A distributed learning method for edge image recognition in intelligent transportation

Abstract

The application discloses a distributed learning method for edge image recognition in intelligent transportation, comprising the following steps: step 1: the server selects a distributed learning task and a deep learning model under an intelligent transportation application scenario; step 2: the server performs model preprocessing; step 3: the server starts a tth round of distributed model training process; step 5: the server selects an edge device based on a selection strategy; step 6: according to information feedback in the existing training, the edge calculates a utility value δ and sends the utility value to the server; and step 7: steps 3 to 8 are repeated until the model parameters converge. k The application has the advantages that fine control granularity and high-performance image recognition model joint training are realized; the number of participants can be reduced and the energy cost can be reduced according to the model training effect and the adjustment of the edge selection frequency; the participation of the edge can be reduced while the accuracy is maintained by dynamically adjusting the selection frequency, and the robustness of the system is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a distributed learning method for edge image recognition in intelligent transportation. Background Technology

[0002] Transportation systems are the foundation of a nation's economic development. The rapid development of communication and network technologies, such as mobile edge computing and 5G networks, has greatly improved the transportation system's ability to collect necessary information and connect its various components. To fully realize the potential of transportation systems, technologies such as machine learning and intelligent analytics are being applied. The field connecting transportation systems and machine learning is the Intelligent Transport System (ITS). By comprehensively applying intelligent technologies to the field of transportation, highly efficient and energy-saving transportation systems are formed.

[0003] Image recognition technology is widely used in Intelligent Transportation Systems (ITS) to identify traffic congestion, detect collisions, and identify pedestrians and bicycles. To maintain service quality, ITS utilizes roadside units (LSUs), which are placed on both sides of the road and equipped with sensors and processing capabilities to acquire image data. Image processing is a crucial function of transportation systems.

[0004] Image recognition applications in intelligent transportation systems (ITS) face a series of unique challenges. Unlike centralized control clusters, ITS involves a large number of edge devices that vary significantly in computing power, robustness, and communication performance. Furthermore, the data across devices is heterogeneous, with substantial differences in data volume and distribution among different edge devices. Beyond the heterogeneity of data and devices, ITS presents a unique challenge to distributed learning applications: a dynamic environment where participants can be highly mobile and unstable connected vehicles, or edge devices with low computing power or energy capacity. Communication networks may fail due to device malfunctions or the intentions of device owners, leading to device disconnections, a problem exacerbated in ITS. Additionally, many edge devices in ITS are highly mobile and exhibit unstable performance. Therefore, certain control methods are needed to maintain system performance stability.

[0005] Image recognition models can be represented as computational graphs composed of a large number of computing nodes and parameters, trained through multiple rounds of forward and backward propagation. Many image recognition models contain a large number of parameters, making direct training on resource-constrained single edge devices using local data impossible. Furthermore, existing parallel distributed methods for models have performance limitations and cannot adapt to the dynamic environments of intelligent transportation systems.

[0006] The technical abbreviations and key technical terms involved in this invention are defined as follows:

[0007] ITS: Intelligent Transport System;

[0008] Forward Propagation: A step in training an intelligent model;

[0009] Backward Propagation: A step in training an intelligent model;

[0010] ResNet18: A convolutional neural network model containing 17 convolutional layers and 1 fully connected layer;

[0011] ReLU: A layer in a neural network that implements a linear rectified function;

[0012] Dropout: A layer in a neural network that randomly discards some training units to prevent overfitting;

[0013] SoftMax: A layer in a neural network, also known as the normalized exponential function, is a type of activation function;

[0014] Oort: A state-of-the-art distributed learning participant selection scheme;

[0015] L2 norm: Used to measure the Euclidean distance between two vectors. For vector x = (x1, x2, ..., x...) n The L2 norm of x is expressed as:

[0016] epoch: One epoch is defined as the process of traversing all data once.

[0017] CIFAR10: A dataset commonly used for image classification, containing 50,000 color images of cats and horses across 10 categories.

[0018] FEMNIST: A handwritten digit dataset, consisting of 805,263 black and white images in 62 categories, including uppercase letters, lowercase letters, and digits.

[0019] IID: Independent Identically Distribution, a distribution characteristic of training data;

[0020] Adam: An algorithm for performing first-order gradient optimization on a stochastic objective function. Summary of the Invention

[0021] The purpose of this invention is to provide a distributed learning method for edge image recognition in intelligent transportation that can overcome the above-mentioned technical problems. The method of this invention includes the following steps:

[0022] Step 1: The server selects the distributed learning task and deep learning model for the intelligent transportation application scenario:

[0023] Step 1.1: The server initializes the distributed learning model to obtain the initial global computation graph G and the initial model parameters w1;

[0024] Step 1.2: Configure the server for training, including the number of edge devices N, the number of training rounds E, the parameters for grouping nodes in the computation graph L, and the hyperparameter of the learning rate used in the distributed learning process;

[0025] Step 2: The server performs model preprocessing and distributes the data to the participating edge devices:

[0026] Step 2.1: Apply constant folding and operator fusing techniques to optimize the computational graph structure of the deep model on the server.

[0027] Step 2.2: The server divides each L-level node into a subgraph according to the level to which each node belongs in the computation graph. The level is defined as the root node in the computation graph being the zeroth level, all child nodes of the root node being the first level, and so on. The value of L is adjusted by the server according to different model sizes, and the parameter is set to L=2.

[0028] Step 3: The server begins the t-th (t = 1, 2, ..., E) round of distributed model training, with t initially set to 1.

[0029] Step 3.1: Each smart device includes a utility value as a numerical attribute. The server selects the K devices with the highest utility values ​​from all candidate available smart edge devices as the training subset C. t Where log2K is an integer, the specific calculation method of the utility value is described in detail in step 6. When t=1, the utility value at the edge is a random number.

[0030] Step 3.2: The server divides the multiple subgraphs into equal groups and sends them to C. t On intelligent edge devices;

[0031] Step 4: The edge devices participating in the learning process perform log(K) rounds of local distributed model training.

[0032] Step 4.1: Each edge device performs forward propagation calculations;

[0033] Step 4.2: Set the current round number as h and the edge device number as k, then determine whether k ≥ 2. h If so, then find the number k-2. h-1 The edge devices are represented by sets A and B, which represent the sets of nodes in the computation graph subgraphs contained in the two devices.

[0034] Step 4.3: For the computation graph nodes in A and B, based on the local data D k ,calculate: b j For a i Parent node}, complete local backpropagation, and obtain the gradient g of each node i;

[0035] Step 4.4: The edge k uses the gradient descent algorithm to update the local weights of the distributed model in round t as shown in the following formula (1):

[0036]

[0037] Where η is the learning rate parameter of the gradient descent algorithm;

[0038] Step 5: The server decides whether to reselect edge devices to participate in distributed learning in the current round based on the edge device selection strategy. The selection frequency p is set according to the dataset attributes and dynamic environment, that is: the edge device is reselected once every p rounds.

[0039] If the edge is not reselected, repeat steps 3 to 4; if the edge is reselected, proceed to step 6.

[0040] Step 6: Based on the feedback from the existing training, the edge device calculates its own utility value δ. k Send to the server:

[0041] Step 6.1: The utility of the edge includes its availability U and usability P. The availability of the edge depends on the time of work submission, the willingness of the device owner to contribute, and distance. Usability refers to the impact of local updates at the edge, including whether it will cause an increase in global model weight shift, slow convergence, or slow transmission due to channel problems, excessively large local training data leading to slow training, or excessively small local training data leading to a large gap between the data and the overall data distribution, as shown in the following formula:

[0042] δ=ω1U+ω2P……(2),

[0043] Where ω1 and ω2 are the weights of availability and practicality in this utility measurement, and their specific values ​​are determined by the user based on the actual smart transportation application scenario.

[0044] Step 6.2: The edge device sends its utility value to the server.

[0045] Step 7: Repeat steps 3 through 8 until the model parameters converge.

[0046] The superior effects of the method described in this invention are:

[0047] 1. The parallel training method proposed in this invention, which uses computation graph nodes as the basic scheduling unit, utilizes resource-constrained edge devices in intelligent transportation systems to achieve joint training of image recognition models with fine control granularity and high performance.

[0048] 2. The method described in this invention can reduce the number of participants by adjusting the selection frequency of the edge end according to the model training effect, thereby reducing energy costs. This is a key objective of intelligent transportation systems. By dynamically adjusting the selection frequency, the participation of the edge end is reduced while maintaining accuracy, effectively improving the robustness of the system.

[0049] 3. The distributed training and edge device selection of the edge-end utility described in this invention provide a large adjustment space for edge-end selection to achieve fine-grained control over the selection frequency.

[0050] 4. The method described in this invention optimizes the selection of edge devices to participate in learning in a dynamic environment and distributes the model to different devices. It trains the model through backpropagation using high-performance distributed computing. By using the selection frequency mechanism of the participants, it highlights the feedback control of the distributed learning process on the selection process and explores the trade-off between accuracy performance and system process in each round. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the basic system architecture for which the method described in this invention is applied;

[0052] Figure 2 This is a schematic diagram comparing the accuracy of edge selection methods on different datasets according to the method described in this invention;

[0053] Figure 3 This is a schematic diagram illustrating the impact of different edge selection frequencies on model performance in the method described in this invention. Detailed Implementation

[0054] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0055] The method of the present invention includes the following steps:

[0056] Step 1: The server selects the distributed learning task and deep learning model for the intelligent transportation application scenario:

[0057] Step 1.1: The server initializes the distributed learning model to obtain the initial global computation graph G and the initial model parameters w1;

[0058] Step 1.2: Configure the server for training, including the number of edge devices N, the number of training rounds E, the parameters for grouping nodes in the computation graph L, and the hyperparameter of the learning rate used in the distributed learning process;

[0059] Step 2: The server performs model preprocessing and distributes the data to the participating edge devices:

[0060] Step 2.1: Apply constant folding and operator fusing techniques to optimize the computational graph structure of the deep model on the server.

[0061] Step 2.2: The server divides each L-level node into a subgraph according to the level to which each node belongs in the computation graph. The level is defined as the root node being the zeroth level, all child nodes of the root node being the first level, and so on. The value of L is adjusted by the server according to different model sizes, and the parameter is set to L=2.

[0062] Step 3: The server begins the t-th (t = 1, 2, ..., E) round of distributed model training, with t initially set to 1.

[0063] Step 3.1: Each smart device includes a utility value as a numerical attribute. The server selects the K devices with the highest utility values ​​from all candidate available smart edge devices as the training subset C. t Where log2K is an integer, the specific calculation method of the utility value is described in detail in step 6. When t=1, the utility value at the edge is a random number.

[0064] Step 3.2: The server divides the multiple subgraphs into equal groups and sends them to C. t On intelligent edge devices;

[0065] Step 4: The edge devices participating in the learning process perform log(K) rounds of local distributed model training.

[0066] Step 4.1: Each edge device performs forward propagation calculations;

[0067] Step 4.2: Set the current round number as h and the edge device number as k, then determine whether k ≥ 2. h If so, then find the number k-2. h-1 The edge devices are represented by sets A and B, which represent the sets of nodes in the computation graph subgraphs contained in the two devices.

[0068] Step 4.3: For the computation graph nodes in A and B, based on the local data Dk ,calculate: b j For a i Parent node}, complete local backpropagation, and obtain the gradient g of each node i;

[0069] Step 4.4: The edge k uses the gradient descent algorithm to update the local weights of the distributed model in round t as shown in the following formula (1):

[0070]

[0071] Where η is the learning rate parameter of the gradient descent algorithm;

[0072] Step 5: The server decides whether to reselect edge devices to participate in distributed learning in the current round based on the edge device selection strategy. The selection frequency p is set according to the dataset attributes and dynamic environment, that is: the edge device is reselected once every p rounds.

[0073] If the edge is not reselected, repeat steps 3 to 4; if the edge is reselected, proceed to step 6.

[0074] Step 6: Based on the feedback from the existing training, the edge device calculates its own utility value δ. k Send to the server:

[0075] Step 6.1: The utility of the edge includes its availability U and usability P. The availability of the edge depends on the time of work submission, the willingness of the device owner to contribute, and the distance. The usability refers to the impact of local updates at the edge, including whether it will cause an increase in global model weight shift, slow convergence, or slow transmission due to channel problems, slow training due to excessively large local training data, or excessively small local training data leading to a large gap between the data and the overall data distribution, as shown in the following formula (2):

[0076] δ=ω1U+ω2P……(2),

[0077] Where ω1 and ω2 are the weights of availability and practicality in this utility measurement, and their specific values ​​are determined by the user based on the actual smart transportation application scenario.

[0078] Step 6.2: The edge device sends its utility value to the server.

[0079] Step 7: Repeat steps 3 through 8 until the model parameters converge.

[0080] The intelligent transportation system architecture on which the method described in this invention is based is as follows: Figure 1As shown, the intelligent transportation system includes a central server and multiple participating intelligent edge terminals, such as smart cameras or vehicles. Communication between the server and the edge terminals, as well as among the edge terminals themselves, occurs via cellular networks or WiFi. The central server sets the initial model and training parameters, while the edge terminals perform forward and backward propagation calculations based on local data. Information during the training process is exchanged through communication between the edge devices. Considering the dynamic environment of the intelligent transportation system, the method described in this invention provides fine-grained control over the edge terminal selection step. Specifically, the server controls whether to reselect an edge terminal in the current round based on the current model training effect, thereby reducing the communication volume between the server and the edge terminals and improving communication efficiency, achieving a trade-off between model accuracy and system progress within a given timeframe.

[0081] The distributed learning method for edge image recognition in intelligent transportation, implemented in the embodiments of the present invention, is deployed and operated according to the following steps:

[0082] Step A: Example Platform Selection and Dataset Selection:

[0083] The implementation example is built on the FedScale open-source platform, which simulates a real-world heterogeneous distributed learning system, provides various training tasks, includes different neural network models and datasets, configures participating devices to have different device runtimes, and uses network performance information from mobile platforms to simulate network connectivity.

[0084] The dataset uses the CIFAR10 dataset and the FEMNIST dataset. Six datasets were created from the original FEMNIST dataset, which have different distribution and size characteristics, as shown in Table 1. Here, s represents the Non-IID degree of the dataset. The smaller s is, the greater the Non-IID degree of the dataset. The split size shows whether the data at different edge ends are average.

[0085] Table 1. Datasets constructed based on the FEMNIST dataset.

[0086] Dataset Name Local data distribution Segment size FEMNIST-IE IID average FEMNIST-IU IID Uneven FEMNIST-NE1024 Non-IID (s=1024) average FEMNIST-NU1024 Non-IID (s=1024) Uneven FEMNIST-NE2048 Non-IID (s=2048) average FEMNIST-NU2048 Non-IID (s=2048) Uneven

[0087] Step B: Construction of the Deep Neural Network:

[0088] For the CIFAR10 dataset, the ResNet18 model is used. The neural network architecture used in the FEMNIST dataset includes two convolutional layers with 32 and 64 output channels, a kernel size of 3×3, and a stride of 1. After both layers, there are ReLU activation layers and DropOut layers, followed by two fully connected layers with an output size of 128 for the first and 64 for the second. Finally, there is a SoftMax layer.

[0089] The local training batch size is set to 10, the local epoch number is 5, the Adam optimizer is used, the learning rate η is 0.001, and the parameters are set to (0.9, 0.999).

[0090] Step C: Distributed learning training design:

[0091] Using the distributed learning method proposed in this invention, as a limitation of step 5, the user manually adjusts the selection rate of new participants using a fixed rate based on dataset attributes and a dynamic distributed training environment. In this embodiment, the participants in the distributed learning are reselected every 1, 5, 10, and 20 rounds.

[0092] As a limitation of step 6, this embodiment uses the L2 norm of the weight change before and after the local update at the edge as the utility measure of the edge. The utility U of edge k used in this embodiment is... k The calculation formula is as follows:

[0093]

[0094] coefficient This is to control the impact of the local data size at the edge on the utility of the edge, where b k It is the size of the data at edge k, w k These are the weights of the new global model after one round of training, while w init It is the weight value at the start of a training round.

[0095] The content and test results of this embodiment are as follows:

[0096] 1) Distributed image recognition model training:

[0097] In the testing, the method described in this invention was compared with the existing Oort method and the benchmark method FedAvg. In the experiment, the total number of available participants was set to 24, and 6 were selected from them in each round. Each experiment lasted for 20 rounds. Figure 2 The results of this test show that the present invention achieves better performance than the benchmark method. When the data is evenly split across all edge points, the performance of this embodiment surpasses the benchmark method. The difference increases with the degree of Non-IID, and the conclusion remains valid even when the data split sizes change from equal to unequal. In different heterogeneous data scenarios, the method proposed in this invention improves the accuracy of the globally distributed learning model in distributed training by 20% compared to the benchmark method.

[0098] 2) Sampling frequency:

[0099] This test evaluates the impact of sampling frequency on the method described in this invention. For each setting, the selection rate of new participants selected every 1, 5, 10, and 20 rounds is used to compare the performance of the globally distributed learning model. The results are as follows: Figure 3 As shown, in a stable environment where local data is edge-end IID, the intelligent transportation system can choose to apply selections at a low frequency without affecting the accuracy of the global distributed learning model. As the degree of Non-IID increases, the differences between these selection frequencies become more pronounced, and the performance gap between the two extreme selection rates is less than 5%. Even with such a highly Non-IID dataset, reducing the selection frequency from every round to every five rounds does not significantly affect the global model accuracy. This indicates that the selection rate dimension provides ample room for adjusting the edge-end selection frequency with little or no impact on the global model accuracy performance.

[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A distributed learning method for edge image recognition in intelligent transportation, characterized in that, Includes the following steps: Step 1: The server selects the distributed learning task and deep learning model for the intelligent transportation application scenario; Step 2: The server performs model preprocessing and distributes the data to the participating edge devices: Step 2.1: Apply constant folding and computational fusion techniques to optimize the computational graph structure of the deep model on the server. Step 2.2: The server divides each L-level node into a subgraph according to the level to which each node belongs in the computation graph. The level is defined as the root node in the computation graph being the zeroth level, all child nodes of the root node being the first level, and so on. The value of L is adjusted by the server according to different model sizes, and the parameter is set to L=2. Step 3: The server begins the t-th (t = 1, 2, ..., E) round of distributed model training, with the initial value of t being 1. Step 4: The edge devices participating in the learning process perform log(K) rounds of local distributed model training. Step 4.1: Each edge device performs forward propagation calculations; Step 4.2: Set the current round number as h and the edge device number as k, then determine whether k ≥ 2. h If so, then find the number k-2. h-1 The edge devices are represented by sets A and B, which represent the sets of nodes in the computation graph subgraphs contained in the two devices. Step 4.3: For the computation graph nodes in A and B, based on the local data D k ,calculate:{ b j For a i Parent node}, complete local backpropagation, and obtain the gradient g of each node i; Step 4.4: The edge k updates the local weights of the distributed model in round t using the gradient descent algorithm as shown in the following formula: Where η is the learning rate parameter of the gradient descent algorithm; Step 5: The server decides whether to reselect edge devices to participate in distributed learning in the current round based on the edge device selection strategy. The selection frequency p is set according to the dataset attributes and dynamic environment. That is, the edge device is reselected once every p rounds. If the edge is not reselected, repeat steps 3 to 4; if the edge is reselected, proceed to step 6. Step 6: Based on the feedback from the existing training, the edge device calculates its own utility value δ. k Send to the server; Step 7: Repeat steps 3 through 6 until the model parameters converge.

2. The distributed learning method for edge image recognition in intelligent transportation according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: The server initializes the distributed learning model to obtain the initial global computation graph G and the initial model parameters w1; Step 1.2: Configure the server for training, including the number of edge devices N, the number of training rounds E, the parameters for grouping nodes in the computation graph L, and the hyperparameter of the learning rate used in the distributed learning process.

3. The distributed learning method for edge image recognition in intelligent transportation according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Each smart device includes a utility value as a numerical attribute. The server selects the K devices with the highest utility values ​​from all candidate available smart edge devices as the training subset C. t , where log2K is an integer, and when t=1, the utility value at the edge is a random number; Step 3.2: The server divides the multiple subgraphs into equal groups and sends them to C. t On intelligent edge devices.

4. A distributed learning method for edge image recognition in intelligent transportation according to claim 1, characterized in that, Step 6 includes the following steps: Step 6.1: The utility of the edge includes its availability U and usability P. The availability of the edge depends on various factors, including the time of work submission, the willingness of the device owner to contribute, and distance. Usability refers to the impact of local updates at the edge, including whether it will cause an increase in global model weight shift, slow convergence, or problems such as slow transmission due to channel issues, slow training due to excessively large local training data, or excessively small local training data leading to a large gap between the data and the overall data distribution, as shown in the following formula: δ=ω1U+ω2P……(2), Where ω1 and ω2 are the weights of availability and practicality in this utility measurement, and their specific values ​​are determined by the user based on the actual smart transportation application scenario. Step 6.2: The edge device sends its utility value to the server.

Images