Multi-objective analysis method for heterogeneous convolutional neural network facing industrial edge scene

Abstract

The application discloses a multi-target analysis method for an industrial edge scene heterogeneous convolutional neural network. The method comprises the following steps: obtaining a plurality of target tasks, wherein different target tasks are processed by different convolutional neural network models; when each target task is unloaded from a terminal device to an edge server, different convolutional neural network models are processed by grouping layer granularity division and optimizer search, so that the number of divisible layers of different convolutional neural network models is equal, wherein each target task corresponds to a terminal device and an edge server. The application solves the technical problem that in the prior art, when a heterogeneous CNN inference task finds the best unloading scheme, frequent initialization is used to cope with the relationship between the search space and the split layer, and each model tailors the analysis strategy, which causes waste of computing resources and loss of correlation between models.

Description

Technical Field

[0001] The present invention relates to the technical fields of Industrial Internet of Things, Convolutional Neural Network Analysis, and Multi-Objective Optimization, and more specifically, to a multi-objective analysis method for heterogeneous convolutional neural networks in industrial edge scenarios. Background Art

[0002] As a key technology in the field of artificial intelligence, convolutional neural networks (CNNs) are used in industrial environments for quality inspection, equipment maintenance, and product classification. The Industrial Internet is characterized by high responsiveness, low-cost efficiency, diverse task processing, and operation in complex environments. CNN performance is closely related to the complexity of the network structure; more complex CNNs generally perform better. However, the high performance requirements of complex CNNs limit their deployment in industrial environments. In particular, in edge environments, lower-performance end devices cannot efficiently perform all CNN inference tasks, significantly limiting the deployment of CNN-based applications.

[0003] To address this issue, algorithms and network architectures optimized for edge device performance must be designed and implemented to fully realize the potential of CNNs in industrial environments. To alleviate the performance limitations of end devices, solutions such as cloud computing, fog computing, and edge computing have emerged. These solutions offload all or part of the inference task to higher-performance servers through parsing. For example, research has shown that fully offloading CNN-based tasks to servers can significantly improve speed and reduce resource consumption. Meanwhile, some studies have explored partial offloading strategies, further accelerating model inference speed and optimizing the allocation of computing resources. Partial offloading, due to its flexibility, stability, and robustness, is gradually becoming a mainstream approach.

[0004] To maximize the performance of parsing strategies, it is necessary to consider different practical scenarios for CNN deployment. These scenarios can be roughly divided into two categories: those with multiple end devices and those with multiple servers. In the case of multiple end devices, it is crucial to consider the logical relationships and constraints between tasks. In contrast, multi-server scenarios involve multiple central nodes, such as edge servers or cloud servers, which poses challenges in task allocation. To address these challenges, improving system performance through effective task allocation has become a new research focus. This research considers various factors and proposes corresponding solutions to achieve effective task scheduling.

[0005] The diverse needs of industrial environments often lead to the deployment of different applications on each end device, utilizing a variety of heterogeneous CNNs. For example, AlexNet and VGG are widely used for image classification tasks. AlexNet is suitable for deployment on devices with low resource requirements and high real-time performance, while VGG is more suitable for processing large datasets. Google's Inception architecture, known as Google LeNet, excels at high-precision tasks. Meanwhile, ResNet, with its deeper structure, is more suitable for industrial scenarios requiring detailed recognition but less demanding real-time processing. These heterogeneous models play different roles in industrial scenarios based on their distinct characteristics.

[0006] However, there are currently limited inventions on parsing continuous heterogeneous CNNs inference tasks. Although many methods have verified the effectiveness of their strategies and achieved remarkable results, these methods usually only parse one type of CNNs inference task at a time, and rarely involve the parsing problem of mixed heterogeneous CNNs inference tasks. When parsing heterogeneous CNN inference tasks to find the best offloading solution, great challenges may be faced due to the dynamic search space and the relationship between the segmentation layers. Existing parsing strategies usually require frequent reinitialization to cope with these challenges, resulting in insufficient flexibility and waste of computing resources. In addition, tailoring the parsing strategy for each model will result in a waste of computing resources and loss of correlation between models. Therefore, how to parse these continuously generated heterogeneous CNNs inference tasks without reinitializing the strategy is an urgent problem to be solved. Summary of the Invention

[0007] An embodiment of the present invention provides a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios, which at least solves the technical problem that when existing technologies parse heterogeneous CNN inference tasks to find the optimal offloading solution, they frequently initialize to cope with the dynamic relationship between the search space and the segmentation layer, and tailor the parsing strategy for each model, which results in a waste of computing resources and a loss of correlation between models.

[0008] According to one aspect of an embodiment of the present invention, a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios is provided. The method may include: obtaining multiple target tasks, wherein different target tasks are processed by different convolutional neural network models; when each target task is unloaded from a terminal device to an edge server, processing different convolutional neural network models through grouping layer granularity division and optimizer search, so that the number of divisible layers of different convolutional neural network models is equal, wherein each target task corresponds to a terminal device and an edge server; when each target task is processed by the convolutional neural network model corresponding to each target task, an initial population of dimensions corresponding to the number of layers of the convolutional neural network model corresponding to each target task is created according to the number of layers of the convolutional neural network model corresponding to each target task, wherein the initial population includes n individuals, and the dimension of each individual is the number of layers of the convolutional neural network model corresponding to each target task plus a dimension representing a server identifier; a differential evolution algorithm is used to perform the multi-objective parsing of the convolutional neural network model corresponding to each target task. The initial population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to the target task is iteratively updated to obtain the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task; the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is converted through the sigmoid function to obtain the 0 / 1 coding population of each target task; the multiple candidate solutions for each target task are obtained by the number of zeros of each individual in the 0 / 1 coding population of each target task, wherein each candidate solution is the number of zeros of each individual plus the server identifier; based on the multiple candidate solutions for each target task, the target number of segmentation layers for each target task is determined; based on the target number of segmentation layers for each target task, the total delay and total energy of each target task are determined; based on the total delay and total energy of each target task, the total delay and total energy of multiple target tasks are obtained.

[0009] Optionally, based on multiple candidate solutions for each target task, the target number of segmentation layers for each target task is determined, including: substituting the information of the number of segmentation layers shown by each candidate solution into the initial delay calculation formula to obtain multiple initial total delays; and determining the candidate solution corresponding to the minimum value of the multiple initial total delays as the target number of segmentation layers for each target task.

[0010] Optionally, the initial delay is calculated as:

[0011]

[0012] in, is the initial delay, Inferring latency for end devices, The transmission delay of the terminal device from the output data of the terminal device to the edge server, After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting, Server inference latency for predictive model output.

[0013] Optionally, based on the target number of segmentation layers of each target task, the total delay and total energy of each target task are determined, including: substituting the target number of segmentation layers of each target task into the target delay calculation formula and the target energy calculation formula to obtain the total delay and total energy of each target task.

[0014] Optionally, the target delay is calculated as:

[0015]

[0016] in, For target delay, Inferring latency for end devices, The transmission delay of the terminal device from the output data of the terminal device to the edge server, After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting, Reasoning about latency for the server.

[0017] Optionally, the target energy calculation formula is:

[0018]

[0019] in, is the target energy, Reasoning about energy consumption for end devices, The transmission energy consumption of the terminal device to transmit the output data of the terminal device to the edge server.

[0020] Beneficial effects of the present invention:

[0021] This paper proposes a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios. First, a delay and energy consumption model is proposed to parse heterogeneous CNN inference tasks in the Internet with multiple terminal devices and edge servers. Secondly, a differential multi-objective optimization method is proposed to simultaneously solve the system delay and energy consumption problems. A lightweight server computing delay prediction model established using historical data reduces the communication frequency between terminal devices and edge servers, thereby reducing the communication burden associated with finding the optimal segmentation layer. Finally, a grouping layer granularity segmentation method is proposed to manage the dynamic search space when determining the optimal segmentation layer. Rigorous theoretical proof and experimental analysis are provided to ensure that the present invention can effectively reduce system delay and energy consumption in heterogeneous model inference tasks, meeting the low latency and low energy consumption requirements in industrial Internet scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0023] Figure 1 is a flowchart of a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios according to an embodiment of the present invention;

[0024] Figure 2 2 is a schematic diagram of a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios according to an embodiment of the present invention;

[0025] Figure 3 2 is a schematic diagram of multi-objective differential evolution coding according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0028] Example 1

[0029] According to an embodiment of the present invention, a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system comprising at least one set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0030] Figure 1 Flowchart of a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:

[0031] Step S101: Acquire multiple target tasks, wherein different target tasks are processed using different convolutional neural network models.

[0032] In the technical solution provided in the above step S101 of the present invention, multiple target tasks are obtained. The target tasks can be image classification, image segmentation, etc., which are not specifically limited here. For example, the convolutional neural network model used for each task is different. For example, the convolutional neural network model is AlexNet, GoogLeNet, ResNet, etc., among which AlexNet is a chain structure, GoogLeNet is an Inception structure, and ResNet is a residual structure.

[0033] The Inception architecture is a key component of the GoogLeNet architecture, effectively addressing the challenge of efficiently capturing multi-scale features. It does this by applying multiple convolutional filters of different sizes (1×1, 3×3, and 5×5) in parallel, followed by a 3×3 max pooling operation, and then concatenating their outputs. The Inception architecture provides CNNs with more flexible convolution options and improves model performance. This paper considers the Inception architecture as a whole, ignoring the complex internal details of each Inception module and focusing solely on offloading the model to the backbone nodes. This approach not only narrows the search space but also significantly reduces the amount of data transmission required for parallel inference, thereby reducing transmission latency.

[0034] The ResNet architecture introduces bottleneck blocks to address the vanishing gradient problem and improve the training efficiency of convolutional neural networks. Each bottleneck block consists of multiple convolutional layers, with a skip connection (residual connection) added from the input to the block's output. This architecture allows the network to learn more complex functions while maintaining a manageable computational cost. This paper treats the bottleneck block as a whole, treating it as a segmentation unit and ignoring its internal structure, thereby simplifying the task.

[0035] In step S102, when each target task is unloaded from the terminal device to the edge server, different convolutional neural network models are processed through grouping layer granularity division and optimizer search to make the number of divisible layers of different convolutional neural network models equal, wherein each target task corresponds to one terminal device and one edge server.

[0036] In the technical solution provided in the above step S102 of the present invention, Figure 2 is a schematic diagram of a multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios according to an embodiment of the present invention. Figure 2 As shown, Figure 2 The two leftmost columns in ,the granularity division of the grouping layers of heterogeneous models,,the optimizer searches to make the number of splittable layers of different,convolutional neural network models equal, e.g. Figure 2 In the third column, the number of separable layers of different convolutional neural network models is 4.

[0037] like Figure 2 As shown in the figure, models of different structures are first treated as a unified chain structure; then, they are grouped according to their different layers. This grouping strategy allows the optimizer to search for the optimal solution within a stable space. Based on the decision information of the optimal solution, the split layer and server for executing subsequent tasks are determined. Subsequently, the output data of the split layer is sent to the designated server. Finally, the subsequent tasks enter the waiting queue of the server, waiting for further execution.

[0038] CNNs are heterogeneous not only in their architectures but also in their lengths. These varying lengths lead to different search spaces for parsing the optimal offloading decision. When dealing with diverse search spaces, the parsing strategy must adapt to the frequent changes in the decision dimension, often requiring a reset of the search algorithm to maintain optimal performance. To address the challenges posed by varying model lengths, this strategy employs layer grouping to ensure robustness.

[0039] When faced with heterogeneous CNNs inference tasks generated by multiple terminal devices, the strategy first determines the search space based on the shortest model length among the models. For example, when deploying AlexNet, GoogLeNet, and VGG-16, the length of the shortest model (AlexNet with a length of 13) is used as the standard. For longer models, such as VGG-16 with a length of 31, the present invention uses random grouping to reduce the search dimension to the standard length. VGG-16 layers are randomly grouped, and each group is treated as a whole. Therefore, the search for the best split layer is only performed between these groups, stabilizing the search space at the standard value of 13.

[0040] Step S103: When each target task is processed by the convolutional neural network model corresponding to each target task, an initial population with a dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is created according to the number of layers of the convolutional neural network model corresponding to each target task, wherein the initial population includes n individuals, and the dimension of each individual is the number of layers of the convolutional neural network model corresponding to each target task plus a dimension representing the server identifier.

[0041] In the technical solution provided in step S103 of the present invention, Figure 3 This is a schematic diagram of multi-objective differential evolution encoding according to an embodiment of the present invention. The CNN parsing process is essentially a discrete problem-solving process. Most current strategies use binary identifiers (0 and 1) to represent parsing decisions. In population encoding, a normalized continuous value in the range [-1, 1] is used to initialize the population. The dimension of each individual in the population corresponds to the number of layers in the model. In edge computing environments with multiple end devices and servers, the choice of which server to send the second half of a task significantly impacts task latency. Therefore, selecting the appropriate server is a key consideration when making parsing decisions. To address this issue, in population encoding, each individual has an additional dimension representing the server identity. This means that each candidate solution must specify not only the parsing decision for each layer but also the server that will execute the second half of the task. Therefore, the dimension of each individual in the population is always twice the number of layers in the model.

[0042] According to the encoding rules, the new generation population is represented by standardized values. The sigmoid function is used to convert each dimension of the individual to 0 or 1 as a parsing identifier. The number of zeros in each individual is calculated to determine the index of the segmentation layer, such as Figure 3 As shown, continue to code the population X1, X2....X n For each individual in the initial population, each individual includes , is the number of layers of the convolutional neural network model corresponding to each target task, A dimension that identifies the server.

[0043] In step S104, the initial population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is iteratively updated through the differential evolution algorithm to obtain the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task.

[0044] In the technical solution provided in the above step S104 of the present invention, the initial population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is iteratively updated by the differential evolution algorithm to obtain the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task.

[0045] Initialization: Create a population with the corresponding dimensions according to the number of layers of the model, and the population contains n individuals. For example, Figure 3 As shown, when the number of layers of the CNN model is 6, the present invention creates a population consisting of n individuals. These individuals are X1, X2....X n Composition, each individual has 6+1 dimensions.

[0046] Evolution: The differential evolution algorithm is used as an optimizer to generate new populations through iteration. The iteration process continues until the termination condition is met. The algorithm outputs a Pareto optimal solution set S1 based on the multi-objective optimization characteristics. Finally, a solution is selected from the solution set S1 as the optimal solution X. best .

[0047] In step S105, the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is converted through a sigmoid function to obtain a 0 / 1 coding population for each target task.

[0048] In the technical solution provided in step S105 of the present invention, Figure 3 As shown in FIG, the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is converted by the sigmoid function to obtain the 0 / 1 coding population for each target task.

[0049] Step S106 , obtaining multiple candidate solutions for each target task by calculating the number of zeros of each individual in the 0 / 1 coding population of each target task, wherein each candidate solution is the number of zeros of each individual plus a server identifier.

[0050] In the technical solution provided in the above step S106 of the present invention, for example, Figure 3 X1 in includes 3 zeros, so the candidate solution for X1 is , and so on for other candidate solutions.

[0051] Step S107 : determining the target number of segmentation levels for each target task based on multiple candidate solutions for each target task.

[0052] In the technical solution provided in the above step S107 of the present invention, the target number of segmentation layers for each target task is obtained based on multiple candidate solutions for each target task.

[0053] Step S108 : determining the total delay and total energy of each target task based on the target segmentation level of each target task.

[0054] In the technical solution provided in the above step S108 of the present invention, the total delay and total energy of each target task are determined according to the target segmentation layer number of each target task.

[0055] Step S109 : obtaining the total delay and total energy of multiple target tasks based on the total delay and total energy of each target task.

[0056] In the technical solution provided in the above step S109 of the present invention, the total delay of each target task is accumulated and summed to obtain the total delay of multiple target tasks, and the total energy of each target task is accumulated and summed to obtain the total energy of multiple target tasks.

[0057] The above method of this embodiment is further introduced below.

[0058] As an optional implementation method, step S107, based on multiple candidate solutions for each target task, determines the target number of segmentation layers for each target task, including: substituting the information of the number of segmentation layers shown by each candidate solution into the initial delay calculation formula to obtain multiple initial total delays; and determining the candidate solution corresponding to the minimum value of the multiple initial total delays as the target number of segmentation layers for each target task.

[0059] As an optional embodiment, the initial delay calculation formula is:

[0060]

[0061] in, is the initial delay, Inferring latency for end devices, The transmission delay of the terminal device from the output data of the terminal device to the edge server, After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting, Server inference latency for predictive model output.

[0062] In this embodiment, in the process of finding the best split layer, each candidate solution must be evaluated to select the best solution. Evaluating each candidate solution requires prior knowledge of the inference latency on the subtask-specified server, which is called L server . Frequently send requests to obtain from the server L server This will cause huge communication overhead and prolong the optimization process. To solve this problem, the present invention uses a lightweight radial basis function network (RBFN) to predict the inference delay on the designated server after segmentation. L server The prediction model uses historical data for incremental training. During the task execution process, the present invention collects the inference delay on the terminal device. L endBased on this information, the present invention constructs a three-layer lightweight neural network. This model captures the implicit relationship between the terminal inference delay, model type, server identification information, and the specified server inference delay. For the analytical decision of the current candidate solution, the input of the prediction proxy model includes the inference delay of the first part of the current task on the terminal device. Input L , CNN information for the generation task Input M And the server information for the second part of the task Input S The output of the prediction surrogate model is the inference latency of the second part of the task on the specified server. The expression is as follows:

[0063]

[0064] As an optional implementation method, step S108, determining the total delay and total energy of each target task based on the target segmentation layer number of each target task, includes: substituting the target segmentation layer number of each target task into the target delay calculation formula and the target energy calculation formula to obtain the total delay and total energy of each target task.

[0065] As an optional embodiment, the target delay calculation formula is:

[0066]

[0067] in, For target delay, Inferring latency for end devices, The transmission delay of the terminal device from the output data of the terminal device to the edge server, After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting, Reasoning about latency for the server.

[0068] In this embodiment, the terminal device infers the delay: For any terminal device, the present invention uses d i Indicates that n In the convolutional neural network model of layer i The floating point computation amount of the layer, where i ∈[1, n]. j Layer as a split layer, j ∈[1,n]. The reasoning from layer 1 to layer j is performed locally, while the reasoning from layer jThe inference process from layer +1 to layer n will be performed on the edge server. f 1 represents the floating point computing capability of the terminal device. The delay formula for executing layer 1 to layer j on the terminal device is as follows. Terminal device inference delay L end The calculation formula is as follows:

[0069]

[0070] The transmission delay of the terminal device to transmit the output data of the terminal device to the edge server: j Layer is the segmentation layer that needs to send output data to the edge server (what does it mean). m j represents the size of the output data of the jth layer. According to Shannon's law, the data transmission rate is ,in, B is the channel bandwidth, Q is the transmission power of the terminal equipment, G is the channel gain between the terminal device and the edge server, is the noise power. Transmission delay L trans The calculation method is as follows:

[0071]

[0072] After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting: the intermediate data of the task is transmitted to the designated server k After that, the task will enter the queue and wait for execution. The waiting time is proportional to the number of tasks before it in the queue. q And each task on the edge server k The calculation formula is as follows:

[0073]

[0074] Server inference delay: Due to j The layer is a segmentation layer, and the edge server needs to perform j +1 layer to n-th layer reasoning tasks. f 3 represents the floating point computing capability of the edge server. j The latency from layer +1 to layer n is expressed as follows: Edge server inference latency L server The calculation formula is as follows:

[0075]

[0076] in,f 3 represents the floating-point computing capability of the edge server.

[0077] As an optional embodiment, the target energy calculation formula is:

[0078]

[0079] in, is the target energy, Reasoning about energy consumption for end devices, The transmission energy consumption of the terminal device to transmit the output data of the terminal device to the edge server.

[0080] In this embodiment, the terminal device infers energy consumption: E end The calculation formula is as follows:

[0081]

[0082] in, p end is the terminal device inference energy consumption rate, L end is the terminal device inference latency for the current task.

[0083] The transmission energy consumption of the terminal device to transmit the output data of the terminal device to the edge server: E trans The calculation formula is as follows:

[0084]

[0085] in, p trans Transmit energy consumption rate for terminal equipment, L trans Transmission delay for the terminal device of the current task.

[0086] In an embodiment of the present invention, a plurality of target tasks are obtained, wherein different target tasks are processed by different convolutional neural network models; when each target task is unloaded from a terminal device to an edge server, different convolutional neural network models are processed by grouping layer granularity division and optimizer search, so that the number of divisible layers of different convolutional neural network models is equal, wherein each target task corresponds to a terminal device and an edge server; when each target task is processed by the convolutional neural network model corresponding to each target task, an initial population of dimensions corresponding to the number of layers of the convolutional neural network model corresponding to each target task is created according to the number of layers of the convolutional neural network model corresponding to each target task, wherein the initial population includes n individuals, and the dimension of each individual is the number of layers of the convolutional neural network model corresponding to each target task plus a dimension representing a server identifier; the initial population of dimensions corresponding to the number of layers of the convolutional neural network model corresponding to each target task is iteratively updated by a differential evolution algorithm to obtain a target population of dimensions corresponding to the number of layers of the convolutional neural network model corresponding to each target task; The target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to the task is converted through the sigmoid function to obtain a 0 / 1 coding population for each target task; multiple candidate solutions for each target task are obtained through the number of zeros of each individual in the 0 / 1 coding population of each target task, where each candidate solution is the number of zeros of each individual plus a server identifier; based on the multiple candidate solutions for each target task, the target number of segmentation layers for each target task is determined; based on the target number of segmentation layers for each target task, the total delay and total energy of each target task are determined; based on the total delay and total energy of each target task, the total delay and total energy of multiple target tasks are obtained, which solves the technical problem that when existing technologies parse heterogeneous CNN inference tasks to find the best offloading solution, they deal with the dynamic relationship between the search space and the segmentation layer through frequent initialization, and tailor-made parsing strategies for each model will cause waste of computing resources and loss of correlation between models, and achieves the technical effect of effectively reducing system delay and energy consumption in heterogeneous model inference tasks and meeting the low latency and low energy consumption requirements in industrial Internet scenarios.

[0087] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0088] In the above embodiments of the present invention, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0089] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0090] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected to achieve the purpose of the present embodiment according to actual needs.

[0091] In addition, the functional units in various embodiments of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0092] The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-objective parsing method for heterogeneous convolutional neural networks in industrial edge scenarios, characterized by: include: Obtain multiple target tasks, where different target tasks are processed using different convolutional neural network models; When each target task is offloaded from the terminal device to the edge server, different convolutional neural network models are processed through group layer granularity division and optimizer search to make the number of splittable layers of different convolutional neural network models equal, where each target task corresponds to one terminal device and one edge server; When each target task is processed by the convolutional neural network model corresponding to each target task, an initial population with a dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is created according to the number of layers of the convolutional neural network model corresponding to each target task, wherein the initial population includes n individuals, and the dimension of each individual is the number of layers of the convolutional neural network model corresponding to each target task plus a dimension representing the server identifier; The initial population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is iteratively updated by the differential evolution algorithm to obtain the target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task; The target population of the dimension corresponding to the number of layers of the convolutional neural network model corresponding to each target task is converted through the sigmoid function to obtain the 0 / 1 coding population for each target task; By encoding the number of zeros of each individual in the 0 / 1 population of each target task, multiple candidate solutions for each target task are obtained, where each candidate solution is the number of zeros of each individual plus the server identifier; Based on multiple candidate solutions for each target task, determine the target number of segmentation layers for each target task; Determine the total latency and total energy of each target task based on the target number of segmentation layers for each target task; Based on the total delay and total energy of each target task, the total delay and total energy of multiple target tasks are obtained.

2. The method according to claim 1, characterized in that The step of determining the target number of segmentation layers for each target task based on multiple candidate solutions for each target task includes: Substitute the number of split layers indicated by each candidate solution into the initial delay calculation formula to obtain multiple initial total delays; The candidate solution corresponding to the minimum value of multiple initial total delays is determined as the target segmentation layer number for each target task.

3. The method according to claim 2, characterized in that The initial delay calculation formula is: in, is the initial delay, Inferring latency for end devices, The transmission delay of the terminal device from the output data of the terminal device to the edge server, After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting, Server inference latency for predictive model output.

4. The method according to claim 1, wherein The determining of the total delay and total energy of each target task based on the target segmentation layer number of each target task includes: Substitute the target number of segmentation layers for each target task into the target delay calculation formula and the target energy calculation formula to obtain the total delay and total energy of each target task.

5. The method according to claim 1, wherein The target delay calculation formula is: in, For the target delay, Inferring latency for end devices, The transmission delay of the terminal device from the output data of the terminal device to the edge server, After the terminal device transmits the output data of the terminal device to the edge server, the waiting delay when the output data of the terminal device enters the queue for waiting, Reasoning about latency for the server.

6. The method according to claim 1, characterized in that The target energy calculation formula is: in, is the target energy, Reasoning about energy consumption for end devices, The transmission energy consumption of the terminal device to transmit the output data of the terminal device to the edge server.

7. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

8. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

9. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the method of claim 1 when the instructions are executed.

Images