Neural network model adaptive splitting method suitable for autonomous mobile platform

Abstract

The invention relates to the technical field of splitting learning, in particular to a neural network model adaptive splitting method suitable for an autonomous mobile platform. According to the method, firstly, a self-adaptive splitting method is provided, and the method comprises the steps of data collection, data correction and decision splitting. According to the method, the time, the energy consumption, the CPU utilization rate, the memory utilization rate and the GPU utilization rate are considered to jointly determine the splitting layer, the computing resources of the autonomous mobile platform can be effectively utilized, and the computing load reduction and the energy consumption reduction of an edge server are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of split learning, and particularly relates to a neural network model adaptive splitting method applicable to an autonomous mobile platform. Background Art

[0002] With the emergence of big data and artificial intelligence, machine learning has become an important tool for data analysis and decision-making in various fields. Traditional machine learning methods rely on centralized data storage and processing, which may lead to privacy leakage and data disclosure. In addition, traditional machine learning methods offload the training task of the neural network model to the cloud server, which consumes a large amount of computing resources of the cloud server and results in high latency of computing tasks. Especially in autonomous mobile platforms, due to the generation of a large amount of data and unreliable communication, it is difficult to train a centralized neural network model. For the problem of unreliable communication, edge computing can effectively alleviate this problem. In recent years, neural network models have become increasingly large. Autonomous mobile platforms have powerful data acquisition capabilities by virtue of the cameras and sensors they carry. However, autonomous mobile platforms generally carry low-performance devices, and the computing resources of these devices are very limited and cannot train neural network models locally alone. An effective way to overcome the resource limitations of terminal devices is to compress the neural network model into a lightweight version, thereby further reducing the amount of computation and the total memory size. The neural network model after compression and pruning can be deployed in low-performance devices in autonomous mobile platforms. However, although model compression can meet the deployment in low-performance devices, it will inevitably reduce the accuracy of classification or detection. Some other researchers have proposed a new machine learning method called split learning (SL). Split learning can divide the model into several sub-models, and the neural network sub-models can be deployed in autonomous mobile platforms that meet their own computing resources. Appropriate machine learning strategies can significantly improve the reliability and practicality of mobile autonomous platforms. Split learning can utilize the computing resources of idle terminal devices to avoid wasting precious computing resources. Using split learning can effectively utilize the computing resources of the terminal, which is of great help in reducing the computing load of the edge server and saving energy consumption.

[0003] The main deficiency of SL is how to efficiently determine the split layer of the neural network model for an autonomous mobile platform with heterogeneous computing resources to adapt to the change of computing resources over time in the autonomous mobile platform and improve the generalization performance. Summary of the Invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a neural network model adaptive splitting method applicable to an autonomous mobile platform, and the specific technical solution adopted is as follows:

[0005] Collect information on each layer of the neural network model to be split in the terminal device and the edge server, including computing time, energy consumption, CPU usage rate, memory usage rate, and GPU usage rate; in each terminal device and edge server, sample the dataset used for the training task of the neural network model to be split to obtain a small sample dataset, and use the small sample dataset to perform calculations in the neural network model to be split, and record the five elements of computing resources required for each layer in the neural network model to be split;

[0006] Construct a decision feedback neural network model, and according to the decision feedback neural network model, correct the collected five elements of computing resources to form a new group of five elements of resources;

[0007] Input the new group of five elements of resources into the decision-making method, which consists of the entropy value method and the comprehensive evaluation method, and output the optimal split layer;

[0008] Send the optimal split layer to the corresponding terminal device and edge server. According to the optimal split layer, divide the neural network model to be split into a first sub-model and a second sub-model, and deploy them on the terminal device and the edge server respectively. The training process is as follows: the forward propagation of the first sub-model starts training, and the output data is sent to the second sub-model. The second sub-model performs forward propagation training, obtains the loss and gradient, and updates the weights of the second sub-model and the first sub-model until the specified number of training times or the accuracy threshold is reached.

[0009] Preferably, the construction of the decision feedback neural network model includes:

[0010] The structure of the decision feedback neural network model is an input layer, a hidden layer, and an output layer;

[0011] Among them, the hidden layer includes batch normalization, fully connected layer, activation function, and fully connected layer.

[0012] Preferably, the construction of the decision feedback neural network model further includes:

[0013] Train the decision feedback neural network model according to the training dataset of the decision feedback neural network model, including five neural network loss models regarding time, energy consumption, memory usage rate, CPU usage rate, and GPU usage rate.

[0014] Preferably, the recording of the five elements of computing resources required for each layer in the neural network model to be split includes:

[0015] Deploy a data acquisition model, and through the data acquisition model, obtain the computing resources required for each layer in the neural network model to be split, and form a vector group from the collected data, which is called a five-tuple. The five-tuple includes: time, energy consumption, CPU usage rate, memory usage rate, and GPU usage rate.

[0016] Preferably, inputting the five resource elements into the decision-making method to output the optimal split layer includes:

[0017] (1.1) Determine the type of the neural network model to be split, whether it is a serial model or a parallel model, and determine the split point positions {1, 2,..., L} of the neural network model to be split;

[0018] (1.2) For the five elements {T, EC, CPU, MEM, GPU} after correction, use all available split points of the neural network to calculate the matrix τ of the required computing resources, and then construct the normalized matrix τ′, as shown in formulas (1) and (2):

[0019]

[0020]

[0021] Among them, T′1 is the normalized value of the time required for the first layer of the split layer; GPU′1 is the normalized value of the GPU usage rate required for the first layer of the split layer; T′2 is the normalized value of the time required for the second layer of the split layer; GPU′2 is the normalized value of the GPU usage rate required for the second layer of the split layer; T′ L is the normalized value of the time required for the Lth layer of the split layer; GPU′ L is the normalized value of the GPU usage rate required for the Lth layer of the split layer; T′ i is the normalized value of the time required for the ith layer of the split layer; EC′ i is the normalized value of the energy consumption required for the ith layer of the split layer; CPU′ i is the normalized value of the CPU usage rate required for the ith layer of the split layer; MEM′ i is the normalized value of the memory usage rate required for the ith layer of the split layer; GPU′ i is the normalized value of the GPU usage rate required for the ith layer of the split layer; T i is the time required for the ith layer of the split layer, EC i is the energy consumption required for the ith layer of the split layer, CPU i is the CPU usage rate required for the ith layer of the split layer, MEM i is the memory usage rate required for the ith layer of the split layer; GPU iThe GPU usage rate required for splitting the i-th layer; min( ) is the minimum value function; max( ) is the maximum value function; min(T) is the minimum value of the time required for splitting the layer; max(T) is the maximum value of the time required for splitting the layer; min(EC) is the minimum value of the energy consumption required for splitting the layer; max(EC) is the maximum value of the energy consumption required for splitting the layer; min(CPU) is the minimum value of the CPU usage rate required for splitting the layer; max(CPU) is the maximum value of the CPU usage rate required for splitting the layer; min(MEM) is the minimum value of the memory usage rate required for splitting the layer; max(MEM) is the maximum value of the memory usage rate required for splitting the layer; min(GPU) is the minimum value of the GPU usage rate required for splitting the layer; max(GPU) is the maximum value of the GPU usage rate required for splitting the layer;

[0022] (1.3) Obtain the weight w of the five elements in the decision-making method, as shown in Formulas (3) to (5):

[0023]

[0024]

[0025]

[0026] where w j represents the corresponding weight of the five elements; p ij represents the proportion of the i-th sample under the j-th index in this index; e j represents the entropy value of the j-th index; e i is the entropy value of the i-th index; i represents the set of split points {1, 2,..., L}; j represents the set of five elements {1, 2, 3, 4, 5}; n is the number of data features collected, which is equal to 5 here;

[0027] (1.4) Construct the weight decision matrix Z ij , as shown in Formula (6):

[0028] Z ij = w j ×τ′ (6);

[0029] i represents the set of split points {1, 2,..., L}, and j represents the set of five elements {1, 2, 3, 4, 5};

[0030] (1.5) Calculate the comprehensive evaluation score S of each split layer i , as shown in Formulas (7) to (11):

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] Among them, Z + represents selecting the maximum value in each column to form a vector as the optimal solution, and Z - represents selecting the minimum value in each column to form a vector as the worst solution. represents calculating the distance from each split layer to the optimal solution. represents calculating the distance from each split layer to the worst solution, and S i represents the comprehensive evaluation score of each split layer; n represents the number of data feature collected, which is equal to 5 here.

[0037] (1.6) Determine the split layer i = max(S i ) between the neural network model to be split on the terminal device and the edge server, (i = 1, 2,..., L).

[0038] The embodiments of the present invention have at least the following beneficial effects:

[0039] The present invention utilizes the advantages of edge computing and split learning. In an unstable communication environment, an edge server deployed closer to the terminal device is used to improve communication stability and efficiency. Split learning can divide the neural network model to be split into several sub-models according to the split points. As an autonomous mobile platform with weak terminals, the sub-models can be deployed on these weak terminals to realize the collaborative training of the neural network model to be split between the terminal device and the edge server. This method not only utilizes the computing resources of idle terminal devices, but also reduces the computing load of the edge server and saves energy consumption. The present invention designs a decision feedback neural network model to correct the collected and estimated information, improve the accuracy of the information, and further improve the effectiveness of the decision-making. The decision model algorithm designed by the present invention can adaptively determine the split layer of the neural network model to be split according to the neural network model to be trained and the device on an autonomous mobile platform with heterogeneous computing resources. This algorithm has good generalization performance for various neural network models and autonomous mobile platforms. Description of the Drawings

[0040] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0041] Figure 1 The flowchart of steps of a neural network model adaptive splitting method applicable to an autonomous mobile platform provided by an embodiment of the present invention;

[0042] Figure 2 The flowchart of a decision feedback neural network model provided by an embodiment of the present invention;

[0043] Figure 3 The schematic diagram of an adaptive splitting learning system provided by an embodiment of the present invention. Detailed implementation manners

[0044] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in combination with the drawings and preferred embodiments, details the specific implementation manners, structures, features and their effects of a neural network model adaptive splitting method applicable to an autonomous mobile platform proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0046] The following specifically describes the specific solution of a neural network model adaptive splitting method applicable to an autonomous mobile platform provided by the present invention in combination with the drawings.

[0047] Please refer to Figure 1 , which shows the flowchart of steps of a neural network model adaptive splitting method applicable to an autonomous mobile platform provided by an embodiment of the present invention. The method includes the following steps:

[0048] Step S100: Collect information of each layer of the neural network model to be split in the terminal device and the edge server, including computing time, energy consumption, CPU usage rate, memory usage rate, and GPU usage rate; in each terminal device and edge server, sample the dataset used for the training task of the neural network model to be split to obtain a small sample dataset, and use the small sample dataset to perform calculations in the neural network model to be split, and record the five elements of computing resources required for each layer in the neural network model to be split.

[0049] With the rapid development of artificial intelligence, there are more and more studies on complex large models. Large models consume a large amount of energy and computing resources during training. Some researchers use edge servers and autonomous mobile platforms to cooperate in training large models. However, for numerous automatic mobile platform devices with heterogeneous computing resources, how to efficiently split the neural network model to effectively balance the training time and energy consumption is an important issue. The embodiment of the present invention proposes an adaptive splitting model method, including data collection, data correction, and decision-making on split layers. It considers five parts: time, energy consumption, CPU usage rate, memory usage rate, and GPU usage rate to jointly determine the split layer, which can effectively utilize the computing resources of the autonomous mobile platform and greatly improve the reduction of the computing load of the edge server and energy conservation.

[0050] (1) Determine the split points of the neural network model to be split.

[0051] The neural network model to be split can be divided into a serial structure and a parallel structure, and the split points of each structure are different. For the neural network model to be split with a parallel structure, the parallel structure part needs to be divided into the same split layer.

[0052] (2) Make a feedback model dataset.

[0053] Collect information on the five elements to be corrected. The five elements to be corrected include: time, energy consumption, CPU usage rate, memory usage rate, GPU usage rate; and obtain the actual required five-element values through experiments. In the embodiment of the present invention, the dataset includes 23,445 pieces of data, and the dataset contains 8 labels, namely: time, energy consumption, memory usage rate, CPU usage rate, GPU usage rate, time loss, energy consumption loss, memory usage rate loss, CPU usage rate loss, GPU usage rate loss; among them, the loss value is obtained by subtracting the actual value from the collected value. The embodiment of the present invention trains five loss models for time, energy consumption, memory usage rate, CPU usage rate, and GPU usage rate respectively through a Deep Neural Network (DNN) model.

[0054] (3) Decision information collection.

[0055] Deploy the acquisition method on the autonomous mobile platform and the edge server respectively, sample the dataset used by the neural network model to be split, and obtain a small sample dataset. Use the small sample dataset to perform a single training in the neural network model to be split. At the same time, record the five elements of the computing resources required for each layer in the neural network model to be split, and obtain the five elements on the autonomous mobile platform and the edge server respectively. Obtain the computing resources required for each layer in the neural network model to be split through the data acquisition model, and form a vector group from the collected data, which is called a five-tuple {time, energy consumption, CPU usage, memory usage, GPU usage}. That is, deploy the data acquisition model, obtain the computing resources required for each layer in the neural network model to be split through the data acquisition model, and form a vector group from the collected data, which is called a five-tuple. The five-tuple includes: time, energy consumption, CPU usage, memory usage, GPU usage.

[0056] (4) Network model adaptive splitting method in the autonomous mobile platform.

[0057] Generally speaking, machine learning methods rely on centralized data storage and processing. Traditional machine learning methods offload the training task of the neural network model to the cloud server, which will consume a large amount of computing resources of the cloud server and lead to high latency of computing tasks. Especially in the autonomous mobile platform, due to the generation of a large amount of data and unreliable communication, it is difficult to train the centralized neural network model. For the problem of unreliable communication, edge computing can effectively alleviate it. The autonomous mobile platform has a powerful data acquisition ability with the help of the equipped cameras and sensors. However, the autonomous mobile platform generally carries weak-performance devices, and the computing resources of these devices are very limited and cannot train the neural network model locally alone. According to a new machine learning method proposed, called split learning, split learning can divide the neural network model to be split into several sub-models.

[0058] The sub-model can be deployed in the autonomous mobile platform that meets the computing resources. Appropriate machine learning strategies can significantly improve the reliability and practicality of the mobile autonomous platform. Split learning can utilize the computing resources of idle terminal devices to avoid wasting precious computing resources. Using the split learning method can effectively utilize the computing resources of the terminal, which is very helpful for reducing the computing load of the edge server and saving energy consumption. Therefore, the present invention proposes a method for adaptively splitting the neural network model to improve the utilization of the computing resources of the terminal device. Considering the heterogeneity of the terminal device and different neural network models, make full use of the computing capabilities of the mobile autonomous platform and the edge server. The embodiment of the present invention can quickly determine the splitting point of the neural network model to be split, and at this splitting point, the resource balance between the mobile autonomous platform and the edge server can be effectively achieved during the neural network training process. Please refer toFigure 3 , Figure 3 is a schematic diagram of an adaptive splitting learning system.

[0059] Step S200: Construct a decision feedback neural network model, and according to the decision feedback neural network model, perform data correction on the five elements of the collected computing resources to form a new group of five resource elements.

[0060] Due to the errors generated in data collection and later data processing, the data is corrected by using the completed and trained decision feedback neural network model. The structure of the decision feedback neural network model is an input layer, a hidden layer, and an output layer. Among them, the hidden layer includes batch normalization, a fully connected layer, an activation function, and a fully connected layer. In the embodiment of the present invention, the Relu algorithm is used to determine the activation function. Please refer to Figure 2 , Figure 2 is a flowchart of the decision feedback neural network model.

[0061] Train the decision feedback neural network model according to the training data set of the decision feedback neural network model, including five neural network loss models regarding time, energy consumption, memory usage, CPU usage, and GPU usage.

[0062] Step S300: Input the new group of five resource elements into the decision-making method, which is composed of the entropy value method and the comprehensive evaluation method, and output the optimal splitting layer.

[0063] The embodiment of the present invention proposes a decision algorithm to determine the optimal splitting layer. The specific steps of this method are as follows:

[0064] (1.1) Determine the type of the neural network model to be split, whether it is a serial model or a parallel model, and determine the split point positions {1, 2,..., L} of the neural network model to be split.

[0065] (1.2) For the five elements {T, EC, CPU, MEM, GPU} after correction, use all available split points of the neural network to calculate the required computing resources to form a matrix τ, and then construct a normalized matrix τ′, as shown in formulas (1) and (2):

[0066]

[0067]

[0068] Among them, T′1 is the normalized value of the time required for the first layer of the split layer; GPU′1 is the normalized value of the GPU usage required for the first layer of the split layer; T′2 is the normalized value of the time required for the second layer of the split layer; GPU′2 is the normalized value of the GPU usage required for the second layer of the split layer; T′ LThe normalized value of the time required to split layer L; GPU' L The normalized value of the GPU usage rate required to split layer L; T' i The normalized value of the time required to split layer i; EC' i The normalized value of the energy consumption required to split layer i; CPU' i The normalized value of the CPU usage rate required to split layer i; MEM' i The normalized value of the memory usage rate required to split layer i; GPU' i The normalized value of the GPU usage rate required to split layer i; T i The time required to split layer i; EC i The energy consumption required to split layer i; CPU i The CPU usage rate required to split layer i; MEM i The memory usage rate required to split layer i; GPU i The GPU usage rate required to split layer i; min() is the minimum value function; max() is the maximum value function; min(T) is the minimum value of the time required for splitting; max(T) is the maximum value of the time required for splitting; min(EC) is the minimum value of the energy consumption required for splitting; max(EC) is the maximum value of the energy consumption required for splitting; min(CPU) is the minimum value of the CPU usage rate required for splitting; max(CPU) is the maximum value of the CPU usage rate required for splitting; min(MEM) is the minimum value of the memory usage rate required for splitting; max(MEM) is the maximum value of the memory usage rate required for splitting; min(GPU) is the minimum value of the GPU usage rate required for splitting; max(GPU) is the maximum value of the GPU usage rate required for splitting.

[0069] (1.3) Obtain the weight w of the five elements in the decision-making method, as shown in formulas (3) to (5):

[0070]

[0071]

[0072]

[0073] Among them, w j Represents the corresponding weight of the five elements; p ij Represents the proportion of the i-th sample under the j-th index in this index; e j Represents the entropy value of the j-th index; e iis the entropy value of the i-th index; i represents the set of split points {1, 2, …, L}; j represents the set of five elements {1, 2, 3, 4, 5}; ln is the natural logarithm; n represents the number of data features collected, which is 5 in the embodiments of the present invention; it should be noted that the j-th index here is the j-th element in the five elements.

[0074] (1.4) Construct the weight decision matrix Z ij , as shown in formula (6):

[0075] Z ij = w j × τ′ (6);

[0076] I represents the set of split points {1, 2, …, L}, and J represents the set of five elements {1, 2, 3, 4, 5};

[0077] (1.5) Calculate the comprehensive evaluation score S of each split layer i , as shown in formulas (7) to (11):

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] Among them, Z + represents selecting the maximum value in each column to form a vector as the optimal solution, and Z - represents selecting the minimum value in each column to form a vector as the worst solution, represents calculating the distance from each split layer to the optimal solution, represents calculating the distance from each split layer to the worst solution, and S i represents the comprehensive evaluation score of each split layer; n represents the number of data features collected, which is equal to 5 here;

[0084] (1.6) Determine the split layer i between the terminal device and the edge server for the neural network model to be split as i = max(S i ), (i = 1, 2, …, L).

[0085] Step S400: Send the optimal split layer to the corresponding terminal device and edge server. According to the optimal split layer, divide the neural network model to be split into a first sub-model and a second sub-model, and deploy them on the terminal device and the edge server respectively. The training process is as follows: The forward propagation of the first sub-model starts training, sends the output data to the second sub-model, the second sub-model performs forward propagation for training, obtains the loss and gradient, and updates the weights of the second sub-model and the first sub-model until the specified number of training times or the accuracy threshold is reached.

[0086] It should be noted that the above sequence of embodiments of the present invention is only for description and does not represent the advantages or disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0087] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the differences between each embodiment and other embodiments are emphasized.

Claims

1. A neural network model adaptive splitting method applicable to an autonomous mobile platform, characterized in that, The method includes the following steps: Collect information of each layer of the neural network model to be split in the terminal device and the edge server, including computing time, energy consumption, CPU usage rate, memory usage rate, and GPU usage rate; in each terminal device and edge server, sample the dataset used for the training task of the neural network model to be split to obtain a small sample dataset, and use the small sample dataset to perform calculations in the neural network model to be split, and record the five elements of computing resources required for each layer in the neural network model to be split; Construct a decision feedback neural network model, and correct the collected five elements of computing resources according to the decision feedback neural network model to form a new group of five elements of resources; Input the new group of five elements of resources into the decision-making method, which consists of the entropy value method and the comprehensive evaluation method, and output the optimal split layer; Send the optimal split layer to the corresponding terminal device and edge server. According to the optimal split layer, divide the neural network model to be split into a first sub-model and a second sub-model, and deploy them on the terminal device and the edge server respectively. The training process is as follows: the forward propagation of the first sub-model starts training, and the output data is sent to the second sub-model. The second sub-model performs forward propagation training, obtains the loss and gradient, and updates the weights of the second sub-model and the first sub-model until the specified number of training times or the accuracy threshold is reached.

2. The adaptive splitting method of a neural network model applicable to an autonomous mobile platform according to claim 1, characterized in that The construction of the decision feedback neural network model includes: The structure of the decision feedback neural network model is an input layer, a hidden layer, and an output layer; Among them, the hidden layer includes batch normalization, a fully connected layer, an activation function, and a fully connected layer.

3. The adaptive splitting method of a neural network model applicable to an autonomous mobile platform according to claim 1, wherein The construction of the decision feedback neural network model also includes: Train the decision feedback neural network model according to the training dataset of the decision feedback neural network model, including five neural network loss models regarding time, energy consumption, memory usage rate, CPU usage rate, and GPU usage rate.

4. A method for adaptively splitting a neural network model applicable to an autonomous mobile platform according to claim 1, characterized in that, The recording of the five elements of computing resources required for each layer in the neural network model to be split includes: Deploy a data collection model, and obtain the computing resources required for each layer in the neural network model to be split through the data collection model. The collected data is composed of a vector group, called a five-tuple, and the five-tuple includes: time, energy consumption, CPU usage rate, memory usage rate, and GPU usage rate.

5. A neural network model adaptive splitting method applicable to an autonomous mobile platform according to claim 1, characterized in that, The input of the five elements of resources into the decision-making method to output the optimal split layer includes: (1.1) Determine the type of the neural network model to be split, whether it is a serial model or a parallel model, and determine the positions of the split points {1, 2,..., L} of the neural network model to be split; (1.2) For the five elements {T, EC, CPU, MEM, GPU} after correction, use all available split points of the neural network to calculate the required computing resources to form a matrix τ, and then construct a normalized matrix τ′, as shown in formulas (1) and (2): Among them, T′1 is the normalized value of the time required for the first layer of the split layer; GPU′1 is the normalized value of the GPU usage rate required for the first layer of the split layer; T′2 is the normalized value of the time required for the second layer of the split layer; GPU′2 is the normalized value of the GPU usage rate required for the second layer of the split layer; T′ L is the normalized value of the time required for the L-th layer of the split layer; GPU′ L is the normalized value of the GPU usage rate required for the L-th layer of the split layer; T′ i is the normalized value of the time required for the i-th layer of the split layer; EC′ i is the normalized value of the energy consumption required for the i-th layer of the split layer; CPU′ i is the normalized value of the CPU usage rate required for the i-th layer of the split layer; MEM′ i is the normalized value of the memory usage rate required for the i-th layer of the split layer; GPU′ i is the normalized value of the GPU usage rate required for the i-th layer of the split layer; T i is the time required for the i-th layer of the split layer, EC i is the energy consumption required for the i-th layer of the split layer, CPU i is the CPU usage rate required for the i-th layer of the split layer, MEM i is the memory usage rate required for the i-th layer of the split layer; GPU i is the GPU usage rate required for the i-th layer of the split layer; min() is the minimum value function; max() is the maximum value function; min(T) is the minimum value of the time required for the split layer; max(T) is the maximum value of the time required for the split layer; min(EC) is the minimum value of the energy consumption required for the split layer; max(EC) is the maximum value of the energy consumption required for the split layer; min(CPU) is the minimum value of the CPU usage rate required for the split layer; max(CPU) is the maximum value of the CPU usage rate required for the split layer; min(MEM) is the minimum value of the memory usage rate required for the split layer; max(MEM) is the maximum value of the memory usage rate required for the split layer; min(GPU) is the minimum value of the GPU usage rate required for the split layer; max(GPU) is the maximum value of the GPU usage rate required for the split layer; (1.3) Obtain the weights w of the five elements in the decision-making method, as shown in formulas (3) to (5): Among them, w j represents the corresponding five-element weight; p ij represents the proportion of the i-th sample under the j-th index in this index; e j represents the entropy value of the j-th index; e i is the entropy value of the i-th index; i represents the set of split points {1, 2, …, L}; j represents the set of five elements {1, 2, 3, 4, 5}; n is the number of data features collected, which is equal to 5 here; (1.4) Construct the weighted decision matrix Z ij , as shown in formula (6): Z ij = w j × τ′ (6); i represents the set of split points {1, 2,..., L}, and j represents the set of five elements {1, 2, 3, 4, 5}; (1.5) Calculate the comprehensive evaluation score S of each split layer i , as shown in formulas (7) to (11): Among them, Z + represents selecting the maximum value in each column to form a vector as the optimal solution, and Z - represents selecting the minimum value in each column to form a vector as the worst solution, represents calculating the distance from the optimal solution for each split layer, represents calculating the distance from the worst solution for each split layer, and S i represents the comprehensive evaluation score of each split layer; n represents the number of data features collected, which is equal to 5 here; (1.6) Determine the split layer i = max(S i ) between the neural network model to be split on the terminal device and the edge server, where (i = 1, 2,..., L).

Images