E-commerce intelligent storage optimization and dynamic scheduling method based on deep learning

Abstract

The invention discloses an e-commerce intelligent storage optimization and dynamic scheduling method based on deep learning, and the method comprises the following steps: S1, collecting data in a storage environment, and carrying out the standardization processing; s2, learning the standardized data, and predicting future order demands; s3, a MoE-Transform model is constructed, and the storage position of the SKU is optimized by using the MoE-Transform model; s4, order tasks are divided, sorting is carried out according to task priorities, and task distribution is carried out; s5, constructing a storage environment topological graph, calculating a path of an execution unit by using the improved GAT, and performing path optimization and dynamic adjustment; and S6, updating parameters of the MoE-Transform model and the improved GAT, optimizing task generation and task allocation, and optimizing inventory layout and storage partition at the same time. According to the method, the MoE-Transform model and the improved GAT are utilized, storage optimization, task scheduling and path planning of the e-commerce storage SKU are achieved, and the method has the advantages of being intelligent, self-adaptive, efficient in cooperation and dynamic in optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce intelligent warehousing optimization and scheduling, and in particular to an e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning. Background Art

[0002] With the rapid development of the e-commerce industry, the scale and complexity of e-commerce warehousing systems are increasing, requiring warehouses to efficiently handle large amounts of order and inventory data. Traditional e-commerce warehousing systems often rely on manual operations and task scheduling based on fixed rules, which has great limitations when facing dynamic factors such as order fluctuations, inventory management, and equipment status changes. Traditional warehousing scheduling systems usually use rule-based path planning and task allocation methods, which cannot flexibly respond to real-time changes in the warehousing environment, resulting in inefficiency and waste of resources.

[0003] In traditional e-commerce warehouse management, the storage location of SKUs is often arranged based on historical experience, and once set, it will not be easily changed. This fixed storage layout often leads to unreasonable storage locations for high-frequency goods when faced with demand fluctuations and order changes, increasing the time and cost of picking. In addition, traditional methods usually do not take into account the correlation between inventory and cannot intelligently predict which goods should be stored closer and which goods should be distributed in different areas to improve overall picking efficiency.

[0004] Task scheduling is another key issue in warehouse management. Traditional task scheduling methods usually rely on static rules or manual experience to determine the execution order of picking tasks, which is prone to resource conflicts and inefficiency when multiple tasks occur simultaneously. For example, multiple automated picking robots, robotic arms, and manual pickers may perform tasks in the warehouse at the same time. Due to the lack of a dynamic scheduling mechanism, they may have path conflicts, resulting in delays in task execution and waste of warehouse resources. Unreasonable task scheduling not only affects warehouse operation efficiency, but also increases equipment idle time and energy consumption.

[0005] Existing warehouse path planning methods also have certain shortcomings. Traditional path planning algorithms, such as Dijkstra and A* algorithms, are mainly based on static network models to calculate the optimal path, which is suitable for scenarios with relatively fixed equipment and environmental conditions. However, in a dynamic and complex e-commerce warehousing environment, path planning needs to take into account multiple factors, such as real-time order demand, warehouse environmental conditions, task urgency, and real-time equipment status. Existing path planning algorithms lack the ability to respond to these dynamic factors in real time, which can easily lead to path conflicts among equipment in the warehouse, affecting overall operational efficiency.

[0006] In recent years, with the development of deep learning and reinforcement learning technologies, the application of artificial intelligence in the field of e-commerce warehousing has gradually increased. Deep learning technology has been widely used in image recognition, natural language processing and other fields by automatically extracting features and learning the underlying laws in the data. However, in e-commerce warehousing systems, traditional deep learning methods mainly rely on historical data for static predictions and lack real-time adaptability to changes in the warehousing environment. For example, factors such as inventory status, order demand, and equipment location have obvious dynamic changes in the time dimension, and existing deep learning models often ignore the temporal relationship of these factors, resulting in limited accuracy and practicality of model predictions.

[0007] In addition, task scheduling and path planning based on deep learning also face problems such as huge data volume and high computational complexity in practical applications. In the e-commerce warehousing environment, the amount of information involved in tasks, equipment, paths, etc. is huge, and traditional deep learning models often find it difficult to efficiently process these large-scale data. While emerging technologies such as graph neural networks and graph attention networks have strong advantages in processing graph structured data, the application of these technologies in the e-commerce warehousing environment is still in the exploratory stage.

[0008] Therefore, the existing technology has the following defects: First, the warehouse layout optimization method lacks intelligence and cannot be dynamically adjusted according to real-time order demand and inventory status; second, the existing task scheduling method relies too much on static rules and cannot adapt to the real-time changing warehouse environment, resulting in resource conflicts and low efficiency; third, the path planning method cannot fully consider dynamic factors and cannot optimize the path in real time, which is prone to path conflicts and affects warehouse efficiency. Therefore, how to provide an e-commerce intelligent warehouse optimization and dynamic scheduling method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One object of the present invention is to propose an e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning. The present invention dynamically optimizes SKU storage location, task scheduling and path planning through deep learning technology, improves the overall efficiency and flexibility of the warehousing system, and realizes dynamic adjustment of the warehousing environment and optimization of path conflicts by combining MoE-Transformer and improved GAT, thereby improving the operational efficiency of the e-commerce warehousing system.

[0010] According to an embodiment of the present invention, a method for e-commerce intelligent warehousing optimization and dynamic scheduling based on deep learning includes the following steps:

[0011] S1. Collect inventory status data, historical order data, equipment location information, task execution records and warehouse environment variables in the warehouse environment, and standardize the collected data to generate standardized data;

[0012] S2. Use the Transformer model to learn standardized data, predict future order demand, and adjust the replenishment strategy based on inventory status data;

[0013] S3. Build the MoE-Transformer model and use it to optimize the SKU storage location and adjust the storage location of the goods according to future order requirements.

[0014] S4. Based on future order demand, optimized inventory status and SKU storage location, order tasks are divided, sorted according to task priority, and task allocation is performed;

[0015] S5. Based on the task allocation results, a storage environment topology map is constructed, the paths of the execution units are calculated using the improved GAT, and a path optimization objective function is constructed. Path optimization and dynamic adjustment are performed according to the path optimization objective function.

[0016] S6. Based on the long-term storage environment operation data, update the parameters of the MoE-Transformer model and the improved GAT, optimize task generation and task allocation, and optimize inventory layout and storage partitioning.

[0017] Optionally, the data includes inventory status data, historical order data, equipment location information, task execution records and warehouse environment variables.

[0018] Optionally, the S3 specifically includes:

[0019] S31, constructing a SKU storage optimization input data set, wherein the SKU storage optimization input data set includes an order demand prediction matrix, a SKU access frequency matrix, and storage area information;

[0020] S32. Calculate the storage area status matrix and the SKU storage optimization feature matrix:

[0021] S={s 1 ,s 2 ,...,s n};

[0022]

[0023] Among them, S represents the storage area state matrix, X t represents the SKU storage optimization feature matrix, s n Indicates the storage location of SKUn in the current warehouse, f k,t represents the SKU access frequency matrix, represents the order demand forecast matrix, Indicates storage area z jAvailable storage capacity, Indicates the current location of the SKU k To storage area z j The shortest path distance, z j Indicates storage area information;

[0024] S33, constructing a MoE-Transformer model, wherein the MoE-Transformer model includes a MoE layer, a gating network and a Transformer network, wherein the MoE layer is composed of a plurality of expert networks, and the gating network is used to select the most suitable expert network;

[0025] S34. Input the SKU storage optimization feature matrix into the MoE-Transformer model to obtain the output H Trans ;

[0026] S35. Predict SKU storage location, set optimization goals, and use Adam for model training:

[0027]

[0028]

[0029] in, represents the optimization goal, represents the probability distribution of SKU in each storage area, W s and b s represents the trainable parameters, y k,j Indicates the actual storage area. represents the predicted value, C j represents the storage capacity of the storage area, λ represents the regularization parameter, θ represents all trainable parameters of the model, α represents the learning rate, represents the gradient of the optimization target, m and n represent the number and location of the storage area respectively, and softmax represents the activation function;

[0030] S36. Determine the final SKU storage location:

[0031]

[0032] Among them, S t+1 Indicates the final SKU storage location.

[0033] Optionally, the S34 specifically includes:

[0034] S341. Input the SKU storage optimization feature matrix into the expert network to obtain a single expert network output:

[0035] M e (Xt )=ReLU(W e1 X t +b e1 );

[0036] H e =ReLU(W e2 M e +b e2 );

[0037] Among them, H e represents the output of the expert network, M e represents the expert network, X t represents the SKU storage optimization feature matrix, W e1 , b e1 , W e2 and b e2 represents trainable parameters;

[0038] S342. Output the weight of each expert network through the gating network:

[0039] G(X t )=softmax(W g X t +b g );

[0040] Among them, G(X t ) represents the weight of each expert network, W g and b g represents the gating network parameters;

[0041] S343. Calculate the output of the MoE layer:

[0042]

[0043] Among them, H MoE Represents the output of the MoE layer;

[0044] S344. Use the multi-head attention mechanism to calculate the output of the Transformer network:

[0045] Q=W q H MoE ,K=W k H MoE ,V=W v H MoE ;

[0046]

[0047] H Trans =LN(ReLU(W t Z+b t ));

[0048] Among them, Z represents the output after the multi-head attention mechanism is calculated, LN represents layer normalization, Q, K and V represent the vectors of query, key and value, and W q , W k and W v Represents the weight matrix of query, key and value, d k represents the dimension of the query vector Q, H Trans represents the final output of the Transformer network, LN represents layer normalization, and W t and b t Represents the Transformer network parameters.

[0049] Optionally, the execution unit includes an AGV, a robotic arm and a manual picker.

[0050] Optionally, the S5 specifically includes:

[0051] S51, construct a warehouse environment topology graph G = (V, E), and define a path weight matrix, where V in the warehouse environment topology graph represents a set of key nodes in the warehouse, and E in the warehouse environment topology graph represents an edge set of traversable paths within the warehouse:

[0052] W i,j =d i,j +βC i,j +γL i,j ;

[0053] Among them, W i,j represents the path weight matrix, which is the cost from path i to path j, d i,j represents the Euclidean distance between path i and path j, C i,j Indicates the channel congestion degree between path i and path j, L i,j represents the historical path usage frequency from path i to path j, β and γ represent the adjustment weights of path congestion and historical usage rate;

[0054] S52. Generate a task path request for the execution unit according to the SKU storage location and the execution unit status:

[0055]

[0056] Among them, P R represents the path request set, e p,t represents the current position of execution unit p at time t, Indicates that SKU k is in storage area z j The storage coordinates of It means "arbitrary";

[0057] S53, using the improved GAT to calculate the path features and generate a node feature matrix;

[0058] S54. Calculate the optimal path score:

[0059]

[0060] Among them, P score (i,j) represents the optimal path score, W (l) represents the feature weights calculated by the improved GAT at layer l, represents the updated representation of node i at layer l, T i,j represents the time weight from path i to path j, represents the frequency of use of path i to path j in historical tasks, ω 1 ,ω 2 and ω 3 represents the trainable path scoring weight parameter, L represents the number of layers of the improved GAT, and C i,j represents the channel congestion degree between path i and path j, e i,j represents edge features, W 2 represents a trainable weight matrix;

[0061] S55. Calculate the optimal path according to the optimal path score:

[0062] P optimal =argmin P ∑ (i,j)∈P P score (i,j);

[0063] Among them, P optimal represents the path with the minimum cost in the path set P, argmin represents the variable value that makes the function achieve the minimum value, and P represents the path set;

[0064] S56. Construct a path optimization objective function and dynamically adjust the path according to the path optimization objective function:

[0065]

[0066] P' optimal =P optimal ∪{(i′,j′)|C i',j' <θ};

[0067] in, represents the path optimization objective function, represents the path conflict cost of execution unit p at time t, λ 1 represents the path conflict adjustment coefficient, P' optimal represents the set of optimized paths, θ represents the path switching cost threshold, Ci',j' represents the channel congestion degree between path i' and path j', {} represents a set, ∪ represents the "and" operation, W i,j represents the path weight matrix;

[0068] S57. Output the final optimized path set:

[0069] P * =P' optimal ;

[0070] Among them, P * Represents the final optimized path set.

[0071] Optionally, the S53 specifically includes:

[0072] S531. Construct the input of graph neural network:

[0073]

[0074] Among them, H (0) represents the initial node feature matrix, represents the initial features of node i, V represents the key node set in the warehouse, include:

[0075] Node Type i :Storage area t i =1, picking station t i =2, conveyor belt interface t i =3, Shipping port t i =4;

[0076] Node coordinate x i ,y i : The absolute position of the node in the warehouse layout;

[0077] Channel width w i : The width of the path where the node is located;

[0078] Picking point flow rate f i : The historical frequency of picking tasks at the node location;

[0079] Historical task completion time τ i : reflects the average time required to execute the task at this point in the past;

[0080] S532, constructing an edge feature matrix;

[0081] E i,j ={e i,j |(i,j)∈E};

[0082] Among them, E i,jrepresents the edge feature matrix, which is the feature of path i to path j. E represents the edge set of the traversable path inside the warehouse. i,j include:

[0083] Physical distance i,j : The Euclidean distance between path i and path j;

[0084] Path channel congestion weight C i,j : The channel congestion degree between path i and path j;

[0085] Path usage frequency L i,j : The historical path usage frequency from path i to path j;

[0086] Path accessibility i,j :If the path is passable, then a i,j =1, otherwise a i,j =0;

[0087] S533, initialize node feature matrix embedding:

[0088]

[0089] Among them, x i represents the original features of node i, represents the initial embedding matrix, represents the bias term;

[0090] S534, using multi-head attention mechanism to calculate the information interaction between nodes:

[0091]

[0092] in, represents the attention weight of the h-th attention head between path i and path j, W h represents the trainable parameter matrix of the h-th attention head, represents the features of node j of the lth attention head, represents the feature of node k of the lth attention head, a h represents the parameter vector of attention calculation, T represents the transposition operation, || represents the vector concatenation operation, N(i) represents the set of neighbor nodes of node i, LeakyReLU represents the activation function, and exp represents the natural exponential function;

[0093] S535, update the node representation, and obtain the node feature matrix H (L) :

[0094]

[0095] in, represents the updated representation of node i after calculation at layer l+1, H represents the number of attention heads, σ represents the nonlinear activation function, and H (L) Represents the node feature matrix calculated by the final L layers.

[0096] The beneficial effects of the present invention are:

[0097] First, in terms of task scheduling, the present invention uses a deep learning model to intelligently divide order tasks, and sort and allocate them based on task priority. Compared with the traditional fixed rule scheduling method, the present invention can optimize tasks according to the urgency of the order, inventory distribution, and the status of the execution unit, effectively avoiding task accumulation and resource waste, and improving the overall throughput of the warehousing system. In addition, the present invention introduces an improved GAT for path optimization. Compared with the static path planning method, the present invention can dynamically calculate the optimal picking path of the execution unit, calculate the path weight in combination with the warehouse environment topology map, and adaptively adjust the path selection through the attention mechanism to reduce path conflicts and channel congestion, and improve the efficiency and stability of task execution.

[0098] Secondly, in terms of path optimization, the present invention breaks through the limitation of traditional path planning algorithms that only consider static weights, combines multi-dimensional data such as historical path utilization, channel congestion, and equipment operating status, dynamically adjusts the path weight matrix, and calculates the optimal path score based on the improved GAT. This method introduces a multi-head attention mechanism in the path search process, so that the path selection can more accurately adapt to the real-time warehousing environment. In addition, the present invention realizes dynamic adjustment of path conflicts by constructing a path optimization objective function, ensuring that multiple execution units will not experience path congestion or resource competition during task execution, thereby improving the overall operation efficiency of the warehousing system.

[0099] Finally, the present invention also optimizes the MoE-Transformer and the improved GAT model through long-term warehouse environment operation data, so that the system can continuously learn and optimize task generation, task allocation and path planning strategies. Compared with the existing warehouse management system, the present invention provides a more adaptive and intelligent method, which can dynamically adjust the warehouse layout, task scheduling and path optimization strategies when the warehouse environment is constantly changing, thereby improving the degree of automation and intelligence of the warehouse system. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0101] Figure 1 This is a flow chart of an e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning proposed by the present invention;

[0102] Figure 2 A schematic diagram of the process of path optimization based on improved GAT for an e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0103] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0104] refer to Figure 1 and Figure 2 , an e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning, comprising the following steps:

[0105] S1. Collect inventory status data, historical order data, equipment location information, task execution records and warehouse environment variables in the warehouse environment, and standardize the collected data to generate standardized data;

[0106] S2. Use the Transformer model to learn standardized data, predict future order demand, and adjust the replenishment strategy based on inventory status data;

[0107] S3. Build the MoE-Transformer model and use it to optimize the SKU storage location and adjust the storage location of the goods according to future order requirements.

[0108] S4. Based on future order demand, optimized inventory status and SKU storage location, order tasks are divided, sorted according to task priority, and task allocation is performed;

[0109] S5. Based on the task allocation results, a storage environment topology map is constructed, the paths of the execution units are calculated using the improved GAT, and a path optimization objective function is constructed. Path optimization and dynamic adjustment are performed according to the path optimization objective function.

[0110] S6. Based on the long-term storage environment operation data, update the parameters of the MoE-Transformer model and the improved GAT, optimize task generation and task allocation, and optimize inventory layout and storage partitioning.

[0111] In this embodiment, the data includes inventory status data, historical order data, equipment location information, task execution records and warehouse environment variables.

[0112] In this implementation, S3 specifically includes:

[0113] S31, constructing a SKU storage optimization input data set, wherein the SKU storage optimization input data set includes an order demand prediction matrix, a SKU access frequency matrix, and storage area information;

[0114] S32. Calculate the storage area status matrix and the SKU storage optimization feature matrix:

[0115] S={s 1 ,s 2 ,...,s n};

[0116]

[0117] Among them, S represents the storage area state matrix, X t represents the SKU storage optimization feature matrix, s n Indicates the storage location of SKUn in the current warehouse, f k,t represents the SKU access frequency matrix, represents the order demand forecast matrix, Indicates storage area z j Available storage capacity, Indicates the current location of the SKU k To storage area z j The shortest path distance, z j Indicates storage area information;

[0118] S33, constructing a MoE-Transformer model, wherein the MoE-Transformer model includes a MoE layer, a gating network and a Transformer network, wherein the MoE layer is composed of a plurality of expert networks, and the gating network is used to select the most suitable expert network;

[0119] S34. Input the SKU storage optimization feature matrix into the MoE-Transformer model to obtain the output H Trans ;

[0120] S35. Predict SKU storage location, set optimization goals, and use Adam for model training:

[0121]

[0122] in, represents the optimization goal, represents the probability distribution of SKU in each storage area, W s and b s represents the trainable parameters, y k,j Indicates the actual storage area. represents the predicted value, C jrepresents the storage capacity of the storage area, λ represents the regularization parameter, θ represents all trainable parameters of the model, α represents the learning rate, represents the gradient of the optimization target, m and n represent the number and location of the storage area respectively, and softmax represents the activation function;

[0123] S36. Determine the final SKU storage location:

[0124]

[0125] Among them, S t+1 Indicates the final SKU storage location.

[0126] In this implementation manner, the S34 specifically includes:

[0127] S341. Input the SKU storage optimization feature matrix into the expert network to obtain a single expert network output:

[0128] M e (X t )=ReLU(W e1 X t +b e1 );

[0129] H e =ReLU(W e2 M e +b e2 );

[0130] Among them, H e represents the output of the expert network, M e represents the expert network, X t represents the SKU storage optimization feature matrix, W e1 , b e1 , W e2 and b e2 represents trainable parameters;

[0131] S342. Output the weight of each expert network through the gating network:

[0132] G(X t )=softmax(W g X t +b g );

[0133] Among them, G(X t ) represents the weight of each expert network, W g and b g represents the gating network parameters, and softmax represents the activation function;

[0134] S343. Calculate the output of the MoE layer:

[0135]

[0136] Among them, H MoE Represents the output of the MoE layer;

[0137] S344. Use the multi-head attention mechanism to calculate the output of the Transformer network:

[0138] Q=W q H MoE ,K=W k H MoE ,V=W v H MoE ;

[0139]

[0140] H Trans =LN(ReLU(W t Z+b t ));

[0141] Among them, Z represents the output after the multi-head attention mechanism is calculated, LN represents layer normalization, Q, K and V represent the vectors of query, key and value, and W q , W k and W v Represents the weight matrix of query, key and value, d k represents the dimension of the query vector Q, H Trans represents the final output of the Transformer network, LN represents layer normalization, and W t and b t Represents the Transformer network parameters.

[0142] In this embodiment, the execution unit includes an AGV, a robotic arm, and a manual picker.

[0143] In this implementation manner, S5 specifically includes:

[0144] S51, construct a warehouse environment topology graph G = (V, E), and define a path weight matrix, where V in the warehouse environment topology graph represents a set of key nodes in the warehouse, and E in the warehouse environment topology graph represents an edge set of traversable paths within the warehouse:

[0145] W i,j =d i,j +βC i,j +γL i,j ;

[0146] Among them, W i,j represents the path weight matrix, which is the cost from path i to path j, di,j represents the Euclidean distance between path i and path j, C i,j Indicates the channel congestion degree between path i and path j, L i,j represents the historical path usage frequency from path i to path j, β and γ represent the adjustment weights of path congestion and historical usage rate;

[0147] S52. Generate a task path request for the execution unit according to the SKU storage location and the execution unit status:

[0148]

[0149] Among them, P R represents the path request set, e p,t represents the current position of execution unit p at time t, Indicates that SKU k is in storage area z j The storage coordinates of It means "arbitrary";

[0150] S53, using the improved GAT to calculate the path features and generate a node feature matrix;

[0151] S54. Calculate the optimal path score:

[0152]

[0153] Among them, P score (i,j) represents the optimal path score, W (l) represents the feature weights calculated by the improved GAT at layer l, represents the updated representation of node i at layer l, T i,j represents the time weight from path i to path j, represents the frequency of use of path i to path j in historical tasks, ω 1 ,ω 2 and ω 3 represents the trainable path scoring weight parameter, L represents the number of layers of the improved GAT, and C i,j represents the channel congestion degree between path i and path j, e i,j represents edge features, W 2 represents a trainable weight matrix;

[0154] S55. Calculate the optimal path according to the optimal path score:

[0155] P optimal =argmin P ∑ (i,j)∈P P score (i,j);

[0156] Among them, Poptimal represents the path with the minimum cost in the path set P, argmin represents the variable value that makes the function achieve the minimum value, and P represents the path set;

[0157] S56. Construct a path optimization objective function and dynamically adjust the path according to the path optimization objective function:

[0158]

[0159] P' optimal =P optimal ∪{(i′,j′)|C i',j' <θ};

[0160] in, represents the path optimization objective function, represents the path conflict cost of execution unit p at time t, λ 1 represents the path conflict adjustment coefficient, P' optimal represents the set of optimized paths, θ represents the path switching cost threshold, C i',j' represents the channel congestion degree between path i' and path j', {} represents a set, ∪ represents the "and" operation, W i,j represents the path weight matrix;

[0161] S57. Output the final optimized path set:

[0162] P * =P' optimal ;

[0163] Among them, P * Represents the final optimized path set.

[0164] In this implementation manner, the S53 specifically includes:

[0165] S531. Construct the input of graph neural network:

[0166]

[0167] Among them, H (0) represents the initial node feature matrix, represents the initial features of node i, V represents the key node set in the warehouse, include:

[0168] Node Type i :Storage area t i =1, picking station t i =2, conveyor belt interface t i =3, Shipping port t i =4;

[0169] Node coordinate x i ,y i : The absolute position of the node in the warehouse layout;

[0170] Channel width w i : The width of the path where the node is located;

[0171] Picking point flow rate f i : The historical frequency of picking tasks at the node location;

[0172] Historical task completion time τ i : reflects the average time required to execute the task at this point in the past;

[0173] S532, constructing an edge feature matrix;

[0174] E i,j ={e i,j |(i,j)∈E};

[0175] Among them, E i,j represents the edge feature matrix, which is the feature of path i to path j. E represents the edge set of the traversable path inside the warehouse. i,j include:

[0176] Physical distance i,j : The Euclidean distance between path i and path j;

[0177] Path channel congestion weight C i,j : The channel congestion degree between path i and path j;

[0178] Path usage frequency L i,j : The historical path usage frequency from path i to path j;

[0179] Path accessibility i,j :If the path is passable, then a i,j =1, otherwise a i,j =0;

[0180] S533, initialize node feature matrix embedding:

[0181]

[0182] Among them, x i represents the original features of node i, represents the initial embedding matrix, represents the bias term;

[0183] S534, using multi-head attention mechanism to calculate the information interaction between nodes:

[0184]

[0185] in, represents the attention weight of the h-th attention head between path i and path j, W h represents the trainable parameter matrix of the h-th attention head, represents the features of node j of the lth attention head, represents the feature of node k of the lth attention head, a h represents the parameter vector of attention calculation, T represents the transposition operation, || represents the vector concatenation operation, N(i) represents the set of neighbor nodes of node i, LeakyReLU represents the activation function, and exp represents the natural exponential function;

[0186] S535, update the node representation, and obtain the node feature matrix H (L) :

[0187]

[0188] in, represents the updated representation of node i after calculation at layer l+1, H represents the number of attention heads, σ represents the nonlinear activation function, and H (L) Represents the node feature matrix calculated by the final L layers.

[0189] Embodiment 1:

[0190] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a smart warehouse of an e-commerce company. The warehouse processes more than 30,000 orders per day, stores up to 50,000 SKUs, has a storage area of ​​about 8,000 square meters, and the automated execution unit includes 300 automatic guided vehicles, 100 robotic arms, and several manual pickers. However, in the actual operation process, the warehouse faces problems such as unreasonable SKU storage layout, low task scheduling efficiency, and serious path planning conflicts, which lead to extended order fulfillment time, limited warehouse throughput capacity, and affected overall operational efficiency.

[0191] In this warehousing environment, the present invention performs SKU storage optimization based on MoE-Transformer and combines the improved GAT for path optimization to improve the operational efficiency of the warehouse. First, the present invention uses Transformer to predict future order demand and dynamically optimizes the SKU storage location, so that high-frequency SKUs have a better storage location and reduce the associated storage conflicts between SKUs, thereby shortening the picking path and reducing the walking distance of automatic guided vehicles and manual pickers. Secondly, the present invention combines the improved GAT to calculate the optimal picking path of the execution unit in the warehouse environment, dynamically adjusts the path weight, avoids automatic guided vehicle path conflicts, and improves equipment coordination efficiency. Finally, through long-term data accumulation, the model parameters are continuously optimized to achieve the intelligent upgrade of the warehousing system.

[0192] During the implementation process, the invention was deployed to the core business system of the e-commerce warehouse and tested for 6 months in actual operation. During the test, multiple key indicators were collected, including SKU picking time, order fulfillment time, path conflict rate, equipment utilization rate, etc., and compared with traditional warehouse scheduling methods.

[0193] Table 1 Experimental data comparison table

[0194]

[0195]

[0196] From the perspective of SKU storage optimization, the present invention optimizes the SKU access frequency and storage area through the MoE-Transformer model, reducing the storage areas of high-frequency SKUs from 4.2 to 2.1, reducing the dispersion of SKUs in the warehouse and improving the efficiency of picking. At the same time, the optimized SKU storage strategy reduces the average order picking time from 28.5 minutes to 19.3 minutes, a decrease of 32.3%, effectively shortening the order processing time. In addition, due to the more reasonable SKU storage location, the average length of the picking path is reduced from 22 meters to 13 meters, a reduction of 37.8%, which reduces the driving distance of pickers and automatic guided vehicles in the warehouse and improves order fulfillment efficiency.

[0197] In terms of path planning, the traditional automated guided vehicle path scheduling method has the problem of path conflict, especially during peak hours, the path conflict rate is as high as 12.4%, resulting in frequent blockage of automated guided vehicles in picking tasks, affecting the warehouse throughput capacity. The present invention adopts improved GAT for dynamic path optimization, combined with the multi-head attention mechanism to calculate the optimal path, so that the path conflict rate is reduced to 4.6%, a reduction of 63.1%, and the stability of AGV task scheduling in the warehouse environment is improved. At the same time, the average waiting time of the automated guided vehicle is reduced from 3.5 minutes to 1.8 minutes, a reduction of 48.6%, indicating that the optimized path planning method significantly reduces the equipment waiting time and makes the automated guided vehicle task scheduling smoother.

[0198] In terms of task scheduling, the present invention combines order priority, execution unit status and other information to allocate tasks, which increases the task completion rate from 72.3% to 88.7%, an increase of 19.4%, fully utilizing warehouse resources and reducing task backlogs. At the same time, the utilization rate of execution units has also increased significantly, from the original 72.3% to 88.7%, an increase of 16.4%, further demonstrating that the optimization of intelligent task scheduling plays a significant role in improving the efficiency of automated guided vehicles and manual pickers. The average daily order throughput of the optimized warehousing system increased from 30,000 orders to 38,500 orders, an increase of 28.3%, indicating that task scheduling and path optimization have a positive impact on the overall throughput capacity of the warehouse.

[0199] In terms of peak order processing capabilities, before the optimization of the present invention, the warehouse's order throughput during big promotions could only be maintained at 25,000 orders / day, and the throughput capacity was lower than usual. However, after the present invention was used to optimize scheduling and path planning, the average daily order processing volume during peak periods increased to 35,000 orders / day, an increase of 36.9%, indicating that the optimization strategy of the present invention has a strong ability to relieve pressure on the warehousing system during peak periods and can better adapt to order fluctuations. In addition, under the condition of intensive automated guided vehicle tasks during peak periods, the optimized scheduling strategy of the present invention reduces the average waiting time of automated guided vehicles from 3.5 minutes to 1.8 minutes, a decrease of 48.6%, which greatly reduces the inefficiency of execution units caused by task accumulation.

[0200] In terms of cost and economic benefits, the traditional warehousing operation mode has caused high operating costs and manual intervention costs due to path conflicts and inefficient task scheduling. The present invention reduces unnecessary equipment operation and labor costs through intelligent scheduling, which reduces the comprehensive operating costs of the warehouse by 11.8%, reduces the order fulfillment costs by 9.4%, and increases the overall profit margin by 6.2%. In addition, due to the significant reduction in order fulfillment time, the user experience has been optimized, and the customer satisfaction score has increased from 4.2 / 5 to 4.57 / 5, an increase of 8.7%, indicating that the optimized warehousing system of the present invention has improved the efficiency while also improving the quality of e-commerce logistics services.

[0201] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for e-commerce intelligent warehousing optimization and dynamic scheduling based on deep learning, characterized in that: The steps include: S1. Collect data in the storage environment and standardize the collected data to generate standardized data; S2. Use the Transformer model to learn standardized data, predict future order demand, and adjust the replenishment strategy based on inventory status data; S3. Build the MoE-Transformer model and use it to optimize the SKU storage location and adjust the storage location of the goods according to future order requirements. S4. Based on future order demand, optimized inventory status and SKU storage location, order tasks are divided, sorted according to task priority, and task allocation is performed; S5. Based on the task allocation results, a storage environment topology map is constructed, the paths of the execution units are calculated using the improved GAT, and a path optimization objective function is constructed. Path optimization and dynamic adjustment are performed according to the path optimization objective function. S6. Based on the long-term storage environment operation data, update the parameters of the MoE-Transformer model and the improved GAT, optimize task generation and task allocation, and optimize inventory layout and storage partitioning.

2. According to claim 1, a method for e-commerce intelligent warehousing optimization and dynamic scheduling based on deep learning is characterized in that: The data includes inventory status data, historical order data, equipment location information, task execution records, and warehouse environment variables.

3. According to the deep learning-based e-commerce intelligent warehousing optimization and dynamic scheduling method of claim 1, it is characterized in that: The S3 specifically includes: S31, constructing a SKU storage optimization input data set, wherein the SKU storage optimization input data set includes an order demand prediction matrix, a SKU access frequency matrix, and storage area information; S32. Calculate the storage area status matrix and the SKU storage optimization feature matrix: S={s1,s2,...,s n }; Among them, S represents the storage area state matrix, X t represents the SKU storage optimization feature matrix, s n Indicates the storage location of SKUn in the current warehouse, f k,t represents the SKU access frequency matrix, represents the order demand forecast matrix, Indicates storage area z j Available storage capacity, Indicates the current location of the SKU k To storage area z j The shortest path distance, z j Indicates storage area information; S33, constructing a MoE-Transformer model, wherein the MoE-Transformer model includes a MoE layer, a gating network and a Transformer network, wherein the MoE layer is composed of a plurality of expert networks, and the gating network is used to select the most suitable expert network; S34. Input the SKU storage optimization feature matrix into the MoE-Transformer model to obtain the output H Trans ; S35. Predict SKU storage location, set optimization goals, and use Adam for model training: in, represents the optimization goal, represents the probability distribution of SKU in each storage area, W s and b s represents the trainable parameters, y k,j Indicates the actual storage area. represents the predicted value, C j represents the storage capacity of the storage area, λ represents the regularization parameter, θ represents all trainable parameters of the model, α represents the learning rate, Represents the gradient of the optimization objective; S36. Determine the final SKU storage location: Among them, S t+1 Indicates the final SKU storage location.

4. According to claim 2, a method for e-commerce intelligent warehousing optimization and dynamic scheduling based on deep learning is characterized in that: The S34 specifically includes: S341. Input the SKU storage optimization feature matrix into the expert network to obtain a single expert network output: M e (X t )=ReLU(W e1 X t +b e1 ): A e =ReLU(W e2 M e +b e2 ); Among them, H e represents the output of the expert network, M e represents the expert network, X t represents the SKU storage optimization feature matrix, W e1 , b e1 , W e2 and b e2 represents trainable parameters; S342. Output the weight of each expert network through the gating network: G(X t )=softmax(W g X t +b g ); Among them, G(X t ) represents the weight of each expert network, W g and b g represents the gating network parameters, and softmax represents the activation function; S343. Calculate the output of the MoE layer: Among them, H MoE Represents the output of the MoE layer; S344. Use the multi-head attention mechanism to calculate the output of the Transformer network: Q=W q H MoE ,K=W k H MoE ,V=W v H MoE ; H Trans =LN(ReLU(W t Z+b t )); Among them, Z represents the output after the multi-head attention mechanism is calculated, LN represents layer normalization, Q, K and V represent the vectors of query, key and value, and W q , W k and W v Represents the weight matrix of query, key and value, d k represents the dimension of the query vector Q, H Trans represents the final output of the Transformer network, LN represents layer normalization, and W t and b t Represents the Transformer network parameters.

5. According to claim 4, a method for e-commerce intelligent warehousing optimization and dynamic scheduling based on deep learning is characterized in that: The execution unit includes an AGV, a robotic arm and a manual picker.

6. The e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning according to claim 1 is characterized in that: The S5 specifically includes: S51, construct a warehouse environment topology graph G = (V, E), and define a path weight matrix, where V in the warehouse environment topology graph represents a set of key nodes in the warehouse, and E in the warehouse environment topology graph represents an edge set of traversable paths within the warehouse: W i,j =d i,j +βC i,j +γL i,j ; Among them, W i,j represents the path weight matrix, which is the cost from path i to path j, d i,j represents the Euclidean distance between path i and path j, C i,j Indicates the channel congestion degree between path i and path j, L i,j represents the historical path usage frequency from path i to path j, β and γ represent the adjustment weights of path congestion and historical usage rate; S52. Generate a task path request for the execution unit according to the SKU storage location and the execution unit status: Among them, P R represents the path request set, e p,t represents the current position of execution unit p at time t, Indicates that SKU k is in storage area z j The storage coordinates of It means "arbitrary"; S53, using the improved GAT to calculate the path features and generate a node feature matrix; S54. Calculate the optimal path score: Among them, P score (i,j) represents the optimal path score, W (l) represents the feature weights calculated by the improved GAT at layer l, represents the updated representation of node i at layer l, T i,j represents the time weight from path i to path j, represents the frequency of use of path i to path j in historical tasks, ω1, ω2 and ω3 represent trainable path scoring weight parameters, L represents the number of layers of the improved GAT, C i,j represents the channel congestion degree between path i and path j, e i,j represents edge features, W2 represents the trainable weight matrix; S55. Calculate the optimal path according to the optimal path score: P optimal =argmin P ∑ (i,j)∈P P score (i,j); Among them, P optimal represents the path with the minimum cost in the path set P, argmin represents the variable value that makes the function achieve the minimum value, and P represents the path set; S56. Construct a path optimization objective function and dynamically adjust the path according to the path optimization objective function: P' optimal =P optimal ∪{(i′,j′)∣C i',j' <θ}; in, represents the path optimization objective function, represents the path conflict cost of execution unit p at time t, λ1 represents the path conflict adjustment coefficient, P' optimal represents the set of optimized paths, θ represents the path switching cost threshold, C i',j' represents the channel congestion degree between path i' and path j', {} represents a set, ∪ represents the "and" operation, W i,j represents the path weight matrix; S57. Output the final optimized path set: P * =P' optimal ; Among them, P * Represents the final optimized path set.

7. The e-commerce intelligent warehousing optimization and dynamic scheduling method based on deep learning according to claim 5 is characterized in that: The S53 specifically includes: S531. Construct the input of graph neural network: Among them, H (0) represents the initial node feature matrix, represents the initial features of node i, V represents the key node set in the warehouse, include: Node Type i :Storage area t i =1, picking station t i =2, conveyor belt interface t i =3, Shipping port t i =4; Node coordinate x i ,y i : The absolute position of the node in the warehouse layout; Channel width w i : The width of the path where the node is located; Picking point flow rate f i : The historical frequency of picking tasks at the node location; Historical task completion time τ i : reflects the average time required to execute the task at this point in the past; S532, constructing an edge feature matrix; HAVE BEEN i,j ={e i,j |(i,j)∈E}; Among them, E i,j represents the edge feature matrix, which is the feature of path i to path j. E represents the edge set of the traversable path inside the warehouse. i,j include: Physical distance i,j : The Euclidean distance between path i and path j; Path channel congestion weight C i,j : The channel congestion degree between path i and path j; Path usage frequency L i,j : The historical path usage frequency from path i to path j; Path accessibility i,j :If the path is passable, then a i,j =1, otherwise a i,j =0; S533, initialize node feature matrix embedding: Among them, x i represents the original features of node i, represents the initial embedding matrix, represents the bias term; S534, using multi-head attention mechanism to calculate the information interaction between nodes: in, represents the attention weight of the h-th attention head between path i and path j, W h represents the trainable parameter matrix of the h-th attention head, represents the features of node j of the lth attention head, represents the feature of node k of the lth attention head, a h represents the parameter vector of attention calculation, T represents the transposition operation, || represents the vector concatenation operation, N(i) represents the set of neighbor nodes of node i, LeakyReLU represents the activation function, and exp represents the natural exponential function; S535, update the node representation, and obtain the node feature matrix H (L) : in, represents the updated representation of node i after calculation at layer l+1, H represents the number of attention heads, σ represents the nonlinear activation function, and H (L) Represents the node feature matrix calculated by the final L layers.

Images