Hot listing method for mobile robot fulfillment system based on improved TabNet algorithm

Abstract

The present invention discloses a method for hot shelving in a mobile robot fulfillment system based on an improved TabNet algorithm. The method comprises the following steps: collecting historical order data; preprocessing the historical order data; improving the TabNet algorithm, and constructing a hot model based on the improved TabNet algorithm; using the preprocessed data set to train the hot model based on the improved TabNet algorithm, calculating the hot score of each SKU, and performing real-time data updates and dynamic recalculation of the hot score; and putting the goods on the shelves in different shelf areas according to the calculated SKU hot score. The present invention provides a novel method for hot shelving in a mobile robot fulfillment system based on an improved TabNet algorithm, so that high-hot SKUs can be effectively identified and centrally shelved through dynamic feature selection and hot score mechanisms, thereby improving the order line hit rate of the RMFS system and optimizing picking efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent logistics robots, and in particular relates to a method for hot listing of a mobile robot fulfillment system based on an improved TabNet algorithm. Background Art

[0002] In recent years, with the rapid development of e-commerce and smart logistics, intelligent logistics robots and mobile robotic fulfillment systems (RMFS) have gradually become the core technologies of warehousing logistics. In RMFS, an automated guided mobile robot (AMR) transports the required goods to a workstation by moving shelves, and then the worker or robotic arm completes the picking task. Since the task unit of the workstation is the order line, how to improve the probability of hitting multiple order lines in a single handling task, that is, the "hitting rate", has become one of the key challenges to improve system efficiency. Existing studies have shown that SKU popularity has an important impact on the distribution of order demand. By dynamically analyzing SKU popularity and concentrating high-popularity goods on specific shelves, the order line hit rate during AMR handling can be significantly improved, the number of shelf handling times can be reduced, and the picking efficiency can be optimized. Therefore, developing an optimized shelf method based on heat analysis is of great significance to improving the system efficiency of RMFS.

[0003] At present, popularity calculation in practical applications usually relies on simple statistical or Excel numerical calculation methods, such as cumulative sales based on SKU or sales frequency in a recent period of time. Although these methods are easy to implement, they do not take into account the multidimensional characteristics of the data, time series trends, and the complexity of SKU popularity changes, and cannot dynamically adapt to changes in order demand. TabNet, as an innovative deep learning model, can realize automatic feature selection and dynamic weight allocation in the end-to-end learning process, and is particularly good at processing tasks with time series characteristics and high-dimensional data. Introducing the TabNet algorithm into SKU popularity prediction can not only significantly improve the accuracy of popularity calculation, but also dynamically adjust model parameters to adapt to real-time order data, thereby better supporting shelf optimization decisions in the RMFS system. Summary of the invention

[0004] The present invention aims to solve one of the technical problems existing in the related art at least to a certain extent.

[0005] The purpose of the present invention is to provide a mobile robot fulfillment system hot shelving method based on an improved TabNet algorithm, so that high-hot SKUs can be effectively identified and centrally shelved through dynamic feature selection and hot scoring mechanisms, thereby improving the order line hit rate in the RMFS system and optimizing picking efficiency.

[0006] In order to achieve the above-mentioned purpose, the present invention provides a method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm, comprising the steps of:

[0007] S100, collecting historical order data including date, SKU identifier, sales volume and order number features;

[0008] S200, preprocessing the historical order data, including missing value processing, outlier processing, feature engineering, data standardization, and data set division;

[0009] S300, improve the TabNet algorithm, and build a heat model based on the improved TabNet algorithm, including:

[0010] Introducing adaptive feature selection optimization based on dynamic heat feedback mechanism, calculating heat error feedback by comparing predicted heat value with actual outbound quantity, and using this feedback to adjust feature selection strategy during model training;

[0011] Perform multi-level feature representation optimization based on task self-supervised learning;

[0012] S400, using the preprocessed data set to train a popularity model based on the improved TabNet algorithm, calculating the popularity score of each SKU, and performing real-time data updates and dynamic recalculation of the popularity score;

[0013] S500: Based on the calculated SKU popularity score, the goods are put on the shelves in different shelf areas and the shelf layout is optimized.

[0014] A further preferred technical solution of the present invention is that in step S200, in preprocessing the historical order data, the method for processing missing values ​​includes:

[0015] Identify the type of missing values. If the SKU is missing, delete the record or fill it in based on the SKU of similar orders. If the order number is missing, fill it in with an automatically generated unique identifier. For missing sales, choose to fill it in with the historical average sales of the same SKU, the sales of the previous order, or the median of historical sales.

[0016] The method for handling outliers includes:

[0017] Identify outlier types, convert sales data into numeric types, and delete data with non-positive integer values; delete data containing fields other than letters and numbers in SKU and order numbers;

[0018] The feature engineering method includes:

[0019] Create new features, aggregate the average daily sales of each SKU, and extract time features from the three dimensions of day, week, and month;

[0020] Use LabelEncoder to encode the data that cannot be directly processed by the deep learning model and convert it into a format suitable for model input;

[0021] The method of data standardization is:

[0022] Use StandardScaler to normalize the sales volume field;

[0023] The method of dividing the data set is:

[0024] The processed data set is divided into a training set and a test set for model training and verification.

[0025] Preferably, in step S300, when improving the TabNet algorithm, a specific method for introducing adaptive feature selection optimization based on a dynamic heat feedback mechanism is as follows:

[0026] S311. Establish a heat feedback mechanism. Whenever there is new order data, the model predicts a real-time heat value for the SKU. and the actual outbound quantity Compare and calculate the thermal error feedback , the formula is:

[0027]

[0028] This feedback is used to adjust the model training process and optimize the feature selection strategy;

[0029] S312. Select dynamic features, introduce time series features in the training process, capture the trend change of SKU popularity, and calculate the 7-day time series trend as follows:

[0030]

[0031] in, Indicates from the time point arrive Sales volume The sum of It’s the starting time. is the current time point;

[0032] The formula for dynamic feature adjustment is:

[0033]

[0034] in, represents the original feature importance, represents the adjusted feature importance;

[0035] S313, instantiate the update mechanism, and dynamically adjust the feature mask according to the error feedback in each decision step to optimize the feature selection strategy of each sample, increase the probability of the features of high-popularity SKUs being selected, and reduce the probability of the features of low-popularity SKUs being selected. The formula for dynamic mask adjustment is:

[0036]

[0037] in, represents the current feature mask matrix, Indicates the adjustment step size.

[0038] Preferably, in step S300, when improving the TabNet algorithm, multi-level feature representation optimization based on task self-supervised learning is performed; the specific method is:

[0039] S321. Establish a self-supervised learning framework, add self-supervised tasks to the decision step of TabNet, use some missing SKU popularity information for prediction training, mask some fields for some SKU popularity data, and then use other known fields to predict these missing fields through self-supervised learning;

[0040] Among them, when masking the sales volume field, the formula for introducing missing values ​​is:

[0041]

[0042] in, Indicates The input feature vector of samples; It means that after applying the self-supervision task, The feature vector of samples; A collection of indexes representing masked fields; Represents missing values, which are used to simulate the situation of missing features;

[0043] The formula for predicting missing values ​​is:

[0044]

[0045] in, Represents the field value predicted by the TabNet model, Represents the TabNet model function;

[0046] The calculation formula of the self-supervised loss function is:

[0047]

[0048] in, is the loss of the self-supervised task, used to measure the prediction value and the true value The mean square error of is the number of samples that were masked;

[0049] S322. In the multi-step decision of TabNet, the important features are dynamically selected through the mask mechanism, and these features are gradually processed with high granularity;

[0050] S323. Implement multi-task learning, combine SKU popularity prediction and sorting tasks, build a multi-task learning framework, and enable the model to optimize popularity prediction and sorting tasks at the same time; the main steps of multi-task learning include:

[0051] Calculate the heat prediction task loss:

[0052]

[0053] in, The loss of the heat prediction task is used to measure the predicted sales and real sales The mean absolute error of

[0054] Calculate the heat sorting task loss:

[0055]

[0056] in, represents the loss for the ranking task, measuring the amount of inconsistency between the predicted ranking and the true ranking; Represents the predicted value rank; Indicates actual value rank; is an indicator function, which takes the value 1 when the predicted ranking is inconsistent with the true ranking, and 0 otherwise;

[0057] Calculate the joint loss function:

[0058]

[0059] in, is the loss function weight, which is used to balance the importance of the three tasks.

[0060] Preferably, in step S400, the preprocessed data set is used to train the heat model based on the improved TabNet algorithm, and when calculating the heat score of each SKU, the feature importance mask of the TabNet multi-step decision is used. and the cumulative attention at the decision step Combined to generate the SKU's heat score, expressed as:

[0061]

[0062] in, Represents a weight vector of a linear mapping, which is used to map the comprehensive features output at each step to the final heat score; Indicates The contribution of the step decision to the overall score is defined according to the output feature vector of the step: ,in For the step output and dimension is The characteristic vector of Ensure that the output is non-negative and emphasize the importance of features; Indicates The feature selection mask of the step is used to control which features are selected to enter the decision of this step and is defined as: ,in represents the feature mask matrix of the previous step, represents real-time error feedback, used to dynamically adjust the mask, Represents the adjustment step size, which is used to control the update speed of the mask.

[0063] Preferably, after calculating the popularity score of each SKU in step S400, real-time data update and dynamic recalculation of the popularity score are performed, and the specific steps are as follows:

[0064] S410, receiving a real-time order data stream from an order management system, and adding the newly acquired data to the analysis;

[0065] S420: Set a sliding time window, perform aggregation calculation on the order data in the window, generate time series features, and dynamically update the SKU popularity;

[0066] S430, based on the sliding window data, using the improved TabNet algorithm to recalculate the SKU heat score, and adjusting the score result in combination with the heat feedback mechanism;

[0067] S440, data caching and batch updating, caches real-time data in small batches and processes them uniformly, ensuring that system resource utilization is optimized without affecting real-time performance.

[0068] Preferably, in step S500, the goods are put on different shelf areas according to the calculated SKU popularity scores, and the shelf layout is optimized at the same time. The specific steps are:

[0069] S510, heat grouping clustering, based on the SKU heat score, uses the K-means clustering algorithm to divide the SKUs into three groups: high heat, medium heat, and low heat, and calculates the total sales share of each group;

[0070] S520, assigning shelf partitions: according to the clustering results, assigning high-heat SKUs to shelves close to the workstations, assigning medium-heat SKUs to middle shelves, and assigning low-heat SKUs to remote shelves;

[0071] S530, dynamic layout adjustment: when the popularity of SKU changes significantly, the shelf re-layout process is automatically triggered to adjust the high-popularity SKU from the remote shelf to the area close to the workstation;

[0072] S540, layout visualization support, generates a visualization interface for heat layout through the robot management system, displays the real-time distribution status of different heat SKUs on the shelves, and supports users to manually fine-tune the layout plan.

[0073] Another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions enable a computer to execute the above-mentioned method for listing the popularity of a mobile robot fulfillment system based on the improved TabNet algorithm.

[0074] Another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned mobile robot fulfillment system heat listing method based on the improved TabNet algorithm.

[0075] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned mobile robot fulfillment system heat listing method based on the improved TabNet algorithm.

[0076] The invention is based on the mobile robot fulfillment system heat listing method of the improved TabNet algorithm. Through a series of innovative technical means, it realizes the accurate prediction and dynamic management of SKU heat, and significantly improves the overall performance and efficiency of the mobile robot fulfillment system. The specific technical effects are as follows:

[0077] Improve order line hit rate:

[0078] By introducing adaptive feature selection optimization based on dynamic heat feedback mechanism, the model can adjust the feature selection strategy in real time, accurately identify high-heat SKUs, and centrally put them on specific shelves. This significantly improves the order line hit rate of the mobile robot fulfillment system (RMFS), reduces the number of times the robot moves shelves, and optimizes picking efficiency. In actual applications, the order line hit rate has increased from the traditional 1.2 to more than 2, and the picking time has been reduced by 30%.

[0079] Enhance the robustness and adaptability of the model:

[0080] By introducing time series features and dynamic feature selection mechanisms, the model is not only based on current data, but also refers to historical trends, dynamically enhancing the importance of SKU features with seasonal changes or significant changes in popularity during promotions. This enables the model to better adapt to changes in order demand and improve its sensitivity and accuracy to changes in popularity. During promotions, the model can quickly identify SKUs with significant changes in popularity and adjust the listing strategy in a timely manner to ensure that high-popularity SKUs are listed first.

[0081] Support real-time data update and dynamic adjustment:

[0082] Through sliding window calculation and real-time score update mechanism, the system can receive new order data in real time and dynamically update the popularity score of SKU. This ensures that the system can reflect changes in order demand in real time, adjust the shelf strategy in time, and maintain efficient picking performance. The system can complete the processing of new order data and the update of popularity score in a short time, ensuring that the shelf layout is always in the optimal state.

[0083] Optimize shelf layout and load balancing:

[0084] By introducing the shelf load balancing mechanism, the system considers the current load of the shelf when allocating SKUs to different shelves, avoiding overloading of some shelves and ensuring the stable operation of the system. The system can automatically adjust the shelf load to ensure the load balance of each shelf and reduce system failures caused by overload.

[0085] By combining the warehouse's spatial layout with the robot's walking path, the shelf partition allocation strategy is optimized to further reduce the robot's moving distance and picking time. The optimized shelf layout reduces the robot's average moving distance by 20% and the picking time by 25%.

[0086] Improve system practicality and user satisfaction:

[0087] Through simulation tests and user feedback mechanisms, the system can evaluate and fine-tune the optimized shelf layout to ensure that the actual effect of the layout plan meets expectations and improve the practicality of the system and user satisfaction. User feedback shows that the optimized shelf layout has increased picking efficiency by 30% and significantly improved the job satisfaction of operators.

[0088] Improve model training efficiency and resource utilization:

[0089] By utilizing incremental learning technology, the model can update model parameters based only on newly received order data without retraining the entire model, significantly improving the system's response speed and resource utilization efficiency. The system can complete incremental learning and update model parameters in a short period of time, while traditional methods take a long time.

[0090] By dynamically adjusting the size of the sliding time window, the system can balance real-time performance and computing resource consumption according to the flow and characteristics of real-time data, ensuring efficient operation of the system. The system automatically adjusts the size of the sliding time window according to the real-time data flow, ensuring efficient processing capabilities even during high-flow periods.

[0091] In summary, the present invention, through a series of innovative technical means, not only improves the order line hit rate and picking efficiency of the mobile robot fulfillment system, but also enhances the robustness and adaptability of the model, supports real-time data update and dynamic adjustment, optimizes shelf layout and load balancing, improves the practicality of the system and user satisfaction, and has significant economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 The present invention is a flow chart of a method for listing the heat of a mobile robot fulfillment system based on an improved TabNet algorithm.

[0093] Figure 2 This is a structural block diagram of the heat model based on the improved TabNet algorithm. DETAILED DESCRIPTION

[0094] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. 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. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0095] Combine the following Figure 1-Figure 2 The invention describes a method for listing the heat of a mobile robot fulfillment system based on an improved TabNet algorithm.

[0096] The Robotic Mobile Fulfillment System (RMFS) is an advanced "goods-to-person" warehousing solution that automatically guides mobile robots to transport shelves to workstations, thereby assisting workers or robotic arms to complete picking tasks. RMFS not only significantly improves the efficiency of warehousing logistics, but also reduces the complexity of manual handling. It has gradually become a core technology in modern intelligent logistics systems. In the RMFS system, the SKU hot shelf strategy plays a vital role in improving efficiency. SKU hotness reflects the frequency of order lines of a product within a period of time. Putting high-hot products on specific shelves can significantly improve the order line hit rate in a single handling task. This means that the robot can meet more order requirements each time it handles, thereby reducing the number of shelf handling times and improving picking efficiency. The optimization strategy of hot shelf can not only improve the overall operating efficiency of the system, but also reduce energy consumption and shorten the order completion time. It is of great significance for the efficient operation of the RMFS system in a complex warehousing environment.

[0097] Embodiment 1: This embodiment provides a method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm. Figure 1 As shown in the figure, the following steps are included: First, collect historical order data, including features such as date, SKU identification, sales volume, order number, etc., and process the data, covering missing value filling, outlier detection and feature engineering to improve data quality. Then, in view of the hot shelf problem of RMFS, the TabNet algorithm is improved, and an adaptive feature selection optimization based on a dynamic heat feedback mechanism is introduced, and a multi-level feature representation optimization based on task self-supervised learning is implemented. Next, the pre-trained model is trained using the historical order dataset, and the heat score of each SKU is calculated based on the improved TabNet algorithm, and real-time data update and heat recalculation are supported. Finally, the SKU is classified according to its heat score, and the goods are put on different shelves according to the heat level, and the shelf layout is optimized to improve the picking efficiency.

[0098] Each step is explained in detail.

[0099] S100: Collect historical order data, where the data includes features such as date, SKU identifier, sales volume, and order number.

[0100] The historical order data in this embodiment refers to the in-and-out order data within a given past time period extracted from the enterprise's inventory management system (WMS) and order management system (OMS). The data format must be an xlsx or csv table file, and should contain at least four fields: date, SKU identifier, sales volume, and order number. The date should be in a standard short date format such as "2024 / 11 / 28", which is convenient for using the pd.to_datetime function in the pandas library to convert it into a date that the model can recognize; the SKU identifier should be a text composed of numbers or uppercase and lowercase letters; the sales volume should be in digital form and must be a non-zero natural number; the order number should be a text composed of numbers or uppercase and lowercase letters, just like the SKU identifier.

[0101] S200, preprocessing the data, including missing value filling, outlier detection and feature engineering, to improve data quality.

[0102] After obtaining the required data, this embodiment first performs comprehensive preprocessing on the extracted data to improve the quality and reliability of the data.

[0103] In data processing, the first step is to identify and fill in missing values. After analyzing the collected data, if the SKU identifier is found to be missing, there are two options for filling: one is to delete the record to maintain data integrity, and the other is to fill in by finding the SKU of similar orders, which can be filled in by historical orders of the same customer or popular SKUs in the same time period. If the order number is missing, it is filled in with an automatically generated unique identifier. In the case of missing sales, it can be filled in according to the historical average sales of the same SKU, the sales of the previous order, or the median to ensure the rationality of the sales data.

[0104] After completing the missing value processing, enter the outlier detection step. This step mainly targets the sales data, converts its type into a numerical format, and checks each line, deleting all records that are not positive integers to ensure the validity of the data. At the same time, for the SKU and order number fields, clear all data containing invalid characters other than letters and numbers to ensure the neatness and consistency of the fields.

[0105] Then, feature engineering is performed. First, new features are created to calculate the average daily sales of each SKU. By aggregating historical order data, the average daily sales of each SKU is obtained. At the same time, time features are extracted from date information, and features are generated from the three dimensions of day, week, and month to capture the seasonal and cyclical changes in SKU sales. For categorical data such as SKU, LabelEncoder is used to encode them, converting these non-numerical features into numerical form to meet the input requirements of the deep learning model.

[0106] Next, data standardization is performed. StandardScaler is used to normalize the sales volume field and convert the data into a normal distribution with a mean of 0 and a standard deviation of 1, ensuring that the impact of different features on the results during model training is at the same order of magnitude, thereby improving the model's convergence speed and prediction accuracy.

[0107] Finally, the processed data set is divided into training set and test set in proportion, with 80% of the data used for training and 20% for testing, to facilitate subsequent model training and performance evaluation.

[0108] S300, improve the TabNet algorithm for RMFS popularity listing, firstly introduce adaptive feature selection optimization based on dynamic popularity feedback mechanism, secondly realize multi-level feature representation optimization based on task self-supervised learning. Construct a popularity model based on the improved TabNet algorithm.

[0109] like Figure 2 As shown, this embodiment first establishes a heat feedback mechanism: whenever new order data arrives, the model will perform real-time heat prediction on the SKU and calculate the heat error feedback. Specifically, the predicted heat value is set to The actual shipment quantity is , through the formula Calculate the thermal error feedback, which will be used to adjust the model training process to optimize the feature selection strategy.

[0110] Next, in the training process, time series features are introduced to capture the trend changes of SKU popularity. Here, a 7-day moving average is used for trend calculation, and the calculation formula is:

[0111]

[0112] in Indicates from the time point arrive The feature selection mechanism of TabNet is not only based on current data, but also refers to historical trends to dynamically enhance the importance of SKU features with seasonal changes or significant changes in popularity during promotions. Realize dynamic adjustment of feature importance.

[0113] In each decision step, TabNet dynamically adjusts the feature mask based on the previously calculated error feedback to optimize the feature selection strategy of the sample. The features of high-popularity SKUs are selected more frequently, while the features of low-popularity SKUs are weakened. The model uses the formula For the current feature mask matrix Make adjustments, including Indicates the adjustment step size.

[0114] This embodiment also establishes a self-supervised learning framework. In the TabNet decision step, a self-supervised task is designed to train by deliberately blocking some SKU popularity information. Here, we take the sales volume field of the category as an example, set some fields as missing values, and use the formula to simulate the situation of missing features. Then, through the formula Predict missing values ​​and use the loss function To measure the predicted value and the true value The mean square error between them is increased, thereby strengthening the model's ability to learn features.

[0115] in, Indicates The input feature vector of samples; It means that after applying the self-supervision task, The feature vector of samples; A collection of indexes representing masked fields; Represents missing values, which are used to simulate the situation of missing features; Represents the field value predicted by the TabNet model; Represents the TabNet model function; is the loss of the self-supervised task, used to measure the prediction value and the true value The mean square error of is the number of masked samples.

[0116] In addition, in the multi-level feature processing part, fine-grained feature processing is performed through self-supervised learning tasks, allowing the TabNet model to capture the complex feature combination of SKU popularity. Specifically, important features are dynamically selected after each decision step, and these features are gradually processed in a more fine-grained manner.

[0117] Finally, the TabNet model can achieve multi-task learning, combining SKU popularity prediction and sorting tasks for optimization. In this process, the model first calculates the loss of the popularity prediction task. and the loss of the heat sorting task , respectively, using the formula and To measure the error between predicted sales and actual sales, as well as the inconsistency between predicted ranking and actual ranking.

[0118] in, The loss of the heat prediction task is used to measure the predicted sales and real sales The mean absolute error of represents the loss for the ranking task, measuring the amount of inconsistency between the predicted ranking and the true ranking; Represents the predicted value rank; Indicates actual value rank; is an indicator function, which takes a value of 1 when the predicted ranking is inconsistent with the true ranking, and 0 otherwise.

[0119] Finally, the joint loss function is constructed by combining the losses of each task:

[0120]

[0121] in, is the weight coefficient used to balance the importance of multiple tasks.

[0122] S400, pre-training model, trains a popularity model based on the improved TabNet algorithm through the historical order data set, and calculates the popularity score of each SKU, while supporting real-time data update and popularity recalculation.

[0123] This embodiment first trains the shelf popularity model based on the improved TabNet algorithm through a large number of historical order data sets, and combines feature masks and cumulative attention in the model to calculate the popularity score of SKU. Optimize at each decision step and calculate the attention value , ensuring that the importance of the feature is emphasized, and the final heat score is calculated by the formula generate.

[0124] in, Represents a weight vector of a linear mapping, which is used to map the comprehensive features output at each step to the final heat score; Indicates The contribution of the first decision to the overall score; For the step output and dimension is The characteristic vector of Ensure that the output is non-negative and emphasize the importance of features; Indicates The feature selection mask of the step is used to control which features are selected to enter the decision of this step and is defined as: ,in represents the feature mask matrix of the previous step, represents real-time error feedback, used to dynamically adjust the mask, Represents the adjustment step size, which is used to control the update speed of the mask.

[0125] Next, the system establishes an interface with the order management system (OMS) to receive order data containing order number, SKU identifier, sales volume and timestamp fields in real time, and formats it and stores it in the in-memory database to support fast access.

[0126] Then, a sliding time window is set to dynamically update the SKU popularity, and time series features are generated through aggregation calculation, including daily average sales and weekly sales. Based on the characteristics of the sliding window, the system uses the improved TabNet algorithm to recalculate the SKU popularity score, and dynamically adjusts the score results in combination with the real-time popularity feedback mechanism. Defined as , and by the formula Adjust the mask matrix to ensure that the heat score reflects the demand for orders in the current stage.

[0127] Finally, to improve processing efficiency, the system establishes a small batch caching mechanism to cache real-time data in small batches. After the cached data reaches the set amount or time threshold, it is processed and analyzed uniformly. This method can improve the real-time response capability of the system while optimizing resource utilization.

[0128] S500: According to the heat score, the goods are classified according to the heat level of the SKU, put on different shelves, and the shelf layout is optimized to improve the picking efficiency.

[0129] This embodiment first divides SKUs into three types according to the calculated SKU heat scores: high heat, medium heat, and low heat, so as to facilitate efficient shelf placement. In specific operations, the system uses the K-means clustering algorithm to group and cluster the SKU heat scores, setting the clustering parameter to 3, with the purpose of dividing the SKUs into three groups. The total sales volume ratio of each group is input in the clustering process to provide a decision basis for subsequent shelf allocation.

[0130] Next, the allocation of shelf partitions will be carried out according to the clustering results. High-hot SKUs are preferentially allocated to shelves close to workstations to shorten the transportation path and improve the efficiency of picking; medium-hot SKUs are allocated to the middle shelf position to ensure a certain access convenience; and low-hot SKUs are arranged to the far-end shelves. This layout strategy can effectively reduce the movement time of AMR during the picking process, thereby improving the overall efficiency of the system.

[0131] Then, the system needs to have a dynamic layout adjustment function. When the popularity of the SKU changes significantly, such as the change rate of the popularity score exceeds the preset threshold, the system will automatically trigger the process of re-layout of the shelves. The specific implementation of this function relies on real-time data monitoring and analysis. When it is found that the popularity score of a certain SKU has increased significantly, the system will immediately send a notification to increase the priority of the SKU the next time it is put on the shelf. At the same time, the sorting weight of the shelf carrying the SKU will be recalculated and timely adjustments will be made to ensure that the picking efficiency is not affected.

[0132] Finally, the system can provide layout visualization support. The Robot Management System (RMS) generates the heat layout into a visualization interface to display the distribution status of different heat SKUs on the shelf in real time. The interface will present the heat distribution of each SKU in an intuitive way such as heat map, and support users to manually fine-tune the layout plan, so that managers can optimize and adjust according to the actual picking situation.

[0133] Embodiment 2: This embodiment provides a non-transitory computer-readable storage medium, on which computer instructions are stored, and the computer instructions enable the computer to execute a method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm, the method comprising the following steps:

[0134] S100, collecting historical order data including date, SKU identifier, sales volume and order number features;

[0135] S200, preprocessing the historical order data, including missing value processing, outlier processing, feature engineering, data standardization, and data set division;

[0136] S300, improve the TabNet algorithm, and build a heat model based on the improved TabNet algorithm, including:

[0137] Introducing adaptive feature selection optimization based on dynamic heat feedback mechanism, calculating heat error feedback by comparing predicted heat value with actual outbound quantity, and using this feedback to adjust feature selection strategy during model training;

[0138] Perform multi-level feature representation optimization based on task self-supervised learning;

[0139] S400, using the preprocessed data set to train a popularity model based on the improved TabNet algorithm, calculating the popularity score of each SKU, and performing real-time data updates and dynamic recalculation of the popularity score;

[0140] S500: Based on the calculated SKU popularity score, the goods are put on the shelves in different shelf areas and the shelf layout is optimized.

[0141] Embodiment 3: This embodiment provides an electronic device, which may include: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. The processor may call the logic instructions in the memory to execute a method for listing the heat of a mobile robot fulfillment system based on an improved TabNet algorithm, and the method includes the following steps:

[0142] S100, collecting historical order data including date, SKU identifier, sales volume and order number features;

[0143] S200, preprocessing the historical order data, including missing value processing, outlier processing, feature engineering, data standardization, and data set division;

[0144] S300, improve the TabNet algorithm, and build a heat model based on the improved TabNet algorithm, including:

[0145] Introducing adaptive feature selection optimization based on dynamic heat feedback mechanism, calculating heat error feedback by comparing predicted heat value with actual outbound quantity, and using this feedback to adjust feature selection strategy during model training;

[0146] Perform multi-level feature representation optimization based on task self-supervised learning;

[0147] S400, using the preprocessed data set to train a popularity model based on the improved TabNet algorithm, calculating the popularity score of each SKU, and performing real-time data updates and dynamic recalculation of the popularity score;

[0148] S500: Based on the calculated SKU popularity score, the goods are put on the shelves in different shelf areas and the shelf layout is optimized.

[0149] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0150] Embodiment 4: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a mobile robot fulfillment system hot listing method based on an improved TabNet algorithm. The method includes the following steps:

[0151] S100, collecting historical order data including date, SKU identifier, sales volume and order number features;

[0152] S200, preprocessing the historical order data, including missing value processing, outlier processing, feature engineering, data standardization, and data set division;

[0153] S300, improve the TabNet algorithm, and build a heat model based on the improved TabNet algorithm, including:

[0154] Introducing adaptive feature selection optimization based on dynamic heat feedback mechanism, calculating heat error feedback by comparing predicted heat value with actual outbound quantity, and using this feedback to adjust feature selection strategy during model training;

[0155] Perform multi-level feature representation optimization based on task self-supervised learning;

[0156] S400, using the preprocessed data set to train a popularity model based on the improved TabNet algorithm, calculating the popularity score of each SKU, and performing real-time data updates and dynamic recalculation of the popularity score;

[0157] S500: Based on the calculated SKU popularity score, the goods are put on the shelves in different shelf areas and the shelf layout is optimized.

[0158] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0159] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm, characterized in that: Includes steps: S100, collecting historical order data including date, SKU identifier, sales volume and order number features; S200, preprocessing the historical order data, including missing value processing, outlier processing, feature engineering, data standardization, and data set division; S300, improve the TabNet algorithm, and build a heat model based on the improved TabNet algorithm, including: Introducing adaptive feature selection optimization based on dynamic heat feedback mechanism, calculating heat error feedback by comparing predicted heat value with actual outbound quantity, and using this feedback to adjust feature selection strategy during model training; Perform multi-level feature representation optimization based on task self-supervised learning; S400, using the preprocessed data set to train a popularity model based on the improved TabNet algorithm, calculating the popularity score of each SKU, and performing real-time data updates and dynamic recalculation of the popularity score; The preprocessed data set is used to train the popularity model based on the improved TabNet algorithm. When calculating the popularity score of each SKU, the feature importance mask M of the TabNet multi-step decision is used. i and the cumulative attention η at the decision step i Combined to generate the SKU's heat score, expressed as: Among them, W final Represents a weight vector of a linear mapping, which is used to map the comprehensive features output at each step to the final heat score; η i It represents the contribution of the decision in step i to the overall score, and is defined as follows according to the output feature vector of this step: where d i,c is the feature vector of the i-th step output and dimension c, ReLU(x)=max(0,x) ensures that the output is non-negative, emphasizing the importance of the feature; M i Represents the feature selection mask of the i-th step, which is used to control which features are selected to enter the decision of this step, and is defined as: M i =M i-1 ·(1+L·e), where M i-1 represents the feature mask matrix of the previous step, represents the thermal error feedback, which is used to dynamically adjust the mask, and L represents the adjustment step size, which is used to control the update speed of the mask; After calculating the popularity score of each SKU, perform real-time data update and dynamic recalculation of the popularity score. The specific steps are as follows: S410, receiving a real-time order data stream from an order management system, and adding the newly acquired data to the analysis; S420: Set a sliding time window, perform aggregation calculation on the order data in the window, generate time series features, and dynamically update the SKU popularity; S430, based on the sliding window data, using the improved TabNet algorithm to recalculate the SKU heat score, and adjusting the score result in combination with the heat feedback mechanism; S440, data caching and batch updating, caches real-time data in small batches and processes them uniformly, ensuring that system resource utilization is optimized without affecting real-time performance; S500: Based on the calculated SKU popularity score, the goods are put on the shelves in different shelf areas and the shelf layout is optimized.

2. According to claim 1, the method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm is characterized in that: In step S200, the historical order data is preprocessed, and the missing value processing method includes: Identify the type of missing values. If the SKU is missing, delete the records with missing SKU or fill it in based on the SKU of similar orders. If the order number is missing, fill it in with an automatically generated unique identifier. For missing sales, choose to fill it in with the historical average sales of the same SKU, the sales of the previous order, or the median of historical sales. The method for handling outliers includes: Identify outlier types, convert sales data into numeric types, and delete data with non-positive integer values; delete data containing fields other than letters and numbers in SKU and order numbers; The feature engineering method includes: Create new features, aggregate the average daily sales of each SKU, and extract time features from the three dimensions of day, week, and month; Use LabelEncoder to encode the data that cannot be directly processed by the deep learning model and convert it into a format suitable for model input; The method of data standardization is: Use StandardScaler to normalize the sales volume field; The method of dividing the data set is: The processed data set is divided into a training set and a test set for model training and verification.

3. The method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm according to claim 1 is characterized in that: Step S300 introduces the specific method of adaptive feature selection optimization based on dynamic heat feedback mechanism when improving the TabNet algorithm: S311. Establish a heat feedback mechanism. Whenever there is new order data, the model predicts a real-time heat value for the SKU. And compare it with the actual outbound quantity y, and calculate the heat error feedback e, the formula is: This feedback is used to adjust the model training process and optimize the feature selection strategy; S312. Select dynamic features, introduce time series features in the training process, capture the trend change of SKU popularity, and calculate the 7-day time series trend as follows: in, represents the sales volume Q from time point t-6 to t i The sum of , t-6 is the starting time point, and t is the current time point; The formula for dynamic feature adjustment is: I′ i =I i ·(1+e) Among them, I i Represents the original feature importance, I′ i represents the adjusted feature importance; S313, instantiate the update mechanism, and dynamically adjust the feature mask according to the error feedback in each decision step to optimize the feature selection strategy of each sample, increase the probability of the features of high-popularity SKUs being selected, and reduce the probability of the features of low-popularity SKUs being selected. The formula for dynamic mask adjustment is: M t+1 =M t ·(1+L·e) Among them, M t Represents the current feature mask matrix, and L represents the adjustment step size.

4. The method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm according to claim 1 is characterized in that: Step S300 performs multi-level feature representation optimization based on task self-supervised learning when improving the TabNet algorithm; the specific method is: S321. Establish a self-supervised learning framework, add self-supervised tasks to the decision step of TabNet, use some missing SKU popularity information for prediction training, mask some fields for some SKU popularity data, and then use other known fields to predict these missing fields through self-supervised learning; Among them, when masking the sales volume field, the formula for introducing missing values ​​is: Among them, X i represents the true value of the i-th sample; M self (X i ) represents the feature vector of the i-th sample after applying the self-supervision task; MI represents the index set of the masked field; NaN represents the missing value, which is used to simulate the situation of missing features; The formula for predicting missing values ​​is: in, Indicates the predicted missing value, f TabNet Represents the TabNet model function; The calculation formula of the self-supervised loss function is: Among them, L self is the loss of the self-supervised task, which is used to measure the prediction of missing values and the true value X i The mean square error of |MI| is the number of masked samples; S322. In the multi-step decision of TabNet, the important features are dynamically selected through the mask mechanism, and these features are gradually processed with high granularity; S323. Implement multi-task learning, combine SKU popularity prediction and sorting tasks, build a multi-task learning framework, and enable the model to optimize popularity prediction and sorting tasks simultaneously; the steps of multi-task learning include: Calculate the heat prediction task loss: Among them, L heat The loss of the heat prediction task is used to measure the predicted sales and real sales volume i The mean absolute error of Calculate the heat sorting task loss: Among them, L rank represents the loss for the ranking task, measuring the amount of inconsistency between the predicted ranking and the true ranking; Indicates forecast sales Ranking of R(y i ) represents the actual sales volume y i The ranking of; I(·) is the indicator function, the value is 1 when the predicted ranking is inconsistent with the true ranking, otherwise it is 0; Calculate the joint loss function: L total =α·L heat +β·L rank +γ·L self Among them, α, β, and γ are the weights of the loss function, which are used to balance the importance of the three tasks.

5. The method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm according to claim 1 is characterized in that: In step S500, the goods are put on different shelf areas according to the calculated SKU popularity score, and the shelf layout is optimized at the same time. The specific steps are as follows: S510, heat grouping clustering, based on the SKU heat score, uses the K-means clustering algorithm to divide the SKUs into three groups: high heat, medium heat, and low heat, and calculates the total sales share of each group; S520, assigning shelf partitions: according to the clustering results, assigning high-heat SKUs to shelves close to the workstations, assigning medium-heat SKUs to middle shelves, and assigning low-heat SKUs to remote shelves; S530, dynamic layout adjustment: when the popularity of SKU changes significantly, the shelf re-layout process is automatically triggered to adjust the high-popularity SKU from the remote shelf to the area close to the workstation; S540, layout visualization support, generates a visualization interface for the heat layout through the robot management system, displays the real-time distribution status of different heat SKUs on the shelves, and supports users to manually fine-tune the layout plan.

6. A non-transitory computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and the computer instructions enable the computer to execute the method for listing the popularity of a mobile robot fulfillment system based on an improved TabNet algorithm as described in any one of claims 1-5.

7. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the mobile robot fulfillment system heat listing method based on the improved TabNet algorithm as described in any one of claims 1-5.

8. A computer program product, characterized in that The computer program product includes a computer program, which is stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the mobile robot fulfillment system hot listing method based on the improved TabNet algorithm as described in any one of claims 1 to 5.

Images