Deep learning cross-border e-commerce warehouse management system

Abstract

The invention relates to the technical field of cross-border e-commerce storage, in particular to a deep learning cross-border e-commerce storage management system, which comprises a data acquisition processing unit, a space-time diagram neural network analysis unit, an alliance management scheduling decision unit and an equipment execution feedback unit, a data acquisition and processing unit acquires and preprocesses storage data, a space-time diagram neural network analysis unit constructs a space-time diagram and predicts task requirements and equipment states, an alliance management scheduling decision-making unit constructs a dynamic alliance and generates a scheduling instruction, and an equipment execution feedback unit receives the instruction and autonomously adjusts a strategy. The problems of unreasonable equipment scheduling, poor collaboration and resource waste in traditional cross-border e-commerce warehouse management are solved, intelligent scheduling and collaborative operation of the warehouse equipment are realized, and the warehouse management efficiency and the resource utilization rate are improved.

Description

Technical Field

[0001] The present invention relates to the field of cross-border e-commerce warehousing technology, and in particular to a deep learning cross-border e-commerce warehousing management system. Background Art

[0002] Cross-border e-commerce warehousing is an important technology. With the booming development of cross-border e-commerce, warehousing management has become a key factor affecting the competitiveness of enterprises. Cross-border e-commerce business involves multiple national markets, complex and changeable order requirements, and a wide variety of goods. Traditional warehousing management systems have drawbacks.

[0003] Traditional cross-border e-commerce warehouse management systems rely on manual experience and simple information technology. Equipment scheduling often adopts a fixed grouping mode. When dealing with massive orders during promotional activities, fixed grouping equipment cannot be flexibly deployed. Equipment in some areas is often idle, while equipment in areas with dense orders is in short supply, resulting in a backlog of goods and delayed order processing. In terms of collaboration, information flow in each link is not smooth, and the storage, sorting, and transportation of goods are independent and lack efficient collaboration. For example, the transportation process of goods from the storage area to the sorting area often wastes a lot of time due to untimely communication and inadequate equipment coordination, resulting in serious waste of storage resources. In order to solve this technical problem, we provide a deep learning cross-border e-commerce warehouse management system. Summary of the invention

[0004] The purpose of the present invention is to provide a deep learning cross-border e-commerce warehouse management system to solve the problems raised in the above background technology.

[0005] To achieve the above purpose, a deep learning cross-border e-commerce warehouse management system is provided, including a data acquisition and processing unit, a spatiotemporal graph neural network analysis unit, an alliance management scheduling decision unit, and an equipment execution feedback unit;

[0006] The data acquisition and processing unit adopts a distributed sensor network based on the Internet of Things, and uses position sensors, weight sensors, state monitoring sensors and environmental sensors to collect the location, operating status, cargo information and storage environment parameters of cross-border e-commerce storage equipment in real time, and uses an adaptive data cleaning algorithm to pre-process the collected data to obtain a storage data set;

[0007] The spatiotemporal graph neural network analysis unit inputs the storage data into the spatiotemporal graph neural network in time series, thereby constructing a spatiotemporal graph containing nodes and edges, wherein the nodes represent storage equipment, cargo storage areas, and sorting areas, and the edges represent the connection relationship between them, and updates the graph structure at each time step to reflect the dynamic changes of the storage environment. By introducing the spatiotemporal attention mechanism, the edge association weights of different nodes and edges in different time and space are automatically calculated to obtain the spatiotemporal characteristics of the storage environment, and the reinforcement transfer learning strategy is used to combine historical storage data and real-time storage data to train the spatiotemporal graph neural network to predict the task requirements and equipment status in the future period;

[0008] The alliance management scheduling decision unit combines the prediction results and the current task with the dynamic alliance formation algorithm of multi-objective optimization, calculates the multi-dimensional matching degree of each device and the task, selects the device with the highest matching degree to form a dynamic alliance and generates a scheduling instruction;

[0009] The device execution feedback unit is used to receive the scheduling instructions generated by the alliance management scheduling decision unit, and autonomously adjust the execution strategy in combination with its own sensor data and environmental perception information.

[0010] As a further improvement of the technical solution, in the data acquisition and processing unit, the specific operation of the adaptive data cleaning algorithm is as follows:

[0011] Define a set of data quality assessment indicators, including data integrity, data accuracy, data consistency, and data timeliness, and assign an initial weight to each indicator;

[0012] For the collected raw data, the quality score of each data sample is calculated according to the initial weight, and the quality score is compared with the preset low-quality threshold and high-quality threshold, and the data is divided into low-quality data, medium-quality data and high-quality data according to the comparison results;

[0013] For low-quality data, data interpolation method is used to repair it. If it cannot be repaired, it is directly deleted. For medium-quality data, data correction is performed, and high-quality data is directly retained to obtain a cleaned warehouse data set.

[0014] As a further improvement of the present technical solution, the spatiotemporal graph neural network analysis unit includes a spatiotemporal attention mechanism module. When the spatiotemporal attention mechanism module constructs the spatiotemporal graph neural network, the edge association weight is calculated as follows:

[0015] For nodes representing storage equipment, cargo storage areas, and sorting areas, calculate the physical distances between them, and calculate the physical distance association weights based on the physical distances between them;

[0016] The business interaction frequency between nodes in historical data is counted, and the business association weight of the edge is calculated based on the business interaction frequency. The physical distance association weight and the business association weight are then substituted into the formula to calculate the edge association weight.

[0017] As a further improvement of the present technical solution, in the spatiotemporal attention mechanism module, when constructing a spatiotemporal graph neural network based on the spatiotemporal attention mechanism, the specific calculation method of the spatiotemporal attention mechanism is as follows:

[0018] For the feature vector of a node in the spatiotemporal graph at a time step, calculate its attention score with other nodes at different time steps;

[0019] The attention scores are normalized to obtain the attention weights, and the attention weights are substituted into the formula to calculate the context vector of the node at this time step. The context vector is used to represent various features and information related to the node and the edge at this time step.

[0020] Finally, the context vector is concatenated and nonlinearly transformed with the node’s own feature vector to obtain the updated node feature vector.

[0021] As a further improvement of the technical solution, the spatiotemporal graph neural network analysis unit includes a reinforcement transfer learning module, and the specific process of the reinforcement transfer learning strategy in the reinforcement transfer learning module is:

[0022] Collect historical warehouse data, clean and interpolate the historical warehouse data, perform feature engineering on the processed historical data, extract the frequency of task peak hours as the key feature, and construct a historical feature data set;

[0023] Extract real-time warehouse data from the data acquisition and processing unit, perform real-time warehouse data cleaning in real time, and then perform feature engineering processing to extract real-time features and construct a real-time feature data set;

[0024] Build a spatiotemporal graph neural network model based on the spatiotemporal attention mechanism, determine the initial representation of nodes and edges, and train it on the source task. The source task is set to predict the task requirements and equipment status changes in subsequent periods in history based on the historical feature dataset.

[0025] Migrate the parameters in the source model to the preset target model. The migrated parameters are frozen during the initial training phase of the target model, and the architecture of the target model is adjusted.

[0026] The target model is adjusted and trained on the target task based on the real-time feature data set. The target task is to accurately predict the task requirements and equipment status in the future period. The reinforcement learning framework is introduced, the reward function is defined, and the proximal strategy optimization algorithm is used to update the target model parameters. In each iteration, the predicted action is generated according to the current model strategy, and the reward is obtained by comparing it with the actual subsequent situation until the reward value reaches the preset threshold, and the target model training is determined to be completed.

[0027] As a further improvement of the technical solution, the specific steps of the dynamic alliance formation algorithm of multi-objective optimization in the alliance management scheduling decision unit are:

[0028] The spatiotemporal characteristics of the storage environment are obtained from the spatiotemporal graph neural network analysis unit, and the current set of tasks to be performed and the set of available equipment are clarified, wherein the set of tasks and the set of available equipment are provided by the storage data set;

[0029] A multi-objective function is defined according to the task set and the available equipment set, wherein the multi-objective function is respectively composed of the shortest task completion time, the lowest equipment energy consumption, the highest equipment utilization rate and the lowest task cost, and a weight is assigned to each objective function;

[0030] For each device and task, its score under each objective function is calculated respectively, and then the multi-dimensional matching degree of the device and the task is calculated according to its score and assigned weight under each objective function. Finally, a genetic algorithm is used to form a dynamic alliance.

[0031] As a further improvement of the technical solution, the specific operation of the alliance management scheduling decision unit using the genetic algorithm to form a dynamic alliance is as follows:

[0032] Integer encoding is used. Devices and The task allocation problem is to encode each individual as a sequence of integers, the first in the sequence The element represents the The device number to which the task is assigned;

[0033] Randomly generate the initial population , the population size is , each individual represents a device-task allocation scheme. At the same time, a heuristic strategy is introduced to generate some individuals. According to the multi-dimensional matching degree, for each task, the one with the highest matching degree is selected. Devices are selected as candidate devices, and then a device is randomly selected from these candidate devices and assigned to the task. This process is repeated to generate multiple individuals to join the initial population.

[0034] For each individual in the population, its fitness is calculated. The fitness function is the sum of the multi-dimensional matching degrees of all tasks and devices under the device-task allocation scheme corresponding to the individual. The tournament selection method is introduced, that is, in each generation, randomly select from the population Individuals form a tournament group, compare the fitness values ​​of individuals in the group, select the individual with the highest fitness as the parent individual, and repeat this process until a preset number of parent individuals are selected for crossover and mutation operations;

[0035] For two parent individuals and , analyze the association relationship between its tasks, and then randomly select a group of associated tasks as the intersection point. The device allocation corresponding to the associated task group is copied to the child individual At the same time, The device allocation information of other tasks except the associated task group is copied to In the formation of a progeny individual, The device allocation corresponding to the associated task group is copied to the child individual At the same time, The device allocation information of other tasks except the associated task group is copied to In the next generation, another individual is generated ;

[0036] Introducing adaptive mutation rate, during mutation operation, randomly select a task and replace its assigned device with another randomly selected device. In addition to reaching the maximum number of iterations as the termination condition, introducing fitness convergence judgment, calculating the fitness value of the best individual in each generation of population, when continuous When the fitness value of the optimal individual of the generation changes less than a threshold, the algorithm is considered to have converged and the iteration process is terminated;

[0037] When the genetic algorithm terminates, the individual with the highest fitness is selected as the optimal equipment-task allocation scheme, and the corresponding scheduling instructions are generated for each task based on the scheme.

[0038] As a further improvement of the technical solution, the device execution feedback unit adopts a fuzzy control algorithm when autonomously adjusting the execution strategy. The specific steps are as follows:

[0039] The sensor data and environmental perception information of the device are used as input variables, and each input variable is fuzzified to divide different fuzzy subsets corresponding to different power intervals, and the membership function of each subset is defined;

[0040] Fuzzy rules are formulated based on historical experience. Each rule has corresponding premises and conclusions. The premises are fuzzy subset combinations of input variables, and the conclusions are fuzzy subsets of actions that the device should take.

[0041] According to the membership value of the input variable and the fuzzy rules, fuzzy reasoning is performed. For each rule, the membership of the premise condition is calculated, and the minimum value is taken as the excitation strength of the rule. Then, the fuzzy subset of the conclusion is truncated according to the excitation strength to obtain the output fuzzy subset of each rule.

[0042] The multiple output fuzzy subsets obtained by fuzzy reasoning are merged and defuzzified using the centroid method to obtain the specific control quantity, that is, the actual action that the device should take. The device adjusts its execution strategy based on this control quantity.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] In a deep learning cross-border e-commerce warehouse management system, the warehouse data is constructed into a space-time graph through the space-time graph neural network analysis unit, and the space-time attention mechanism and enhanced transfer learning strategy are introduced to accurately obtain the space-time characteristics of the warehouse environment, predict future task requirements and equipment status, which is helpful to plan scheduling in advance and improve the foresight and initiative of warehousing operations. The alliance management scheduling decision-making unit uses a dynamic alliance formation algorithm and genetic algorithm with multi-objective optimization to form a dynamic alliance and generate scheduling instructions based on the prediction results and current tasks, realize efficient matching of equipment and tasks, achieve the effect of optimizing resource allocation, reduce equipment energy consumption, shorten task completion time, improve equipment utilization and reduce task costs. The equipment execution feedback unit receives the scheduling instructions and uses the fuzzy control algorithm to autonomously adjust the execution strategy, so that the equipment can flexibly respond to complex environments and ensure accurate and efficient execution of tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is an overall block diagram of the present invention.

[0046] The meaning of each number in the figure is:

[0047] 1. Data acquisition and processing unit; 2. Spatiotemporal graph neural network analysis unit; 21. Spatiotemporal attention mechanism module; 22. Reinforcement transfer learning module; 3. Alliance management and scheduling decision-making unit; 4. Equipment execution feedback unit. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The present invention provides a deep learning cross-border e-commerce warehouse management system, please refer to Figure 1 As shown, it includes a data acquisition and processing unit 1, a spatiotemporal graph neural network analysis unit 2, an alliance management and scheduling decision unit 3, and a device execution feedback unit 4;

[0050] The data acquisition and processing unit 1 adopts a distributed sensor network based on the Internet of Things, and uses position sensors, weight sensors, status monitoring sensors and environmental sensors to collect the location, operating status, cargo information and storage environment parameters of cross-border e-commerce warehousing equipment in real time, and uses an adaptive data cleaning algorithm to pre-process the collected data to obtain a warehousing data set.

[0051] In the data acquisition and processing unit 1, the specific operation of the adaptive data cleaning algorithm is as follows:

[0052] In order to comprehensively and objectively evaluate the quality of the collected data, it is necessary to clarify specific evaluation indicators and assign weights to each indicator to reflect its importance in the overall data quality assessment.

[0053] Define a set of data quality assessment indicators ,in Represents data integrity, Represents data accuracy, Represents data consistency, Represents the timeliness of data and assigns an initial weight to each indicator , and satisfies Through clear indicators and weights, data quality can be quantitatively evaluated, avoiding the uncertainty of subjective judgment, making data quality assessment more scientific and accurate, and providing a reliable basis for subsequent data cleaning operations.

[0054] By calculating the quality score of each data sample, we can intuitively understand the data quality of each sample. Classification based on the score helps to adopt different processing strategies for data of different quality.

[0055] For each data sample in the collected original data, its score under each indicator is calculated, and then the quality score is calculated based on the initial weight and its score under each indicator. The low-quality threshold and the high-quality threshold are set to divide the data into low-quality data, medium-quality data and high-quality data. This classification method can process data of different qualities in a targeted manner and improve the efficiency and effectiveness of data cleaning.

[0056] After the above processing for data of different qualities, the processed data are integrated together to form a cleaned storage data set for subsequent analysis and application. The processed low-quality data, medium-quality data and high-quality data are merged to obtain a cleaned storage data set.

[0057] The spatiotemporal graph neural network analysis unit 2 inputs the warehousing data into the spatiotemporal graph neural network in time series, and constructs a spatiotemporal graph containing nodes and edges, wherein the nodes represent warehousing equipment, cargo storage areas, and sorting areas, and the edges represent the connection relationship between them. The graph structure is updated at each time step to reflect the dynamic changes of the warehousing environment. By introducing the spatiotemporal attention mechanism, the edge association weights of different nodes and edges in different time and space are automatically calculated to obtain the spatiotemporal characteristics of the warehousing environment. The reinforcement transfer learning strategy is used to train the spatiotemporal graph neural network in combination with historical warehousing data and real-time warehousing data to predict task requirements and equipment status in future time periods.

[0058] The spatiotemporal graph neural network analysis unit 2 includes a spatiotemporal attention mechanism module 21. When the spatiotemporal attention mechanism module 21 constructs the spatiotemporal graph neural network, the edge association weight is calculated as follows:

[0059] In the cross-border e-commerce warehouse management system, the physical distance between storage equipment, cargo storage areas and sorting areas is an important factor affecting the association between them. The closer the physical distance, the more frequent and convenient the interaction between them may be. Therefore, it is necessary to consider the physical distance to calculate the association weight.

[0060] For nodes representing storage equipment, cargo storage areas, and sorting areas and nodes , using the coordinate information to calculate the physical distance between them , define the physical distance association weight calculation function. Generally, the closer the physical distance, the higher the association weight. The influence of physical space factors on the association between nodes is considered, making the weight calculation more in line with the actual warehousing environment and more accurately reflecting the degree of association between nodes based on physical distance, which helps the model better capture the spatial information in the warehousing system.

[0061] In addition to physical distance, the frequency of business interactions between nodes is also a key indicator to measure the strength of their association. Nodes with frequent business interactions are more closely connected and have a greater impact on the overall operation of the warehousing system. Therefore, it is necessary to count the frequency of business interactions to calculate the association weight.

[0062] In historical data, statistical nodes and nodes Number of business interactions between At the same time, the total number of business interactions between all nodes is counted , and calculate the nodes accordingly and nodes The business interaction frequency between nodes, the business association weight can directly adopt the value of the business interaction frequency, combined with the actual business situation, measure the association between nodes from the perspective of business interaction, so that the weight can better reflect the importance of the node in the business process.

[0063] The physical distance association weight and business association weight reflect the degree of association between nodes from different perspectives respectively. Combining the two can calculate the edge association weight more comprehensively and accurately, thereby better constructing a spatiotemporal graph neural network.

[0064] Associating weights with physical distance and business association weight Substitute the formula to calculate the edge association weight , using the weighted summation method, the weight coefficient of the physical distance association weight is , the weight coefficient of the business association weight is , then the edge association weight ; It comprehensively considers two important factors, physical distance and business interaction, making the calculation of edge association weight more comprehensive and accurate.

[0065] In the spatiotemporal attention mechanism module 21, when constructing a spatiotemporal graph neural network based on the spatiotemporal attention mechanism, the specific calculation method of the spatiotemporal attention mechanism is as follows:

[0066] Suppose the nodes in the space-time graph At time step The eigenvector of ,node At time step The eigenvector of , in order to calculate the attention score between them , firstly, the feature vector is mapped to a new space through linear transformation, that is, , ,in and is a learnable weight matrix, then, the attention score ;here is the dimension of the feature vector, divided by This is to prevent the dot product result from being too large. This calculation method takes into account the linear transformation and scale normalization of node features, which can more accurately capture the temporal and spatial correlation information between nodes and avoid the numerical instability problem that may be caused by direct dot products.

[0067] Attention score Normalize and use the softmax function to get the attention weight ;node At time step The context vector , which represents the node At time step The various features and information associated with the edge are weighted aggregations of information from other nodes at different time steps.

[0068] The normalized attention weights have probabilistic meaning and can clearly indicate the importance of each node and time step to the current node. The context vector effectively integrates relevant information in the spatiotemporal graph and provides a rich source of information for updating node features.

[0069] The spatiotemporal graph neural network analysis unit 2 includes a reinforcement transfer learning module 22. The specific process of the reinforcement transfer learning strategy in the reinforcement transfer learning module 22 is as follows:

[0070] Collect historical warehouse data and real-time warehouse data, clean them, remove obviously erroneous or abnormal data records, perform feature engineering on the processed historical data, extract the frequency of occurrence during the task peak as the key feature, combine the extracted key features with other related features, construct historical feature data sets and real-time feature data sets.

[0071] The quality of real-time data and its consistency with historical data are ensured, so that the model can better combine real-time information for prediction, enhancing the real-time and adaptability of the model. The spatiotemporal graph neural network model based on the spatiotemporal attention mechanism can effectively capture the spatiotemporal characteristics in the warehousing system. The source task is set to allow the model to first learn the laws and patterns in historical data, providing a knowledge basis for subsequent transfer learning.

[0072] A spatiotemporal graph neural network model based on the spatiotemporal attention mechanism is constructed, the initial representation of nodes and edges is determined, and training is performed on the source task. The source task is set to predict the task requirements and equipment status changes in subsequent time periods in history based on the historical feature data set. The parameters of the model are updated through optimization algorithms such as stochastic gradient descent. The spatiotemporal graph neural network model can make full use of the spatiotemporal structure information of the warehousing system. The training of the source task enables the model to master the inherent laws of historical data, providing an effective starting point for knowledge migration to the target task.

[0073] Migrate some parameters in the source model to the preset target model. In the initial training stage of the target model, freeze the migrated parameters to prevent over-adjustment in the early stage of training. At the same time, adjust the architecture of the target model according to the characteristics of real-time data and the requirements of the target task.

[0074] The target task is to accurately predict the task requirements and equipment status in the future period. The introduction of the reinforcement learning framework can give rewards or penalties according to the prediction effect of the model, encourage the model to continuously optimize the strategy, and use the proximal strategy optimization algorithm to update the target model parameters. It can train the model more stably and efficiently, concatenate the context vector with the node's own feature vector, and then update the node feature vector through nonlinear transformation. The concatenation operation effectively integrates the node's own information and context information. The nonlinear transformation increases the complexity and expression ability of the model, and can learn more complex spatiotemporal features.

[0075] The alliance management scheduling decision unit 3 combines the prediction results and the current task with the dynamic alliance formation algorithm of multi-objective optimization, calculates the multi-dimensional matching degree of each device and the task, selects the device with the highest matching degree to form a dynamic alliance and generates scheduling instructions.

[0076] The specific steps of the dynamic alliance formation algorithm of multi-objective optimization in the alliance management scheduling decision unit 3 are:

[0077] Comprehensive and accurate data is the basis for the formation of dynamic alliances. The spatiotemporal characteristics of the warehousing environment can reflect the real-time status and potential laws of warehousing. Clarifying the task set and equipment set can define the scope of the problem and provide a basis for subsequent calculations and decisions. The spatiotemporal characteristics of the warehousing environment are obtained from the spatiotemporal graph neural network analysis unit 2, and the current set of tasks to be executed and the available equipment set are clarified, ensuring the integrity and accuracy of the data used to form a dynamic alliance, so that subsequent calculations and decisions are more in line with actual warehousing conditions.

[0078] Defining multi-objective functions ,in Indicates the shortest task completion time. Indicates that the device has the lowest energy consumption. Indicates the highest equipment utilization rate, Indicates that the task cost is minimized and a weight is assigned to each objective function ,Warehouse management needs to consider multiple factors comprehensively. ,The multi-objective function can comprehensively measure the matching degree between ,equipment and tasks, and the allocation weights can weigh different goals ,according to the actual business focus, making the alliance formation more in line with ,actual needs.

[0079] By quantifying the matching scores between the device and the task under each objective function, we can clearly understand the adaptation of the device to the task in all aspects, providing a basis for comprehensive evaluation. and tasks , calculate the score under each objective function respectively , converting abstract goals into specific values ​​to facilitate subsequent calculations and comparisons, making the evaluation more objective and accurate.

[0080] Computing equipment and tasks Multidimensional matching ; The multidimensional matching degree comprehensively considers multiple objective functions and reflects the comprehensive adaptability of the device and the task. The score of a single objective cannot fully reflect the matching situation between the device and the task. The multidimensional matching degree integrates the scores of multiple objective functions and can more accurately measure the overall adaptability of the device and the task.

[0081] Coding is to convert the equipment-task allocation scheme into a form suitable for genetic algorithm processing. Integer coding is intuitive and concise, and can easily express the correspondence between tasks and equipment, which is convenient for subsequent genetic operations. Devices and The task allocation problem is to encode each individual as a sequence of integers, the first in the sequence The element represents the The device number assigned to each task is encoded in a simple and easy-to-understand way, easy to implement and operate, and can effectively represent the device-task allocation scheme and improve the efficiency of the algorithm.

[0082] Randomly generating the initial population can ensure the extensiveness of the search space, while using the heuristic strategy to generate some individuals can make the initial population contain some high-quality individuals, thus accelerating the convergence of the algorithm.

[0083] Randomly generate the initial population , the population size is , each individual represents a device-task allocation scheme. At the same time, a heuristic strategy is introduced to generate some individuals according to the multi-dimensional matching degree. , for each task , select the one with the highest matching degree devices as candidate devices, and then randomly select one device from these candidate devices and assign it to the task. Repeat this process to generate multiple individuals to join the initial population. It combines the advantages of random search and heuristic search, which not only ensures the comprehensiveness of the search, but also improves the convergence efficiency of the algorithm and reduces unnecessary searches.

[0084] The fitness function is the standard for genetic algorithms to evaluate the quality of individuals. By calculating the sum of multi-dimensional matching degrees, it can comprehensively measure the quality of each device-task allocation plan and provide a basis for selection, crossover and mutation operations.

[0085] For each individual in the population , calculate its fitness , the fitness function is the sum of the multi-dimensional matching degrees of all tasks and devices under the device-task allocation scheme corresponding to the individual, that is, ;in Is an individual Middle The device number assigned to each task is randomly selected from the population in each generation. Individuals form a tournament group, compare the fitness values ​​of individuals in the group, select the individual with the highest fitness as the parent individual, and repeat this process until a preset number of parent individuals are selected for crossover and mutation operations. The selection process has a certain degree of competitiveness and can screen out individuals with higher fitness, thus ensuring the quality and evolutionary direction of the population.

[0086] Traditional crossover operations may destroy the association between tasks. Crossover operations based on task association can preserve the logical relationship between tasks, making the generated child individuals more in line with the actual warehousing business needs. and , analyze the association relationship between its tasks, and then randomly select a group of associated tasks as the intersection point. The device allocation corresponding to the associated task group is copied to the child individual At the same time, The device allocation information of other tasks except the associated task group is copied to In the formation of a progeny individual, The device allocation corresponding to the associated task group is copied to the child individual At the same time, The device allocation information of other tasks except the associated task group is copied to In the next generation, another individual is generated , the correlation between tasks is taken into account, unreasonable allocation schemes are avoided, and the quality and feasibility of offspring individuals are improved.

[0087] The adaptive mutation rate can dynamically adjust the probability of mutation according to the fitness of the individual. For individuals with higher fitness, the mutation rate is reduced to prevent the destruction of their excellent characteristics. For individuals with lower fitness, the mutation rate is increased to increase the possibility of searching for new solutions.

[0088] Assume that the average fitness of the current population is , the fitness of the optimal individual is , then the individual The fitness of , mutation rate Calculated according to the following formula: ;in and is a preset constant. During the mutation operation, a task is randomly selected and its assigned device is replaced with another randomly selected device, making the mutation operation more flexible and intelligent, which can not only ensure the diversity of the population but also avoid the inefficient search caused by excessive mutation.

[0089] In addition to reaching the maximum number of iterations as the termination condition, the fitness convergence judgment is introduced to calculate the fitness value of the best individual in each generation of the population. When the fitness value of the optimal individual in each generation changes by less than a threshold, the algorithm is considered to have converged and the iterative process is terminated, which improves the efficiency of the algorithm, avoids unnecessary calculations, and ensures the quality of the solution found by the algorithm.

[0090] When the genetic algorithm terminates, the individual with the highest fitness is selected as the optimal equipment-task allocation scheme. According to the scheme, a corresponding scheduling instruction is generated for each task, and the instruction content includes information such as the assigned equipment number, task start time, and task execution order. These scheduling instructions are sent to the equipment execution feedback unit 4 so that the equipment can execute the corresponding tasks, thereby improving the efficiency and accuracy of warehouse management and achieving the ultimate goal of dynamic alliance formation.

[0091] The device execution feedback unit 4 is used to receive the scheduling instructions generated by the alliance management scheduling decision unit 3, and autonomously adjust the execution strategy in combination with its own sensor data and environmental perception information.

[0092] The device execution feedback unit 4 uses a fuzzy control algorithm when autonomously adjusting the execution strategy. The specific steps are as follows:

[0093] The sensor data and environmental perception information of the equipment are usually precise numerical values, but in actual decision-making, these precise values ​​are difficult to be used directly to formulate flexible execution strategies. Through fuzzy processing, the precise input data is converted into fuzzy subsets, which can better simulate the human way of thinking and make the control strategy more flexible and adaptable.

[0094] The sensor data and environmental perception information of the equipment are used as input variables. Taking the remaining power as an example, different fuzzy subsets are divided, such as "high", "medium" and "low" corresponding to different power ranges, and the membership function of each subset is defined. Fuzzy processing enables precise input data to be represented by fuzzy concepts, which is more in line with human cognition and decision-making habits, and increases the flexibility and robustness of the control strategy.

[0095] Fuzzy rules are the core of fuzzy control algorithms. Based on historical experience, they link the fuzzy subset combination of input variables with the actions that the equipment should take to provide a basis for decision-making. Fuzzy reasoning is performed based on the membership values ​​of the input variables and fuzzy rules. For each rule, the membership of the premise conditions is calculated, and the minimum value is taken as the excitation intensity of the rule. Then, the fuzzy subset of the conclusion is truncated according to the excitation intensity to obtain the output fuzzy subset of each rule. Fuzzy rules can make full use of historical experience and expert knowledge, simplify complex decision-making processes into a series of condition-action rules, and make control strategies easy to understand and adjust.

[0096] Through fuzzy reasoning, according to the membership value of the input variables and the fuzzy rules, the excitation intensity of each rule is calculated, and the corresponding output fuzzy subset is obtained, so as to comprehensively consider the influence of multiple rules and obtain a reasonable decision result. According to the membership value of the input variables and the fuzzy rules, fuzzy reasoning is performed. For each rule, the membership of the premise condition is calculated, that is, for the rule "IF E is "high" AND is "suitable" THEN v is "high speed", then the membership degree of the antecedent condition is , Indicates the remaining power. Expressed as ambient temperature, Expressed as the running speed, It is expressed as taking the minimum value, and taking the minimum value as the excitation intensity of the rule. Then, the fuzzy subset of the conclusion is truncated according to the excitation intensity to obtain the output fuzzy subset of each rule. Fuzzy reasoning can comprehensively consider the influence of multiple fuzzy rules, avoid the limitations of single rule decision-making, and make the decision result more comprehensive and reasonable.

[0097] The output of fuzzy reasoning is a fuzzy subset, which cannot be directly used to control the actual action of the device. Defuzzification converts the fuzzy subset into a specific control quantity, so that the device can adjust the execution strategy according to this control quantity. The multiple output fuzzy subsets obtained by fuzzy reasoning are merged and defuzzified using the centroid method. The membership function after the merger is set as , then the specific control quantity The calculation formula is The device adjusts its execution strategy according to this control quantity. The centroid defuzzification processing can comprehensively consider all the information of the fuzzy subset to obtain a representative specific control quantity, making the control strategy more accurate and effective, and improving the device's adaptability and operating efficiency.

[0098] In the present invention, the data acquisition and processing unit 1 collects and preprocesses warehouse data, the spatiotemporal graph neural network analysis unit 2 constructs a spatiotemporal graph to predict task requirements and equipment status, the alliance management scheduling decision unit 3 forms a dynamic alliance and generates scheduling instructions, and the equipment execution feedback unit 4 receives instructions and adjusts strategies independently, thereby solving the problems of unreasonable equipment scheduling, poor coordination and resource waste in traditional cross-border e-commerce warehouse management, realizing intelligent scheduling and collaborative operation of warehouse equipment, and improving warehouse management efficiency and resource utilization.

[0099] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A deep learning cross-border e-commerce warehouse management system, characterized in that: It includes a data acquisition and processing unit (1), a spatiotemporal graph neural network analysis unit (2), an alliance management and scheduling decision unit (3), and a device execution feedback unit (4); The data acquisition and processing unit (1) adopts a distributed sensor network based on the Internet of Things, uses position sensors, weight sensors, status monitoring sensors and environmental sensors to collect the position, operating status, cargo information and storage environment parameters of cross-border e-commerce storage equipment in real time, and uses an adaptive data cleaning algorithm to pre-process the collected data to obtain a storage data set; The spatiotemporal graph neural network analysis unit (2) inputs the storage data into the spatiotemporal graph neural network in time series, thereby constructing a spatiotemporal graph containing nodes and edges, wherein the nodes represent storage equipment, cargo storage areas and sorting areas, and the edges represent the connection relationship between them, and updates the graph structure at each time step to reflect the dynamic changes of the storage environment. By introducing a spatiotemporal attention mechanism, the edge association weights of different nodes and edges in different time and space are automatically calculated to obtain the spatiotemporal characteristics of the storage environment, and the spatiotemporal graph neural network is trained by combining historical storage data and real-time storage data using a reinforcement transfer learning strategy to predict task requirements and equipment status in future time periods; The alliance management scheduling decision unit (3) combines the prediction results and the current task with the dynamic alliance formation algorithm of multi-objective optimization, calculates the multi-dimensional matching degree between each device and the task, selects the device with the highest matching degree to form a dynamic alliance and generates a scheduling instruction; The device execution feedback unit (4) is used to receive the dispatch instruction generated by the alliance management dispatch decision unit (3), and autonomously adjust the execution strategy in combination with its own sensor data and environmental perception information.

2. According to claim 1, a deep learning cross-border e-commerce warehouse management system is characterized by: In the data acquisition and processing unit (1), the specific operation of the adaptive data cleaning algorithm is as follows: Define a set of data quality assessment indicators, including data integrity, data accuracy, data consistency, and data timeliness, and assign an initial weight to each indicator; For the collected raw data, the quality score of each data sample is calculated according to the initial weight, and the quality score is compared with the preset low-quality threshold and high-quality threshold, and the data is divided into low-quality data, medium-quality data and high-quality data according to the comparison results; For low-quality data, data interpolation method is used to repair it. If it cannot be repaired, it is directly deleted. For medium-quality data, data correction is performed, and high-quality data is directly retained to obtain a cleaned warehouse data set.

3. According to claim 2, a deep learning cross-border e-commerce warehouse management system is characterized by: The spatiotemporal graph neural network analysis unit (2) comprises a spatiotemporal attention mechanism module (21). When the spatiotemporal attention mechanism module (21) constructs the spatiotemporal graph neural network, the edge association weight is calculated as follows: For nodes representing storage equipment, cargo storage areas, and sorting areas, calculate the physical distances between them, and calculate the physical distance association weights based on the physical distances between them; The business interaction frequency between nodes in historical data is counted, and the business association weight of the edge is calculated based on the business interaction frequency. The physical distance association weight and the business association weight are then substituted into the formula to calculate the edge association weight.

4. According to claim 3, a deep learning cross-border e-commerce warehouse management system is characterized by: In the spatiotemporal attention mechanism module (21), when constructing a spatiotemporal graph neural network based on the spatiotemporal attention mechanism, the specific calculation method of the spatiotemporal attention mechanism is as follows: For the feature vector of a node in the spatiotemporal graph at a time step, calculate its attention score with other nodes at different time steps; The attention scores are normalized to obtain the attention weights, and the attention weights are substituted into the formula to calculate the context vector of the node at this time step. The context vector is used to represent various features and information related to the node and the edge at this time step. Finally, the context vector is concatenated and nonlinearly transformed with the node’s own feature vector to obtain the updated node feature vector.

5. According to claim 4, a deep learning cross-border e-commerce warehouse management system is characterized by: The spatiotemporal graph neural network analysis unit (2) comprises a reinforcement transfer learning module (22), and the specific process of the reinforcement transfer learning strategy in the reinforcement transfer learning module (22) is as follows: Collect historical warehouse data, clean and interpolate the historical warehouse data, perform feature engineering on the processed historical data, extract the frequency of task peak hours as the key feature, and construct a historical feature data set; Extracting real-time warehouse data from the data acquisition and processing unit (1), immediately cleaning the real-time warehouse data, and then performing feature engineering processing to extract real-time features and construct a real-time feature data set; Build a spatiotemporal graph neural network model based on the spatiotemporal attention mechanism, determine the initial representation of nodes and edges, and train it on the source task. The source task is set to predict the task requirements and equipment status changes in subsequent periods in history based on the historical feature dataset. Migrate the parameters in the source model to the preset target model. The migrated parameters are frozen during the initial training phase of the target model, and the architecture of the target model is adjusted. The target model is adjusted and trained on the target task based on the real-time feature data set. The target task is to accurately predict the task requirements and equipment status in the future period. The reinforcement learning framework is introduced, the reward function is defined, and the proximal strategy optimization algorithm is used to update the target model parameters. In each iteration, the predicted action is generated according to the current model strategy, and the reward is obtained by comparing it with the actual subsequent situation until the reward value reaches the preset threshold, and the target model training is determined to be completed.

6. A deep learning cross-border e-commerce warehouse management system according to claim 5, characterized in that: The specific steps of the multi-objective optimization dynamic alliance formation algorithm in the alliance management scheduling decision unit (3) are: The spatiotemporal characteristics of the storage environment are obtained from the spatiotemporal graph neural network analysis unit (2), and the current set of tasks to be performed and the set of available equipment are determined, wherein the set of tasks and the set of available equipment are provided by the storage data set; A multi-objective function is defined according to the task set and the available equipment set, wherein the multi-objective function is respectively composed of the shortest task completion time, the lowest equipment energy consumption, the highest equipment utilization rate and the lowest task cost, and a weight is assigned to each objective function; For each device and task, its score under each objective function is calculated respectively, and then the multi-dimensional matching degree of the device and the task is calculated according to its score and assigned weight under each objective function. Finally, a genetic algorithm is used to form a dynamic alliance.

7. A deep learning cross-border e-commerce warehouse management system according to claim 6, characterized in that: The specific operation of the alliance management scheduling decision unit (3) using the genetic algorithm to form a dynamic alliance is as follows: Integer encoding is used. Devices and The task allocation problem is to encode each individual as a sequence of integers, the first in the sequence The element represents the The device number to which the task is assigned; Randomly generate the initial population , the population size is , each individual represents a device-task allocation scheme. At the same time, a heuristic strategy is introduced to generate some individuals. According to the multi-dimensional matching degree, for each task, the one with the highest matching degree is selected. Devices are selected as candidate devices, and then a device is randomly selected from these candidate devices and assigned to the task. This process is repeated to generate multiple individuals to join the initial population. For each individual in the population, its fitness is calculated. The fitness function is the sum of the multi-dimensional matching degrees of all tasks and devices under the device-task allocation scheme corresponding to the individual. The tournament selection method is introduced, that is, in each generation, randomly select from the population Individuals form a tournament group, compare the fitness values ​​of individuals in the group, select the individual with the highest fitness as the parent individual, and repeat this process until a preset number of parent individuals are selected for crossover and mutation operations; For two parent individuals and , analyze the association relationship between its tasks, and then randomly select a group of associated tasks as the intersection point. The device allocation corresponding to the associated task group is copied to the child individual At the same time, The device allocation information of other tasks except the associated task group is copied to In the formation of a progeny individual, The device allocation corresponding to the associated task group is copied to the child individual At the same time, The device allocation information of other tasks except the associated task group is copied to In the next generation, another individual is generated ; Introducing adaptive mutation rate, during mutation operation, randomly select a task and replace its assigned device with another randomly selected device. In addition to reaching the maximum number of iterations as the termination condition, introducing fitness convergence judgment, calculating the fitness value of the best individual in each generation of population, when continuous When the fitness value of the optimal individual of the generation changes less than a threshold, the algorithm is considered to have converged and the iteration process is terminated; When the genetic algorithm terminates, the individual with the highest fitness is selected as the optimal equipment-task allocation scheme, and the corresponding scheduling instructions are generated for each task based on the scheme.

8. A deep learning cross-border e-commerce warehouse management system according to claim 7, characterized in that: The device execution feedback unit (4) adopts a fuzzy control algorithm when autonomously adjusting the execution strategy. The specific steps are as follows: The sensor data and environmental perception information of the device are used as input variables, and each input variable is fuzzified to divide different fuzzy subsets corresponding to different power intervals, and the membership function of each subset is defined; Fuzzy rules are formulated based on historical experience. Each rule has corresponding premises and conclusions. The premises are fuzzy subset combinations of input variables, and the conclusions are fuzzy subsets of actions that the device should take. According to the membership value of the input variable and the fuzzy rules, fuzzy reasoning is performed. For each rule, the membership of the premise condition is calculated, and the minimum value is taken as the excitation strength of the rule. Then, the fuzzy subset of the conclusion is truncated according to the excitation strength to obtain the output fuzzy subset of each rule. The multiple output fuzzy subsets obtained by fuzzy reasoning are merged and defuzzified using the centroid method to obtain the specific control quantity, that is, the actual action that the device should take. The device adjusts its execution strategy based on this control quantity.

Images