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

Abstract

This invention belongs to the field of IoT data processing technology. It discloses an IoT gateway data processing method based on edge computing, comprising: acquiring topology connection data and baseline statistical data of each device; collecting real-time sensor data and normalizing it to obtain a synchronization signal vector, and generating a confidence vector based on the data quality status; performing signal correlation mining on topology connection devices, and fusing correlation metrics with topology connection relationships to construct a weighted graph structure; performing graph domain spectral decomposition on the synchronization signal vector to obtain low-frequency and high-frequency components, and correcting the high-frequency components using the confidence vector; extracting node anomaly indices and global low-frequency energy, and matching them with preset thresholds to obtain anomaly type labels; and performing differentiated data upload based on the anomaly type labels. This method effectively distinguishes between single-point faults and environmental disturbances, avoids false alarms caused by low-quality data, and improves the resource utilization efficiency of the data upload strategy.

Description

Technical Field

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

[0002] With the large-scale deployment of IoT devices in data centers, industrial plants, municipal infrastructure, agricultural greenhouses, and other scenarios, edge gateways deployed on-site need to simultaneously connect to multiple sensors such as temperature, humidity, vibration, and air pressure, leading to increasing pressure on data aggregation. To reduce the indiscriminate transmission of raw data to the cloud, edge-side preprocessing has gradually become the mainstream architectural choice—the gateway performs preliminary data filtering and anomaly detection locally, reporting only valuable information. However, in practical implementation, the effectiveness of this architecture largely depends on the gateway's ability to accurately determine which data is worth reporting, and existing technologies still have the following shortcomings in this aspect.

[0003] First, current edge gateway anomaly detection typically models each device independently, using its own historical statistical characteristics to determine whether current readings deviate from the normal range. This is acceptable in scenarios where devices are isolated, but in IoT environments, sensors often coexist in the same physical environment, and the readings of adjacent devices naturally have a correlation. If the ambient temperature suddenly rises or the air pressure changes abruptly, all surrounding sensors will respond synchronously. The single-point statistical model will confuse this global environmental fluctuation with a hardware failure of a single device—both scenarios manifest as readings deviating from the historical average from a single device's perspective, but require drastically different handling: the former only requires recording the environmental event, while the latter requires immediate device maintenance. Existing methods are unable to distinguish between these two fundamentally different sources of anomalies, leading to frequent alarms of uncertain nature for maintenance personnel and inefficient handling.

[0004] Secondly, precisely because of the lack of effective identification of anomaly types, existing gateways can only adopt a crude approach to data upload strategies—either uploading all data at fixed intervals or triggering uploads when readings exceed a preset threshold. The former indiscriminately reports normal fluctuations and abnormal data together, while the latter, although having triggering conditions, cannot determine the form and granularity of the uploaded content based on the nature and scope of the anomaly. When environmental disturbances affect all devices, uploading the raw readings of each device individually results in almost zero information increment compared to uploading only a summary information reflecting the overall trend, while consuming several times the transmission resources. Conversely, when a single device fails, only uploading the original sampling sequence of that device is truly useful for subsequent fault location; uploading all data dilutes the proportion of critical information. Existing methods lack an effective adaptation mechanism between upload strategies and anomaly types, failing to achieve reasonable data simplification during stable periods and failing to ensure that uploaded data is targeted for problem diagnosis during anomaly periods. Summary of the Invention

[0005] To overcome the aforementioned problems in the prior art, this invention proposes an IoT gateway data processing method based on edge computing to solve the above problems.

[0006] This invention provides the following technical solution: A data processing method for an IoT gateway based on edge computing includes: Obtain topology connection data and baseline statistics for each device; Real-time sensor data from each device is collected, and time window alignment and normalization are performed in conjunction with the baseline statistical data to obtain a synchronization signal vector; a confidence vector is generated based on the data quality status of each device within the current time window. Based on the historical sequence of the synchronization signal vector, for the device pairs with connection relationships in the topology connection data, signal correlation mining is performed to obtain the correlation metric. The correlation metric is then constrained and fused with the topology connection relationship to construct a weighted graph structure. Based on the weighted graph structure, the synchronization signal vector is decomposed into a graph domain spectrum, separating the signal into a low-frequency component vector and a high-frequency component vector; the high-frequency component vector is then corrected for data confidence by combining the confidence vector, resulting in a corrected high-frequency component vector. Based on the low-frequency component vector and the corrected high-frequency component vector, the node anomaly index and the global low-frequency energy are extracted respectively; the node anomaly index and the global low-frequency energy are matched with a preset discrimination threshold to obtain the anomaly type label of each device. Based on the aforementioned anomaly type label, differentiated upload processing is performed on the sensor data of each device.

[0007] Preferably, the generation of the confidence vector includes: Within the current time window, the arrival completeness of real-time sensor data of each device, the compliance of signal value range based on the baseline statistical data, and the sampling timestamp deviation are statistically analyzed to obtain multi-dimensional quality evaluation indicators for each device. The multidimensional quality assessment indicators of each device are weighted and fused to obtain the confidence value of each device in the interval [0,1]. When a device does not receive data or the amount of data received is lower than the preset integrity threshold in the current time window, the confidence value of the device is set to zero. The confidence values ​​of each device are combined to obtain the confidence vector.

[0008] Preferably, the correlation metric obtained by performing the signal correlation mining includes: From the historical sequence of the synchronization signal vector, extract the signal values ​​of each device within a preset historical window length to form a historical signal matrix; For each pair of devices in the historical signal matrix that have a topological connection relationship, calculate the absolute value of the linear correlation metric between the corresponding row sequences; set the linear correlation metric of the device pairs whose absolute value is lower than a preset correlation threshold to zero to obtain the correlation metric.

[0009] Preferably, the step of constraining and fusing the correlation metric with the topological connectivity relationship to construct a weighted graph structure includes: The correlation metric of each device pair is multiplied element-wise with the topological connection data to obtain a weighted adjacency matrix, such that the weight of device pairs that have no topological connection relationship is zero. The weighted degree of each device is calculated based on the weighted adjacency matrix, a degree matrix is ​​constructed, and the weighted adjacency matrix and the degree matrix are combined to obtain the weighted graph structure.

[0010] Preferably, the step of performing graph-domain spectral decomposition on the synchronization signal vector based on the weighted graph structure to separate the signal into low-frequency component vectors and high-frequency component vectors includes: A normalized graph Laplacian operator is constructed based on the weighted graph structure; scaling parameters are determined according to the topological properties of the weighted graph structure, and the normalized graph Laplacian operator is scaled to obtain a scaled graph Laplacian operator; The synchronization signal vector and the scaling graph Laplacian operator are recursively operated on by a preset order to obtain a preset number of basis vectors; preset weight coefficients are applied to the preset number of basis vectors and superimposed to obtain a low-frequency component vector; the difference between the synchronization signal vector and the low-frequency component vector is taken to obtain a high-frequency component vector.

[0011] Preferably, the step of combining the confidence vector with the high-frequency component vector to perform data confidence correction on the high-frequency component vector to obtain the corrected high-frequency component vector includes: Using the confidence values ​​corresponding to each device in the confidence vector as weights, each component in the high-frequency component vector is scaled element-wise; the component corresponding to a device with a confidence value of zero is scaled to zero; the scaled components of each device are combined to obtain the corrected high-frequency component vector.

[0012] Preferably, the extraction of node anomaly index and global low-frequency energy respectively includes: For devices with a confidence value greater than zero, the node anomaly index of each device is determined based on the difference between the corrected high-frequency component vector and the low-frequency component vector in adjacent time windows. The global low-frequency energy is determined based on the weighted summation of the components of each device in the low-frequency component vector.

[0013] Preferably, the step of matching the node anomaly index with the global low-frequency energy and a preset discrimination threshold to obtain anomaly type labels for each device includes: When the node anomaly index of a device exceeds the first discrimination threshold and the global low-frequency energy does not exceed the second discrimination threshold, a single-point fault label is assigned to the device. When the global low-frequency energy exceeds the second discrimination threshold, an environmental disturbance label is assigned to all devices with a confidence value greater than zero. When the node anomaly index does not exceed the first discrimination threshold and the global low-frequency energy does not exceed the second discrimination threshold, a normal label is assigned to each device.

[0014] Preferably, the method for determining the preset discrimination threshold includes: Perform the graph domain spectrum decomposition on the historical normal operation data of each device to obtain the node anomaly index sequence and the global low-frequency energy sequence during normal operation, and statistically analyze the distribution characteristics of the sequence; Perform the graph domain spectrum decomposition on the historical alarm data with labeled anomaly types to obtain the node anomaly index and global low-frequency energy during the anomaly period; Based on the distribution characteristics during normal operation and the corresponding values ​​during abnormal periods, the first discrimination threshold and the second discrimination threshold are determined respectively.

[0015] Preferably, the step of performing differentiated uploading processing on the sensor data of each device based on the anomaly type label includes: When there is a device with the anomaly type label "single point of failure", upload the original sensor data of the device within the current time window and generate a single point of failure alarm; When all devices are labeled as having an environmental disturbance and there is no single point of failure, the low-frequency component vector is used as the aggregated representative of all device data for uploading, and an environmental disturbance notification is generated; when the duration of the environmental disturbance exceeds a preset duration threshold, the independent uploading of the original sensor data of each device is restored. When all devices have normal anomaly type labels, the sensor data of each device is aggregated and uploaded according to the preset compression strategy.

[0016] This invention provides a data processing method for IoT gateways based on edge computing, which has the following beneficial effects: This invention constructs an adaptive weighted graph structure under topological constraints and performs graph-domain spectral decomposition to separate the signals of each device into low-frequency and high-frequency components. Combined with a confidence correction mechanism, it suppresses false anomaly signals from low-quality data. Finally, based on a dual-indicator discrimination of node anomaly index and global low-frequency energy, it effectively distinguishes between single-point faults and environmental disturbances. Compared with existing single-point statistical models, this invention avoids conflating device hardware faults with global environmental fluctuations, enabling edge gateways to provide anomaly type information with clear semantics for operational decisions. Simultaneously, the confidence-weighted correction effectively suppresses false alarms caused by missing or degraded data, improving the reliability of anomaly detection.

[0017] Building upon the ability to differentiate between anomaly types, this invention adaptively binds anomaly type tags to data upload strategies. Based on three states—single point of failure, environmental disturbance, and normal operation—it performs raw data upload, low-frequency component aggregation upload, and compressed summary upload, respectively. This differentiated upload mechanism matches the format and granularity of uploaded data with the nature of the anomaly event. This ensures the timing details required for fault diagnosis while avoiding the uploading of highly correlated redundant data during environmental disturbances. It guarantees targeted diagnosis during anomaly periods while optimizing transmission resources during normal operation, improving the targeted nature and resource utilization efficiency of edge gateway data uploads under different operating states. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an IoT gateway data processing method based on edge computing according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0020] This embodiment uses an IoT gateway deployed in a data center as an example. The gateway connects to several temperature and humidity sensors distributed in different locations and cabinets within the data center. These sensors are interconnected via an industrial Ethernet network. The gateway is responsible for aggregating data from all sensors and reporting it to the cloud. The specific implementation of this invention will be described in detail below with reference to the accompanying drawings.

[0021] Please see Figure 1 In this embodiment, an IoT gateway data processing method based on edge computing includes: Step S1: Obtain the topology connection data and baseline statistics of each device; In this embodiment, topology connection data is obtained by the gateway reading a pre-set engineering configuration file during the initial device registration. The configuration file records sensor devices that are physically in the same cabinet, in the same cooling aisle, or have control linkages in the form of device ID pairs, and stores them in the form of an adjacency matrix. A value of 1 in the matrix indicates a connection between the corresponding device pairs, while a value of 0 indicates no connection. Baseline statistical data is taken from the gateway's local database, storing the mean and standard deviation of sensor signals for each device under historical normal operating conditions, as well as the rated sampling data volume for each device within a preset time window. For example, the baseline mean for temperature sensor T01 is 24.5℃, the standard deviation is 0.8℃, and the rated sampling period is 1s. Baseline statistical data is calculated offline based on sufficient historical normal operating data during the system initialization phase and stored locally on the gateway. It is called as needed during subsequent operation for subsequent normalization processing and data integrity assessment.

[0022] Step S2: Collect real-time sensor data from each device, and perform time window alignment and normalization processing in conjunction with the baseline statistical data to obtain a synchronization signal vector; generate a confidence vector based on the data quality status of each device within the current time window. In this embodiment, the time window length is set to 60 seconds, meaning the gateway aggregates the sampling data from each device every 60 seconds and executes subsequent processing. In actual deployment, this can be adjusted appropriately based on sensor sampling frequency and anomaly response time requirements. At the end of each time window, the data collected by each device within that window is aligned: if a device reports multiple times within the window, the arithmetic mean is used as the representative value; if it does not report once, it is marked as empty. Subsequently, the representative value is z-score normalized using the historical mean and standard deviation of each device in the baseline statistics, converting sensor signals of different dimensions such as temperature and humidity into dimensionless values, eliminating dimensional differences, and making the signal values ​​of different types of sensors comparable. It should be noted that when the historical standard deviation of a device is close to zero, it indicates that the device's signal has been stable with almost no fluctuations under historical normal operating conditions. In this case, the normalization denominator uses a preset minimum standard deviation lower limit to avoid numerical overflow. The normalized representative values ​​of each device are arranged in order of device number to form the synchronization signal vector for the current time window. Each component of the vector corresponds to the current normalized signal value of a device.

[0023] The generation of the confidence vector includes: Within the current time window, the arrival completeness of real-time sensor data of each device, the compliance of signal value range based on the baseline statistical data, and the sampling timestamp deviation are statistically analyzed to obtain multi-dimensional quality evaluation indicators for each device. The multidimensional quality assessment indicators of each device are weighted and fused to obtain the confidence value of each device in the interval [0,1]. When a device does not receive data or the amount of data received is lower than the preset integrity threshold in the current time window, the confidence value of the device is set to zero. The confidence values ​​of each device are combined to obtain the confidence vector.

[0024] In this embodiment, arrival integrity is measured by the ratio of the actual amount of sampled data arriving within the current window to the rated amount of sampled data in the baseline statistics; signal value range compliance is measured by whether the normalized signal value falls within the reasonable range determined by the baseline statistics, with a larger deviation indicating lower compliance; sampling timestamp deviation is measured by the maximum deviation between the timestamp of each sampling point within the window and the expected sampling time, with a larger deviation indicating lower timeliness. The evaluation results of the above three dimensions are weighted and fused according to preset weights. In this embodiment, the fusion weights of arrival integrity, signal value range compliance, and sampling timestamp deviation are set to 0.5, 0.3, and 0.2, respectively. Arrival integrity has the highest weight because complete data loss is the most serious quality problem, followed by signal value range compliance. Sampling timestamp deviation has a relatively small impact in most industrial scenarios. In actual deployment, the weights of each dimension can be adjusted according to the specific sensor type and communication environment. The weighted fusion result is normalized to the [0,1] interval to obtain the confidence value of each device. The higher the confidence value, the more reliable the current data of the device. In this embodiment, the preset integrity threshold is set to 50% of the rated sampled data volume. When a device fails to receive any data or receives less than 50% of the rated data within the current time window, the confidence value of that device is directly set to zero, indicating that the current data of that device is unavailable. The confidence values ​​of each device are arranged in the same device order as the synchronization signal vector to form the confidence vector for the current time window. The confidence vector corresponds one-to-one with the synchronization signal vector, providing a basis for differentiated processing of device data with different confidence levels in subsequent steps.

[0025] Step S3: Based on the historical sequence of the synchronization signal vector, perform signal correlation mining on the device pairs with connection relationships in the topology connection data to obtain correlation metrics, and fuse the correlation metrics with the topology connection relationships to construct a weighted graph structure; The correlation metrics obtained from the execution signal correlation mining include: From the historical sequence of the synchronization signal vector, extract the signal values ​​of each device within a preset historical window length to form a historical signal matrix; For each pair of devices in the historical signal matrix that have a topological connection relationship, calculate the absolute value of the linear correlation metric between the corresponding row sequences; set the linear correlation metric of the device pairs whose absolute value is lower than a preset correlation threshold to zero to obtain the correlation metric.

[0026] In this embodiment, the gateway caches the synchronization signal vectors of the most recent 30 time windows, forming a historical signal matrix with devices as rows and time windows as columns. If the historical window is too short, the statistical reliability of the correlation calculation is insufficient; if it is too long, the response to recent changes in the correlation between devices is sluggish. In this embodiment, the 30 windows correspond to approximately 30 minutes of historical data, achieving a balance between statistical stability and dynamic response. In practical applications, this can be adjusted appropriately based on the number of devices and the signal change cycle. For each pair of devices with topological connections in the matrix, the absolute value of the Pearson correlation coefficient between the corresponding row sequences is calculated as a linear correlation measure. The larger the absolute value, the higher the linear coordination degree of the two device signals. The purpose of taking the absolute value is that both positive and negative correlations in the graph structure indicate a linkage between the two device signals, and both should be assigned higher edge weights. In this embodiment, the preset correlation threshold is set to 0.3. The linear correlation measure of device pairs with an absolute value lower than 0.3 is set to zero to filter out accidental correlations that may occur when the historical sequences are short, ensuring that the edge weights of the graph structure are statistically significant.

[0027] The step of constraining and fusing the correlation metric with the topological connectivity relationship to construct a weighted graph structure includes: The correlation metric of each device pair is multiplied element-wise with the topological connection data to obtain a weighted adjacency matrix, such that the weight of device pairs that have no topological connection relationship is zero. The weighted degree of each device is calculated based on the weighted adjacency matrix, a degree matrix is ​​constructed, and the weighted adjacency matrix and the degree matrix are combined to obtain the weighted graph structure.

[0028] In this embodiment, the linear correlation metrics of each device pair are arranged into a matrix of the same dimension as the topological connection data. This matrix is ​​then multiplied element-wise with the adjacency matrix of the topological connection data to obtain a weighted adjacency matrix, forcing the weights of device pairs without physical connections to zero. The significance of this constraint fusion operation is that purely data-driven correlation calculations may produce false associations without physical basis due to accidental factors. By masking the correlation metrics with topological connections, it ensures that edges in the graph structure exist only between device pairs that are physically connected, giving the weighted graph structure both data adaptability and physical interpretability. Based on the weighted adjacency matrix, the weighted degree of each device (i.e., the sum of the edge weights of the device and all its neighboring devices) is calculated. A degree matrix with the weighted degree of each device as its diagonal element is constructed. The weighted adjacency matrix and the degree matrix are combined to obtain the weighted graph structure. The introduction of the degree matrix provides the necessary structural information for the subsequent construction of the normalized graph Laplacian operator. The weighted graph structure is periodically updated by repeatedly performing the above calculation process according to a preset update cycle to reflect the dynamic changes in the signal correlation between devices over time.

[0029] Step S4: Perform graph domain spectral decomposition on the synchronization signal vector based on the weighted graph structure to separate the signal into low-frequency component vectors and high-frequency component vectors; combine the confidence vector to perform data confidence correction on the high-frequency component vector to obtain the corrected high-frequency component vector. The step of performing graph-domain spectral decomposition on the synchronization signal vector based on the weighted graph structure to separate the signal into low-frequency component vectors and high-frequency component vectors includes: A normalized graph Laplacian operator is constructed based on the weighted graph structure; scaling parameters are determined according to the topological properties of the weighted graph structure, and the normalized graph Laplacian operator is scaled to obtain a scaled graph Laplacian operator; The synchronization signal vector and the scaling graph Laplacian operator are recursively operated on by a preset order to obtain a preset number of basis vectors; preset weight coefficients are applied to the preset number of basis vectors and superimposed to obtain a low-frequency component vector; the difference between the synchronization signal vector and the low-frequency component vector is taken to obtain a high-frequency component vector.

[0030] In this embodiment, before performing graph domain spectral decomposition, the components corresponding to devices with a confidence value of zero in the synchronization signal vector are set to zero to prevent abnormal signal values ​​of these devices from affecting the calculation of low-frequency components of topological neighboring devices through neighborhood diffusion in the graph structure. The normalized graph Laplacian operator is constructed from the degree matrix and the weighted adjacency matrix according to the standard graph signal processing method. Its eigenvalues ​​reflect the frequency distribution characteristics of the signal in the graph structure. The smaller the eigenvalue, the lower the frequency of the graph signal component, that is, the smoother the change between adjacent devices. The scaling parameter is determined based on the largest eigenvalue of the normalized graph Laplacian operator. The normalized graph Laplacian operator is linearly scaled to map its eigenvalue range to the interval [-1,1] to obtain the scaled graph Laplacian operator, so that subsequent multi-order recursive operations remain numerically stable.

[0031] The synchronization signal vector is recursively processed by the scaling graph Laplacian operator at a preset order. In this embodiment, the preset order of the recursive operation is set to 3, which means calculating four basis vectors from the zeroth to the third order. Higher orders result in smoother low-frequency components, but also increase computational complexity. The third order strikes a reasonable balance between decomposition effect and computational cost. In practical applications, this can be adjusted appropriately based on device scale and gateway computing power. Specifically, the zeroth-order basis vector is the synchronization signal vector itself; the first-order basis vector is the result of the scaling graph Laplacian operator applied to the synchronization signal vector; the second-order basis vector is the result of the scaling graph Laplacian operator applied to the first-order basis vector; and the third-order basis vector is the result of the scaling graph Laplacian operator applied to the second-order basis vector. Each order of basis vector captures the diffusion characteristics of the synchronization signal vector within different order neighborhoods in the graph structure. Lower-order basis vectors reflect more local information, while higher-order basis vectors reflect more global cooperative information. Preset weighting coefficients are applied to the basis vectors of each order and then superimposed to obtain the low-frequency component vector. The weighting coefficients decrease with the order to suppress high-frequency components contained in higher-order basis vectors, making the superposition result biased towards the low-frequency signal components, i.e., the parts of each device signal that change in tandem with its topological neighbors. In this embodiment, the weighting coefficients corresponding to the four orders are set to 0.4, 0.3, 0.2, and 0.1, respectively. The high-frequency component vector is obtained by subtracting the synchronization signal vector from the low-frequency component vector. The high-frequency component vector captures the local abrupt changes in each device signal that cannot be explained by the tandem trends of its topological neighbors.

[0032] The step of combining the confidence vector with the high-frequency component vector to perform data confidence correction, resulting in the corrected high-frequency component vector, includes: Using the confidence values ​​corresponding to each device in the confidence vector as weights, each component in the high-frequency component vector is scaled element-wise; the component corresponding to a device with a confidence value of zero is scaled to zero; the scaled components of each device are combined to obtain the corrected high-frequency component vector.

[0033] In this embodiment, the confidence values ​​corresponding to each device in the confidence vector are used as weights to perform element-wise scaling on each component in the high-frequency component vector. The lower the confidence value, the greater the compression of the high-frequency component of the corresponding device; the high-frequency component corresponding to a device with a confidence value of zero is naturally zero after scaling. The physical meaning of this correction is that the high-frequency component represents a local abrupt change in a device's signal relative to its topological neighbors. If the data quality of the device is low (such as incomplete sampling, large timestamp deviation, or the signal exceeding the reasonable value range), the high frequency component is likely due to data quality problems rather than real physical anomalies. Without processing, subsequent steps will misjudge it as a single point of failure. Through confidence-weighted correction, the false high-frequency components introduced by low-quality data are suppressed, so that subsequent anomaly detection is based on reliable signals. It should be noted that the low-frequency component vector does not need such correction because the low-frequency component is obtained by weighted fusion of multiple device signals through a graph structure. The impact of low data quality of a certain device on the fusion result has been diluted by the signals of other topological neighbor devices, which has a natural anti-interference capability. The scaled components of each device are combined to obtain the corrected high-frequency component vector.

[0034] Step S5: Based on the low-frequency component vector and the corrected high-frequency component vector, extract the node anomaly index and the global low-frequency energy respectively; perform anomaly pattern matching between the node anomaly index and the global low-frequency energy and a preset discrimination threshold to obtain the anomaly type label of each device; The extraction of node anomaly index and global low-frequency energy respectively includes: For devices with a confidence value greater than zero, the node anomaly index of each device is determined based on the difference between the corrected high-frequency component vector and the low-frequency component vector in adjacent time windows. The global low-frequency energy is determined based on the weighted summation of the components of each device in the low-frequency component vector.

[0035] In this embodiment, for devices with a confidence value greater than zero, the absolute value of the difference between the corrected high-frequency component and the previous time window, and the absolute value of the difference between the low-frequency component and adjacent time windows are calculated. The weighted sum of these two values ​​is used as the node anomaly index for the device. In this embodiment, the weights for the absolute values ​​of the high-frequency and low-frequency component differences are set to 0.7 and 0.3, respectively. Changes in the high-frequency component contribute more to the node anomaly index because the most direct manifestation of a single-point fault is a local mutation, i.e., a significant change in the high-frequency component. The change in the low-frequency component serves as an auxiliary criterion to enhance the index's sensitivity to gradual anomalies. The node anomaly index comprehensively measures the degree of local mutation in the device's signal and the magnitude of its co-current trend. A higher index indicates a more significant deviation between the device's signal behavior within the current time window and historical patterns. For devices with a confidence value greater than zero, the squares of their low-frequency components are weighted by confidence, summed, and the square root is taken to obtain the global low-frequency energy. The purpose of squaring the value is to make the contribution of larger low-frequency components to the global low-frequency energy more significant, thereby enhancing the sensitivity to environmental disturbances with large cooperative offset amplitudes. The global low-frequency energy reflects the overall strength of the cooperative change trend of the entire device network within the current time window. When multiple devices are synchronously affected by environmental factors and produce cooperative offsets, the low-frequency components rise as a whole, and the global low-frequency energy increases significantly accordingly.

[0036] The step of matching the node anomaly index with the global low-frequency energy and a preset discrimination threshold to obtain anomaly type labels for each device includes: When the node anomaly index of a device exceeds the first discrimination threshold and the global low-frequency energy does not exceed the second discrimination threshold, a single-point fault label is assigned to the device. When the global low-frequency energy exceeds the second discrimination threshold, an environmental disturbance label is assigned to all devices with a confidence value greater than zero. When the node anomaly index does not exceed the first discrimination threshold and the global low-frequency energy does not exceed the second discrimination threshold, a normal label is assigned to each device.

[0037] It should be noted that the physical basis of the above dual-index discrimination mechanism is as follows: A single-point fault manifests as a localized abrupt change in the signal of a certain device, with a significantly elevated high-frequency component and the corresponding node anomaly index exceeding the first discrimination threshold. However, other device signals are unaffected, the low-frequency component remains at a normal level overall, and the global low-frequency energy does not exceed the second discrimination threshold. Environmental disturbances manifest as synchronous signal shifts in multiple devices affected by the same environmental factor, with an overall increase in low-frequency components and the global low-frequency energy exceeding the second discrimination threshold. In this case, the differences between the signals of each device relative to its topological neighbors are not significant, the high-frequency component is not significant, and single-point fault discrimination is not triggered. This mechanism, based on the spatial distribution characteristics of signals, distinguishes between local anomalies of a single device and global anomalies across devices, solving the problem of existing single-point detection methods conflating two fundamentally different types of anomalies.

[0038] The method for determining the preset discrimination threshold includes: Perform the graph domain spectrum decomposition on the historical normal operation data of each device to obtain the node anomaly index sequence and the global low-frequency energy sequence during normal operation, and statistically analyze the distribution characteristics of the sequence; Perform the graph domain spectrum decomposition on the historical alarm data with labeled anomaly types to obtain the node anomaly index and global low-frequency energy during the anomaly period; Based on the distribution characteristics during normal operation and the corresponding values ​​during abnormal periods, the first discrimination threshold and the second discrimination threshold are determined respectively.

[0039] In this embodiment, after performing graph domain spectral decomposition on historical normal operation data, the 99th percentile of the node anomaly index sequence and the global low-frequency energy sequence during normal operation can be used as candidate values ​​for the corresponding discrimination thresholds. Then, graph domain spectral decomposition is performed on historical alarm data with anomaly types already labeled to verify whether the node anomaly index and the global low-frequency energy during each anomaly period exceed the above candidate values. The thresholds are adjusted appropriately based on the verification results to ensure that the two types of thresholds have effective distinguishing ability between normal and abnormal conditions, and the first discrimination threshold and the second discrimination threshold are determined respectively.

[0040] Step S6: Based on the anomaly type label, perform differentiated upload processing on the sensor data of each device.

[0041] The step of performing differentiated upload processing on the sensor data of each device based on the anomaly type label includes: When there is a device with the anomaly type label "single point of failure", upload the original sensor data of the device within the current time window and generate a single point of failure alarm; When all devices are labeled as having an environmental disturbance and there is no single point of failure, the low-frequency component vector is used as the aggregated representative of all device data for uploading, and an environmental disturbance notification is generated; when the duration of the environmental disturbance exceeds a preset duration threshold, the independent uploading of the original sensor data of each device is restored. When all devices have normal anomaly type labels, the sensor data of each device is aggregated and uploaded according to the preset compression strategy.

[0042] In this embodiment, the differentiated upload strategy is designed based on the fundamental differences in information requirements corresponding to different anomaly types. When a single point of failure exists, the cloud needs to obtain the complete original sampling sequence of the device for fault location and root cause analysis. Compression or aggregation processing would lose key timing details. Therefore, the original sensor data of the device within the current time window is uploaded, and a single point of failure alarm is generated. When all devices are experiencing environmental disturbances and there are no single points of failure, the changes in the signals of each device are highly coordinated. The low-frequency component vector has fully captured this global trend. The low-frequency component vector is used as the aggregated representative of all device data for upload. This significantly reduces the amount of uploaded data while retaining the core information of the environmental event, and generates an environmental disturbance notification to alert maintenance personnel. It should be noted that when the duration of the environmental disturbance exceeds a preset duration threshold, the independent upload of the original sensor data of each device is restored to prevent the gradual changes in the individual device status from being masked by long-term replacement with aggregated data. When all devices are normal, the sensor data of each device is aggregated and uploaded according to a preset compression strategy, achieving reasonable control of data traffic under normal operating conditions.

[0043] As shown above, this embodiment constructs an adaptive weighted graph structure under topological constraints, organically combining the physical connection relationships between IoT devices with data-driven signal correlations, enabling the graph structure to possess both physical interpretability and data adaptability. Based on this, graph domain spectral decomposition separates the signals of each device into co-changing low-frequency components and locally abrupt high-frequency components. A confidence correction mechanism is used to suppress false anomaly signals introduced by low-quality data. Furthermore, a dual-indicator approach using node anomaly index and global low-frequency energy effectively distinguishes between single-point faults and environmental disturbances. Finally, data upload strategies are differentiated based on anomaly type labels, ensuring the integrity of anomaly event data while simplifying data under normal and disturbed states. This effectively solves the problems of insufficient anomaly identification accuracy and lack of anomaly type adaptability in existing edge gateways.

[0044] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0045] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0046] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An edge-computing-based Internet of Things gateway data processing method, characterized in that, include: Obtain topology connection data and baseline statistics for each device; Real-time sensor data from each device is collected, and time window alignment and normalization are performed in conjunction with the baseline statistical data to obtain a synchronization signal vector; a confidence vector is generated based on the data quality status of each device within the current time window. Based on the historical sequence of the synchronization signal vector, for the device pairs with connection relationships in the topology connection data, signal correlation mining is performed to obtain the correlation metric. The correlation metric is then constrained and fused with the topology connection relationship to construct a weighted graph structure. Based on the weighted graph structure, the synchronization signal vector is subjected to graph domain spectral decomposition to separate the signal into low-frequency component vectors and high-frequency component vectors; The high-frequency component vector is corrected for data confidence by combining the confidence vector to obtain the corrected high-frequency component vector. Based on the low-frequency component vector and the corrected high-frequency component vector, the node anomaly index and global low-frequency energy are extracted respectively. The node anomaly index is matched with the global low-frequency energy and a preset discrimination threshold to obtain the anomaly type label for each device. Based on the aforementioned anomaly type label, differentiated upload processing is performed on the sensor data of each device.

2. The IoT gateway data processing method based on edge computing according to claim 1, characterized in that, The generation of the confidence vector includes: Within the current time window, the arrival completeness of real-time sensor data of each device, the compliance of signal value range based on the baseline statistical data, and the sampling timestamp deviation are statistically analyzed to obtain multi-dimensional quality evaluation indicators for each device. The multidimensional quality assessment indicators of each device are weighted and fused to obtain the confidence value of each device in the interval [0,1]. When a device does not receive data or the amount of data received is lower than the preset integrity threshold in the current time window, the confidence value of the device is set to zero. The confidence values ​​of each device are combined to obtain the confidence vector.

3. The IoT gateway data processing method based on edge computing according to claim 2, characterized in that, The correlation metrics obtained from the execution signal correlation mining include: From the historical sequence of the synchronization signal vector, extract the signal values ​​of each device within a preset historical window length to form a historical signal matrix; For each pair of devices in the historical signal matrix that have a topological connection relationship, calculate the absolute value of the linear correlation metric between the corresponding row sequences; set the linear correlation metric of the device pairs whose absolute value is lower than a preset correlation threshold to zero to obtain the correlation metric.

4. The IoT gateway data processing method based on edge computing according to claim 3, characterized in that, The step of constraining and fusing the correlation metric with the topological connectivity relationship to construct a weighted graph structure includes: The correlation metric of each device pair is multiplied element-wise with the topological connection data to obtain a weighted adjacency matrix, such that the weight of device pairs that have no topological connection relationship is zero. The weighted degree of each device is calculated based on the weighted adjacency matrix, a degree matrix is ​​constructed, and the weighted adjacency matrix and the degree matrix are combined to obtain the weighted graph structure.

5. The IoT gateway data processing method based on edge computing according to claim 4, characterized in that, The step of performing graph-domain spectral decomposition on the synchronization signal vector based on the weighted graph structure to separate the signal into low-frequency component vectors and high-frequency component vectors includes: A normalized graph Laplacian operator is constructed based on the weighted graph structure; scaling parameters are determined according to the topological properties of the weighted graph structure, and the normalized graph Laplacian operator is scaled to obtain a scaled graph Laplacian operator; The synchronization signal vector and the scaling graph Laplacian operator are recursively operated on by a preset order to obtain a preset number of basis vectors; preset weight coefficients are applied to the preset number of basis vectors and superimposed to obtain a low-frequency component vector; the difference between the synchronization signal vector and the low-frequency component vector is taken to obtain a high-frequency component vector.

6. The IoT gateway data processing method based on edge computing according to claim 5, characterized in that, The step of combining the confidence vector with the high-frequency component vector to perform data confidence correction, resulting in the corrected high-frequency component vector, includes: Using the confidence values ​​corresponding to each device in the confidence vector as weights, each component in the high-frequency component vector is scaled element-wise; the component corresponding to a device with a confidence value of zero is scaled to zero; the scaled components of each device are combined to obtain the corrected high-frequency component vector.

7. The IoT gateway data processing method based on edge computing according to claim 6, characterized in that, The extraction of node anomaly index and global low-frequency energy respectively includes: For devices with a confidence value greater than zero, the node anomaly index of each device is determined based on the difference between the corrected high-frequency component vector and the low-frequency component vector in adjacent time windows. The global low-frequency energy is determined based on the weighted summation of the components of each device in the low-frequency component vector.

8. The IoT gateway data processing method based on edge computing according to claim 7, characterized in that, The step of matching the node anomaly index with the global low-frequency energy and a preset discrimination threshold to obtain anomaly type labels for each device includes: When the node anomaly index of a device exceeds the first discrimination threshold and the global low-frequency energy does not exceed the second discrimination threshold, a single-point fault label is assigned to the device. When the global low-frequency energy exceeds the second discrimination threshold, an environmental disturbance label is assigned to all devices with a confidence value greater than zero. When the node anomaly index does not exceed the first discrimination threshold and the global low-frequency energy does not exceed the second discrimination threshold, a normal label is assigned to each device.

9. The IoT gateway data processing method based on edge computing according to claim 8, characterized in that, The method for determining the preset discrimination threshold includes: Perform the graph domain spectrum decomposition on the historical normal operation data of each device to obtain the node anomaly index sequence and the global low-frequency energy sequence during normal operation, and statistically analyze the distribution characteristics of the sequence; Perform the graph domain spectrum decomposition on the historical alarm data with labeled anomaly types to obtain the node anomaly index and global low-frequency energy during the anomaly period; Based on the distribution characteristics during normal operation and the corresponding values ​​during abnormal periods, the first discrimination threshold and the second discrimination threshold are determined respectively.

10. The IoT gateway data processing method based on edge computing according to claim 8, characterized in that, The step of performing differentiated upload processing on the sensor data of each device based on the anomaly type label includes: When there is a device with the anomaly type label "single point of failure", upload the original sensor data of the device within the current time window and generate a single point of failure alarm; When all devices are labeled as having an environmental disturbance and there is no single point of failure, the low-frequency component vector is used as the aggregated representative of all device data for uploading, and an environmental disturbance notification is generated; when the duration of the environmental disturbance exceeds a preset duration threshold, the independent uploading of the original sensor data of each device is restored. When all devices have normal anomaly type labels, the sensor data of each device is aggregated and uploaded according to the preset compression strategy.

Images