Resin grinding wheel safety early warning system based on edge computing

Abstract

The present application relates to a kind of resin grinding wheel safety early warning system based on edge computing, the system includes: edge collection node, message construction module, priority remapping module and path selection and sending module.The system is by time window cutting to operating data and constructs layered encapsulation message, in combination with abnormal score and data integrity determines hierarchical identification;Based on network operation index, construct network state vector, and accordingly dynamically remap message priority and queue allocation;Further in combination with candidate path historical statistics adaptive path selection is carried out, and the efficient transmission of layered message is realized.At the same time, to be checked abnormality is introduced to improve the reliability of judgment, and in the case of link instability, high-risk alarm is executed multiple-path redundant transmission, so as to improve the accuracy of early warning and transmission reliability.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and specifically to a resin grinding wheel safety early warning system based on edge computing. Background Technology

[0002] Resin-bonded grinding wheels are widely used in metal processing and precision grinding. However, under high-speed rotation and complex operating conditions, they are susceptible to material fatigue, localized defect propagation, clamping deviations, and load fluctuations, posing a risk of breakage or abnormal wear. To reduce the probability of accidents, current technologies typically collect spindle power, vibration signals, and machining status information from the equipment side, combining this with threshold judgments or simple statistical analysis to achieve anomaly warnings. However, in practical applications, the dangerous signs of resin-bonded grinding wheels often exhibit complex patterns of multi-dimensional feature coupling, short-term evolution, and cross-time window correlations. A single indicator or isolated window cannot accurately reflect potential risks, easily leading to false alarms or missed alarms.

[0003] With the development of the Industrial Internet, some solutions have introduced edge computing architectures, pushing data processing capabilities down to the device side or near-end nodes to reduce latency and improve real-time performance. However, existing edge-side early warning methods mostly focus on local anomaly detection, lacking the utilization of multi-node collaborative information. Especially when multiple data acquisition nodes simultaneously exhibit weak anomaly signals, it is difficult to perform timely and effective fusion and judgment. In addition, existing solutions generally adopt fixed priority or static queue mechanisms at the network transmission layer, failing to fully consider the impact of network state fluctuations on message transmission. In the case of network congestion or unstable links, a large amount of low-value raw data consumes bandwidth, which may lead to delays or even loss of critical alarm information transmission.

[0004] Furthermore, in existing technologies, the confirmation of abnormal results typically relies on a single judgment or simple continuous threshold judgment, lacking a dynamic supplementary verification mechanism for uncertain abnormal states, resulting in insufficient reliability of early warning results. Simultaneously, during alarm message transmission, there is a lack of a sending strategy based on dynamic path quality adjustment, and a lack of redundancy protection mechanisms for critical alarms, making it difficult to guarantee the reliable delivery of high-risk information in complex network environments.

[0005] Therefore, how to achieve hierarchical data representation, dynamic priority scheduling, network status awareness transmission, and multi-node collaborative verification for dangerous signs with time correlation and uncertainty during the operation of resin grinding wheels in an edge computing environment has become an urgent technical problem to be solved. Summary of the Invention

[0006] In response, this application provides a resin grinding wheel safety early warning system based on edge computing to at least partially solve the above-mentioned technical problems.

[0007] This application provides a safety early warning method for resin grinding wheels based on edge computing, comprising the following steps: The edge acquisition node receives the operating data of the resin grinding wheel equipment, divides the operating data into time windows, and generates window data segments; Based on the window data fragment, a hierarchical identifier is determined, and a hierarchical message header containing the hierarchical identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier, and window index is generated, as well as a hierarchical encapsulated message body corresponding to the hierarchical identifier. Edge aggregation nodes collect network operation metrics, construct network state vectors, and determine network state labels; based on the level identifier, the confidence parameter corresponding to the data integrity identifier, and the network state label, message priorities are remapped, and layered encapsulated messages are allocated to the corresponding sending queues. Based on the network state vector, the remapped message priority, and the historical statistics of the candidate path, a sending path is selected, and the layered encapsulated message is sent. Among them, for the dangerous symptom level message to be verified, a supplementary verification request is initiated according to the supplementary verification conditions and the level is updated according to the supplementary verification return message. For the confirmed high-risk alarm level message, multi-path redundant sending is performed in the link unstable state.

[0008] In one possible embodiment, the hierarchical identifier includes a normal raw data level, a feature summary level, a danger symptom to be verified level, and a confirmed high-risk alarm level; wherein, the normal raw data level corresponds to the raw window data or raw sampling fragment, the feature summary level corresponds to the statistical features extracted from the window data fragment, the danger symptom to be verified level corresponds to the message that the abnormal score is higher than the first threshold but does not meet the high-risk confirmation condition, and the confirmed high-risk alarm level corresponds to the message that the continuous window confirmation results meet the confirmation ratio threshold.

[0009] In one possible embodiment, the anomaly score is output by a temporal convolutional network model, which includes an input layer, multiple temporal convolutional blocks, a global average pooling layer, a fully connected layer, and a score output layer connected in sequence. The input layer receives a temporal feature matrix composed of standardized feature vectors corresponding to multiple consecutive windows, and the score output layer outputs the anomaly score of the current window to be evaluated. The standardized feature vector is composed of at least two of the following: spindle power features, vibration features, feed error features, and state position features in the window data segment.

[0010] In one possible embodiment, the confidence parameter is determined comprehensively based on the current window data integrity, the local packet loss rate of the edge acquisition node in the most recent time period, the timestamp continuity, and the sampling stability; in the layered encapsulated message body, the danger symptom level message to be verified includes an anomaly score, level determination basis, the confidence parameter, and supplementary evidence request identifier, and the confirmed high-risk alarm level message includes alarm level, suggested action, evidence chain summary, traceability index, and redundant transmission identifier.

[0011] In one possible embodiment, the network operation metrics include at least four of the following: queue backlog length, packet loss rate, average round-trip time, latency jitter, retransmission count, and link reachability. The edge aggregation node normalizes the network operation metrics, constructs a network state vector, and classifies the current network state into a stable state, a lightly congested state, a heavily congested state, and a link unstable state based on the network state vector. The switching of the network state label adopts a switching rule that satisfies the state conditions for multiple consecutive detection cycles.

[0012] In one possible embodiment, the message priority remapping includes: determining a basic priority based on a hierarchy identifier, adjusting the basic priority based on a confidence parameter to obtain an original priority, and remapping the original priority based on a network state label; maintaining the original priority in a stable state, performing downsampling or merging transmission on ordinary raw data hierarchy messages in a lightly congested state, suspending the transmission of ordinary raw data hierarchy main data in a heavily congested state, and adding redundant transmission identifiers to confirmed high-risk alarm hierarchy messages in an unstable link state.

[0013] In one possible embodiment, the step of allocating the layered encapsulated message to the corresponding sending queue includes: mapping the layered encapsulated message to a high-priority queue, a medium-priority queue, or a low-priority queue based on the remapped message priority and the queue boundary threshold; and retrieving messages from each sending queue and sending them according to a weighted round-robin scheduling rule, wherein the service weight of the high-priority queue is higher than the service weight of the medium-priority queue, and the service weight of the medium-priority queue is higher than the service weight of the low-priority queue.

[0014] In one possible embodiment, the selection of the sending path adopts a contextual multi-armed gambling machine model; the input of the contextual multi-armed gambling machine model includes a network state vector, a hierarchy identifier, a remapped message priority, a message timeliness parameter, a message length, and historical statistics of candidate paths, and the output is the path identifier in the candidate paths; the historical statistics of candidate paths include the average successful sending rate, average latency, and latency fluctuation of the candidate paths in the recent window; the model updates the reward parameters of the corresponding candidate paths according to the acknowledgment receipts and actual latency after message sending.

[0015] In one possible embodiment, initiating a supplementary verification request based on the supplementary verification conditions and updating the level based on the supplementary verification return message includes: When the anomaly score of the danger symptom level message to be verified is between the first threshold and the second threshold and the confidence parameter is higher than the supplementary verification confidence threshold, or when multiple edge acquisition nodes generate danger symptom level messages to be verified within a similar time window, the edge aggregation node generates a supplementary verification request message; the edge acquisition node receiving the supplementary verification request extracts the minimum necessary feature summary corresponding to the target time window from its local cache and returns it; the edge aggregation node associates and merges the supplementary verification return message with the original danger symptom level message to be verified, and maintains the original level or updates it to the confirmed high-risk alarm level based on the fusion result.

[0016] In another aspect, this application also provides a resin grinding wheel safety early warning system based on edge computing, comprising: Edge acquisition nodes are used to receive operating data from resin wheel grinding equipment and to divide the operating data into time windows to generate window data segments; The message construction module is used to determine the level identifier based on the window data fragment, and generate a hierarchical message header containing the level identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier and window index, as well as a layered encapsulated message body corresponding to the level identifier; The priority remapping module is used by edge aggregation nodes to collect network operation indicators, construct network state vectors and determine network state labels; based on the level identifier, the confidence parameter corresponding to the data integrity identifier and the network state label, the message priority is remapped and the layered encapsulated messages are allocated to the corresponding sending queues. The path selection and sending module is used to select a sending path based on the network state vector, the remapped message priority, and the historical statistics of the candidate path, and to send the layered encapsulated message; wherein, for the dangerous symptom level message to be verified, a supplementary verification request is initiated according to the supplementary verification conditions and the level is updated according to the supplementary verification return message, and multi-path redundant sending is performed on the confirmed high-risk alarm level message in the state of link instability.

[0017] This application constructs a layered encapsulation and dynamic scheduling mechanism under an edge computing architecture to structurally represent resin grinding wheel operation data according to risk levels. It then combines anomaly scoring and confidence parameters to achieve graded judgment, enabling early identification and verifiability of dangerous signs. Simultaneously, by introducing a network state-based priority remapping and sending queue scheduling strategy, it prioritizes the transmission of high-value alarm information under network congestion or link fluctuation conditions, reducing the bandwidth consumption of low-value data. Furthermore, a supplementary verification mechanism performs cross-node information fusion for anomalies to be verified, improving the accuracy and reliability of early warning results. A multi-path redundant sending strategy is adopted under unstable link conditions to improve the delivery success rate of high-risk alarms. In addition, a context-based multi-armed gambling machine model is used to adaptively select transmission paths, allowing the message transmission strategy to be dynamically optimized according to the network environment. Thus, timely, accurate, and reliable early warning of resin grinding wheel safety risks is achieved under complex industrial network conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of a resin grinding wheel safety early warning method based on edge computing, provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the layered encapsulation message body generation process provided in an embodiment of this disclosure.

[0021] Figure 3 This is a schematic diagram of the message priority remapping process provided in the embodiments of this disclosure.

[0022] Figure 4 This is a schematic diagram of the structure of a resin grinding wheel safety early warning system based on edge computing provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In one embodiment, this implementation applies to the edge computing network where the resin wheel grinding equipment is located. The network includes multiple edge acquisition nodes, one edge aggregation node, and a host control system. Each edge acquisition node is connected to the grinding equipment controller or an existing acquisition channel, receiving at least two types of operational data from spindle power data, vibration characteristic data, feed error data, and status status data. The edge aggregation node is communicatively connected to each edge acquisition node and to the host control system. The edge acquisition nodes are responsible for operational data access, time window segmentation, anomaly scoring, hierarchical determination, and message encapsulation; the edge aggregation node is responsible for network status monitoring, message priority remapping, sending queue allocation, path selection, supplementary scheduling, redundant sending, and traceability data coordination. The host control system receives alarm messages, sends traceability requests, and performs deduplication processing.

[0025] In actual deployment, both edge acquisition nodes and edge aggregation nodes pre-establish unified data dictionaries and parameter tables. The data dictionary includes at least data type identifiers, node identifiers, hierarchy identifiers, network status labels, queue identifiers, and alarm identifiers. The parameter table includes at least window length, window step size, first threshold, second threshold, confirmation ratio threshold, supplementary evidence confidence threshold, queue boundary threshold, message timeliness parameter, queue service weight, reward weight coefficient, and exploration rate parameter.

[0026] The following detailed description, in conjunction with specific embodiments, illustrates the implementation process of the edge computing-based resin grinding wheel safety early warning method described in this application. It should be noted that these embodiments are merely for explaining this application and not for limiting its scope of protection. Any conventional adjustments or substitutions made by those skilled in the art to the steps without departing from the concept of this application should be included within the scope of protection of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a schematic diagram of a resin grinding wheel safety early warning method based on edge computing, which includes the following method steps: In step S1, the edge acquisition node receives the operating data of the resin grinding wheel equipment, divides the operating data into time windows, and generates window data segments.

[0028] Specifically, after the grinding equipment starts operating, the edge acquisition nodes continuously receive operational data. Each piece of operational data is first formed into a basic data unit before entering the processing link. The basic data unit includes at least the acquisition time, source node identifier, data type identifier, data payload, and sequence number. The acquisition time is used for time sorting and cross-node alignment; the source node identifier is used to determine the message source; the data type identifier is used for subsequent feature extraction strategy selection; the data payload carries the original value or intermediate features; and the sequence number is used to detect packet loss and duplicate reception. To ensure that data from different sources can be processed uniformly within the same link, the edge acquisition nodes also perform dimensionality adjustment, abnormal empty packet removal, and missing sample checking on the basic data units. If the proportion of missing samples in a continuous sampling interval exceeds a preset upper limit, a data integrity decrease marker is added to the corresponding interval, and this marker is subsequently used in the confidence parameter calculation.

[0029] For the pre-processed runtime data, edge acquisition nodes perform segmentation according to time windows. The object of time window segmentation is a sequence of samples of the same data type on a continuous time axis. Each time window includes a start time, an end time, a window index, and a set of samples within the window. The window index is generated sequentially by the edge acquisition nodes and retained in subsequent hierarchical message headers for supplementary verification retrieval and retrospective transmission. The window length represents the sampling time range for each statistical and anomaly scoring operation, and the window step size represents the interval between the starting positions of two adjacent windows. The window length and window step size can be jointly determined by the equipment operating cycle, sampling frequency, and real-time warning requirements. For example, in one example, the window length corresponding to the spindle power data can be configured to be greater than the single load fluctuation cycle, and the window length corresponding to the vibration characteristic data can be configured to cover several high-frequency fluctuation segments.

[0030] For a window data segment, the edge acquisition node first extracts statistical features. These features include at least one or more of the following: mean, variance, peak difference, rate of change, fluctuation count, and abnormal state position count. For spindle power data, the mean power, power slope, and peak power difference can be extracted; for vibration feature data, short-time energy, kurtosis, kurtosis, and fluctuation count can be extracted; for feed error data, the mean error, mean absolute error, and error dispersion can be extracted; and for state position data, the state jump count and abnormal state duration can be extracted. The extracted statistical features form the window feature vector. To ensure consistent calculation of subsequent anomaly scores across different data volumes, the edge acquisition node also performs standardization on the window feature vector. Let the first... The mean of each feature in the stable operating samples is The standard deviation is The original value of this feature in the current window is The standardized eigenvalues ​​are denoted as:

[0031] in, To prevent extremely small positive numbers with a denominator of zero, a preset value is used during system initialization. The stable operation sample consists of a historical window set representing the equipment's normal production period, with no abnormal downtime records and no high-risk alarm records. Mean and standard deviation This data was obtained from statistics of the stable operating sample.

[0032] To further determine the level of danger signs, the edge acquisition node performs anomaly scoring on the standardized feature vectors. The anomaly score is provided by a temporal convolutional network model. This model is deployed locally on the edge acquisition node, receiving a temporal feature matrix formed by multiple consecutive windows and outputting the anomaly score for the current window to be evaluated. The input to this model is the standardized feature vectors of multiple adjacent windows arranged in chronological order. This setup is because danger signs in resin grinding wheels often exhibit short-term, continuous evolution, and a single window is insufficient to fully describe the temporal correlation of anomaly development.

[0033] In step S2, a hierarchical identifier is determined based on the window data fragment, and a hierarchical message header containing the hierarchical identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier, and window index is generated, as well as a hierarchical encapsulated message body corresponding to the hierarchical identifier.

[0034] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the layered encapsulation message body generation process provided in an embodiment of this disclosure. Figure 2 As shown, in step S201, the edge acquisition node determines the level identifier based on window data fragments, statistical features, and anomaly scores.

[0035] After obtaining continuous windows, the edge acquisition nodes form a temporal feature matrix in a sliding manner. Each row of this matrix corresponds to a time window, and each column corresponds to a standardized feature. Assume the number of features extracted in a single window is... The number of consecutive windows participating in anomaly scoring is Then the size of the time series feature matrix is .in, The time context length indicates the number of historical windows referenced in the current window's anomaly scoring; The feature dimension is determined by the statistical terms actually used from the spindle power characteristics, vibration characteristics, feed error characteristics, and state position characteristics. The time context length is determined by the duration of the danger symptom evolution and the inference delay that edge nodes can tolerate. If... If too small, time-related information is insufficient; if If the value is too large, edge inference latency and cache usage will increase. Therefore, in this implementation, the value is determined by offline sample testing during system initialization. The range of values ​​is determined and remains stable after deployment.

[0036] The temporal convolutional network model consists of an input layer, multiple temporal convolutional blocks, a global average pooling layer, a fully connected layer, and a scoring output layer, all connected sequentially. The input layer receives the temporal feature matrix and feeds it into the first temporal convolutional block. Each temporal convolutional block comprises a one-dimensional convolutional layer, a batch normalization layer, a linear rectified activation layer, and residual connections. The one-dimensional convolutional layer of the first temporal convolutional block performs convolution operations along the time axis, with the convolutional kernel sliding sequentially along the window to locally model the feature changes of adjacent time segments. The feature map output from this convolutional layer enters the batch normalization layer, which normalizes the mean and variance of the current batch output to reduce the differences in sample distribution between different batches. Subsequently, the linear rectified activation layer truncates negative value regions. Finally, the residual connections add the input of the temporal convolutional block to the convolution result, forming a block output to mitigate gradient decay during deep network training.

[0037] Between multiple temporal convolutional blocks, the receptive field of the convolution is expanded layer by layer through dilated convolutional structures. Specifically, the first temporal convolutional block uses the basic convolutional span, and subsequent temporal convolutional blocks expand the sampling interval according to a preset dilation factor. The dilation factor represents the sampling interval of the convolutional kernel on the time axis. The dilation factor is preset during model design, and its value can optionally be set in a layer-by-layer incrementing manner to cover the temporal dependence from the nearest neighbor window to the farthest window. After multiple temporal convolutional blocks are concatenated, the output of the last block enters a global average pooling layer. The global average pooling layer averages and aggregates the features at each position on the time axis to obtain a fixed-length aggregated feature vector. The aggregated feature vector is then input into a fully connected layer, which performs a linear mapping through a weight matrix and a bias vector. The scoring output layer receives the output of the fully connected layer and generates an anomaly score for the current window to be evaluated through a sigmoid activation function. The anomaly score is between 0 and 1.

[0038] The temporal convolutional network model takes as input a temporal feature matrix composed of standardized feature vectors corresponding to multiple consecutive windows, and outputs the anomaly score corresponding to the last window. The connection order between modules is: input layer to the first temporal convolutional block, then through each temporal convolutional block, into the global average pooling layer, then into the fully connected layer, and finally into the score output layer. The input of each temporal convolutional block is the output feature map of the previous layer, and the output is a new feature map after convolution, normalization, activation, and residual stacking.

[0039] In one embodiment, model training is performed offline. Training samples are derived from window sequences in historical operational data. Sample labeling is as follows: continuous window sequences are extracted as normal samples within intervals of stable equipment production, no abnormal shutdowns, and normal manual verification; continuous window sequences are extracted as abnormal samples within intervals where recorded hazardous events, protective actions are triggered, or high-risk situations are manually confirmed. During training, feature extraction and standardization are performed on each sample sequence in the same manner as in the online phase to ensure consistency between training and online inputs. The loss function is binary cross-entropy, and the optimization algorithm is adaptive gradient descent. During training, the temporal feature matrix is ​​input in batches, anomaly scores are obtained through forward propagation, and then compared with the labels to calculate the loss. The parameters of the convolutional layer, fully connected layer, and output layer are updated through backpropagation. After training, the model parameters are fixed into an inference file that can be loaded by edge nodes.

[0040] Edge acquisition nodes determine their level identifiers based on window data fragments, statistical features, and anomaly scores. These level identifiers include: ordinary raw data level, feature summary level, pending verification danger symptom level, and confirmed high-risk alarm level. The ordinary raw data level corresponds to the original window data or original sampling fragments; the feature summary level corresponds to the statistical features extracted from the window data fragments; the pending verification danger symptom level corresponds to messages where the anomaly score is higher than the first threshold but does not meet the high-risk confirmation criteria; and the confirmed high-risk alarm level corresponds to messages where the continuous window confirmation results meet the confirmation ratio threshold.

[0041] The first threshold is the trigger threshold for the level of danger symptom to be verified. This threshold is derived from the validation sample statistics and can be determined by the intersection of the score distributions of normal and abnormal samples on the validation set, or by on-site calibration. The second threshold is the high-risk confirmation score threshold, which must be greater than the first threshold. The confirmation ratio threshold is used to determine the proportion of windows in a continuous window that reach the second threshold. Let the total number of continuous decision windows be... The number of windows with anomaly scores not lower than the second threshold is . The confirmation ratio is:

[0042] When the confirmation rate is not lower than the confirmation rate threshold, the edge acquisition node marks the message corresponding to the current state as a confirmed high-risk alarm level. The values ​​of the first threshold, the second threshold, and the confirmation rate threshold can be determined through historical sample verification.

[0043] In step S202, a hierarchical message header is generated, which includes a hierarchy identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier, and window index.

[0044] After determining the hierarchical identifier, the edge acquisition node generates a hierarchical message header. The hierarchical message header includes the hierarchical identifier, generation time, source node identifier, message validity period identifier, data integrity identifier, and window index. The generation time is the moment the current node generates the message; the message validity period identifier is mapped from the hierarchical identifier, for example, higher-level messages correspond to shorter validity periods, and lower-level messages correspond to longer validity periods; the data integrity identifier reflects the missing sample status of the current window; and the window index is consistent with the aforementioned time window segmentation result.

[0045] In step S203, a layered encapsulated message body corresponding to the layer identifier is generated. The message body for the ordinary raw data layer includes the original window data fragment or the cache address of the original window data fragment; the message body for the feature summary layer includes a statistical feature summary; the message body for the danger symptom layer to be verified includes an anomaly score, layer determination criteria (e.g., a brief description of the anomaly score being higher than a first threshold but lower than a second threshold, or the changing trend of anomaly scores in a continuous window), confidence parameters, and a supplementary verification request identifier; the message body for the confirmed high-risk alarm layer includes the alarm level, suggested action, evidence chain summary, traceability index, and redundant transmission identifier. The evidence chain summary referred to here represents the minimum set of information retained to support alarm determination, including at least the key feature name, the direction of key feature change, the corresponding window range, and the source node identifier. The traceability index is used to locate locally cached data when subsequently initiating a raw data supplementation request.

[0046] The confidence level parameter is determined by a combination of factors, including the data integrity of the current window, the local packet loss rate of the edge acquisition nodes in the most recent time period, timestamp continuity, and sampling stability. Let the data integrity score be... The local packet loss rate is The timestamp continuity score is The sampling stability score is The confidence parameter can then be calculated using a weighted summation method:

[0047] in, , , , These are weighting coefficients, and the sum of the four factors must be 1. Each weighting coefficient represents the degree of influence of the four indicators on the confidence parameter. The weighting coefficients are determined by correlation analysis of historical data or by calibration using operational samples. If calibration is used, the correlation between the four indicators and the confidence level of alarms can be analyzed in samples with known alarm accuracy to determine the initial weight values. During operation, if adjustments are needed, the contribution of each indicator can be recalculated based on recent correct alarm samples during the maintenance cycle, and the weighting coefficients can be fine-tuned. To maintain stability, the magnitude of a single fine-tuning is limited by a preset upper limit.

[0048] For ordinary raw data level messages, raw window data fragments are not always sent immediately. Edge acquisition nodes typically write them to a local buffer first, recording the buffer address or buffer key in the message body. This buffer stores data in window index order, along with a retention period. If a traceability request or a request for supplementary certification is received subsequently, the corresponding data can be retrieved from the local buffer based on the window index. This approach reduces the real-time transmission bandwidth consumed by low-priority raw data.

[0049] After constructing the hierarchical message header and layered encapsulated message body, the edge acquisition node combines the two into a layered encapsulated message and sends it to the edge aggregation node. Thus, the operational data is transformed into a structured message that can participate in network scheduling and path selection.

[0050] In step S3, the edge aggregation node collects network operation indicators, constructs a network state vector, and determines a network state label. Based on the level identifier, the confidence parameter corresponding to the data integrity identifier, and the network state label, the message priority is remapped, and the layered encapsulated messages are allocated to the corresponding sending queues.

[0051] Edge aggregation nodes continuously receive layered and encapsulated messages from various edge acquisition nodes, while simultaneously monitoring the operational status of the current industrial network. Changes in the industrial network status directly impact message transmission latency, packet loss, and retransmission behavior. Therefore, edge aggregation nodes need to construct network state vectors and determine network state labels in real time to facilitate subsequent message priority remapping and sending queue allocation.

[0052] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the message priority remapping process provided in this embodiment of the disclosure. For example... Figure 3 As shown, in step S301, network state vectors are formed by collecting and normalizing network operation indicators.

[0053] Edge aggregation nodes collect network operation metrics according to the detection period. These metrics include at least four of the following: queue backlog length, packet loss rate, average round-trip time (RTT), latency jitter, retransmission count, and link reachability. Queue backlog length is given by the number of unprocessed messages or unsent bytes in the current send and receive queues; packet loss rate is calculated as the ratio of lost packets to sent packets within the detection period; average RTT is obtained from probe messages or service acknowledgment receipts; latency jitter is a statistical measure of the degree of change between adjacent latency samples; retransmission count is obtained from repeated transmission records within the detection period; and link reachability is given by the ratio of successful probes to the total number of probes.

[0054] Different network performance metrics have different dimensions, and edge aggregation nodes perform normalization processing before constructing the network state vector. For a given network performance metric... Let the lower bound of the historical baseline be The upper limit of the historical baseline is The normalization result can then be expressed as:

[0055] in, To prevent extremely small positive numbers with a denominator of zero, values ​​exceeding the upper bound of the historical baseline can be truncated to 1. The historical baseline is generated from initial network deployment stress test data and stable operation statistics. If network hardware or topology changes after long-term operation, the baseline can be re-collected and updated. and The normalized indices are arranged in a predetermined order to form the network state vector.

[0056] In step S302, the edge aggregation node determines the network state label based on the network state vector. The network state label includes stable state, lightly congested state, heavily congested state, and unstable link state. A stable state corresponds to low levels of queue backlog, packet loss, latency, and jitter; a lightly congested state corresponds to an increase in queue backlog or latency but not yet reaching a severe level; a heavily congested state corresponds to a significant increase in queue backlog, latency, or packet loss; and an unstable link state corresponds to significant fluctuations in latency, jitter, or reachability. To avoid frequent state transitions, the edge aggregation node adopts a switching rule where conditions are met within multiple consecutive detection cycles. That is, a candidate state must be met within a certain number of consecutive detection cycles before being updated to a new network state label. The detection cycle length and the number of consecutive detection cycles are preset during system initialization. This process allows the priority strategy to remain stable under short-term transient fluctuations.

[0057] In step S303, message priorities are remapped based on the layer identifier, the confidence parameter corresponding to the data integrity identifier, and the network status label. After receiving the layered encapsulated message, the edge aggregation node first determines the basic priority based on the layer identifier. The basic priority of the ordinary raw data layer is the lowest, followed by the feature summary layer, then the danger symptom layer, and finally the confirmed high-risk alarm layer. Then, the edge aggregation node adjusts the basic priority based on the confidence parameter to obtain the original priority. The higher the confidence parameter, the more stable the data quality and local link conditions of the current message, and thus a higher original priority can be obtained within the same layer.

[0058] Furthermore, edge aggregation nodes remap the original priorities based on network status labels. Under stable conditions, the original priorities are maintained. Under light congestion conditions, ordinary raw data level messages are downsampled or merged for transmission, feature summary level messages maintain medium priority, and message symptom-to-be-verified and confirmed high-risk alarm level messages are given higher queue contention order. Under heavy congestion conditions, transmission of ordinary raw data level main data is suspended, only cached indexes or necessary summaries are retained, feature summary level messages are transmitted using remaining bandwidth, and message symptom-to-be-verified and confirmed high-risk alarm level messages retain high priority. Under unstable link conditions, confirmed high-risk alarm level messages are appended with redundant transmission flags, and their reliability requirements in subsequent path selection are correspondingly increased.

[0059] Downsampling here means that instead of sending ordinary raw data level messages in every window, partial window indices or summaries are sent at predetermined sampling intervals. Merged sending means aggregating ordinary raw data level messages from multiple adjacent windows into a single summary message. Through this priority remapping, in the event of network congestion or unstable links, edge aggregation nodes can structurally control the transmission order of messages at different levels, preventing large amounts of low-value data from occupying the real-time transmission opportunities of high-value messages.

[0060] In step S304, the layered encapsulated messages are allocated to the corresponding sending queues. The edge aggregation node maps the layered encapsulated messages to high-priority, medium-priority, or low-priority queues based on the remapped message priority and the queue boundary threshold. The queue boundary threshold is preset during system initialization and can also be maintained and updated based on average waiting latency statistics during operation. The high-priority queue mainly stores confirmed high-risk alarm level messages and some high-priority pending-verification danger symptom level messages; the medium-priority queue mainly stores general pending-verification danger symptom level messages and key feature summary level messages; the low-priority queue mainly stores ordinary raw data level messages and other feature summary level messages.

[0061] Edge aggregation nodes use a weighted round-robin scheduling rule to retrieve messages from each sending queue. Let the service weights corresponding to the high-priority queue, medium-priority queue, and low-priority queue be respectively... , and If the service weight of a high-priority queue is greater than that of a medium-priority queue, and the service weight of a medium-priority queue is greater than that of a low-priority queue, then the service weight of the high-priority queue satisfies the condition that the service weight of the high-priority queue is greater than that of the low-priority queue. The service weight represents the relative number of times each queue is served within a scheduling period. Its initial value can be given by a preset strategy, such as an estimate based on the expected latency of alarm messages and the network's capacity. During operation, the edge aggregation node can adjust the weight based on the average waiting latency and timeout drop rate of each queue. If the average waiting latency of the high-priority queue is too high, the weight of the high-priority queue will be increased. If low-priority queues are experiencing excessive message backlog while high-priority queues are under low pressure, the priority level can be appropriately increased. Updates are performed during maintenance cycles, not with each packet release, to ensure a stable queue service rhythm.

[0062] In step S4, a sending path is selected based on the network state vector, the remapped message priority, and the historical statistics of the candidate path, and the layered encapsulated message is sent; wherein, for the dangerous symptom level message to be verified, a supplementary verification request is initiated according to the supplementary verification conditions and the level is updated according to the supplementary verification return message, and multi-path redundant sending is performed on the confirmed high-risk alarm level message in the state of link instability.

[0063] After a layered encapsulated message is selected to enter the sending phase, the edge aggregation node still needs to choose the current sending path from multiple candidate paths. The latency, packet loss, and jitter performance of different candidate paths change over time. To ensure that the path selection matches the current network state, message level, and message timeliness, this implementation adopts a context-based multi-armed gambler model.

[0064] The context-based multi-armed gambling machine model is deployed in edge aggregation nodes and includes a context encoding module, a reward estimation module, and an action selection module. The context encoding module receives the network state vector, hierarchical identifier, remapped message priority, message timeliness parameters, message length, and historical statistics of candidate paths, and concatenates them into a context feature vector. Historical statistics of candidate paths include the average successful transmission rate, average latency, and latency fluctuation of the candidate path within the recent window. The average successful transmission rate is determined by the ratio of the number of messages for which acknowledgments are successfully received to the total number of messages sent within the recent window; the average latency is determined by the average time difference between the delivery of a successful acknowledgment message and the delivery receipt; and the latency fluctuation is given by the dispersion of the latency sequence. The recent window length is defined by the path historical statistics period parameter, which is jointly determined by network dynamics and statistical stability requirements.

[0065] The reward estimation module maintains a set of reward parameters for each candidate path and estimates the expected reward of that path based on the current context feature vector. The action selection module performs path decisions based on the expected rewards of each candidate path. In this embodiment, the action selection module adopts... Greedy strategy. Exploration rate parameter. This represents the probability of selecting a non-optimal estimated path. The initial value of the exploration rate parameter can be preset and gradually decays with each run. The purpose of decay is to retain more path exploration opportunities in the early stages and to rely more on historical rewards for path selection in the later stages of the run. The exploration rate parameter ranges from zero to one. If the network topology or link quality is stable over a long period, the exploration rate parameter can be reduced; if the network environment changes frequently, the exploration rate parameter can be kept at a higher level to continuously discover better paths.

[0066] After sending a message and receiving an acknowledgment, the edge aggregation node updates the reward parameters of the corresponding candidate path based on the sending result. The reward parameters represent the sending revenue of a particular path in the current context. Let the path reward be:

[0067] in, This indicates whether a confirmation receipt has been received within the specified timeframe; a value of 1 indicates success, and a value of 0 indicates failure. This represents the normalized actual transmission delay; This represents the normalized path delay fluctuation. , , This refers to the reward weighting coefficient. The initial values ​​for the successful transmission reward weight, latency penalty weight, and jitter penalty weight are obtained from historical operational samples. If the security alert scenario places greater emphasis on delivery rate, then this is increased. If the current system has stronger constraints on arrival delay, then improve... If link jitter causes timing correlation errors, then improve... These weights are adjusted periodically during operation and are not reset immediately after a single message is sent. In this way, the context-based multi-armed gambling machine model can continuously update its path selection preferences based on historical sending results.

[0068] When an edge aggregation node receives a message indicating a potential danger level to be verified, it first determines whether the verification conditions are met. Verification conditions fall into two categories. One is a single-node message condition, where the anomaly score of the message indicates a danger level to be verified is between the first and second thresholds, and the confidence parameter is higher than the verification confidence threshold. The verification confidence threshold represents the minimum level of data credibility required to trigger verification, and its value can be determined by statistical analysis of historical successful verification samples. The other is a multi-node association condition, where multiple edge acquisition nodes generate messages indicating potential danger levels to be verified within a similar time window. The similar time window is defined by a time difference threshold, which is determined by statistical analysis of cross-node synchronization errors and operational propagation time.

[0069] For example, the sum of the maximum synchronization error across nodes and the maximum propagation time of the operating condition can be taken, or the distribution of the time difference of the same abnormal event occurring in different nodes can be statistically analyzed through historical data, and the 95th percentile can be taken as the threshold.

[0070] Once the supplementary verification conditions are met, the edge aggregation node generates a supplementary verification request message. This message includes the target time window range, the required feature types, the request source node identifier, and the supplementary verification request identifier. The edge acquisition node receiving the request retrieves the corresponding original window data segment from its local cache according to the target time window range, extracts the minimum necessary feature summary, and returns it. The minimum necessary feature summary represents a simplified set of features sufficient to support hierarchical update judgment without returning the complete original data. For example, when the message for verifying a danger symptom level is triggered by spindle power anomalies, the minimum necessary feature summary may include the power change rate, peak difference, and fluctuation count for the corresponding window; when the message is triggered by vibration feature anomalies, the minimum necessary feature summary may include necessary items from short-term energy changes, kurtosis, and short-term energy changes.

[0071] Upon receiving a supplementary verification return message, the edge aggregation node correlates and merges it with the original message at the danger symptom level to be verified. The correlation is based on the source node identifier, the target time window range, and the supplementary verification request identifier. After merging, the edge aggregation node calculates a comprehensive confirmation result. This comprehensive confirmation result can be obtained based on rule-based fusion: if the minimum necessary feature digest in the supplementary verification return message is consistent with the evidence chain digest in the original message in terms of key feature direction and time range, the confirmation level of the current message is increased; otherwise, the original level is maintained. If the comprehensive confirmation result meets the high-risk confirmation criteria, the edge aggregation node updates the original message at the danger symptom level to a confirmed high-risk alarm level message and re-enters it into the high-priority queue. Through this process, even with limited network bandwidth, the system can complete the level update using a relatively small amount of supplementary verification information, shortening the time required for danger confirmation.

[0072] For confirmed high-risk alarm messages, when the network status label indicates a link instability, the edge aggregation node performs multi-path redundancy transmission. This process first determines the primary path based on the path selection results of the contextual multi-armed gambling machine model, and then selects an alternative path from the remaining candidate paths. The selection of the alternative path can consider recent reachability, average latency, and path independence. In this implementation, path independence represents the degree of overlap between the candidate path and the primary path at network switching nodes, gateway channels, or congested locations; the lower the overlap, the higher the path independence.

[0073] To determine path independence, edge aggregation nodes can pre-acquire network topology information, record the gateways and switching nodes traversed by each candidate path, and calculate the number of common nodes or links between the candidate path and the main path. The fewer the common nodes or links, the higher the path independence. Alternatively, in the absence of topology information, path independence can be estimated by the correlation of historical congestion events. For example, if the historical latency and packet loss changes of two paths are highly correlated, they are considered to likely share congestion points, resulting in low path independence. When selecting alternative paths, paths with high path independence and satisfactory recent reachability and average latency should be prioritized.

[0074] The edge aggregation node then generates a primary replica and a redundant replica with the same alarm identifier, which are sent via the primary path and the backup path, respectively. The alarm identifier is uniformly assigned by the edge aggregation node when the message is generated, and together with the source node identifier and the time window range, it represents the same high-risk event.

[0075] To prevent the receiver from performing duplicate processing on duplicate copies, it performs deduplication based on a deduplication key. The deduplication key consists of an alarm identifier, a source node identifier, a time window range, and a sequence number. The receiver can write the deduplication key into a short-term cache table. When a new high-risk message copy is received, this short-term cache table is searched first. If a duplicate key exists, it is determined to be a duplicate copy, and only the earlier arriving copy or the copy with more complete fields is retained. If no duplicate key exists, it is received as a new message and written to the cache table. This process prevents redundant multi-path transmission from causing the upper-level control system to repeatedly trigger subsequent actions.

[0076] Optionally, this implementation also includes retrospective data transmission processing. When the upper-level control system needs to trace the original data after receiving a confirmed high-risk alarm level message, it can send a tracing request to the edge aggregation node. The edge aggregation node initiates an original data transmission request to the corresponding edge acquisition node based on the window index in the confirmed high-risk alarm level message. The edge acquisition node retrieves the corresponding original data segment from its local cache based on the window index and sends it to the edge aggregation node or the upper-level control system. It can be understood that the local cache has pre-stored the original data segment or its cache address during the aforementioned ordinary original data level processing. Retrospective data transmission is usually performed after network pressure decreases to avoid competing for bandwidth with the real-time transmission of high-risk alarms.

[0077] In some embodiments, edge aggregation nodes can also uniformly manage message timeouts. Message timeout parameters are mapped from hierarchical identifiers and given in the hierarchical message header. Edge aggregation nodes periodically check message timeouts within each sending queue. If a low-priority message has exceeded its validity period, only its index information is retained or the message body is discarded directly; if a message at a critical symptom level is approaching its timeout limit, its scheduling order is promoted or priority is given to triggering supplementary verification. Through message timeout management, stale messages can be prevented from remaining in the queue for extended periods, thus avoiding impacting subsequent message processing.

[0078] In some embodiments, edge acquisition nodes and edge aggregation nodes can also reduce cross-node window alignment errors through a clock synchronization mechanism. The clock synchronization mechanism can be implemented through network time synchronization or unified time synchronization by the controller. The synchronized timestamps are used for judging adjacent time windows and cross-node supplementary verification. If a continuous decrease in the timestamp continuity score of an edge acquisition node is detected, this score will be reflected in the confidence parameter, thereby indirectly reducing its message priority. In this way, time synchronization quality is consistently correlated with the message scheduling link.

[0079] Through the above implementation method, resin grinding wheel operation data is converted into layered encapsulated messages with hierarchical identifiers and structured message bodies. Edge aggregation nodes dynamically adjust message priorities, sending queues, and sending paths based on network conditions. In scenarios with pending verification of dangerous signs, hierarchical updates are completed through supplementary verification messages. In scenarios with unstable links, high-risk alarm transmission is ensured through multi-path redundant transmission. When necessary, original data is traced and retransmitted based on the window index. Thus, regardless of changes in edge computing network conditions, the generation, confirmation, transmission, and tracing processes of security warning messages maintain consistent data source relationships and clear processing.

[0080] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a resin grinding wheel safety early warning system 4 based on edge computing, provided in an embodiment of this application. Figure 4 As shown, the system includes: Edge acquisition node 401 is used to receive the operating data of the resin grinding wheel equipment, and to divide the operating data into time windows to generate window data segments; The message construction module 402 is used to determine the level identifier based on the window data fragment, and generate a hierarchical message header containing the level identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier and window index, as well as a layered encapsulated message body corresponding to the level identifier; The priority remapping module 403 is used to collect network operation indicators at the edge aggregation node, construct a network state vector and determine a network state label; based on the level identifier, the confidence parameter corresponding to the data integrity identifier and the network state label, the message priority is remapped and the layered encapsulated messages are allocated to the corresponding sending queues. The path selection and sending module 404 is used to select a sending path based on the network state vector, the remapped message priority, and the historical statistics of the candidate path, and send the layered encapsulated message; wherein, for the dangerous symptom level message to be verified, a supplementary verification request is initiated according to the supplementary verification conditions and the level is updated according to the supplementary verification return message, and multi-path redundant sending is performed on the confirmed high-risk alarm level message in the state of link instability.

[0081] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0082] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0083] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0084] The embodiments disclosed herein are preferred embodiments, but are not limited thereto. Those skilled in the art can readily grasp the spirit of the present invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A safety early warning method for resin grinding wheels based on edge computing, characterized in that, include: The edge acquisition node receives the operating data of the resin grinding wheel equipment, divides the operating data into time windows, and generates window data segments; Based on the window data fragment, a hierarchical identifier is determined, and a hierarchical message header containing the hierarchical identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier, and window index is generated, as well as a hierarchical encapsulated message body corresponding to the hierarchical identifier. Edge aggregation nodes collect network operation metrics, construct network state vectors, and determine network state labels; Based on the level identifier, the confidence parameter corresponding to the data integrity identifier, and the network status label, the message priority is remapped, and the layered encapsulated messages are allocated to the corresponding sending queues. Based on the network state vector, the remapped message priority, and the historical statistics of the candidate path, a sending path is selected, and the layered encapsulated message is sent. Among them, for the dangerous symptom level message to be verified, a supplementary verification request is initiated according to the supplementary verification conditions and the level is updated according to the supplementary verification return message. For the confirmed high-risk alarm level message, multi-path redundant sending is performed in the link unstable state.

2. The method according to claim 1, characterized in that, The hierarchical identifiers include ordinary raw data level, feature summary level, danger symptom to be verified level, and confirmed high-risk alarm level; wherein, ordinary raw data level corresponds to raw window data or raw sampling fragments, feature summary level corresponds to statistical features extracted from window data fragments, danger symptom to be verified level corresponds to messages with an abnormal score higher than the first threshold but not meeting the high-risk confirmation conditions, and confirmed high-risk alarm level corresponds to messages with continuous window confirmation results meeting the confirmation ratio threshold.

3. The method according to claim 2, characterized in that, The anomaly score is output by a temporal convolutional network model, which includes an input layer, multiple temporal convolutional blocks, a global average pooling layer, a fully connected layer, and a score output layer connected in sequence. The input layer receives a temporal feature matrix composed of standardized feature vectors corresponding to multiple consecutive windows, and the score output layer outputs the anomaly score of the current window to be evaluated. The standardized feature vector is composed of at least two of the following: spindle power features, vibration features, feed error features, and state position features in the window data segment.

4. The method according to claim 1, characterized in that, The confidence level parameter is determined comprehensively based on the data integrity of the current window, the local packet loss rate of the edge acquisition node in the most recent period, the timestamp continuity, and the sampling stability. In the layered encapsulated message body, the danger symptom level message to be verified includes an anomaly score, level determination basis, the confidence level parameter, and supplementary evidence request identifier. The confirmed high-risk alarm level message includes alarm level, suggested action, evidence chain summary, traceability index, and redundant transmission identifier.

5. The method according to claim 1, characterized in that, The network operation metrics include at least four of the following: queue backlog length, packet loss rate, average round-trip time, latency jitter, number of retransmissions, and link reachability. The edge aggregation node normalizes the network operation indicators, constructs a network state vector, and classifies the current network state into stable state, lightly congested state, heavily congested state, and unstable link state based on the network state vector; the switching of the network state label adopts a switching rule that meets the state conditions for multiple consecutive detection cycles.

6. The method according to claim 5, characterized in that, The message priority remapping process includes: determining a basic priority based on a hierarchy identifier, adjusting the basic priority based on a confidence parameter to obtain an original priority, and remapping the original priority based on a network status label; maintaining the original priority in a stable state, performing downsampling or merging transmission on ordinary raw data hierarchy messages in a lightly congested state, suspending the transmission of ordinary raw data hierarchy main data in a heavily congested state, and adding redundant transmission identifiers to confirmed high-risk alarm hierarchy messages in an unstable link state.

7. The method according to claim 6, characterized in that, The step of allocating the layered encapsulated messages to the corresponding sending queues includes: mapping the layered encapsulated messages to high-priority queues, medium-priority queues, or low-priority queues based on the remapped message priority and queue boundary thresholds; and retrieving messages from each sending queue and sending them according to a weighted round-robin scheduling rule, wherein the service weight of the high-priority queue is higher than that of the medium-priority queue, and the service weight of the medium-priority queue is higher than that of the low-priority queue.

8. The method according to claim 1, characterized in that, The selection of the sending path adopts a contextual multi-armed gambling machine model; the input of the contextual multi-armed gambling machine model includes network state vector, hierarchy identifier, remapped message priority, message timeliness parameter, message length, and historical statistics of candidate paths, and the output is the path identifier in the candidate paths; the historical statistics of candidate paths include the average successful sending rate, average latency, and latency fluctuation of candidate paths in the recent window; the model updates the reward parameters of the corresponding candidate paths according to the acknowledgment receipt and actual latency after message sending.

9. The method according to claim 1, characterized in that, The step of initiating a supplementary certification request based on the supplementary certification conditions and updating the level based on the supplementary certification return message includes: When the anomaly score of the danger symptom level message to be verified is between the first threshold and the second threshold and the confidence parameter is higher than the supplementary verification confidence threshold, or when multiple edge acquisition nodes generate danger symptom level messages to be verified within a similar time window, the edge aggregation node generates a supplementary verification request message; the edge acquisition node receiving the supplementary verification request extracts the minimum necessary feature summary corresponding to the target time window from its local cache and returns it; the edge aggregation node associates and merges the supplementary verification return message with the original danger symptom level message to be verified, and maintains the original level or updates it to the confirmed high-risk alarm level based on the fusion result.

10. A resin grinding wheel safety early warning system based on edge computing, characterized in that, include: Edge acquisition nodes are used to receive operating data from resin wheel grinding equipment and to divide the operating data into time windows to generate window data segments; The message construction module is used to determine the level identifier based on the window data fragment, and generate a hierarchical message header containing the level identifier, generation time, source node identifier, message timeliness identifier, data integrity identifier and window index, as well as a layered encapsulated message body corresponding to the level identifier; The priority remapping module is used by edge aggregation nodes to collect network operation indicators, construct network state vectors, and determine network state labels. Based on the level identifier, the confidence parameter corresponding to the data integrity identifier, and the network status label, the message priority is remapped, and the layered encapsulated messages are allocated to the corresponding sending queues. The path selection and sending module is used to select a sending path based on the network state vector, the remapped message priority, and the historical statistics of the candidate path, and to send the layered encapsulated message; wherein, for the dangerous symptom level message to be verified, a supplementary verification request is initiated according to the supplementary verification conditions and the level is updated according to the supplementary verification return message, and multi-path redundant sending is performed on the confirmed high-risk alarm level message in the state of link instability.

Images