Fault early warning method and system based on edge computing and artificial intelligence

Abstract

This invention provides a fault early warning method and system based on edge computing and artificial intelligence. It captures multi-source continuous feature streams of device operation through edge computing nodes, performs feature temporal anchoring and cross-dimensional correlation processing, and generates anchored feature sequences and other data. These are then input into an artificial intelligence fault trajectory model for multi-scale feature trajectory mining and anomaly tracing, yielding multi-scale anomaly feature tracing results and a set of feature anomaly anchor points. Based on the feature anomaly anchor point set and cross-dimensional correlation graph, the relevant rules and strategies of the edge computing nodes are dynamically calibrated, generating calibration anchoring parameters and optimized correlation strategies. The multi-source continuous feature streams are reprocessed by the edge computing nodes using the calibration parameters and strategies to obtain calibrated relevant data. Finally, the multi-scale anomaly feature tracing results, calibration anchored feature sequences, and optimized cross-dimensional correlation graph are combined to generate fault early warning commands and transmit them to the device management platform. This invention can improve the accuracy of device fault early warning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a fault early warning method and system based on edge computing and artificial intelligence. Background Technology

[0002] In the field of industrial production and equipment operation management, timely early warning of equipment failures is crucial for ensuring production safety, improving production efficiency, and reducing maintenance costs. Traditional equipment failure early warning methods mainly rely on regular manual inspections and simple threshold judgments based on a single data source. Manual inspections are not only costly in terms of manpower and resources, but also difficult to monitor in real time and cannot promptly detect potential equipment failures. Threshold judgment methods based on a single data source only consider one aspect of the equipment's operating data, ignoring the complex correlations and temporal changes between multiple data sources during equipment operation, resulting in poor accuracy and timeliness of failure warnings, and a high likelihood of false alarms and missed alarms. With the rapid development of IoT, edge computing, and AI technologies, existing failure early warning solutions based on these technologies lack effective feature temporal anchoring and cross-dimensional correlation processing mechanisms when handling multi-source continuous feature streams, making it difficult to accurately mine multi-scale feature trajectories and anomaly tracing information during equipment operation. Furthermore, they are insufficient in dynamically calibrating the parameters and strategies of edge computing nodes, failing to adjust them in a timely manner according to actual operating conditions, thus affecting the reliability and adaptability of failure early warnings. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a fault early warning method based on edge computing and artificial intelligence, the method comprising: By capturing multi-source continuous feature streams running on devices through edge computing nodes, feature temporal anchoring and cross-dimensional association processing are performed on the multi-source continuous feature streams to generate anchored feature sequences, feature temporal anchor points and cross-dimensional association maps. The anchored feature sequence, feature time-series anchor points and cross-dimensional correlation graph are input into the artificial intelligence fault trajectory model to perform multi-scale feature trajectory mining and anomaly tracing, and obtain multi-scale anomaly feature tracing results and feature anomaly anchor point set. Based on the feature anomaly anchor point set and cross-dimensional correlation map, the feature temporal anchoring rules and correlation mapping strategies of edge computing nodes are dynamically calibrated to generate calibration anchoring parameters and optimized correlation strategies. By applying calibration anchoring parameters and optimization association strategies to edge computing nodes, multi-source continuous feature streams are reprocessed to obtain calibration anchoring feature sequences, calibration feature time-series anchor points, and optimized cross-dimensional association maps. By combining the results of multi-scale anomaly feature tracing, calibrating and anchoring feature sequences, and optimizing cross-dimensional correlation maps, fault warning commands are generated and transmitted to the equipment management platform.

[0004] In another aspect, embodiments of the present invention also provide a fault early warning system based on edge computing and artificial intelligence, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0005] Based on the above, this embodiment of the invention captures multi-source continuous feature streams of device operation through edge computing nodes and performs feature temporal anchoring and cross-dimensional correlation processing. This enables comprehensive and accurate extraction of key feature information during device operation, generating anchored feature sequences, feature temporal anchor points, and cross-dimensional correlation maps. Inputting this data into an artificial intelligence fault trajectory model for multi-scale feature trajectory mining and anomaly tracing allows for precise identification of abnormal situations during device operation, yielding multi-scale anomaly feature tracing results and a set of feature anomaly anchor points, effectively improving the accuracy of fault warnings. Dynamically calibrating the feature temporal anchoring rules and correlation mapping strategies of edge computing nodes based on the feature anomaly anchor point set and cross-dimensional correlation maps allows edge computing nodes to better adapt to changes in actual device operation, generating calibration anchoring parameters and optimized correlation strategies, enhancing the system's adaptability. By reprocessing the multi-source continuous feature streams using calibration parameters and strategies, calibration anchored feature sequences, calibration feature temporal anchor points, and optimized cross-dimensional correlation maps are obtained, further optimizing the data processing process. Finally, by combining the results of multi-scale anomaly feature tracing, calibrated anchor feature sequences, and optimized cross-dimensional correlation maps, fault warning instructions are generated and transmitted to the equipment management platform, enabling timely warning of equipment faults. This helps improve the stability and safety of equipment operation, reduce losses caused by equipment faults, and improve the reliability, accuracy, and adaptability of equipment fault warnings overall. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the execution flow of the fault early warning method based on edge computing and artificial intelligence provided in the embodiments of the present invention.

[0007] Figure 2 This is a schematic diagram of exemplary hardware and software components of a fault early warning system based on edge computing and artificial intelligence provided in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating a fault early warning method based on edge computing and artificial intelligence provided in an embodiment of the present invention. The following is a detailed description of the fault early warning method based on edge computing and artificial intelligence.

[0009] Step S110: Capture the multi-source continuous feature stream running on the device through the edge computing node, perform feature temporal anchoring and cross-dimensional association processing on the multi-source continuous feature stream, and generate anchored feature sequences, feature temporal anchor points and cross-dimensional association maps.

[0010] In this embodiment, a large centrifugal water pump in an industrial production scenario is used as the target monitoring object. This water pump is the core power unit of the production line's water circulation system, and its operating status directly affects the stable operation of the entire production line. Edge computing nodes are deployed in the water pump room and connected to the water pump's control system and sensor network through an industrial bus interface to achieve real-time acquisition and processing of water pump operating data.

[0011] Step S111: Through the multi-source feature acquisition module of the edge computing node, connect to the device's operation status output port, load monitoring port and environmental sensing port respectively, synchronously receive the device operation raw data, load fluctuation raw data and environmental related raw data continuously output by the port, and perform timestamp alignment processing on the device operation raw data, load fluctuation raw data and environmental related raw data to keep the time dimension of the device operation raw data, load fluctuation raw data and environmental related raw data consistent.

[0012] In this embodiment, the multi-source feature acquisition module of the edge computing node includes three independent acquisition channels, corresponding to the device's operating status output port, load monitoring port, and environmental sensing port, respectively. The operating status output port is connected to the PLC data interface of the water pump control system, continuously outputting raw operating data such as the water pump's spindle speed, motor stator temperature, inlet and outlet pressure, and flow rate. The load monitoring port is connected to the smart meter in the motor distribution cabinet, continuously outputting raw load fluctuation data such as the motor's active power, reactive power, and line current. The environmental sensing port is connected to temperature and humidity sensors and vibration sensors deployed in the computer room, continuously outputting raw environmental data such as the computer room's ambient temperature, relative humidity, and water pump casing vibration acceleration. Since the sampling frequencies of different types of sensors differ—for example, the sampling frequency for operating status data is multiple times per unit time, load data is once per unit time, and environmental data is twice per unit time—timestamp alignment processing is required. Specifically, using the timestamp of the operating status data as a reference, linear interpolation is used to resample the load data and environmental data, ensuring that the three are synchronized in the time dimension and that each point in time has corresponding data of the three types. The timestamp is generated based on the high-precision clock module built into the edge computing node, ensuring that the time accuracy reaches the microsecond level.

[0013] Step S112: Perform feature extraction processing on the raw equipment operation data, extracting equipment operation status features according to the preset extraction standards of the core equipment operation indicators; perform feature extraction processing on the raw load fluctuation data, extracting load fluctuation features according to the preset extraction standards of load changing over time; perform feature extraction processing on the raw environmental correlation data, extracting environmental correlation features according to the preset extraction standards of the impact of environmental factors on equipment operation, and integrate the equipment operation status features, load fluctuation features, and environmental correlation features in chronological order to form a multi-source continuous feature stream.

[0014] In this embodiment, the preset extraction criteria for the core operating indicators of the equipment are determined by the operation manual provided by the pump equipment manufacturer. These criteria include extraction rules for features such as the stability index of the spindle speed, the rate of change of the motor stator temperature, and the fluctuation range of the inlet and outlet pressure difference. For the spindle speed data in the raw equipment operating data, the standard deviation of the speed within each window is calculated using the sliding window method as a stability feature; the temperature difference between adjacent time points is calculated for the motor stator temperature data as a temperature change rate feature; and the pressure difference is calculated from the difference between the inlet and outlet pressure data to obtain the pressure difference feature, with the maximum and minimum values ​​of the pressure difference extracted as fluctuation range features. These features together constitute the equipment operating state feature vector, with each element of the vector corresponding to a specific feature value. For the raw load fluctuation data, features such as the trend of active power change (calculated using first-order difference), the ratio of reactive power to active power, and the harmonic distortion rate of the line current are extracted according to the preset extraction criteria for load changes over time, forming a load fluctuation feature vector. For the raw environmental data, features such as the moving average of ambient temperature, the range of relative humidity change, and the effective value of vibration acceleration are extracted, forming an environmental correlation feature vector. The three feature vectors are arranged in chronological order to form a multi-source continuous feature stream, where each time step corresponds to a high-dimensional feature vector composed of three sub-vectors.

[0015] Step S113: Set time-series segmentation rules according to the continuous time process of equipment operation, and uniformly segment the multi-source continuous feature stream with a fixed time span as the segmentation benchmark to obtain multiple continuous time-series feature segments; extract the core feature units that represent the core state of equipment operation in the corresponding time-series feature segments, and add a unique time-dimensional associated position mark to the core feature units based on the relative position information of the core feature units in the corresponding time-series feature segments to generate the feature time-series anchor points corresponding to the core feature units.

[0016] In this embodiment, based on the operating cycle characteristics of the water pump equipment, the time-series segmentation rule is set as follows: the time span of a complete working cycle of the equipment is used as a fixed time span to uniformly segment the multi-source continuous characteristic flow. For example, if the working cycle of the water pump is a certain fixed duration, then the length of each time-series characteristic segment is equal to that duration. Through this segmentation method, each time-series characteristic segment can fully reflect the operating characteristics of the water pump within a working cycle. For each time-series characteristic segment, a feature importance evaluation algorithm (such as feature importance scoring based on random forest) is used to select several features that contribute the most to equipment fault early warning from equipment operating status features, load fluctuation features, and environmental correlation features as core feature units. For example, the four features selected as core feature units are spindle speed stability, motor stator temperature change rate, active power change trend, and ambient temperature moving average. The relative position information of the core feature unit in the time-series characteristic segment is represented by the ratio of its time index within the segment to the segment length. For example, if a core feature unit appears at the Nth time step within the segment, and the segment length is M time steps, then its relative position is N / M. The relative position is combined with the start timestamp of the time-series feature segment to generate a unique time-dimensional associated position marker, such as the string form of "start timestamp + relative position". This marker is the feature time-series anchor point, used to uniquely identify the position of the core feature unit in the time dimension.

[0017] Step S114: Based on the feature attributes of equipment operating status characteristics, load fluctuation characteristics, and environmental correlation characteristics, select the numerical change trend, fluctuation frequency, and correlation influence degree of the features, and calculate the feature attribute similarity between equipment operating status characteristics and load fluctuation characteristics, equipment operating status characteristics and environmental correlation characteristics, and load fluctuation characteristics and environmental correlation characteristics.

[0018] In this embodiment, the selection of feature attributes is based on the analysis of the operating mechanism of the water pump equipment. The numerical change trend is characterized by the first derivative sequence of the feature data, reflecting the direction and rate of change of the feature over time; the fluctuation frequency is characterized by the main frequency components obtained after Fourier transforming the feature data, reflecting the periodic change characteristics of the feature; the degree of correlation is characterized by the cross-correlation coefficient between two features, reflecting the degree of linear correlation between the features. For the spindle speed stability feature in the equipment operating status features and the active power change trend feature in the load fluctuation features, when calculating the similarity of their numerical change trends, the cosine similarity of their first derivative sequences is calculated; when calculating the fluctuation frequency similarity, the overlap of the main frequency component sets after Fourier transform is compared; when calculating the degree of correlation similarity, the cross-correlation coefficient is directly used. Through similarity calculations in these three dimensions, the similarity between the two features is comprehensively evaluated.

[0019] Step S1141: Select the numerical change trend, fluctuation frequency, peak occurrence timing, and correlation effect delay duration of the features, and calculate the similarity of the equipment operating status features and load fluctuation features, the equipment operating status features and environmental correlation features, and the load fluctuation features and environmental correlation features in the feature attribute dimension.

[0020] In this embodiment, two additional feature attributes, peak occurrence timing and correlation effect delay duration, are added to step S114. Peak occurrence timing refers to the time point when the feature data reaches a local maximum within a time-series feature segment. Similarity is calculated by comparing the relative positions of the peak occurrence times of the two features within the segment. Correlation effect delay duration refers to the time interval between a significant change in one feature and a subsequent change in another feature. It is determined by the peak position difference of the cross-correlation function. For example, when calculating the similarity of the peak occurrence timing between the motor stator temperature change rate in the equipment operating status feature and the ambient temperature moving average in the environmental correlation feature, all peak points of both are first identified within the time-series feature segment. Then, the average value of the relative position difference between the corresponding peak points is calculated and normalized to the [0,1] interval as the similarity value. When calculating the similarity of the correlation effect delay duration, the time interval during which the change in the ambient temperature moving average precedes the change in the motor stator temperature change rate is identified through sliding cross-correlation analysis. This interval is compared with a preset typical delay duration to obtain the similarity value.

[0021] Step S1142: Use a weighted summation algorithm to comprehensively calculate the similarity on the feature attribute dimension to obtain the overall feature attribute similarity between equipment operating status features and load fluctuation features, equipment operating status features and environmental correlation features, and load fluctuation features and environmental correlation features; set a similarity threshold, classify feature combinations with overall feature attribute similarity higher than the threshold as direct correlation types, and classify those with similarity lower than the threshold as indirect correlation types, and assign a preset high value range of correlation strength value to the direct correlation type, and assign a preset low value range of correlation strength value to the indirect correlation type.

[0022] In this embodiment, for each feature combination (such as the spindle speed stability in the equipment operating status feature and the active power change trend in the load fluctuation feature), the similarity values ​​of four dimensions—numerical change trend, fluctuation frequency, peak occurrence timing, and correlation effect delay duration—have been obtained in step S1141. When using a weighted summation algorithm for synthesis, the weights of each dimension are determined based on expert experience and historical data training results. For example, the weight of numerical change trend is 0.3, the weight of fluctuation frequency is 0.25, the weight of peak occurrence timing is 0.2, and the weight of correlation effect delay duration is 0.25. The similarity values ​​of each dimension are multiplied by their corresponding weights and then summed to obtain the overall feature attribute similarity. A similarity threshold of 0.6 is set. When the overall similarity value is greater than 0.6, it is determined to be a direct association type, and the association strength value range is assigned to 0.7-0.9; when the overall similarity value is less than or equal to 0.6, it is determined to be an indirect association type, and the association strength value range is assigned to 0.3-0.5. For example, if the similarity of the four dimensions of a certain feature combination is 0.8, 0.7, 0.6, and 0.75, and the weights are 0.3, 0.25, 0.2, and 0.25, then the overall similarity is 0.8×0.3+0.7×0.25+0.6×0.2+0.75×0.25=0.24+0.175+0.12+0.1875=0.7225, which is greater than the threshold of 0.6. Therefore, it is determined to be a direct association type, and the association strength can be selected from a specific value in the range of 0.7-0.9, such as 0.8.

[0023] Step S1143: Establish bidirectional pointing mappings from equipment operating status characteristics to load fluctuation characteristics and from load fluctuation characteristics to equipment operating status characteristics, respectively, and mark the feature transmission direction of the pointing mappings.

[0024] In this embodiment, for feature combinations determined to be directly related, a bidirectional mapping is established. Taking the inlet and outlet pressure difference in the equipment operating status feature and the line current in the load fluctuation feature as an example, on the one hand, changes in the inlet and outlet pressure difference will lead to changes in the pump load, which in turn will cause changes in the line current. Therefore, a mapping from the inlet and outlet pressure difference to the line current is established. On the other hand, abnormal fluctuations in the line current may also reflect faults in the impeller or bearings inside the pump, thus affecting the inlet and outlet pressure difference. Therefore, a mapping from the line current to the inlet and outlet pressure difference is established. The feature transmission direction is marked in the mapping relationship. For example, the bidirectional mapping direction is represented by "inlet and outlet pressure difference → line current" and "line current → inlet and outlet pressure difference".

[0025] Step S1144: In the established bidirectional association mapping, label the corresponding association type and association strength value, and record the temporal feature segment identifier corresponding to the bidirectional association mapping, and label the time distribution range of the bidirectional association mapping.

[0026] In this embodiment, the attribute description of the bidirectional correlation mapping explicitly marks the correlation type (direct or indirect) and correlation strength value. For example, "Inlet / outlet pressure difference → Line current (direct correlation, correlation strength 0.8)". Simultaneously, the time-series feature segment identifier corresponding to this bidirectional correlation mapping is recorded. The time-series feature segment identifier is a unique number for each time-series feature segment, automatically generated by the edge computing nodes according to the segmentation order. Based on the start and end times of the time-series feature segments, the time distribution range of the bidirectional correlation mapping is marked, for example, "Time-series feature segment 10: Start time - End time", thus clarifying the time period within which the correlation mapping is valid.

[0027] Step S1145: Classify and summarize all the established bidirectional association mappings according to feature dimensions, remove duplicate bidirectional association mapping records, retain the complete attribute information of bidirectional association mappings, and form an association mapping set containing association type, association strength, and temporal distribution information.

[0028] In this embodiment, all established bidirectional association mappings are classified according to feature dimensions, such as association mappings between sub-features within the device operating status feature, association mappings between the device operating status feature and load fluctuation feature, association mappings between the device operating status feature and environmental related features, and association mappings between load fluctuation feature and environmental related features. During the classification process, duplicate mapping records are checked and removed. For example, the same pair of features may be associated with multiple times in different time-series feature segments, but only the latest or most representative record needs to be retained. The final set of association mappings contains complete attribute information for each mapping, including source feature, target feature, transmission direction, association type, association strength, time-series feature segment identifier, and time distribution range.

[0029] Step S115: Based on the similarity of feature attributes, establish bidirectional correlation mappings between equipment operating status features and load fluctuation features, equipment operating status features and environmental correlation features, and load fluctuation features and environmental correlation features. Concatenate the time-series feature segments with the feature time-series anchor points corresponding to the core feature units in chronological order. Eliminate the gaps between time-series feature segments through feature interpolation technology to complete the continuous fusion of features. Combine the type, intensity, and time-series distribution information of all established bidirectional correlation mappings to form an ordered anchored feature sequence and a complete cross-dimensional correlation map.

[0030] In this embodiment, after establishing the bidirectional association mapping in step S114, the core feature units with feature time-series anchor points in each time-series feature segment are extracted and arranged sequentially according to time order. Since there may be small time gaps between time-series feature segments (e.g., due to data acquisition or processing delays), linear interpolation is used to fill the feature values ​​at the gaps to ensure the continuity of the entire feature sequence. For example, if two adjacent time-series feature segments have a time step gap, and the last core feature unit of the previous segment is value A, and the first core feature unit of the subsequent segment is value B, then the feature value at the gap is calculated using (A+B) / 2. After completing the feature continuity fusion, an ordered anchored feature sequence is formed, where each element contains the specific value of the core feature unit and the corresponding feature time-series anchor point. Simultaneously, the association mapping set obtained in step S1145 is integrated according to feature dimensions and time order to construct a complete cross-dimensional association graph. Nodes in the graph represent each feature dimension, edges represent bidirectional association mappings between features, and the edges are labeled with association type, association strength, and time-series distribution information, thus intuitively displaying the association relationships between features of different dimensions.

[0031] Step S120: Input the anchored feature sequence, feature time-series anchor points and cross-dimensional correlation graph into the artificial intelligence fault trajectory model, perform multi-scale feature trajectory mining and anomaly tracing, and obtain multi-scale anomaly feature tracing results and feature anomaly anchor point set.

[0032] In this embodiment, the AI ​​fault trajectory model is an end-to-end model built on a deep learning framework, specifically designed to mine potential fault trajectories from device operating characteristics and trace anomalies. This model is deployed in the AI ​​acceleration module of an edge computing node, enabling real-time inference using the computing power resources of edge computing.

[0033] Step S121: According to the input data format requirements preset by the artificial intelligence fault trajectory model, the anchored feature sequence is converted into a fixed-dimensional feature vector, the feature time-series anchor point is converted into a time index identifier, and the cross-dimensional correlation graph is converted into an adjacency matrix form, forming standardized input data that can be recognized by the artificial intelligence fault trajectory model. During the format conversion process, the correlation and binding relationship between the anchored feature sequence, the feature time-series anchor point, and the cross-dimensional correlation graph is maintained.

[0034] In this embodiment, the AI ​​fault trajectory model requires the input anchor feature sequence to be a fixed-length vector. Therefore, the original anchor feature sequence needs to be normalized. For example, if the model requires the input vector length to be a fixed value, the anchor feature sequence can be converted into a feature vector of that fixed length using sliding window sampling or feature dimensionality reduction (such as principal component analysis). When converting feature time-series anchors to time index identifiers, the timestamp information of each anchor point is converted into an integer index relative to a certain start time. For example, based on the start time of a day, the timestamp is converted into a second-level integer index. When converting cross-dimensional correlation graphs to adjacency matrices, the rows and columns of the matrix correspond to different feature dimensions, and the values ​​of the matrix elements represent the correlation strength between the corresponding feature dimensions. If two feature dimensions are not correlated, the element value is 0. During the format conversion process, additional index information is added to maintain the correlation binding relationship between the three. For example, the same batch number or timestamp label is added to each feature vector element, time index identifier, and adjacency matrix element to ensure that the model can correctly associate the correspondence between the three during processing.

[0035] Step S122: Through the multi-scale feature extraction layer of the artificial intelligence fault trajectory model, the fine-grained perception module, medium-grained perception module and coarse-grained perception module built into the multi-scale feature extraction layer are called to extract features at different scales from the anchor feature sequence in the standardized input data. The fine-grained perception module extracts the local detail change information of the features, the medium-grained perception module extracts the mid-term trend change information of the features, and the coarse-grained perception module extracts the overall global change information of the features, so as to obtain the fine-grained feature trajectory, the medium-grained feature trajectory and the coarse-grained feature trajectory.

[0036] In this embodiment, the multi-scale feature extraction layer of the AI ​​fault trajectory model consists of three parallel perception modules. The fine-grained perception module uses a convolutional neural network with small kernels (e.g., 3×3) to extract local detail changes in the anchored feature sequence over a short time period through multi-layer convolutional operations, such as instantaneous fluctuations and abrupt changes in features. The medium-grained perception module uses a convolutional neural network with medium-sized kernels (e.g., 5×5) and a stride, combined with pooling operations to extract trend changes over a medium time period, such as the upward or downward trend of features within a few minutes. The coarse-grained perception module uses a convolutional neural network with large kernels (e.g., 7×7) and a larger stride, or uses a recurrent neural network (e.g., LSTM) to globally model the entire anchored feature sequence, extracting overall change information over a long time period, such as the periodic change pattern of features within a few hours. Each perception module outputs a corresponding feature trajectory, namely, a fine-grained feature trajectory, a medium-grained feature trajectory, and a coarse-grained feature trajectory. These trajectories describe the change patterns of equipment operating characteristics at different time scales.

[0037] Step S123: Retrieve the preset normal feature trajectory library in the artificial intelligence fault trajectory model. The normal feature trajectory library contains standard fine-grained feature trajectories, standard medium-grained feature trajectories, and standard coarse-grained feature trajectories of the equipment under different operating conditions. Compare the extracted fine-grained, medium-grained, and coarse-grained feature trajectories with the corresponding standard fine-grained, medium-grained, and coarse-grained feature trajectories point by point, record the trajectory deviation information at each point, and use the deviation accumulation algorithm to quantify the trajectory deviation information to obtain the overall deviation value corresponding to the fine-grained, medium-grained, and coarse-grained feature trajectories. Set a deviation threshold and filter feature trajectories with overall deviation values ​​exceeding the threshold as abnormal scale feature trajectories.

[0038] In this embodiment, the pre-set normal feature trajectory library in the artificial intelligence fault trajectory model is constructed by extracting features and training the model from historical data of the water pump equipment under normal operating conditions. This library contains standard feature trajectories under various typical operating conditions (such as rated load, light load, startup phase, shutdown phase, etc.). When comparing trajectories, the corresponding standard trajectory is first selected from the normal feature trajectory library based on the current operating condition of the equipment. For example, if the water pump is currently under rated load, the standard fine-grained, medium-grained, and coarse-grained feature trajectories under rated load are selected. Point-by-point comparison refers to comparing the feature values ​​of the extracted feature trajectory and the standard feature trajectory at the same time point, and calculating the absolute difference between the two as the trajectory deviation information for that point. The deviation accumulation algorithm uses a weighted summation method to quantify the deviation information at all time points, assigning higher weights to recent deviations and lower weights to older deviations to highlight the impact of recent changes. The overall deviation value is the result of the weighted summation. Set deviation thresholds (fine-grained threshold, medium-grained threshold, coarse-grained threshold) for feature trajectories at different scales. When the overall deviation value of a certain scale exceeds its corresponding threshold, the feature trajectory at that scale is determined to be an abnormal scale feature trajectory.

[0039] Step S124: Combining the adjacency matrix information in the cross-dimensional correlation graph, trace the initial feature dimension of the abnormal feature unit in the abnormal scale feature trajectory, find the first-level correlation feature dimension that is directly related to the initial feature dimension from the cross-dimensional correlation graph, then find the second-level correlation feature dimension that is related to the first-level correlation feature dimension, and so on, to trace the complete propagation path of the abnormal feature between different dimensions; based on the order of abnormal occurrence of feature dimensions in the propagation path, locate the starting position of the abnormality and the propagation node.

[0040] In this embodiment, the abnormal feature unit in the abnormal scale feature trajectory refers to the feature point in the trajectory whose deviation value exceeds a set local threshold. First, the initial feature dimension to which the abnormal feature unit belongs is determined in the cross-dimensional correlation graph based on the feature identifier of the abnormal feature unit. For example, if the abnormal feature unit belongs to the line current feature in the load fluctuation feature, then the initial feature dimension is line current. Then, according to the adjacency matrix of the cross-dimensional correlation graph, the first-level correlated feature dimension that has a direct correlation with the line current (correlation strength greater than the set threshold) is found, such as the inlet and outlet pressure difference, motor stator temperature, etc. Next, using these first-level correlated feature dimensions as the starting point, their correlated feature dimensions are found as the second-level correlated feature dimensions, such as ambient temperature, vibration acceleration, etc. In this process, the time sequence of the occurrence of anomalies in each correlated feature dimension is recorded, thereby tracing the propagation path of abnormal features between different dimensions. The anomaly starting position refers to the feature dimension that first appears abnormally in the propagation path and its corresponding time point, while the propagation node is the other abnormal feature dimension in the propagation path other than the starting position.

[0041] Step S1241: Extract the core attribute information of the abnormal feature unit from the abnormal scale feature trajectory, and locate the initial feature dimension to which the abnormal feature unit belongs in the cross-dimensional correlation graph based on the core attribute information.

[0042] In this embodiment, the core attribute information of the abnormal feature unit includes feature name, feature value, timestamp, etc. For example, the feature name of the abnormal feature unit is "line current", the feature value is a certain abnormal value, and the timestamp is a certain specific moment. Based on the feature name "line current", the feature dimension list of the cross-dimensional correlation graph is searched to determine that its initial feature dimension is the line current dimension under the load fluctuation feature.

[0043] Step S1242: Find all associated feature dimensions that have an association mapping with the initial feature dimension from the cross-dimensional association graph. Sort the associated feature dimensions in order of priority from largest to smallest according to the association type and association strength value of the association mapping.

[0044] In this embodiment, in the cross-dimensional correlation map, the correlated feature dimensions that have a correlation mapping with the initial feature dimension (line current) include the inlet and outlet pressure difference under the equipment operating state characteristics, the motor stator temperature, and the vibration acceleration under the environmental correlation characteristics. Based on the correlation type (direct or indirect correlation) and correlation strength values ​​of these correlation mappings, directly correlated feature dimensions are preferentially selected and sorted from largest to smallest correlation strength values. For example, if the correlation strength between the inlet and outlet pressure difference and the line current is 0.8 (direct correlation), the correlation strength between the motor stator temperature and the line current is 0.7 (direct correlation), and the correlation strength between the vibration acceleration and the line current is 0.6 (indirect correlation), then the sorting result is: inlet and outlet pressure difference, motor stator temperature, and vibration acceleration.

[0045] Step S1243: In order of priority, retrieve the feature trajectories corresponding to the associated feature dimensions in sequence, compare them with the preset standard feature trajectories, and detect whether there are any abnormal deviations; mark the associated feature dimensions with abnormal deviations as the first-level propagation nodes.

[0046] In this embodiment, following the priority order of step S1242, the feature trajectory corresponding to the inlet and outlet pressure difference is first retrieved and compared with the corresponding standard feature trajectory in the normal feature trajectory library to calculate the overall deviation value. If the overall deviation value exceeds the deviation threshold corresponding to the inlet and outlet pressure difference, it is determined that there is an abnormal deviation in the inlet and outlet pressure difference, and it is marked as a first-level propagation node. If there is no abnormality in the inlet and outlet pressure difference, the feature trajectory corresponding to the motor stator temperature is retrieved and compared, and so on.

[0047] Step S1244: Continue to search for the associated feature dimensions that are associated with the first-level propagation node from the cross-dimensional association graph, repeat the process of retrieving the feature trajectory corresponding to the associated feature dimension and comparing it with the preset standard feature trajectory, and identify the second-level propagation node, the third-level propagation node and subsequent propagation nodes in sequence, and record the propagation order of abnormal features from the initial dimension to the propagation node.

[0048] In this embodiment, taking the inlet and outlet pressure difference of the first-level propagation node as an example, the associated feature dimensions, such as flow rate and pump impeller speed, are searched from the cross-dimensional correlation map. The feature trajectories corresponding to these feature dimensions are retrieved and compared with standard feature trajectories to detect any abnormal deviations. Feature dimensions with abnormal deviations are marked as second-level propagation nodes. Then, starting from the second-level propagation node, its associated feature dimensions are searched, and the same comparison and detection are performed to identify the third-level propagation node, and so on. Throughout the process, the time sequence in which each propagation node is identified is recorded to determine the propagation order of abnormal features.

[0049] Step S1245: Based on the feature time-series anchor points corresponding to the core feature units, find the time index identifier of the first occurrence of an anomaly in the initial feature dimension of the abnormal feature unit, and locate the anomaly starting position; extract the feature time-series anchor points corresponding to the core feature units corresponding to the anomaly starting position and propagation nodes, and include them in the feature anomaly anchor point set.

[0050] In this embodiment, the time point at which the abnormal feature unit first exhibits an anomaly in the initial feature dimension (line current) can be determined through its feature timing anchor point. The feature timing anchor point contains a time index identifier, which allows location of the specific time position at which the anomaly first appears, i.e., the anomaly initiation position. Then, the feature timing anchor points of the core feature unit corresponding to the anomaly initiation position, as well as the feature timing anchor points of the core feature units corresponding to each propagation node (such as inlet / outlet pressure difference, flow rate, etc.), are extracted, and these anchor points are compiled into a feature anomaly anchor point set.

[0051] Step S125: Integrate the detailed information of the anomaly scale feature trajectory, the time and dimension identifier of the anomaly starting position, the complete propagation path description, the feature information of the propagation nodes, and the feature time-series anchor points corresponding to the core feature units corresponding to the anomaly starting position and propagation nodes to form a multi-scale anomaly feature tracing result containing multi-dimensional anomaly information; summarize the feature time-series anchor points corresponding to all core feature units related to the anomaly to form a feature anomaly anchor point set.

[0052] In this embodiment, the multi-scale anomaly feature tracing result is a comprehensive summary of anomaly information, including the type of anomaly scale feature trajectory (fine-grained, medium-grained, or coarse-grained), the distribution of anomaly points on the trajectory, the specific time and feature dimension identifier of the anomaly's starting position, a detailed description of the anomaly propagation path (including propagation nodes at each level and their order of appearance), the feature names of the propagation nodes, and the degree of anomaly. The feature anomaly anchor point set is a collection of the feature time-series anchor points of the core feature units corresponding to the anomaly's starting position and all propagation nodes, forming a set containing multiple anchor points. Each anchor point uniquely identifies the position of an anomaly-related core feature unit in the time dimension.

[0053] Step S130: Based on the feature anomaly anchor point set and cross-dimensional correlation map, dynamically calibrate the feature temporal anchoring rules and correlation mapping strategies of edge computing nodes, and generate calibration anchoring parameters and optimized correlation strategies.

[0054] In this embodiment, the purpose of dynamic calibration is to improve the accuracy of edge computing nodes in capturing abnormal features and identifying correlations, thereby enhancing the reliability of fault warning.

[0055] Step S131: Traverse the feature anomaly anchor points in the feature anomaly anchor point set, extract the core feature units corresponding to the feature anomaly anchor points, determine the dimension type to which the core feature units belong based on their attribute information, count the frequency of occurrence of dimension types in the feature anomaly anchor point set, and determine the dimension type with the highest frequency of occurrence as the feature dimension with high anomaly incidence.

[0056] In this embodiment, each feature anomaly anchor point in the feature anomaly anchor point set is traversed, and the corresponding core feature unit is extracted based on the core feature unit identifier contained in the anchor point. The attribute information of the core feature unit includes its dimension type, such as equipment operating status features, load fluctuation features, environmental correlation features, etc., as well as more specific sub-dimensional types, such as line current, inlet and outlet pressure difference, etc. The frequency of occurrence of each sub-dimensional type in the feature anomaly anchor point set is counted. For example, if the line current dimension occurs 10 times, the inlet and outlet pressure difference dimension occurs 8 times, and the ambient temperature dimension occurs 5 times, then the line current dimension, which has the highest frequency of occurrence, is identified as the feature dimension with high anomaly incidence.

[0057] Step S132: Call the current feature temporal anchoring rules and association mapping strategies stored in the storage module of the edge computing node, and parse the preset anchoring dimension list and anchoring interval duration of the anchoring dimensions in the feature temporal anchoring rules, as well as the preset inter-dimensional association weight allocation scheme in the association mapping strategy.

[0058] In this embodiment, the storage module of the edge computing node stores the currently used feature temporal anchoring rules and association mapping strategies. The feature temporal anchoring rules include a list of anchoring dimensions (i.e., a list of feature dimensions that need to be temporally anchored) and the anchoring interval for each anchoring dimension (i.e., how often the core feature unit is extracted). The association mapping strategy includes a weight allocation scheme for inter-dimensional associations, i.e., the initial weight settings for the association strength between different feature dimensions. By parsing these stored rules and strategies, the current anchoring dimensions, anchoring intervals, and association weights are obtained.

[0059] Step S133: Set the feature dimensions with high anomaly occurrence as key anchor dimensions, and increase the coverage of sub-dimensions of key anchor dimensions in the anchor dimension list to achieve comprehensive capture of feature information of key anchor dimensions; extract the time position information of core feature units corresponding to all feature anomaly anchor points under key anchor dimensions, analyze the time distribution, optimize the anchor interval based on the time distribution, shorten the anchor interval during periods of concentrated anomalies, and improve the anchoring accuracy of time dimensions.

[0060] In this embodiment, after designating the feature dimension with the highest frequency of anomalies (such as line current) as the key anchoring dimension, the anchoring dimension list not only includes the basic characteristics of line current but also adds its related sub-dimensions, such as the harmonic components of line current and the three-phase current imbalance, to achieve comprehensive capture of line current feature information. For optimizing the anchoring interval, the time position information of the core feature units corresponding to all feature anomaly anchor points under the key anchoring dimension is first extracted. This information is obtained through the timestamps in the feature time-series anchor points. The distribution of this time position information is analyzed to identify the periods when anomalies are concentrated. For example, if it is found that line current anomalies are mainly concentrated in a few fixed hour periods each day, the anchoring interval is shortened during these periods, such as from once per unit time to multiple times per unit time. During other non-anomaly concentrated periods, the original anchoring interval is maintained or appropriately extended to balance data acquisition accuracy and system resource consumption.

[0061] Step S1331: Extract the temporal position information of the core feature units corresponding to all feature anomaly anchor points under the key anchoring dimension in the feature anomaly anchor point set, and arrange them in chronological order to form a temporal position sequence.

[0062] In this embodiment, the key anchoring dimension is line current. All feature anomaly anchors belonging to the line current dimension are selected from the feature anomaly anchor point set, and the timestamp information of the core feature unit corresponding to each anchor point is extracted as time position information. These time position information are sorted in chronological order to form a time position sequence, for example, [t1, t2, t3, ..., tn], where t1...tn... <t2<...<tn。

[0063] Step S1332: Perform statistical analysis on the time position sequence, calculate the time interval between two adjacent time positions, and obtain the value of the adjacent time interval; determine the concentrated distribution interval of the adjacent time interval values ​​through frequency distribution statistics.

[0064] In this embodiment, the difference between adjacent time positions is calculated for the time position sequence [t1,t2,t3,...,tn], i.e., t2-t1, t3-t2,...,tn-t(n-1), to obtain the numerical sequence of adjacent time intervals. Frequency distribution statistics are performed on this sequence, and a frequency histogram is plotted to identify the interval with the highest frequency, i.e., the concentrated distribution interval. For example, if statistics show that adjacent time intervals are mainly distributed within a certain time period, then that range is the concentrated distribution interval.

[0065] Step S1333: Using the median value of the concentrated distribution interval as a benchmark, set the initial calibration anchoring interval for the key anchoring dimension; calculate the degree of deviation between the initial calibration anchoring interval and the adjacent time interval values, and count the number of adjacent time intervals with a deviation degree less than the preset deviation threshold.

[0066] In this embodiment, the median value of the concentrated distribution interval is used as the reference value for the initial calibration anchoring interval. For example, if the concentrated distribution interval is [a, b], then the median value is (a+b) / 2, which is used as the initial calibration anchoring interval. The absolute value of the difference between each adjacent time interval value and the initial calibration anchoring interval is calculated, and the ratio of this absolute value to the initial calibration anchoring interval is used as the degree of deviation. A preset deviation threshold (e.g., 0.2) is set, and the number of adjacent time intervals with a deviation degree less than this threshold is counted.

[0067] Step S1334: If the statistical quantity ratio does not reach the preset ratio threshold, the initial calibration anchoring interval is reduced by a fixed ratio, and the deviation degree and quantity ratio are recalculated; the adjustment process of reducing the initial calibration anchoring interval by a fixed ratio and calculating the deviation degree and quantity ratio is repeated until the quantity ratio reaches the preset ratio threshold. The calibration anchoring interval at this time is the optimized anchoring interval of the key anchoring dimension.

[0068] In this embodiment, the quantity ratio refers to the ratio of the number of adjacent time intervals with a deviation less than a preset deviation threshold to the total number of adjacent time intervals. A preset percentage threshold (e.g., 0.8) is used. If the current quantity ratio does not reach this threshold, the initial calibration anchoring interval is reduced by a fixed proportion (e.g., 0.9), and the deviation and quantity ratio are recalculated. This process is repeated until the quantity ratio reaches or exceeds the preset percentage threshold; the calibration anchoring interval at this point is the optimized anchoring interval.

[0069] Step S1335: Extract the current anchoring interval of non-key anchoring dimensions in the edge computing nodes, detect whether the optimized key anchoring interval and the remaining anchoring interval conflict in the distribution of the time dimension, adjust the conflicting anchoring intervals so that the anchoring nodes of the dimension do not overlap and are evenly distributed; at the same time, detect whether the optimized anchoring interval is within the processing capacity of the edge computing nodes, adjust the anchoring intervals that exceed the processing capacity range so that the anchoring accuracy of the time dimension meets the preset requirements of anomaly tracing.

[0070] In this embodiment, the current anchoring interval of non-key anchoring dimensions may overlap with the optimized key anchoring interval in time, leading to data acquisition conflicts. By checking the timestamps of anchoring nodes for each dimension, if overlap exists, the anchoring interval of non-key anchoring dimensions is adjusted, for example, by increasing its anchoring interval by a small offset, so that the anchoring nodes of each dimension are evenly distributed in time. Meanwhile, the processing capacity of edge computing nodes is limited, and it is necessary to ensure that the optimized anchoring interval does not cause data processing overload. By monitoring indicators such as CPU utilization and memory usage of edge computing nodes, if the optimized anchoring interval causes resource utilization to exceed a preset threshold, the anchoring interval is appropriately extended until resource utilization is within a reasonable range, while ensuring that the anchoring accuracy of the time dimension meets the requirements for anomaly tracing.

[0071] Step S134: Combine the correlation relationships corresponding to the abnormal propagation path in the cross-dimensional correlation graph to identify the key correlation links on the abnormal propagation path. Based on the influence of the key correlation links on the abnormal propagation, adjust the correlation weight values ​​between the corresponding dimensions and adjust the correlation weights of the key correlation links to a preset value range to enhance the sensitivity of abnormal propagation path identification.

[0072] In this embodiment, the associated links on the anomaly propagation path refer to the association mappings connecting the propagation nodes. By analyzing the role of these associated links in the anomaly propagation process, links that play a crucial role in the anomaly propagation are identified, namely, key associated links. For example, the associated link between line current and the inlet / outlet pressure difference plays an important bridging role in anomaly propagation and is therefore identified as a key associated link. Based on the degree of influence of the key associated links on anomaly propagation (which can be measured by indicators such as the frequency of the link's appearance in the anomaly propagation path and the magnitude of changes in association strength), the association weight values ​​between corresponding dimensions are adjusted. The association weights of the key associated links are adjusted to a preset high value range (e.g., 0.8-0.9) to enhance the model's attention to these links and improve the sensitivity of anomaly propagation path identification.

[0073] Step S135: Integrate the adjusted list of anchoring dimensions and the optimized anchoring interval duration to form calibration anchoring parameters; integrate the adjusted inter-dimensional correlation weight allocation scheme to form an optimized correlation strategy; perform a collaborative verification process on the calibration anchoring parameters and the optimized correlation strategy to verify their compatibility and lack of logical conflict in actual operation.

[0074] In this embodiment, the calibration anchoring parameters include an adjusted list of anchoring dimensions (with added sub-dimensions of key anchoring dimensions) and optimized anchoring intervals for each dimension. The optimized association strategy is the adjusted inter-dimension association weight allocation scheme. The collaborative verification process simulates data acquisition and association analysis to check whether the calibration anchoring parameters can accurately capture abnormal features, whether the optimized association strategy can correctly identify abnormal propagation paths, and whether there are logical conflicts between the two during data processing (such as mismatch between anchoring and association dimensions). If problems are found during verification, steps S133 and S134 are returned for readjustment until the calibration anchoring parameters and optimized association strategy work together normally.

[0075] Step S140: By applying calibration anchoring parameters and optimization association strategies through edge computing nodes, the multi-source continuous feature stream is reprocessed to obtain the calibration anchoring feature sequence, calibration feature time-series anchor points, and optimized cross-dimensional association map.

[0076] In this embodiment, calibrated parameters and strategies are applied to reprocess the multi-source continuous feature stream to improve the quality of the feature data and the accuracy of the correlation.

[0077] Step S141: Transmit the calibration anchoring parameters and optimization association strategy to the feature processing module through the internal communication bus of the edge computing node, replace the original anchoring parameters and association strategy configuration in the feature processing module, and update the running parameter settings of the feature processing module.

[0078] In this embodiment, the internal communication bus of the edge computing node adopts the Industrial Ethernet protocol to ensure the real-time performance and reliability of data transmission. The calibrated anchoring parameters and optimized association strategies are packaged in a structured data format (such as JSON) and transmitted to the feature processing module via the bus. Upon receiving the new parameters and strategies, the feature processing module stores them in its local cache and updates its internal operating parameter configurations, such as updating the anchoring dimension list, anchoring interval timer, and association weight matrix.

[0079] Step S142: Based on the list of anchoring dimensions in the calibration anchoring parameters, filter the feature dimensions of the multi-source continuous feature stream of the device operation, retain all feature information of the dimensions contained in the list, and remove feature information of irrelevant dimensions; according to the optimized anchoring interval duration in the calibration anchoring parameters, perform time dimension segmentation processing on the filtered multi-source continuous feature stream to generate multiple continuous calibration time sequence feature segments.

[0080] In this embodiment, the feature dimension screening process extracts all the feature dimension information that needs to be anchored from the multi-source continuous feature stream based on the anchor dimension list in the calibration anchoring parameters. Examples include line current and its sub-dimensions, inlet and outlet pressure differences, etc., while removing feature dimension information that is irrelevant to anomaly detection, such as certain environmental noise data. The time dimension segmentation process divides the screened feature stream into multiple continuous calibration time-series feature segments according to the optimized anchoring interval duration. The length of each calibration time-series feature segment is equal to the anchoring interval duration.

[0081] Step S143: Extract the core feature units in the generated calibration time series feature segment, and add calibrated time dimension associated position markers to the core feature units in the calibration time series feature segment based on the time dimension associated position marker rules set in the calibration anchoring parameters, thereby generating calibration feature time series anchor points corresponding to the core feature units in the calibration time series feature segment.

[0082] In this embodiment, the extraction method of the core feature units is similar to that in step S113, but the updated feature importance evaluation criteria in the calibration anchoring parameters are adopted. The time dimension associated position marking rules have been optimized based on the original ones, for example, considering the sub-dimension information of the key anchoring dimension, making the marking more accurate. After adding the calibrated time dimension associated position markings, calibrated feature time-series anchor points are generated, which contain more detailed feature dimension information and more accurate time position information.

[0083] Step S144: According to the inter-dimensional association weight allocation scheme set in the optimization association strategy, calculate the feature attribute similarity between equipment operating status features and load fluctuation features, equipment operating status features and environmental association features, and load fluctuation features and environmental association features in the calibration time series feature segment. Based on the similarity and association weight, establish a bidirectional association mapping between equipment operating status features and load fluctuation features, equipment operating status features and environmental association features, and load fluctuation features and environmental association features in the calibration time series feature segment.

[0084] In this embodiment, the method for calculating feature attribute similarity is similar to step S114, but the adjusted association weight allocation scheme in the optimized association strategy is used. When calculating the overall feature attribute similarity, the association weight of key association links is assigned a higher weight value, thereby affecting the similarity calculation result. Based on the new similarity result and association weight, a bidirectional association mapping is re-established, making the mapping relationship more consistent with the current operating status and anomaly propagation characteristics of the device.

[0085] Step S145: The calibration time-series feature segments with calibration feature time-series anchor points corresponding to the core feature units in the calibration time-series feature segments are sequentially connected in chronological order. The seamless fusion between the calibration time-series feature segments is achieved through feature fusion algorithm. Combined with the update information of all established bidirectional correlation mappings, the calibration anchor feature sequence and optimized cross-dimensional correlation map are integrated to form a calibration anchor feature sequence and optimized cross-dimensional correlation map.

[0086] In this embodiment, the feature fusion algorithm uses a weighted average method to smooth the feature values ​​at the boundaries between calibration time-series feature segments, ensuring the continuity of the calibration anchor feature sequence. Optimizing the cross-dimensional correlation map involves updating the correlation type, correlation strength, and temporal distribution information of the bidirectional correlation mapping based on the original map, enabling it to more accurately reflect the real-time correlation relationships between feature dimensions.

[0087] Step S150: Combine the multi-scale anomaly feature tracing results, calibrated anchor feature sequences, and optimized cross-dimensional correlation maps to generate fault warning commands and transmit them to the equipment management platform.

[0088] In this embodiment, the generation of fault warning commands is based on comprehensive analysis of multi-source information to ensure the accuracy and timeliness of the warnings, while the transmission process ensures the security and reliability of the data.

[0089] Step S151: Extract the time and dimension identifiers of the anomaly starting position, the complete description of the anomaly propagation path, the feature information of the propagation nodes, and the definition of the anomaly's impact range from the multi-scale anomaly feature tracing results; retrieve the calibration anchor feature sequence and the optimized cross-dimensional correlation map, and find the corresponding feature segments of the anomaly starting position and propagation path in the calibration anchor feature sequence, as well as the corresponding correlation links in the optimized cross-dimensional correlation map, through the feature matching algorithm.

[0090] In this embodiment, the scope of anomaly impact is determined based on the anomaly propagation path and the feature importance assessment of each propagation node. For example, if the anomaly propagates to the motor stator temperature dimension, the scope of impact may include the motor's insulation system. The feature matching algorithm employs Dynamic Time Warping (DTW) to match the anomaly's starting position and propagation path in the multi-scale anomaly feature tracing results with feature segments in the calibration anchored feature sequence to find the most similar segment; simultaneously, it searches for the correlation link corresponding to the propagation path in the optimized cross-dimensional correlation graph.

[0091] Step S152: Classify and integrate the matched anomaly start location information, propagation path information, propagation node information, impact range information, corresponding feature fragments and related link information, and organize them into a structured fault early warning core dataset according to the time sequence of anomaly occurrence and dimensional association logic.

[0092] In this embodiment, the classification and integration process categorizes different types of information into different data fields. For example, the anomaly start location information includes timestamps and feature dimension names; the propagation path information includes the propagation node sequence and propagation order; and the impact scope information includes affected equipment components and possible fault types. This information is organized according to time sequence and dimensional association logic to form a structured fault early warning core dataset, which is stored in the form of relational database tables.

[0093] Step S153: Call the preset warning instruction generation template. The warning instruction generation template includes anomaly basic information field, feature data field, correlation graph field, instruction priority field, and transmission identifier field. Fill the anomaly starting position, propagation path, propagation node, and impact range information from the fault warning core dataset into the anomaly basic information field. Fill the corresponding feature fragments into the feature data field. Fill the corresponding correlation link information into the correlation graph field. Match the corresponding instruction priority according to the anomaly impact range and severity, and fill it into the instruction priority field.

[0094] In this embodiment, the early warning instruction generation template is a predefined XML format file containing multiple fields. The anomaly basic information field is used to fill in descriptive text information, such as "the anomaly started in the line current dimension, at a certain time"; the feature data field is used to store the numerical data of the feature segments, represented in array form; the association map field is used to store the identification and association strength information of the associated links; the instruction priority field is divided into three levels: emergency, important, and general according to the scope and severity of the anomaly's impact. For example, anomalies that affect the safe operation of the motor are marked as emergency priority.

[0095] Step S154: Standardize the format of the completed early warning instruction generation template, standardize the expression format, data type and arrangement order of the fields, so that the field information meets the requirements of clear readability and logical coherence; according to the instruction data format requirements preset by the equipment management platform, convert the standardized template into a binary instruction format that the platform can recognize, and generate a fault early warning instruction.

[0096] In this embodiment, the format standardization process includes converting text information into a unified encoding format (such as UTF-8), converting numerical data into a specified data type (such as floating-point numbers or integers), and arranging the fields in a preset order. The device management platform's preset instruction data format is a specific binary protocol format, including an instruction header, data length, data body, checksum, etc. The standardized XML template is converted into this binary format using a format conversion tool to generate a fault warning instruction.

[0097] Step S155: Activate the communication module of the edge computing node, retrieve the Internet Protocol (IP) address, port number, and authentication key of the device management platform, and send a link establishment request to the device management platform through the communication module. The request includes the identity identifier of the edge computing node and the priority information of the fault warning command. Receive the link establishment response from the device management platform and verify the validity of the platform identity identifier in the response information. After successful verification, negotiate the transmission protocol and data encryption method. Establish a dedicated communication link between the edge computing node and the device management platform based on the negotiation result. Encrypt the fault warning command to generate an encrypted data packet. Transmit the encrypted data packet in blocks to the device management platform through the dedicated communication link, and monitor the transmission progress and status of the data blocks in real time. Receive data block reception confirmation information from the device management platform. After all data blocks have been transmitted, send a transmission completion notification to the device management platform to complete the transmission of the fault warning command.

[0098] In this embodiment, the communication module of the edge computing node supports multiple network protocols, such as TCP / IP and MQTT. The Internet protocol communication address, port number, and authentication key are stored in the secure storage area of ​​the edge computing node and read using encryption. The edge computing node identity identifier included in the link establishment request is used by the device management platform for authentication of the edge node. Platform identity identifier validity verification is achieved by comparing it with a preset platform public key certificate. The highly reliable TCP protocol is prioritized in the transmission protocol negotiation, and the AES-256 encryption algorithm is selected in the data encryption method negotiation. A dedicated communication link is established by setting a TCP connection and a unique port number. The encryption processing of fault warning commands uses the negotiated AES-256 algorithm, and the encryption key is transmitted through a secure channel. The size of the encrypted data packet block is determined based on the network bandwidth and MTU value, typically 1024 bytes or 4096 bytes. Transmission progress and status are monitored through an acknowledgment mechanism after each data block is sent; if no acknowledgment is received within a preset time, a retransmission mechanism is triggered. After all data blocks have been transmitted, the device management platform sends a final receipt confirmation to the edge computing node, which then sends a transmission completion notification, thus ending the entire fault warning command transmission process.

[0099] For example, in step S1551: Start the communication module of the edge computing node, retrieve the Internet protocol communication address, port number and identity authentication public key of the device management platform from the configuration file of the edge computing node; construct a link establishment request containing the edge computing node identity, communication protocol version and fault warning command priority, and send it to the device management platform through the communication module.

[0100] In this embodiment, the configuration file for the edge computing node is an encrypted JSON file stored on a non-volatile storage medium. After the communication module starts, it reads the device management platform's Internet Protocol (IP) address (e.g., IPv4 address), port number (e.g., a specific port), and authentication public key (RSA public key) from the configuration file through a dedicated decryption interface. The link establishment request is constructed according to a preset protocol format, including a fixed-length header and a variable-length data segment. The data segment contains the edge computing node's unique identifier (e.g., MAC address), supported communication protocol version number (e.g., V1.0), and the priority of the fault warning command (e.g., emergency). The request is sent to the device management platform through the communication module's network interface.

[0101] Step S1552: Receive the link establishment response information from the device management platform, parse the platform identity identifier, protocol confirmation information and encryption method suggestion in the response information; verify whether the platform identity identifier is consistent with the preset device management platform identity information, and perform a communication object legality determination.

[0102] In this embodiment, after receiving a link establishment request, the device management platform returns a link establishment response. The response includes a platform identity identifier (such as the hash value of the platform certificate), a confirmed communication protocol version, and a list of supported encryption methods (such as AES-256, DES, etc.). After parsing the response, the edge computing node compares the platform identity identifier with the preset device management platform identity information stored locally. If they match, the communication object is deemed legitimate; otherwise, link establishment is refused.

[0103] Step S1553: Based on the encryption method suggestions in the response information, negotiate and determine the final data encryption method; generate the corresponding encryption key pair, including a public key for encryption and a private key for decryption; send the encryption public key to the device management platform, receive the public key reception confirmation information from the device management platform, and perform a key transmission success determination.

[0104] In this embodiment, the edge computing node selects AES-256, the most secure encryption method from the list of encryption methods supported by the device management platform, as the final data encryption method. An AES-256 encryption key pair (public and private keys) is generated using the built-in encryption chip. The public key is used by the device management platform to encrypt response data, and the private key is stored locally by the edge computing node for decryption. The encrypted public key is sent to the device management platform through the currently unencrypted link. Upon receiving the public key, the device management platform returns a public key reception confirmation message. After receiving the confirmation message, the edge computing node determines that the key transmission was successful.

[0105] Step S1554: Using the encryption method determined through negotiation, the generated fault warning instruction is segmented and encrypted. The fault warning instruction is split into multiple data blocks of fixed size, and after encrypting the data blocks, a check code is added to generate an encrypted data packet.

[0106] In this embodiment, the size of the fault warning command may exceed the maximum data packet length for network transmission, thus requiring segmentation. The command is split into multiple fixed-size data blocks (e.g., 4096 bytes), with the last block potentially smaller than the fixed size. Each data block is encrypted using the AES-256 algorithm in CBC mode. An initialization vector (IPV) is randomly generated and sent along with the encrypted data block. A CRC32 checksum is appended to each encrypted data block for the device management platform to verify its integrity. All encrypted data blocks, along with their corresponding IPVs and checksums, are combined to form an encrypted data packet.

[0107] Step S1555: Establish a dedicated communication link between the edge computing node and the device management platform through the negotiated transmission protocol. Mark the established dedicated communication link as a dedicated channel for fault early warning and set access permissions to prohibit unrelated data transmission. Transmit encrypted data packets to the device management platform block by block through the dedicated communication link in the order of data blocks. Monitor the transmission progress, transmission delay and packet loss of data blocks in real time, and trigger the packet loss retransmission process to handle lost data blocks. Receive data block reception confirmation information and verification results from the device management platform. After all data blocks have been transmitted and verified, send a transmission completion notification to the device management platform to complete the encrypted transmission of the fault early warning command.

[0108] In this embodiment, the negotiated transmission protocol is TCP, and a TCP connection is established as a dedicated communication link through a three-way handshake. At the operating system level, this link is marked as a dedicated channel for fault warning and an access control list (ACL) is configured to allow only fault warning-related data between the edge computing node and the device management platform to be transmitted through this channel. The encrypted data packets are transmitted sequentially according to the data block order. After each data block is sent, the device management platform confirms receipt. The confirmation message includes the data block number and CRC32 checksum result. If the checksum fails or no confirmation is received within a timeout period, a packet loss retransmission process is triggered, and the data block is retransmitted. After all data blocks have been transmitted and verified, the edge computing node sends a transmission completion notification to the device management platform. The notification includes information such as the total length of the data packets and the number of data blocks. After the device management platform confirms that everything is correct, the entire encrypted transmission process is complete.

[0109] Figure 2 The illustration shows exemplary hardware and software components of a fault warning system 100 based on edge computing and artificial intelligence, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the fault warning system 100 based on edge computing and artificial intelligence and to perform the functions in this application.

[0110] The fault early warning system 100 based on edge computing and artificial intelligence can be a general-purpose server or a special-purpose server; both can be used to implement the fault early warning method based on edge computing and artificial intelligence of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0111] For example, the fault warning system 100 based on edge computing and artificial intelligence may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the fault warning system 100 based on edge computing and artificial intelligence may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The fault warning system 100 based on edge computing and artificial intelligence also includes an I / O interface 150 between the computer and other input / output devices.

[0112] For ease of illustration, only one processor is described in the fault warning system 100 based on edge computing and artificial intelligence. However, it should be noted that the fault warning system 100 based on edge computing and artificial intelligence in this application may also include multiple processors. Therefore, the steps performed by one processor as described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the fault warning system 100 based on edge computing and artificial intelligence performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0113] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned fault warning method based on edge computing and artificial intelligence is implemented.

[0114] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A fault early warning method based on edge computing and artificial intelligence, characterized in that, The method includes: By capturing multi-source continuous feature streams running on devices through edge computing nodes, feature temporal anchoring and cross-dimensional association processing are performed on the multi-source continuous feature streams to generate anchored feature sequences, feature temporal anchor points and cross-dimensional association maps. The anchored feature sequence, feature time-series anchor points and cross-dimensional correlation graph are input into the artificial intelligence fault trajectory model to perform multi-scale feature trajectory mining and anomaly tracing, and obtain multi-scale anomaly feature tracing results and feature anomaly anchor point set. Based on the feature anomaly anchor point set and cross-dimensional correlation map, the feature temporal anchoring rules and correlation mapping strategies of edge computing nodes are dynamically calibrated to generate calibration anchoring parameters and optimized correlation strategies. By applying calibration anchoring parameters and optimization association strategies to edge computing nodes, multi-source continuous feature streams are reprocessed to obtain calibration anchoring feature sequences, calibration feature time-series anchor points, and optimized cross-dimensional association maps. By combining the results of multi-scale anomaly feature tracing, calibrating and anchoring feature sequences, and optimizing cross-dimensional correlation maps, fault warning commands are generated and transmitted to the equipment management platform.

2. The fault early warning method based on edge computing and artificial intelligence according to claim 1, characterized in that, The process of capturing multi-source continuous feature streams running through edge computing nodes and performing feature temporal anchoring and cross-dimensional correlation processing on the multi-source continuous feature streams generates anchored feature sequences, feature temporal anchor points, and cross-dimensional correlation maps, including: Through the multi-source feature acquisition module of the edge computing node, the device's operation status output port, load monitoring port and environmental sensing port are connected to the device respectively. The device operation raw data, load fluctuation raw data and environmental related raw data continuously output by the port are received synchronously. The device operation raw data, load fluctuation raw data and environmental related raw data are processed to perform timestamp alignment processing to keep the time dimension of the device operation raw data, load fluctuation raw data and environmental related raw data consistent. Feature extraction processing is performed on the raw equipment operation data to extract equipment operation status features according to the preset extraction standards of the core equipment operation indicators; feature extraction processing is performed on the raw load fluctuation data to extract load fluctuation features according to the preset extraction standards of load changing over time; feature extraction processing is performed on the raw environmental data to extract environmental correlation features according to the preset extraction standards of the impact of environmental factors on equipment operation; and the equipment operation status features, load fluctuation features, and environmental correlation features are integrated in chronological order to form a multi-source continuous feature stream. According to the continuous time process of equipment operation, the time segmentation rules are set, and the multi-source continuous feature stream is uniformly segmented with a fixed time span as the segmentation benchmark to obtain multiple continuous time-series feature segments. The core feature units that represent the core state of equipment operation in the corresponding time-series feature segments are extracted. Based on the relative position information of the core feature units in the corresponding time-series feature segments, a unique time dimension associated position mark is added to the core feature units to generate the feature time-series anchor points corresponding to the core feature units. Based on the feature attributes of equipment operating status characteristics, load fluctuation characteristics, and environmental correlation characteristics, the numerical change trend, fluctuation frequency, and correlation influence of the features are selected to calculate the feature attribute similarity between equipment operating status characteristics and load fluctuation characteristics, equipment operating status characteristics and environmental correlation characteristics, and load fluctuation characteristics and environmental correlation characteristics. Based on the similarity of feature attributes, a bidirectional correlation mapping is established between equipment operating status features and load fluctuation features, equipment operating status features and environmental correlation features, and load fluctuation features and environmental correlation features. The time-series feature segments with feature time-series anchor points corresponding to core feature units are sequentially concatenated in chronological order. The gaps between time-series feature segments are eliminated by feature interpolation technology to complete the continuous fusion of features. Combining the type, intensity, and time-series distribution information of all established bidirectional correlation mappings, an ordered anchored feature sequence and a complete cross-dimensional correlation map are integrated to form an ordered feature sequence and a complete cross-dimensional correlation map.

3. The fault early warning method based on edge computing and artificial intelligence according to claim 1, characterized in that, The process involves inputting anchored feature sequences, feature time-series anchor points, and cross-dimensional correlation maps into an artificial intelligence fault trajectory model to perform multi-scale feature trajectory mining and anomaly tracing, resulting in multi-scale anomaly feature tracing results and a set of feature anomaly anchor points, including: According to the input data format requirements preset by the artificial intelligence fault trajectory model, the anchored feature sequence is converted into a fixed-dimensional feature vector, the feature time-series anchor point is converted into a time index identifier, and the cross-dimensional correlation graph is converted into an adjacency matrix form, forming standardized input data that can be recognized by the artificial intelligence fault trajectory model. During the format conversion process, the correlation and binding relationship between the anchored feature sequence, the feature time-series anchor point, and the cross-dimensional correlation graph is maintained. By using the multi-scale feature extraction layer of the artificial intelligence fault trajectory model, the fine-grained perception module, medium-grained perception module, and coarse-grained perception module built into the multi-scale feature extraction layer are invoked to extract features at different scales from the anchor feature sequence in the standardized input data. The fine-grained perception module extracts local detail change information of the features, the medium-grained perception module extracts mid-term trend change information of the features, and the coarse-grained perception module extracts overall global change information of the features, thus obtaining fine-grained feature trajectories, medium-grained feature trajectories, and coarse-grained feature trajectories. The system retrieves a pre-defined normal feature trajectory library from the AI ​​fault trajectory model. This library contains standard fine-grained, medium-grained, and coarse-grained feature trajectories of the equipment under different operating conditions. The extracted fine-grained, medium-grained, and coarse-grained feature trajectories are then compared point-by-point with their corresponding standard fine-grained, medium-grained, and coarse-grained feature trajectories. The trajectory deviation information at each point is recorded. A deviation accumulation algorithm is used to quantify the trajectory deviation information, yielding the overall deviation value for each fine-grained, medium-grained, and coarse-grained feature trajectory. A deviation threshold is set, and feature trajectories with overall deviation values ​​exceeding the threshold are filtered as abnormal scale feature trajectories. By combining the adjacency matrix information in the cross-dimensional correlation graph, the initial feature dimension of the anomalous feature unit in the anomalous scale feature trajectory is traced. The first-level correlation feature dimension that is directly related to the initial feature dimension is found in the cross-dimensional correlation graph. Then, the second-level correlation feature dimension that is related to the first-level correlation feature dimension is found, and so on, to trace the complete propagation path of the anomalous feature between different dimensions. Based on the order of the occurrence of the anomalous feature dimension in the propagation path, the starting position of the anomalous feature and the propagation node are located. By integrating detailed information of anomaly scale feature trajectories, time and dimension identifiers of anomaly initiation positions, complete propagation path descriptions, feature information of propagation nodes, and feature time-series anchor points corresponding to core feature units corresponding to anomaly initiation positions and propagation nodes, a multi-scale anomaly feature tracing result containing multi-dimensional anomaly information is formed; and all feature time-series anchor points corresponding to core feature units related to anomalies are summarized to form a feature anomaly anchor point set.

4. The fault early warning method based on edge computing and artificial intelligence according to claim 1, characterized in that, The method, based on a set of feature anomaly anchor points and a cross-dimensional correlation graph, dynamically calibrates the feature temporal anchoring rules and correlation mapping strategies of edge computing nodes, generating calibration anchoring parameters and optimized correlation strategies, including: Traverse the feature anomaly anchor points in the feature anomaly anchor point set, extract the core feature units corresponding to the feature anomaly anchor points, determine the dimension type to which the core feature units belong based on the attribute information of the core feature units, count the frequency of occurrence of dimension types in the feature anomaly anchor point set, and determine the dimension type with the highest frequency of occurrence as the feature dimension with high anomaly incidence. The current feature time-series anchoring rules and association mapping strategies stored in the storage module of the edge computing node are called to parse the list of anchoring dimensions and the anchoring interval duration of the anchoring dimensions in the feature time-series anchoring rules, as well as the association weight allocation scheme between dimensions in the association mapping strategy. The feature dimensions with high anomalies are set as key anchor dimensions. The sub-dimension coverage of key anchor dimensions is increased in the anchor dimension list to achieve comprehensive capture of feature information of key anchor dimensions. The time position information of the core feature units corresponding to all feature anomaly anchor points under key anchor dimensions is extracted, the time distribution is analyzed, and the anchoring interval is optimized based on the time distribution to shorten the anchoring interval during the period of concentrated anomalies and improve the anchoring accuracy of the time dimension. By combining the correlation relationships corresponding to the anomaly propagation path in the cross-dimensional correlation graph, key correlation links on the anomaly propagation path are identified. Based on the influence of the key correlation links on the anomaly propagation, the correlation weight values ​​between the corresponding dimensions are adjusted, and the correlation weights of the key correlation links are adjusted to a preset value range to enhance the sensitivity of anomaly propagation path identification. The adjusted list of anchoring dimensions and the optimized anchoring interval duration are integrated to form the calibration anchoring parameters; the adjusted inter-dimensional correlation weight allocation scheme is integrated to form the optimized correlation strategy; a collaborative verification process is performed on the calibration anchoring parameters and the optimized correlation strategy to verify their compatibility and lack of logical conflict in actual operation.

5. The fault early warning method based on edge computing and artificial intelligence according to claim 1, characterized in that, The process involves applying calibration anchoring parameters and optimizing correlation strategies through edge computing nodes to reprocess multi-source continuous feature streams, resulting in calibration anchoring feature sequences, calibration feature time-series anchor points, and optimized cross-dimensional correlation maps. This includes: The calibration anchoring parameters and optimization association strategies are transmitted to the feature processing module through the internal communication bus of the edge computing node, replacing the original anchoring parameters and association strategy configurations in the feature processing module and updating the running parameter settings of the feature processing module. Based on the list of anchoring dimensions in the calibration anchoring parameters, the multi-source continuous feature stream of the device operation is filtered for feature dimensions, retaining all feature information of the dimensions in the list and removing feature information of irrelevant dimensions; according to the optimized anchoring interval duration in the calibration anchoring parameters, the filtered multi-source continuous feature stream is segmented in the time dimension to generate multiple continuous calibration time series feature segments. Extract the core feature units from the generated calibration time series feature segments, and add calibrated time dimension associated position markings to the core feature units in the calibration time series feature segments based on the time dimension associated position marking rules set in the calibration anchoring parameters, thereby generating calibration feature time series anchor points corresponding to the core feature units in the calibration time series feature segments. According to the inter-dimensional association weight allocation scheme set in the optimized association strategy, the similarity of feature attributes between equipment operating status features and load fluctuation features, equipment operating status features and environmental association features, and load fluctuation features and environmental association features in the calibration time series feature segment is calculated. Based on the similarity and association weight, a bidirectional association mapping is established between equipment operating status features and load fluctuation features, equipment operating status features and environmental association features, and load fluctuation features and environmental association features in the calibration time series feature segment. The calibration time-series feature segments with calibration feature time-series anchor points corresponding to the core feature units in the calibration time-series feature segments are concatenated in chronological order. The seamless fusion between the calibration time-series feature segments is achieved through feature fusion algorithm. Combined with the update information of all established bidirectional correlation mappings, a calibration anchor feature sequence and an optimized cross-dimensional correlation map are formed.

6. The fault early warning method based on edge computing and artificial intelligence according to claim 1, characterized in that, The process of combining multi-scale anomaly feature tracing results, calibrated anchored feature sequences, and optimized cross-dimensional correlation maps to generate fault warning commands and transmit them to the equipment management platform includes: Extract the time and dimension identifiers of the anomaly initiation position, the complete description of the anomaly propagation path, the feature information of the propagation nodes, and the definition of the anomaly's impact range from the multi-scale anomaly feature tracing results; retrieve the calibration anchor feature sequence and the optimized cross-dimensional correlation map, and find the corresponding feature segments of the anomaly initiation position and propagation path in the calibration anchor feature sequence, as well as the corresponding correlation links in the optimized cross-dimensional correlation map through feature matching algorithms; The matched anomaly start location information, propagation path information, propagation node information, impact range information, corresponding feature fragments and related link information are classified and integrated, and organized into a structured fault early warning core dataset according to the time sequence of anomaly occurrence and the dimensional association logic. The system calls a preset warning instruction generation template, which includes fields for basic anomaly information, feature data, correlation graph, instruction priority, and transmission identifier. Information such as the anomaly's starting location, propagation path, propagation nodes, and impact range from the fault warning core dataset is filled into the basic anomaly information field. Corresponding feature fragments are filled into the feature data field. Corresponding correlation link information is filled into the correlation graph field. The corresponding instruction priority is matched according to the anomaly's impact range and severity, and then filled into the instruction priority field. The format of the completed early warning instruction template is standardized, including the expression format, data type, and arrangement order of the fields, so that the field information meets the requirements of clear readability and logical coherence. According to the instruction data format requirements preset by the equipment management platform, the standardized template is converted into a binary instruction format that the platform can recognize to generate fault early warning instructions. The edge computing node's communication module is activated, retrieving the Internet Protocol (IP) address, port number, and authentication key from the device management platform. A link establishment request is sent to the device management platform via the communication module, containing the edge computing node's identity identifier and the priority information of the fault warning command. The node receives the link establishment response from the device management platform and verifies the validity of the platform's identity identifier in the response information. After successful verification, the node negotiates the transmission protocol and data encryption method. Based on the negotiation results, a dedicated communication link is established between the edge computing node and the device management platform. The fault warning command is encrypted, generating encrypted data packets. These encrypted data packets are transmitted in blocks to the device management platform via the dedicated communication link, with real-time monitoring of the data block transmission progress and status. The node receives data block reception confirmation information from the device management platform. After all data blocks have been transmitted, a transmission completion notification is sent to the device management platform, completing the transmission of the fault warning command.

7. The fault early warning method based on edge computing and artificial intelligence according to claim 2, characterized in that, The method of establishing a bidirectional correlation mapping between equipment operating status features and load fluctuation features, equipment operating status features and environmental correlation features, and load fluctuation features and environmental correlation features based on the similarity of feature attributes includes: Select the numerical change trend, fluctuation frequency, peak occurrence timing, and correlation effect delay duration of the features, and calculate the similarity of the equipment operating status features and load fluctuation features, the equipment operating status features and environmental correlation features, and the load fluctuation features and environmental correlation features in terms of feature attribute dimensions. A weighted summation algorithm is used to comprehensively calculate the similarity across the feature attribute dimensions, resulting in the overall feature attribute similarity between equipment operating status features and load fluctuation features, equipment operating status features and environmental correlation features, and load fluctuation features and environmental correlation features. A similarity threshold is set, and feature combinations with an overall feature attribute similarity higher than the threshold are classified as directly correlated, while those lower than the threshold are classified as indirectly correlated. A preset high-value range of correlation strength is assigned to the directly correlated type, and a preset low-value range of correlation strength is assigned to the indirectly correlated type. Establish bidirectional pointing mappings from equipment operating status characteristics to load fluctuation characteristics and from load fluctuation characteristics to equipment operating status characteristics, respectively, and mark the feature transmission direction of the pointing mappings; In the established bidirectional association mapping, the corresponding association type and association strength value are labeled, and the temporal feature segment identifiers corresponding to the bidirectional association mapping are recorded, and the time distribution range of the bidirectional association mapping is labeled. All established bidirectional association mappings are categorized and summarized according to feature dimensions. Duplicate bidirectional association mapping records are removed, and complete attribute information of bidirectional association mappings is retained to form an association mapping set containing association type, association strength, and temporal distribution information.

8. The fault early warning method based on edge computing and artificial intelligence according to claim 3, characterized in that, The method of combining cross-dimensional correlation maps to trace the propagation path of abnormal features in different dimensions and locate the starting point and propagation nodes of the anomaly includes: The core attribute information of the abnormal feature unit is extracted from the abnormal scale feature trajectory, and the initial feature dimension to which the abnormal feature unit belongs is located in the cross-dimensional correlation graph based on the core attribute information. Find all associated feature dimensions that have an association mapping with the initial feature dimension from the cross-dimensional association graph. Sort the associated feature dimensions in order of priority from largest to smallest according to the association type and association strength value of the association mapping. According to priority, the feature trajectories corresponding to the associated feature dimensions are retrieved in turn and compared with the preset standard feature trajectories to detect whether there are any abnormal deviations; the associated feature dimensions with abnormal deviations are marked as first-level propagation nodes. Continue to search for associated feature dimensions that are associated with the first-level propagation node from the cross-dimensional association graph, repeatedly retrieve the feature trajectory corresponding to the associated feature dimension and compare it with the preset standard feature trajectory, and identify the second-level propagation node, the third-level propagation node and subsequent propagation nodes in sequence, and record the propagation order of abnormal features from the initial dimension to the propagation node. Based on the feature time-series anchor points corresponding to the core feature units, the time index identifier of the first occurrence of an anomaly in the initial feature dimension of the anomaly feature unit is found to locate the anomaly starting position; the feature time-series anchor points corresponding to the core feature units corresponding to the anomaly starting position and propagation nodes are extracted and included in the feature anomaly anchor point set; the anomaly propagation order, initial feature dimension, propagation nodes, and anomaly starting position are integrated to form a complete anomaly propagation path description.

9. The fault early warning method based on edge computing and artificial intelligence according to claim 4, characterized in that, The step of setting frequently occurring feature dimensions as key anchoring dimensions, adjusting the anchoring coverage and optimizing the anchoring interval to improve the anchoring accuracy of the time dimension includes: Extract the temporal location information of the core feature units corresponding to all feature anomaly anchor points under the key anchoring dimension in the feature anomaly anchor point set, and arrange them in chronological order to form a temporal location sequence. Statistical analysis is performed on the time position sequence to calculate the time interval between two adjacent time positions, thus obtaining the value of the adjacent time interval; the concentrated distribution interval of the adjacent time interval values ​​is determined by frequency distribution statistics. Using the median value of the concentrated distribution interval as a benchmark, set the initial calibration anchoring interval for the key anchoring dimension; calculate the degree of deviation between the initial calibration anchoring interval and the values ​​of adjacent time intervals, and count the number of adjacent time intervals with a deviation degree less than the preset deviation threshold. If the statistical proportion does not reach the preset proportion threshold, the initial calibration anchoring interval is reduced by a fixed ratio, and the degree of deviation and the proportion of quantity are recalculated. The adjustment process of reducing the initial calibration anchoring interval by a fixed ratio and calculating the degree of deviation and the proportion of quantity is repeated until the proportion of quantity reaches the preset proportion threshold. The calibration anchoring interval at this time is the optimized anchoring interval of the key anchoring dimension. Extract the current anchoring intervals of non-key anchoring dimensions in edge computing nodes, detect whether the optimized key anchoring intervals and the remaining anchoring intervals conflict in the distribution of the time dimension, adjust the conflicting anchoring intervals to make the anchoring nodes of the dimension non-overlapping and evenly distributed; at the same time, detect whether the optimized anchoring intervals are within the processing capacity of the edge computing nodes, adjust the anchoring intervals that exceed the processing capacity, so that the anchoring accuracy of the time dimension meets the preset requirements for anomaly tracing.

10. A fault early warning system based on edge computing and artificial intelligence, characterized in that, The fault warning system based on edge computing and artificial intelligence includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the fault warning method based on edge computing and artificial intelligence as described in any one of claims 1-9.

Images