Federal learning-based cross-logistics transit site resource collaborative allocation method

Abstract

The invention discloses a cross-logistics transit field resource collaborative allocation method based on federated learning, and the method comprises the following steps: constructing a set of multi-source heterogeneous data in each logistics transit field, and carrying out the feature extraction to obtain a feature vector of each logistics transit field; generating a local data set of each logistics transit site; a federal learning model is constructed, and each logistics transit site and the central server cooperate for training; and based on the federated learning model, constructing a cross-transit-site resource collaborative allocation optimization problem with a time span of T, and solving the optimization problem to obtain a collaborative allocation strategy. According to the invention, through the distributed training and security aggregation technology, cooperative scheduling of resources in different logistics transit fields is realized, and the resource utilization efficiency is improved while the privacy of each logistics transit field is protected.

Description

Technical Field

[0001] The present invention relates to the field of logistics sorting, and in particular to a method for collaboratively allocating resources across logistics transfer sites based on federated learning. Background Art

[0002] With the rapid development of the logistics industry, how to effectively coordinate the allocation of resources in different logistics transfer stations has gradually become the key to improving logistics efficiency. The traditional centralized resource scheduling method requires uploading heterogeneous data in different logistics transfer stations, such as vehicle scheduling data, personnel operation data, equipment operation data, etc., to the central server. This not only makes each logistics transfer station face the risk of data privacy leakage, but also leads to low system processing efficiency, poor compatibility, and insufficient generalization due to data heterogeneity. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for collaborative allocation of resources across logistics transfer stations based on federated learning. Through distributed training and security aggregation technology, the collaborative scheduling of resources in different logistics transfer stations is realized, while protecting the privacy of each logistics transfer station, while improving resource utilization efficiency.

[0004] The object of the present invention is achieved through the following technical solutions: constructing a collection of multi-source heterogeneous data in each logistics transfer station, and performing feature extraction to obtain a feature vector of each logistics transfer station; Generate local data sets for each logistics transfer station; Build a federated learning model, which is trained collaboratively by various logistics transfer stations and the central server; Based on the federated learning model, a time span is constructed T The collaborative allocation optimization problem of cross-transit station resources is solved, and the collaborative allocation strategy is obtained by solving the optimization problem.

[0005] The beneficial effect of the present invention is that the present invention realizes the coordinated scheduling of resources in different logistics transfer stations through distributed training and security aggregation technology, thereby improving resource utilization efficiency while protecting the privacy of each logistics transfer station. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0007] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0008] As a distributed machine learning framework, federated learning can collaboratively train models without sharing original data between logistics transfer stations, which can effectively solve data silos and privacy protection problems. Therefore, the present invention applies federated learning technology to the collaborative allocation of resources across logistics transfer stations, which can optimize the global model and improve the efficiency of resource collaborative scheduling while ensuring the data security of each logistics transfer station.

[0009] like Figure 1 As shown in FIG. 1 , a method for collaboratively allocating resources across logistics transfer stations based on federated learning includes the following steps: By deploying sensors, RFID, video surveillance and other equipment in the logistics transfer yard and using the logistics management system to collect key data such as vehicle scheduling, personnel operations, equipment operation, and cargo flow, a multi-source heterogeneous data set as shown below is constructed in each logistics transfer yard: (1) Vehicle dispatching dataset: Collect vehicle ID, real-time location, average vehicle load factor, average driving speed, and route complexity (number of turns) in each logistics transfer station and construct the first The vehicle scheduling dataset of logistics transfer stations is denoted as ,in Indicates The first logistics transfer station The collected data of each vehicle, Indicates The number of vehicles in a logistics transfer station.

[0010] (2) Personnel work data set: collect information such as personnel ID, average working hours, task completion efficiency, historical performance scores, etc. in each logistics transfer station, and construct the first The personnel operation data set of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station The data collected by each person, Indicates The number of personnel in a logistics transfer station.

[0011] (3) Equipment operation data set: collect information such as equipment ID, equipment operation time, real-time operation status, cumulative failure time, remaining days of maintenance cycle, etc. in each logistics transfer station, and construct the first The equipment operation data set of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station The collected data of each device, Indicates The number of equipment in a logistics transfer station.

[0012] (4) Cargo flow data set: Collect information such as cargo ID, category, weight, in / out timestamp, destination transfer station ID, etc. in each logistics transfer station, and construct the first The cargo flow data set of logistics transfer stations is recorded as ,in Indicates The first logistics transfer station The collection data of each cargo, Indicates The quantity of goods in a logistics transfer station.

[0013] In an embodiment of the present application, raw data is collected from various data sources (such as IoT sensors, field control systems, enterprise management systems, logistics monitoring platforms, etc.). These data may have different formats and diverse sources, and may contain various types of information such as numerical, textual, and time types. Data preprocessing and cleaning: The collected raw data is formatted, missing values ​​are filled, and outliers are detected to improve data quality. In this process, it is necessary to uniformly define the data from different data sources to ensure that the meaning of each field is consistent during subsequent processing. Format unification and data fusion: Due to the heterogeneity of data sources, it is necessary to map the data from each data source to a predefined unified data structure. Using data fusion technology, the key information in each source data (such as operating indicators, equipment status, environmental information, etc. of the logistics transfer station) is integrated into a feature vector. The generated feature vector can accurately reflect the specific attributes and operating conditions of each logistics transfer station. Construct a local data set: By processing the data after the above steps, a local data set is divided according to the specific business logic of each logistics transfer station. This dataset generally contains multiple feature vectors (each vector represents a sample or the state at a moment) and corresponding category data information (such as resource allocation strategy, module type, etc.), providing raw data support for the subsequent construction of a local training set.

[0014] In order to facilitate the subsequent local model training and parameter upload and download steps, the numerical data in the data set constructed in each transfer station is standardized, the categorical data is one-hot encoded, and the timestamp is converted into time features; The multi-source heterogeneous data such as vehicle dispatch, personnel operation, equipment operation, and cargo flow are standardized, and their core features are extracted and uniformly represented as feature vectors. ,in Indicates The first logistics transfer station feature sample vectors, The total number of feature samples sampled for this transfer field ( =4), each Specifically, it is composed of vehicle, personnel, equipment and cargo data sets extracted and combined through feature engineering, which can be expressed as: in, It is a custom feature engineering module, including feature extraction, normalization, and principal component analysis algorithm. Principal component analysis is a commonly used linear dimensionality reduction technique. Its main purpose is to project high-dimensional data into a low-dimensional space while retaining most of the variation information in the original data as much as possible. It is a classic method.

[0015] is the statistical interval parameter. For the feature extraction part, the specific operations are as follows: For logistics transfer stations, respectively, from their corresponding vehicle scheduling data sets , Personnel Job Dataset , equipment operation data set , cargo flow data set Extract the following features: Normalize the average vehicle load rate, average driving speed, and route complexity index (number of turns) in the vehicle scheduling dataset and form a sub-vector ; Normalize the three indicators of average working hours, task completion efficiency, and historical performance score in the personnel operation data set and form a sub-vector ; Normalize the equipment operation time, average maintenance cycle of a single device, and real-time status ratio (operation / standby / fault) in the equipment operation data set and form a sub-vector The average residence time of goods in the goods flow data set (calculated by the in-and-out timestamp), the category proportion distribution (calculated by the number of recorded categories and the overall number), and the frequency of in-and-out per unit time calculated by the in-and-out timestamp) are normalized and form sub-vectors ; Then splice into a complete feature vector ;Right now Among them, the normalization part is mainly the standardization of numerical data: Min-Max normalization: map the value to a fixed interval according to the minimum and maximum values, the formula is x' = (x - min) / (max - min); In an embodiment of the present application, categorical data can be one-hot encoded: For categorical data, in order to avoid the size order problem caused by directly assigning numerical codes, one-hot encoding technology is used. The specific steps are: Determine the category: For each categorical variable, first determine the set of all possible categories. Code generation: Generate a binary feature for each category in the set. If the category to which a sample belongs matches the current feature, the feature is assigned a value of 1, otherwise it is assigned a value of 0. This encoding method converts categorical data into a numerical vector, so that the machine learning model can process discrete variable information.

[0016] In the embodiment of the present application, the timestamp can also be converted into time features: time information often implies obvious periodicity and trend. In order to better capture the impact of time, the original timestamp is usually converted into multiple time features in the preprocessing stage, including: time decomposition: extracting basic components such as year, month, day, hour, minute, second, etc.; periodic features: for periodic scenarios, such as fluctuations within a day, information such as hours is converted into periodic features using sine and cosine functions; classification features such as weekdays / weekends: judging whether it is a weekday, weekend, holiday, etc. according to the timestamp, so as to capture business characteristics of different time periods; Generate The local data set of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station feature vectors, Indicates The number of data categories in the logistics transfer station, and the local training set is constructed based on the local data set ,in Indicates A set of resource allocation strategies for logistics transfer stations, where Indicates A logistics transfer station for the first A real resource allocation strategy.

[0017] Specifically, the platform makes a set of specific deployment plans for key dispatchable resources including vehicles, manpower, equipment and goods according to the current status and task requirements of the logistics transfer station, including the binding relationship between transport vehicles and routes; the sorting area where the sorting personnel are located; the use and rotation of sorting equipment; the sorting system for the specific allocation of sorted goods, etc., as the supervision signal for the training of the federated learning model. The present invention uses the real resource allocation results corresponding to each sample Standardized encoding is performed, including: 1) For scenarios where the policy set is limited and each scheduling corresponds to a unique policy number, one-hot vector encoding is used, which is suitable for classification model training; 2) For policies composed of a combination of multi-dimensional resource dimensions, multi-dimensional discrete vector representation is used, and each type of resource scheduling result is used as an independent output prediction target; 3) If a certain type of resource has multiple valid choices in the same policy (such as multi-person collaboration, parallel operation of equipment, etc.), a multi-hot encoding structure is used. Through the above structured encoding method, the deep model training and reasoning process of the present invention can be effectively supported, and the learning accuracy and system adaptability of the resource scheduling strategy can be improved.

[0018] Next, a multi-layer perceptron model is used to perform local model training at each logistics transfer station, and the loss function is expressed as: (1) in, For the multi-layer perception model The first logistics transfer station k resource allocation strategy, For the The parameter vector in the logistics transfer station. And the batch size is B The small batch stochastic gradient descent method is used to update the parameters. The update process can be expressed as: (2) in, is the learning rate, is the gradient of a single batch; , Respectively represent Logistics transfer station t The moment and t +1 time instant local model parameters; Then, the central server uses the probability from Randomly select from logistics transfer stations The logistics transfer station is used for training and parameter aggregation. The parameter update process is as follows: Since each transfer field node collects the feature vector Divide into small batches, each batch size is B , used for the gradient descent update process during local model training; for each logistics transfer station, combined with the updated rounds, calculate the iterative update parameters of the selected s-th logistics transfer station ; (3) in, Indicates the selected s-th logistics transfer station The parameters after round iteration, Represents global parameters; Z represents the current iteration round; The central server updates the global parameters. The update process is as follows: (4) in, To consider the data weight of each logistics transfer station The global update amount is calculated as follows: (5) It represents the number of characteristic samples of the sth logistics transfer station among the selected S logistics transfer stations; Let t = t + 1, the central server uses the updated global model parameters as the new , sent to each logistics transfer station for the next round of training, until the set round is reached and the training results are obtained. The global model parameters at this time are recorded as , and distributed to each logistics transfer node.

[0019] Finally, based on the federated learning model, a time span of T The problem of cross-transit resource collaborative allocation optimization: (6) in, Indicates Transfer stations during the period Call The number of class resources, Indicates Transfer stations during the period Call The cost of class resources, To call The priority weight of the class resource; the first constraint is the resource capacity constraint, Indicates Class resources in time period The available capacity of the second constraint is the task demand constraint. Indicates Logistics transfer stations during the period The third constraint is the soft constraint of the federated learning model, which ensures that the number of resource allocations is consistent with the recommended trend of the federated learning model. is the allowable probability deviation range. By solving this problem through integer programming algorithms such as branch and bound method, the optimal resource allocation solution can be obtained. .

[0020] The above is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for collaborative resource allocation across logistics transfer stations based on federated learning, characterized by: The following steps are involved: Construct a collection of multi-source heterogeneous data in each logistics transfer station, and perform feature extraction to obtain the feature vector of each logistics transfer station; Generate local data sets for each logistics transfer station; Build a federated learning model, which is trained collaboratively by various logistics transfer stations and the central server; Based on the federated learning model, we construct a T The collaborative allocation optimization problem of cross-transfer station resources is solved, and the collaborative allocation strategy is obtained by solving the optimization problem.

2. According to claim 1, a method for collaboratively allocating resources across logistics transfer stations based on federated learning is characterized in that: The multi-source heterogeneous data set constructed in each logistics transfer station includes: Construct a vehicle dispatching dataset: Collect vehicle ID, real-time location, average vehicle load factor, average driving speed, and route complexity in each logistics transfer station; build the The vehicle scheduling dataset of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station The collected data of each vehicle, Indicates The number of vehicles in a logistics transfer station; Build the personnel job dataset: Collect the personnel ID, average working hours, task completion efficiency, and historical performance scores in each logistics transfer station; build the The personnel operation data set of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station The data collected by each person, Indicates The number of personnel in a logistics transfer station; Build the device operation data set: Collect the equipment ID, equipment operation time, real-time operation status, accumulated failure time, and remaining days of maintenance cycle in each logistics transfer station; build the The equipment operation data set of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station The collected data of each device, Indicates The number of equipment in a logistics transfer station; Construct cargo flow data set: collect cargo ID, category, weight, in and out timestamp, and destination transfer station ID in each logistics transfer station, and construct the first The cargo flow data set of a logistics transfer station is recorded as ,in Indicates The first logistics transfer station The collection data of each cargo, Indicates The quantity of goods in a logistics transfer station.

3. According to claim 2, a method for collaboratively allocating resources across logistics transfer stations based on federated learning is characterized in that: The feature extraction to obtain the feature vector of each logistics transfer site includes: For the i-th logistics transfer station: Feature extraction is performed on the vehicle dispatching dataset, personnel operation dataset, equipment operation dataset, and cargo flow dataset to obtain feature vectors. ; in Indicates The first logistics transfer station feature sample vectors, The total number of feature samples obtained for the transfer site; Custom feature engineering modules, including feature extraction, normalization or principal component analysis algorithms; is the statistical interval parameter.

4. According to claim 1, a method for collaboratively allocating resources across logistics transfer stations based on federated learning is characterized in that: In the local data sets generated for each logistics transfer station, the local data set of the i-th logistics transfer station is recorded as ,in Indicates The first logistics transfer station feature vectors, Indicates The number of data categories in the logistics transfer station, and the local training set is constructed based on the local data set ,in Indicates A set of resource allocation strategies for logistics transfer stations, where Indicates A logistics transfer station for the first A real resource allocation strategy.

5. According to claim 4, a method for collaboratively allocating resources across logistics transfer stations based on federated learning is characterized in that: Build a federated learning model, which is trained collaboratively by various logistics transfer stations and the central server, including; A1. Each logistics transfer station is used as a node of federated learning. The central server sends the current global model parameters to each logistics transfer station. , the global model adopts a multi-layer perceptron model, and the global model parameters are the parameters of the multi-layer perceptron model; where t represents the tth time period, and t=1 in the initial state. That is , are the initialized model parameters; A2. Each logistics transfer station performs several rounds of small batch training based on local data to update local model parameters; The training process is based on each transfer field i The feature sample vector set As model input, a set of resource allocation strategies obtained from real historical records As a supervisory label; The training goal is to learn a prediction function that maps feature vectors to optimal resource allocation strategies. ,This function achieves model fitting by minimizing the cross entropy loss function, and its local training objective function is: in, represents the policy probability distribution predicted by the model, is the actual strategy label; the overall loss function is （1） in, For the multi-layer perception model The first logistics transfer station k The probability distribution of strategies, For the The parameter vector in the logistics transfer station is used. B The small batch stochastic gradient descent method is used to update the parameters. The update process of the local model parameters is expressed as: （2） in, is the learning rate, is the gradient of a single batch; , Respectively represent Logistics transfer station t The moment and t +1 time instant local model parameters; A3. Each logistics transfer station uploads the updated local model parameters to the central server; A4. The central server calculates the new global model parameters based on the weighted average of the data size of each node. , and distribute it to each node; The central server uses probability from Randomly select from logistics transfer stations The logistics transfer stations are trained and parameter aggregated, and the global parameters are updated. The update process is as follows: Since each transfer field node collects the feature vector Divide into small batches, each batch size is B , used for the gradient descent update process during local model training; for each logistics transfer station, combined with the updated rounds, calculate the selected s Iterative update parameters of logistics transfer stations ; （3） in, Indicates the selected s Logistics transfer station The parameters after round iteration, Represents global parameters; Z represents the current iteration round; （4） in, To consider the data weight of each logistics transfer station The global update amount is calculated as follows: （5） It represents the number of characteristic samples of the sth logistics transfer station among the selected S logistics transfer stations; A5. Let t = t + 1, and the central server uses the updated global model parameters as the new , sent to each logistics transfer station for the next round of training, until the set round is reached and the training results are obtained. The global model parameters at this time are recorded as , and distributed to each logistics transit node.

6. The method for collaboratively allocating resources across logistics transfer stations based on federated learning according to claim 5 is characterized by: Based on the federated learning model, a time span is constructed. T The problem of cross-transit resource collaborative allocation optimization: （6） in, Indicates Transfer stations during the period Call The number of class resources, Indicates Transfer stations during the period Call The cost of class resources, To call The priority weight of the class resource; the first constraint is the resource capacity constraint, Indicates Class resources in time period The available capacity of the second constraint is the task demand constraint. Indicates Logistics transfer stations during the period The third constraint is the soft constraint of the federated learning model, which ensures that the number of resource allocations is consistent with the recommended trend of the federated learning model. Indicates that the parameter at time t is The federated learning model is based on the local dataset The recommended resource allocation is is the permissible probability deviation range; The problem is solved by the branch-and-bound integer programming algorithm to obtain the optimal resource allocation solution. .

Images