Generator set anomaly detection method based on edge calculation

Abstract

The invention discloses a generator set anomaly detection method based on edge calculation. The method comprises the following steps: S1, generating multichannel working condition data of a standardized generator set; s2, deploying a lightweight long short-term memory network prediction model in an industrial edge computing node; s3, constructing a multi-mode residual error map; s4, performing generator set abnormity judgment on the local residual error sequence and outputting generator set abnormity candidate events; and S5, generating a generator set anomaly causal interpretation graph for the generator set anomaly candidate events by using an interpretive analysis algorithm, wherein the generator set anomaly causal interpretation graph reveals the contribution relationship of each working condition feature to the generator set anomaly confidence coefficient. According to the method, the traceability and operability of fault positioning are remarkably improved, and operation and maintenance personnel are assisted in accurately formulating a maintenance strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generator set, and particularly relates to a generator set abnormality detection method based on edge computing. BACKGROUND

[0002] With the wide application of new generation information technology in the energy industry, the intelligent operation and maintenance level of the generator set is continuously improved, and the key tasks of equipment state monitoring, fault early warning and abnormality diagnosis gradually transit from manual inspection to automation and intelligence. Traditional unit abnormality detection usually relies on centralized SCADA system or remote monitoring platform, and abnormality identification and fault positioning are performed on the back-end server through the collection of operation data of the control system and sensor network. However, the centralized processing architecture has the following typical problems: firstly, the raw data needs to be transmitted to the remote server for processing, which is easy to cause bandwidth bottleneck, especially in the multi-channel high-frequency data scene, the data transmission delay is high, and it is difficult to meet the requirements of high real-time and response speed; secondly, the unified analysis and processing on the remote server lacks the adaptation ability to the specific operation environment of the edge scene, and it is difficult to respond sensitively to local abnormal changes; thirdly, the fault features of some key components are early subtle statistical disturbances or frequency domain signal changes, which lack sufficient time resolution and spatial granularity, causing misjudgment or missed detection.

[0003] In terms of models, most existing methods use static feature extraction and fixed window statistical analysis to model and judge the unit operation state, which fails to fully utilize the dynamic evolution characteristics of long time series data, and cannot deeply capture the working condition trend change and short-time disturbance signal pattern. At the same time, most existing models are deployed on the cloud or central server, which occupies high computing resources and is not suitable for deployment on edge nodes, making it difficult to meet the deployment needs of on-site online processing. In addition, some abnormality detection methods rely on preset fixed thresholds, which are difficult to achieve adaptive judgment in a variable working condition background, resulting in high false positive rate and false negative rate.

[0004] In addition, although some researches try to introduce deep learning models to improve prediction and identification accuracy, they often face problems such as high model calculation complexity, unclear sensitivity to input channel information, and inability to dynamically adjust the prediction range when deployed in actual industrial sites. At the same time, the abnormality detection results lack causal explanation mechanism, which makes it difficult to effectively assist dispatch personnel in rapid positioning and fault troubleshooting, and is not conducive to the landing of actual operation and maintenance.

[0005] Therefore, the existing generator set abnormality detection technology has different degrees of deficiencies in real-time, deployment flexibility, detection accuracy and explainability, and a new method that integrates edge computing, efficient sequence modeling and causal explanation analysis is urgently needed to improve the comprehensiveness, accuracy and practicality of on-site abnormality detection. SUMMARY

[0006] An object of the present application is to provide an edge computing-based generator set anomaly detection method.

[0007] According to an embodiment of the present application, an edge computing-based generator set anomaly detection method comprises the following steps:

[0008] S1. An industrial edge computing node is deployed on site to collect and preprocess multi-channel working condition data of the generator set, and to generate standardized multi-channel working condition data of the generator set;

[0009] S2. A light-weight long short-term memory network prediction model is deployed in the industrial edge computing node, which receives the standardized multi-channel working condition data of the generator set and outputs a rolling predicted generator set state estimation sequence and a corresponding prediction confidence interval;

[0010] S3. The difference between the standardized multi-channel working condition data of the generator set and the rolling predicted generator set state estimation sequence is calculated to obtain a local residual sequence, and a sliding time window is used to calculate statistical features and spectral energy to construct a multi-modal residual atlas;

[0011] S4. Based on the residual atlas and historical benchmark data, a device health index is calculated to form a global threshold baseline of the device health index, an adaptive local statistical threshold is generated according to the current statistical distribution of the residual atlas, and a dual-threshold discrimination mechanism is constructed by fusing the global threshold baseline of the device health index and the adaptive local statistical threshold to discriminate the local residual sequence for generator set anomaly and output generator set anomaly candidate events;

[0012] S5. An explanatory analysis algorithm is used to generate a generator set anomaly causal explanation graph for the generator set anomaly candidate events, which reveals the contribution relationship of each working condition feature to the confidence of the generator set anomaly.

[0013] Optionally, the S1 comprises the following steps:

[0014] S11. Through the industrial edge computing node deployed on site, real-time working condition raw data of the generator set control system, sensor network and SCADA system are collected to obtain each piece of multi-channel working condition raw data of the generator set, which together form a multi-channel working condition raw data set of the generator set;

[0015] S12. The multi-channel working condition raw data set of the generator set is processed for time sequence alignment to obtain a time sequence aligned multi-channel working condition data matrix;

[0016] S13. For the multi-channel operating condition data matrix after time-series alignment, the long-term trend component is extracted by using a moving average model for the data sequence of each channel, and the trend residual value of the operating condition data after trend removal is obtained by subtracting the moving average trend component of the corresponding channel at the time step from each measurement value.

[0017] S14. The residual values ​​of the operating condition data after trend removal are normalized by Z-score normalization method to obtain normalized operating condition data.

[0018] S15. Combine all normalized operating condition data into a standardized operating condition data matrix.

[0019] Optionally, S2 includes the following steps:

[0020] S21. Organize the data from each time step in the standardized operating condition data matrix into a time step input vector according to channel order. Each time step input vector contains the standardized measurement values ​​of all channels. For each time step input vector, calculate the channel attention weight coefficient vector using weight parameters and bias parameters. Multiply the channel attention weight coefficient vector element-wise with the time step input vector to obtain the channel-aware weighted input vector.

[0021] S22. Weight the channel-aware input vector Input lightweight Long Short-Term Memory (LSTM) network prediction model lite In the middle, the hidden state h is output at each time step. k And with the most recent W h The hidden state sequence within each time step constitutes the hidden state window H. k ={h k -W h +1,…,h k};

[0022] S23. Based on the hidden state window H k The residual perturbation changes in the hidden state are used to construct the state perturbation intensity index S. k The state disturbance intensity index is used to measure the overall magnitude of hidden state changes within the current window. It is based on the state disturbance intensity index S within the current window. k Given a specified maximum prediction step size and control coefficient, the rolling prediction step size H is adaptively adjusted using an exponential decay method. p (k), the prediction step size is automatically shortened when the state disturbance intensity index is greater than the threshold, and automatically extended when the state disturbance intensity index is less than the threshold;

[0023] S24. Based on the state disturbance intensity index S k Simultaneously, construct pruning importance constraint indicators;

[0024] S25. Using the pruned lightweight long short-term memory network prediction model LSTM at current time step k lite with a rolling prediction horizon H p (k) outputting a sequence of future generator set state predictions and combining the most recent W h step state disturbance intensity indicator S k computing a confidence intensity indicator γ k = 1 - tanh(ηS k ), and generating a prediction confidence interval:

[0025]

[0026] where σ is the historical residual standard deviation, γ k reflects the dynamic adjustment capability of prediction reliability with disturbance changes;

[0027] S26. Storing the sequence of predictions and the corresponding prediction confidence interval {I k+h} into the edge node cache area.

[0028] Optionally, the pruning importance constraint indicator is the sensitivity of each weight parameter of the lightweight long short-term memory network prediction model to the state disturbance intensity indicator, the sensitivity is defined as the product of the partial derivative of the weight parameter to the state disturbance intensity indicator and the weight parameter itself, all weight parameters with a sensitivity greater than a pruning threshold are retained, and other weight parameters are set to zero.

[0029] Optionally, the S3 comprises the following steps:

[0030] S31. Based on the standardized generator set multi-channel operating condition data matrix and the sequence of rolling prediction generator set state predictions, calculating the prediction error at each time step to obtain a sequence of local residuals where x k+h is the actual measurement vector at time step k+h, is the corresponding predicted value vector;

[0031] S32. For the sequence of local residuals e k+h construct a single-channel residual sequence e Within a sliding window length W r , calculate the residual sliding mean and the residual sliding variance

[0032] S33. For each channel residual sequence, use short-time Fourier transform to calculate the spectral energy density value E (j)(f, t);

[0033] S34. Jointly encode the residual sliding mean, residual sliding variance and spectral energy density value of all channels within the current sliding window, to construct the residual atlas R k ;

[0034] S35. Calculate the relative change rate of each channel's current measurement value with respect to the mean value of the channel's most recent stable window within the same sliding time window:

[0035]

[0036] wherein, is the mean value of the channel j's most recent stable window, v" k,j represents the normalized measurement value of the multi-channel operating condition data of the generator set at the current time step k and channel number j, and ∈ is a constant;

[0037] Assemble the relative change rates of all channels into the operating condition relative change atlas D k ;

[0038] S36. Calculate the Pearson correlation coefficient ρ i,j between each pair of residual time series of all channels within the same sliding time window, and assemble the Pearson correlation coefficients between all channels into the structural consistency atlas C k ;

[0039] S37. Fuse the residual atlas R k , the operating condition relative change atlas D k , and the structural consistency atlas C k to form a multi-modal residual atlas

[0040] Optionally, the S4 comprises the following steps:

[0041] S41. Based on the multi-modal residual atlas, sum the absolute values of the residual sliding means, sum the residual sliding variances, and sum the spectral energy density values at all frequencies and time windows of all channels, according to the weighted aggregation coefficient vector w = [w1, w2, w3] T to obtain the multi-modal residual comprehensive indicator A k of the current time step;

[0042] S42. Under the action of a set exponential smoothing coefficient α, recursively calculate the generator set equipment health index EHI k-1 based on the historical multi-modal residual comprehensive indicator sequence {A k-2 , …}. k = αA k+ (1 - a)EHI k-1 , wherein EHI0takes the mean value of {A -N} in the historical reference period, EHI k The larger the value, the lower the overall generator set equipment health level.

[0043] S43. Utilize the historical sliding window W g The maximum-minimum normalization of the generator set equipment health index in the window is performed to obtain the normalized health score The normalized health score is multiplied by a preset amplification coefficient and added to a safety margin constant to map and generate the global health threshold baseline G k ;

[0044] S44. In the current sliding window, the local statistical mean and local statistical standard deviation of the multi-modal residual comprehensive index are calculated, the local statistical standard deviation is multiplied by the quantile constant corresponding to the significance level under the standard normal distribution and added to the local statistical mean to generate the adaptive local statistical threshold L k ;

[0045] S45. The global health threshold baseline and the adaptive local statistical threshold are fused to construct a double-threshold discrimination mechanism, and the determination result Flag k ;

[0046] S46. The determination result Flag k of S45 is packaged and output with the corresponding multi-modal residual comprehensive index A k , the global threshold G k , and the local threshold L k

[0047] Optionally, the double-threshold discrimination mechanism is specifically: if the multi-modal residual comprehensive index of a certain time step exceeds both the global health threshold baseline and the adaptive local statistical threshold, it is determined that there is a generator set anomaly at the corresponding time step, otherwise it is considered to be normal.

[0048] Optionally, the S5 includes the following steps:

[0049] S51. When receiving an abnormal candidate event and the abnormality discrimination flag is abnormal, the standardized generator set multi-channel operating condition input vector x k = [v″ k,1 , v″ k,2 , …, v″ k,C ] T of the current time step is extracted, and the multi-modal residual map corresponding to the current time step is recorded.

[0050] S52. A lightweight long short-term memory network prediction model LSTM lite ​An explanation objective function of the forward model The normalized multi-channel working condition input vector of the current time step is input into the forward model, and the integrated gradient attribution analysis algorithm is used to calculate the prediction contribution degree of each working condition feature in the normalized multi-channel working condition input vector of the generator set;

[0051] S53. The prediction contribution degrees of all working condition features are combined into a feature contribution degree vector, and are fused with the structural consistency features in the multi-modal residual atlas of the current time step to obtain a generator set abnormal causal explanation graph with feature contribution degrees and structural consistency features;

[0052] S54. All nodes in the generator set abnormal causal explanation graph are ranked according to the feature contribution degrees, and the top N channels in the feature contribution degree ranking are selected as abnormal causal path nodes, and an abnormal causal path set is constructed in accordance with the order from high to low of the feature contribution degrees;

[0053] S55. The generator set abnormal causal explanation graph and the abnormal causal path set are structured and visualized to generate a generator set abnormal causal explanation atlas, which is used to assist dispatchers in fault confirmation and maintenance strategy decision-making.

[0054] Optionally, the integrated gradient attribution analysis algorithm accumulates the partial derivative of the output of the forward model to each working condition feature on the path from the baseline input interpolation to the actual input of the normalized multi-channel working condition input vector of the generator set, multiplies the difference between the working condition feature and the baseline input, and obtains the corresponding prediction contribution degree.

[0055] The beneficial effects of the present application are:

[0056] (1) The present application deploys a lightweight long short-term memory network prediction model on an industrial edge computing node, introduces an adaptive rolling prediction mechanism based on state disturbance and a pruning importance constraint mechanism, dynamically adjusts the prediction step while ensuring the real-time of the model, can more accurately capture the early signals of state mutation or fluctuation of the device, can adaptively shorten or lengthen the prediction range according to the hidden state disturbance intensity index, thereby improving the sensitivity when the working condition changes dramatically and reducing the false alarm rate when running smoothly, in addition, the network structure is constrained by the pruning importance index, only the weight parameters sensitive to the state disturbance are retained, the redundant calculation of the model is reduced, and the efficiency of the edge deployment is improved.

[0057] (2) The application first constructs a multi-modal residual atlas that fuses residual statistics, spectral features and structural consistency, and comprehensively depicts the microscopic fluctuation characteristics and channel structure coordination of the generator set operating state, which is particularly effective in discriminating critical states or early weak faults. The health index based on the atlas can not only reflect single-point abnormalities, but also depict the structural instability trend at the system level. Combined with the adaptive statistical threshold, a double-threshold discrimination mechanism is formed, which effectively improves the detection accuracy.

[0058] (3) The application introduces an abnormal causal explanation graph generation algorithm based on an integrated gradient attribution mechanism, which analyzes the explainability of the contribution of the input multi-channel working condition features to the forward prediction model, constructs a structured abnormal explanation atlas with causal relationships, and automatically generates an abnormal causal path set by analyzing the fusion results of feature contribution and structural consistency atlas, significantly improving the traceability and operability of fault location, and assisting maintenance personnel in accurately formulating maintenance strategies. BRIEF DESCRIPTION OF DRAWINGS

[0059] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0060] Figure 1 A flowchart of an abnormality detection method for a generator set based on edge computing is provided. DETAILED DESCRIPTION

[0061] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically illustrate the basic structure of the application, and therefore only show the components related to the application.

[0062] REFERENCE Figure 1 An abnormality detection method for a generator set based on edge computing includes the following steps:

[0063] S1. An industrial edge computing node is deployed on site to collect and locally cache multi-channel working condition data of the generator set, and performs time alignment, missing interpolation, trend removal and normalization preprocessing on the multi-channel working condition data of the generator set using the industrial edge computing node to generate standardized multi-channel working condition data of the generator set;

[0064] S2. A lightweight long short-term memory network prediction model is deployed in the industrial edge computing node, which receives the standardized multi-channel working condition data of the generator set and outputs a rolling prediction generator set state estimation sequence and a corresponding prediction confidence interval;

[0065] S3. Calculate the difference between the standardized multi-channel operating condition data of the generator set and the rolling predicted generator set state estimation sequence to obtain a local residual sequence, apply a sliding time window to the local residual sequence to calculate statistical features and spectral energy, and construct a multi-modal residual atlas;

[0066] S4. Calculate the equipment health index based on the residual atlas and historical benchmark data to form a global threshold baseline of the equipment health index, generate an adaptive local statistical threshold according to the current statistical distribution of the residual atlas, fuse the global threshold baseline of the equipment health index and the adaptive local statistical threshold to construct a double-threshold discrimination mechanism, and perform generator set anomaly discrimination on the local residual sequence and output generator set anomaly candidate events;

[0067] S5. Use an explanatory analysis algorithm to generate a generator set anomaly causal explanation graph for the generator set anomaly candidate event, and the generator set anomaly causal explanation graph reveals the contribution relationship of each operating condition feature to the confidence of the generator set anomaly.

[0068] In this embodiment, S1 includes the following steps:

[0069] S11. Collect real-time operating condition raw data of the generator set control system, sensor network and SCADA system through an industrial edge computing node deployed on site of the generator set to obtain each piece of multi-channel operating condition raw data of the generator set, wherein the multi-channel operating condition raw data of the generator set includes a collection timestamp, a sensor measurement value and a channel type identifier, all the multi-channel operating condition raw data of the generator set collectively form a multi-channel operating condition raw data set of the generator set, the collection timestamp is used to mark a specific time of data collection, the sensor measurement value is used to represent a specific operating state of the generator set, and the channel type identifier is used to distinguish different data source types;

[0070] S12. Perform time series alignment processing on the multi-channel operating condition raw data set of the generator set, synchronize and interpolate missing data of all channels using a unified time reference sequence, so that each channel has a corresponding measurement value at each unified time point, and all channel data finally form a multi-channel operating condition data matrix with a specified total number of time steps and a specified total number of channels, each measurement value in the multi-channel operating condition data matrix corresponds to a unique time step and a channel;

[0071] S13. For the time series aligned multi-channel operating condition data matrix, a sliding average model is used to extract a long-term trend component for each channel data sequence, and a trend-removed operating condition data residual value is obtained by subtracting the sliding average trend component of the corresponding channel at the time step from each measurement value, and the trend-removed operating condition data residual value is used to eliminate the long-term change trend in the data;

[0072] S14. The trend-removed operating condition data residual values are normalized by using the Z-score normalization method. For each channel, the residual mean and residual standard deviation of the channel in the sliding window are calculated. Each trend-removed operating condition data residual value is subtracted by the residual mean of the channel and divided by the residual standard deviation of the channel to obtain the normalized operating condition data;

[0073] S15. All the normalized operating condition data are combined to form a standardized operating condition data matrix, and the standardized operating condition data matrix is cached in the local cache area of the industrial edge computing node.

[0074] In the embodiment, S2 includes the following steps:

[0075] S21. The data of each time step in the standardized operating condition data matrix is combined into a time step input vector in the order of the channels. Each time step input vector contains the standardized measurement values of all channels. For each time step input vector, a channel attention weight coefficient vector is calculated using the weight parameter and the bias parameter. The channel attention weight coefficient vector is used to measure the dominant influence degree of each channel measurement value on the evolution of the generator set state. The channel attention weight coefficient vector is multiplied by the time step input vector element by element to obtain a channel perception weighted input vector.

[0076] S22. The channel perception weighted input vector is input into the lightweight long short-term memory network prediction model LSTM lite At each time step, the output hidden state h k is obtained, and the hidden state sequence in the last W h time steps is used to form a hidden state window H k ={h k-Wh+1 ,…,h k}, which is used to describe the current short-time dynamic evolution.

[0077] S23. A state disturbance intensity indicator is constructed based on the residual disturbance change of the hidden state in the hidden state window H k . The state disturbance intensity indicator is used to measure the overall amplitude of the change of the hidden state in the current window:

[0078]

[0079] wherein, is the average hidden state vector in the window, and S k represents the state disturbance intensity indicator in the current window.

[0080] According to the state disturbance intensity indicator S kand the specified maximum prediction step size and control coefficient, the rolling prediction step size H is adaptively adjusted in an exponential decay manner p (k), the prediction step size is automatically shortened when the state disturbance intensity index is greater than a threshold, and is automatically lengthened when the state disturbance intensity index is less than the threshold:

[0081]

[0082] wherein β is a control coefficient, H max is the maximum prediction step size;

[0083] S24. Based on the state disturbance intensity index S k Meanwhile, a pruning importance constraint index is constructed;

[0084] S25. At the current time step k, the light-weight long short-term memory network prediction model LSTM lite with the rolling prediction step size H p (k) outputs a future generator set state prediction sequence and combines the recent W h step state disturbance intensity index S k to calculate a confidence intensity index γ k = 1-tanh(ηS k ), and generates a prediction confidence interval, the confidence intensity index is used to dynamically adjust the width of the prediction confidence interval, the greater the confidence intensity index, the smaller the width of the prediction confidence interval, indicating the higher the prediction credibility:

[0085]

[0086] wherein σ is the historical residual standard deviation, γ k reflects the dynamic adjustment ability of the prediction credibility with the disturbance;

[0087] S26. The prediction sequence and the corresponding prediction confidence interval {I k+h} are stored in the edge node cache area.

[0088] In the embodiment, the pruning importance constraint index is the sensitivity of each weight parameter of the light-weight long short-term memory network prediction model to the state disturbance intensity index, the sensitivity is defined as the product of the partial derivative of the weight parameter to the state disturbance intensity index and the weight parameter itself, all weight parameters with a sensitivity greater than a pruning threshold are retained, and other weight parameters are set to zero, realizing structural sparse optimization for the current state disturbance sensitive features.

[0089] The lightweight long short-term memory network prediction model in the embodiment not only performs standardization processing on the data of each channel, but also dynamically allocates the contribution degrees of the channels through an attention weight mechanism, and can automatically identify signals that are more sensitive to abnormalities in different working condition stages.

[0090] In combination with the actual state disturbance intensity, the importance of the network weight parameters is dynamically evaluated to perform structured pruning on the connections with low contribution degrees, so as to reduce the model size and the calculation overhead under the premise of ensuring the prediction accuracy. The future prediction window is dynamically contracted or expanded according to the change of the model hidden state, intelligent scheduling of high-frequency prediction in fast-changing working conditions and low-frequency prediction in stable working conditions is realized, and the response capability to sudden abnormalities and early trend changes is effectively improved. The prediction confidence interval is dynamically adjusted by using multi-level information, so that the adaptability and alarm reliability of the abnormal threshold are obviously better than those of a conventional threshold scheme based on only single-layer variance or an empirical coefficient.

[0091] In the embodiment, S3 includes the following steps:

[0092] S31. Based on the standardized generator set multi-channel working condition data matrix and the rolling prediction generator set state prediction sequence, the prediction error of each time step is calculated to obtain a local residual error sequence where x k+h is an actual measurement vector at time step k+h, is a corresponding predicted value vector;

[0093] S32. The local residual error sequence e k+h is constructed according to the channel number j In a sliding window length W r , the residual sliding mean and the residual sliding variance of each channel are calculated. The residual sliding mean represents the average level of the residual of the channel in the current sliding window, and the residual sliding variance represents the fluctuation degree of the residual of the channel in the current sliding window.

[0094] S33. The short-time Fourier transform is adopted for each channel residual error sequence, and the spectral energy density value E (j) (f,t) of the channel at the frequency component and the time window position is calculated at each frequency component and each time window position. The spectral energy density value is used to measure the energy distribution of the channel at different frequencies and time windows.

[0095] S34. The residual sliding mean, the residual sliding variance and the spectral energy density value of all channels in the current sliding window are jointly coded to construct a residual atlas for comprehensively depicting the statistical features and frequency domain features of the residual signals of the channels at the current time step:

[0096]

[0097] wherein C is the total number of channels of the generator set multi-channel operating condition data;

[0098] S35. Calculate the relative change rate of each channel current measured value relative to the mean value of the current channel recent stable window within the same sliding time window:

[0099]

[0100] wherein, is the mean value of the channel j recent stable window, and ∈ is a constant;

[0101] Assemble the operating condition relative change graph of all channels:

[0102]

[0103] S36. Calculate the Pearson correlation coefficient between the residual time series of all channels two by two within the same sliding time window:

[0104]

[0105] wherein, i,j represents the Pearson correlation coefficient of the residual series of the i-th channel and the j-th channel in the current sliding window in the multi-channel operating condition of the generator set, used to measure the structural consistency of the change of different channel residuals, the value range is [-1, 1], the closer to 1 represents that the two channels are highly positively correlated, the closer to -1 represents highly negatively correlated, and close to 0 represents no significant correlation, represents the local residual value of the i-th channel at time step k-h moment, i.e., the difference between the measured value and the predicted value of the i-th channel of the generator set, represents the residual mean value of the i-th channel within the current sliding window length W r , used to remove the mean value effect when calculating the correlation coefficient, W r is the sliding window length of residual analysis;

[0106] Assemble the structural consistency graph C k =[ρ i,j ] C×C between all channels, used to describe the residual coupling structure between channels;

[0107] S37. Fuse the residual graph R k , the operating condition relative change graph D k , and the structural consistency graph C k to form a multi-modal residual graph and store the multi-modal residual graph in the edge computing node graph buffer.

[0108] In this embodiment, S4 includes the following steps:

[0109] S41. Based on the multi-modal residual atlas, sum the residual sliding mean absolute value of all channels, sum the residual sliding variance, and sum the spectral energy density value under all frequencies and time windows, according to the weighted aggregation coefficient vector w = [w1, w2, w3] T The weighted accumulation obtains the multi-modal residual comprehensive index A of the current time step k ;

[0110] S42. Under the action of the set exponential smoothing coefficient a, the generator set equipment health index EHI is recursively calculated according to the historical multi-modal residual comprehensive index sequence {A k-1 ,A k-2 ,…}:

[0111] EHI k = aA k + (1-a) EHI k-1 ;

[0112] Wherein, EHI0 takes the mean value of {A -N} in the historical reference period, and EHI k The larger the value is, the lower the overall generator set equipment health level is;

[0113] S43. The generator set equipment health index is maximum-minimum normalized in the window W g to obtain the normalized health score After multiplying the normalized health score by the preset amplification coefficient and adding the safety margin constant, the global health threshold baseline G k is generated;

[0114] S44. In the current sliding window, the local statistical mean and the local statistical standard deviation of the multi-modal residual comprehensive index are calculated, the local statistical standard deviation is multiplied by the quantile constant corresponding to the significance level under the standard normal distribution and added to the local statistical mean to generate the adaptive local statistical threshold L k The local statistical mean represents the central tendency of the residual level in the time window, and the local statistical standard deviation represents the fluctuation degree of the residual level in the time window;

[0115] S45. The global health threshold baseline and the adaptive local statistical threshold are fused to construct a double-threshold discrimination mechanism, and the determination result Flag k is obtained;

[0116] S46. The determination result Flag k of S45 is fused with the corresponding multi-modal residual comprehensive index A k , the global threshold Gk and local threshold L k packaging output abnormal candidate events and push to the edge computing node alarm queue.

[0117] In this embodiment, the double-threshold discrimination mechanism is specifically: if the multi-modal residual comprehensive index of a certain time step simultaneously exceeds the global health threshold baseline and the adaptive local statistical threshold, it is determined that the corresponding time step has a generator set abnormality, otherwise it is considered to have no abnormality. The double-threshold discrimination mechanism combines long-term trend and local fluctuation information, and improves the sensitivity and reliability of the generator set discrimination.

[0118] In this embodiment, S5 includes the following steps:

[0119] S51. When receiving an abnormal candidate event and the abnormality discrimination flag is abnormal, extracting the standardized generator set multi-channel working condition input vector x of the current time step k = [v" k,1 ,v" k,2 ,…,v" k,C ] T , and recording the multi-modal residual spectrum corresponding to the current time step;

[0120] S52. Constructing a lightweight long short-term memory network prediction model LSTM lite as the explanation objective function of the forward model Input the standardized generator set multi-channel working condition input vector of the current time step into the forward model, and calculate the prediction contribution degree of each working condition feature in the standardized generator set multi-channel working condition input vector using the integrated gradient attribution analysis algorithm. The prediction contribution degree represents the attribution intensity of each working condition feature to the model output prediction error. The integrated gradient attribution analysis algorithm accumulates the partial derivative of the forward model output to each working condition feature on the path from the baseline input interpolation to the actual input of the standardized generator set multi-channel working condition input vector, multiplies the difference between the working condition feature and the baseline input, and obtains the corresponding prediction contribution degree;

[0121] S53. Grouping the prediction contribution degrees of all working condition features into a feature contribution degree vector, and fusing it with the structural consistency features in the multi-modal residual spectrum of the current time step to obtain a generator set abnormality causal explanation graph with feature contribution degrees and structural consistency features. The structural consistency features are composed of correlation coefficients between all channels. All channel numbers are corresponded to nodes of the explanation graph. Each node is assigned its feature contribution degree, and the edge between any two nodes is assigned a structural consistency correlation coefficient.

[0122] S54. Rank all nodes in the generator set anomaly causal explanation graph according to feature contribution, select the top N channels with the highest feature contribution as the anomaly causal path nodes, and construct an anomaly causal path set in descending order of feature contribution;

[0123] S55. Structure and visualize the generator set anomaly causal explanation graph and the anomaly causal path set to generate a generator set anomaly causal explanation atlas, jointly package the generator set anomaly causal explanation atlas with the abnormal candidate event, and push it to the generator set operation and maintenance system for assisting dispatchers in fault confirmation and maintenance strategy decision-making.

[0124] Example 1: A 600 MW coal-fired power plant in East China No. 1 generator set is in a load operation state, the running load is about 580 MW, at 14:08 that day, the edge computing node deployed in the main control room of the unit began to detect that the prediction errors of the “main steam pressure” and “reheater outlet temperature” two channels continued to rise, which had exceeded the upper limit of the 90% confidence interval for 8 consecutive time steps.

[0125] The system records that the residual values of the two channels reach 0.126 and 0.147, respectively, which are about 2.6 times and 3.1 times higher than the historical stable period mean, respectively. The multi-modal residual atlas analysis found that the “spectrum energy density (3.3Hz-4.5Hz)” in the corresponding time window was amplified synchronously on the above two channels, and the system automatically marked this time as a suspected disturbance enhancement point.

[0126] The system automatically shortens the prediction step from the original 12 steps to 6 steps according to the state disturbance intensity index, in order to enhance the response ability of the prediction model to small disturbances, and recalculates the health index EHI as 0.782, which is higher than the historical 95% threshold of 0.733, and triggers the double threshold mechanism for the first time, judging that there is a potential anomaly.

[0127] Within 0.6 seconds after triggering the system, an abnormal candidate event is automatically generated and the explanation module is called, and the integrated gradient algorithm is used for attribution analysis of the current input vector (a total of 118 channels). The system found that the contribution degrees of “main steam flow”, “denitration ammonia injection flow” and “air preheater differential pressure” were significantly increased, reaching 0.284, 0.193 and 0.112, respectively, which were much higher than the average level of other channels (about 0.021).

[0128] The abnormal causal explanation atlas is visualized and uploaded to the dispatching and control platform by the system to generate a report file “EVD-20241018-01”. The atlas shows that the sudden increase in denitration ammonia injection flow causes the increase in ammonia escape, which leads to the increase in thermal deviation in the boiler tail heat exchange area, and then affects the stability of the reheater and the main steam system.

[0129] The dispatcher contacted the operation shift immediately after receiving the system report and called the on-site camera video to confirm that the ammonia injection control loop had a short-term fluctuation, the ammonia injection valve opening degree suddenly increased from the original 35% to 61%, and lasted for about 3 minutes. Further checking found that the sudden increase was caused by the amplification of the ammonia concentration feedback error.

[0130] The shift engineer manually intervened in the ammonia injection PID parameter and recalibrated the target value. The system monitored the residual of each main thermal channel for 5 minutes and found that the health index fell to 0.543, which was lower than the baseline threshold. The system determined that the abnormality was removed, and a closed-loop event record was formed.

[0131] This abnormality did not cause direct shutdown or alarm to the unit operation, but it was a typical weak disturbance-thermal deviation chain reaction event. If not intervened in time, it would evolve into tail heat exchanger blockage or main steam temperature overrun.

[0132] In this cycle, the system processed 2.136 million data samples. Compared with the traditional SCADA fixed threshold mechanism:

[0133] Table 1 Comparison of traditional SCADA fixed threshold mechanism

[0134] Comparison index System of the present invention Traditional SCADA discrimination mechanism Detection response time (seconds) 5.6 42.3 Abnormal report generation time (seconds) 1.8 Unable to generate Abnormal positioning accuracy (main channel) ± 2 channels ± 7 channels Abnormal intervention effective window 3.7 minutes <1 minute or no warning Abnormal explanation atlas generation Yes No

[0135] Based on the historical data set of the power plant in the past three months (containing about 120 million time series records), after using the adaptive pruning mechanism, the LSTMlite inference time is reduced by 61%, and the GPU resource occupation of the edge device is not more than 12%. The system can run efficiently. Within 7 days after deployment, the system cumulatively determined 12 candidate abnormal events, of which 9 were confirmed by the operation and maintenance department to have early signals of working condition fluctuation.

[0136] The present application deploys a lightweight long short-term memory network prediction model on an industrial edge computing node, and introduces an adaptive rolling prediction mechanism based on state disturbance and a pruning importance constraint mechanism. While ensuring the real-time performance of the model, the prediction step length is dynamically adjusted, which can more accurately capture the early signals of device state mutation or fluctuation. The prediction range can be shortened or lengthened adaptively according to the hidden state disturbance intensity index, thereby improving the sensitivity when the working condition changes dramatically and reducing the false alarm rate when running smoothly. In addition, by constraining the network structure through the pruning importance index, only the weight parameters sensitive to the state disturbance are retained, reducing the redundant calculation of the model and improving the deployment efficiency of the edge.

[0137] The application firstly constructs a multi-modal residual atlas fusing residual statistics, spectral features and structural consistency, and comprehensively depicts the micro-fluctuation characteristics and channel structure coordination of the generator set operation state by combining the residual atlas, relative change atlas and inter-channel Pearson correlation to form a multi-dimensional fusion atlas, which is particularly effective in discriminating critical states or early weak faults. The health index based on the atlas can not only reflect single-point abnormalities, but also depict the structural instability trend at the system level. Combined with the adaptive statistical threshold, a double-threshold discrimination mechanism is formed to effectively improve the detection accuracy.

[0138] The application introduces an abnormal causal explanation graph generation algorithm based on an integrated gradient attribution mechanism, combines the contribution of the input multi-channel working condition features to the forward prediction model for explainability analysis, constructs a structured abnormal explanation atlas with causal relationships, automatically generates an abnormal causal path set by analyzing the fusion results of the feature contribution and the structural consistency atlas, and significantly improves the traceability and operability of fault location, thereby assisting operation and maintenance personnel in accurately formulating maintenance strategies.

[0139] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes within the technical range disclosed by the application according to the technical solutions and inventive concepts of the application, which should be covered within the protection scope of the application.

Claims

1. An edge computing-based generator set anomaly detection method, characterized in that, Comprise the following steps: S1. Collecting and preprocessing the multi-channel operating condition data of the generator set on-site deployment of industrial edge computing nodes, generating standardized multi-channel operating condition data of the generator set; S2. Deploying a lightweight long short-term memory network prediction model in the industrial edge computing node, the lightweight long short-term memory network prediction model receiving the standardized multi-channel operating condition data of the generator set and outputting a rolling predicted generator set state estimation sequence and a corresponding prediction confidence interval; S3. Calculate the difference between the standardized multi-channel operating condition data of the generator set and the rolling predicted generator set state estimation sequence, obtain the local residual sequence, and calculate the statistical characteristics and spectral energy by applying a sliding time window, construct a multi-modal residual atlas; S4. Calculate the equipment health index based on the residual atlas and historical benchmark data, form the global threshold baseline of the equipment health index, generate the adaptive local statistical threshold according to the current statistical distribution of the residual atlas, fuse the global threshold baseline of the equipment health index and the adaptive local statistical threshold to construct a double threshold discrimination mechanism, and output the generator set abnormal candidate event through the generator set abnormal discrimination of the local residual sequence; S5. Using an explanatory analysis algorithm to generate a generator set abnormal causal explanation graph, which reveals the contribution relationship of each operating condition feature to the generator set abnormal confidence.

2. The method of claim 1, wherein, The S1 comprises the following steps: S11. Collecting real-time operating condition raw data of the generator set control system, sensor network and SCADA system through the industrial edge computing node deployed on-site, obtaining each piece of multi-channel operating condition raw data of the generator set, and jointly forming a multi-channel operating condition raw data set of the generator set; S12. Time alignment processing is performed on the multi-channel operating condition raw data set to obtain a time-aligned multi-channel operating condition data matrix; S13. For the time-aligned multi-channel operating condition data matrix, the sliding average model is used to extract the long-term trend component for each channel data sequence, and the trend-removed operating condition data residual value is obtained by subtracting the sliding average trend component of the corresponding channel at the time step from each measurement value; S14. The trend-removed operating condition data residual value is normalized by using the Z-score normalization method for normalization preprocessing to obtain normalized operating condition data; S15. All normalized operating condition data are combined to form a standardized operating condition data matrix.

3. The method of claim 2, wherein the method further comprises: The S2 comprises the following steps: S21. Group the data of each time step in the standardized working condition data matrix into a time step input vector in channel order, each time step input vector containing the standardized measurement values of all channels, for each time step input vector, calculate a channel attention weight coefficient vector using the weight parameter and the bias parameter, multiply the channel attention weight coefficient vector and the time step input vector element by element to obtain a channel perception weighted input vector S22. The channel-aware weighted input vector input light-weight long short-term memory network prediction model LSTM lite At each time step, the hidden state h k is updated as follows: h The hidden state window is constructed from the hidden state sequence of the last W S23. Based on the hidden state window H k The residual perturbation changes in the hidden state are used to construct the state perturbation intensity index S. k The state disturbance intensity index is used to measure the overall magnitude of hidden state changes within the current window. It is based on the state disturbance intensity index S within the current window. k Given a specified maximum prediction step size and control coefficient, the rolling prediction step size H is adaptively adjusted using an exponential decay method. p (k), the prediction step size is automatically shortened when the state disturbance intensity index is greater than the threshold, and automatically extended when the state disturbance intensity index is less than the threshold; S24. Based on the state disturbance intensity index S k At the same time, the pruning importance constraint index is constructed. S25. Predicting model LSTM using dynamic pruned light-weight long short-term memory network at current time step k lite with a rolling prediction horizon H p (k) outputting a sequence of future generator set state predictions combined with the most recent W h step state disturbance intensity indicator S k computing a confidence intensity indicator γ k = 1 - tanh(ηS k ), and generating a prediction confidence interval: where σ is the historical residual standard deviation, γ k reflects the dynamic adjustment capability of the prediction confidence with the disturbance change; S26. Store the predicted sequence and the corresponding prediction confidence interval {I k+h in the edge node cache.

4. The method according to claim 3, wherein, The pruning importance constraint index is to calculate the sensitivity of each weight parameter of the lightweight long short-term memory network prediction model to the state disturbance intensity index, the sensitivity is defined as the product of the partial derivative of the weight parameter to the state disturbance intensity index and the weight parameter itself, all weight parameters with sensitivity greater than the pruning threshold are retained, and other weight parameters are set to zero.

5. The method of claim 3, wherein the method further comprises: The S3 comprises the following steps: S31. Based on the standardized generator set multi-channel working condition data matrix and the rolling prediction generator set state prediction sequence, the prediction error of each time step is calculated to obtain a local residual sequence where x k+h is the actual measurement vector at time step k+h, is the corresponding predicted value vector; S32. The local residual sequence e k+h Constructing a single-channel residual sequence e Within a sliding window length W r Compute the sliding mean of the residual e And the sliding variance of the residual e S33. Applying a short-time Fourier transform to each of the channel residual sequences, calculating a spectral energy density value E of the channel at the frequency component and the time window position for each frequency component and each time window position (j) (f,t); S34. Jointly encode the residual sliding mean, residual sliding variance and spectral energy density values of all channels within the current sliding window, to construct a residual atlas R for comprehensively characterizing the statistical and frequency-domain features of the residual signals of all channels at the current time step k ; S35. Calculate the relative change rate of each channel current measurement value relative to the mean value of the current channel in the same sliding time window: wherein, is the mean value of the most recent stable window for channel j, v" k,j denotes the measured value of the normalized multi-channel operating condition data of the generator set at the current time step k and channel number j, and ∈ is a constant; The relative change rates of all channels are composed into a working condition relative change map D k ; S36. Within the same sliding time window, compute the Pearson correlation coefficient p between all pairs of channel residual time series i,j The Pearson correlation coefficients p between all pairs of channels are assembled into a structure consistency map C k ; S37. The residual atlas R k , the working condition relative change atlas D k , the structure consistency atlas C k are fused to form a multi-modal residual atlas 6. The edge computing-based generator set anomaly detection method of claim 5, wherein, The S4 comprises the following steps: S41. Based on the multi-modal residual atlas, sum the residual sliding mean absolute value of all channels, sum the residual sliding variance, sum the spectral energy density value under all frequencies and time windows, and according to the weighted aggregation coefficient vector w = [w1, w2, w3] T The weighted accumulation obtains the multi-modal residual comprehensive index A of the current time step k ; S42. Under the action of the set exponential smoothing coefficient a, the generator set equipment health index EHI is recursively calculated according to the historical multi-modal residual comprehensive index sequence {A k-1 ,A k-2 ,…} EHI k = aA k +(1-a)EHI k-1 , wherein EHI0 takes the average of {A -N} in the historical reference period, and EHI k The greater the value is, the lower the overall generator set equipment health level is. S43. Utilize a historical sliding window W g Max-min normalize the health index of the generator set equipment within the window to obtain a normalized health score Multiply the normalized health score by a preset amplification coefficient and add a safety margin constant to map and generate a global health threshold baseline G k ; S44. Within the current sliding window, calculate the local statistical mean and the local statistical standard deviation of the multi-modal residual synthesis index, multiply the local statistical standard deviation by the quantile constant corresponding to the significance level under the standard normal distribution, and add it to the local statistical mean to generate an adaptive local statistical threshold L k ; S45. Merge the global health threshold baseline with the adaptive local statistical threshold to construct a double-threshold discrimination mechanism, and obtain a determination result Flag k ; S46. Flag = 1 if the result of S45 is positive k with the corresponding multimodal residual synthesis index A k , a global threshold G k and a local threshold L k packaging the output abnormal candidate events 7. The edge computing-based generator set anomaly detection method of claim 6, wherein, The double-threshold discrimination mechanism is specifically: if the multi-modal residual comprehensive index of a certain time step simultaneously exceeds the global health threshold baseline and the adaptive local statistical threshold, it is determined that the corresponding time step has a generator set abnormality, otherwise it is considered to have no abnormality.

8. The edge computing-based generator set anomaly detection method of claim 6, wherein, The S5 includes the following steps: S51. When an abnormal candidate event is received and the abnormality discrimination flag is abnormal, the standardized generator set multi-channel operating condition input vector x of the current time step is extracted k = [v" k,1 ,v" k,2 ,…,v" k,C ] T , and the multi-modal residual atlas corresponding to the current time step is recorded; S52. Constructing a lightweight long short-term memory network prediction model LSTM lite The explanation objective function of the forward model The normalized multi-channel operating condition input vector of the current time step is input into the forward model, and the integrated gradient attribution analysis algorithm is used to calculate the prediction contribution degree of each operating condition feature in the normalized multi-channel operating condition input vector. S53. The predicted contribution degrees of all working condition characteristics are combined into a feature contribution degree vector, and are fused with the structural consistency features in the multi-modal residual map of the current time step to obtain a generator set abnormality causal explanation graph with feature contribution degrees and structural consistency features; S54. All nodes in the generator set abnormality causal explanation graph are ranked according to the feature contribution degrees, the top N channel nodes with the highest feature contribution degrees are selected as the abnormality causal path nodes, and an abnormality causal path set is constructed according to the order from high to low of the feature contribution degrees; S55. Abnormal causal explanation graph of the generator set and an abnormal causal path set The structured visual coding is performed to generate the abnormal causal explanation graph of the generator set, which is used for assisting the dispatch personnel in fault confirmation and maintenance strategy decision-making.

9. The edge computing-based generator set anomaly detection method of claim 7, wherein, The integrated gradient attribution analysis algorithm accumulates the partial derivative of the forward model output to each working condition characteristic on the path from the baseline input interpolation to the actual input of the standardized generator set multi-channel working condition input vector, multiplies the difference between the working condition characteristic and the baseline input, and obtains the corresponding predicted contribution degree.

Images