Battery replacement recommendation method and system based on multi-target deep learning

Abstract

The invention relates to a battery replacement recommendation method and system based on multi-target deep learning, an electronic device and a storage medium, and the method comprises the steps: employing the basic attributes of a battery, the related data of a battery replacement cabinet, the personal information of a rider, and the interaction record between the rider and the battery, and carrying out the multi-dimensional basic information through a multi-target recommendation system based on deep learning, potential variables in the data are deeply mined, the basic information and the mined hidden variables are integrated into a recommendation system model, and battery replacement cabinet and battery suggestions which most meet the requirements of a rider are output in real time. The problems of low user matching degree and low resource utilization rate of a battery replacement recommendation method in related technologies are solved, the cost and the production pressure are reduced, and the balance of a supply-demand relationship is enhanced.

Description

Technical Field

[0001] The present application relates to the field of machine learning, and in particular to a battery replacement recommendation method, system, computer device, and computer-readable storage medium based on multi-objective deep learning. Background Art

[0002] With the acceleration of urbanization and the rapid development of e-commerce, electric bicycles have become an indispensable means of transportation for industries such as express delivery and food delivery. In order to ensure that the delivery services of these industries can continue to operate efficiently, fast and convenient battery replacement services are particularly important.

[0003] The traditional battery replacement method relies on manual operation at fixed sites, which is not only inefficient but also difficult to meet the growing service needs. In recent years, smart battery swap cabinets have gradually become popular as a new type of infrastructure, allowing riders to replace batteries by themselves, greatly improving convenience and efficiency. The existing battery swap cabinet system still has certain limitations in battery management and distribution, and cannot fully meet the personalized needs of riders. At the same time, due to differences in battery release time, usage environment and other factors, there are differences in performance between batteries. In addition, the number of batteries is limited, and it is impossible to guarantee that every user can get the best performing battery. Summary of the invention

[0004] The embodiments of the present application provide a battery replacement recommendation method, system, computer device and computer-readable storage medium based on multi-objective deep learning, so as to at least solve the problem of low user matching degree of the battery replacement recommendation method in the related art.

[0005] In a first aspect, an embodiment of the present application provides a battery replacement recommendation method based on multi-objective deep learning, which is implemented based on an improved ESMM model. The method includes: Integrate the multi-dimensional original features into shared underlying features, wherein the multi-dimensional original features include: user features, battery features and electrical cabinet features; Based on the shared underlying features, feature representations related to different types of specific tasks are learned through an expert network to obtain multi-dimensional feature learning results, and the multi-dimensional feature learning results are weighted and fused; Through the task tower, predictions are made based on the weighted combined results to obtain battery replacement recommendation results that match user habits.

[0006] In some embodiments, integrating the multi-dimensional original features into a unified feature representation through a feature processing layer includes: Receiving the multi-dimensional original features through an input layer; Through multiple multi-layer perceptrons, the original features of each dimension are converted into high-dimensional feature representations; The multi-dimensional high-dimensional feature representations are integrated to obtain the shared underlying features.

[0007] In some of the embodiments, the expert network includes a plurality of expert layers that are shared in structure but trained independently, and each expert layer is used to focus on learning feature representations related to a specific task; The weight parameters corresponding to each expert layer are generated through the gating network, and all feature representations are weighted and fused according to the weight parameters to obtain the multi-dimensional feature learning result.

[0008] In some embodiments, feature representations related to different types of specific tasks are learned through an expert network, and multi-dimensional feature learning results are obtained, including: Through the first fully connected layer, the shared underlying features are linearly transformed, and the shared underlying features are mapped to a new feature space using an activation function to obtain potential features corresponding to the specific task; Through the second fully connected layer, the latent features are linearly transformed and activated, the feature patterns related to the specific task are strengthened, and the feature learning results are output.

[0009] In some embodiments, weighted fusion of the multi-dimensional feature learning results includes: Obtain feature learning results output by each expert layer through a gating network, process the feature learning results respectively through a fully connected layer, and convert the results output by the fully connected layer into gating signals corresponding to each expert layer through a Softmax function; According to the gating signal, the feature learning results output by each expert layer are weighted and combined to obtain a multi-dimensional feature learning result.

[0010] In some embodiments, the task tower includes a first task tower and a second task tower, wherein the first task tower is used for recommending tasks for the electric cabinet to output a first prediction value, and the second task tower is used for recommending tasks for the battery to output a second prediction value; By combining the first prediction value and the second prediction value, a comprehensive indicator reflecting the performance on the battery cabinet recommendation task is obtained.

[0011] In some embodiments, the prediction results of the task are supervised using cross entropy loss, and the improved ESMM model is optimized and trained with the goal of minimizing the cross entropy between the predicted value and the true value of the cabinet recommendation task and the battery recommendation task, wherein the loss function of the first task tower and the second task tower is expressed by the following formula:

[0012] in, represents the loss function of the cabinet-related tasks, represents the loss function of battery-related tasks, H represents the cross entropy function, To calculate the predicted value and the cross entropy loss between the true value L1, To calculate the predicted value and the cross entropy loss between the true label distribution, represents the i-th training sample, N represents the number of training samples, , They represent the predicted values ​​of the power cabinet recommendation task and the battery recommendation task respectively, and L1 and L2 represent the true values ​​of the power cabinet recommendation task and the battery recommendation task respectively.

[0013] In some of these embodiments, the method further comprises: calculating the uncertainty and accuracy of the cabinet recommended task and the battery recommended task respectively; In the process of optimizing and training the improved ESMM model, the uncertainty and the accuracy are incorporated into the overall loss function of the model to dynamically allocate the weights of the cabinet feature-related tasks and the battery feature-related tasks, wherein the overall function is expressed by the following formula:

[0014] Among them, L is the overall loss function, are the accuracies corresponding to the cabinet feature-related tasks and the battery feature-related tasks, They are the uncertainties corresponding to the tasks related to the electrical cabinet characteristics and the tasks related to the battery characteristics.

[0015] In a second aspect, an embodiment of the present application provides a battery replacement recommendation system based on multi-objective deep learning, characterized in that the system includes: an input module, a feature processing module and a recommendation module, wherein: The input module is used to integrate multi-dimensional original features into shared underlying features, wherein the multi-dimensional original features include: user features, battery features and electrical cabinet features; The feature processing module is used to learn feature representations related to different types of specific tasks through an expert network based on the shared underlying features, obtain multi-dimensional feature learning results, and perform weighted fusion on the multi-dimensional feature learning results; The recommendation module is used to make predictions based on the weighted combination results through the task tower to obtain battery replacement recommendation results that match user habits.

[0016] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect above when executing the computer program.

[0017] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0018] Compared with the traditional method of recommending battery replacement based solely on battery power and fixed rules, the battery replacement recommendation method provided in the embodiment of the present application makes full use of multi-dimensional basic information such as the basic properties of the battery, relevant data of the battery replacement cabinet, the personal information of the rider, and the interaction records between the rider and the battery. Through a multi-objective recommendation system based on deep learning, the potential variables in these data are deeply mined, and these basic information and the mined implicit variables are integrated into the recommendation system model to output the battery replacement cabinet and battery recommendations that best meet the rider's needs in real time. The problems of low user matching and low resource utilization in the battery replacement recommendation method in the related technology are solved, which not only reduces costs and production pressure, but also improves the balance of supply and demand. In addition, the rider-battery interaction information and the hidden variable information of deep learning are included to fully explore the relationship between the rider and the battery. Since the model mines the rider's habitual battery replacement data information and generates recommendation information based on the rider's personal habits and preferences, it can also meet the rider's needs more comprehensively, further alleviating the pressure on the supply side of high-quality batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is a flowchart of a battery replacement recommendation method based on multi-objective deep learning according to an embodiment of the present application; Figure 2 It is an architectural diagram of an improved ESMM model according to an embodiment of the present application; Figure 3 It is a structural block diagram of a battery replacement recommendation system based on multi-objective deep learning according to an embodiment of the present application; Figure 4 It is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0021] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. In addition, it can also be understood that although the efforts made in this development process may be complicated and lengthy, for ordinary technicians in this field related to the content disclosed in this application, some changes in design, manufacturing or production based on the technical content disclosed in this application are just conventional technical means, and should not be understood as insufficient content disclosed in this application.

[0022] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0023] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantity limitation, and may indicate the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0024] With the acceleration of urbanization and the rapid development of e-commerce, electric bicycles have become an indispensable means of transportation for industries such as express delivery and takeout. In order to ensure that the delivery services of these industries can continue to operate efficiently, fast and convenient battery replacement services are particularly important. The traditional battery replacement method relies on manual operation at fixed sites, which is not only inefficient but also difficult to meet the growing service needs. In recent years, smart battery swap cabinets have gradually become popular as a new type of infrastructure, allowing riders to replace batteries by themselves, greatly improving convenience and efficiency. However, the existing battery swap cabinet system still has certain limitations in battery management and distribution, and cannot fully meet the personalized needs of riders. The performance of batteries varies depending on factors such as the time of battery delivery and the environment in which they are used. In addition, the number of batteries is limited, and there is no guarantee that every user can get the best performing battery.

[0025] In the prior art, battery recommendation methods usually adopt random distribution strategies or rely on simple rules, such as: 1. The system randomly selects a battery with the highest power from the battery swap cabinet for recommendation Most battery swapping facilities on the market currently adopt a random distribution strategy, that is, randomly selecting a fully charged battery from the existing ones for users to use. Although this approach is simple and fast, it fails to take into account the unique needs of each rider. For example, riders who perform long-distance delivery tasks are more suitable for using large-capacity batteries, while for those who frequently make short-distance deliveries, carrying lighter batteries may be more ideal. In addition, there are differences in performance between batteries of different brands or models, which makes the random distribution method more likely to produce a mismatch between supply and demand.

[0026] 2. Fixed rule allocation Some improved systems optimize the battery allocation process by introducing basic rules, such as giving priority to the battery with the most remaining power, or determining the priority based on the number of times the battery has been used. Although such an approach can improve the user experience to some extent, it still lacks sufficient flexibility to adapt to diverse actual needs. Especially when serving a large number of users, it is difficult to fully meet various different demand scenarios by relying on a single rule.

[0027] In actual applications, battery recommendation is a more complicated issue. It is not reasonable to rely solely on the power or capacity of the battery to determine whether it is suitable for a certain rider, and it is also impossible to achieve effective use of resources. Therefore, in the operation of the battery swap system, it is necessary to establish a multi-factor comprehensive consideration strategy model to better meet the personalized needs of riders and provide accurate support for battery swap needs.

[0028] In view of this, the embodiment of the present application provides a battery replacement recommendation method based on multi-objective deep learning. Figure 1 is a flowchart of a battery replacement recommendation method based on multi-objective deep learning according to an embodiment of the present application, such as Figure 1 As shown, the process includes the following steps: S101, integrating multi-dimensional original features into shared underlying features, wherein the multi-dimensional original features include: user features, battery features, and electrical cabinet features; Specifically, this step includes the following subdivision steps: Step 1, receiving multi-dimensional original features through the input layer; The original features include: user features, battery features and electrical cabinet features; It can be understood that user characteristics reflect the past behavior patterns of riders, which may be but are not limited to trip length, energy consumption, and recharging cycles; further, battery characteristics include but are not limited to battery capacity, remaining power, and information related to battery type; electrical cabinet characteristics may be electrical cabinet location and distribution information, resource information within the device, and device status information, wherein the device status information includes: operating status, fault status, and maintenance records.

[0029] Step 2, through multiple multi-layer perceptrons, the original features of each dimension are converted into high-dimensional feature representations; Figure 2 is an architecture diagram of an improved ESMM model according to an embodiment of the present application, such as Figure 2 As shown in the figure, user features, battery features and cabinet features are sent to the shared layer (share bottom) as original inputs. The shared layer includes multiple multilayer perceptrons (MLPs for short). Each MLP processes the input features through its internal structure and converts the original features into a higher-level abstract representation to facilitate the understanding of the subsequent deep learning model.

[0030] Specifically, MLP consists of multiple neurons, including an input layer, a hidden layer, and an output layer. After the input layer receives the original feature data, the data undergoes a series of linear transformations and nonlinear activation functions (such as ReLU, etc.) in the hidden layer.

[0031] Among them, the original features often have the characteristics of high dimension, sparsity and complexity, which is not conducive to the model to directly mine useful information. In this embodiment, the conversion of the original features to a higher level of abstract representation enables the subsequent deep learning model to better understand the data. The abstract representation after MLP processing reduces the complexity of the data, highlights the key information, and is more suitable for deep learning model processing.

[0032] For example, in subsequent components such as shared layers and expert networks, the model can more efficiently learn the complex relationships between users, batteries, and power cabinets based on these abstract features, thereby performing power cabinet recommendation (CTR) and battery recommendation (CVR) tasks more accurately.

[0033] Step 3: Integrate the multi-dimensional high-dimensional feature representations to obtain shared underlying features.

[0034] It can be understood that the multi-dimensional high-dimensional feature representations obtained after processing by each MLP are integrated together. This process combines the high-dimensional feature vectors of the three dimensions of user, battery, and cabinet into a unified feature representation, that is, sharing the underlying features.

[0035] In this embodiment, the shared underlying features can fully reflect the complex relationship between users, batteries, and power cabinets. For example, it may include the relationship between the rider's demand for a certain battery performance on a specific delivery route and the possibility of a nearby power cabinet providing the battery. This shared underlying feature will serve as the input of the expert network in the subsequent deep learning model, providing a unified and rich information foundation for the entire battery swap recommendation system, enabling the system to perform more accurate power cabinet recommendation (CTR) and battery recommendation (CVR) tasks based on this.

[0036] S102, based on the shared underlying features, through the expert network, respectively learn the feature representations related to different types of specific tasks to obtain multi-dimensional feature learning results; like Figure 2 As shown in the figure, the expert network (Expert) includes multiple expert layers that are shared in structure but trained independently. Each expert layer is used to focus on learning feature representations related to a specific task. After the "basic feature representation" processed by the multi-layer perceptron (MLP) is generated from the shared bottom layer, it will be distributed to multiple expert layers. Each expert layer focuses on learning the feature representation related to a specific task and performs in-depth analysis and processing based on the specific type of features it is responsible for.

[0037] For example, the first expert layer focuses on processing features related to battery performance, such as battery capacity, remaining power, charging speed, etc.; the second expert layer focuses on analyzing user behavior characteristics, such as trip length, energy consumption, recharging cycle, and historical battery replacement habits; the third expert layer will conduct feature mining and analysis on cabinet-related features, such as cabinet location, number and distribution of batteries in the cabinet, etc.

[0038] It can be understood that in this embodiment, different expert layers focus on learning different task features. This specialized division of labor enables the system to fully understand and mine key information in the data from different angles, providing a rich and targeted feature basis for accurate recommendations.

[0039] In an exemplary embodiment, the specific processing flow of a certain expert layer includes the following steps: Step 1: The expert layer receives the "shared bottom layer features" from the shared bottom layer. These features have been processed in the early stage and integrated into a unified form, including user features, battery features, cabinet features and other aspects of preliminarily abstracted information.

[0040] Step 1: Through the first fully connected layer, the shared underlying features are linearly transformed, and the activation function is used to map the shared underlying features to the new feature space to obtain the potential features corresponding to the specific task; The input features enter the first fully connected layer of the expert network. Each neuron in this fully connected layer is connected to the neurons of the shared bottom layer. By linearly transforming the input features and applying activation functions (such as ReLU, etc.) to introduce nonlinear factors, the shared bottom layer features of the input are mapped to a new feature space, thereby preliminarily mining the potential feature patterns related to the specific type of features that the expert network is responsible for.

[0041] Step 2: Through the second fully connected layer, the features mapped to the new feature space are linearly transformed and activated to strengthen the feature patterns related to the specific task and output the feature learning results; It can be understood that the features processed by the first fully connected layer are further used as the input of the second fully connected layer. The second fully connected layer continues to perform similar linear transformation and activation operations on the features, further strengthening and refining the feature patterns related to specific tasks, making the feature representation more accurately reflect the information related to specific tasks, and completing deeper feature extraction.

[0042] After deep processing by two layers of fully connected layers, each expert layer network outputs a feature representation of fixed dimension. For those skilled in the art, this feature representation is a highly refined and abstracted version of the shared underlying features of the input from a specific task perspective.

[0043] Step 3, obtain the feature learning results output by each expert network through the gating network, and process the feature learning results respectively through the fully connected layer, and convert the results output by the fully connected layer into the gating signals corresponding to each expert network through the Softmax function, where the gating signals are used to indicate how to weight the output results of each expert layer; Among them, the role of the gating network is to determine how the output of each expert layer should affect the final prediction result. Specifically, in this embodiment, two gating layers (Gate1 and Gate2) are set, and the gating layer will generate a series of gating signals (such as G_11, G_12, ..., G_24), which indicate how the output of the expert layer should be weighted and combined, so that when generating prediction tasks such as cabinet recommendation (CTR) and battery recommendation (CVR), the key information extracted by each expert network can be comprehensively considered.

[0044] For example, if an expert network performs well in processing user behavior characteristics and its output feature representation is of great value in reflecting the user's personalized needs for batteries and power cabinets, then the gating network may assign it a higher weight, so that the final recommendation results are more inclined to meet the specific needs of the user.

[0045] It can be understood that through the gating mechanism, the output of the expert network works in conjunction with the gating network to provide a solid foundation for the system to generate accurate and personalized battery swap recommendations, ensuring that the recommendation results can meet the actual needs of the riders to the greatest extent and improve the efficiency and quality of the entire battery swap service.

[0046] Specifically, the gating network first processes the input expert network feature representation through a fully connected layer, and then the results after the fully connected layer are input into the Softmax function. The Softmax function converts these values ​​into weights corresponding to each expert network, and the sum of all weights is 1. These weights reflect the relative importance of each expert network output in the final prediction result.

[0047] Step 4: According to the gating signal, the feature learning results output by each expert layer are weighted and combined to obtain multi-dimensional feature learning results.

[0048] According to the generated weights, the gating network performs a weighted combination of the outputs of each expert layer to obtain a comprehensive feature representation. This comprehensive feature representation integrates the key information extracted from different angles by each expert network, and reflects the importance of different information in the current task according to the weight distribution.

[0049] Through the above step S102, multiple expert networks are used to focus on processing specific feature learning respectively, and the shared features are converted into precise abstract features corresponding to different types of learning tasks, providing a rich information basis for the system. On this basis, the gated network is used to assign weights to the outputs of each expert network to highlight the contribution of important features. The collaboration between the two enables the system to accurately grasp the relationship between users, batteries and electric cabinets. The system accurately analyzes the relationship between users, batteries and electric cabinets, realizes accurate recommendations for electric cabinets and batteries, effectively improves distribution efficiency, improves customer satisfaction, and ensures efficient and reliable operation of the battery replacement recommendation system.

[0050] S103, making predictions based on the weighted combination results of multi-dimensional feature learning results to obtain prediction results for recommending battery replacement.

[0051] It should be noted that the task towers in this embodiment include a first task tower Tower 1 and a second task tower Tower 2; the weighted combination results output by the gated network are input to Task Tower 1 and Task Tower 2, respectively.

[0052] Among them, Task Tower 1 focuses on the cabinet recommendation task (CTR) prediction, and Task Tower 2 performs the battery recommendation task (CVR) prediction; both input the first prediction result and the second prediction result respectively, and the first prediction result and the second prediction result are then multiplied and integrated to form a comprehensive indicator CVCTR; Specifically, in this deep learning network architecture, both CTR and CVR tasks have their own independent task-specific output layers. The output layer is constructed with two fully connected layers. The first fully connected layer performs preliminary feature transformation and integration on the input data, further processes the feature information passed in after weighted merging of the expert layer, and extracts more task-relevant feature representations. The second fully connected layer performs more in-depth calculations and mappings on this basis. The last layer uses the Sigmoid activation function, which has the characteristic of being able to convert any real number input into a probability value between 0 and 1, thereby generating the final prediction value.

[0053] The prediction results of the two task towers are organically combined through multiplication operations to generate a comprehensive index CVCTR, which comprehensively reflects the overall performance of the model in the cabinet battery recommendation task. It should be noted that this comprehensive index is not just a simple combination of results, but considers the synergy and comprehensive matching degree between cabinet recommendation and battery recommendation as a whole, providing a key and effective quantitative basis for comprehensively and accurately evaluating the actual effectiveness of the model in multi-objective recommendation tasks.

[0054] In addition, it should be noted that in this embodiment, in the model training process, the cross entropy loss is used to supervise the prediction results of the task, and the error between the predicted value and the true value of the cabinet recommendation task and the battery recommendation task is minimized to train the improved ESMM model, wherein the loss function of the first task tower and the second task tower is expressed by the following formula:

[0055] in, represents the loss function of the cabinet-related tasks, represents the loss function of battery-related tasks, H represents the cross entropy function, To calculate the predicted value and the cross entropy loss between the true value L1, To calculate the predicted value and the cross entropy loss between the true label distribution, represents the i-th training sample, N represents the number of training samples, , They represent the predicted values ​​of the power cabinet recommendation task and the battery recommendation task respectively, and L1 and L2 represent the true values ​​of the power cabinet recommendation task and the battery recommendation task respectively.

[0056] It should be noted that cross entropy is a measurement method used to measure the difference between two probability distributions. During the model training process, the present application scheme continuously adjusts the model parameters by minimizing the cross entropy loss between the predicted value and the true value, so that the predicted result is closer to the actual situation.

[0057] In addition, in a multi-objective recommendation system, when optimizing the cabinet recommendation (CTR) and battery recommendation (CVR) tasks simultaneously, a negative transfer situation may occur in which the performance of one task improves while the performance of the other task decreases, like a "seesaw" that is difficult to balance at both ends at the same time.

[0058] Taking the above problems into consideration, the technical solution of this application introduces uncertain weight loss calculation, and adjusts the weight of each task in the total loss by comprehensively considering the uncertainty (variance) and precision of the task. When the uncertainty of a task is high (large variance, low precision), its weight in the total loss will be adjusted accordingly, so that the model will not be overly biased towards the optimization of the task, thereby avoiding the performance of another task being damaged due to over-optimization of one task, and effectively alleviating negative transfer and seesaw phenomena.

[0059] Specifically, we first define uncertainty, and then calculate the uncertainty of the first and second tasks respectively. Taking the i-th task as an example, its uncertainty is obtained by taking the logarithm of the variance of the task, and is calculated by the following formula:

[0060] in, is the uncertainty of the ith task, is the variance of the task; Furthermore, based on the uncertainty, the accuracy is calculated. For the i-th task, its accuracy is It is uncertainty The reciprocal of the exponential can be calculated by the following formula:

[0061] in, represents the accuracy of the i-th task.

[0062] Finally, the overall loss L of the model is obtained by adding the losses of the two tasks and the uncertainty-related terms, which can be expressed by the following formula:

[0063] The overall loss function of the model set in the above way is composed of the loss functions of the CTR and CVR tasks. Among them, the cross entropy loss function has a significant penalty on model prediction errors. When the predicted probability differs greatly from the true value, the loss value increases rapidly, prompting the model to avoid wrong predictions. At the same time, its gradient characteristics in the optimization process are conducive to the rapid convergence of the model and can effectively guide the update of model parameters.

[0064] Furthermore, in a multi-task environment, a reasonable combination of CTR and CVR loss functions based on uncertainty can balance the relationship between the cabinet recommendation and battery recommendation tasks, avoiding the "seesaw" phenomenon where only focusing on one task leads to a decline in the performance of the other task. By combining the loss functions of the two tasks, the performance of the model in cabinet and battery recommendation can be evaluated as a whole, allowing the model to provide more accurate recommendation solutions.

[0065] How to select the battery that best suits individual driving habits from nearby battery swap stations to enhance delivery efficiency and customer satisfaction. There are a variety of battery options on the market today, which vary in specifications, capacity, and production batches. In the past, users were usually asked to randomly select a fully charged battery for replacement, but this method often ignores the specific preferences of individual users, resulting in insufficient resource allocation.

[0066] Through the above steps S101 to S103, the present application scheme makes full use of multi-dimensional basic information such as the basic properties of the battery, the relevant data of the battery swap cabinet, the personal information of the rider, and the interaction records between the rider and the battery. Through a multi-objective recommendation system based on deep learning, the potential variables in these data are deeply mined, and these basic information and the mined implicit variables are integrated into the recommendation system model to output the battery swap cabinet and battery recommendations that best meet the rider's needs in real time. It solves the problems of low user matching and low resource utilization in the battery swap recommendation method in the related technology, which not only reduces costs and production pressure, but also improves the balance of supply and demand; in addition, it contains rider-battery interaction information and deep learning hidden variable information to fully explore the relationship between riders and batteries. Since the model mines the rider's habitual battery swap data information and generates recommendation information based on the rider's personal habits and preferences, it can also meet the rider's needs more comprehensively, further alleviating the pressure on the supply side of high-quality batteries.

[0067] In a second aspect, the embodiment of the present application also provides a battery replacement recommendation system based on multi-objective deep learning. Figure 3 is a structural block diagram of a battery replacement recommendation system based on multi-objective deep learning according to an embodiment of the present application, such as Figure 3 As shown, the system includes: an input module 30, a feature processing module 31 and a recommendation module 32, wherein: The input module 30 is used to integrate the multi-dimensional original features into shared underlying features, wherein the multi-dimensional original features include: user features, battery features and electrical cabinet features; Among them, user characteristics reflect the past behavior patterns of riders, which may include but are not limited to trip length, energy consumption, and recharging cycle, etc.; further, battery characteristics include but are not limited to battery capacity, remaining power, and information related to battery type; power cabinet characteristics may include power cabinet location and distribution information, resource information in the device, and device status information, wherein the device status information includes: operating status, fault status, and maintenance records, etc.; In this embodiment, the shared underlying features can fully reflect the complex relationship between users, batteries, and power cabinets. For example, it may include the relationship between the rider's demand for a certain battery performance on a specific delivery route and the possibility of a nearby power cabinet providing the battery. This shared underlying feature will serve as the input of the expert network in the subsequent deep learning model, providing a unified and rich information foundation for the entire battery swap recommendation system, enabling the system to perform more accurate power cabinet recommendation (CTR) and battery recommendation (CVR) tasks based on this.

[0068] The feature processing module 31 is used to learn feature representations related to different types of specific tasks through an expert network based on shared underlying features, obtain multi-dimensional feature learning results, and perform weighted fusion on the multi-dimensional feature learning results; This module uses multiple expert networks to focus on specific feature learning, converting shared features into precise abstract features corresponding to different types of learning tasks, providing a rich information basis for the system. On this basis, the gated network is used to assign weights to the outputs of each expert network to highlight the contribution of important features. The collaboration between the two enables the system to accurately grasp the relationship between users, batteries and electric cabinets. The system accurately analyzes the relationship between users, batteries and electric cabinets, and achieves accurate recommendations for electric cabinets and batteries, effectively improving distribution efficiency, increasing customer satisfaction, and ensuring efficient and reliable operation of the battery replacement recommendation system. The recommendation module 32 is used to make predictions based on the weighted combined results through the task tower to obtain battery replacement recommendation results that match user habits.

[0069] It should be noted that the task towers in this embodiment include the first task tower Tower1 and the second task tower Tower2; the weighted merged results output by the gated network are input to Task Tower1 and Task Tower2 respectively. Among them, Task Tower1 focuses on the cabinet recommendation task (CTR) prediction, and Task Tower2 performs the battery recommendation task (CVR) prediction; the first prediction result and the second prediction result are input to the two respectively, and the first prediction result and the second prediction result are then multiplied and integrated to form the comprehensive index CVCTR; This indicator comprehensively reflects the overall performance of the model in the task of recommending batteries for power cabinets. It should be noted that this comprehensive indicator is not just a simple combination of results, but considers the synergy and comprehensive matching degree between power cabinet recommendation and battery recommendation as a whole, providing a key and effective quantitative basis for comprehensively and accurately evaluating the actual effectiveness of the model in multi-objective recommendation tasks.

[0070] Through the above system, we use multi-dimensional basic information such as the basic properties of the battery, the relevant data of the battery swap cabinet, the personal information of the rider, and the interaction records between the rider and the battery. Through the multi-objective recommendation system based on deep learning, we deeply mine the potential variables in these data, and integrate these basic information and the mined implicit variables into the recommendation system model, and output the battery swap cabinet and battery recommendations that best meet the rider's needs in real time. This solves the problems of low user matching and low resource utilization in the battery swap recommendation method in related technologies, which not only reduces costs and production pressure, but also improves the balance of supply and demand.

[0071] In one embodiment, Figure 4 is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present application, such as Figure 4 As shown, an electronic device is provided, which may be a server, and its internal structure diagram may be as shown in Figure 4 As shown. The electronic device includes a processor, a network interface, an internal memory and a non-volatile memory connected by an internal bus, wherein the non-volatile memory stores an operating system, a computer program and a database. The processor is used to provide computing and control capabilities, the network interface is used to communicate with an external terminal through a network connection, the internal memory is used to provide an environment for the operation of the operating system, the computer program is executed by the processor to implement a battery replacement recommendation method based on multi-objective deep learning, and the database is used to store data.

[0072] Those skilled in the art will understand that Figure 4 The structure shown in the figure is merely a block diagram of a partial structure related to the scheme of the present application, and does not constitute a limitation on the electronic device to which the scheme of the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0073] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0074] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A battery replacement recommendation method based on multi-objective deep learning, characterized in that: Based on the improved ESMM model, the method includes: Integrate the multi-dimensional original features into shared underlying features, wherein the multi-dimensional original features include: user features, battery features and electrical cabinet features; Based on the shared underlying features, feature representations related to different types of specific tasks are learned through an expert network to obtain multi-dimensional feature learning results, and the multi-dimensional feature learning results are weighted and fused; Through the task tower, predictions are made based on the weighted combined results to obtain battery replacement recommendation results that match user habits.

2. The method according to claim 1, characterized in that Integrating multi-dimensional original features into a unified feature representation includes: Receiving the multi-dimensional original features through an input layer; Through multiple multi-layer perceptrons, the original features of each dimension are converted into high-dimensional feature representations; The multi-dimensional high-dimensional feature representations are integrated to obtain the shared underlying features.

3. The method according to claim 1, characterized in that The expert network includes a plurality of expert layers that are shared in structure but trained independently, and each expert layer is used to focus on learning feature representations related to a specific task; The weight parameters corresponding to each expert layer are generated through the gating network, and all feature representations are weighted and fused according to the weight parameters to obtain the multi-dimensional feature learning result.

4. The method according to claim 3, characterized in that Through the expert network, feature representations related to different types of specific tasks are learned respectively, and the multi-dimensional feature learning results are obtained, including: Through the first fully connected layer, the shared underlying features are linearly transformed, and the shared underlying features are mapped to a new feature space using an activation function to obtain potential features corresponding to the specific task; Through the second fully connected layer, the latent features are linearly transformed and activated, the feature patterns related to the specific task are strengthened, and the feature learning results are output.

5. The method according to claim 4, characterized in that The weighted fusion of the multi-dimensional feature learning results includes: Obtain feature learning results output by each expert layer through a gating network, process the feature learning results respectively through a fully connected layer, and convert the results output by the fully connected layer into gating signals corresponding to each expert layer through a Softmax function; According to the gating signal, the feature learning results output by each expert layer are weighted and combined to obtain a multi-dimensional feature learning result.

6. The method according to claim 1, characterized in that The task tower includes a first task tower and a second task tower, wherein the first task tower is used for recommending tasks for the electric cabinet to output a first prediction value, and the second task tower is used for recommending tasks for the battery to output a second prediction value; By combining the first prediction value and the second prediction value, a comprehensive indicator reflecting the performance on the battery cabinet recommendation task is obtained.

7. The method according to claim 5, characterized in that The prediction results of the task are supervised by using cross entropy loss, and the improved ESMM model is optimized and trained with the goal of minimizing the cross entropy between the predicted value and the true value of the cabinet recommendation task and the battery recommendation task, wherein the loss function of the first task tower and the second task tower is expressed by the following formula: in, represents the loss function of the cabinet-related tasks, represents the loss function of battery-related tasks, H represents the cross entropy function, To calculate the predicted value and the cross entropy loss between the true value L1, To calculate the predicted value and the cross entropy loss between the true label distribution, represents the i-th training sample, N represents the number of training samples, , They represent the predicted values ​​of the power cabinet recommendation task and the battery recommendation task respectively, and L1 and L2 represent the true values ​​of the power cabinet recommendation task and the battery recommendation task respectively.

8. The method according to claim 7, characterized in that The method further comprises: Calculating the uncertainty and accuracy of the electric cabinet recommended task and the battery recommended task respectively; In the process of optimizing and training the improved ESMM model, the uncertainty and the accuracy are incorporated into the overall loss function of the model to dynamically allocate the weights of the cabinet feature-related tasks and the battery feature-related tasks, wherein the overall function is expressed by the following formula: Among them, L is the overall loss function, are the accuracies corresponding to the cabinet feature-related tasks and the battery feature-related tasks, They are the uncertainties corresponding to the tasks related to the electrical cabinet characteristics and the tasks related to the battery characteristics.

9. A battery replacement recommendation system based on multi-objective deep learning, characterized in that: include: The system comprises: an input module, a feature processing module and a recommendation module, wherein: The input module is used to integrate multi-dimensional original features into shared underlying features, wherein the multi-dimensional original features include: user features, battery features and electrical cabinet features; The feature processing module is used to learn feature representations related to different types of specific tasks through an expert network based on the shared underlying features, obtain multi-dimensional feature learning results, and perform weighted fusion on the multi-dimensional feature learning results; The recommendation module is used to make predictions based on the weighted combination results through the task tower to obtain battery replacement recommendation results that match user habits.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

Images