Equipment operation state anomaly detection method based on edge computing

Abstract

The invention provides an equipment operation state anomaly detection method based on edge computing. According to the method, the accuracy, the robustness and the real-time response capability of equipment anomaly detection are improved. The method comprises the following steps: firstly, setting a channel confidence coefficient screening mechanism at an edge calculation end, and eliminating an abnormal transmission channel with drifting or fluctuation to guarantee data quality; then, fusing and modeling a structural dependency relationship among multiple channels by utilizing a graph neural network and a self-attention mechanism, and extracting a global context sensing vector; and finally, constructing a dynamic criterion system based on a multi-task state index, and realizing efficient anomaly judgment in combination with state difference scoring. According to the invention, through the integration of edge-side high-quality data screening, depth feature modeling and a fusion type detection mechanism, the intelligent level and field adaptability of equipment operation state abnormity identification are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of device running state anomaly detection, and particularly relates to a device running state anomaly detection method based on edge computing. BACKGROUND

[0002] With the rapid development of industrial Internet of Things, the monitoring and anomaly detection of device running state have become a key link to ensure the safe and efficient operation of production systems. Traditional monitoring methods mostly rely on centralized cloud computing mode, which uploads running data collected by various sensors to a remote cloud server for processing and analysis by a backend model. However, such methods often face problems such as high bandwidth consumption, large data transmission delay, and insufficient rapid response capability to anomalies.

[0003] To overcome the shortcomings of traditional technologies, existing technologies introduce edge computing into device anomaly detection, realizing remote unified monitoring and centralized management of device state. Compared with traditional cloud computing mode, the existing technology has the advantages of having a certain degree of intelligent judgment capability, being able to identify some common anomalies through preset rules, and supporting remote unified management of multiple devices. However, the existing technology does not introduce a channel quality screening mechanism, and cannot identify and eliminate device running data that has deviation, drift, or distortion, affecting the reliability of device anomaly judgment. Moreover, the existing technology does not model the correlation structure between channels, and can only make isolated judgments on a single channel, making it difficult to identify systematic anomalies. SUMMARY

[0004] To solve the technical problems mentioned in the background art, the present application proposes a device running state anomaly detection method based on edge computing.

[0005] To this end, the technical solution adopted by the present application is as follows: A device running state anomaly detection method based on edge computing, the method comprising: S1, setting a channel confidence screening mechanism in the edge computing end, evaluating the confidence of the transmission channel of device running data, eliminating transmission channels that do not meet the quality requirements, and expanding the transmission matrix constructed by the running data of the remaining transmission channels into a three-dimensional feature tensor; S2, modeling and fusing the three-dimensional feature tensor based on a graph neural network and a self-attention mechanism, generating a channel context perception vector, and further constructing a multi-task state indicator; S3, designing a fusion type anomaly judgment mechanism in the edge computing end, constructing a historical state sequence based on the multi-task state indicator and generating a criterion parameter, and combining a state variation score to perform anomaly detection.

[0006] Further, the operation data is collected by M sensors, and the sensors and transmission channels are one-to-one bound, and the operation data is transmitted to the edge computing end through the corresponding transmission channel.

[0007] Further, the specific steps of the channel confidence screening mechanism are, 1) At time step t, collect the operation data of M transmission channels in T time windows, and the transmission matrix is represented as: Wherein, is the operation data of M channels t time step in time window T; is the operation data of the mth channel in time window T, represented as: Wherein, represents the operation data of the mth channel at the start of the time window; 2) Set the distribution consistency index, and define the KL divergence for the mth channel, the formula is: Wherein, is the KL divergence of channel m; is the a-th interval of the operation data in the histogram bin; is the probability of channel m in the a-th interval on the historical period; is the probability of channel m in the a-th interval in the current time window; 3) Set the fluctuation intensity index, calculate the coefficient of variation value, the formula is: Wherein, is the coefficient of variation value; is the standard deviation of channel m in the current time window; is the mean value of channel m in the current time window, is a positive number; 4) Combine the distribution consistency index and the fluctuation intensity index to construct the channel confidence score, represented as: Wherein, is the confidence score of channel m; and is a weight factor.

[0008] Further, according to the comparison between the confidence score and the set confidence threshold , the quality of all transmission channels is judged, and when , then the quality of channel m does not meet the requirements and needs to be eliminated. At the same time, the corresponding sensor should be replaced; The operating data of the remaining transmission channels form the reconstructed transmission matrix, which is expressed as: Where N is the number of channels that meet the quality requirements.

[0009] Furthermore, based on the graph neural network, a channel graph is constructed , the nodes of the channel graph correspond to transmission channels, and the edges are defined by edge weights, which can be expressed as: in, and are the running data of the bth and cth channels in the time window T respectively; is the Pearson correlation coefficient; The three-dimensional feature tensor is flattened and mapped into a node feature vector using the following formula: in, represents the dimensionality reduction function; It is the feature vector of the transmission channel n in the 0th layer of the graph neural network, and the 0th layer of the graph neural network is the input layer.

[0010] Furthermore, through a multi-layer graph convolutional neural network, the feature vectors between nodes are propagated and aggregated on the channel graph. The formula is: in, It is the first Feature vector representation of node n in the layer; is a nonlinear activation function; is the set of neighbor nodes of node n; is the edge weight between node n and its neighbor node c; and are the degrees of node n and node c, i.e. the number of neighbors; It is The graph convolution weight matrix of the layer; It is a transmission channel In the graph neural network The feature vector of the layer; According to the propagation and aggregation operations, the final feature vector sequence of the transmission channel is obtained, which is expressed as: in, It is the embedded feature vector of transmission channel n in the Lth layer of the graph neural network.

[0011] Further, the self-attention mechanism generates a channel context awareness vector sequence by fusing the global dependency between the embedding feature vectors is expressed as: wherein, is the channel context awareness vector of the transmission channel n; According to the channel context awareness vector sequence, a multi-task state index is calculated, respectively, state score , score trend change rate and state reconstruction residual .

[0012] Further, the historical state sequence is constructed by collecting P historical state scores, score trend change rates and state reconstruction residuals, The generation formula of the criterion parameter is: wherein, , and are the weighted factors of state score, score trend change rate and state reconstruction residual, respectively; is the abnormality judgment threshold of the current time step t; is a light neural network model; is a historical state sequence; The residual value between the current time step t and the P historical time steps of the fused channel context awareness vector sequence is calculated, and the state difference score is obtained through light convolution network and pooling operation; The criterion parameter and the state difference score are fused to calculate the final abnormality score Q, and the formula is: wherein, is the fusion weight of the state variation score; when , it indicates that the running state of the device is abnormal.

[0013] Compared with the prior art, the advantages of the present application are: 1. The present application first introduces a channel confidence screening mechanism at the edge computing end, which can effectively eliminate low-quality transmission channels with drift, fluctuation or distortion, thereby significantly improving the input quality and stability of subsequent anomaly detection. The mechanism guarantees the reliability of the running data from the source, ensures that the detection result is more accurate and reliable, and the response is fast in the edge deployment, which is suitable for high-frequency sampling and high-timeliness scenarios in industrial field.

[0014] 2. The present invention constructs a channel graph with transmission channels as nodes and the correlation between channels as edge weights, and implements multi-layer propagation and feature aggregation in the graph neural network, effectively extracting the potential collaborative relationships and abnormal propagation paths between transmission channels, compensating for the local deviation caused by the traditional method of independent channel modeling, and improving the ability to identify system-level faults and linkage anomalies.

[0015] 3. Based on the output of the graph neural network, the present invention introduces a self-attention mechanism to perform global dependency modeling on the embedded features of different channels, forming a channel context perception vector, effectively capturing the deep interactive influence relationship between channels, and improving the expression ability and generalization adaptability of complex and changeable abnormal patterns.

[0016] 4. The present invention combines three indicators: state score, trend change rate, and reconstruction residual. It dynamically generates judgment parameters through a lightweight neural network and constructs a state difference score based on historical residuals to form an anomaly scoring system that integrates explicit and implicit information. This design enhances the ability to distinguish trend-type, degradation-type, and sudden anomalies, and has good detection accuracy and anti-interference ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0018] Fig. 1 This is a flow chart of abnormality detection of equipment operating status according to the present invention; Fig. 2 Flowchart of the channel confidence screening mechanism of the present invention; Fig. 3 This is a flowchart of the three-dimensional feature tensor modeling and fusion of the present invention. DETAILED DESCRIPTION

[0019] To achieve the above objectives, the present invention is implemented through the following technical solutions: the present invention provides a device operation status abnormality detection method based on edge computing, please refer to Figs. 1 to 3 , the method comprising: S1. Set up a channel confidence screening mechanism in the edge computing end to perform confidence assessment on the transmission channel of the device operation data, eliminate the transmission channels that do not meet the quality requirements, and expand the transmission matrix constructed by the operation data of the remaining transmission channels into a three-dimensional feature tensor. The operating data of the equipment is collected through sensors. The sensors and transmission channels are bound one-to-one. The collected operating data is transmitted to the edge computing end through their respective transmission channels. In the running data transmission process, due to the influence of installation position, environmental conditions, service life and other factors, the corresponding transmission channel will have abnormal fluctuations, offset drift or distortion noise and other problems in the transmission process. If all running data is input into the edge computing terminal without distinction, it will seriously affect the accuracy of subsequent anomaly detection. In order to reduce the interference caused by this situation, a channel confidence screening mechanism is designed, specifically, 1) The running data is transmitted through M transmission channels, and the current time step is t. Each channel continuously transmits the running data of T time windows, and the transmission matrix is represented as: Among them, is the running data of M channels t time steps in the time window T; is the running data of the mth channel in the time window T, which is represented as: Among them, represents the running data of the mth channel at the start of the time window; In the embodiment, M is 8, T is 60s, m is 4, t is 70s, the current time step is 70s, and the running data of 60s time window is continuously transmitted through 8 transmission channels, i.e. from 11s to 70s.

[0020] 2) Set the distribution consistency index, measure the difference between the current observation distribution and the historical reference distribution of the transmission channel, and reflect whether the channel has offset or drift phenomenon. For the mth channel, define the KL divergence, the formula is: Among them, is the KL divergence of channel m, which means that the current data deviates from the history more seriously, and there may be an anomaly; is the a interval of the running data in the histogram bin; is the probability of channel m in the a interval in the historical period; is the probability of channel m in the a interval in the current time window; 3) Set the fluctuation intensity index to evaluate the stability of the current channel data, the formula is: Among them, is the coefficient of variation value, which means that the channel data fluctuation is more intense, and the stability is poor; is the standard deviation of channel m in the current time window; is the mean value of channel m in the current time window, is a positive number; 4) Combined with the distribution consistency index and the fluctuation intensity index, a channel confidence score formula is constructed, which is expressed as: wherein, is the confidence score of channel m, the higher the better the transmission channel quality; and is the weight factor; when , is the set confidence threshold, indicating that the quality of channel m does not meet the requirements and needs to be removed, and finally the reconstructed transmission matrix is formed, which is expressed as: wherein, N is the number of channels meeting the quality requirements; The running data of each transmission channel is subjected to min-max normalization preprocessing, and the normalized running data facilitates fast convergence and avoids training bias caused by dimensional differences between channels.

[0021] In the embodiment, and are 0.6 and 0.4 respectively, is 0.45, and in the running data collection process, it is detected that the confidence score of channel 4 is less than the set confidence threshold 0.45, and finally channel 4 is removed and the sensor corresponding to channel 4 is replaced.

[0022] Finally, in order to retain the complete time sequence and multi-dimensional features of each channel, the processed transmission matrix is expanded into a three-dimensional tensor form suitable for neural network input, and the three dimensions are channel number N, time sequence length T of each channel and feature dimension D at each time step; The standardization processing of the running data and the three-dimensional tensor construction operation have very small calculation amount, are suitable for resource-limited edge computing node environment, and only involve dimension expansion and basic normalization operation on two-dimensional matrix, so that the subsequent neural network can receive input in a unified and standardized format, thereby meeting the requirements of stability and high efficiency of inference interface at the edge.

[0023] S2, based on the graph neural network and the self-attention mechanism, the three-dimensional feature tensor is modeled and fused to generate a channel context perception vector, and a multi-task state index is further constructed, After obtaining the three-dimensional feature tensor, in order to further improve the accuracy and robustness of anomaly detection, a state representation method combining graph neural network (GNN) structure and self-attention mechanism is introduced in the neural network at the edge computing end, which is used to learn the potential correlation between channels and enhance the sensitivity expression of the three-dimensional feature tensor, thereby improving the anomaly perception ability of the edge computing end.

[0024] Due to the correlation between sensors in physical location, sensing type or historical statistical pattern, the corresponding N transmission channel set is constructed as a channel graph , where each transmission channel is a node in the graph, and the edge is defined by the correlation of the cooperative change between channels, and the edge weight is expressed as: , where and are the running data of the bth and cth channels in the time window T; is the Pearson correlation coefficient, and the value is , which represents the consistency of the behavior of channels b and c in the T time window, and the larger the value, the stronger the current cooperation of the two channels, and the higher the interaction weight will be given in the graph neural network propagation; Since the graph neural network requires the input form to be a two-dimensional node feature matrix, it cannot directly process the time dimension, so it is necessary to perform dimensionality reduction and flattening operation on the three-dimensional feature tensor. Specifically, for each transmission channel, use sliding statistics (such as mean, maximum, spectral energy, etc.) or one-dimensional convolution pooling to map the corresponding three-dimensional feature tensor to a node feature vector, denoted as: , where represents the dimensionality reduction function; is the feature vector of transmission channel n at the 0th layer of the graph neural network, and the 0th layer of the graph neural network is the input layer; After summarizing the feature vectors of all transmission channels, a node feature matrix is constructed, denoted as: On the channel graph , use a multi-layer graph convolutional neural network (GCN) to propagate and aggregate the features between nodes, and the propagation and aggregation formula of the first layer is: , where is the embedding feature vector representation of the nth node in the layer of the graph neural network; is a nonlinear activation function; is the neighbor node set of node n; is the edge weight between node n and neighbor node c, representing the correlation between channels; and are the degrees of node n and node c, respectively, that is, the number of neighbors; is the graph convolution weight matrix of the first layer; is the transmission channel the embedding feature vector of the nth transmission channel at the Lth layer of the graph neural network. Through propagation and aggregation, the final embedding feature vector sequence is output , denoted as: wherein, is the embedding feature vector of the nth transmission channel at the Lth layer of the graph neural network.

[0025] The embedding feature vector sequence is input into a self-attention mechanism (Transformer), and the global dependency between each embedding feature vector is modeled through multi-head self-attention mechanism, thereby further strengthening the perception ability of anomaly commonness and difference. The calculation formula of the attention mechanism is: wherein, , and are the query (Query), key (Key) and value (Value) vector matrices in the embedding feature vector sequence; is the dimension of the key vector; after attention weighting, the fused channel context perception vector sequence is obtained, denoted as: wherein, is the channel context perception vector of the nth transmission channel after fusing global information (the state of other transmission channels); is the channel context perception vector sequence after fusing all channels, which retains the global dependency information.

[0026] Based on the fused channel context perception vector sequence , a multi-task state indicator is constructed, and multiple indicators are jointly output to reflect the running performance in different dimensions, specifically, a state score is calculated, and the fused channel context perception vector sequence is extracted through an average pooling operation to obtain a global state, denoted as: wherein, is a pooling operation function, and average pooling is adopted to compress the vectors of multiple channels into a global embedding; is a global state representation vector, which is a global context representation after summarizing the channel dimension; then is input into a fully connected layer and mapped into a state score result by a Sigmoid function, denoted as: wherein, is the state score result of time step t, is a Sigmoid activation function; is a weight vector of the state score; is a score bias term; a score trend change rate is calculated, the score trend change rate represents a time change trend of the state score, reflects whether the running state is continuously deteriorating or improving, and is defined as a score difference between consecutive time points, and is expressed as: a state reconstruction residual is calculated based on the embedding feature vector of each channel, the running data is reconstructed by a reconstruction function, and the overall state reconstruction deviation is measured in the form of mean square error, and is expressed as: wherein, is the reconstructed running data; by introducing a joint loss function, the output results of the multi-task state indicators are simultaneously optimized.

[0027] Finally, by combining graph structure modeling, self-attention fusion and multi-index task decoupling design, efficient and refined device running state description is realized on the edge computing end, and the forward-looking and accuracy of the anomaly detection are effectively improved.

[0028] S3, a fusion type anomaly judgment mechanism is designed on the edge computing end, a historical state sequence is constructed based on the multi-task state indicators, and criterion parameters are generated, and combined with the state variation score for anomaly detection, to enhance the dynamic adaptability of anomaly detection, first, a self-evolution criterion generator is designed, the optimal weighting and judgment threshold are automatically learned according to the multi-task state indicators; second, through an attention memory reasoning mechanism, the context correction of the historical state to the current state is introduced to enhance the recognition ability of the trend type anomaly; finally, the response action is matched according to different risk levels, specifically, First, P historical state scores, score trend change rates and state reconstruction residuals are collected to form a historical state sequence a lightweight neural network criterion generator is used to learn the evolution law of the historical state sequence, and all criterion parameters of the current time step t are output, and is expressed as: wherein, , and are the weighting factors of the state score, the score trend change rate and the state reconstruction residual, respectively; is the anomaly judgment threshold of the current time step t; is a lightweight neural network model used for dynamic criterion generation.

[0029] Then, the fused channel context perception vector sequence at time step t is obtained Referring to, the residual values between P historical time steps are calculated, and a residual sequence is composed, the residual sequence is input into a light convolutional network (1D CNN) to extract local change trends, and then global statistics are extracted through a pooling operation, and a state difference score is output , The greater, the more obvious the difference between the current device running state and the history, and the more likely there are trend drift, function degradation or potential abnormalities.

[0030] Finally, the criterion parameters of the multi-task state indicators (explicit criteria) and the state difference score (implicit perception) are fused to calculate the final abnormal score , the formula is: wherein, is the fusion weight of the state variation score, and the value is set according to the environment in which the device is deployed.

[0031] When , it indicates that the device is abnormal, and at the same time, the abnormality is divided into multiple levels according to the size of the abnormal score, and the corresponding response action is performed according to the corresponding level.

[0032] The application proposes a device running state abnormality detection method based on edge computing, which innovatively integrates a channel confidence screening mechanism, a graph neural network structure modeling, a self-attention global fusion mechanism and a multi-task state indicator determination system at the edge computing end. Through accurate evaluation and screening of the quality of the transmission data, combined with the structural dependence relationship between multiple channels and the context information fusion, the application greatly improves the accuracy, robustness and real-time performance of the device abnormality detection on the basis of ensuring the efficiency of edge deployment.

[0033] In summary, the application realizes the key technical leap from "low-quality data + shallow analysis" to "high-confidence input + deep modeling" in the field of device running state abnormality detection, and can be widely applied to industrial automation, intelligent manufacturing, energy monitoring and other application scenarios that have strict requirements for high reliability and high timeliness of abnormality identification, and has significant technological progress and industrial application value.

[0034] The above is merely a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for detecting abnormal operation status of equipment based on edge computing, characterized in that: The method includes: S1. Setting a channel confidence screening mechanism in the edge computing end to perform confidence assessment on the transmission channels of the device operation data, eliminating transmission channels that do not meet the quality requirements, and expanding the transmission matrix constructed from the operation data of the remaining transmission channels into a three-dimensional feature tensor; S2. Based on a graph neural network and a self-attention mechanism, the three-dimensional feature tensor is modeled and fused to generate a channel context-aware vector, and a multi-task status indicator is further constructed. S3. Design a fusion anomaly judgment mechanism at the edge computing end, construct a historical state sequence based on the multi-task state indicators and generate judgment parameters, and perform anomaly detection in combination with the state variation score.

2. The method for detecting abnormal operation status of a device based on edge computing according to claim 1, characterized in that: The operating data is collected by M sensors, and the sensors and transmission channels are bound one-to-one. The operating data is transmitted to the edge computing end through the corresponding transmission channel.

3. The method for detecting abnormal operation status of a device based on edge computing according to claim 1, characterized in that: The specific steps of setting up the channel confidence screening mechanism are: 1) When the time step is t, M transmission channels are gathered to continuously transmit the running data of T time windows. The transmission matrix is ​​expressed as: in, It is the running data of M channels with t time steps in the time window T; is the running data of the mth channel in the time window T, expressed as: in, Represents the running data of the mth channel at the starting point of the time window; 2) Set the distribution consistency index and define the KL divergence for the mth channel. The formula is: in, is the KL divergence of channel m; is the ath interval of the running data in the histogram bin; is the probability of channel m being in the ath interval in the historical period; is the probability of channel m in the ath interval in the current time window; 3) Set the volatility strength indicator and calculate the coefficient of variation. The formula is: in, is the coefficient of variation value; is the standard deviation of channel m in the current time window; is the mean value of channel m in the current time window, is a positive number; 4) Combining the distribution consistency index and the volatility strength index, we construct a channel confidence score, which is expressed as: in, is the confidence score of channel m; and is the weighting factor.

4. The method for detecting abnormal operation status of a device based on edge computing according to claim 3, characterized in that: Based on the confidence score and the confidence threshold Compare and judge the quality of all transmission channels. , then the quality of channel m does not meet the requirements and needs to be eliminated. At the same time, the corresponding sensor should be replaced; The operating data of the remaining transmission channels form the reconstructed transmission matrix, which is expressed as: Where N is the number of channels that meet the quality requirements.

5. The method for detecting abnormal operation status of a device based on edge computing according to claim 1 or 4, characterized in that: According to the graph neural network, a channel graph is constructed , the nodes of the channel graph correspond to transmission channels, and the edges are defined by edge weights, which can be expressed as: in, and are the running data of the bth and cth channels in the time window T respectively; is the Pearson correlation coefficient; The three-dimensional feature tensor is flattened and mapped into a node feature vector using the following formula: in, represents the dimensionality reduction function; It is the feature vector of the transmission channel n in the 0th layer of the graph neural network, and the 0th layer of the graph neural network is the input layer.

6. The method for detecting abnormal operation status of equipment based on edge computing according to claim 5, characterized in that: Through the multi-layer graph convolutional neural network, the feature vectors between nodes are propagated and aggregated on the channel graph. The formula is: in, It is the first Feature vector representation of node n in the layer; is a nonlinear activation function; is the set of neighbor nodes of node n; is the edge weight between node n and its neighbor node c; and are the degrees of node n and node c, i.e. the number of neighbors; It is The graph convolution weight matrix of the layer; It is a transmission channel In the graph neural network The feature vector of the layer; According to the propagation and aggregation operations, the final feature vector sequence of the transmission channel is obtained, which is expressed as: in, It is the embedded feature vector of transmission channel n in the Lth layer of the graph neural network.

7. A method for detecting abnormal operation status of equipment based on edge computing according to claim 1 or 6, characterized in that: The self-attention mechanism generates a channel context-aware vector sequence by fusing the global dependencies between the embedded feature vectors Expressed as: in, is the channel context awareness vector of transmission channel n; The multi-task state indicators are calculated based on the channel context perception vector sequence, and the state scores are , score trend change rate and state reconstruction residual .

8. The method for detecting abnormal operation status of equipment based on edge computing according to claim 7, characterized in that: The historical state sequence is constructed by collecting P historical state scores, score trend change rates and state reconstruction residuals. The generation formula of the criterion parameter is: in, 、 and are the weighting factors for the state score, the rate of change of the score trend, and the state reconstruction residual, respectively; is the abnormality determination threshold at the current time step t; It is a lightweight neural network model; is a historical state sequence; Calculate the residual value between the current time step t and P historical time steps of the fused channel context perception vector sequence, and obtain the state difference score through a lightweight convolutional network and pooling operation ; The criterion parameters and the state difference score are integrated to calculate the final anomaly score Q, which is: in, is the fusion weight of the state variation score; when , indicating that the device's operating status is abnormal.

Images