neighboring data source j at time t, respectively. The number of neighboring data sources; To prevent extremely small positive numbers from being divided by zero; The overall health score The calculation method is as follows: in, and The fusion weights for the time decay factor and the spatial correlation score are respectively, satisfying... .

4. The method according to claim 2, characterized in that, The dynamic health threshold The calculation method is as follows: in, is the sensitivity coefficient, with a value range of (0, 1); N is the total number of currently connected data sources; Let be the overall health score of the i-th data source at time t; when the overall health score of the i-th data source... Below the dynamic health threshold When the data source is determined to be in a low-confidence state, the graph attention network adaptive interpolation compensation is triggered.

5. The method according to claim 2, characterized in that, The specific steps of the graph attention network adaptive interpolation compensation are as follows: A graph is constructed based on the logistics network topology, with nodes as data sources and edges representing the road network connectivity between nodes; after performing a linear transformation on the feature vector of each node, the normalized attention coefficient of the target low-confidence node i to its neighbor node j is calculated. in, This is a learnable attention parameter vector; and These are the feature vectors of nodes i and j after linear transformation, respectively. This represents a vector concatenation operation; Let k be the set of neighboring nodes of node i; k is the set of nodes of node i. The traversal index of all neighboring nodes is obtained; the features of neighboring nodes are weighted and aggregated using the normalized attention coefficient, and a compensation estimate for low-confidence node i is generated through the ELU activation function, which replaces the original low-confidence data in the subsequent improved DS evidence theory fusion.

6. The method according to claim 2, characterized in that, The specific method for dynamically adjusting the basic probability allocation by using the comprehensive health score of each data source as a weight is as follows: the i-th data source is used to weight the basic probability allocation of the health score of proposition A. The calculation method is as follows: in, Let i be the original basic probability assignment of proposition A for the i-th data source; For identification framework; The size of the power set is given; the second term on the right-hand side of the equation distributes the confidence of low-health data sources evenly across all propositions, reducing their interference with the final fusion result; the basic probability allocation of all data sources after health weighting is substituted into the Dempster merging rule for pairwise fusion, outputting the final high-confidence unified data stream.

7. The method according to claim 2, characterized in that, The three-dimensional reward vector consists of three components: time-saving cost reward, operating cost reward, and carbon emission reward, all in vector form. These three components are not weighted and linearized. The non-dominated ranking deep reinforcement learning algorithm maintains a Pareto archive. It checks the dominance relationship between each newly acquired cumulative reward vector and existing solutions in the Pareto archive, adding non-dominated solutions to the Pareto archive. The algorithm uses the crowding distance of actions corresponding to each non-dominated solution in the Pareto archive as additional selection pressure to guide the Actor-Critic network parameters towards diverse Pareto optimal directions. After a sufficient number of online interaction rounds, it outputs a Pareto frontier policy set, allowing decision-makers to select the final scheduling scheme based on current business priorities.

8. The method according to claim 2, characterized in that, The method for calculating the distance between the current data distribution and the benchmark distribution based on the maximum mean difference is as follows: in, To map the original features to the regenerating kernel Hilbert space The radial basis function kernel; and These are the sample sizes for the baseline distribution and the current distribution, respectively. and These are the sample feature vectors of the baseline distribution and the current distribution, respectively; The norm in the regenerated kernel Hilbert space is represented; when the maximum mean difference exceeds the preset virtual drift threshold, a distribution drift alarm is triggered to increase the monitoring frequency of model prediction error, but without triggering model parameter updates.

9. The method according to claim 2, characterized in that, The loss function for the elastic weight consolidation incremental update is: in, The gradient loss of the non-dominated ranking deep reinforcement learning policy for the model on the new scene dataset; These are fixed snapshot values ​​of each parameter after training for the old task; The Fisher information value for the j-th parameter is calculated by taking the mean of the squared partial derivatives of the log-likelihood over a representative subset of samples from the old task. A larger value indicates that the parameter is more important for older tasks; To reinforce the regularization strength hyperparameter for elastic weights, the loss function is iteratively updated using gradient descent with the Adam optimizer until the average reward prediction error drops below the true drift threshold on the new scenario validation set and the scheduling performance degradation in the old scenario does not exceed 5%. Then, the incremental update is completed, and the baseline distribution and parameter snapshot are refreshed simultaneously.

Images