Iot signal repair method and apparatus

Abstract

The method and device for repairing Internet of Things signals complete the whole process of loss statistics, noise modeling, weight generation and sample separation locally at each Internet of Things node, do not need to upload the original data of the nodes, fully fit the core privacy protection characteristics of federated learning, realize the fine quantification of noise through data-driven loss distribution statistics, break through the ideal data hypothesis limitations of traditional schemes on sample balance, data isomorphism and modal integrity, can strongly adapt to complex non-ideal working conditions in actual deployment of the Internet of Things, and the finally output weight parameters and sample separation results can suppress the interference of noise on federated collaborative training from the data source level, provide high-reliability sample support and weight basis for subsequent robust federated model training, and effectively solve the pain points that the existing technology cannot adapt to heterogeneous scenes and has great difficulty in cross-domain adaptation.

Description

Technical Field

[0001] This application relates to the field of the Internet of Things (IoT), specifically to an IoT signal repair method and apparatus. Background Technology

[0002] Currently, IoT technology has been deployed on a large scale in many fields such as the Industrial Internet, smart cities, and environmental monitoring. These scenarios typically involve a large number of distributed subnets and heterogeneous sensor nodes, generating massive amounts of multimodal sensor data. Intelligent analysis and pattern mining based on multimodal data are the core foundation for realizing intelligent control and decision-making in IoT. Meanwhile, federated learning technology, with its privacy-preserving characteristic of "data remains still, model moves," can achieve cross-node collaborative model training without disclosing the original data of each subnet, and has become the mainstream technical solution for multi-source data collaborative modeling in IoT scenarios.

[0003] In related technologies, federated learning schemes for multimodal data are mostly designed around the ideal scenario of homogeneous data and complete modalities, and achieve collaborative updates of the global model through standard parameter aggregation mechanisms.

[0004] However, it is difficult to adapt to complex scenarios such as heterogeneous signal distribution, data imbalance, and modal missing among IoT subnets. The ideal data assumptions it relies on are seriously inconsistent with actual working conditions. It cannot accurately model the complex noise in multi-source heterogeneous data, resulting in the inability to effectively unify the representation space of multi-modal data and making cross-domain adaptation extremely difficult. Summary of the Invention

[0005] This application provides an IoT signal repair method and apparatus, which can solve the technical problems in related technologies that make it difficult to adapt to complex scenarios such as heterogeneous signal distribution, data imbalance, and modal missing among IoT subnets, and make it impossible to accurately model complex noise in multi-source heterogeneous data, thus leading to the inability to effectively unify the multimodal data representation space and the great difficulty of cross-domain adaptation.

[0006] In a first aspect, embodiments of this application provide an Internet of Things (IoT) signal repair method, the IoT signal repair method comprising: Obtain the local datasets of each IoT node in the federated learning system, as well as the local models corresponding to each node; Based on the prediction results of the local dataset and the local model, the data loss distribution statistics and noise distribution modeling are completed, and weight parameters that are adapted to the degree of data noise are generated. Based on the weight parameters, the local dataset is purified and separated to obtain a clean sample set and a noisy sample set, and the weight parameters and sample separation results are output for training the federated model. The global model of federated learning is initialized. Based on the output weight parameters, clean sample set and noisy sample set, the local model of each node is repaired and trained through knowledge distillation to obtain the model update amount of each node.

[0007] In conjunction with the first aspect, in one implementation, the step of completing the data loss distribution statistics based on the prediction results of the local dataset and the local model includes: For each IoT node, the prediction loss of a single sample in the local dataset is calculated based on the prediction results of the local model and the sample labels. Using data category as the statistical dimension and combining it with the overall loss distribution benchmark within the node, the relative loss deviation corresponding to each category of data is calculated and used as the statistical result of loss distribution.

[0008] In conjunction with the first aspect, in one implementation, the step of calculating the relative loss deviation corresponding to each data category, using data category as the statistical dimension and combining it with the overall loss distribution benchmark within the node, includes: Calculate the mean and variance of class loss for each category of samples within each node, based on the data category dimension. Based on the loss statistics of all categories within each node, the overall loss distribution benchmark corresponding to that node is calculated. The overall loss distribution benchmark includes the node's average category loss and the standard deviation of the node's category loss. By combining the category loss mean of each category with the overall loss distribution benchmark of the corresponding node, the relative loss deviation of each category is calculated.

[0009] In conjunction with the first aspect, in one implementation, the step of completing the data loss distribution statistics and noise distribution modeling includes: Based on the statistical results of the loss distribution, the data loss is modeled as a hybrid distribution model consisting of the superposition of the clean sample distribution and the noisy sample distribution; The model parameters of the mixed distribution model are solved by iterative optimization. Based on the solved model parameters, the noise probability corresponding to each data point is calculated, and the noise probability is determined as the noise quantization result.

[0010] In conjunction with the first aspect, in one implementation, generating weight parameters adapted to the level of data noise includes: Based on the noise quantization results, the noise uncertainty of the corresponding data is calculated; Based on the noise uncertainty, a weight parameter negatively correlated with the noise level is generated, wherein the higher the noise level of the data, the lower the corresponding weight parameter.

[0011] In conjunction with the first aspect, in one implementation, the step of purifying and separating the local dataset based on the weight parameters to obtain a clean sample set and a noisy sample set includes: Based on the generated weight parameters, combined with nonlinear filtering, a clean sample set is obtained from the local dataset. By using set difference operations, a noisy sample set that has no intersection with the clean sample set is partitioned from the local dataset.

[0012] In conjunction with the first aspect, in one implementation, the step of completing the repair training of the local model at each node through knowledge distillation to obtain the model update amount for each node includes: A knowledge distillation training framework is constructed using the global model as the teacher model and the node local model as the student model. Based on a clean sample set, the task loss of the local model and the guided distillation loss of the global model are calculated separately. By combining the proportion of clean and noisy samples in the local dataset, the task loss and distillation loss are weighted and fused to obtain the total training loss; The local model is iteratively trained based on the total training loss, and the parameter update amount of the local model relative to the global model is calculated after training is completed.

[0013] In conjunction with the first aspect, in one embodiment, the method further includes: The model update volume of each node is evaluated for reliability in multiple dimensions. Based on the evaluation results, adaptive aggregation weight adjustment is completed, and the global model is updated in combination with the adjusted weight.

[0014] In conjunction with the first aspect, in one implementation, the step of performing a multi-dimensional credibility assessment on the model update amount of each node, completing adaptive aggregation weight adjustment based on the assessment results, and updating the global model with the adjusted weights includes: Calculate the fusion similarity, historical update consistency and node data quality index of the model update volume of each node, and then weightedly fuse them to obtain the credibility score of the model update volume of each node. Based on the credibility scores of all participating training nodes, the aggregate weights of each node are adaptively reduced. Based on the adjusted aggregation weights, the model update amounts of each node are weighted and aggregated to update the parameters of the global model; Based on the updated global model, the loss distribution statistics, noise distribution modeling, and sample purification and separation steps are re-executed to complete the closed-loop update of sample partitioning and enter the next round of federated learning iteration.

[0015] Secondly, embodiments of this application provide an Internet of Things (IoT) signal repair device, comprising: The data acquisition module is used to acquire the local datasets of each IoT node in the federated learning system, as well as the local models corresponding to each node. The noise modeling and weight generation module is used to complete the data loss distribution statistics and noise distribution modeling based on the prediction results of the local dataset and the local model, and generate weight parameters that are adapted to the degree of data noise. The sample purification and separation module is used to purify and separate the local dataset based on the weight parameters to obtain a clean sample set and a noisy sample set, and output the weight parameters and sample separation results for training the federated model. The model repair and training module is used to initialize the global model of federated learning. Based on the output weight parameters, clean sample set and noisy sample set, it completes the repair training of the local model of each node through knowledge distillation, and obtains the model update amount of each node.

[0016] The beneficial effects of the technical solutions provided in this application include: By completing the entire process of loss statistics, noise modeling, weight generation, and sample separation locally on each IoT node, without uploading the original node data, it fully aligns with the core privacy protection characteristics of federated learning. Through data-driven loss distribution statistics, it achieves refined noise quantification, breaking through the limitations of traditional solutions' ideal data assumptions regarding sample balance, data isomorphism, and modal integrity. This allows for strong adaptation to complex and non-ideal working conditions in actual IoT deployments. The final output weight parameters and sample separation results can suppress noise interference in federated collaborative training at the data source level, providing highly reliable sample support and weight basis for subsequent robust federated model training. This effectively solves the pain points of existing technologies being unable to adapt to heterogeneous scenarios and facing significant difficulties in cross-domain adaptation. Based on the sample separation results and weight parameters obtained in the aforementioned steps, it completes noise-robust federated collaborative training and model repair. Through knowledge distillation, it achieves the transfer of globally reliable knowledge to the local model, correcting feature biases in the local model caused by noise. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the IoT signal repair method of this application; Figure 2 This is a flowchart illustrating another embodiment of the IoT signal repair method of this application; Figure 3 This is a schematic diagram of the structure of the IoT signal repair device in the embodiments of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0019] It's important to understand that IoT technology has been deployed on a large scale in many fields such as the Industrial Internet, smart cities, and environmental monitoring. These scenarios typically involve numerous distributed subnets and heterogeneous sensor nodes, generating massive amounts of multimodal sensor data. Intelligent analysis and pattern mining based on this multimodal data are the core foundation for achieving intelligent control and decision-making in the IoT. Meanwhile, federated learning technology, with its privacy-preserving characteristic of "the model moves while the data remains still," can achieve cross-node collaborative model training without disclosing the original data of each subnet, and has become the mainstream technical solution for multi-source data collaborative modeling in IoT scenarios.

[0020] Among them, federated learning schemes for multimodal data are mostly designed around the ideal scenario of homogeneous data and complete modalities, and achieve collaborative updates of the global model through standard parameter aggregation mechanisms.

[0021] However, it is difficult to adapt to complex scenarios such as heterogeneous signal distribution, data imbalance, and modal missing among IoT subnets. The ideal data assumptions it relies on are seriously inconsistent with actual working conditions. It cannot accurately model the complex noise in multi-source heterogeneous data, resulting in the inability to effectively unify the representation space of multi-modal data and making cross-domain adaptation extremely difficult.

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0023] In one aspect, embodiments of this application provide an Internet of Things (IoT) signal repair method.

[0024] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the IoT signal repair method of this application. Figure 1 As shown, the IoT signal repair methods include: S100: Obtain the local datasets of each IoT node in the federated learning system, as well as the local models corresponding to each node; S200: Based on the prediction results of the local dataset and the local model, complete the statistical analysis of data loss distribution and the modeling of noise distribution, and generate weight parameters that are compatible with the degree of data noise. S300: Based on the weight parameters, perform sample purification and separation on the local dataset to obtain a clean sample set and a noisy sample set, and output the weight parameters and sample separation results for federated model training; S400: Initialize the global model for federated learning. Based on the output weight parameters, clean sample set, and noisy sample set, complete the repair training of the local model of each node through knowledge distillation to obtain the model update amount of each node.

[0025] In this embodiment, the core problems commonly encountered in IoT federated learning scenarios, such as heterogeneous signal distribution, data imbalance, and inaccurate modeling of complex noise due to modal missing features, as well as sample contamination interfering with model training, are addressed. This is achieved by completing the entire process of loss statistics, noise modeling, weight generation, and sample separation locally on each IoT node, without uploading the original node data, thus fully aligning with the privacy-preserving core characteristic of federated learning. Data-driven loss distribution statistics enable refined noise quantification, overcoming the limitations of traditional solutions that rely on ideal data assumptions such as balanced samples, isomorphic data, and complete modal features. This approach strongly adapts to the complex and non-ideal working conditions in actual IoT deployments. The final output weight parameters and sample separation results suppress noise interference in federated collaborative training from the data source level, providing highly reliable sample support and weight basis for subsequent robust federated model training. This effectively solves the pain points of existing technologies being unable to adapt to heterogeneous scenarios and facing significant difficulties in cross-domain adaptation. Based on the sample separation results and weight parameters obtained in the aforementioned steps, noise-robust federated collaborative training and model repair are completed. Knowledge distillation enables the transfer of globally reliable knowledge to the local model, correcting feature biases caused by noise in the local model.

[0026] Furthermore, in one embodiment, step S200 includes the following steps: S201: For each IoT node, calculate the prediction loss of a single sample in the local dataset based on the prediction results of the local model and the sample labels; S202: Using data category as the statistical dimension and combining it with the overall loss distribution benchmark within the node, the relative loss deviation corresponding to each category of data is calculated and used as the statistical result of loss distribution.

[0027] In this embodiment, hierarchical loss statistics based on nodes and categories are used to eliminate statistical biases caused by data heterogeneity and category imbalance between nodes, providing an accurate statistical basis for subsequent noise modeling.

[0028] In step S201, it is assumed that the entire federated learning system includes The IoT nodes (subnets) are categorized as follows: Single sample loss The calculation formula is as follows: (1) In the formula, , Let represent the set of all samples of the c-th category in the k-th IoT node, and and That is, the sample set consists of clean samples and noisy samples that have no overlap; This represents the local model of the k-th IoT node. The preset loss function is used. Step S201 quantifies the prediction effect of a single sample, providing the smallest granular loss data for subsequent stratified statistics.

[0029] In step S202, by normalizing the loss statistics of the category dimension and the overall distribution benchmark of the nodes, a relative loss deviation that can characterize the degree of loss deviation at the category level is obtained. This effectively adapts to the data distribution characteristics within the nodes, avoids statistical deviations caused by data heterogeneity between nodes, and provides unified and comparable statistical features for subsequent noise modeling.

[0030] Furthermore, in one embodiment, S202 includes the following steps: S202-1: Calculate the mean and variance of the class loss for each category of samples within each node, according to the data category dimension. S202-2: Based on the loss statistics of all categories within each node, calculate the overall loss distribution benchmark corresponding to that node. The overall loss distribution benchmark includes the node's average category loss and the standard deviation of the node's category loss. S202-3: By combining the average loss of each category with the overall loss distribution benchmark of the corresponding node, the relative loss deviation of each category is calculated.

[0031] In this embodiment, a detailed characterization of the loss distribution of each category within a node is achieved through hierarchical and progressive statistical calculations. At the same time, the impact of data distribution differences between different nodes on loss statistics is eliminated by normalizing the overall distribution benchmark of the nodes, ensuring that the loss characteristics of different nodes and different categories are comparable.

[0032] In step S202-1, the average loss of the c-th category of the k-th IoT node. The calculation formula is as follows: (2) In the formula, This represents the number of samples of the c-th category in the k-th IoT node.

[0033] Variance of corresponding category loss The calculation formula is as follows: (3) The above calculations can accurately depict the central tendency and dispersion of sample loss within each category, providing basic data for subsequent benchmark calculations of the overall node distribution.

[0034] In step S202-2, the average class loss of the k-th IoT node The calculation formula is as follows: (4) The standard deviation of the category loss of the kth IoT node The calculation formula is as follows: (5) Through the above calculations, benchmark parameters characterizing the overall loss distribution of nodes are obtained, providing a normalized benchmark for subsequent relative deviation calculations.

[0035] In step S202-3, the relative loss bias of the c-th category of the k-th IoT node The calculation formula is as follows: (6) By calculating the relative loss deviation, the influence of the difference in loss magnitude between different nodes is eliminated, which can objectively characterize the degree of deviation of a single category relative to the overall distribution of nodes. The degree of loss deviation is directly related to the sample noise level, providing core input features for subsequent noise distribution modeling.

[0036] Furthermore, in one embodiment, step S200 includes the following steps: S203: Based on the statistical results of the loss distribution, the data loss is modeled as a hybrid distribution model consisting of the superposition of the clean sample distribution and the noisy sample distribution; S204: Solve the model parameters of the mixed distribution model through iterative optimization, calculate the noise probability corresponding to each data based on the solved model parameters, and determine the noise probability as the noise quantization result.

[0037] In this embodiment, based on the distribution characteristics of relative loss bias, a two-component Gaussian mixture model is used to jointly model the loss distribution of clean samples and noisy samples, thereby achieving refined and quantifiable modeling of complex noise and solving the problem that traditional solutions cannot effectively characterize complex noise in heterogeneous data.

[0038] In step S203, the relative loss deviation obtained in step S202 is... The model is a linear mixture of Gaussian distributions for clean samples and Gaussian distributions for noisy samples, with the mixture probability distribution being... The calculation formula is as follows: (7) In the formula, The mixing coefficient is the distribution of the clean sample. Let be the mixing coefficient of the noise sample distribution, satisfying + =1; , These are the mean and variance of the Gaussian distribution corresponding to the clean sample, respectively. , Let be the mean and variance of the Gaussian distribution corresponding to the noise sample, respectively, and satisfy . < This ensures that the model converges to the feature distribution of clean samples; This represents the probability density function of a normal distribution. By constructing a bicomponent mixture distribution, the distributional differences in loss bias between clean and noisy samples can be accurately characterized, providing a model basis for noise quantization.

[0039] In step S204, the Expectation-Maximization (EM) algorithm is used to estimate the parameters of the mixed distribution model. Iterative solutions are performed, specifically including: Step E: Calculate the posterior probability, i.e., the relative loss bias. Posterior probability of belonging to a clean sample distribution and the posterior probability belonging to the noise sample distribution. The calculation formula is as follows: (8) (9) M-step: Based on the calculated posterior probabilities, update all parameters of the mixture distribution model, where the update formula for parameters corresponding to clean samples is as follows: (10) (11) (12) The parameters corresponding to the noise samples are updated in the same way as the above formula; through iterative execution of the E step and M step, the model parameters converge, and the optimal model parameters are output; based on the converged model parameters, the noise probability corresponding to each node and each category is calculated. This result is used as the noise quantification result, which enables accurate quantification of the noise level in the data and provides a direct quantitative basis for subsequent weight generation and sample separation.

[0040] Furthermore, in one embodiment, step S200 includes the following steps: S205: Based on the noise quantization results, calculate the noise uncertainty of the corresponding data; S206: Based on the noise uncertainty, generate a weight parameter that is negatively correlated with the noise level, wherein the higher the noise level of the data, the lower the corresponding weight parameter.

[0041] In this embodiment, the uncertainty of noise is modeled by information entropy, thereby generating noise-sensitive adaptive weight parameters to achieve differentiated management of data with different levels of noise and suppress the interference of noise on subsequent model training from the parameter level.

[0042] In step S205, based on the noise quantization results, the noise uncertainty of the k-th IoT node in the c-th category is calculated using information entropy. The calculation formula is as follows: (13) By calculating the noise uncertainty, the confidence level of the noise probability can be further characterized. The higher the uncertainty, the more difficult it is to determine the noise level of this type of data, and the higher the risk of interference to model training.

[0043] In step S206, weight parameters for the corresponding category are generated based on noise uncertainty. The calculation formula is as follows: (14) In the formula, This represents the initial parameters of the k-th IoT node belonging to the c-th category. The weight adjustment coefficient is sensitive to noise, which can adaptively adjust the weight of the corresponding category to be lower when the noise level is higher. This provides a precise control basis for weight allocation in subsequent sample separation and federated training, and preemptively suppresses the negative impact of high noise data on model training.

[0044] Furthermore, in one embodiment, step S300 includes the following steps: S301: Based on the generated weight parameters, combined with nonlinear filtering, a clean sample set is obtained from the local dataset; S302: By using set difference operations, a noisy sample set that has no intersection with the clean sample set is partitioned from the local dataset.

[0045] In this embodiment, based on noise-adaptive weight parameters, the clean samples and noisy samples are accurately separated, eliminating the misleading influence of noisy samples on model feature learning from the data source level, and providing high-purity clean samples for subsequent federated model training.

[0046] In step S301, from the original sample set A clean sample set is obtained by filtering. The calculation formula is as follows: (15) In the formula, σ( ) represents a nonlinear filtering function used to filter high-noise samples based on weight parameters, while retaining clean samples with low noise.

[0047] In step S302, based on the mutual exclusion property of the sample set ( and The noise sample set is obtained through difference operations. The calculation formula is as follows: (16) Through the above steps, a precise separation between clean and noisy samples without overlap is achieved, completely avoiding the feature learning bias caused by the participation of noisy samples in local model training, and laying a data foundation for subsequent robust model training.

[0048] Furthermore, in one embodiment, step S400 includes the following steps: S401: Construct a knowledge distillation training framework using the global model as the teacher model and the node local model as the student model; S402: Based on a clean sample set, calculate the task loss of the local model and the guided distillation loss of the global model respectively; S403: Combining the proportion of clean and noisy samples in the local dataset, the task loss and distillation loss are weighted and fused to obtain the total training loss; S404: Iteratively train the local model based on the total training loss, and calculate the parameter update amount of the local model relative to the global model after training is completed.

[0049] In this embodiment, knowledge distillation based on the teacher-student architecture utilizes the globally reliable knowledge learned by the global model to guide the training of the local model on clean samples, thereby achieving noise robustness repair of the local model. At the same time, it adapts to the distribution characteristics of local data and solves the feature learning bias problem caused by noise sample interference in the local model.

[0050] In step S402, local task loss The calculation is based on a clean sample set to ensure the local model's ability to fit the task objective. The calculation formula is as follows: (17) In the formula, Represents cross-entropy loss, This indicates the number of samples in the clean sample set.

[0051] Global model-guided distillation loss This is used to achieve knowledge transfer from the global model to the local model, and the calculation formula is as follows: (18) In the formula, This represents the Kullback-Leibler divergence function. Represents a non-linear activation function. For temperature hyperparameters, This is a global model.

[0052] In step S403, the total training loss of the local model obtained by weighted fusion is... The calculation formula is as follows: (19) In the formula, and Let be the weight hyperparameter, satisfying + =1; This is a sample quality adaptive coefficient that can adaptively adjust the weight of distillation loss based on the noise ratio of local data, thereby achieving a balance between task fitting and global knowledge transfer.

[0053] In step S404, backpropagation iterative optimization is performed based on the total training loss. After training the local model, the parameter update amount of the local model relative to the initial global model in this round is calculated. ,in These are the parameters of the local model after training. This update represents the global model parameters from the previous iteration. This update will be uploaded to the federated server for aggregated updates of the global model.

[0054] Furthermore, in one embodiment, such as Figure 2 As shown, the method also includes the following steps: S500: Performs multi-dimensional credibility assessment on the model update amount of each node, completes adaptive aggregation weight adjustment based on the assessment results, and updates the global model in combination with the adjusted weights.

[0055] In this embodiment, adaptive aggregation based on multi-dimensional credibility assessment suppresses the negative impact of low-credibility local updates on the global model, ultimately achieving robust optimization of the global model and completing the construction of a unified representation space for cross-node multimodal data, thus solving the problems of low accuracy and poor robustness of the global model in the prior art.

[0056] Furthermore, in one embodiment, S500 includes the following steps: S501: Calculate the fusion similarity, historical update consistency and node data quality index of the model update volume of each node, and weightedly fuse them to obtain the credibility score of the model update volume of each node. S502: Based on the credibility scores of all participating training nodes, the aggregate weights of each node are adaptively reduced. S503: Based on the adjusted aggregation weights, the model update amounts of each node are weighted and aggregated to update the parameters of the global model; S504: Based on the updated global model, re-execute the loss distribution statistics, noise distribution modeling, and sample purification and separation steps to complete the closed-loop update of sample partitioning and enter the next round of federated learning iteration.

[0057] In this embodiment, the reliability of local model updates is fully quantified through multi-dimensional credibility assessment, thereby achieving adaptive weight adjustment and global aggregation. This avoids low-quality, high-noise local updates from polluting the global model. At the same time, through a closed-loop iterative mechanism, the sample partitioning and model performance are continuously optimized, achieving continuous optimization of the global model and gradual unification of the multimodal data representation space.

[0058] In step S501, the evaluation indicators for each dimension are first calculated: First, a similarity metric is used to measure the difference in direction and magnitude between local model updates and global model updates. This is achieved by first calculating the cosine similarity between the local update and the global historical updates. The calculation formula is as follows: (20) In the formula, Let $\mathbf{k}$ be the model update value for the $k$-th node in the $t$-th iteration. The update amount of the global model in the previous round. Denotes the vector dot product, ∥ ∥ represents the vector norm.

[0059] Then calculate the weighted Euclidean distance. The calculation formula is as follows: (twenty one) In the formula, L represents the total number of layers in the neural network. For the first Preset weights of layers For the k-th node Layer parameter update amount, This represents the update amount of the previous round for the l-th layer of the global model.

[0060] The fusion similarity is calculated based on cosine similarity and weighted Euclidean distance. The calculation formula is as follows: (twenty two) In the formula, , To measure the weight hyperparameters, satisfying , For scaling hyperparameters.

[0061] Second, historical consistency measurement This is used to evaluate the stability of historical updates of local nodes, and the calculation formula is as follows: (twenty three) In the formula, T is the window length of the historical round. This represents the model update amount of the k-th node in the i-th iteration. The lower the historical consistency, the greater the fluctuation in local updates and the lower the reliability.

[0062] Third, node data quality indicators The class weights generated in the first stage are used to characterize the overall quality of the local training data. The calculation formula is as follows: (twenty four) In the formula, This represents the number of samples of the c-th category at the k-th node.

[0063] Finally, the three dimensions of indicators are weighted and fused to obtain the credibility score of the model update quantity of the k-th node. The calculation formula is as follows: (25) In the formula, , , Let be the weight hyperparameter, satisfying By integrating multiple dimensions, the reliability of local updates can be comprehensively and objectively evaluated.

[0064] In step S502, the aggregation weight is adaptively adjusted based on the credibility score, and the adjustment rules are as follows: (26) In the formula, This represents the average credibility score of all participating training nodes in the current round. For the corresponding standard deviation, The preset weighting factor, and , This is the initial aggregation weight for the nodes; through this rule, updates to nodes with extremely low credibility are subject to strong weight reduction to prevent them from polluting the global model.

[0065] In step S503, the global model parameters are updated based on the adjusted aggregate weights, and the calculation formula is as follows: (27) In the formula, These are the global model parameters updated in the t-th iteration.

[0066] In step S504, the updated global model is distributed to each participating node. Based on the latest global model, each node re-executes the entire process of loss distribution statistics, noise distribution modeling, and sample purification and separation, and updates the local clean sample set, noisy sample set, and class weights, forming a closed-loop iterative mechanism of "noise modeling-sample separation-model training-global update-sample optimization" to continuously optimize the model performance, gradually unify the representation space of cross-node multimodal data, and improve the accuracy and generalization ability of the global model.

[0067] Secondly, such as Figure 3 As shown in the embodiments of this application, an IoT signal repair device is also provided, comprising: a data acquisition module, which is used to acquire the local datasets of each IoT node in the federated learning system, and the local models corresponding to each node; a noise modeling and weight generation module, which is used to complete data loss distribution statistics and noise distribution modeling based on the prediction results of the local datasets and local models, and generate weight parameters that are adapted to the degree of data noise; a sample purification and separation module, which is used to perform sample purification and separation on the local dataset based on the weight parameters, to obtain a clean sample set and a noisy sample set, and output the weight parameters and sample separation results for federated model training; and a model repair training module, which is used to initialize the global model of federated learning, and based on the output weight parameters, clean sample set and noisy sample set, to complete the repair training of the local models of each node through knowledge distillation, and obtain the model update amount of each node.

[0068] The functions of each module in the above-mentioned IoT signal repair device correspond to the steps in the above-mentioned IoT signal repair method embodiment, and their functions and implementation processes will not be described in detail here.

[0069] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0070] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0071] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0072] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0073] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0074] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0075] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for repairing Internet of Things (IoT) signals, characterized in that, The IoT signal repair method includes: Obtain the local datasets of each IoT node in the federated learning system, as well as the local models corresponding to each node; Based on the prediction results of the local dataset and the local model, the data loss distribution statistics and noise distribution modeling are completed, and weight parameters that are adapted to the degree of data noise are generated. Based on the weight parameters, the local dataset is purified and separated to obtain a clean sample set and a noisy sample set, and the weight parameters and sample separation results are output for training the federated model. The global model of federated learning is initialized. Based on the output weight parameters, clean sample set and noisy sample set, the local model of each node is repaired and trained through knowledge distillation to obtain the model update amount of each node.

2. The IoT signal repair method as described in claim 1, characterized in that, The step of performing data loss distribution statistics based on the prediction results of the local dataset and the local model includes: For each IoT node, the prediction loss of a single sample in the local dataset is calculated based on the prediction results of the local model and the sample labels. Using data category as the statistical dimension and combining it with the overall loss distribution benchmark within the node, the relative loss deviation corresponding to each category of data is calculated and used as the statistical result of loss distribution.

3. The IoT signal repair method as described in claim 2, characterized in that, The method of using data category as the statistical dimension and combining it with the overall loss distribution benchmark within the node to calculate the relative loss deviation corresponding to each category of data includes: Calculate the mean and variance of class loss for each category of samples within each node, based on the data category dimension. Based on the loss statistics of all categories within each node, the overall loss distribution benchmark corresponding to that node is calculated. The overall loss distribution benchmark includes the node's average category loss and the standard deviation of the node's category loss. By combining the category loss mean of each category with the overall loss distribution benchmark of the corresponding node, the relative loss deviation of each category is calculated.

4. The IoT signal repair method as described in claim 2, characterized in that, The completion of data loss distribution statistics and noise distribution modeling includes: Based on the statistical results of the loss distribution, the data loss is modeled as a hybrid distribution model consisting of the superposition of the clean sample distribution and the noisy sample distribution; The model parameters of the mixed distribution model are solved by iterative optimization. Based on the solved model parameters, the noise probability corresponding to each data point is calculated, and the noise probability is determined as the noise quantization result.

5. The IoT signal repair method as described in claim 4, characterized in that, The weight parameters that are adapted to the level of data noise include: Based on the noise quantization results, the noise uncertainty of the corresponding data is calculated; Based on the noise uncertainty, a weight parameter negatively correlated with the noise level is generated, wherein the higher the noise level of the data, the lower the corresponding weight parameter.

6. The IoT signal repair method as described in claim 1, characterized in that, The step of purifying and separating the local dataset based on the weight parameters to obtain a clean sample set and a noisy sample set includes: Based on the generated weight parameters, combined with nonlinear filtering, a clean sample set is obtained from the local dataset. By using set difference operations, a noisy sample set that has no intersection with the clean sample set is partitioned from the local dataset.

7. The IoT signal repair method as described in claim 1, characterized in that, The process of repairing and training the local model at each node through knowledge distillation to obtain the model update amount for each node includes: A knowledge distillation training framework is constructed using the global model as the teacher model and the node local model as the student model. Based on a clean sample set, the task loss of the local model and the guided distillation loss of the global model are calculated separately. By combining the proportion of clean and noisy samples in the local dataset, the task loss and distillation loss are weighted and fused to obtain the total training loss; The local model is iteratively trained based on the total training loss, and the parameter update amount of the local model relative to the global model is calculated after training is completed.

8. The IoT signal repair method as described in claim 1, characterized in that, The method further includes: The model update volume of each node is evaluated for reliability in multiple dimensions. Based on the evaluation results, adaptive aggregation weight adjustment is completed, and the global model is updated in combination with the adjusted weights.

9. The IoT signal repair method as described in claim 8, characterized in that, The process of performing a multi-dimensional credibility assessment of the model update amount for each node, adjusting the adaptive aggregation weights based on the assessment results, and updating the global model using the adjusted weights includes: Calculate the fusion similarity, historical update consistency and node data quality index of the model update volume of each node, and then weightedly fuse them to obtain the credibility score of the model update volume of each node. Based on the credibility scores of all participating training nodes, the aggregate weights of each node are adaptively reduced. Based on the adjusted aggregation weights, the model update amounts of each node are weighted and aggregated to update the parameters of the global model; Based on the updated global model, the steps of loss distribution statistics, noise distribution modeling, and sample purification and separation are re-executed to complete the closed-loop update of sample partitioning and enter the next round of federated learning iteration.

10. An Internet of Things (IoT) signal repair device, characterized in that, include: The data acquisition module is used to acquire the local datasets of each IoT node in the federated learning system, as well as the local models corresponding to each node. The noise modeling and weight generation module is used to complete the data loss distribution statistics and noise distribution modeling based on the prediction results of the local dataset and the local model, and generate weight parameters that are adapted to the degree of data noise. The sample purification and separation module is used to purify and separate the local dataset based on the weight parameters to obtain a clean sample set and a noisy sample set, and output the weight parameters and sample separation results for training the federated model. The model repair and training module is used to initialize the global model of federated learning. Based on the output weight parameters, clean sample set and noisy sample set, it completes the repair training of the local model of each node through knowledge distillation, and obtains the model update amount of each node.

Images