Power equipment data anomaly detection method and system based on deep residual network

Abstract

The present invention relates to the technical field of data anomaly detection, specifically to a method and system for detecting anomaly in power equipment data based on a deep residual network. The method obtains multidimensional time series data of the power equipment and performs preprocessing; a deep residual network model is constructed to analyze the multidimensional time series data to obtain abnormal data of the power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoding abnormality metric analysis unit, and a lightweight reasoning optimization unit; an abnormality interpretability analysis is performed based on the abnormal data of the power equipment, including locating the abnormal data of the power equipment, generating an abnormal heat map of the power equipment, and generating the fault type of the power equipment. The present invention improves the accuracy of anomaly identification by analyzing the power equipment data based on a deep residual network.

Description

Technical Field

[0001] The present invention relates to the technical field of data anomaly detection, and specifically to a method and system for detecting anomaly in power equipment data based on a deep residual network. Background Art

[0002] The safe and stable operation of power systems is crucial to the national economy and social well-being. With the advancement of sensor technology and data acquisition systems, power equipment generates massive amounts of high-dimensional operational data. This data, in the form of time series, records the dynamic characteristics of the equipment in various aspects, such as voltage, current, temperature, vibration, and partial discharge. Accurately and promptly detecting abnormal conditions from this complex, high-dimensional time series data is a major technical challenge facing the power industry.

[0003] Prior research and application of anomaly detection methods for power equipment data have been conducted. However, power equipment data typically exhibits highly nonlinear, multi-component coupled, and complex dynamic time series characteristics. When faced with high-dimensional data, these traditional methods often struggle to effectively extract deep features and hidden correlations that fully reflect subtle changes in equipment status. This results in feature information loss or redundancy, resulting in insufficient anomaly detection accuracy and sensitivity.

[0004] Traditional machine learning methods, such as support vector machines, K-nearest neighbors, decision trees, and ensemble learning methods, have also been applied to power equipment anomaly detection. However, for high-dimensional, complex time series data, designing highly discriminative features is a laborious and expert-reliant task. This makes it difficult to ensure the completeness and optimality of the extracted features, limiting the model's ability to extract sufficiently rich discriminative information from high-dimensional data and hindering further improvement in anomaly detection accuracy.

[0005] Furthermore, in practical applications, power systems place extremely high demands on real-time anomaly detection. Many complex models incur high computational overhead during training and inference, potentially failing to meet the real-time monitoring needs of large-scale power equipment clusters. Many supervised learning methods rely on large amounts of high-quality, precisely labeled (normal / abnormal) training data. In the power equipment sector, abnormal events themselves are low-probability events, making obtaining comprehensive and accurately labeled anomaly samples extremely difficult and costly. Data quality issues, such as noise and missing values, can also significantly impact model performance, increasing the complexity and implementation cost of data preprocessing.

[0006] To this end, a power equipment data anomaly detection method and system based on deep residual network is proposed. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and system for detecting anomalies in power equipment data based on a deep residual network. The method obtains multidimensional time series data of power equipment and performs preprocessing; constructs a deep residual network model to analyze the multidimensional time series data to obtain power equipment anomaly data; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoder anomaly metric analysis unit, and a lightweight inference optimization unit; and performs anomaly interpretability analysis based on the power equipment anomaly data, including locating the power equipment anomaly data, generating a power equipment anomaly heat map, and generating power equipment fault types. The present invention improves the accuracy of anomaly identification by analyzing power equipment data based on a deep residual network.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The power equipment data anomaly detection method based on deep residual network includes:

[0010] Acquire multi-dimensional time series data of power equipment and perform preprocessing;

[0011] Constructing a deep residual network model to analyze the multidimensional time series data to obtain abnormal data of power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoding abnormality measurement analysis unit and a lightweight reasoning optimization unit;

[0012] Anomaly explainability analysis is performed based on the abnormal data of the power equipment, including locating the abnormal data of the power equipment, generating an abnormal heat map of the power equipment, and generating a fault type of the power equipment.

[0013] Preferably, the multidimensional time series data includes voltage, current, temperature, vibration and partial discharge; and the preprocessing includes data cleaning, normalization, missing value interpolation and time window segmentation.

[0014] Preferably, the multi-scale residual analysis unit includes a multi-branch parallel structure feature extraction layer, a depth-separable convolution layer configuration layer, a residual connection path setting layer and a feature fusion layer; the multi-branch parallel structure feature extraction layer obtains the short-term time dependence features, medium-term time dependence features and long-term time dependence features of the power equipment by adopting convolution kernels of different sizes for each branch; the depth-separable convolution layer configuration layer reduces the computational complexity of the short-term time dependence features, medium-term time dependence features and long-term time dependence features of the power equipment based on the depth-separable convolution layer; the residual connection path setting layer connects the multi-branch input features to the branch output end by direct addition through the residual connection path, and retains the original signal of the power equipment in the deep network; the feature fusion layer generates the comprehensive characteristics of the operating status of the power equipment by adaptively fusing the short-term time dependence features, medium-term time dependence features and long-term time dependence features of the power equipment.

[0015] Preferably, the adaptive dimensional attention mechanism analysis unit includes a global information extraction layer, a channel attention module construction layer and an inter-dimensional dependency modeling layer; the global information extraction layer calculates the statistical features of each dimension of the multidimensional time series data through global average pooling and global maximum pooling; the channel attention module construction layer is composed of two layers of fully connected networks, which compresses the statistical feature input dimension and restores the compressed statistical features to the input dimension, generates the importance score of each dimension and applies the Sigmoid activation function for processing; the inter-dimensional dependency modeling layer calculates the correlation matrix between different parameters of the power equipment based on the importance scores of each dimension, obtains the coupling relationship between the parameters, and verifies the comprehensive characteristics of the operating status.

[0016] Preferably, the autoencoding anomaly measurement analysis unit includes an autoencoding network structure of an encoder and a decoder, wherein the encoder extracts the implicit features of the normal operating state of the power equipment based on a multi-scale residual analysis unit, and the decoder uses a transposed convolution structure to reconstruct the implicit features into original input data; a joint optimization objective function of reconstruction loss and contrast loss is constructed to analyze the comprehensive features and implicit features of the operating state to obtain minor anomaly data of the power equipment; through multi-scale reconstruction error measurement, combined with the time domain reconstruction error and frequency domain reconstruction error of the power equipment data, the degree of anomaly of the minor anomaly data of the power equipment is obtained.

[0017] Preferably, the lightweight reasoning optimization unit includes a network pruning layer, a network parameter quantization layer, a knowledge distillation analysis layer and an reasoning acceleration layer;

[0018] The network pruning layer calculates the sensitivity score of each branch to the degree of abnormality of the minor abnormal data of the power equipment output by the self-encoding abnormality measurement analysis unit, sorts the network connections according to the sensitivity score, and removes the connections with sensitivity lower than the threshold according to the preset threshold; the sensitivity score is obtained by observing the comprehensive operating status characteristics output by the model after perturbing and / or shielding the connection; the network parameter quantization layer determines the optimal quantization scaling factor by performing statistical analysis on the comprehensive operating status characteristics to obtain the comprehensive abnormal status data; the knowledge distillation analysis layer identifies the comprehensive abnormal status data by combining the comprehensive distillation objective function of hard label loss, soft label loss and feature representation loss to generate abnormal characteristics of the power equipment; the inference acceleration layer optimizes the abnormal characteristics of the power equipment through batch processing optimization, computational graph optimization and memory access mode optimization to generate abnormal data of the power equipment.

[0019] Preferably, the anomaly explainability analysis includes an anomaly location layer, an anomaly heat map generation layer, and a fault type mapping layer;

[0020] The anomaly location layer analyzes the abnormal data of the power equipment based on the gradient weighted class activation mapping technology, performs global average pooling on the gradient of the final feature map and the gradient of each channel to obtain the channel importance weight, combines the weight with the corresponding feature activation map weight to generate a class activation heat map, and upsamples the class activation heat map to the same time dimension as the input data to locate the time point and data dimension that contribute most to the abnormal data of the power equipment; the anomaly heat map generation layer uses a two-dimensional visualization method for the abnormal data of the power equipment, and uses a gradient color scale to represent the degree of abnormality, where dark areas represent high abnormality and light areas represent low abnormality; the fault type mapping layer obtains the similarity between the abnormal pattern of the abnormal data of the power equipment and each fault pattern in the knowledge base based on the fault feature pattern description stored in the power equipment expert knowledge base, and identifies the fault type.

[0021] The power equipment data anomaly detection system based on deep residual network includes:

[0022] Data acquisition module, used to obtain multi-dimensional time series data of power equipment and perform preprocessing;

[0023] A deep residual network analysis module is used to construct a deep residual network model to analyze the multi-dimensional time series data and obtain abnormal data of power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoding abnormality measurement analysis unit and a lightweight inference optimization unit;

[0024] The abnormality explainability analysis module is used to perform abnormality explainability analysis based on the abnormality score of the power equipment, including locating abnormal data of the power equipment, generating an abnormal heat map of the power equipment, and generating a fault type of the power equipment.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention constructs a deep residual network model that integrates multi-scale residual analysis, adaptive dimensional attention mechanism, autoencoding anomaly measurement and lightweight inference optimization, and conducts joint modeling and analysis of multi-dimensional time series data including voltage, current, temperature, vibration, partial discharge, etc., so as to more comprehensively describe the operating status of the equipment. The short-term, medium-term and long-term time-dependent features are extracted through multi-scale convolution, and the original features are retained in combination with the residual structure, which significantly enhances the feature capture capability of the model. At the same time, the autoencoding structure jointly optimizes the implicit features of the normal state and the reconstruction error, and integrates the reconstruction errors of the time domain and frequency domain to effectively identify minor anomalies. The overall detection accuracy is higher, especially in the early stage of the anomaly or in the presence of interference data, it has good detection sensitivity and robustness, which improves the reliability and practicality of power equipment data anomaly detection.

[0027] 2. Based on the realization of accurate anomaly detection, the present invention further enhances the interpretability of anomaly results. Through the anomaly interpretability analysis module, a gradient-weighted class activation mapping mechanism is introduced to perform channel-level importance weighting on the final feature map, generate a class activation heat map, and locate the most critical data dimensions and time points. The heat map clearly reflects the degree of anomaly in different dimensions through gradient color visualization, making it easier for operation and maintenance personnel to quickly understand the distribution of anomalies. At the same time, similarity matching is performed in combination with the fault modes defined in the power equipment expert knowledge base to achieve the identification and classification of specific fault types, thereby enhancing the practical value of the detection results.

[0028] 3. The present invention integrates four strategies, namely network pruning, parameter quantization, knowledge distillation and reasoning acceleration, by setting up a lightweight reasoning optimization unit to compress and accelerate the model from a structural level. Specifically, network pruning removes redundant connections based on output sensitivity, effectively reducing computational complexity; parameter quantization achieves low-precision compression of feature data, reducing model storage space; knowledge distillation transfers the feature representation of the teacher model to the student model, achieving a lightweight network with similar performance but a more compact structure; and reasoning acceleration optimizes the computational graph and memory access mode to improve operational efficiency. These optimization methods work together to enable the model proposed in the present invention to run stably in resource-constrained embedded devices or edge computing terminals, supporting real-time online anomaly detection of power field equipment and meeting the stringent requirements of actual deployment and industrial-grade applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1A schematic diagram of the process flow of the power equipment data anomaly detection method based on deep residual network provided by the present invention;

[0030] Figure 2 A schematic diagram of the structure of the power equipment data anomaly detection system based on the deep residual network provided by the present invention;

[0031] Figure 3 Schematic diagram of the deep residual network model structure provided by the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] The present invention provides a method for detecting anomaly in power equipment data based on a deep residual network. This method is applied to a system for detecting anomaly in power equipment data based on a deep residual network. The flowchart of the specific method and system is shown in FIG. Figure 1 and Figure 2 .

[0034] Example 1:

[0035] See also Figure 1 The present invention provides a method for detecting anomaly in power equipment data based on a deep residual network. The technical solution is as follows:

[0036] Acquire multi-dimensional time series data of power equipment and perform preprocessing;

[0037] Acquire multi-dimensional time series data of power equipment and perform preprocessing;

[0038] Construct a deep residual network model to analyze the multi-dimensional time series data to obtain abnormal data of power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimension attention mechanism analysis unit, an autoencoding abnormality measurement analysis unit and a lightweight reasoning optimization unit, see Figure 3 ;

[0039] Anomaly explainability analysis is performed based on the abnormal data of the power equipment, including locating the abnormal data of the power equipment, generating an abnormal heat map of the power equipment, and generating a fault type of the power equipment.

[0040] Specifically, the multidimensional time series data includes voltage, current, temperature, vibration and partial discharge; the preprocessing includes data cleaning, normalization, missing value interpolation and time window segmentation.

[0041] In this embodiment, by introducing five types of multidimensional time series data, namely voltage, current, temperature, vibration and partial discharge, and comprehensively implementing data cleaning, normalization, missing value interpolation and time window segmentation in the preprocessing stage, the present invention can more comprehensively characterize the operating status of power equipment under the complementary effect of multi-source information, and significantly improve the sensitivity and accuracy of anomaly detection; the normalization and missing value interpolation steps eliminate the interference caused by dimensional differences and incomplete data, ensuring the continuity and consistency of model input; and time window segmentation enables the model to capture the dynamic change characteristics of the equipment at different time scales, thereby more timely and reliable identification and location of potential fault modes.

[0042] Furthermore, the multi-scale residual analysis unit includes a multi-branch parallel structure feature extraction layer, a depth-separable convolution layer configuration layer, a residual connection path setting layer and a feature fusion layer; the multi-branch parallel structure feature extraction layer obtains the short-term time dependence features, medium-term time dependence features and long-term time dependence features of the power equipment by adopting convolution kernels of different sizes for each branch; the depth-separable convolution layer configuration layer reduces the computational complexity of the short-term time dependence features, medium-term time dependence features and long-term time dependence features of the power equipment based on the depth-separable convolution layer; the residual connection path setting layer connects the multi-branch input features to the branch output end by direct addition through the residual connection path, and retains the original signal of the power equipment in the deep network; the feature fusion layer generates the comprehensive characteristics of the operating status of the power equipment by adaptively fusing the short-term time dependence features, medium-term time dependence features and long-term time dependence features of the power equipment.

[0043] In this embodiment, by introducing a multi-branch parallel structure and a depth-separable convolution into the multi-scale residual analysis unit, it is possible not only to extract features for the short-term, medium-term, and long-term time dependencies of power equipment, but also to significantly reduce the computational complexity while maintaining the model's expressiveness; the introduction of residual connection paths effectively avoids information loss and gradient degradation in deep networks, allowing key details of the original signal to be losslessly transmitted between layers; and the adaptive feature fusion mechanism can dynamically weigh dependency features at different time scales to generate more comprehensive and discriminative operating status comprehensive features. Therefore, while improving the accuracy and robustness of anomaly detection, the present invention takes into account the lightweight and real-time performance of the model, significantly enhancing the ability to identify fault modes under complex working conditions.

[0044] Furthermore, the adaptive dimensional attention mechanism analysis unit includes a global information extraction layer, a channel attention module construction layer and an inter-dimensional dependency modeling layer; the global information extraction layer calculates the statistical features of each dimension of the multidimensional time series data through global average pooling and global maximum pooling; the channel attention module construction layer is composed of two layers of fully connected networks, which compresses the statistical feature input dimension and restores the compressed statistical features to the input dimension, generates the importance score of each dimension and applies the Sigmoid activation function for processing; the inter-dimensional dependency modeling layer calculates the correlation matrix between different parameters of the power equipment based on the importance scores of each dimension, obtains the coupling relationship between the parameters, and verifies the comprehensive characteristics of the operating status.

[0045] In this embodiment, by introducing an adaptive dimensional attention mechanism, the present invention is able to grasp the global statistical characteristics of multi-dimensional parameters as a whole, and based on this, dynamically assign the importance weights of each channel, thereby highlighting key dimensions and suppressing redundant information in feature expression; the compression-then-recovery structure of the two-layer fully connected network not only improves the nonlinear modeling ability of channel attention, but also effectively reduces the number of parameters and computational overhead; the inter-dimensional dependency matrix subsequently constructed further reveals the coupling and constraint relationship between parameters such as voltage, current, temperature, vibration and partial discharge, providing an internal consistency test for the comprehensive features. After integrating these mechanisms, the model shows higher robustness and discrimination in capturing weak abnormal signals and dealing with multi-parameter interaction interference, significantly improving the accuracy and interpretability of anomaly detection while taking into account computational efficiency.

[0046] Furthermore, the autoencoding anomaly measurement analysis unit includes an autoencoding network structure of an encoder and a decoder, wherein the encoder extracts implicit features of the normal operating state of the power equipment based on a multi-scale residual analysis unit, and the decoder uses a transposed convolution structure to reconstruct the implicit features into original input data; a joint optimization objective function of reconstruction loss and contrast loss is constructed to analyze the comprehensive features and implicit features of the operating state to obtain minor anomaly data of the power equipment; through multi-scale reconstruction error measurement, combined with the time domain reconstruction error and frequency domain reconstruction error of the power equipment data, the degree of anomaly of the minor anomaly data of the power equipment is obtained.

[0047] In this embodiment, by combining multi-scale residual feature extraction with the autoencoder network, the present invention can accurately mine the implicit features of the normal operating state of the power equipment, and use the transposed convolution decoder to reconstruct the original input with high fidelity, thereby generating a significant reconstruction error signal when a minor anomaly occurs; the optimization goal of the joint reconstruction loss and contrast loss not only improves the model's sensitivity to the distribution difference between normal and abnormal samples, but also enhances the ability to discriminate implicit features, making the capture of weak fault signs more reliable; in addition, by fusing the multi-scale reconstruction error metrics of the time domain and frequency domain, the system can comprehensively characterize the performance of anomalies in different feature spaces, and then accurately quantify the degree of anomaly of minor anomaly data, significantly improving the accuracy of anomaly detection and early warning capabilities.

[0048] Furthermore, the lightweight inference optimization unit includes a network pruning layer, a network parameter quantization layer, a knowledge distillation analysis layer, and an inference acceleration layer;

[0049] The network pruning layer calculates the sensitivity score of each branch to the degree of abnormality of the minor abnormal data of the power equipment output by the self-encoding abnormality measurement analysis unit, sorts the network connections according to the sensitivity score, and removes the connections with sensitivity lower than the threshold according to the preset threshold; the sensitivity score is obtained by observing the comprehensive operating status characteristics output by the model after perturbing and / or shielding the connection; the network parameter quantization layer determines the optimal quantization scaling factor by performing statistical analysis on the comprehensive operating status characteristics to obtain the comprehensive abnormal status data; the knowledge distillation analysis layer identifies the comprehensive abnormal status data by combining the comprehensive distillation objective function of hard label loss, soft label loss and feature representation loss to generate abnormal characteristics of the power equipment; the inference acceleration layer optimizes the abnormal characteristics of the power equipment through batch processing optimization, computational graph optimization and memory access mode optimization to generate abnormal data of the power equipment.

[0050] In this embodiment, by integrating lightweight optimization strategies such as network pruning, parameter quantization, knowledge distillation, and inference acceleration, the present invention significantly reduces the computational and storage overhead of deep residual networks, while maintaining or even improving anomaly detection performance. Sensitivity-driven pruning can remove redundant connections to ensure that the model retains key feature paths after compression; parameter quantization based on statistical analysis maps floating-point weights to low-bit-width representations while minimizing precision loss, significantly reducing model storage and bandwidth requirements; knowledge distillation uses multiple constraints of soft and hard labels and feature representations to enable the streamlined student model to inherit the discriminative ability of the teacher model and remain highly sensitive to the recognition of abnormal patterns; in addition, the joint optimization of batch processing with computational graphs and memory access patterns accelerates the online inference process and provides millisecond-level response for on-site deployment. Overall, this unit not only achieves lightweight and efficient power equipment anomaly detection models, but also takes into account real-time and accuracy, meeting application requirements in resource-constrained industrial environments.

[0051] The anomaly explainability analysis includes an anomaly location layer, an anomaly heat map generation layer, and a fault type mapping layer;

[0052] The anomaly location layer analyzes the abnormal data of the power equipment based on the gradient weighted class activation mapping technology, performs global average pooling on the gradient of the final feature map and the gradient of each channel to obtain the channel importance weight, combines the weight with the corresponding feature activation map weight to generate a class activation heat map, and upsamples the class activation heat map to the same time dimension as the input data to locate the time point and data dimension that contribute most to the abnormal data of the power equipment; the anomaly heat map generation layer uses a two-dimensional visualization method for the abnormal data of the power equipment, and uses a gradient color scale to represent the degree of abnormality, where dark areas represent high abnormality and light areas represent low abnormality; the fault type mapping layer obtains the similarity between the abnormal pattern of the abnormal data of the power equipment and each fault pattern in the knowledge base based on the fault feature pattern description stored in the power equipment expert knowledge base, and identifies the fault type.

[0053] In this embodiment, by introducing an anomaly location mechanism based on gradient-weighted class activation mapping, the present invention can accurately identify the time points and parameter dimensions that have the greatest impact on the detection results, providing clear guidance on the source of the fault for operation and maintenance personnel; and combined with the visualization method of a two-dimensional gradient heat map, complex multi-dimensional anomaly information is presented in intuitive color depths, making the degree of anomaly clear at a glance, further shortening the on-site diagnosis time; finally, with the help of the fault mode description in the expert knowledge base, the detected anomaly mode is efficiently matched with the existing fault type through similarity calculation, which not only enhances the accuracy of the fault category determination, but also provides a reliable decision-making basis for the subsequent maintenance strategy selection. Overall, the interpretability analysis module of the present invention not only improves the transparency and comprehensibility of the anomaly detection results, but also greatly accelerates the fault diagnosis and response speed.

[0054] This invention comprehensively preprocesses multi-source time-series data, including voltage, current, temperature, vibration, and partial discharge, from power equipment operation. This includes data cleaning, normalization, missing value interpolation, and time window segmentation. This eliminates interference caused by dimensional differences and data incompleteness at the source, laying a solid foundation for stable training and efficient inference of subsequent deep models. The introduction of multi-scale residual branches and depthwise separable convolution not only enables the parallel mining of dynamic dependency features of equipment in the short, medium, and long term, but also ensures lossless information transfer and element complementarity through residual connections and adaptive fusion, fully preserving and amplifying the detailed signals of abnormal patterns. The adaptive dimensional attention mechanism dynamically assigns attention weights to each parameter channel based on global statistical features and verifies the coupling relationship between different physical quantities using a dependency matrix. This further eliminates the redundant effects of multi-parameter interactions and improves the ability to identify subtle anomalies. The autoencoder anomaly measurement unit, through multi-scale reconstruction and contrast loss, can more clearly distinguish between normal and abnormal sample distributions. It also integrates time-domain and frequency-domain errors to provide a more refined quantitative indicator of the degree of anomaly, significantly enhancing the sensitivity of early warning. Combining network pruning, parameter quantization, knowledge distillation and reasoning acceleration strategies, the present invention significantly compresses the model size and computational complexity while maintaining or even improving detection performance and response speed, achieving millisecond-level online reasoning in the restricted hardware environment of industrial sites. Finally, based on gradient-weighted class activation mapping and two-dimensional heat map visualization, it can not only accurately locate the most contributing fault time points and channel dimensions, but also quickly match abnormal patterns to fault types with the help of expert knowledge bases, greatly improving diagnostic transparency and decision-making efficiency. Overall, the method of the present invention has the advantages of high precision, high robustness, strong interpretability and lightweight deployment, and effectively meets the actual needs of the power system for real-time, reliable anomaly detection and diagnosis. Please refer to Table 1 for details.

[0055] Table 1 compares the performance of the 3σ rule, random forest, standard CNN, and the deep residual network method proposed in this paper on key indicators.

[0056] Table 1 Performance comparison of power equipment data anomaly detection methods

[0057]

[0058] The core technical advantages of the present invention are mainly reflected in the following aspects:

[0059] The multi-scale residual analysis unit can simultaneously capture the short-term, medium-term, and long-term time-dependent characteristics of power equipment, and utilize residual connections to retain key information from the original signal. The adaptive dimensional attention mechanism analysis unit automatically learns the importance weights of different parameter dimensions, explores the complex coupling relationships between equipment parameters, and improves sensitivity to abnormal patterns. Combining multi-scale reconstruction error metrics in the time and frequency domains enables the model to identify subtle abnormal changes, significantly improving early fault warning capabilities. Through network pruning, parameter quantization, knowledge distillation, and inference acceleration, computational complexity and memory usage are significantly reduced while maintaining high detection performance, enabling real-time deployment of edge devices. The anomaly interpretability analysis module not only detects anomalies but also locates the time points and data dimensions that contribute the most, generates intuitive heat maps, and maps them to specific fault types, providing strong support for operation and maintenance decisions.

[0060] Example 2:

[0061] This embodiment provides a specific application of a method and system for detecting anomalies in power equipment data based on a deep residual network in a thermal power equipment scenario. Thermal power equipment, such as boilers, steam turbines, generators and their ancillary systems in thermal power plants, is an important part of the power system, and the stability and safety of its operating status are of vital importance. This embodiment takes the boiler system in a thermal power plant as a specific application scenario for thermal power equipment, and explains in detail how the present invention is applied to the detection of anomalies in the operating data of the boiler system. Figure 2 ,Specifically: a data acquisition module is used to acquire multi-dimensional time series data of ,power equipment and perform pre-processing;

[0062] A deep residual network analysis module is used to construct a deep residual network model to analyze the multi-dimensional time series data and obtain abnormal data of power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoding abnormality measurement analysis unit and a lightweight inference optimization unit;

[0063] The abnormality explainability analysis module is used to perform abnormality explainability analysis based on the abnormality score of the power equipment, including locating abnormal data of the power equipment, generating an abnormal heat map of the power equipment, and generating a fault type of the power equipment.

[0064] Specifically, during the operation of the boiler system, a large amount of multidimensional time series data is generated. The present invention first obtains this data, which may include:

[0065] Voltage and current include the operating voltage and current data of the motors of boiler auxiliary equipment, such as induced draft fans, forced draft fans, feed water pumps, coal mills, etc.;

[0066] Temperature data include boiler furnace temperature, steam temperature of each level of superheater and reheater, exhaust gas temperature, furnace tube wall temperature, and bearing temperature key measurement point temperature data;

[0067] Vibration data includes bearing vibration and casing vibration data of the main rotating equipment of the boiler, such as fans, water pumps, coal mills, etc.

[0068] Pressure data such as drum pressure, main steam pressure, reheat steam pressure and furnace pressure;

[0069] Flow data such as feed water flow, fuel flow, steam flow and flue gas flow;

[0070] Partial discharge data refers to partial discharge monitoring data of high-voltage electrical components related to the boiler system, such as water feed pump motors, large fan motors, etc. Although the boiler itself is not a high-voltage component, its associated drive equipment involves high voltage and can therefore be included in the monitoring scope.

[0071] After obtaining the above multidimensional time series data, perform the following preprocessing operations:

[0072] Data cleaning removes obvious outliers or noise data caused by sensor failure, communication interference, etc. For example, for boiler wall temperature data, instantaneous jump values ​​that far exceed physical possibilities will be identified and processed;

[0073] Normalization is performed by scaling data with different physical dimensions and value ranges, such as temperature, pressure, and vibration values, to similar numerical ranges through maximum and minimum normalization or Z-score normalization, thereby eliminating the impact of dimensional differences on model training.

[0074] Missing value interpolation ensures data integrity by filling in missing data caused by temporary sensor failure or data transmission interruptions using mean interpolation, median interpolation, time series-based interpolation methods such as linear interpolation and spline interpolation, or predictive interpolation methods based on machine learning. For example, if the superheater outlet steam temperature data at a certain moment is missing, it can be interpolated based on the data before and after the moment and other relevant parameters such as fuel quantity and feedwater flow.

[0075] Time window segmentation divides continuous time series data into fixed time window lengths, forming a series of fixed-length data segments as input to the deep residual network model. This segmentation helps the model capture dynamic features and dependencies within a specific time scale.

[0076] The pre-processed multi-dimensional time series data is input into the constructed deep residual network model for analysis to obtain abnormal data of boiler equipment. The core units of the deep residual network model function as follows in the boiler system scenario:

[0077] The multi-scale residual analysis unit includes a multi-branch parallel structure feature extraction layer, a depth-separable convolution layer configuration layer, a residual connection path setting layer and a feature fusion layer;

[0078] The multi-branch parallel structure feature extraction layer targets the complex dynamic characteristics of boiler operating parameters. This layer uses convolution kernels of different sizes to extract features at different time scales in parallel. For example, small-sized convolution kernels can capture rapid fluctuations in boiler combustion conditions, such as short-term changes in flame detector signals and small, rapid pressure fluctuations. Medium-sized convolution kernels can analyze periodic changes or medium-term trends in parameters, such as the response characteristics of key parameters during load adjustment and the short-term improvement in efficiency after soot blowing on the heating surface. Large-sized convolution kernels focus on the long-term evolution of parameters, such as the gradual decline in heat transfer efficiency caused by gradual soot accumulation on the heating surface and the long-term trend changes in vibration signals caused by equipment wear. This method captures the short-term, medium-term, and long-term temporal dependencies of boiler operating conditions.

[0079] To ensure feature extraction capabilities while reducing the computational complexity of the model, the depthwise separable convolution layer uses depthwise separable convolution to process the temporal dependency data extracted by each branch. This is particularly important for boiler systems that require real-time monitoring of a large number of measurement points. It allows for efficient data processing within limited computing resources, yielding short-term, medium-term, and long-term temporal dependency features for the boiler.

[0080] Residual connection path setting layer: Through residual connections, the input features of each branch are directly superimposed on the output end, ensuring that important information in the boiler's original operating signal, such as the baseline values ​​of key temperatures and pressures, is not lost during the deep network transmission process, helping the model better learn small changes and deviations.

[0081] The feature fusion layer adaptively weights and fuses short-term, medium-term, and long-term time-dependent features extracted from different time scales to generate comprehensive features that fully reflect the current boiler's overall operating status. For example, when determining whether there is a water-wall tube leak, the model comprehensively considers short-term fluctuations in furnace pressure, drum water level, and the difference between feedwater flow and steam flow, as well as the long-term trends in the relevant wall temperature.

[0082] The adaptive dimension attention mechanism analysis unit includes a global information extraction layer, a channel attention module construction layer and an inter-dimensional dependency modeling layer;

[0083] The global information extraction layer performs global average pooling and global maximum pooling on the boiler's multi-dimensional time series data, such as temperature, pressure, flow, vibration, etc., in each time window, calculates the overall statistical characteristics of each monitoring dimension, and grasps the macroscopic state of each parameter.

[0084] The channel attention module construction layer utilizes a two-layer fully connected network to learn the importance scores of each monitoring channel parameter. For example, during the initial startup of a boiler, parameters such as water temperature and drum water level may be more important; whereas during stable full-load operation, parameters related to main steam temperature, pressure, and combustion stability may be more important. After processing with a sigmoid activation function, these scores are used to adjust the weights of each parameter feature.

[0085] The inter-dimensional dependency modeling layer calculates a correlation matrix between different boiler operating parameters based on the importance scores of each parameter. This helps the model understand the coupling and mutual constraints between parameters, such as the complex relationship between fuel supply rate, air supply rate, main steam temperature, and exhaust gas temperature. This method verifies and calibrates the comprehensive operating status features generated by the multi-scale residual analysis unit, improving the reliability of the features.

[0086] The autoencoder network structure consists of an encoder that learns the deep implicit features of the boiler under normal operating conditions, such as the normal temperature field distribution pattern, pressure fluctuation range, vibration spectrum characteristics, etc. The decoder attempts to use these implicit features to reconstruct the original boiler input data.

[0087] A joint optimization objective function uses a reconstruction loss to measure the difference between the reconstructed data and the original data, and a contrast loss to narrow the implicit features of normal samples and narrow the implicit features of abnormal samples. This joint optimization objective function analyzes the comprehensive and implicit features of the boiler's operating status. When a minor anomaly occurs in the boiler, such as a slight overheating of the heating surface tube wall or local temperature fluctuations caused by uneven combustion, the data pattern deviates from normal, increasing the reconstruction error and potentially being identified.

[0088] Multi-scale reconstruction error metric: This method combines time-domain reconstruction errors (such as direct deviations in temperature and pressure values) with frequency-domain reconstruction errors (such as changes in the spectral components of the vibration signal) to comprehensively assess the severity of minor anomalies. For example, early bearing wear may not be apparent in the time-domain vibration signal, but specific fault frequency components may appear in the frequency domain. Combining these two factors allows for a more accurate assessment of the anomaly's severity.

[0089] The lightweight reasoning optimization unit includes a network pruning layer, a network parameter quantization layer, a knowledge distillation analysis layer, and an reasoning acceleration layer;

[0090] The network pruning layer analyzes the sensitivity of each connection in the deep residual network model to the output of boiler anomaly detection results, removes connections and neurons that contribute less to model performance, and simplifies the model structure.

[0091] The network parameter quantization layer quantizes the floating-point parameters in the model, namely weights and activation values, into low-bit representations (such as 8-bit integers), reducing the model storage space and computational complexity. At the same time, it determines the optimal quantization scaling factor through statistical analysis to reduce accuracy loss.

[0092] The knowledge distillation analysis layer transfers knowledge from a pre-trained, more complex "teacher model" to a simpler "student model" (the currently optimized model). By combining hard label loss, soft label loss (the probabilistic output of the teacher model), and feature representation loss, the lightweight student model can approach the performance of the teacher model and effectively identify boiler anomaly characteristics.

[0093] Inference acceleration layer: Through batch processing optimization, computation graph optimization (operator fusion), and memory access pattern optimization (reducing data movement), the inference speed of the model in actual deployment is improved, ensuring rapid response to changes in boiler system status and timely generation of boiler anomaly data.

[0094] The anomaly explainability analysis includes an anomaly location layer, an anomaly heat map generation layer, and a fault type mapping layer;

[0095] Based on the abnormal data of boiler equipment output by the deep residual network model, anomaly explainability analysis is performed to provide decision support for operators:

[0096] Anomaly Localization Layer: Based on Gradient Weighted Class Activation Mapping (GLM), it analyzes the input data dimensions—sensor parameters and time points—that contribute most to the final anomaly determination. For example, if the system identifies a boiler water wall leak, this layer can pinpoint the area where the wall temperature sensor data first shows a significant drop, and the time at which the difference between the feedwater flow rate and the steam flow rate begins to increase dramatically.

[0097] The Abnormal Heat Map Generation Layer visualizes located abnormal information in a two-dimensional format, such as a heat map. For example, a temperature distribution heat map can be generated for a boiler furnace cross section or a specific heating surface expansion. Different colors, such as darker colors for high abnormality and lighter colors for low abnormality, clearly mark abnormal areas, such as localized high or low temperatures, allowing operations and maintenance personnel to intuitively understand the location and severity of the abnormality.

[0098] Fault type mapping layer: Based on a pre-built expert knowledge base of boiler equipment, which stores characteristic descriptions of various known fault modes, such as typical parameter change patterns corresponding to different types of tube wall leakage, unstable combustion, fan stall, and feedwater pump vaporization, the currently detected abnormal data pattern is compared with the fault patterns in the knowledge base through similarity calculations such as cosine similarity and Euclidean distance, thereby identifying the most likely specific fault type and assigning a corresponding confidence level. For example, if an abnormal drop in main steam temperature is detected, accompanied by a drop in reheat steam temperature and an increase in exhaust gas temperature, it may be mapped to a fault type such as "inadvertent closure of the boiler main steam valve" or "severe deterioration of combustion."

[0099] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting anomaly in power equipment data based on a deep residual network, characterized in that: include: Acquire multi-dimensional time series data of power equipment and perform preprocessing; The multidimensional time series data includes voltage, current, temperature, vibration and partial discharge; Constructing a deep residual network model to analyze the multidimensional time series data to obtain abnormal data of power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoding abnormality measurement analysis unit and a lightweight reasoning optimization unit; The multi-scale residual analysis unit includes a multi-branch parallel structure feature extraction layer, a depth-separable convolution layer configuration layer, a residual connection path setting layer and a feature fusion layer; the multi-branch parallel structure feature extraction layer obtains the short-term time dependency feature, medium-term time dependency feature and long-term time dependency feature of the power equipment by using convolution kernels of different sizes for each branch; the depth-separable convolution layer configuration layer reduces the computational complexity of the short-term time dependency feature, medium-term time dependency feature and long-term time dependency feature of the power equipment based on the depth-separable convolution layer; the residual connection path setting layer connects the multi-branch input features to the branch output end by direct addition through the residual connection path, and retains the original signal of the power equipment in the deep network; The feature fusion layer generates a comprehensive feature of the operating status of the power equipment by adaptively fusing the short-term time-dependent features, medium-term time-dependent features and long-term time-dependent features of the power equipment; The adaptive dimension attention mechanism analysis unit includes a global information extraction layer, a channel attention module construction layer and an inter-dimensional dependency modeling layer; The global information extraction layer calculates the statistical features of each dimension of the multidimensional time series data through global average pooling and global maximum pooling. The channel attention module construction layer consists of two layers of fully connected networks, which compress the input dimension of the statistical features and restore the compressed statistical features to the input dimension, generate the importance score of each dimension and apply the Sigmoid activation function. The inter-dimensional dependency modeling layer calculates the correlation matrix between different parameters of the power equipment based on the importance scores of each dimension, obtains the coupling relationship between the parameters, and verifies the comprehensive characteristics of the operating status. The autoencoding anomaly measurement analysis unit includes an autoencoding network structure of an encoder and a decoder, wherein the encoder extracts implicit features of the normal operating state of the power equipment based on a multi-scale residual analysis unit, and the decoder uses a transposed convolution structure to reconstruct the implicit features into original input data; Anomaly explainability analysis is performed based on the abnormal data of the power equipment, including locating the abnormal data of the power equipment, generating an abnormal heat map of the power equipment, and generating a fault type of the power equipment.

2. The method for detecting anomaly in power equipment data based on a deep residual network according to claim 1, characterized in that: The preprocessing includes data cleaning, normalization, missing value interpolation and time window segmentation.

3. The method for detecting anomaly in power equipment data based on a deep residual network according to claim 1, wherein: A joint optimization objective function of reconstruction loss and contrast loss is constructed to analyze the comprehensive characteristics and implicit characteristics of the operating status and obtain the minor anomaly data of the power equipment; through multi-scale reconstruction error measurement, combined with the time domain reconstruction error and frequency domain reconstruction error of the power equipment data, the abnormality degree of the minor anomaly data of the power equipment is obtained.

4. The method for detecting anomaly in power equipment data based on a deep residual network according to claim 3 is characterized in that: The lightweight reasoning optimization unit includes a network pruning layer, a network parameter quantization layer, a knowledge distillation analysis layer, and an reasoning acceleration layer; The network pruning layer calculates the sensitivity score of each branch to the degree of abnormality of the minor abnormal data of the power equipment output by the self-encoding abnormality measurement analysis unit, sorts the network connections according to the sensitivity score, and removes the connections with sensitivity lower than the threshold according to the preset threshold; the sensitivity score is obtained by observing the comprehensive operating status characteristics output by the model after perturbing and / or shielding the connection; the network parameter quantization layer determines the optimal quantization scaling factor by performing statistical analysis on the comprehensive operating status characteristics to obtain the comprehensive abnormal status data; the knowledge distillation analysis layer identifies the comprehensive abnormal status data by combining the comprehensive distillation objective function of hard label loss, soft label loss and feature representation loss to generate abnormal characteristics of the power equipment; the inference acceleration layer optimizes the abnormal characteristics of the power equipment through batch processing optimization, computational graph optimization and memory access mode optimization to generate abnormal data of the power equipment.

5. The method for detecting anomaly in power equipment data based on a deep residual network according to claim 1, wherein: The anomaly explainability analysis includes an anomaly location layer, an anomaly heat map generation layer, and a fault type mapping layer; The anomaly location layer analyzes the abnormal data of the power equipment based on the gradient weighted class activation mapping technology, performs global average pooling on the gradient of the final feature map and the gradient of each channel to obtain the channel importance weight, combines the weight with the corresponding feature activation map weight to generate a class activation heat map, and upsamples the class activation heat map to the same time dimension as the input data to locate the time point and data dimension that contribute most to the abnormal data of the power equipment; the anomaly heat map generation layer uses a two-dimensional visualization method for the abnormal data of the power equipment, and uses a gradient color scale to represent the degree of abnormality, where dark areas represent high abnormality and light areas represent low abnormality; the fault type mapping layer obtains the similarity between the abnormal pattern of the abnormal data of the power equipment and each fault pattern in the knowledge base based on the fault feature pattern description stored in the power equipment expert knowledge base, and identifies the fault type.

6. Power equipment data anomaly detection system based on deep residual network, characterized by: The method for detecting anomaly in power equipment data based on a deep residual network according to claim 1 comprises: Data acquisition module, used to obtain multi-dimensional time series data of power equipment and perform preprocessing; A deep residual network analysis module is used to construct a deep residual network model to analyze the multi-dimensional time series data and obtain abnormal data of power equipment; the deep residual network model includes a multi-scale residual analysis unit, an adaptive dimensional attention mechanism analysis unit, an autoencoding abnormality measurement analysis unit and a lightweight inference optimization unit; The abnormality explainability analysis module is used to perform abnormality explainability analysis based on the abnormal data of the power equipment, including locating the abnormal data of the power equipment, generating an abnormal heat map of the power equipment and generating a fault type of the power equipment.

Images