An edge-computing-based internet of things data collaborative processing method

Abstract

The application discloses an Internet of Things data collaborative processing method based on edge computing and relates to the technical field of edge intelligent processing of the Internet of Things, and comprises the following steps: an edge node receives Internet of Things data, extracts local data features, node state information and data state fingerprints; adjacent edge nodes exchange data state fingerprints and node state summaries, generate repeated candidate data labels and abnormal state summaries; an improved ModernTCN edge state evaluation model is used to output repeated data scores, collaborative request scores and task undertaking scores; a main processing node is determined according to the scores and reasoning is performed; a target collaborative node is selected by using a task constraint type UCB node selection strategy; collaborative later-stage model calculation is completed by feature slicing, and a final collaborative processing result is generated by fusion; and relevant models and rules are updated according to collaborative execution feedback. The application can reduce repeated processing and communication overhead, improve edge node collaborative reasoning efficiency and processing stability.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) data processing technology, and in particular to an IoT data collaborative processing method based on edge computing. Background Technology

[0002] With the increasing demand for Industrial Internet of Things (IIoT), networked equipment in industrial parks, and intelligent on-site operation and maintenance, edge computing has been widely used to perform data processing, task scheduling, and collaborative inference at the device side or on-site network side. Public patent CN120263650A proposes a collaborative allocation method for edge computing resources for IoT, generating a collaborative allocation scheme for edge computing resources through IoT terminal association graphs and historical computing resource demand data; public patent CN112799823A proposes an online assignment and scheduling method for edge computing tasks, selecting a target edge server from several edge servers using a multi-armed slot machine model; and public patent CN116306943A proposes a multi-task local collaborative inference method for AIoT, achieving local collaborative inference through model slicing and intermediate result transmission. These technologies demonstrate that edge resource scheduling, online node selection, and model-sliced ​​collaborative inference all have publicly available technological foundations.

[0003] However, existing technologies typically focus on balanced allocation of computing resources or single task offloading, lacking pre-identification of duplicate data collection and inference from adjacent edge nodes. Node selection often relies on historical revenue or load status, failing to uniformly model duplicate data states, abnormal state summaries, and task handling capabilities. In collaborative inference, some solutions still primarily rely on client-server collaboration, making it difficult to adapt to low-signaling collaborative processing between edge nodes. In actual business applications, when multiple edge gateways cover the same area, issues frequently arise such as the same device state being processed repeatedly by multiple nodes, abnormal summaries being suppressed by duplicate data, and unstable collaborative node selection after sudden node overload. This leads to wasted edge computing power, increased response latency, and increased risk of exposing on-site data.

[0004] Therefore, how to provide an IoT data collaborative processing method based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an IoT data collaborative processing method based on edge computing. This invention fully utilizes edge node collaboration, an improved ModernTCN edge state evaluation model, a task-constrained UCB node selection strategy, and model splitting inference technology. It describes in detail the processing flow of IoT data deduplication, anomaly verification, collaborative node selection, and collaborative feedback update, and has the advantages of less redundant computation, low communication overhead, high collaborative inference efficiency, and good edge processing stability.

[0006] An IoT data collaborative processing method based on edge computing according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Edge nodes receive IoT data, generate local standard data units according to time slices, and extract local data features, node status information, and data status fingerprints;

[0008] Step 2: Adjacent edge nodes exchange data state fingerprints and node state summaries to generate duplicate candidate data markers and abnormal state summaries;

[0009] Step 3: Input the local data features, node status information, duplicate candidate data labels and abnormal status summaries into the improved ModernTCN edge state evaluation model. After temporal convolution, duplicate judgment channel and node status fusion layer, the duplicate data score, collaborative request score and task acceptance score are output.

[0010] Step 4: Determine the main processing node and perform inference based on the duplicate candidate data marking, duplicate data score and task acceptance score. The remaining nodes enter the auxiliary retention state. When the abnormal state summary meets the abnormal retention rules, a review is triggered.

[0011] Step 5: When the collaboration request score meets the collaboration request conditions, a set of candidate nodes is formed based on the collaboration task information, task acceptance score and historical collaboration records. The target collaboration node is selected using the task-constrained UCB node selection strategy.

[0012] Step 6: Generate collaborative features based on model segmentation points, obtain feature slices through perturbation and slicing, and send them to the target collaborative node to perform collaborative post-processing model calculations. The task initiating node then fuses intermediate results to generate the final collaborative processing result.

[0013] Step 7: Generate collaborative execution feedback based on the final collaborative processing results, execution status, processing latency, and communication overhead, and update the improved ModernTCN edge state assessment model, main processing node determination rules, exception retention rules, and task-constrained UCB node selection strategy.

[0014] Optionally, step one specifically includes:

[0015] According to the IoT data access configuration, the edge node extracts the device identifier, measurement point type, collection time, data value and coverage area identifier from the IoT data, performs timestamp correction on the collection time, and classifies the data with the same device identifier and the same measurement point type into the corresponding time slice according to the preset time slice length to form a local standard data unit.

[0016] Data quality is marked for local standard data units. Data for which no data value is received is marked as missing quality, data for which the value exceeds the allowable range is marked as out of bounds quality, data for which the difference between adjacent time slices reaches the mutation threshold is marked as mutation quality, data for which no device message is received within the acquisition period is marked as communication abnormality quality, and the rest of the data is marked as normal quality.

[0017] Data change trends are generated based on comparisons of data values ​​within continuous time slices, fluctuation levels are generated based on the dispersion of data values ​​within continuous time slices, short-term mutation states are generated based on mutation quality markers, and these are combined to form local data features.

[0018] The node load status is generated based on the processing load, the task queuing status is generated based on the number of tasks to be processed, and the communication availability status is generated based on the connection status of adjacent edge nodes. These are combined to form node status information.

[0019] The fingerprint input record is generated by combining local data features, measurement point type, time slice identifier, and coverage area identifier according to the preset field order. Locality-sensitive hashing is then performed on the fingerprint input record to obtain the data status fingerprint.

[0020] Optionally, step two specifically includes:

[0021] Adjacent edge nodes send data status fingerprints and node status summaries according to a preset exchange cycle. The node status summary is formed by a combination of node load status, task queuing status, and communication availability status.

[0022] Time slice identification and measurement point type matching are performed on the data state fingerprints of this node and its adjacent edge nodes, and the matching data state fingerprints are used as the fingerprints to be compared.

[0023] Calculate the Hamming distance between the fingerprint to be compared at this node and the fingerprint to be compared at the adjacent edge node, and use the Hamming distance as the fingerprint distance; when the fingerprint distance is not greater than the preset repetition distance threshold and the two coverage area identifiers have an overlapping relationship, mark the corresponding local standard data unit as duplicate candidate data and generate duplicate candidate data markers;

[0024] An anomalous state summary is generated when the data quality marker belongs to any of the following: missing quality marker, out-of-bounds quality marker, mutation quality marker, or communication anomaly quality marker; or when the data change trend shows an alternation between rising and falling change markers within a continuous time slice.

[0025] Optionally, step three specifically includes:

[0026] Local data features, node status information, duplicate candidate data markers, and abnormal state summaries are aligned into a multivariate state sequence according to time slice identifiers. The multivariate state sequence is then segmented and embedded through a state input layer to generate a state segment embedding sequence.

[0027] The state fragments are embedded into the sequence input temporal convolutional processing layer. The temporal convolutional processing layer performs layer normalization, pointwise convolutional dimensionality increase, deep large kernel temporal convolution, GELU activation, pointwise convolutional dimensionality reduction and residual connection in sequence according to the ModernTCN convolutional structure to generate temporal convolutional features.

[0028] The pointwise convolution dimensionality upscaling maps the fields corresponding to local data features to high-dimensional channels of data state, the fields corresponding to node state information to high-dimensional channels of node state, and the fields corresponding to duplicate candidate data labels and abnormal state summaries to high-dimensional channels of labeled state. The deep large-kernel temporal convolution extracts the change features of the high-dimensional channels of data state and node state within continuous time slices. The pointwise convolution dimensionality downscaling fuses the high-dimensional channels of labeled state with the high-dimensional channels of data state and node state to generate temporal convolution features.

[0029] The duplicate detection channel converts duplicate candidate data labels into duplicate label values, converts abnormal state summaries into abnormal label values, concatenates duplicate label values, abnormal label values, and temporal convolutional features into channels, and generates duplicate data scores through pointwise convolution, fully connected scoring layers, and Sigmoid normalization.

[0030] The node state fusion layer converts node state information into node state values ​​and performs normalization and embedding mapping to generate node state embeddings. The node state embeddings, temporal convolutional features, and expanded repeated data scores are concatenated by channels and processed by pointwise convolution, Sigmoid channel gating, and temporal pooling to generate node fusion state features.

[0031] The status scoring output layer generates a collaborative request score based on node fusion status features, anomaly marker values, and node status embedding. It also generates a task acceptance score based on node fusion status features, duplicate data scores, and node status embedding, and outputs the duplicate data score, collaborative request score, and task acceptance score.

[0032] Optionally, step four specifically includes:

[0033] Edge nodes with consistent time slice identifiers, the same measurement point type, overlapping coverage areas, and duplicate candidate data markers are grouped into the same duplicate candidate node group.

[0034] The duplicate data scores of each edge node in the duplicate candidate node group are judged by threshold, and the edge nodes whose duplicate data scores are not less than the preset duplicate score threshold are retained in the main processing node selection process.

[0035] In the main processing node selection process, edge nodes are sorted from high to low according to their task acceptance scores. The edge node with the highest score is selected as the main processing node. If there are two or more edge nodes with the same task acceptance score, they are sorted again according to the order of communication availability meeting collaborative processing requirements and the number of tasks to be processed from small to large.

[0036] The main processing node calls the locally deployed inference model to perform inference processing on the local standard data units corresponding to the duplicate candidate node group and generates the main processing result; the edge nodes in the duplicate candidate node group other than the main processing node enter the auxiliary retention state.

[0037] When the abnormal status summary in the auxiliary retention state contains the abnormal type corresponding to the missing quality marker, out-of-bounds quality marker, mutation quality marker, or communication abnormality quality marker, the abnormal status summary is sent to the main processing node; the main processing node compares the abnormal status summary with the main processing result and generates the verified main processing result.

[0038] Optionally, step five specifically includes:

[0039] When the collaborative request score is not less than the preset collaborative request threshold, collaborative task information is generated, which includes task type, real-time level and privacy level.

[0040] Adjacent edge nodes are filtered based on task type, response time limit, privacy level, and communication availability to form a candidate node set;

[0041] For each candidate node in the candidate node set, a node selection record is generated. The node selection record includes the candidate node identifier, task acceptance score, historical average collaborative benefit, number of times selected, and the most recent collaborative execution feedback. When a candidate node does not have a historical collaborative record, the task acceptance score is used as the initial benefit value, and the number of times selected is set to 1.

[0042] The selection value of candidate nodes is calculated using a task-constrained UCB node selection strategy. The selection value includes a basic benefit item, a state correction item, and an exploration compensation item. The basic benefit item is generated by the historical average collaborative benefit, the state correction item is generated by the task acceptance score, and the exploration compensation item is obtained by multiplying the difference between the maximum number of times the candidate node is selected in the candidate node set and the number of times the current candidate node is selected by a preset exploration coefficient.

[0043] The basic benefit item, status correction item, and exploration compensation item are added together to generate a selection value. The candidate node set is sorted from high to low according to the selection value. The candidate node with the highest ranking is selected as the target collaboration node. If the target collaboration node returns a task rejection message, fails to return a confirmation message within the response time limit, or changes the communication availability status to unavailable, the selection result is recorded as a negative collaboration record, and the next candidate node is selected as the new target collaboration node according to the ranking result.

[0044] Optionally, step six specifically includes:

[0045] The task initiating node matches the model splitting point between the local front-end model and the collaborative back-end model based on the task type of the task to be processed, the target collaborative node identifier, and the collaborative back-end model identifier supported by the target collaborative node.

[0046] When the intermediate layer output corresponding to the model split point can be sliced ​​according to the channel dimension or time slice dimension, the feature slicing process is entered. When the intermediate layer output corresponding to the model split point does not meet the slicing processing conditions, the feature slicing process is stopped, and the local lightweight model is used to complete the processing.

[0047] The task initiation node inputs the corresponding local standard data unit into the local front-end model, which then runs to the model splitting point to generate collaborative features.

[0048] Channel rearrangement and feature normalization are performed on the co-operative features to generate perturbed co-operative features; the perturbed co-operative features are sliced ​​according to the slice order index to generate feature slices;

[0049] The task initiating node sends the feature slice, task identifier, model split point identifier, and collaborative backend model identifier to the target collaborative node; the target collaborative node calls the corresponding collaborative backend model to perform forward computation on the feature slice and generate intermediate results.

[0050] The task initiating node receives the intermediate results returned by the target collaborating node, sorts the intermediate results according to the slice number and slice order index, and generates the final collaborative processing result according to the fusion rules corresponding to the model split points.

[0051] If the target collaboration node fails to return intermediate results within the response time limit, or if the returned intermediate results lack task identifiers, slice numbers, or model split point identifiers, the task initiating node will mark this collaboration process as a failure and re-execute step five to select a new target collaboration node.

[0052] Optionally, step seven specifically includes:

[0053] After generating the final collaborative processing result, the task initiating node generates collaborative execution feedback based on the task identifier, target collaborative node identifier, main processing node identifier, execution status, processing latency, and communication overhead.

[0054] When the execution status is successful and the processing delay does not exceed the response time limit in the collaborative task information, the corresponding target collaborative node is recorded as a valid acceptance result. When the execution status is task rejection, response timeout, calculation failure, or fusion failure, the corresponding target collaborative node is recorded as an invalid acceptance result, and the invalid acceptance result is used to reduce the basic benefit item of the target collaborative node in the task-constrained UCB node selection strategy.

[0055] The model update sample is composed of valid acceptance results, invalid acceptance results, duplicate candidate data labels, abnormal state summaries, and final collaborative processing result summaries. The valid acceptance results and invalid acceptance results are used as supervision labels to update the parameters of the state scoring output layer of the improved ModernTCN edge state evaluation model.

[0056] When the main processing result does not contain the abnormal type in the abnormal state summary under the auxiliary retention state, the corresponding abnormal type will be added to the set of mandatory review types of the abnormal retention rule. When the main processing result contains the abnormal type in the abnormal state summary, a preset sorting bonus item will be added to the subsequent sorting of the same measurement point type and coverage area corresponding to the current main processing node.

[0057] The processing latency and communication overhead corresponding to the effective acceptance results are included in the historical average collaboration benefit, the invalid acceptance results are included in the negative collaboration record, and the number of times the candidate node is selected, the historical average collaboration benefit, and the feedback of the most recent collaboration execution are updated.

[0058] The updated and improved ModernTCN edge state assessment model, master processing node determination rules, anomaly retention rules, and task-constrained UCB node selection strategy will be used for collaborative processing of IoT data in the next time slice.

[0059] The beneficial effects of this invention are:

[0060] This invention enables lightweight exchange between adjacent edge nodes using data state fingerprints and node state summaries. It first generates duplicate candidate data markers and abnormal state summaries, addressing the issues of duplicate data collection, duplicate inference, and direct filtering of abnormal information when multiple nodes cover the same area. This allows duplicate data to be processed uniformly at the main processing node while retaining an entry point for abnormal data verification. Through an improved ModernTCN edge state assessment model, local data features, node state information, duplicate candidate data markers, and abnormal state summaries are jointly modeled to output duplicate data scores, collaborative request scores, and task acceptance scores. This addresses the problem that relying solely on fixed thresholds or load indicators cannot accurately determine node processing status, providing continuous state-based criteria for main processing node determination and collaborative request triggering. The task-constrained UCB node selection strategy combines collaborative task information, task acceptance scores, and historical collaborative records to select target collaborative nodes, addressing issues of sudden node overload and fixed scheduling being unsuitable for changing field conditions. Finally, through feature slice collaborative processing, only perturbed feature slices and intermediate results are transmitted, reducing the risk of exposing raw IoT data. Its significance lies in transforming the IoT edge from independent processing into a verifiable, collaborative, and updatable on-site data processing system, thereby improving edge computing power utilization, collaborative response stability, and data processing security. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of an IoT data collaborative processing method based on edge computing proposed in this invention;

[0063] Figure 2 This is a schematic diagram of an IoT data collaborative processing method based on edge computing proposed in this invention;

[0064] Figure 3 This is a framework diagram of the improved ModernTCN edge state evaluation model in the IoT data collaborative processing method based on edge computing proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figures 1-3 An IoT data collaborative processing method based on edge computing includes the following steps:

[0067] Step 1: Edge nodes receive IoT data, generate local standard data units according to time slices, and extract local data features, node status information, and data status fingerprints;

[0068] Step 2: Adjacent edge nodes exchange data state fingerprints and node state summaries to generate duplicate candidate data markers and abnormal state summaries;

[0069] Step 3: Input the local data features, node status information, duplicate candidate data labels and abnormal status summaries into the improved ModernTCN edge state evaluation model. After temporal convolution, duplicate judgment channel and node status fusion layer, the duplicate data score, collaborative request score and task acceptance score are output.

[0070] Step 4: Determine the main processing node and perform inference based on the duplicate candidate data marking, duplicate data score and task acceptance score. The remaining nodes enter the auxiliary retention state. When the abnormal state summary meets the abnormal retention rules, a review is triggered.

[0071] Step 5: When the collaboration request score meets the collaboration request conditions, a set of candidate nodes is formed based on the collaboration task information, task acceptance score and historical collaboration records. The target collaboration node is selected using the task-constrained UCB node selection strategy.

[0072] Step 6: Generate collaborative features based on model segmentation points, obtain feature slices through perturbation and slicing, and send them to the target collaborative node to perform collaborative post-processing model calculations. The task initiating node then fuses intermediate results to generate the final collaborative processing result.

[0073] Step 7: Generate collaborative execution feedback based on the final collaborative processing results, execution status, processing latency, and communication overhead, and update the improved ModernTCN edge state assessment model, main processing node determination rules, exception retention rules, and task-constrained UCB node selection strategy.

[0074] In this embodiment, step one specifically includes:

[0075] Edge nodes read IoT data access configurations. Edge nodes are edge computing devices deployed on the IoT device side or on the field network side, capable of receiving IoT data, processing local data, collecting node status, and communicating between nodes. From the received IoT data, they extract device identifier, measurement point type, collection time, data value, and coverage area identifier. They perform timestamp correction on the collection time and group data with the same device identifier and the same measurement point type into the corresponding time slice according to the preset time slice length to form local standard data units.

[0076] In practice, the preset time slice length is set to 1 second, 5 seconds, or 10 seconds. For measurement points with high frequency of change, such as vibration, current, and pressure, the preset time slice length is set to 1 second; for measurement points with slower change, such as temperature, humidity, and energy consumption, the preset time slice length is set to 5 seconds or 10 seconds. The local standard data unit includes a device identification field, a measurement point type field, a time slice identification field, a data value field, a coverage area identification field, and a data quality marker field.

[0077] Data quality is marked for local standard data units. If no data value is received in the corresponding time slice, it is judged as a missing quality mark; if the data value exceeds the allowable value range for the measurement point type, it is judged as an out-of-bounds quality mark; if the difference between the data value of the current time slice and the data value of the previous time slice reaches the mutation threshold corresponding to the measurement point type, it is judged as a mutation quality mark; if the edge node does not receive the data packet of the corresponding device within the acquisition period, it is judged as a communication anomaly quality mark; if the above quality marks are not triggered, it is judged as a normal quality mark.

[0078] Read local standard data units within a continuous time slice, compare the current time slice data value with the previous time slice data value, generate rising change markers, falling change markers, or unchanged change markers, and use them as data change trends; calculate the dispersion of data values ​​of the same measurement point type within a continuous time slice, and use the calculation result as the degree of fluctuation; read abrupt change quality markers and generate short-term abrupt change states; combine data change trends, fluctuation degrees, and short-term abrupt change states to form local data features;

[0079] In practice, the number of continuous time slices is set to 8 or 16. The degree of fluctuation can be represented by the variance, range or normalized coefficient of variation of the data values ​​within the continuous time slice. Short-term mutation states are generated based on mutation quality markers. When a mutation quality marker appears within a continuous time slice, the corresponding measurement point is marked as a short-term mutation state.

[0080] Read the running status of edge nodes, generate node load status based on processing load, generate task queuing status based on the number of tasks to be processed, generate communication availability status based on the connection status with adjacent edge nodes, and combine node load status, task queuing status and communication availability status to form node status information.

[0081] The fingerprint input record is generated by combining local data features, measurement point type, time slice identifier and coverage area identifier according to the preset field order. Local sensitive hashing is then performed on the fingerprint input record to obtain the data status fingerprint.

[0082] In practice, locality-sensitive hashing uses SimHash or MinHash encoding, and the data state fingerprint length is set to 64 bits or 128 bits. The data state fingerprint does not contain the original data value and is used to determine duplicate data between adjacent edge nodes.

[0083] In this embodiment, step two specifically includes:

[0084] Neighboring edge nodes send their data status fingerprints and node status summaries according to a preset exchange cycle. The node status summary is formed by combining the node load status, task queuing status, and communication availability status in the node status information. After receiving the data status fingerprints and node status summaries sent by neighboring edge nodes, they are matched according to time slice identifiers and measurement point types. Data status fingerprints with the same time slice identifiers and the same measurement point type or belonging to the same associated measurement point group are used as fingerprints to be compared. The associated measurement point group is a pre-configured set of measurement points belonging to the same equipment, the same process section, or the same coverage area.

[0085] In practice, the preset exchange cycle can be set to one time slice or two time slices. The associated measurement point group is a pre-configured set of measurement points belonging to the same equipment, the same process section, or the same coverage area. For example, vibration measurement points, current measurement points, and temperature measurement points of the same machine tool spindle can be configured as the same associated measurement point group; adjacent cameras or adjacent environmental sensors in the same storage area can be configured as the same associated measurement point group.

[0086] Calculate the Hamming distance between the fingerprint to be compared at this node and the fingerprint to be compared at adjacent edge nodes, and use the Hamming distance as the fingerprint distance; read the coverage area identifier of this node and the coverage area identifier of adjacent edge nodes; when the fingerprint distance is not greater than the preset repetition distance threshold and the two coverage area identifiers overlap, mark the corresponding local standard data unit as a duplicate candidate data and generate a duplicate candidate data label; when the fingerprint distance is greater than the preset repetition distance threshold, or the two coverage area identifiers do not overlap, do not generate a duplicate candidate data label.

[0087] In specific implementation, when the data status fingerprint length is 64 bits, the preset repetition distance threshold is set to 6 to 10 bits; when the data status fingerprint length is 128 bits, the preset repetition distance threshold is set to 10 to 18 bits. The coverage area identifier can be the equipment area number, production line section number, warehouse area number, or gateway coverage number.

[0088] Read the data quality markers of local standard data units, short-term mutation states and data change trends in local data characteristics; when the data quality marker is determined to be any of the missing quality marker, out-of-bounds quality marker, mutation quality marker or communication anomaly quality marker, or when the data change trend shows an alternating state of rising and falling change markers within a continuous time slice, generate an anomaly state summary, which includes the device identifier, measurement point type, time slice identifier, anomaly type and corresponding duplicate candidate data markers;

[0089] In practice, even if a piece of data is marked as a duplicate candidate data, as long as its corresponding abnormal state summary contains a mutation quality marker, out-of-bounds quality marker, or communication abnormal quality marker, the corresponding abnormal state summary will still be retained for subsequent review by the main processing node.

[0090] In this embodiment, step three specifically includes:

[0091] The local data features, node state information, duplicate candidate data markers, and abnormal state summaries are used as inputs to the improved ModernTCN edge state evaluation model and aligned according to time slice identifiers to form a multivariate state sequence.

[0092] The multivariate state sequence is input into the state input layer. The state input layer performs segmentation on the multivariate state sequence, dividing the local data features, node state information, duplicate candidate data markers and abnormal state summaries within the continuous time slice into state segments, and performs embedding mapping on the state segments to generate state segment embedding sequences.

[0093] In practical implementation, the state segment length is set to 4 or 8 time slices, and the embedding dimension is set to 32 or 64 dimensions. For different measurement point types and node state fields, the state input layer is configured with field embedding matrices to convert discrete markers and continuous values ​​into state segment embeddings with the same embedding dimension.

[0094] State segments are embedded into the sequence input temporal convolutional processing layer. This layer, following the ModernTCN convolutional structure, sequentially performs layer normalization, pointwise convolutional dimensionality increase, deep large-kernel temporal convolution, GELU activation, pointwise convolutional dimensionality reduction, and residual connection processing to generate temporal convolutional features. Specifically, in the pointwise convolutional dimensionality increase process, fields corresponding to local data features are mapped to high-dimensional data state channels, fields corresponding to node state information are mapped to high-dimensional node state channels, and fields corresponding to duplicate candidate data labels and abnormal state summaries are mapped to high-dimensional labeled state channels. In the deep large-kernel temporal convolution process, change features within consecutive time slices are extracted from the high-dimensional data state channels and the high-dimensional node state channels. In the pointwise convolutional dimensionality reduction process, the high-dimensional labeled state channels are fused with the high-dimensional data state channels and the high-dimensional node state channels to generate temporal convolutional features.

[0095] In practice, the dimensionality increase factor of pointwise convolution is set to 4 times, the kernel length of depthwise large kernel temporal convolution is set to 7 or 15, and the residual connection adds the embedded features before and after the convolution process to preserve the original time slice state information.

[0096] In the duplicate detection channel, duplicate candidate data labels are converted into duplicate label values, and abnormal state summaries are converted into abnormal label values. After expanding the duplicate label values ​​and abnormal label values ​​to the same time dimension as the temporal convolutional features, they are concatenated with the temporal convolutional features. The concatenated results are processed by pointwise convolution, and then input into a fully connected scoring layer for feature compression and score mapping. Finally, the duplicate data scores are generated by Sigmoid normalization.

[0097] In practice, time slices with duplicate candidate data markers correspond to a duplicate marker value of 1, while time slices without duplicate candidate data markers correspond to a duplicate marker value of 0; time slices with abnormal state summaries correspond to an abnormal marker value of 1, while time slices without abnormal state summaries correspond to an abnormal marker value of 0. The duplicate judgment channel participates in the scoring mapping through the abnormal marker value, so that data with abnormal state summaries, even if they belong to duplicate candidate data, will not be directly treated as ordinary duplicate data.

[0098] In the node state fusion layer, the node load state, task queuing state, and communication availability state in the node state information are converted into load values, queuing values, and communication values, respectively. Normalization and embedding mapping are then performed on these values ​​to generate node state embeddings. The node state embeddings are extended to the same time dimension as the temporal convolutional features and then concatenated with the temporal convolutional features through channels to generate node state concatenated features. Duplicate data scores are extended to the same time dimension as the node state concatenated features and then concatenated with the node state concatenated features through channels again to generate node score concatenated features. Pointwise convolution processing is performed on the node score concatenated features, and channel gating coefficients are generated using the Sigmoid function. These channel gating coefficients are then multiplied by the node score concatenated features channel-wise to generate weighted node state features. Finally, temporal pooling is performed on the weighted node state features to generate fused node state features.

[0099] In the state scoring output layer, the node fusion state features, the anomaly label value obtained from the anomaly state summary, and the node state embedding obtained from the node state information are concatenated, and then processed by a fully connected scoring layer for feature compression and score mapping, and finally normalized by Sigmoid to generate a collaborative request score. The node fusion state features, duplicate data scores, and the node state embedding obtained from the node state information are concatenated, and then processed by a fully connected scoring layer for feature compression and score mapping, and finally normalized by Sigmoid to generate a task acceptance score. The duplicate data score, collaborative request score, and task acceptance score are used as the output results of the improved ModernTCN edge state evaluation model.

[0100] In this embodiment, step four specifically includes:

[0101] Read the duplicate candidate data markers, duplicate data scores, and task acceptance scores, and group edge nodes with the same time slice identifier, the same measurement point type, or belonging to the same related measurement point group, overlapping coverage areas, and duplicate candidate data markers into the same duplicate candidate node group.

[0102] For each edge node in the duplicate candidate node group, a threshold judgment is made on the duplicate data score. Edge nodes with duplicate data scores not less than the preset duplicate score threshold are retained in the main processing node selection process. When there are no edge nodes with duplicate data scores not less than the preset duplicate score threshold in the duplicate candidate node group, the corresponding edge node maintains the local processing process and does not enter the auxiliary retention state.

[0103] In practice, the preset duplicate scoring threshold can be set to 0.60 or 0.70. When the score of duplicate data is not less than the threshold, it means that the corresponding data has duplicate processing value and can enter the main processing node selection process.

[0104] In the main processing node selection process, the task acceptance score of each edge node in the duplicate candidate node group is read and sorted from high to low according to the task acceptance score. The edge node with the highest score is selected as the main processing node. When there are two or more edge nodes with the same task acceptance score, they are sorted in a second order according to the communication availability status meeting the collaborative processing requirements and the number of tasks to be processed from small to large. The edge node with the highest score in the second sort is selected as the main processing node.

[0105] The main processing node calls the locally deployed inference model to perform inference processing on the local standard data units corresponding to the duplicate candidate node group and generates the main processing result. The edge nodes in the duplicate candidate node group other than the main processing node enter the auxiliary retention state. The auxiliary retention state does not perform the same inference processing as the main processing node, and retains the data state fingerprint, duplicate candidate data marker and abnormal state summary of the corresponding time slice.

[0106] Read the abnormal state summary of the edge node in the auxiliary retention state. When the abnormal state summary contains the abnormal type corresponding to the missing quality mark, out-of-bounds quality mark, mutation quality mark or communication abnormal quality mark, it is determined that the abnormal retention rule is met, and the abnormal state summary is sent to the main processing node.

[0107] After receiving the abnormal status summary, the main processing node compares the device identifier, measurement point type, time slice identifier, and abnormal type in the abnormal status summary with the main processing result. When the main processing result does not contain the corresponding abnormal type, the abnormal status summary is incorporated into the main processing result. When the main processing result already contains the corresponding abnormal type, the main processing result is retained and an abnormal review mark is added to generate the reviewed main processing result.

[0108] In this embodiment, step five specifically includes:

[0109] Read the collaborative request score. When the collaborative request score is not less than the preset collaborative request threshold, generate collaborative task information, which includes task type, real-time level and privacy level. Match whether the corresponding collaborative backend model is deployed on adjacent edge nodes according to the task type, match the response time limit of adjacent edge nodes according to the real-time level, and match whether adjacent edge nodes support feature slicing processing according to the privacy level.

[0110] In practice, the preset collaborative request threshold is set to 0.65 or 0.70. Task types include anomaly detection tasks, device status recognition tasks, classification tasks, and regression tasks. Real-time performance levels include high real-time, medium real-time, and low real-time. The privacy level is used to determine whether feature slicing processing is necessary.

[0111] Read the task acceptance score, node status summary and historical collaboration record of adjacent edge nodes. The historical collaboration record includes the execution status, processing latency, communication overhead, number of task rejections and number of response timeouts within the most recent time window. Add adjacent edge nodes that meet the requirements of task type, response time limit, privacy level and communication availability to the candidate node set, and remove adjacent edge nodes that do not meet any of the above requirements from the candidate node set.

[0112] For each candidate node in the candidate node set, a node selection record is generated. The node selection record includes the candidate node identifier, task acceptance score, historical average collaboration benefit, number of times it was selected, and the feedback of the most recent collaboration execution. The historical average collaboration benefit is generated based on the execution status, processing latency, and communication overhead within the most recent time window. When a candidate node does not have a historical collaboration record, the task acceptance score of the candidate node is used as the initial benefit value, and the number of times the candidate node was selected is set to 1.

[0113] A task-constrained UCB node selection strategy is used to calculate the selection value of candidate nodes. The selection value includes a basic benefit item, a state correction item, and an exploration compensation item. The basic benefit item is generated from the historical average collaborative benefit, and the state correction item is generated from the task acceptance score. The number of times each candidate node in the candidate node set within the most recent time window is read, the maximum number of times selected is selected, the difference between the maximum number of times selected and the current candidate node's selection count is calculated, and the difference is multiplied by a preset exploration coefficient to obtain the exploration compensation item for the current candidate node. The basic benefit item, state correction item, and exploration compensation item are added together to generate the selection value for the current candidate node.

[0114] The candidate node set is sorted from highest to lowest according to the selection value, and the candidate node with the highest ranking is selected as the target collaboration node. When the target collaboration node returns a task rejection message, fails to return an acknowledgment message within the response time limit, or changes its communication availability status to unavailable, the selection result of the target collaboration node is recorded as a negative collaboration record, and the next candidate node is selected as the new target collaboration node according to the ranking result.

[0115] In this embodiment, step six specifically includes:

[0116] The task initiating node reads the task type, target collaboration node identifier, and collaboration backend model identifier supported by the target collaboration node for the task to be processed, and matches the model splitting points of the local frontend model and the collaboration backend model according to the task type.

[0117] In practical implementation, the local front-end model can be the front-end network structure of an anomaly detection model, device status recognition model, classification model, or regression model, while the collaborative back-end model is the subsequent network structure that matches the local front-end model structure. The model splitting point can be set after convolutional layers, temporal convolutional layers, embedding layers, or fully connected layers.

[0118] When the intermediate layer output corresponding to the model split point can be sliced ​​according to the channel dimension or time slice dimension, the feature slicing process is initiated; when the intermediate layer output corresponding to the model split point does not meet the slicing processing conditions, the feature slicing process is stopped, and the local lightweight model is used to complete the processing. The local lightweight model is a lightweight inference model deployed on the task initiation node for the task to be processed.

[0119] The task initiation node inputs the corresponding local standard data unit into the local front-end model, which then runs to the model split point to generate collaborative features. The collaborative features are the intermediate layer features output by the model split point, including channel number, time slice number, and feature value.

[0120] The task initiating node performs perturbation processing on the collaborative features. The perturbation processing includes channel rearrangement and feature normalization. Channel rearrangement adjusts the channel arrangement order of the collaborative features according to the slice order index generated by the task initiating node. Feature normalization performs scale transformation on the feature values ​​of the collaborative features according to preset normalization parameters to generate perturbed collaborative features.

[0121] The task initiation node performs slicing processing on the perturbed collaborative features according to the slice order index to generate feature slices; the feature slices include slice number, task identifier, model split point identifier, channel range, time slice range, and slice feature value;

[0122] The task initiating node sends the feature slice, task identifier, model split point identifier, and collaborative back-end model identifier to the target collaborative node; the target collaborative node calls the corresponding collaborative back-end model according to the collaborative back-end model identifier, loads the corresponding back-end model entry according to the model split point identifier, inputs the received feature slice into the collaborative back-end model to perform forward calculation, and generates intermediate results;

[0123] The target collaboration node returns the intermediate results, slice numbers, and task identifiers to the task initiation node. The task initiation node matches the task to be fused based on the task identifier, sorts each intermediate result according to the slice number and slice order index, and fuses the sorted intermediate results according to the fusion rules corresponding to the model split points to generate the final collaborative processing result. The fusion rules include concatenation by channel number, concatenation by time slice number, or aggregation according to the output dimension of the collaborative back-end model.

[0124] When the target collaboration node fails to return intermediate results within the response time limit, or when the returned intermediate results lack task identifier, slice number, or model split point identifier, the task initiating node will mark this collaboration process as a failure and re-execute step five to select a new target collaboration node.

[0125] In practice, when the task is to detect device anomalies, the local front-end model encodes the features of the local standard data units, and the model segmentation point is set after the feature embedding layer. The target collaborative node performs the subsequent classification or anomaly judgment. When the task is to identify device status, the local front-end model outputs time slice features, and the target collaborative node performs the subsequent status identification on the feature slices.

[0126] In this embodiment, step seven specifically includes:

[0127] After generating the final collaborative processing result, the task initiating node reads the task identifier, target collaborative node identifier, main processing node identifier, execution status, processing latency, and communication overhead. The execution status includes execution success, task rejection, response timeout, computation failure, and fusion failure.

[0128] The collaborative execution process is associated with the duplicate data score, collaborative request score, and task acceptance score output in step three based on the task identifier, generating collaborative execution feedback; the collaborative execution feedback includes the target collaborative node identifier, collaborative task information, execution status, processing latency, communication overhead, and a summary of the final collaborative processing result;

[0129] When the execution status is successful and the processing delay does not exceed the response time limit in the collaborative task information, the corresponding target collaborative node is recorded as a valid acceptance result; when the execution status is task rejection, response timeout, computation failure, or fusion failure, the corresponding target collaborative node is recorded as an invalid acceptance result, and the invalid acceptance result is used to reduce the basic benefit item of the target collaborative node in the task-constrained UCB node selection strategy.

[0130] Based on the improved ModernTCN edge state assessment model updated by collaborative execution feedback, the model update sample is composed of effective acceptance results, invalid acceptance results, duplicate candidate data labels, abnormal state summaries and final collaborative processing result summaries. The effective acceptance results and invalid acceptance results are used as supervision labels to update the output parameters of the state scoring output layer.

[0131] The rules for determining the main processing node are updated based on the feedback from collaborative execution. When the main processing result does not contain the abnormal type in the abnormal state summary under the auxiliary retention state, the corresponding abnormal type is added to the set of mandatory review types of the abnormal retention rules. When the main processing result corresponding to the main processing node contains the abnormal type in the abnormal state summary, a preset sorting bonus item is added to the subsequent sorting of the same measurement point type and coverage area corresponding to the main processing node.

[0132] The task-constrained UCB node selection strategy is updated based on the collaborative execution feedback. The processing latency and communication overhead corresponding to the effective acceptance results are included in the historical average collaborative benefits, and the invalid acceptance results are included in the negative collaborative records. The number of times the candidate node is selected, the historical average collaborative benefits, and the most recent collaborative execution feedback are also updated.

[0133] The updated and improved ModernTCN edge state assessment model, master processing node determination rules, anomaly retention rules, and task-constrained UCB node selection strategy will be used for collaborative processing of IoT data in the next time slice.

[0134] Example 1: To verify the feasibility of this invention in practice, it was applied to a networked monitoring scenario for equipment in an industrial park. The park deployed 12 edge nodes, connecting to CNC machine tools, air compressors, conveyor motors, environmental sensors, and energy consumption acquisition terminals, totaling 168 monitoring points. Edge nodes generated local standard data units in 5-second time slices. The abrupt change threshold for vibration and current monitoring points was set to exceed 25% of the historical stable interval average in adjacent time slices, while the abrupt change threshold for temperature and energy consumption monitoring points was set to exceed 15%. Data state fingerprints used 64-bit locality-sensitive hashing, with a preset repetition distance threshold of 8 bits. Node state summaries were exchanged every 10 seconds. The improved ModernTCN edge state evaluation model had a state segment length of 8 time slices, an embedding dimension of 64 dimensions, and a depthwise large-kernel temporal convolution kernel length of 15. The preset repetition scoring threshold was set to 0.70, and the preset collaborative request threshold was set to 0.65. In the task-constrained UCB node selection strategy, the nearest time window was set to 30 minutes, and the preset exploration coefficient was set to 0.08.

[0135] During implementation, multiple edge nodes simultaneously collect vibration, current, and temperature data from adjacent equipment on the same production line. Traditionally, each node infers independently and uploads the results, easily leading to duplicate processing of the same equipment status. In this invention, adjacent edge nodes first exchange data status fingerprints and node status summaries, generating duplicate candidate data markers and abnormal status summaries. Then, the improved ModernTCN edge status assessment model outputs duplicate data scores, collaboration request scores, and task acceptance scores. When the data status fingerprints of two edge nodes within a certain area are close and their coverage areas overlap, the node with the higher task acceptance score becomes the primary processing node to perform inference, while the remaining nodes enter an auxiliary retention state. If out-of-bounds quality markers or communication abnormality quality markers still exist in the auxiliary retention state, the primary processing node is triggered for review. For nodes with high inference pressure, a task-constrained UCB node selection strategy is used to select target collaboration nodes based on collaboration task information, task acceptance scores, and historical collaboration records. The subsequent model calculation is then completed through feature slicing collaboration processing.

[0136] To verify the effectiveness, this invention is compared with three other methods. Method A is an independent processing scheme for each edge node; Method B is an edge collaboration scheme based on central scheduling; Method C is a scheme that uses ordinary ModernTCN for node status scoring but does not set up a duplicate judgment channel and a node status fusion layer; This invention is a scheme that uses data status fingerprinting, an improved ModernTCN edge status evaluation model, a task-constrained UCB node selection strategy, and feature slicing collaborative processing.

[0137] Table 1 Comparison of IoT Data Collaborative Processing Effects

[0138] Method A 438 186 0.00 9 124.6 78.3 Method B 261 164 86.2 6 91.4 69.8 Method C 193 142 89.5 4 68.7 63.5 This invention 82 109 95.8 1 31.9 52.6

[0139] As shown in Table 1, Method A has the highest number of repeated inferences due to the lack of duplicate candidate data marking and main processing node mechanism, with an average node load of 78.3%. Although Method B reduces some duplicate processing through central scheduling, it still requires more node status reporting and unified allocation, with an average processing latency of 164ms. Method C can identify some node load changes after introducing ordinary ModernTCN, but because it does not include duplicate candidate data marking and abnormal status summary in the duplicate judgment channel, the number of abnormal missed checks is still 4 times / day. This invention uses data state fingerprints for pre-processing duplicate detection, inputting duplicate candidate data tags and abnormal state summaries into an improved ModernTCN edge state assessment model. This enables duplicate data scoring, collaborative request scoring, and task acceptance scoring to directly serve the determination of the main processing node and the selection of the target collaborative node. The number of repeated inferences is reduced to 82 times / hour, a reduction of approximately 81.3% compared to Method A; the average processing latency is reduced to 109ms, a reduction of approximately 33.5% compared to Method B; and the collaborative task success rate reaches 95.8%, indicating that the task-constrained UCB node selection strategy can reliably select acceptable nodes. The amount of raw data transmitted is reduced to 31.9MB / hour, mainly because step six only transmits perturbed feature slices and intermediate results, rather than directly transmitting complete IoT data.

[0140] In combination with the above embodiments, the present invention can reduce redundant inference and redundant transmission in scenarios with dense deployment of multiple edge nodes, reduce the risk of anomalies being masked by repeated data processing through an anomaly state summary retention mechanism, improve the stability of collaborative node selection through a task-constrained UCB node selection strategy, and reduce the amount of original IoT data transmitted outward through feature slicing collaborative processing. It is suitable for IoT edge collaborative processing scenarios such as industrial equipment monitoring, park energy consumption management, and on-site intelligent operation and maintenance.

[0141] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An edge-computing-based Internet of Things data collaborative processing method, characterized in that, Includes the following steps: Step 1: Edge nodes receive IoT data, generate local standard data units according to time slices, and extract local data features, node status information, and data status fingerprints; Step 2: Adjacent edge nodes exchange data state fingerprints and node state summaries to generate duplicate candidate data markers and abnormal state summaries; Step 3: Input the local data features, node status information, duplicate candidate data labels and abnormal status summaries into the improved ModernTCN edge state evaluation model. After temporal convolution, duplicate judgment channel and node status fusion layer, the duplicate data score, collaborative request score and task acceptance score are output. Step 4: Determine the main processing node and perform inference based on the duplicate candidate data marking, duplicate data score and task acceptance score. The remaining nodes enter the auxiliary retention state. When the abnormal state summary meets the abnormal retention rules, a review is triggered. Step 5: When the collaboration request score meets the collaboration request conditions, a set of candidate nodes is formed based on the collaboration task information, task acceptance score and historical collaboration records. The target collaboration node is selected using the task-constrained UCB node selection strategy. Step 6: Generate collaborative features based on model segmentation points, obtain feature slices through perturbation and slicing, and send them to the target collaborative node to perform collaborative post-processing model calculations. The task initiating node then fuses intermediate results to generate the final collaborative processing result. Step 7: Generate collaborative execution feedback based on the final collaborative processing results, execution status, processing latency, and communication overhead, and update the improved ModernTCN edge state assessment model, main processing node determination rules, exception retention rules, and task-constrained UCB node selection strategy. 2.The edge computing based data collaborative processing method of Internet of Things according to claim 1, characterized in that, Step one specifically includes: According to the IoT data access configuration, the edge node extracts the device identifier, measurement point type, collection time, data value and coverage area identifier from the IoT data, performs timestamp correction on the collection time, and classifies the data with the same device identifier and the same measurement point type into the corresponding time slice according to the preset time slice length to form a local standard data unit. Data quality is marked for local standard data units. Data for which no data value is received is marked as missing quality, data for which the value exceeds the allowable range is marked as out of bounds quality, data for which the difference between adjacent time slices reaches the mutation threshold is marked as mutation quality, data for which no device message is received within the acquisition period is marked as communication abnormality quality, and the rest of the data is marked as normal quality. Data change trends are generated based on comparisons of data values ​​within continuous time slices, fluctuation levels are generated based on the dispersion of data values ​​within continuous time slices, short-term mutation states are generated based on mutation quality markers, and these are combined to form local data features. The node load status is generated based on the processing load, the task queuing status is generated based on the number of tasks to be processed, and the communication availability status is generated based on the connection status of adjacent edge nodes. These are combined to form node status information. The fingerprint input record is generated by combining local data features, measurement point type, time slice identifier, and coverage area identifier according to the preset field order. Locality-sensitive hashing is then performed on the fingerprint input record to obtain the data status fingerprint. 3.The edge computing based data collaborative processing method of Internet of Things according to claim 1, characterized in that, Step two specifically includes: Adjacent edge nodes send data status fingerprints and node status summaries according to a preset exchange cycle. The node status summary is formed by a combination of node load status, task queuing status, and communication availability status. Time slice identification and measurement point type matching are performed on the data state fingerprints of this node and its adjacent edge nodes, and the matching data state fingerprints are used as the fingerprints to be compared. Calculate the Hamming distance between the fingerprint to be compared at this node and the fingerprint to be compared at the adjacent edge node, and use the Hamming distance as the fingerprint distance; when the fingerprint distance is not greater than the preset repetition distance threshold and the two coverage area identifiers have an overlapping relationship, mark the corresponding local standard data unit as duplicate candidate data and generate duplicate candidate data markers; An anomalous state summary is generated when the data quality marker belongs to any of the following: missing quality marker, out-of-bounds quality marker, mutation quality marker, or communication anomaly quality marker; or when the data change trend shows an alternation between rising and falling change markers within a continuous time slice. 4.The edge computing based data collaborative processing method of Internet of Things according to claim 1, characterized in that, Step three specifically includes: Local data features, node status information, duplicate candidate data markers, and abnormal state summaries are aligned into a multivariate state sequence according to time slice identifiers. The multivariate state sequence is then segmented and embedded through a state input layer to generate a state segment embedding sequence. The state fragments are embedded into the sequence input temporal convolutional processing layer. The temporal convolutional processing layer performs layer normalization, pointwise convolutional dimensionality increase, deep large kernel temporal convolution, GELU activation, pointwise convolutional dimensionality reduction and residual connection in sequence according to the ModernTCN convolutional structure to generate temporal convolutional features. The pointwise convolution dimensionality upscaling maps the fields corresponding to local data features to high-dimensional channels of data state, the fields corresponding to node state information to high-dimensional channels of node state, and the fields corresponding to duplicate candidate data labels and abnormal state summaries to high-dimensional channels of labeled state. The deep large-kernel temporal convolution extracts the change features of the high-dimensional channels of data state and node state within continuous time slices. The pointwise convolution dimensionality downscaling fuses the high-dimensional channels of labeled state with the high-dimensional channels of data state and node state to generate temporal convolution features. The duplicate detection channel converts duplicate candidate data labels into duplicate label values, converts abnormal state summaries into abnormal label values, concatenates duplicate label values, abnormal label values, and temporal convolutional features into channels, and generates duplicate data scores through pointwise convolution, fully connected scoring layers, and Sigmoid normalization. The node state fusion layer converts node state information into node state values ​​and performs normalization and embedding mapping to generate node state embeddings. The node state embeddings, temporal convolutional features, and expanded repeated data scores are concatenated by channels and processed by pointwise convolution, Sigmoid channel gating, and temporal pooling to generate node fusion state features. The status scoring output layer generates a collaborative request score based on node fusion status features, anomaly marker values, and node status embedding. It also generates a task acceptance score based on node fusion status features, duplicate data scores, and node status embedding, and outputs the duplicate data score, collaborative request score, and task acceptance score. 5.The edge computing based data collaborative processing method of Internet of Things according to claim 1, characterized in that, Step four specifically includes: Edge nodes with consistent time slice identifiers, the same measurement point type, overlapping coverage areas, and duplicate candidate data markers are grouped into the same duplicate candidate node group. The duplicate data scores of each edge node in the duplicate candidate node group are judged by threshold, and the edge nodes whose duplicate data scores are not less than the preset duplicate score threshold are retained in the main processing node selection process. In the main processing node selection process, edge nodes are sorted from high to low according to their task acceptance scores. The edge node with the highest score is selected as the main processing node. If there are two or more edge nodes with the same task acceptance score, they are sorted again according to the order of communication availability meeting collaborative processing requirements and the number of tasks to be processed from small to large. The main processing node calls the locally deployed inference model to perform inference processing on the local standard data units corresponding to the duplicate candidate node group and generates the main processing result; the edge nodes in the duplicate candidate node group other than the main processing node enter the auxiliary retention state. When the abnormal status summary in the auxiliary retention state contains the abnormal type corresponding to the missing quality marker, out-of-bounds quality marker, mutation quality marker, or communication abnormality quality marker, the abnormal status summary is sent to the main processing node; the main processing node compares the abnormal status summary with the main processing result and generates the verified main processing result.

6. The IoT data collaborative processing method based on edge computing according to claim 1, characterized in that, Step five specifically includes: When the collaborative request score is not less than the preset collaborative request threshold, collaborative task information is generated, which includes task type, real-time level and privacy level. Adjacent edge nodes are filtered based on task type, response time, privacy level, and communication availability to form a candidate node set; For each candidate node in the candidate node set, a node selection record is generated. The node selection record includes the candidate node identifier, task acceptance score, historical average collaborative benefit, number of times selected, and the most recent collaborative execution feedback. When a candidate node does not have a historical collaborative record, the task acceptance score is used as the initial benefit value, and the number of times selected is set to 1. The selection value of candidate nodes is calculated using a task-constrained UCB node selection strategy. The selection value includes a basic benefit item, a state correction item, and an exploration compensation item. The basic benefit item is generated by the historical average collaborative benefit, the state correction item is generated by the task acceptance score, and the exploration compensation item is obtained by multiplying the difference between the maximum number of times the candidate node is selected in the candidate node set and the number of times the current candidate node is selected by a preset exploration coefficient. The basic benefit item, status correction item, and exploration compensation item are added together to generate a selection value. The candidate node set is sorted from high to low according to the selection value. The candidate node with the highest ranking is selected as the target collaboration node. If the target collaboration node returns a task rejection message, fails to return a confirmation message within the response time limit, or changes the communication availability status to unavailable, the selection result is recorded as a negative collaboration record, and the next candidate node is selected as the new target collaboration node according to the ranking result.

7. The IoT data collaborative processing method based on edge computing according to claim 1, characterized in that, Step six specifically includes: The task initiating node matches the model splitting point between the local front-end model and the collaborative back-end model based on the task type of the task to be processed, the target collaborative node identifier, and the collaborative back-end model identifier supported by the target collaborative node. When the intermediate layer output corresponding to the model split point can be sliced ​​according to the channel dimension or time slice dimension, the feature slicing process is entered. When the intermediate layer output corresponding to the model split point does not meet the slicing processing conditions, the feature slicing process is stopped, and the local lightweight model is used to complete the processing. The task initiation node inputs the corresponding local standard data unit into the local front-end model, which then runs to the model splitting point to generate collaborative features. Channel rearrangement and feature normalization are performed on the co-operative features to generate perturbed co-operative features; the perturbed co-operative features are sliced ​​according to the slice order index to generate feature slices; The task initiating node sends the feature slice, task identifier, model split point identifier, and collaborative backend model identifier to the target collaborative node; the target collaborative node calls the corresponding collaborative backend model to perform forward computation on the feature slice and generate intermediate results. The task initiating node receives the intermediate results returned by the target collaborating node, sorts the intermediate results according to the slice number and slice order index, and generates the final collaborative processing result according to the fusion rules corresponding to the model split points. If the target collaboration node fails to return intermediate results within the response time limit, or if the returned intermediate results lack task identifiers, slice numbers, or model split point identifiers, the task initiating node will mark this collaboration process as a failure and re-execute step five to select a new target collaboration node.

8. The IoT data collaborative processing method based on edge computing according to claim 1, characterized in that, Step seven specifically includes: After generating the final collaborative processing result, the task initiating node generates collaborative execution feedback based on the task identifier, target collaborative node identifier, main processing node identifier, execution status, processing latency, and communication overhead. When the execution status is successful and the processing delay does not exceed the response time limit in the collaborative task information, the corresponding target collaborative node is recorded as a valid acceptance result. When the execution status is task rejection, response timeout, calculation failure or fusion failure, the corresponding target collaborative node is recorded as an invalid acceptance result, and the invalid acceptance result is used to reduce the basic benefit item of the target collaborative node in the task-constrained UCB node selection strategy. The model update sample is composed of valid acceptance results, invalid acceptance results, duplicate candidate data labels, abnormal state summaries, and final collaborative processing result summaries. The valid acceptance results and invalid acceptance results are used as supervision labels to update the parameters of the state scoring output layer of the improved ModernTCN edge state evaluation model. When the main processing result does not contain the abnormal type in the abnormal state summary under the auxiliary retention state, the corresponding abnormal type will be added to the set of mandatory review types of the abnormal retention rule. When the main processing result contains the abnormal type in the abnormal state summary, a preset sorting bonus item will be added to the subsequent sorting of the same measurement point type and coverage area corresponding to the current main processing node. The processing latency and communication overhead corresponding to the effective acceptance results are included in the historical average collaboration benefit, the invalid acceptance results are included in the negative collaboration record, and the number of times the candidate node is selected, the historical average collaboration benefit, and the feedback of the most recent collaboration execution are updated. The updated and improved ModernTCN edge state assessment model, master processing node determination rules, anomaly retention rules, and task-constrained UCB node selection strategy will be used for collaborative processing of IoT data in the next time slice.

Images