An edge-cloud data collaborative processing method based on a dynamic sliding window and federated learning

Abstract

The application discloses a kind of based on dynamic sliding window and federated learning's edge cloud data collaborative processing method, specifically related to edge computing and data collaborative processing technical field;By constructing time stamp sliding window, based on local mutation intensity self-adaptive adjustment window length, extract key state fragment, generate local feature abstract stream;Characteristic abstract is input into local federal submodel, and the response value of trusted feature is calculated, and collaborative sparse atlas is constructed to identify abnormal nodes, and trigger incremental retraining;Under bandwidth constraint, upload lightweight model update and feature abstract, cloud fusion constructs global model and issues;According to node mutation rate and network state, adjust the next round sliding window parameter;The application can effectively improve model training precision and edge cloud collaborative efficiency, with low bandwidth, high dynamic and self-adaptive advantage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of edge computing and data collaborative processing, and particularly relates to an edge-cloud data collaborative processing method based on dynamic sliding window and federated learning. BACKGROUND

[0002] Currently, in application scenarios such as industrial internet, smart manufacturing and remote monitoring, a large number of edge nodes are usually deployed to collect and pre-process production data in real time. Since these edge nodes are often in a complex environment with limited network bandwidth and unstable communication delay, traditional centralized data processing schemes cannot balance communication overhead and processing timeliness.

[0003] On the one hand, directly uploading all the original data of the edge nodes to the cloud not only causes serious bandwidth congestion, but also has the risk of data security and privacy leakage; on the other hand, the data synchronization mechanism based on static strategy cannot adapt to the sudden state of the industrial site (such as device abnormality or mutation event), and it is difficult to realize efficient, dynamic and differentiated data collaboration.

[0004] In addition, in the process of multi-edge node collaborative learning, due to different sampling frequencies and data time sequence misalignment between nodes, the global model training error is often accumulated, which seriously affects the federated learning effect. SUMMARY

[0005] The purpose of the present application is to provide an edge-cloud data collaborative processing method based on dynamic sliding window and federated learning to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: an edge-cloud data collaborative processing method based on dynamic sliding window and federated learning, comprising:

[0007] For the time series data collected by each edge node, a sliding data window with a time stamp is constructed, and a dynamic sliding window adjustment method is used to adjust the window length;

[0008] Based on the data mutation characteristics in the sliding window, the key state segments in each time window are extracted to generate a local feature abstract stream F_i;

[0009] The local federated sub-model is input with the local feature abstract stream F_i, and the output result of the local federated sub-model is weighted confidence evaluated to obtain a local trusted feature response value R_i;

[0010] The R_i of all edge nodes is summarized to construct a collaborative sparse graph, and a local data enhancement and retraining request is triggered according to the abnormal mutation nodes in the graph structure;

[0011] The selected local federated sub-model update results and the local feature abstract stream F_i are uploaded to the cloud server to construct a global model M_g by cloud aggregation under the condition of meeting the bandwidth constraint;

[0012] The updated M_g is issued to the edge node, and the initial parameters of the next round of sliding window are adjusted according to the mutation rate and network state of the node itself.

[0013] Preferably, the dynamic sliding window adjustment method adjusts the window length, including:

[0014] The original time series data collected by the edge node is time-stamped and preliminarily segmented at a fixed time granularity;

[0015] Based on the statistical volatility indicators in each data segment, including the mean change rate, standard deviation increase and mutation frequency, a local mutation intensity vector is constructed;

[0016] According to the local mutation intensity vector and the current network bandwidth state of the edge node, a bandwidth-sensitive window adjustment algorithm is used to dynamically update the sliding window length.

[0017] Preferably, the extraction of the key state segment in each time window generates a local feature abstract stream F_i, including:

[0018] In the data segment covered by the current sliding window, it is judged whether each data point meets the mutation condition based on a preset threshold, and the mutation condition includes a mutation amplitude threshold that the data increase exceeds three times the standard deviation of the historical mean;

[0019] The continuous data points meeting the mutation condition are divided into state segments;

[0020] Each state segment is processed for feature compression, including calculating the average change rate, maximum fluctuation range and segment duration to form a feature vector;

[0021] All feature vectors in the current time window are spliced in time sequence to generate a local feature abstract stream F_i corresponding to the window.

[0022] Preferably, the local feature abstract stream F_i is input into the local federated sub-model, and the local federated sub-model output result is weighted and confidence evaluated to obtain a local trusted feature response value R_i, including:

[0023] The local feature abstract stream F_i generated in the current sliding window is input into the pre-initialized local federated sub-model, and the federated sub-model is a lightweight deep neural network structure, which uses a one-dimensional convolutional neural network to extract time series features;

[0024] The prediction output of the model for F_i is obtained, and a corresponding confidence score is calculated, which is obtained by a Softmax activation function of the output layer, indicating the credibility of the current prediction result;

[0025] The confidence score is weighted and adjusted, and the weight factor is calculated according to the fluctuation intensity and duration of the state segment in F_i;

[0026] The weighted confidence score is taken as the local trusted feature response value R_i.

[0027] Preferably, the R_i of all edge nodes is aggregated to construct a collaborative sparse graph, and a local data enhancement and retraining request is triggered according to the abnormal mutation nodes in the graph structure, including:

[0028] The trusted feature response values R_i uploaded by multiple edge nodes are collected, and the edge weight values between nodes are calculated according to the physical distance, functional similarity and feature response correlation between nodes;

[0029] A non-directional graph structure between edge nodes is constructed, the edge weight value is taken as the weight of the connecting edge, and sparse processing is performed on the graph structure, only the edges with edge weight values higher than a set threshold are reserved;

[0030] Abnormal mutation regions of nodes in the sparse graph structure are identified, specifically, nodes with significant response value difference and center degree drop from adjacent nodes are detected as potential abnormal points;

[0031] For the potential abnormal points, the re-extraction and expansion of the local original data are triggered, and the incremental retraining of the local sub-model is performed based on the enhanced data.

[0032] Preferably, the selected local federal sub-model update result and local feature summary stream F_i are uploaded to the cloud server, including:

[0033] The current local federal sub-model parameter update result is processed by sparse coding, and the part with gradient change amplitude exceeding a set change threshold is reserved to form a lightweight model update vector;

[0034] According to the current available bandwidth of the edge node, it is judged whether the minimum uploading condition is met, if the available bandwidth is higher than the uploading threshold value, the lightweight model update vector and the corresponding local feature summary stream F_i are bound and packaged;

[0035] The packaged result is uploaded to the cloud server, and the cloud server uses an aggregation algorithm to weight and fuse the model update results uploaded by multiple edge nodes to generate a global model M_g.

[0036] Preferably, the updated M_g is issued to the edge node, and the initial parameters of the next round of sliding window are adjusted according to the mutation rate of the node itself and the network state, including:

[0037] The cloud-end aggregated global model M_g is distributed to each participating edge node;

[0038] After the edge node receives M_g, the frequency of mutation events in the last round of sliding window is counted, and the mutation rate index V_i is calculated, which is the ratio of the number of data points meeting the mutation judgment condition to the total number of data points;

[0039] The network state information of the current edge node is collected, including bandwidth B_i and average delay D_i, and a joint regulation function of mutation rate and network state is established;

[0040] According to the output result of the joint regulation function, the initial length and moving step of the next round of sliding window are adjusted.

[0041] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0042] 1. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning provided by the present application can realize data hierarchical screening, model precision guarantee and communication efficiency balance in a complex environment with extremely small network bandwidth and frequent state fluctuations of edge nodes. By constructing a mutation-aware dynamic sliding window mechanism, combining statistical volatility analysis and adaptive window adjustment strategy, the response ability of edge nodes to local sudden state is effectively improved, and the cumulative risk of invalid data redundancy and model training error is reduced.

[0043] 2. The present application realizes the state cooperative perception and abnormality identification between edge nodes by introducing a trusted feature response evaluation and collaborative sparse graph construction mechanism, so as to dynamically trigger local data enhancement and incremental retraining operation without increasing the communication load. The cloud uses bandwidth-aware compression upload strategy and differential weighted aggregation algorithm to further enhance the stability and generalization ability of the global model in the multi-source heterogeneous data environment, which is suitable for various edge cloud collaborative computing scenes such as industrial manufacturing, intelligent transportation and remote monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments or prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0045] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0046] The technical solutions and advantages of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0047] Embodiments, please refer to Figure 1 As shown in the drawings, the edge cloud data collaborative processing method based on a dynamic sliding window and federated learning comprises the following steps.

[0048] For the time series data collected by each edge node, a sliding data window with a timestamp is constructed, and a dynamic sliding window adjustment method is used to adjust the window length.

[0049] First, for the original time series data collected by the edge node, time stamp marking processing is performed. Specifically, for each piece of original data received within the sampling period of the edge node, a current local timestamp is attached according to the data collection time. The timestamp accuracy can be set to millisecond level to ensure the accurate controllability of the time sequence order of the data in the sliding window.

[0050] After completing the time stamp marking, the continuous data stream is preliminarily segmented according to a fixed time granularity. The time granularity is a preset parameter, preferably 5 seconds or 10 seconds, indicating that every 5 seconds or 10 seconds is an initial data segment. The purpose of this step is to provide a basic data structure for subsequent mutation analysis and dynamic window adjustment. Each data segment contains a number of data points with timestamps, forming an intra-segment data set.

[0051] Next, for the data in each data segment described above, statistical volatility indicators are extracted for subsequent mutation intensity modeling. Specifically, the following three statistical indicators are included:

[0052] Mean change rate: the change ratio between the average value of each data dimension in the current data segment and the average value of the previous data segment, reflecting the long-term trend change;

[0053] Standard deviation increase: the ratio of the standard deviation in the current data segment to the historical average standard deviation, used to evaluate the abnormality of the fluctuation amplitude;

[0054] Mutation frequency: the proportion of data points that meet a certain mutation determination condition (such as exceeding a set threshold ΔT) in the current data segment. The mutation determination threshold ΔT is a value set according to domain experience, such as temperature change exceeding 5 degrees Celsius, voltage fluctuation exceeding 2 volts, etc.

[0055] The three indicators are combined to form a three-dimensional numerical vector, denoted as the local mutation intensity vector. This vector is used to represent the comprehensive dynamic characteristics of the current data segment in terms of stability, continuity, and trend of change.

[0056] After obtaining the local mutation intensity vector, the length of the sliding data window is adjusted according to the current network bandwidth status of the edge node. The bandwidth status of the edge node is determined by continuously monitoring the size of the successfully uploaded data packets per second and the transmission delay of the node, denoted as the current bandwidth utilization B, with the unit of bits per second.

[0057] According to the mutation intensity vector and the bandwidth utilization B, a bandwidth-sensitive window adjustment algorithm is executed. This algorithm is implemented through the following logic:

[0058] When any dimension of the mutation intensity vector exceeds its set change threshold (for example, the mean change rate exceeds 0.3, the standard deviation increase exceeds 0.5, and the mutation frequency exceeds 20%), and the bandwidth utilization B is lower than the set bandwidth threshold B_min (for example, 1000 bits per second), the sliding window length is shortened to the minimum allowed value T_min (for example, 5 seconds);

[0059] When the mutation intensity is low, and the bandwidth utilization B is higher than the set upper limit B_max (for example, 8000 bits per second), the window length is expanded to the maximum allowed value T_max (for example, 60 seconds);

[0060] In other cases, the window length remains unchanged, or it is adjusted linearly between T_min and T_max.

[0061] Based on the data mutation characteristics within the sliding window, key state segments are extracted within each time window to generate a local feature summary stream F_i.

[0062] After the dynamic adjustment of the sliding window is completed, key state segments are extracted for the data segment covered by the current sliding window to generate a local feature summary stream for subsequent model input. First, all data points in the window are subjected to mutation condition judgment. The mutation condition is set as follows: for any data dimension, if the increase of the value at a certain time point compared to the historical mean of that dimension is greater than 3 times the historical standard deviation, it is determined that the data point has mutated.

[0063] Specifically, let the historical mean of a certain dimension in the current sliding window be μ, the standard deviation be σ, and the observed value of the current data point be x_i. If the absolute value of x_i minus μ is greater than 3 times σ, i.e., it satisfies the point is marked as a mutation point. This method can effectively distinguish between normal fluctuations and abnormal mutations, improving the accuracy of state recognition.

[0064] The continuous data points marked as mutation points are divided into several state segments. Continuity refers to the case where multiple data points satisfying the mutation condition are adjacent to each other without interruption in the time stamp. In order to avoid interference caused by short-term abnormalities or signal noise, only when the continuous mutation duration is greater than the set minimum segment duration threshold (for example, 1 second), it is determined as a valid state segment.

[0065] For each valid state segment, a feature compression process is performed to condense the original multiple time points into a feature vector. Feature compression includes the following three index calculations:

[0066] Average change rate: represents the average speed of data change in the state segment, calculated as the sum of the differences between adjacent data points in the segment divided by the duration;

[0067] Maximum fluctuation range: represents the difference between the maximum and minimum values in the segment, used to measure the mutation amplitude;

[0068] Segment duration: the time difference between the start time and end time of the state segment, in seconds.

[0069] The feature compression result corresponding to each state segment forms a set of three-dimensional feature vectors, which are arranged in order according to their time stamps. Finally, the feature vectors of all state segments in the current sliding window are spliced to form the local feature summary stream corresponding to the window, denoted as F_i.

[0070] The local feature summary stream F_i is input into the local federated sub-model, and the output result of the local federated sub-model is weighted and confidence evaluated to obtain the local trusted feature response value R_i.

[0071] After the local feature summary stream F_i is constructed, it is passed as input to the federated sub-model deployed locally on the edge node. The federated sub-model is a lightweight deep neural network structure. To adapt to the calculation resource limitation in the edge computing environment, a one-dimensional convolutional neural network structure is preferably used. The network includes an input layer, two one-dimensional convolutional layers, a fully connected layer, and an output layer, with a convolution kernel size of 3 and a step size of 1, using a ReLU activation function to extract the local patterns and trends of the time series features in F_i.

[0072] After the model receives the input F_i, it performs forward propagation and outputs a corresponding prediction result vector, representing the model's judgment of the state in the current time window. To measure the credibility of the prediction result, a Softmax activation function is applied to the output result, normalizing each dimension in the prediction vector to a probability value between 0 and 1. The greater the probability, the more confident the model is in its classification. This confidence score is denoted as C_i, representing the basic trusted response of the model to the current state classification result.

[0073] To further enhance the adaptability of the model output to the mutation state, the confidence score C_i needs to be weighted and adjusted. The weighting factor W_i is generated according to the statistical characteristics of each state segment in the local feature summary stream F_i, specifically: calculate the weighted product of the fluctuation intensity (i.e. the maximum fluctuation range) and the duration of each state segment, and take the average of all segments to obtain the overall fluctuation index S_i, and then construct the weighting function where α is an empirical adjustment parameter, preferably a floating-point value between 0.5 and 1.

[0074] Multiply the confidence score C_i by the weighting factor W_i to obtain the final weighted confidence response value R_i, which represents the reliable output of the edge node in the current time window for the feature state judgment.

[0075] All edge node R_i is summarized to construct a collaborative sparse graph, and according to the abnormal mutation node in the graph structure, the local data enhancement and retraining request is triggered.

[0076] After all edge nodes complete the calculation of the reliable feature response value R_i in the current round, upload it to the central aggregation processing end. First, all R_i values collected are summarized and processed, and based on this, the connection weight between edge nodes is calculated as the edge weight value for constructing the graph structure. In the specific calculation process, the following three factors are considered:

[0077] Physical distance similarity: the closer the geographical position between edge nodes, the higher the relevance; a normalized Euclidean distance inverse function is used for modeling;

[0078] Functional similarity: the consistency of the type of equipment monitored or the task executed by the edge node, using classification matching or functional label similarity score;

[0079] Feature response correlation: the Pearson correlation coefficient between the reliable feature response values R_i and R_j of different nodes is calculated to reflect the degree of synchronization of state changes.

[0080] According to the above three indicators, an edge weight matrix is constructed to generate an undirected weighted graph structure, where each node in the graph represents an edge node, and each edge in the graph represents the state association between two nodes, and the edge weight value is used to describe the connection strength. To compress the size of the graph structure and highlight the significant association relationship, sparse processing is performed on the graph structure. The specific way is to set an edge weight threshold T_e, for example T_e=0.6, only the edges with edge weight value higher than T_e are retained, and the remaining edge connection relationships are removed, forming a sparse collaborative graph.

[0081] On the basis of the above sparse graph structure, an abnormal node identification operation is performed. The identification process includes the following two judgment criteria:

[0082] Response value difference: the difference between R_i of a certain node and the average R_j value of all its adjacent nodes is greater than the set abnormal judgment threshold ΔR, for example, ΔR > 0.4;

[0083] Graph structure centrality decline: the degree centrality, eigenvector centrality or betweenness centrality of the node in the current graph structure is significantly lower than the historical average level, indicating that its connectivity or transmission ability in the local subgraph is weakened.

[0084] The node meeting any of the above conditions is marked as a potential abnormal mutation node.

[0085] For the nodes judged as abnormal mutations, the local raw data re-extraction and sample expansion operations are triggered immediately. The data expansion includes extending the sliding window time range, increasing the mutation point density sampling frequency, and introducing historical pair photos to construct an enhanced sample set. After completing the data expansion, the local feature summary stream is re-input into the local federal sub-model, and an incremental retraining operation is performed based on the original model parameters.

[0086] Under the condition of meeting the bandwidth constraint, the selected local federal sub-model update result and the local feature summary stream F_i are uploaded to the cloud server, and the global model M_g is constructed by cloud aggregation.

[0087] To reduce communication cost and adapt to the bandwidth limitation of edge nodes, the model parameter update result generated after each round of local training is processed by sparse coding. Specifically, let the model parameter change vector obtained by the edge node after local training be ΔP_i, and the absolute value of the gradient change amplitude of each parameter dimension is calculated, and a change amplitude threshold T_g is set, which is preferably set to ten percent of the maximum value of the original gradient vector. Only the parameters whose change amplitudes are greater than T_g are retained, and the parameters with weak changes are removed to form a sparse vector ΔP_i', which is the lightweight model update vector.

[0088] After obtaining the lightweight model update vector ΔP_i', the edge node monitors the real-time network bandwidth state. The bandwidth state can be obtained by the ratio of the total number of successfully transmitted data packets per unit time to the transmission delay, and the result is recorded as the available bandwidth B_i, which is measured in bits per second. Set a minimum upload threshold B_min, which is preferably set to 2000 bits per second. When B_i ≥ B_min, it is judged that the node meets the upload condition.

[0089] If the upload condition is met, the lightweight model update vector ΔP_i' and the local feature summary stream F_i in the current sliding window are packaged together as an upload data packet U_i. The packaging format uses a structured binary format, including a model parameter part and a feature summary part, to ensure data integrity and parsing efficiency. U_i is uploaded to the cloud aggregation server through an encrypted transmission channel.

[0090] After receiving the upload data packets U_i from multiple edge nodes, the cloud extracts ΔP_i' from them and performs model aggregation operations. The aggregation process adopts a weighted average algorithm, and the aggregation weight of each edge node i is W_i. The calculation of W_i is based on the historical upload frequency of the node, the model update sparsity, and the stability index of the feature summary stream F_i. Finally, by calculating the weighted sum of all ΔP_i' and corresponding W_i, the new global model parameter P_g is generated as the output of the global model M_g of this round.

[0091] The updated M_g is distributed to the edge nodes, and the initial parameters of the next round sliding window are adjusted according to the mutation rate and network state of the node itself.

[0092] After completing the edge collaborative training of this round and forming the global model parameter M_g by the cloud server, the distribution process of the global model is started. Specifically, the global model M_g generated by the cloud is distributed to all edge nodes participating in this round of training through a secure transmission channel. Each node receives M_g and immediately replaces its local model parameters to synchronize into the next round of training cycle.

[0093] After updating the model parameters, the edge node counts the number of mutation events that occurred during the running of the last sliding window, which is used to measure the degree of change in the current node state. The mutation event determination criteria are consistent with the previously set mutation amplitude threshold. If the change amplitude of a data point compared to the historical mean of the sliding window is greater than three times the standard deviation, it is considered a mutation point. Let the total number of data points be N and the total number of mutation points be N_v, then the mutation rate index V_i is defined as N_v divided by N, i.e. .

[0094] Synchronously, the edge node performs real-time perception of the current communication network state, collecting the current available bandwidth B_i (unit: bits per second) and the average communication delay D_i (unit: milliseconds). The available bandwidth B_i can be calculated by the total amount of successfully transmitted data per unit time, and the average delay D_i can be obtained by measuring the upload feedback time multiple times and taking the arithmetic mean.

[0095] To realize the adaptive adjustment of the sliding window parameters, a joint control function U_i is constructed based on the above mutation rate V_i and network state indicators B_i and D_i. The form of the joint control function is as follows: where α, β, γ are positive weight factors set empirically to balance the influence of mutation intensity and network load factors on window adjustment.

[0096] According to the value range of the joint regulation function U_i, the initial length L_i of the sliding window and the moving step S_i are dynamically adjusted. If U_i exceeds the upper threshold T_high (for example, set to 1.0), it indicates that the current node state changes dramatically and the network is unstable, then L_i is reduced (for example, to 10 seconds) and S_i is reduced (for example, to 2 seconds); if U_i is lower than the lower threshold T_low (for example, set to 0.3), then the window length (for example, increased to 60 seconds) and the step (for example, increased to 15 seconds) are appropriately enlarged to improve the processing efficiency and reduce the resource occupation.

[0097] The adjusted sliding window parameters will be used in the next round of edge node data collection and key state segment extraction process, ensuring that the model training process continuously adapts to changes in node state and network conditions, and improving the dynamic response capability of edge-cloud collaboration.

[0098] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An edge cloud data collaborative processing method based on a dynamic sliding window and federated learning, characterized in that: The method comprises the following steps: For each edge node, the time series data collected is used to construct a sliding data window with a timestamp, and a dynamic sliding window adjustment method is used to adjust the window length; Based on the data mutation characteristics in the sliding window, the key state segments in each time window are extracted to generate a local feature abstract stream F_i; The local feature abstract stream F_i is input into the local federated sub-model, and the output results of the local federated sub-model are subjected to weighted confidence evaluation to obtain a local trusted feature response value R_i; All R_i of the edge nodes are collected to construct a collaborative sparse graph, and the abnormal mutation nodes in the graph structure trigger local data enhancement and retraining requests; Under the condition of satisfying the bandwidth constraint, the selected local federated sub-model update results and the local feature abstract stream F_i are uploaded to the cloud server, and the cloud server aggregates to construct a global model M_g; The updated M_g is distributed to the edge nodes, and the initial parameters of the next round of sliding window are adjusted according to the mutation rate and network state of the node itself.

2. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning according to claim 1, characterized in that: The method for adjusting the window length by using the dynamic sliding window adjustment method comprises the following steps: The original time series data collected by the edge node is marked with a timestamp, and the data stream is preliminarily segmented at a fixed time granularity; Based on the statistical volatility indicators in each data segment, including the mean change rate, the standard deviation increase and the mutation frequency, a local mutation intensity vector is constructed; According to the local mutation intensity vector and the current network bandwidth state of the edge node, a bandwidth-sensitive window adjustment algorithm is used to dynamically update the sliding window length.

3. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning according to claim 1, characterized in that: The method for extracting key state segments in each time window to generate a local feature abstract stream F_i comprises the following steps: In the data segment covered by the current sliding window, it is judged whether each data point meets the mutation condition based on a preset threshold, and the mutation condition includes a mutation amplitude threshold that the data increase exceeds three times the standard deviation of the historical mean; The continuous data points meeting the mutation condition are divided into state segments; Each state segment is subjected to feature compression processing, including calculating the average change rate, the maximum fluctuation range and the segment duration to form a feature vector; All feature vectors in the current time window are spliced in time sequence to generate a local feature abstract stream F_i corresponding to the window.

4. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning according to claim 3, characterized in that: The method for inputting the local feature abstract stream F_i into the local federated sub-model and performing weighted confidence evaluation on the output results of the local federated sub-model to obtain a local trusted feature response value R_i comprises the following steps: The local feature abstract stream F_i generated in the current sliding window is input into a pre-initialized local federated sub-model, and the federated sub-model is a lightweight deep neural network structure using a one-dimensional convolutional neural network to extract time series features; The prediction output of the model to F_i is obtained, and the corresponding confidence score is calculated, which is obtained by the Softmax activation function of the output layer and represents the credibility of the current prediction result; The confidence score is subjected to weighted adjustment, and the weight factor is calculated according to the fluctuation intensity and duration of the state segment in F_i; The weighted confidence score is taken as the local trusted feature response value R_i.

5. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning according to claim 4, characterized in that: The R_i of all edge nodes is aggregated to construct a collaborative sparse graph, and a local data enhancement and retraining request is triggered according to the abnormal mutation nodes in the graph structure, including: Collecting the trusted feature response values R_i uploaded by multiple edge nodes, and calculating the edge weight between nodes according to the physical distance, functional similarity and feature response correlation between nodes; Constructing an undirected graph structure between edge nodes, taking the edge weight as the weight of the connecting edge, and performing sparse processing on the graph structure, only retaining the edges with edge weight higher than the set threshold; Identify the node abnormal mutation area in the sparse graph structure, specifically detect the nodes with significant response value difference and center degree decline with adjacent nodes as potential abnormal points; For potential abnormal points, trigger the re-extraction and expansion of local raw data, and perform incremental retraining of the local sub-model based on the enhanced data.

6. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning according to claim 5, characterized in that: The selected local federated sub-model update result and local feature summary stream F_i are uploaded to the cloud server, including: Sparse coding processing is performed on the current local federated sub-model parameter update result, and the part with gradient change amplitude exceeding the set change threshold is retained to form a lightweight model update vector; According to the current available bandwidth of the edge node, it is judged whether the minimum upload condition is met, if the available bandwidth is higher than the upload threshold, the lightweight model update vector and the corresponding local feature summary stream F_i are bound and packaged; The packaged result is uploaded to the cloud server, and the cloud server uses an aggregation algorithm to weight and fuse the model update results uploaded by multiple edge nodes to generate a global model M_g.

7. The edge cloud data collaborative processing method based on dynamic sliding window and federated learning according to claim 1, characterized in that: The updated M_g is distributed to the edge nodes, and the initial parameters of the next round of sliding window are adjusted according to the mutation rate and network state of the node itself, including: The global model M_g aggregated by the cloud server is distributed to each participating edge node; After receiving M_g, the edge node calculates the mutation rate V_i by counting the frequency of mutation events in the last round of sliding window, the mutation rate is the ratio of the number of data points meeting the mutation determination condition to the total number of data points; Collect the network state information of the current edge node, including bandwidth B_i and average delay D_i, and establish a joint regulation function of mutation rate and network state; According to the output result of the joint regulation function, adjust the initial length and moving step of the next round of sliding window.

Images