A wireless federated segmentation learning algorithm for collaborative intelligence in the Internet of Things

Abstract

A wireless federated segmentation learning algorithm for collaborative intelligence of the Internet of Things, involving the field of federated segmentation learning, including: the edge server constructs a device scheduling optimization model and selects participating user devices; determines whether the local model parameters of the participating user devices are the latest, if not the latest, the participating user devices need to download the local model of the current global iterative training from the edge server; otherwise, no download is required; performs forward propagation, segmentation layer data upload, back propagation, gradient averaging, segmentation layer average gradient download and return; the participating user device with the best channel quality uploads the updated local model to the edge server; repeats the above steps until the global model converges or reaches the maximum communication round limit. The present invention can reduce the communication delay and transmission energy consumption caused by frequent interactions between the edge server and the participating user devices, and can effectively overcome the heterogeneity problem caused by the training of participating user devices under the premise of ensuring learning performance, and reduce system energy consumption.

Description

Technical Field

[0001] The present invention relates to the technical field of federated segmentation learning, and in particular to a wireless federated segmentation learning algorithm for collaborative intelligence of the Internet of Things. Background Art

[0002] With the widespread deployment of wireless devices and the development of sensing technology, mobile devices are generating unprecedented amounts of data every day, which leads to considerable latency, high power consumption, security and privacy issues, making cloud-based solutions inefficient. Distributed Collaborative Machine Learning (DCML) is an important machine learning architecture that can effectively solve the problem of collaborative training between nodes, protect the privacy of the generated data, and improve the efficiency of cloud servers.

[0003] In distributed collaborative machine learning for privacy protection, Federated Learning (FL) has been widely studied in edge computing due to its privacy protection, reduced network overhead, and excellent learning performance. Federated learning enables multiple edge devices to collaboratively train a model and perform model parameter aggregation by transmitting model parameters to the edge server. It avoids the leakage of user data and protects data privacy while having a better model than the participating edge devices training alone. However, federated learning requires uploading and downloading the parameters of the entire model in each round of training, which generates a lot of communication overhead; and for edge mobile devices (such as smart watches or bracelets, smart mobile phones, tablets, etc.) with limited resources (such as computing, energy, etc.) in the Internet of Things (IoT) scenario, it is usually difficult to undertake the training tasks of large deep learning models.

[0004] In order to train large deep learning models on resource-constrained edge devices in the Internet of Things, Split Learning (SL) was proposed. Its main idea is to split the global model of federated learning into two parts. One part of the underlying model is placed on the mobile device for local calculation, and the other part of the high-level model is placed on the edge server for global update. Split learning transmits the output results of the split layer during each round of training, which reduces the communication overhead between each edge mobile device and the edge server during a single training process, and also protects data privacy. However, the earliest proposed split learning mechanism is serial, that is, each participating edge mobile device executes model training with the edge server in sequence. At each moment, only one edge mobile device interacts with the edge server, and the other edge mobile devices are idle. Although this method can achieve high accuracy, it will cause the model training time to be too long and inefficient. Summary of the invention

[0005] In view of the problem of limited wireless resources of participating user devices in the Internet of Things scenario, the present invention aims to develop an efficient wireless federated segmentation learning algorithm for collaborative intelligence of the Internet of Things. The present invention reduces the communication overhead and computing overhead caused by training on the basis of ensuring learning performance. In addition, on this basis, the present invention combines energy consumption and delay to perform optimal scheduling of wireless resources (computing, communication, energy, etc.), selects appropriate participating user devices to participate in training in each communication round, and minimizes the weighted sum of the delay cost and energy consumption cost of the entire training process, meeting the compromise requirements of low delay and low energy consumption of participating user devices with limited resources in the Internet of Things.

[0006] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0007] A wireless federation segmentation learning algorithm for collaborative intelligence of the Internet of Things of the present invention comprises the following steps:

[0008] Step S1: The edge server constructs a device scheduling optimization model based on the weighted sum of the system's minimized delay cost and energy consumption cost, generates a scheduling strategy for participating user devices, and selects participating user devices;

[0009] Step S2: Determine whether the local model parameters of the selected participating user device are the latest. If not, the selected participating user device needs to download the local model of the current global iterative training from the edge server; if it is the latest, no download is required;

[0010] Step S3: Execute forward propagation, segmentation layer data upload, back propagation, gradient averaging, segmentation layer average gradient download and back transmission;

[0011] S3.1: All participating user devices simultaneously perform forward propagation to obtain the output of the segmentation layer and upload it and the corresponding sample labels to the edge server;

[0012] S32: After receiving the output result of the segmentation layer, the edge server performs forward propagation and backward propagation in parallel, and then the edge server averages the gradients formed by all participating user devices, and transmits the averaged result back to each participating user device, and then the edge server updates the global model;

[0013] S3.3: After receiving the returned gradient, each participating user device performs back propagation and then performs local model update;

[0014] S4: The participating user device with the best channel quality among all participating user devices uploads the updated local model to the edge server;

[0015] S5: Repeat steps S1 to S4 until the global model converges or the maximum communication round limit is reached.

[0016] Furthermore, in step S1, the equipment scheduling optimization model constructed is:

[0017]

[0018]

[0019]

[0020] N min ≤m≤N

[0021]

[0022] In the formula, represents the scheduling index of the user equipment n participating in the tth communication round, ρ 1 and ρ 2 Represent the weighted coefficients of delay cost and energy consumption cost, τ t and E t They represent the delay and energy consumption of the tth communication round, m represents the number of participating user devices selected in the tth communication round, and N min represents the minimum limit of the number of participating user devices selected in the tth communication round, N represents the total number of participating user devices, Represents the channel state of user equipment n participating in the tth communication round.

[0023] Furthermore, the specific operation process of step S2 is as follows:

[0024] The edge server uses the channel state information uploaded by the participating user devices to calculate the channel gain of the participating user device n as:

[0025]

[0026] In the formula, h n represents the channel gain of participating user equipment n, c n represents a complex Gaussian random variable, and c n ~CN(0,1) represents the attenuation of small-scale signals, d n represents the distance from the participating user device n to the edge server, and β represents the path attenuation index;

[0027] definition represents the channel state of user device n participating in the tth communication round. When , it means that the participating user device n is selected for the second time in a row; when , it means that the participating user device n is not selected for the second time in a row, and it is assumed that the participating user device n selected in the first round Determine whether the local model parameters of the selected participating user device n are the latest. If not, the selected participating user device n needs to download the local model w of the current global iterative training from the edge server accordingly. u,t On the contrary, if it is the latest version, no download is required.

[0028] Furthermore, the specific operation process of step S3.1 is as follows:

[0029] The m participating user devices simultaneously perform the forward propagation of a small batch of data samples, namely Mini-Batch, to obtain the output activation data of the segmentation layer. And upload it to the edge server with the corresponding data sample label.

[0030] Furthermore, the specific operation process of step S3.2 is as follows:

[0031] The edge server receives the output activation data of the segmentation layer After that, forward propagation and back propagation are performed in parallel, and then the edge server forms the segmentation layer gradients for all participating user devices. Average The average result is The data is sent back to each participating user device, where l represents the segmentation layer and l = 1, 2, …, L-1; the edge server updates the global model, and the global model w e,t The update process is:

[0032]

[0033]

[0034] In the formula, represents the loss function of participating user device n, η represents the learning rate, represents the high-level model deployed on the edge server side of the user device n participating in the tth and t+1th communication rounds, respectively, and w e,t+1 It represents the edge server model after all participating devices aggregate and update in t+1 communication rounds.

[0035] The global loss function expression is:

[0036]

[0037] In the formula, w represents the entire network model, expressed as w = {w u ;w e},w uand w e They represent the models deployed on the edge device and edge server, respectively. n (w) represents the loss function of the participating user device n, which is expressed as:

[0038]

[0039] In the formula, represents the local data set of participating user device n, and the local data set is denoted as N = {1, 2, ..., N}, where D n =|D n |、x i ∈R Q×1 With y i ∈R 1×1 denote the size of the local data set, the i-th input data sample, and the label of the i-th input data sample, respectively. Q denotes the dimension of the i-th input data sample. f(x i ,y i ; w) represents the loss function of each data sample.

[0040] Furthermore, the specific operation process of step S3.3 is as follows:

[0041] Each participating user device receives the average gradient of the segmentation layer transmitted back by the edge server After that, back propagation is performed, and then the local model is updated. The process of local model update is:

[0042]

[0043]

[0044] In the formula, They represent the low-level model of the user device n participating in the tth and t+1th communication rounds deployed on the edge device side, represents the gradient of the low-level model parameters of the user device n deployed on the edge device side in the tth communication round, w u,t+1 Represents the edge device model after aggregation and update of all participating user devices in the t+1th communication round.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention designs a wireless federation segmentation learning algorithm framework for IoT collaborative intelligence, and performs device scheduling on this basis, so that resource-constrained participating user devices can efficiently participate in global model training. On the one hand, the present invention can reduce the communication delay and transmission energy consumption caused by frequent interactions between edge servers and participating user devices; on the other hand, the present invention implements the participating user device scheduling strategy based on the designed wireless federation segmentation learning algorithm framework, which can effectively overcome the heterogeneity problems (such as device power failure and poor channel conditions) brought about by the training of participating user devices while ensuring learning performance, and reduce system energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 The present invention is a flow chart of a wireless federation segmentation learning algorithm for collaborative intelligence of the Internet of Things.

[0049] Figure 2 This is an architectural diagram of a wireless federated segmentation learning algorithm framework for collaborative intelligence of the Internet of Things according to the present invention.

[0050] Figure 3 This is a learning performance curve obtained in computer simulation.

[0051] Figure 4 This is a graph of delay and energy consumption cost under different system bandwidths obtained in computer simulation.

[0052] Figure 5 This is a graph of latency and energy cost under different edge server computing powers obtained in computer simulation. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments 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 creative work are within the scope of protection of the present invention.

[0054] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0055] The present invention proposes a wireless federated segmentation learning algorithm for collaborative intelligence of the Internet of Things, which is mainly based on a wireless federated segmentation learning algorithm framework for collaborative intelligence of the Internet of Things (Parallel Split Learning with Gradient Average, PSLAVG). By performing gradient averaging and selective downloading of local models of participating user devices, the communication delay and transmission energy consumption are greatly reduced. In addition, the present invention also overcomes the problem of device heterogeneity by executing scheduling selection of participating user devices, makes more full use of distributed resources and wireless resources, and selects the best participating user devices to participate in global model training by constructing a scheduling strategy for participating user devices.

[0056] See also Figure 1 To illustrate, the present invention provides a wireless federation segmentation learning algorithm for IoT collaborative intelligence, and the specific operation process is as follows:

[0057] Step S1: The edge server constructs a participating user device scheduling strategy and selects participating user devices;

[0058] The edge server builds a device scheduling optimization model based on the weighted sum of the system's minimized latency cost and energy consumption cost, generates a scheduling strategy for participating user devices, and selects participating user devices; the specific operation steps are as follows:

[0059] Step S1.1: Calculate the computation delay of participating user equipment;

[0060] The decision of user equipment scheduling is determined by the total system energy consumption including learning delay and learning energy consumption. Learning delay mainly consists of two parts: computing delay and transmission delay. Correspondingly, learning energy consumption mainly consists of two parts: computing energy consumption and transmission energy consumption.

[0061] definition and They represent the number of CPU cycles (in cycles / sample) required by the participating user device to process a data sample during the forward propagation and backward propagation of the local model segmentation layer l, respectively. The calculation delay of the participating user device n is:

[0062]

[0063]

[0064] in, and They represent the forward propagation calculation delay and the reverse propagation calculation delay of the participating user equipment n, respectively. n Indicates the size of Mini-Batch, f n Indicates the CPU main frequency of participating user device n (in cycles / s).

[0065] Step S1.2: Calculate the uplink transmission rate of the participating user equipment;

[0066] In order to obtain high spectrum efficiency and ensure communication quality, non-orthogonal multiple access (NOMA) technology is used for parameter transmission. For the uplink, the strength of the signal received by each participating user device depends on the channel gain. Therefore, the edge servers are arranged in descending order according to the channel gain as follows:

[0067] |h 1 | 2 ≥|h 2 | 2 ≥...≥|h n | 2 ≥...≥|h N | 2

[0068] Among them, h 1 ,h 2 …h n …h N They represent the channel gains of participating user equipments 1, 2…n…N respectively.

[0069] Ignoring the propagation error, and according to Shannon's formula, the uplink transmission rate of participating user equipment n is for:

[0070]

[0071] Where W represents the system bandwidth, p n represents the transmission power of participating user equipment n, p k Indicates that the channel gain is less than or equal to the transmit power of the participating user equipment k of the participating user equipment n, h k represents the channel gain of participating user equipment k that is less than or equal to the channel gain of participating user equipment n, σ 2 Represents the noise power received by the edge server.

[0072] Step S1.3: Calculate the downlink transmission rate of the participating user equipment;

[0073] Unlike the uplink, the signal strength received by each participating user device no longer depends solely on the channel gain, because the edge server will reasonably allocate the transmission power according to the channel status of each participating user device. Therefore, the downlink transmission rate of participating user device n for:

[0074]

[0075] Among them, p' n represents the transmission power of the edge server to the participating user device n and satisfies P max represents the maximum transmission power of the edge server, p' k represents the transmission power of the edge server to the participating user device k whose channel gain is greater than or equal to the participating user device n, σ n 2 Indicates the noise power received by each participating user equipment.

[0076] Step S1.3: Calculate the uplink transmission delay of the participating user equipment;

[0077] definition represents the output size (in bits) of a data sample when the participating user equipment performs the local model segmentation layer l, then the uplink transmission delay of the participating user equipment n is for:

[0078]

[0079] In the formula, represents the uplink transmission rate of user equipment n participating in the tth communication round.

[0080] Step S1.4: Calculate the downlink transmission delay of the participating user equipment;

[0081] definition represents the average gradient size (in bits) of the data samples of the local model layer l+1 of the participating user device n, then the downlink transmission delay of the participating user device n is for:

[0082]

[0083] Where p' n,t represents the transmission power of the edge server to the participating user device n in the tth communication round, represents the channel gain of user equipment n participating in the tth communication round.

[0084] Step S1.5: Calculate the uplink and downlink parameter transmission delays of the participating user equipment;

[0085] definition represents the size of the local model of the participating user equipment n (in bits), and the uplink and downlink parameter transmission delays of the participating user equipment n are:

[0086]

[0087]

[0088] in, and They represent the uplink parameter transmission delay and downlink parameter transmission delay of user equipment n participating in the tth communication round, respectively, and p n,t represents the transmission power of user device n participating in the tth communication round to the edge server, represents the channel gain of user equipment n participating in the tth communication round.

[0089] Step S1.6: Calculate the computational delay of each participating user device for training a Mini-Batch;

[0090] definition represents the number of CPU cycles (in cycles / sample) required by the edge server to calculate a data sample (including forward propagation and back propagation processes). The computational delay of the edge server for training a Mini-Batch for each participating user device is:

[0091]

[0092] Where m represents the number of participating user devices selected for scheduling, and f e Indicates the CPU main frequency of the edge server (in cycles / s) and Indicates the maximum CPU frequency of the edge server (in cycles / s).

[0093] Step S1.7: In summary, for all participating user devices in the tth communication round, the overall delay cost and energy consumption cost can be modeled as:

[0094]

[0095]

[0096]

[0097] in, represents the energy consumption of all participating user devices in the tth communication round, They represent the transmission energy consumption and computing energy consumption of all participating user devices in the tth communication round respectively. κB represents the effective capacitance exchange coefficient, which is related to the chip architecture.

[0098] Step S1.8: Based on the above definition and analysis, the constructed equipment scheduling optimization model can be expressed as:

[0099]

[0100]

[0101]

[0102] N min ≤m≤N

[0103]

[0104] in, represents the scheduling index of the user equipment n participating in the tth communication round, ρ 1 and ρ 2 Represent the weighted coefficients of delay cost and energy consumption cost, τ t and E t They represent the delay and energy consumption of the tth communication round, m represents the number of participating user devices selected in the tth communication round, and N min represents the minimum limit of the number of participating user devices selected in the tth communication round, N represents the total number of participating user devices, Represents the channel state of user equipment n participating in the tth communication round.

[0105] By solving the above optimization problem about device scheduling, the optimal decision-making scheme for the selection of participating user devices is obtained. The above optimization problem about device scheduling is a mixed integer nonlinear programming (MINLP) that is difficult to solve directly, so the Dinkelbach transform (DT) method is used to convert the problem into a solvable convex optimization problem, and then iteratively solve it to obtain the optimal solution. After obtaining the optimal solution of the tth communication round, the edge server selects the corresponding participating user devices to participate in the global iterative training.

[0106] Step S2: downloading local model;

[0107] Determine whether the local model parameters of the selected participating user device are the latest. If not, the corresponding selected participating user device needs to download the local model of the current global iterative training from the edge server; otherwise, if it is the latest, it does not need to be downloaded. In addition, the edge server needs to select the participating user device with the best channel quality from all participating user devices according to the quality of the participating user device channel. best , that is, making decisions based on the channel status of participating user devices.

[0108] Specifically, the edge server uses the channel state information uploaded by the participating user equipment to calculate the channel gain of the participating user equipment n as:

[0109]

[0110] Among them, h n represents the channel gain of participating user equipment n, c n represents a complex Gaussian random variable, and c n ~CN(0,1) represents the attenuation of small-scale signals, d n represents the distance from the participating user device n to the edge server, and β represents the path attenuation exponent.

[0111] definition represents the channel state of user device n participating in the tth communication round. When , it means that the participating user device n is selected for the second time; on the contrary, when , it means that the participating user device n is not selected for the second time in a row, and it is assumed that the participating user device n selected in the first round Determine whether the local model parameters of the selected participating user device n are the latest. If not, the selected participating user device n needs to download the local model w of the current global iterative training from the edge server accordingly. u,t On the contrary, if it is the latest version, no download is required.

[0112] Step S3: Based on the wireless federated segmentation learning algorithm framework for IoT collaborative intelligence, forward propagation, segmentation layer data upload, back propagation, gradient averaging, segmentation layer average gradient download and backhaul are performed; the specific operation steps are as follows:

[0113] See also Figure 2 To illustrate, the wireless federated segmentation learning algorithm framework for IoT collaborative intelligence of the present invention is mainly composed of an edge platform and multiple resource-constrained participating user devices (1...n...N), and the edge platform is composed of a base station and an edge server. Among them, the participating user devices are used to perform local model The local data set (D 1 ...Dn ...D N ) and the output data of the segmentation layer (a 1 ……a n ……a N ) is uploaded to the edge server for subsequent global iterative training; the edge server is responsible for the subsequent forward propagation of the local model, the reverse propagation of the local model, and the average gradient of the segmentation layer Download and upload, global model Update and aggregate.

[0114] Step S3.1: m participating user devices simultaneously perform forward propagation of a small batch of data samples, i.e., Mini-Batch, to obtain the output activation data of the segmentation layer And upload it to the edge server with the corresponding data sample label.

[0115] Step S3.2: The edge server receives the output activation data of the segmentation layer After that, forward propagation and back propagation are performed in parallel, and then the edge server forms the segmentation layer gradients for all participating user devices. Average The average result is is sent back to each participating user device, where l represents the segmentation layer, and l = 1, 2, ..., L-1. Subsequently, the edge server updates the global model. Accordingly, the edge server performs global model w e,t The update process is:

[0116]

[0117]

[0118] In the formula, represents the loss function of participating user device n, η represents the learning rate, represents the high-level model deployed on the edge server side of the user device n participating in the tth and t+1th communication rounds, respectively, and w e,t+1 It represents the edge server model after all participating devices aggregate and update in t+1 communication rounds.

[0119] The goal of federated segmentation learning is to train a global model that has good inference performance for all participating user devices. Mathematically, the goal of training is to train a minimum global loss function. Therefore, the global loss function is expressed as:

[0120]

[0121] In the formula, w represents the entire network model, expressed as w = {w u ;w e},w u and w e They represent the models deployed on the edge device and edge server, respectively. n (w) represents the loss function of participating user device n, which is expressed as:

[0122]

[0123] in, represents the local data set of participating user device n, and the local data set is denoted as N = {1, 2, ..., N}, where D n =|D n |、x i ∈R Q×1 With y i ∈R 1×1 denote the size of the local data set, the i-th input data sample, and the label of the i-th input data sample, respectively. Q denotes the dimension of the i-th input data sample. f(x i ,y i ; w) represents the loss function of each data sample.

[0124] Step S3.3: Each participating user device receives the average gradient of the segmentation layer transmitted back by the edge server Then perform back propagation and then update the local model. Accordingly, the local model update process is:

[0125]

[0126]

[0127] In the formula, They represent the low-level model of the user device n participating in the tth and t+1th communication rounds deployed on the edge device side, represents the gradient of the low-level model parameters of the user device n deployed on the edge device side in the tth communication round, w u,t+1 Represents the edge device model after aggregation and update of all participating user devices in the t+1th communication round.

[0128] S4: The participating user equipment n with the best channel quality among all participating user equipments best Upload the updated local model to the edge server.

[0129] S5: Repeat the above steps S1 to S4 until the global model converges or the maximum communication round limit is reached.

[0130] The present invention can reduce the communication delay and transmission energy consumption caused by frequent interactions between the edge server and the participating user devices by performing training, gradient averaging, and selective downloading of local models of participating user devices in parallel. At the same time, through the transmission of the output results of the split layer and the collaborative calculation of the participating user devices and the edge server, the computing and resource storage utilization of the edge platform are effectively improved.

[0131] The application of a wireless federation segmentation learning algorithm for collaborative intelligence of the Internet of Things provided by the present invention can realize the scheduling of different types of participating user devices, which helps to improve the utilization rate of wireless network resources and alleviate the problem of limited resources of participating user devices under the Internet of Things.

[0132] The feasibility of a wireless federation segmentation learning algorithm for collaborative intelligence of the Internet of Things provided by the present invention is verified by a specific computer simulation test.

[0133] In computer simulation, an IoT network with 60 resource-constrained participating user devices is considered. The learning performance of the proposed wireless federation segmentation learning algorithm is evaluated by training ResNet-18 on the ChestXRay 2017 dataset.

[0134] (1) To verify the learning performance of the wireless federation segmentation learning algorithm proposed in the present invention, the following four comparison schemes are considered:

[0135] Baseline 1 (CL): All participating user devices send their local datasets to the edge platform for centralized training;

[0136] Baseline 2 (FL): In the federated learning method, all participating user devices jointly train a model, and then transmit their trained models to the edge server to perform model aggregation in each training round;

[0137] Baseline 3 (SSL): All participating user devices conduct model training with the edge server one by one;

[0138] Baseline 4 (PSL): All participating user devices perform model training in parallel with the edge server;

[0139] PSLAVG: Model training is performed according to the wireless federation segmentation learning algorithm of the present invention.

[0140] See also Figure 3Based on the above computer simulation settings, the learning performance achieved by various comparison schemes is demonstrated. On the one hand, the simulation results show that the wireless federated segmentation learning algorithm proposed in the present invention achieves learning performance similar to that of baseline 2 (FL) and baseline 3 (SSL), which is worse than baseline 1 (CL) and baseline 2 (FL); however, the energy consumption cost and delay cost consumed by the wireless federated segmentation learning algorithm proposed in the present invention are the lowest. On the other hand, the simulation results show that the wireless federated segmentation learning algorithm proposed in the present invention performs gradient averaging and selective downloading of local models, which reduces communication overhead and computational overhead compared to baseline 3 (SSL), while the prediction accuracy remains basically unchanged.

[0141] (2) To verify the cost of device scheduling using the wireless federation segmentation learning algorithm proposed in the present invention, the following three comparison schemes are considered:

[0142] Baseline 1: All participating user devices participate in training, and no optimization scheduling is performed;

[0143] Baseline 2: In each training round, all participating user devices are randomly selected to participate in training;

[0144] Baseline 3: In each training round, the user equipment with the best channel quality is always randomly selected to participate in the training;

[0145] PSLAVG: Model training is performed according to the wireless federation segmentation learning algorithm of the present invention.

[0146] See also Figure 4 Based on the above computer simulation settings, the costs of various schemes under different system bandwidths are demonstrated. On the one hand, the simulation results show that the wireless federation segmentation learning algorithm proposed in the present invention can achieve the minimum energy consumption cost and delay cost, and the energy consumption and delay costs are always kept at a low level, which reveals the superiority and necessity of participating in the user equipment scheduling selection. On the other hand, the simulation results show that the performance gap between the wireless federation segmentation learning algorithm proposed in the present invention and other schemes narrows with the increase of system bandwidth. This is because the system energy consumption and delay mainly depend on the communication delay and transmission energy consumption. Increasing the system bandwidth is conducive to reducing communication delay and transmission energy consumption. This also indirectly shows that the wireless federation segmentation learning algorithm proposed in the present invention mainly optimizes the energy consumption and delay caused by the frequent data exchange between the participating user equipment and the edge server.

[0147] See also Figure 5Based on the above computer simulation settings, the costs of various schemes under different edge server computing powers are demonstrated. On the one hand, the simulation results show that the wireless federation segmentation learning algorithm proposed in the present invention can achieve the minimum energy consumption cost and delay cost, and the energy consumption and delay costs are always kept at a low level, which also reveals the superiority and necessity of participating in the user equipment scheduling selection. On the other hand, the simulation results show that the performance gap between the wireless federation segmentation learning algorithm proposed in the present invention and other schemes maintains a balance as the computing power of the edge server increases. This is because the system energy consumption and delay mainly depend on the communication delay and transmission energy consumption. Increasing the computing power of the edge server is conducive to reducing the computing delay and energy consumption of the edge server, but when it increases to a certain extent, the reduction in computing delay and energy consumption brought about by this improvement is very small compared to the communication energy consumption and transmission delay.

[0148] The above is only a preferred embodiment 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 principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A wireless federated segmentation learning algorithm for collaborative intelligence in the Internet of Things. It is characterized in that The following steps are involved: Step S1: The edge server constructs a device scheduling optimization model based on the weighted sum of the system's minimized delay cost and energy consumption cost, generates a scheduling strategy for participating user devices, and selects participating user devices; The equipment scheduling optimization model constructed is: N min ≤m≤N In the formula, represents the scheduling index of the user equipment n participating in the tth communication round, ρ 1 and ρ 2 Represent the weighted coefficients of delay cost and energy consumption cost, τ t and E t They represent the delay and energy consumption of the tth communication round, m represents the number of participating user devices selected in the tth communication round, and N min represents the minimum limit of the number of participating user devices selected in the tth communication round, N represents the total number of participating user devices, represents the channel state of user equipment n participating in the tth communication round; Step S2: Determine whether the local model parameters of the selected participating user device are the latest. If not, the selected participating user device needs to download the local model of the current global iterative training from the edge server; if it is the latest, no download is required; Step S3: Execute forward propagation, segmentation layer data upload, back propagation, gradient averaging, segmentation layer average gradient download and back transmission; S3.1: All participating user devices simultaneously perform forward propagation to obtain the output of the segmentation layer and upload it and the corresponding sample labels to the edge server; S3.2: After receiving the output of the segmentation layer, the edge server performs forward propagation and backward propagation in parallel. Then, the edge server averages the gradients formed by all participating user devices and transmits the averaged result back to each participating user device. Then, the edge server performs a global model update. The edge server receives the output activation data of the segmentation layer After that, forward propagation and back propagation are performed in parallel, and then the edge server forms the segmentation layer gradients for all participating user devices. Average The average result is The data is sent back to each participating user device, where l represents the segmentation layer and l = 1, 2, …, L-1; the edge server updates the global model, and the global model w e,t The update process is: In the formula, represents the loss function of participating user device n, η represents the learning rate, represents the high-level model deployed on the edge server side of the user device n participating in the tth and t+1th communication rounds, respectively, and w e,t+1 Represents the edge server model after all participating devices aggregate and update in t+1 communication rounds; The global loss function expression is: In the formula, w represents the entire network model, expressed as w = {w u ;w e },w u and w e They represent the models deployed on the edge device and edge server, respectively. n (w) represents the loss function of the participating user device n, which is expressed as: In the formula, represents the local data set of participating user device n. The local data set is represented as in and denote the size of the local data set, the i-th input data sample, and the label of the i-th input data sample, respectively. Q denotes the dimension of the i-th input data sample. f(x i ,y i ; w) represents the loss function of each data sample; S3.3: After receiving the returned gradient, each participating user device performs back propagation and then performs local model update; Each participating user device receives the average gradient of the segmentation layer transmitted back by the edge server After that, back propagation is performed, and then the local model is updated. The process of local model update is: In the formula, They represent the low-level model of the user device n participating in the tth and t+1th communication rounds deployed on the edge device side, represents the gradient of the low-level model parameters of the user device n deployed on the edge device side in the tth communication round, w u,t+1 Represents the edge device model after aggregation and update of all participating user devices in the t+1th communication round; S4: The participating user device with the best channel quality among all participating user devices uploads the updated local model to the edge server; S5: Repeat steps S1 to S4 until the global model converges or the maximum communication round limit is reached.

2. According to the wireless federation segmentation learning algorithm for collaborative intelligence of the Internet of Things according to claim 1, It is characterized in that The specific operation process of step S2 is as follows: The edge server uses the channel state information uploaded by the participating user devices to calculate the channel gain of the participating user device n as: In the formula, h n represents the channel gain of participating user equipment n, c n represents a complex Gaussian random variable, and c n ~CN(0,1) represents the attenuation of small-scale signals, d n represents the distance from the participating user device n to the edge server, and β represents the path attenuation index; definition represents the channel state of user device n participating in the tth communication round. When , it means that the participating user device n is selected for the second time in a row; when , it means that the participating user device n is not selected for the second time in a row, and it is assumed that the participating user device n selected in the first round Determine whether the local model parameters of the selected participating user device n are the latest. If not, the selected participating user device n needs to download the local model w of the current global iterative training from the edge server accordingly. u,t ; On the contrary, if it is the latest version, no download is required.

3. According to the wireless federation segmentation learning algorithm for collaborative intelligence of the Internet of Things according to claim 1, It is characterized in that The specific operation process of step S3.1 is as follows: The m participating user devices simultaneously perform the forward propagation of a small batch of data samples, namely Mini-Batch, to obtain the output activation data of the segmentation layer. And upload it to the edge server with the corresponding data sample label.

Images