A collection reminder and life cycle abnormal data analysis method

Abstract

The present application discloses a method for collecting reminders and analyzing abnormal data in the life cycle, which relates to the field of power equipment, including: collecting data of power equipment; obtaining physical features of power equipment through mapping fusion methods based on extracted structured features and unstructured features, as well as physical mechanism models of power equipment; building an anomaly detection model; using the anomaly detection model to perform anomaly detection on the collected power equipment data; based on the anomaly detection results, using a Bayesian network causal reasoning graph model to build a causal relationship between abnormal monitoring quantities, and obtaining an abnormal causal relationship graph; based on the abnormal causal relationship graph, using a Markov chain Monte Carlo probability reasoning algorithm to trace the abnormal propagation path layer by layer, obtain the abnormal cause with the highest probability, and output it as the fault cause of the power equipment. In view of the low accuracy of abnormal cause analysis of power equipment in the prior art, the present application improves the accuracy through Bayesian network causal reasoning and Markov Monte Carlo probability reasoning.

Description

Technical Field

[0001] The present application relates to the technical field of full life cycle management and control of power equipment, and in particular to a method for collecting reminders and analyzing abnormal life cycle data. Background Art

[0002] In recent years, with the rapid development of sensor technology, communication technology, big data technology, and the increasing maturity of artificial intelligence theory, power equipment anomaly detection and diagnosis have begun to develop in the direction of data-driven and intelligent. By deploying various sensors on the equipment, a large amount of equipment operating condition data can be collected. Through machine learning algorithms, data fusion modeling and intelligent analysis can be carried out to mine the degradation laws and abnormal patterns contained in the data, significantly improving the accuracy and real-time performance of anomaly detection.

[0003] However, abnormality detection and diagnosis of power equipment still face many challenges. On the one hand, massive heterogeneous data puts tremendous pressure on data processing and calculation, requiring efficient data management and analysis technology; on the other hand, abnormal situations are complicated, with similar symptoms, abnormal mixing and other problems. It is difficult to accurately locate the root cause by relying solely on data-driven methods. Therefore, the existing data-driven anomaly detection methods are not accurate in abnormal cause analysis and diagnosis, and it is often difficult to give explainable and credible diagnostic conclusions, and cannot provide strong support for subsequent maintenance decisions.

[0004] In the related art, for example, Chinese patent document CN117745054A provides a full life cycle management and control platform and method for equipment abnormal events, including: obtaining monitoring signals, activating events according to monitoring signals; conducting abnormal analysis and tracking of activation events. Establishing a communication briefing, recording the full life cycle data through the communication briefing, transmitting the communication briefing as a sustainable data link, and changing the signature of the communication briefing as the abnormal event changes. After conducting abnormal analysis and tracking of the activation event, an analysis report is generated, and the analysis report records the entire process of the abnormal event; the analysis report is stored in a time series as a case for selection when the event is activated again. This scheme conducts abnormal analysis and tracking of activation events, and generates an analysis report to record the entire process of the abnormal event. However, there are complex causal dependencies between abnormal monitoring quantities, and the occurrence of abnormal root causes may lead to a series of chain reactions and abnormal propagation. Therefore, the accuracy of abnormal cause analysis of this scheme needs to be further improved. Summary of the invention

[0005] In response to the problem of low accuracy in the analysis of causes of abnormalities in power equipment in the prior art, the present application provides a method for collecting reminders and analyzing lifecycle abnormality data. For the detected anomalies, the accuracy of the analysis of causes of abnormalities is improved through Bayesian network causal reasoning and Markov Monte Carlo probability reasoning.

[0006] Technical solution, the purpose of this application is achieved through the following technical solution.

[0007] The present specification provides a collection reminder and life cycle abnormal data analysis method, including: collecting data of power equipment, performing preprocessing such as cleaning and standardization on the collected data to obtain preprocessed structured data and unstructured data; wherein the structured data includes telemetry data and power data; the unstructured data includes infrared image data and vibration audio data; feature extraction is performed based on the preprocessed structured data and unstructured data to obtain structured features and unstructured features; based on the extracted structured features and unstructured features, and the physical mechanism model of the power equipment, the physical features of the power equipment are obtained through a mapping fusion method; based on the extracted structured features and unstructured features, and the obtained physical features, a random forest algorithm is used to perform model training to obtain an anomaly detection model; anomaly detection is performed on the collected power equipment data using the anomaly detection model; based on the anomaly detection results, a Bayesian network causal reasoning graph model is used to construct a causal relationship between abnormal monitoring quantities to obtain an abnormal causal relationship graph; based on the abnormal causal relationship graph, a Markov chain Monte Carlo probability reasoning algorithm is used to trace the abnormal propagation path layer by layer to obtain the abnormal cause with the highest probability, which is output as the fault cause of the power equipment.

[0008] Among them, telemetry data refers to the operating status data of power equipment collected by remote measurement. It includes real-time measurement values ​​or time series records of monitoring quantities such as voltage, current, power, frequency, and temperature. Telemetry data reflects the real-time operating conditions of power equipment. Electricity data refers to the electricity consumption or power generation data of power equipment within a certain period of time. It includes the cumulative value or time-sharing record of measurement indicators such as active power, reactive power, and electricity charges. Electricity data reflects the energy consumption or production of power equipment. Infrared image data refers to image data obtained by thermal imaging of power equipment using an infrared thermal imager. Infrared images reflect the temperature distribution on the surface of the equipment and can be used to detect abnormal phenomena such as overheating and partial discharge of the equipment. Infrared image data is stored in the form of image files, containing information such as pixel matrix and temperature scale. Vibration audio data refers to data obtained by collecting vibration and sound signals of power equipment using vibration sensors and audio sensors. Vibration signals reflect the mechanical movement state of the equipment and can be used to detect abnormal conditions such as imbalance, looseness, and friction of the equipment. Audio signals reflect the acoustic characteristics of the equipment during operation and can be used to detect abnormal phenomena such as discharge and leakage of the equipment.

[0009] The physical mechanism model of power equipment refers to a mathematical model or simulation model established based on the physical structure, working principle and operation mechanism of power equipment. The physical mechanism model describes the physical processes such as energy conversion, signal transmission, and state change inside the power equipment, and reveals the inherent laws and relationships of the equipment. Common physical mechanism models of power equipment include circuit models, magnetic field models, thermal network models, mechanical dynamics models, etc. The physical characteristics of power equipment refer to quantitative indicators that reflect the physical properties and operating status of power equipment. Physical characteristics can be extracted from multi-source heterogeneous data such as telemetry data, power data, infrared image data, vibration audio data, etc., or they can be obtained through the mapping and fusion of physical mechanism models. Common physical characteristics of power equipment include: electrical characteristics: voltage, current, power, power factor, harmonics, etc. Thermal characteristics: temperature, temperature rise, heat flux, etc. Mechanical characteristics: vibration amplitude, vibration frequency, speed, pressure, etc. Acoustic characteristics: sound pressure level, spectrum characteristics, octave characteristics, etc. Optical characteristics: discharge intensity, flicker index, etc.

[0010] Furthermore, the structured features are obtained, including: inputting the pre-processed telemetry data and electricity data into the seasonal and trend decomposition using Loess (STL) algorithm, decomposing the telemetry data and electricity data into seasonal terms, trend terms and random noise terms through the STL algorithm, and extracting features reflecting the multi-scale periodicity of the data; inputting the telemetry data and electricity data decomposed by the STL algorithm into the multi-scale discrete wavelet transform model (Multi-scale Discrete Wavelet Transform). Transform, MSDWT); the Daubechies wavelet basis function is used to perform multi-level discrete wavelet decomposition on the telemetry data and electricity data decomposed by the STL algorithm, and the low-frequency information and high-frequency information of the telemetry data and electricity data are extracted simultaneously in the time domain and frequency domain to obtain the wavelet coefficient characteristics that characterize the transient impact; among which, the low-frequency trend information represents the slowly changing trend components of the telemetry data and electricity data, the high-frequency mutation information represents the fast-changing impact components of the telemetry data and electricity data, and the transient impact represents the mutation and short-term spike information of the telemetry data and electricity data; the extracted multi-scale periodic features and wavelet coefficient features are spliced ​​as structured features.

[0011] Among them, in the telemetry data and power data of power equipment, the seasonal term may correspond to the periodic patterns of equipment operation on Sundays, months, quarters, etc. The seasonal term reflects the regular changes of data on a specific time scale. In the telemetry data and power data of power equipment, the trend term may correspond to the gradual evolution of the equipment operation status, such as equipment aging, performance degradation, etc. The trend term reflects the direction and rate of change of data on a longer time scale. Random noise terms usually come from uncertain factors such as measurement errors and environmental interference, which interfere with data analysis and anomaly detection. Through seasonal and trend decomposition algorithms, random noise terms can be separated from the original data to improve the signal-to-noise ratio of the data and the accuracy of feature extraction. Multi-scale periodic features refer to the periodic change characteristics of time series data on different time scales. The operating status of power equipment may be affected by multiple periodic factors, such as daily load cycles, weekly load cycles, seasonal load cycles, etc.

[0012] The multi-scale discrete wavelet transform model MSDWT is a time-frequency analysis method used to extract multi-scale time-frequency features of time series data. MSDWT performs multi-level wavelet decomposition on the data and maps the data to wavelet coefficients of different scales to achieve joint feature extraction in the time domain and frequency domain. In each level of decomposition, MSDWT decomposes the data into low-frequency approximate coefficients and high-frequency detail coefficients, which represent the low-frequency trend information and high-frequency mutation information of the data respectively. The Daubechies wavelet basis function is a commonly used orthogonal wavelet basis function with excellent characteristics such as compact support, symmetry and vanishing moment. Low-frequency information refers to the components in time series data that change slowly and have obvious trends. In the multi-scale discrete wavelet transform, low-frequency information corresponds to the low-frequency approximate coefficients obtained by wavelet decomposition. The low-frequency approximate coefficients characterize the overall trend and average characteristics of the data on a longer time scale, reflecting the long-term change law of the data. High-frequency information refers to the components in time series data that change rapidly and fluctuate significantly. In the multi-scale discrete wavelet transform, high-frequency information corresponds to the high-frequency detail coefficients obtained by wavelet decomposition. The high-frequency detail coefficients characterize the local fluctuations and mutation characteristics of the data on a finer time scale, and reflect the short-term change pattern of the data. Transient shocks refer to the sudden, short-term and large changes in time series data. In the telemetry data and power data of power equipment, transient shocks may correspond to abnormal conditions such as sudden failures and instantaneous overloads during equipment operation. Wavelet coefficient characteristics refer to the time-frequency domain characteristics obtained by multi-scale discrete wavelet transform. Wavelet coefficients reflect the distribution characteristics of time series data on different time scales and frequency scales, including low-frequency approximate coefficients and high-frequency detail coefficients. The low-frequency approximate coefficients characterize the trend information of the data, and the high-frequency detail coefficients characterize the mutation information of the data.

[0013] Furthermore, unstructured features are obtained, including: taking the preprocessed infrared image data as input, using the pretrained ResNet50 convolutional neural network model, freezing the convolutional layer and part of the fully connected layer parameters of ResNet50 through the transfer learning method, and only performing parameter adjustment training on the last fully connected layer to extract the texture feature vector of the infrared image; taking the preprocessed vibration audio data as input, using the pretrained LSTM-Attention recurrent neural network model, freezing the LSTM layer and Attention layer parameters of the model through the transfer learning method, and only performing parameter adjustment training on the last fully connected layer to extract the time series feature vector of the vibration audio; and concatenating the extracted texture feature vector and the time series feature vector as the unstructured feature.

[0014] Among them, the LSTM-Attention recurrent neural network is a deep learning model for processing time series data, which combines the long short-term memory network (LSTM) and the attention mechanism (Attention). LSTM is a special recurrent neural network that can effectively capture long-distance dependencies in time series data by introducing gating mechanisms and memory units, overcoming the gradient vanishing problem of traditional recurrent neural networks. The Attention mechanism can adaptively focus on key information in time series data by introducing attention weights, improving the accuracy and interpretability of feature extraction.

[0015] Furthermore, the physical characteristics of the power equipment are obtained, including: inputting the extracted structured features and unstructured features into a feature mapping neural network established based on the physical mechanism of the power equipment, and obtaining the nonlinear mapping representation of the structured features and unstructured features in the physical feature space through a forward propagation algorithm; wherein the feature mapping neural network adopts a multi-layer perceptron structure to convert the physical mechanism of the power equipment into the network's inter-layer connection weight parameters and activation function parameters; the nonlinear mapping representation of the structured features and unstructured features output by the feature mapping neural network in the physical feature space is input as a physical feature subset into a feature fusion autoencoder network constructed in the physical feature space, and the shared representation features of the physical feature subset are extracted; the extracted shared representation features are input into the constructed physical feature pooling network, and the weights of different physical features are calculated through an attention mechanism; a fused physical feature vector is obtained based on the shared representation features and the corresponding weights; the fused physical feature vector is processed by a softmax normalization function to obtain a normalized physical feature vector as the physical feature output of the power equipment in the current state.

[0016] Among them, the feature mapping neural network model maps the original structured features (such as telemetry data features, power data features) and unstructured features (such as infrared image features, vibration audio features) to a unified physical feature space, and obtains the nonlinear mapping representation of these features in the physical feature space, namely the physical feature subset. The physical feature subset represents the physical meaning of the original feature data under the constraints of the physical mechanism of the power equipment, and reflects the physical quantity information of the equipment operation status, such as temperature, pressure, vibration frequency, etc. However, these physical feature subsets may have information redundancy and low correlation problems, and their shared representation features need to be further extracted. Therefore, the physical feature subset is used as the input of the feature fusion autoencoder network model. Through unsupervised pre-training, under the action of the encoder and decoder, the shared information between the physical feature subsets is extracted to obtain a more abstract and high-level physical feature representation for subsequent feature fusion and anomaly analysis tasks.

[0017] Among them, the feature mapping neural network is an independent multi-layer perceptron structure, which is used to map the original structured features and unstructured features to the physical feature space to obtain a physical feature subset. The network parameters of this model are determined by the physical mechanism of the power equipment. The feature fusion autoencoder network is an independent symmetric encoder-decoder structure, which is used to extract the shared representation features of the physical feature subset. This model optimizes the network parameters through unsupervised pre-training to achieve feature fusion. The physical feature pooling network is an independent attention mechanism module, which is used to calculate the importance weights of different physical features and perform weighted fusion of the shared representation features to obtain the final physical feature vector. These three neural network structures are connected in sequence to form a hierarchical physical feature extraction and fusion process. Among them, the output of the feature mapping neural network model is used as the input of the feature fusion autoencoder network model, and the output of the feature fusion autoencoder network model is used as the input of the physical feature pooling layer, and finally the fused physical feature vector output is obtained.

[0018] Among them, the multi-layer perceptron (MLP) is a feedforward neural network structure, which consists of an input layer, a hidden layer, and an output layer. In the acquisition of physical features of power equipment, the feature mapping neural network adopts the MLP structure, takes structured features and unstructured features as input, and maps them to the physical feature space through the forward propagation algorithm. The inter-layer connection weight parameters and activation function parameters of the MLP are set according to the physical mechanism of the power equipment, so that the network can learn the nonlinear mapping relationship between the feature space and the physical space. Through the feature mapping of the MLP structure, heterogeneous structured features and unstructured features can be uniformly mapped to the physical feature space to achieve physical interpretation and fusion of features. The number of hidden layers and the number of neurons of the MLP can be adjusted according to the complexity of the power equipment and the data scale to balance the accuracy of feature mapping and computational efficiency.

[0019] Shared representation features refer to the common representations extracted from features of different sources or modalities during the feature fusion process. In the acquisition of physical features of power equipment, structured features and unstructured features are nonlinearly mapped by feature mapping neural networks to obtain their representations in the physical feature space, which are called physical feature subsets. In order to further fuse these physical feature subsets, a feature fusion autoencoder network is introduced. The autoencoder network maps physical feature subsets to a low-dimensional shared representation space through unsupervised learning and extracts their shared representation features. The shared representation features characterize the intrinsic associations and statistical laws between different physical feature subsets. The extraction of shared representation features is achieved through the encoder part of the autoencoder network. The encoder compresses the physical feature subsets into a low-dimensional feature space through multiple layers of nonlinear transformations to obtain compact and expressive shared representation features. The decoder part reconstructs the shared representation features back to the original physical feature subsets, ensuring the information fidelity of the shared representation features. In the physical feature pooling network, the shared representation features are used to calculate the weights of different physical features, and dynamically focus on key physical features through the attention mechanism to obtain the fused physical feature vector.

[0020] Furthermore, the nonlinear mapping representation of structured features and unstructured features in the physical feature space is obtained, including: constructing a multi-physical field coupling simulation model of power equipment, coupling and solving the electromagnetic field, temperature field and stress-strain field of the power equipment through the finite element analysis method, and obtaining the physical quantity mapping relationship that reflects the internal physical mechanism of the power equipment; converting the function form in the physical quantity mapping relationship into the inter-layer connection weight parameter of the neural network; converting the boundary conditions and constraints in the physical quantity mapping relationship into the activation function parameters of the neural network; using the obtained inter-layer connection weight parameters and activation function parameters to construct a feature mapping neural network; using the historical operating conditions of the power equipment The structured features and unstructured feature samples are selected, and the error between the output of the feature mapping neural network and the corresponding physical feature true value is taken as the optimization target. The inter-layer connection weights and activation function of the feature mapping neural network are supervised and adjusted through the gradient descent algorithm; the inter-layer connection weights and activation function parameters of the supervised adjustment feature mapping neural network are fixed as the nonlinear mapping function that maps the structured features and unstructured features to the physical feature space; the extracted structured features and unstructured features are input into the feature mapping neural network with fixed parameters, and the nonlinear mapping representation of the structured features and unstructured features in the physical feature space is obtained through the forward propagation algorithm.

[0021] Among them, the multi-physics coupling simulation model of power equipment refers to a mathematical physics model established by comprehensively considering the interaction and coupling effects between multiple physical fields (such as electromagnetic fields, temperature fields, stress and strain fields, etc.) inside the power equipment. This model describes the physical phenomena and processes inside the power equipment through mathematical expressions such as partial differential equations, boundary conditions and constraints, and reflects the energy conversion and mutual influence between different physical fields. The multi-physics coupling simulation model is usually based on information such as the geometric structure, material properties and operating conditions of the power equipment, and uses the finite element analysis method to discretize and numerically solve the model. By coupling and solving the control equations of different physical fields, the spatiotemporal evolution laws of physical quantities such as electromagnetic field distribution, temperature distribution, stress and strain distribution inside the power equipment can be obtained, revealing the internal physical mechanism of the power equipment.

[0022] Among them, the physical quantity mapping relationship of the physical mechanism inside the power equipment refers to the functional relationship between different physical quantities in the multi-physics field coupling simulation model. These physical quantities can be electric field strength, magnetic induction strength, temperature, displacement, stress, strain, etc. They are interrelated in space and time, and constitute the complex physical process inside the power equipment. The physical quantity mapping relationship is expressed in the form of a mathematical function, describing the quantitative relationship between a physical quantity and other physical quantities. For example, the electric field strength and charge density satisfy Gauss's law, the temperature and heat flux density satisfy Fourier's law, and the stress and strain satisfy Hooke's law. These physical quantity mapping relationships reflect the physical mechanism inside the power equipment and embody basic physical laws such as conservation of energy and conservation of momentum. In the construction of the feature mapping neural network, the physical quantity mapping relationship is converted into the inter-layer connection weight parameters and activation function parameters of the neural network. The inter-layer connection weight parameters characterize the linear relationship between different physical quantities, and the activation function parameters characterize the nonlinear relationship and constraints between physical quantities. By embedding the physical quantity mapping relationship into the neural network structure, the feature mapping neural network has physical interpretability and can learn the physical mechanism inside the power equipment.

[0023] Among them, when converting the function form in the physical quantity mapping relationship into the interlayer connection weight parameter of the neural network, the function form can be selected according to the physical mechanism of the power equipment and the characteristics of the multi-physics field coupling simulation model. Specifically, partial differential equations: The electromagnetic field, temperature field and stress-strain field of the power equipment can be described by partial differential equations, such as Maxwell's equations, heat conduction equations, elastic mechanics equations, etc. These partial differential equations characterize the spatial distribution and time evolution relationship between physical quantities. After discretizing the partial differential equations, the numerical solution of the physical quantity at the finite element node can be obtained. The discretized partial differential equation coefficient matrix is ​​converted into the interlayer connection weight of the neural network, and the source term and boundary condition of the equation are converted into the bias term and activation function parameter of the neural network, so as to construct a neural network model driven by physical information. Constitutive relations: The material properties and physical characteristics of power equipment can be described by constitutive relations, such as the stress-strain relationship, conductivity-temperature relationship, magnetic permeability-magnetic field intensity relationship, etc. of the material. The constitutive relationship is usually a nonlinear function that reflects the physical properties of the material. Converting the constitutive relationship function into the activation function of the neural network can introduce material nonlinear effects and improve the accuracy of physical field coupling solutions. For example, the stress-strain curve can be fitted as a piecewise linear function or a polynomial function, and the conductivity-temperature curve can be fitted as an exponential function or a power function, and then these functions can be used as the activation function of the neural network. Physical field coupling equation: The multi-physical field coupling process of power equipment can be described by coupling equations, such as the electric-thermal coupling equation, the electric-mechanical coupling equation, the thermal-mechanical coupling equation, etc. The coupling equation reflects the interaction and influence mechanism between different physical fields. Converting the coupling term coefficients in the coupling equation into the connection weights between different layers of the neural network can establish the interaction relationship between the physical fields. For example, the Joule heat term in the electric-thermal coupling equation can be converted into the connection weight between the electric field intensity and the temperature field, and the thermal stress term in the thermal-mechanical coupling equation can be converted into the connection weight between the temperature field and the stress-strain field.

[0024] Furthermore, the shared representation features of the physical feature subsets are extracted, including: inputting the physical feature subset output by the feature mapping neural network into a feature fusion autoencoder network constructed in the physical feature space; wherein the feature fusion autoencoder network adopts a symmetric encoder-decoder structure; inputting the physical feature subset into the encoder of the feature fusion autoencoder network, the encoder adopts a three-layer fully connected neural network, mapping the physical feature subset to a low-dimensional latent space composed of the last layer of hidden nodes of the encoder, and obtaining a hidden layer representation of the physical feature subset as a shared representation of different physical feature subsets in the latent space; inputting the hidden layer representation of the physical feature subset output by the encoder into the decoder of the feature fusion autoencoder network, the decoder adopts A three-layer transverse fully connected neural network reconstructs and maps the hidden layer representation from the latent space constructed by the encoder to the feature space where the physical feature subset is located, and obtains the reconstructed representation of the physical feature subset; the mean square error between the reconstructed representation of the physical feature subset and the corresponding original physical feature subset is used as the reconstruction error loss function, and the encoder and decoder parameters of the feature fusion autoencoder network are optimized by the stochastic gradient descent algorithm to minimize the reconstruction error loss function; the encoder of the trained feature fusion autoencoder network is used as the feature encoder, and the physical feature subset is input into the feature encoder, and the hidden layer representation of the physical feature subset in the latent space constructed by the encoder is obtained by the forward propagation algorithm as the shared representation feature of different physical feature subsets.

[0025] Among them, the low-dimensional latent space refers to a low-dimensional feature space to which the physical feature subset is mapped after multiple layers of nonlinear transformation in the encoder of the feature fusion autoencoder network. This latent space usually has a lower dimension than the original physical feature subset and can represent the intrinsic structure and characteristics of the physical feature subset in a more compact and abstract way. In the network structure of the encoder, the low-dimensional latent space corresponds to the feature space composed of the last layer of hidden nodes of the encoder.

[0026] Hidden layer representation refers to the intermediate feature representation obtained on the hidden layer of the encoder of the feature fusion autoencoder network after the physical feature subset undergoes multiple layers of nonlinear transformation. Each hidden layer can be regarded as an abstraction and transformation of the physical feature subset, and the hidden layer representation is the specific representation of the physical feature subset in these abstract spaces. During the forward propagation process of the encoder, the physical feature subset enters the network through the input layer, and after multiple fully connected layers and nonlinear activation function transformations, a series of hidden layer representations are generated.

[0027] The reverse fully connected neural network refers to a fully connected layer structure in the decoder of the feature fusion autoencoder network, which is opposite to the encoder, and gradually maps the hidden layer representation from the low-dimensional latent space back to the high-dimensional feature space where the original physical feature subset is located. The goal of the decoder is to reconstruct the hidden layer representation into a reconstructed representation that is as similar as possible to the input physical feature subset to verify the fidelity of the hidden layer representation to the original feature. In the network structure of the decoder, the hidden layer representation is first dimensionally expanded through a fully connected layer to map the low-dimensional hidden layer representation to a higher-dimensional space. Then, through multiple reverse fully connected layers and nonlinear activation functions, the expanded hidden layer representation is gradually transformed and restored to the dimension and scale of the original physical feature subset. The number of layers and neurons of the decoder usually corresponds to that of the encoder, forming a symmetrical encoder-decoder structure. The reverse fully connected neural network utilizes the idea of ​​layer-by-layer restoration and reconstruction in deep learning, and generates a reconstructed representation similar to the original feature by gradually amplifying and decoding the hidden layer representation.

[0028] Furthermore, an abnormal causal relationship graph is obtained, including: extracting abnormal monitoring quantities of power equipment based on the abnormal detection results of the anomaly detection model; the abnormal monitoring quantities include structured feature anomalies, unstructured feature anomalies and physical feature anomalies; using the physical mechanism of the power equipment to perform causal analysis on the abnormal monitoring quantities, and obtaining the initial causal relationship between the abnormal monitoring quantities; using the abnormal monitoring quantities as nodes of the Bayesian network causal reasoning graph model, and using the obtained initial causal relationship as the directed edge of the Bayesian network causal reasoning graph model; using the abnormal monitoring quantity data samples under the historical operating conditions of the power equipment, based on the parameter learning algorithm of the Bayesian network, and through the maximum likelihood estimation method, obtaining the conditional probability and prior probability of each node in the Bayesian network. Using the abnormal monitoring data samples under the historical operating conditions of power equipment, the structural learning algorithm based on the Bayesian network is used to score the directed edges through the minimum description length criterion MDL, and the Bayesian network structure with the highest score is selected as the causal relationship graph between the abnormal monitoring quantities; according to the Bayesian network node conditional probability and node prior probability obtained by parameter learning, and the causal relationship graph obtained by structure learning, an abnormal causal relationship graph model is constructed for causal reasoning analysis under abnormal conditions of power equipment.

[0029] Among them, abnormal monitoring quantities can come from different types of features, including structured features, unstructured features and physical features. Structured feature anomalies refer to abnormalities in structured data such as operating parameters and environmental parameters of power equipment, such as voltage, current, temperature, etc. that are beyond the normal range. Unstructured feature anomalies refer to abnormalities in unstructured data such as vibration, discharge, and image of power equipment, such as abnormal increase in vibration amplitude and abnormal increase in discharge frequency. Physical feature anomalies refer to abnormalities in physical characteristics such as internal physical quantities and physical field distribution of power equipment, such as abnormal electric field strength and stress strain. The extraction of abnormal monitoring quantities needs to be based on the historical operating data of power equipment and learned and identified through anomaly detection models. The anomaly detection model can use statistical methods, machine learning methods, or deep learning methods to identify abnormal data that deviates from the normal state by establishing the data distribution or feature representation of the normal state.

[0030] The parameter learning algorithm based on Bayesian network refers to an algorithm that estimates and learns the conditional probability distribution of each node in the network using data samples when the structure of the Bayesian network is known. Bayesian network is a probabilistic graph model consisting of a directed acyclic graph and node probability distribution, which is used to describe the conditional dependency and joint probability distribution between variables. In parameter learning, it is assumed that the structure of the Bayesian network is known, that is, the causal relationship between nodes has been determined. Nodes can represent abnormal monitoring quantities, and directed edges represent the causal relationship between abnormal monitoring quantities. The goal of parameter learning is to estimate the conditional probability distribution of each node under different combinations of its parent node values, as well as the prior probability distribution of each node. Commonly used parameter learning algorithms include maximum likelihood estimation (MLE) and Bayesian estimation (BE). Maximum likelihood estimation solves the optimal estimate of network parameters by maximizing the likelihood function of data samples. Bayesian estimation introduces the prior distribution of parameters and obtains the posterior estimate of parameters by maximizing the posterior probability. The conditional probability distribution and prior probability distribution obtained by parameter learning reflect the causal strength between abnormal monitoring quantities and the probability of abnormal occurrence.

[0031] The Bayesian network-based structural learning algorithm refers to an algorithm that learns the optimal Bayesian network structure from the data by searching and evaluating different network structures when the data samples are known. The network structure reflects the causal relationship between the abnormal monitoring quantities and determines the decomposition method of the joint probability distribution. The goal of structural learning is to find a network structure that best matches the data sample so that the network can accurately represent the causal dependency between the abnormal monitoring quantities. Common structural learning algorithms include scoring search-based methods and constraint-based methods. The scoring search-based method defines a scoring function, such as BIC, AIC, MDL, etc., which is used to evaluate the degree of fit between the network structure and the data. By searching different network structures, such as greedy search, heuristic search, etc., the network structure with the highest score is found. The constraint-based method introduces prior knowledge or causal constraints, such as IC algorithm, PC algorithm, etc., to infer the network structure that meets the constraints from the data.

[0032] The minimum description length criterion (MDL) is a commonly used scoring criterion. The MDL criterion is based on the principles of information theory and regards the learning problem as a data compression problem. The MDL scoring takes into account the complexity of the network structure and the degree of fit of the data, and selects the network structure that can describe the data with the least amount of information as the optimal structure. The MDL criterion can well balance the network's fitting ability and generalization ability, avoiding overfitting and underfitting problems.

[0033] Furthermore, the highest-scoring Bayesian network structure is selected as the causal relationship diagram between abnormal monitoring quantities, including: dividing the abnormal monitoring quantity data samples under the historical operating conditions of the power equipment into a training set and a verification set; based on the training set, the directed edges are combined and enumerated using the Bayesian network structure learning algorithm to obtain a candidate Bayesian network structure; for each candidate Bayesian network structure, the corresponding structural complexity and fitting error are calculated based on the verification set by the minimum description length criterion MDL; wherein the structural complexity represents the shortest coding length of the Bayesian network structure; the error fitting represents the prediction error of the Bayesian network structure for the verification set; each candidate Bayesian network structure is scored by MDL according to the calculated structural complexity and fitting error, and the Bayesian network structure with the highest MDL score is selected as the causal relationship diagram between the abnormal monitoring quantities; wherein, when the corresponding structural complexity is calculated by the minimum description length criterion MDL, a penalty term coefficient is set to control the sparsity of the Bayesian network structure; when the corresponding fitting error is calculated by the minimum description length criterion MDL, a sample quantity penalty term is set to control the generalization ability of the Bayesian network structure.

[0034] Furthermore, according to the abnormal causal relationship graph, the Markov chain Monte Carlo probability inference algorithm is used to trace the abnormal propagation path layer by layer to obtain the abnormal cause with the highest probability, including: dividing the abnormal monitoring nodes in the constructed abnormal causal relationship graph model into known nodes and unknown nodes; wherein the known nodes represent the monitoring quantities in which the abnormalities are detected in the abnormal monitoring results; the unknown nodes represent the monitoring quantities in the causal relationship graph between the abnormal monitoring quantities that are connected to the known nodes but not detected; based on the abnormal causal relationship graph model, the abnormal state of the unknown nodes is probabilistically predicted by the Markov chain Monte Carlo probability inference algorithm to obtain the unknown nodes with predicted abnormalities; the unknown nodes with predicted abnormalities are added to the known node set to form a new known node set; according to ... abnormal state of the unknown nodes is probabilistically predicted by the Markov chain Monte Carlo probability inference algorithm to obtain the unknown nodes with predicted abnormalities; the abnormal state of the unknown nodes is probabilistically predicted by the Markov chain Monte Carlo probability inference algorithm to obtain the unknown nodes The model determines the new unknown nodes connected to the new known node set, recursively executes the Markov chain Monte Carlo probability inference process, and traces the abnormal propagation path upward layer by layer until the abnormal states of all nodes in the abnormal causal relationship graph model are known; according to the traced abnormal propagation path, the joint probability corresponding to the abnormal state combination of the initial known nodes in the abnormal propagation path is statistically analyzed, and the node with the largest joint probability is selected as the abnormal root cause of the power equipment failure; wherein, the abnormal state of the initial known node is the abnormal state of the corresponding monitoring quantity in the abnormal detection result, and the joint probability between the nodes is calculated according to the obtained node conditional probability and the node prior probability; in the abnormal propagation path, the physical monitoring quantity corresponding to the abnormal root source node with the largest probability is output as the fault cause of the power equipment.

[0035] Furthermore, based on the abnormal causal relationship graph model, the abnormal state of the unknown node is probabilistically predicted through the Markov chain Monte Carlo probabilistic inference algorithm to obtain the unknown node with predicted abnormality, including: taking the abnormal state of the known node in the abnormal causal relationship graph model as the initial state of the Markov chain; taking the node conditional probability and the node prior probability obtained by parameter learning as the Markov chain state transition probability; performing random walk sampling on the Markov chain through the Gibbs sampling method to obtain random sampling samples of the abnormal state of the unknown node; performing statistical estimation of the posterior probability of the abnormal state of the unknown node based on the random sampling samples, and taking the abnormal state with the largest posterior probability as the predicted state of the unknown node.

[0036] Compared with the prior art, the advantages of this application are:

[0037] By collecting structured data (including telemetry data and electricity data) and unstructured data (including infrared image data and vibration audio data) of power equipment and performing multi-dimensional abnormal feature extraction, the operating status of the equipment can be fully characterized, overcoming the problem of insufficient feature representation capabilities caused by a single data source, and providing rich feature support for subsequent anomaly detection.

[0038] The seasonal trend decomposition STL algorithm and multi-scale discrete wavelet transform MSDWT are used to extract the multi-scale periodic characteristics and wavelet coefficient characteristics of structured data. The convolutional neural network is used to extract the texture characteristics of infrared images. The recurrent neural network is used to extract the time series characteristics of vibration audio. The intrinsic characteristics of data can be mined in different dimensions such as time domain, frequency domain, and spatial domain, and the ability to characterize abnormal characteristics can be enhanced.

[0039] The physical mechanism of the equipment is modeled using multi-physical field coupling simulation of power equipment, and the physical quantity mapping relationship is converted into the connection weights and activation functions of the feature mapping neural network. The structured features and unstructured features are mapped to a unified physical feature space for fusion through a feature fusion autoencoder network, so that the extracted abnormal features have clear physical meanings and the interpretability of abnormal diagnosis is improved.

[0040] Based on the extracted multi-source anomaly features, the random forest algorithm is used to build an anomaly detection model. The prediction results of different feature subsets are integrated through ensemble learning, which can effectively improve the accuracy and robustness of anomaly detection.

[0041] For the anomaly detection results, by constructing a Bayesian network causal reasoning graph model, the minimum description length criterion MDL is used to learn the causal structure between anomaly monitoring quantities from the data, and the causal strength is obtained through parameter learning. This can reveal the inherent mechanism of anomaly propagation and provide a causal logical basis for anomaly tracing.

[0042] Based on the abnormal causal relationship graph, the Markov Chain Monte Carlo probability inference algorithm is used to trace the abnormal propagation path upward layer by layer by setting known nodes and unknown nodes, and infer the root cause of the abnormality based on the joint probability of the abnormal propagation path. This can overcome the uncertainty of the diagnostic results caused by abnormal confounding and find the most likely cause of the abnormality. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is an exemplary flow chart of a collection reminder and life cycle abnormal data analysis method according to some embodiments of this specification;

[0044] Figure 2 is an exemplary flow chart of obtaining structural features according to some embodiments of this specification;

[0045] Figure 3 is an exemplary flow chart of obtaining unstructured features according to some embodiments of this specification;

[0046] Figure 4 is an exemplary flow chart of obtaining physical characteristics of an electric power device according to some embodiments of this specification;

[0047] Figure 5is an exemplary flow chart of obtaining a nonlinear mapping representation according to some embodiments of the present specification;

[0048] Figure 6 It is an exemplary flowchart for obtaining an abnormal causal relationship graph model according to some embodiments of this specification. DETAILED DESCRIPTION

[0049] The method and system provided in the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0050] Figure 1 The present invention is an exemplary flow chart of a method for collecting, reminding and analyzing lifecycle abnormal data according to some embodiments of the present specification, and the method comprises: collecting data of power equipment and performing preprocessing to obtain preprocessed structured data and unstructured data; wherein the structured data comprises telemetry data and power data; the unstructured data comprises infrared image data and vibration audio data; performing feature extraction based on the preprocessed structured data and unstructured data to obtain structured features and unstructured features; obtaining the physical features of the power equipment through a mapping fusion method based on the extracted structured features and unstructured features and the physical mechanism model of the power equipment; performing model training using a random forest algorithm based on the extracted structured features and unstructured features and the obtained physical features to obtain an anomaly detection model; performing anomaly detection on the collected power equipment data using the anomaly detection model; based on the anomaly detection results, constructing the causal relationship between abnormal monitoring quantities using a Bayesian network causal reasoning graph model to obtain an abnormal causal relationship graph; based on the abnormal causal relationship graph, using a Markov chain Monte Carlo probability reasoning algorithm to trace the abnormal propagation path layer by layer to obtain the abnormal cause with the highest probability, which is output as the fault cause of the power equipment.

[0051] Data collection and preprocessing. Deploy various types of sensors and data collection devices on power equipment to collect various monitoring data in real time during the operation of the equipment. For structured data, install temperature, pressure, flow, current and other sensors at key parts of the equipment to collect telemetry data at a certain sampling frequency (such as once per second); use electric energy meters to record the equipment's active power, reactive power, voltage, current and other electrical data to form a time series. For unstructured data, use infrared thermal imagers to conduct regular inspections of the equipment to obtain infrared image data on the surface of the equipment; use vibration sensors to collect vibration acceleration signals at bearings, gearboxes and other parts of the equipment to obtain vibration audio data. Upload the collected structured and unstructured data to the data management platform and classify and store them according to equipment type, collection time, etc.

[0052] Due to the complex on-site environment, the collected data will inevitably contain noise, outliers, missing values, etc., and the data needs to be cleaned before analysis. For structured data, outliers that exceed the range are eliminated according to the reasonable threshold of the physical quantity; missing values ​​are filled using linear interpolation, spline interpolation and other methods; moving average filtering, wavelet threshold filtering and other methods are used to denoise and smooth the data. For unstructured data, infrared images are preprocessed by deblurring and contrast enhancement; short-time Fourier transform is performed on vibration audio data to filter out background noise of specific frequencies, and endpoint detection, amplitude normalization and other processing are performed.

[0053] The numerical ranges of different physical quantities vary greatly. In order to eliminate the influence of dimension, it is necessary to normalize the structured data. Common normalization methods include maximum and minimum value normalization, zero mean normalization, etc. Maximum and minimum value normalization maps the data to the interval [0, 1], and the formula is: Where x is the original data, x max and x min are the maximum and minimum values ​​of the samples respectively. Zero mean normalization is to normalize the data to a standard normal distribution with a mean of 0 and a variance of 1. The formula is: Where μ is the sample mean and σ is the sample standard deviation. The normalized telemetry data and power data constitute the preprocessed structured data.

[0054] The collected infrared image data and vibration audio data are time-stamped and aligned with the structured data at the corresponding time to form a heterogeneous data set with consistent time and equipment. Each sample contains the structured and unstructured features of the equipment at the same time, which facilitates the subsequent multi-source data fusion modeling and analysis.

[0055] Figure 2 This is an exemplary flow chart for obtaining structured features according to some embodiments of this specification, obtaining structured features, seasonality and trend decomposition: inputting the preprocessed telemetry data and power data into the STL seasonality and trend decomposition algorithm. The STL algorithm is a time series decomposition method based on Loess (local weighted regression scatter point smoothing), which can decompose the time series data Yt into three parts: seasonal term St, trend term Tt and random noise term Et, that is: Yt = St + Tt + Et, where the seasonal term reflects the pattern of periodic fluctuations in the time series, the trend term reflects the long-term change trend of the time series, and the random noise term is the random fluctuation after removing the seasonal term and the trend term.

[0056] MSDWT uses Daubechies wavelet basis functions, which can provide good localization capabilities in both time domain and frequency domain. Commonly used Daubechies wavelet basis functions include db1 (Haar wavelet), db2, db3, etc. According to the sampling frequency, data length and other characteristics of telemetry data and power data, select appropriate Daubechies wavelet basis functions, such as db4 wavelet basis functions. The telemetry data and power data obtained by STL decomposition are decomposed by multi-level discrete wavelets using the Mallat algorithm.

[0057] Perform MSDWT on the trend item data obtained by STL decomposition: Input the trend item data reflecting the long-term change trend of telemetry data and power data into the MSDWT model, perform multi-level discrete wavelet decomposition on the trend item data using the Daubechies wavelet basis function, extract the wavelet coefficients representing the low-frequency change characteristics of the trend item, and reflect the long-term trend and slow-changing information of equipment operation. Perform MSDWT on the seasonal item data obtained by STL decomposition: Input the seasonal item data reflecting the periodic fluctuation pattern of telemetry data and power data into the MSDWT model, perform multi-level discrete wavelet decomposition on the seasonal item data using the Daubechies wavelet basis function, extract the wavelet coefficients representing the periodic change characteristics of the seasonal item, and reflect the periodic fluctuation and rhythm information of equipment operation.

[0058] Perform MSDWT on the random noise item data obtained by STL decomposition: Input the noise item data reflecting the random fluctuations of telemetry data and power data into the MSDWT model, use the Daubechies wavelet basis function to perform multi-level discrete wavelet decomposition on the random noise item data, extract the wavelet coefficients that characterize the transient impact characteristics of the random noise item, and reflect the sudden abnormality and transient fault information of the equipment operation. Perform feature selection and dimensionality reduction on the extracted trend item wavelet coefficients, seasonal item wavelet coefficients, and random noise item wavelet coefficients to obtain the compressed wavelet coefficient feature vector. The trend item wavelet coefficient features, seasonal item wavelet coefficient features, and random noise item wavelet coefficient features are spliced ​​to obtain the complete wavelet coefficient features as the structured features that characterize transient impacts.

[0059] The multi-scale periodic features extracted by the STL algorithm and the wavelet coefficient features extracted by the MSDWT are aligned according to the device ID and timestamp, and the features are spliced ​​to form a structured feature vector. Assuming that the dimension of the multi-scale periodic features extracted by the STL algorithm is m, and the dimension of the wavelet coefficient features extracted by the MSDWT is n, the dimension of the structured features after splicing is m+n. The structured features contain information on the periodicity, trend, transient nature, and other aspects of telemetry data and power data, and comprehensively describe the dynamic changes in the operating status of power equipment over time, providing a richer and more effective feature representation for subsequent anomaly detection and diagnosis.

[0060] Figure 3 This is an exemplary flowchart for obtaining unstructured features according to some embodiments of the present specification, and obtaining unstructured features includes: inputting the preprocessed infrared image data into the pre-trained ResNet50 convolutional neural network model. ResNet50 is a classic deep convolutional neural network structure, which includes 50 convolutional layers, pooling layers and fully connected layers. It has been pre-trained on large image datasets such as ImageNet and has strong feature extraction capabilities. The transfer learning method is adopted to freeze the parameters of the convolutional layers and some fully connected layers of ResNet50, and only the parameters of the last fully connected layer are adjusted and trained. Transfer learning utilizes the general image features learned by the pre-trained model on a large-scale dataset, and can quickly adapt to image classification tasks in specific fields through a small amount of parameter adjustment.

[0061] The output of the last fully connected layer of ResNet50 is used as the texture feature vector of the infrared image. This feature vector extracts high-level semantic information such as the edge, texture, and shape of the infrared image, and can characterize the appearance state and thermal distribution pattern of the equipment components. The preprocessed vibration audio data is input into the pre-trained LSTM-Attention recurrent neural network model. LSTM (Long Short-Term Memory) is a recurrent neural network structure that can model long-term dependencies. The Attention mechanism can adaptively focus on key fragments in time series data. The combination of the two can efficiently extract the dynamic features of time series data. Using the transfer learning method, the parameters of the LSTM layer and the Attention layer of the LSTM-Attention model are frozen, and only the parameters of the last fully connected layer are adjusted and trained. The LSTM layer and the Attention layer extract the universal time series features of the vibration audio data. Through transfer learning and parameter adjustment, it can quickly adapt to the vibration audio analysis tasks of specific equipment.

[0062] The output of the last fully connected layer of the LSTM-Attention model is used as the time series feature vector of the vibration audio. This feature vector extracts the dynamic information of the vibration audio signal, such as frequency, amplitude, and phase, and can characterize the vibration and noise patterns during the operation of the device. The infrared image texture feature vector extracted by ResNet50 and the vibration audio time series feature vector extracted by LSTM-Attention are aligned according to the device ID and timestamp, and the features are spliced ​​to obtain an unstructured feature vector. Assuming that the dimension of the infrared image texture feature vector is m and the dimension of the vibration audio time series feature vector is n, the dimension of the spliced ​​unstructured feature is m+n. The unstructured feature integrates the feature representation of the device appearance image and the operating audio, and comprehensively characterizes the static morphology and dynamic behavior information of the device components.

[0063] Figure 4 It is an exemplary flow chart for obtaining the physical characteristics of power equipment shown in some embodiments of this specification, physical feature mapping and fusion. Construct a physical feature mapping neural network: According to the physical mechanism of the power equipment, determine the input and output of the physical feature mapping neural network. Taking the transformer as an example, the input features may include structured features such as winding temperature trend items, winding temperature season items, winding temperature wavelet coefficients, oil temperature trend items, oil temperature season items, oil temperature wavelet coefficients, current trend items, current season items, current wavelet coefficients, and unstructured features such as infrared image texture features and vibration audio timing features. The output features are transformer physical quantities, such as winding temperature, oil temperature, current, load rate, etc. According to the dimensions of the input and output features, the number of layers and neurons of the physical feature mapping neural network are designed. For example, a 3-layer MLP structure can be used, the number of neurons in the input layer is the structured feature dimension + the unstructured feature dimension, the number of neurons in the hidden layer is 1 / 2 of the input layer, and the number of neurons in the output layer is the physical quantity dimension. According to the physical mechanism equation of the power equipment, the weight matrix and bias vector of the physical feature mapping neural network are designed. Taking the thermal balance equation of the transformer as an example, the physical relationship between the winding temperature, oil temperature, and current can be converted into the connection weight between the hidden layer and the output layer. For example, the winding temperature is proportional to the square of the current and proportional to the difference in oil temperature, so the corresponding weight parameter should be a positive value. Similarly, the electromagnetic field equation can be converted into the connection weight from the input layer to the hidden layer. Select a suitable activation function, such as ReLU, Sigmoid, Tanh, etc., to perform a nonlinear transformation on the weighted input of the neuron. The selection of the activation function needs to consider the value range and change law of the physical quantity. For example, the ReLU function is suitable for non-negative physical quantities, and the Sigmoid function is suitable for physical quantities with upper and lower bounds.

[0064] Physical feature mapping: Align the extracted structured features and unstructured features according to the device ID and timestamp to form a feature vector. Taking the transformer as an example, structured features such as winding temperature trend items, oil temperature seasonal items, current wavelet coefficients, and unstructured features such as infrared image texture features and vibration audio time series features can be spliced ​​into a high-dimensional feature vector. The feature vector is input into the physical feature mapping neural network, and the activation value of the neuron is calculated by the forward propagation algorithm. The forward propagation starts from the input layer and calculates the weighted input and activation value of the neuron layer by layer until the output layer. The activation value of the output layer is the nonlinear mapping representation of the structured features and unstructured features in the physical feature space. Taking the transformer as an example, each neuron in the output layer represents the predicted value of the physical quantities such as winding temperature, oil temperature, and current in the current state. The result of physical feature mapping reflects the nonlinear correspondence between the original features and the physical quantities of the power equipment. For example, there is a complex coupling relationship between the transformer winding temperature feature and the physical quantities such as winding current, oil temperature, and load rate. Through physical feature mapping, a mapping function between structured features, unstructured features, and physical quantities can be adaptively established.

[0065] Figure 5 This is an exemplary flow chart for obtaining a nonlinear mapping representation according to some embodiments of the present specification, and the acquisition of a physical feature subset: the nonlinear mapping representation of the structured features and unstructured features output by the physical feature mapping neural network in the physical feature space is used as a physical feature subset. The physical feature subset reflects the correspondence between the original features and the physical quantities of the power equipment, and is the basis for subsequent feature fusion. Taking the transformer as an example, the physical feature subset may include the predicted values ​​of multiple physical quantities such as winding temperature, oil temperature, and current in the current state. These predicted values ​​are obtained by mapping structured features such as winding temperature trend terms, oil temperature seasonal terms, current wavelet coefficients, and unstructured features such as infrared image texture features and vibration audio timing features to the physical feature space through a physical feature mapping neural network.

[0066] Structural design of feature fusion autoencoder network: A feature fusion autoencoder network is constructed in the physical feature space, and a symmetric encoder-decoder structure is adopted. Both the encoder and the decoder adopt a three-layer fully connected neural network, including an input layer, a hidden layer, and an output layer. The number of nodes in the input layer of the encoder is equal to the dimension of the physical feature subset, such as the winding temperature, oil temperature, current, and other physical quantities of the transformer. The number of nodes in the hidden layer is less than the number of nodes in the input layer, which plays a role in compression and feature extraction. The number of nodes in the output layer is equal to the number of nodes in the hidden layer, which serves as the hidden layer representation of the physical feature subset. The number of nodes in the input layer of the decoder is equal to the number of nodes in the output layer of the encoder, which receives the hidden layer representation of the physical feature subset. The number of nodes in the hidden layer is greater than the number of nodes in the input layer, which plays a role in restoring and reconstructing features. The number of nodes in the output layer is equal to the dimension of the physical feature subset, generating a reconstructed representation of the physical feature subset. The activation functions of the encoder and decoder can select nonlinear functions such as ReLU and Sigmoid to introduce nonlinear transformation capabilities. A Dropout regularization layer can be added between the hidden layer and the output layer to prevent overfitting.

[0067] Encoder forward propagation and hidden layer representation extraction: The physical feature subset is input to the input layer of the encoder, and the input signal is passed to the hidden layer and output layer layer by layer through the forward propagation algorithm. The hidden layer converts the input signal into a higher-level feature representation through matrix multiplication and nonlinear activation function. Let the weight matrix from the input layer to the hidden layer be We, the bias vector of the hidden layer be be, and the activation function be f, then the output h of the hidden layer is: h = f (We × x + be), where x is the physical feature subset and × represents matrix multiplication. The output layer of the encoder is transformed through similar matrix multiplication and activation function, and the output h of the hidden layer is mapped to a low-dimensional latent space to obtain the hidden layer representation z of the physical feature subset: z = f (Wz × h + bz), where Wz and bz are the weight matrix and bias vector from the hidden layer to the output layer, respectively. The hidden layer representation z is a shared representation of different physical feature subsets in the latent space, which describes the inherent commonality of the physical feature subsets, such as the correlation pattern between transformer winding temperature, oil temperature, and current. The dimension of the hidden layer representation is much smaller than the original physical feature space, reflecting the dimensionality reduction and feature extraction capabilities of the autoencoder network.

[0068] Decoder forward propagation and reconstructed representation generation: The hidden layer representation z output by the encoder is input to the input layer of the decoder, and is passed layer by layer to the hidden layer and output layer through the forward propagation algorithm. The hidden layer of the decoder converts the hidden layer representation z into a higher-dimensional feature representation through matrix multiplication and nonlinear activation function. Assume that the weight matrix from the input layer to the hidden layer is Wd, the bias vector of the hidden layer is bd, and the activation function is f, then the output of the hidden layer of the decoder is for: The output layer of the decoder undergoes a similar transformation, converting the output of the hidden layer Reconstruct the map to the original physical feature space to obtain the reconstructed representation of the physical feature subset Among them, Wx and bx are the weight matrix and bias vector from the hidden layer to the output layer respectively. Reconstructed representation The original physical feature subset x should be as close as possible to ensure the information integrity of the hidden layer representation z. The goal of the decoder is to minimize the reconstruction error, that is, The difference between .

[0069] Reconstruction error loss function and parameter optimization: Reconstruction representation The mean square error (MSE) between the original physical feature subset x and the original physical feature subset x is used as the reconstruction error loss function L: Where N is the number of samples in the physical feature subset, and x i are the reconstructed representation and original representation of the i-th sample respectively. The stochastic gradient descent (SGD) algorithm is used to optimize the parameters of the encoder and decoder to minimize the reconstruction error loss function L. The gradient of the loss function L to the encoder weights We, Wz and biases be, bz and the decoder weights Wd, Wx and biases bd, bx is calculated by the back propagation algorithm. The parameters are updated at a certain learning rate η. For example, the update formula of the encoder weight We is: The update method of other parameters is similar. Iterate multiple epochs, perform forward propagation and backward propagation on all training samples in each epoch, and continuously update the parameters until the reconstruction error loss function L converges or reaches the preset number of epochs. Use the trained encoder as the feature encoder, fix its parameters and no longer update it. Input the physical feature subset x into the feature encoder, and obtain the hidden layer representation z of the physical feature subset in the latent space through the forward propagation algorithm. Using the hidden layer representation z as the shared representation feature of different physical feature subsets reveals the association pattern between physical feature subsets and reflects the inherent law of the device status. The shared representation feature z can be used for subsequent tasks such as physical feature pooling, fault diagnosis, and health assessment to provide a comprehensive feature description of the device status.

[0070] Construction and fusion of physical feature pooling network: A physical feature pooling network is constructed, with the shared representation features extracted by the feature encoder as input, and the weight coefficients of different physical features are adaptively calculated through the attention mechanism. An attention mechanism layer is designed in the physical feature pooling network, and the shared representation features are mapped to the physical feature weights through the fully connected layer and the Softmax activation function. The Softmax function normalizes the weights to the range of [0, 1], and the sum of the weights is 1, which represents the relative importance of different physical features. According to the shared representation features and the corresponding physical feature weights, the fused physical feature vector is obtained by weighted averaging or weighted concatenation. The fused physical feature vector integrates the physical mapping information of structured features and unstructured features, and takes into account the importance differences of different physical quantities. The fused physical feature vector is normalized, and the maximum and minimum normalization or zero mean unit variance normalization is used to scale the feature values ​​to a uniform scale to eliminate the dimension effect. The normalized physical feature vector is used as the standardized physical feature output of the power equipment in the current state.

[0071] Anomaly detection model construction. The extracted structured features, unstructured features and fused physical features are used as the input of the anomaly detection model to form a comprehensive feature vector x. Structured features include numerical or categorical variables such as operating parameters and environmental factors of power equipment, such as voltage, current, power, temperature, humidity, etc. Unstructured features include statistical features extracted from time series data such as vibration signals and discharge signals, such as mean, variance, peak, kurtosis, etc. Physical features are standardized physical indicators obtained by pooling through feature fusion autoencoder network and attention mechanism. The three types of features are spliced ​​into a comprehensive feature vector x in a predefined order, and each sample corresponds to a feature vector. The comprehensive feature vector integrates the information of multi-source heterogeneous data, comprehensively describes the operating status of power equipment, and provides rich feature representation for anomaly detection.

[0072] The random forest model is used to train the comprehensive feature vector x, and the power equipment anomaly detection model is constructed by ensemble learning. Random forest is an ensemble learning algorithm based on decision trees. By constructing multiple decision trees and combining their prediction results, the generalization ability and robustness of the model are improved. Set the hyperparameters of the random forest, such as the number of decision trees, the maximum depth of each decision tree, the minimum number of samples for each node, and the sampling ratio of feature subsets. These hyperparameters control the complexity and randomness of the random forest model and need to be tuned according to the data scale and problem characteristics. Randomly sample the training data to generate multiple bootstrap sample sets (BootstrapSamples). Each bootstrap sample set is obtained by extracting N samples from the original training set with replacement, where N is the number of samples in the original training set. Bootstrap sampling introduces random perturbations in the data and enhances the diversity of the model. For each bootstrap sample set, a decision tree is constructed. At each node of the decision tree, a feature subset is randomly selected from all features, and the optimal features and optimal split points are selected through a certain partitioning criterion (such as the Gini index, information gain, etc.), and child nodes are recursively generated until the predefined stopping conditions are met (such as reaching the maximum depth, the number of node samples is less than the threshold, etc.). Repeat the above steps to build multiple decision trees to form a random forest. Each decision tree predicts the comprehensive feature vector x and gives a judgment result on whether the sample is abnormal. The prediction results of all decision trees in the random forest are combined, and the final anomaly detection result is obtained through a voting mechanism. When the voting ratio of the abnormal sample exceeds the preset threshold (such as 50%), it is judged as abnormal; otherwise, it is judged as normal. Ensemble learning improves the accuracy and robustness of anomaly detection by combining the predictions of multiple weak learners.

[0073] The constructed anomaly detection model is used to perform anomaly detection on the collected power equipment data. For each sample to be tested, its structured features, unstructured features and physical features are extracted and spliced ​​into a comprehensive feature vector x* in the same order. The comprehensive feature vector x* is input into the trained random forest model, and the prediction results of each tree are obtained through the path judgment of the decision tree. The prediction results of all decision trees in the random forest are voted, and the final anomaly detection results are obtained based on the comparison between the voting ratio and the preset threshold. If the voting ratio exceeds the threshold, the sample is judged to be in an abnormal state; otherwise, it is judged to be in a normal state. The anomaly detection results can be used to trigger subsequent decisions such as alarms and dispatch maintenance, timely discover and handle abnormal conditions of power equipment, and ensure the safe and stable operation of the power grid.

[0074] Figure 6It is an exemplary flow chart for obtaining an abnormal causal relationship graph model according to some embodiments of this specification, and an abnormal causal relationship graph is constructed. Extraction of abnormal monitoring quantity and acquisition of initial causal relationship: Based on the detection results of the abnormal detection model in step 4, the abnormal monitoring quantity of the power equipment is extracted. The abnormal monitoring quantity includes structured feature anomalies, unstructured feature anomalies and physical feature anomalies, reflecting the abnormal performance of the power equipment in different feature domains. The abnormal monitoring quantity is causally analyzed using the physical mechanism of the power equipment to obtain the initial causal relationship between the abnormal monitoring quantities. The physical mechanism reveals the mutual influence and transmission mechanism between the various components and parameters inside the power equipment, and provides prior knowledge for the causal relationship of the abnormal monitoring quantity. According to the physical structure, energy flow, signal transmission and other mechanisms of the power equipment, the causal dependency between the abnormal monitoring quantities is analyzed. For example, voltage anomalies may cause current anomalies, current anomalies may cause temperature anomalies, and vibration anomalies may be the result of bearing failures. The causal dependency between the abnormal monitoring quantities obtained by analysis is used as the initial causal relationship to guide the construction of the Bayesian network causal reasoning graph model. The initial causal relationship provides a priori dependency structure between the abnormal monitoring quantities, reducing the causal relationship search space.

[0075] Node and edge construction of the Bayesian network causal reasoning graph model: The abnormal monitoring quantity is used as the node of the Bayesian network causal reasoning graph model, and each node represents an abnormal monitoring quantity. The abnormal monitoring quantity can be a binary variable (normal / abnormal) or a multi-valued variable (different abnormal levels). The initial causal relationship obtained in step 1 is used as the directed edge of the Bayesian network causal reasoning graph model. The directed edge indicates the causal dependency between the abnormal monitoring quantities. The directed edge points from the cause node to the result node, indicating that the state of the cause abnormal monitoring quantity will affect the state of the result abnormal monitoring quantity. The constructed Bayesian network causal reasoning graph model is a directed acyclic graph (DAG), where the nodes represent the abnormal monitoring quantities and the directed edges represent the causal dependency between the abnormal monitoring quantities. This graph model describes the causal propagation mechanism of abnormal states of power equipment.

[0076] Parameter learning and node probability estimation of Bayesian network: Using the abnormal monitoring data samples under the historical operating conditions of power equipment, based on the parameter learning algorithm of Bayesian network, the conditional probability and prior probability of each node in the Bayesian network are parameter learned through the maximum likelihood estimation method. For each abnormal monitoring node, estimate its probability of occurrence under different abnormal state combinations. The abnormal state combination takes into account the state of the node itself and the state of its parent node (cause node). Maximum likelihood estimation estimates the probability parameters by maximizing the likelihood function of the sample data. For node X i , its conditional probability P(X i |P a (X i The maximum likelihood estimate of )) is: Among them, Count(x i ,pa(X i )) represents the node X in the sample i The value is x i And its parent node P a (X i ) takes the value of pa(X i ) samples, Count(pa(X i )) represents the parent node P in the sample a (X i ) takes the value of pa(X i ) samples. For nodes without parent nodes (root nodes), estimate their prior probability P(X i )for: Among them, Count(x i ) represents the node X in the sample i The value is x i The number of samples is N, and N is the total number of samples. Through parameter learning, the occurrence probability estimation of abnormal monitoring nodes under different abnormal state combinations is obtained, the dependence strength between abnormal monitoring quantities is quantified, and a probabilistic basis is provided for causal reasoning.

[0077] Structural learning of Bayesian network and optimization of causal relationship graph: Using the data samples of abnormal monitoring quantities under the historical operating conditions of power equipment, the structure learning algorithm based on the Bayesian network is used to score the directed edges through the minimum description length criterion (MDL), and the Bayesian network structure with the highest score is selected as the causal relationship graph between abnormal monitoring quantities. The data samples of abnormal monitoring quantities under the historical operating conditions of power equipment are divided into training set and validation set. The training set is used for searching and learning the Bayesian network structure, and the validation set is used to evaluate the quality of the learned structure. Based on the training set, the Bayesian network structure learning algorithm is used to combine and enumerate the directed edges to obtain the candidate Bayesian network structure. The candidate structure is obtained by adding, deleting or reversing the directed edges on the basis of the initial causal relationship, covering the possible causal relationship hypothesis space. For each candidate Bayesian network structure, the corresponding structural complexity and fitting error are calculated based on the validation set by the minimum description length criterion (MDL). The MDL criterion considers the balance between model complexity and data fitting, and selects the structure with low complexity and small fitting error as the optimized causal relationship graph. Structural complexity represents the shortest encoding length of the Bayesian network structure, which is related to the sparsity of the network structure. By setting the penalty coefficient, the sparsity of the Bayesian network structure is controlled to avoid overly complex structures. The fitting error represents the prediction error of the Bayesian network structure for the validation set, reflecting the degree of fit of the structure to the data. By setting the sample number penalty, the generalization ability of the Bayesian network structure is controlled to avoid overfitting. The MDL score is performed on each candidate Bayesian network structure based on the calculated structural complexity and fitting error, and the Bayesian network structure with the highest MDL score is selected as the optimized causal relationship diagram between abnormal monitoring quantities. The MDL score comprehensively considers the structural complexity and fitting error, and balances the simplicity and expressiveness of the causal relationship diagram.

[0078] Construction and application of abnormal causal relationship graph model: According to the conditional probability of Bayesian network nodes and node prior probability obtained by parameter learning, and the optimized causal relationship graph obtained by structure learning, the abnormal causal relationship graph model is constructed. The abnormal causal relationship graph model integrates the causal structure and probabilistic dependency between abnormal monitoring quantities, and can be used for causal reasoning analysis under abnormal working conditions of power equipment. Given the observed values ​​of certain abnormal monitoring quantities, the posterior probability of other abnormal monitoring quantities can be calculated through the reasoning algorithm of the Bayesian network (such as variable elimination, belief propagation, etc.), the abnormal cause and abnormal consequence can be determined. The abnormal causal relationship graph model provides a causal basis for abnormal diagnosis and prevention, helping operation and maintenance personnel to quickly locate the root cause of the abnormality, formulate targeted maintenance and preventive measures, and improve the reliability and safety of power equipment.

[0079] Abnormal cause tracing and diagnosis. Division of abnormal monitoring nodes: The abnormal monitoring nodes in the constructed abnormal causal relationship graph model are divided into two categories: known nodes and unknown nodes. Known nodes represent the monitoring quantities in which abnormalities are detected in the abnormal monitoring results, that is, the monitoring quantities determined to be abnormal by the anomaly detection model. The abnormal states of these nodes are determined and serve as the starting point for abnormal tracing. Unknown nodes represent the monitoring quantities that are connected to known nodes but not detected in the causal relationship graph between abnormal monitoring quantities, that is, the monitoring quantities with unknown abnormal states. The abnormal states of these nodes need to be predicted through probabilistic reasoning. Setting the initial state of the Markov chain: According to the anomaly detection results, the monitoring nodes in which abnormalities are detected in the abnormal causal relationship graph model are marked as known nodes, and the monitoring nodes in which abnormalities are not detected are marked as unknown nodes. The abnormal state of the known node is used as the initial state of the Markov chain. The abnormal state of the known node is a definite value, indicating the starting point of abnormal propagation. The initial state is a node state vector, denoted as x{(0)}=(x1{(0)},x2{(0)},......,x m {(0)}), where x i {(0)} represents the abnormal state of the i-th known node, and m is the number of known nodes.

[0080] Setting of Markov chain state transition probability: The node conditional probability and node prior probability obtained by parameter learning are used as the Markov chain state transition probability. The state transition probability describes the probabilistic dependency of the propagation of abnormal states in the causal graph. For any two node states x and x', the state transition probability P(x'|x) represents the probability of transitioning from state x to state x'. The state transition probability is calculated based on the local dependency property of the Bayesian network: Where n is the total number of nodes, x i ' represents the value of node i in state x', p a (x i ') represents the value of the parent node of node i in state x'. i '|p a (x i ')) is the conditional probability of node i.

[0081] Application of Gibbs sampling method: Use Gibbs sampling method to perform random walk sampling on Markov chain and generate random sampling samples of abnormal states of unknown nodes. Gibbs sampling updates the abnormal state of each unknown node in turn according to the conditional probability distribution of the node to obtain a sample sequence of abnormal state combinations. The steps of Gibbs sampling are as follows: Initialize the current state x{(0)} of the Markov chain to the abnormal state of the known node. For each iteration t=1, 2, ..., T: For each unknown node i=1, 2, ..., k: Fix the states of other nodes and update them according to the conditional probability distribution of node i Sample and update the abnormal state of node i to obtain a new state

[0082] Update the current state of the Markov chain to Get the sample sequence {x {0} ,x {1} ,......,x {T}}, where each sample x {t} Represents a possible abnormal state combination. Posterior probability estimation of unknown node abnormal state: Statistical estimation of the posterior probability of unknown node abnormal state based on random sampling. The posterior probability reflects the abnormal possibility of the unknown node given the abnormal state of the known node. For each unknown node i, the frequency of its taking different abnormal states in the sample sequence is counted to estimate its posterior probability distribution: where x obs represents the observed abnormal state of a known node, is the indicator function, which takes 1 when the condition is true and 0 otherwise. For each unknown node, the abnormal state with the largest posterior probability is selected as its predicted state:

[0083] Recursive tracing of the abnormal propagation path and updating the known node set: add the unknown nodes predicted to be abnormal to the known node set to form a new known node set. The unknown nodes predicted to be abnormal are predicted by the Markov Chain Monte Carlo method, and their abnormal states have been determined. The new known node set includes the initial known nodes and the unknown nodes predicted to be abnormal, indicating the node set with the currently determined abnormal state. The unknown nodes predicted to be abnormal become the new starting point of abnormal propagation, which is used to further trace the cause of the abnormality. The abnormal state of these nodes may be caused by the abnormal state of their parent nodes or ancestor nodes, and need to be traced up along the causal relationship graph. Determine the new unknown nodes: According to the abnormal causal relationship graph model, starting from the new known node set, determine the new unknown nodes connected to it. Traverse each node in the new known node set and check its adjacent nodes in the abnormal causal relationship graph. For each known node, mark the nodes whose abnormal states have not been predicted or determined in its adjacent nodes as new unknown nodes. These new unknown nodes are the next layer of nodes on the abnormal propagation path, and their abnormal states need to be further predicted. They are directly connected to the current known nodes and may be the cause or influencing factor of the abnormal state of the current known nodes.

[0084] Recursively perform Markov chain Monte Carlo probability inference: predict the abnormal state of the new unknown node and recursively perform the Markov chain Monte Carlo probability inference process. Repeat the sampling and prediction process, use the new unknown node as the initial state of the Markov chain, and use the Gibbs sampling method to generate random sampling samples of abnormal states. Estimate the posterior probability of the abnormal state of the new unknown node based on the sampled samples, and select the abnormal state with the largest posterior probability as the predicted state of the new unknown node. Obtain the abnormal state prediction result of the new unknown node, add the new unknown node with the predicted abnormality to the known node set, and form an updated known node set.

[0085] Tracing the abnormal propagation path layer by layer: Tracing the abnormal propagation path upward layer by layer, repeating the process of steps 1-3. Starting from the updated known node set, determine the new unknown nodes in the next layer, recursively perform Markov chain Monte Carlo probability inference, and predict the abnormal state of the new unknown node. Each recursion adds the node predicted to be abnormal to the known node set, and traces the abnormal propagation path upward along the causal graph. The recursive process continues until the abnormal state of all nodes in the abnormal causal graph model is known, that is, all nodes are classified as known nodes, and the abnormal propagation path is completely reconstructed. Output the abnormal propagation path: When the abnormal state of all nodes in the abnormal causal graph model has been determined, the recursive process terminates. Output the complete abnormal propagation path, including the initial known nodes, the unknown nodes predicted to be abnormal, and all abnormal nodes determined in the recursive tracing process. The abnormal propagation path reflects the propagation and evolution of the abnormal state in the causal graph, tracing from the abnormal manifestation node to the abnormal root node.

[0086] Determination of the root cause of the anomaly and output of the fault cause: Calculate the joint probability of the abnormal state combination of the initial known nodes: According to the abnormal propagation path obtained by tracing, extract the abnormal state combination of the initial known nodes. The initial known node is the starting point of the abnormal propagation path and reflects the initial manifestation of the anomaly. The abnormal state combination refers to the state value combination of the initial known node in the abnormal propagation path, expressed as X obs =(X1=x1,X2=x2,.....,X m =x m ), where X i represents the i-th initial known node, x i represents its abnormal state value, and m is the number of initially known nodes. According to the node conditional probability and node prior probability in the abnormal causal relationship graph model, the initial known node abnormal state combination X is calculated. obs The joint probability P(X obs ).

[0087] The formula for calculating the joint probability is: Among them, P a (X i ) represents node X i The parent node set of P(X i =x i |P a (X i )) is node X i Conditional probability under a given parent node state. Node conditional probability P(X i =x i |P a (X i )) describes the causal dependency between nodes, which has been estimated in the parameter learning process of step 5. The node prior probability P(X i =x i ) describes the abnormal probability of the node itself and can be set based on prior knowledge or statistical data.

[0088] Determine the abnormal root node: Compare the joint probabilities of the abnormal state combinations of different initial known nodes in the abnormal propagation path, and select the node corresponding to the combination with the largest joint probability as the abnormal root node of the power equipment fault. The abnormal root node is the most likely cause of abnormal propagation, and its abnormal state corresponds to the cause of the fault. The maximum joint probability means that the abnormal state combination of the node has the highest probability of appearing among all possible abnormal causes. The abnormal root node is denoted as X. root , and its corresponding abnormal state is x root ,satisfy: Output fault cause: In the abnormal propagation path, the abnormal root node X rootThe corresponding physical monitoring quantity is output as the fault cause of the power equipment. The physical monitoring quantity corresponds to the abnormal root node one by one, reflecting the physical manifestation and quantitative indicators of the fault. The physical monitoring quantity of the abnormal root node exceeds the normal range, resulting in the generation and propagation of the abnormal state. The output fault cause includes the name of the abnormal root node, the corresponding physical monitoring quantity and the abnormal state value x root , expressed as (X root ,x root ).

Claims

1. A collection reminder and life cycle abnormal data analysis method, comprising: Collect data from power equipment and perform preprocessing to obtain preprocessed structured data and unstructured data; wherein the structured data includes telemetry data and power data; the unstructured data includes infrared image data and vibration audio data; Feature extraction is performed based on the preprocessed structured data and unstructured data to obtain structured features and unstructured features; based on the extracted structured features and unstructured features, as well as the physical mechanism model of the power equipment, the physical features of the power equipment are obtained through a mapping fusion method; Based on the extracted structured features and unstructured features, as well as the acquired physical features, a random forest algorithm is used to train the model to obtain an anomaly detection model; the anomaly detection model is used to perform anomaly detection on the collected power equipment data; According to the anomaly detection results, the Bayesian network causal reasoning graph model is used to construct the causal relationship between the abnormal monitoring quantities and obtain the abnormal causal relationship graph; According to the abnormal causal relationship diagram, the Markov chain Monte Carlo probability inference algorithm is used to trace the abnormal propagation path layer by layer to obtain the abnormal cause with the highest probability, which is output as the fault cause of the power equipment; Obtain physical characteristics of electrical equipment, including: The extracted structured features and unstructured features are input into the feature mapping neural network established based on the physical mechanism of the power equipment, and the nonlinear mapping representation of the structured features and unstructured features in the physical feature space is obtained through the forward propagation algorithm; wherein, the feature mapping neural network adopts a multi-layer perceptron structure to convert the physical mechanism of the power equipment into the network's inter-layer connection weight parameters and activation function parameters; The nonlinear mapping representation of structured features and unstructured features output by the feature mapping neural network in the physical feature space is used as a physical feature subset and input into the physical feature space to construct a feature fusion autoencoder network to extract the shared representation features of the physical feature subset; The extracted shared representation features are input into the constructed physical feature pooling network, and the weights of different physical features are calculated through the attention mechanism; according to the shared representation features and the corresponding weights, the fused physical feature vector is obtained; The fused physical feature vector is normalized to obtain a normalized physical feature vector as the physical feature output of the power equipment in the current state.

2. The method for collecting reminders and analyzing abnormal life cycle data according to claim 1, characterized in that: Get structured features, including: The preprocessed telemetry data and electricity data are input into the seasonality and trend decomposition algorithm STL. The telemetry data and electricity data are decomposed into seasonal terms, trend terms and random noise terms through the STL algorithm, and the features reflecting the multi-scale periodicity of the data are extracted; The telemetry data and power data decomposed by the STL algorithm are input into the multi-scale discrete wavelet transform model MSDWT; the telemetry data and power data decomposed by the STL algorithm are decomposed by multi-level discrete wavelet using Daubechies wavelet basis function, and the low-frequency information and high-frequency information of the telemetry data and power data are extracted in the time domain and frequency domain at the same time to obtain the wavelet coefficient characteristics that characterize the transient impact; among which the low-frequency information represents the trend part; the high-frequency information represents the mutation part; The extracted multi-scale periodic features and wavelet coefficient features are concatenated as structured features.

3. The collection reminder and life cycle abnormal data analysis method according to claim 2 is characterized by: Obtain unstructured features, including: The preprocessed infrared image data is used as input, and the pre-trained ResNet50 convolutional neural network model is used. The parameters of the convolutional layer and some fully connected layers of ResNet50 are frozen through the transfer learning method. Only the parameter adjustment training of the last fully connected layer is performed to extract the texture feature vector of the infrared image. The preprocessed vibration audio data is used as input, and the pre-trained LSTM-Attention recurrent neural network model is adopted. The parameters of the LSTM layer and Attention layer of the model are frozen through the transfer learning method, and only the parameter adjustment training of the last fully connected layer is performed to extract the time series feature vector of the vibration audio. The extracted texture feature vector and temporal feature vector are concatenated as unstructured features.

4. The collection reminder and life cycle abnormal data analysis method according to claim 1 is characterized by: Obtain nonlinear mapping representations of structured and unstructured features in physical feature space, including: Construct a multi-physical field coupling simulation model for power equipment, and use the finite element analysis method to couple and solve the electromagnetic field, temperature field, and stress-strain field of the power equipment to obtain the physical quantity mapping relationship that reflects the internal physical mechanism of the power equipment; Convert the function form in the physical quantity mapping relationship into the inter-layer connection weight parameter of the neural network; convert the boundary conditions and constraint conditions in the physical quantity mapping relationship into the activation function parameters of the neural network; construct the feature mapping neural network using the obtained inter-layer connection weight parameters and activation function parameters; Using the structured and unstructured feature samples of the historical operating conditions of the power equipment, the error between the output of the feature mapping neural network and the corresponding physical feature true value is used as the optimization target, and the inter-layer connection weights and activation functions of the feature mapping neural network are supervised and adjusted through the gradient descent algorithm. The inter-layer connection weights and activation function parameters of the supervised adjusted feature mapping neural network are fixed as a nonlinear mapping function for mapping structured features and unstructured features to the physical feature space; The extracted structured features and unstructured features are input into a feature mapping neural network with fixed parameters, and the nonlinear mapping representation of structured features and unstructured features in the physical feature space is obtained through the forward propagation algorithm.

5. The collection reminder and life cycle abnormal data analysis method according to claim 1 is characterized by: Extract shared representation features of a subset of physical features, including: Input the physical feature subset output by the feature mapping neural network into the feature fusion autoencoder network constructed in the physical feature space; wherein the feature fusion autoencoder network adopts a symmetric encoder-decoder structure; The physical feature subset is input into the encoder of the feature fusion autoencoder network. The encoder uses a three-layer fully connected neural network to map the physical feature subset to a low-dimensional latent space composed of the last layer of hidden nodes of the encoder, and obtains the latent layer representation of the physical feature subset as the shared representation of different physical feature subsets in the latent space. The hidden layer representation of the physical feature subset output by the encoder is input into the decoder of the feature fusion autoencoder network. The decoder uses a three-layer transverse fully connected neural network to reconstruct the hidden layer representation from the latent space constructed by the encoder to the feature space where the physical feature subset is located, and obtain the reconstructed representation of the physical feature subset; The mean square error between the reconstructed representation of the physical feature subset and the corresponding original physical feature subset is used as the reconstruction error loss function. The encoder and decoder parameters of the feature fusion autoencoder network are optimized by the stochastic gradient descent algorithm to minimize the reconstruction error loss function. The encoder of the trained feature fusion autoencoder network is used as the feature encoder, and the physical feature subset is input into the feature encoder. The hidden layer representation of the physical feature subset in the latent space constructed by the encoder is obtained through the forward propagation algorithm as the shared representation feature of different physical feature subsets.

6. The method for collecting reminders and analyzing abnormal life cycle data according to any one of claims 1 to 5, characterized in that: Get the abnormal causal relationship diagram, including: Based on the anomaly detection results of the anomaly detection model, the abnormal monitoring quantity of the power equipment is extracted; the abnormal monitoring quantity includes structured feature anomalies, unstructured feature anomalies and physical feature anomalies; Use the physical mechanism of power equipment to conduct causal analysis on abnormal monitoring quantities and obtain the initial causal relationship between abnormal monitoring quantities; The abnormal monitoring amount is used as the node of the Bayesian network causal reasoning graph model, and the obtained initial causal relationship is used as the directed edge of the Bayesian network causal reasoning graph model; Using the abnormal monitoring data samples under the historical operating conditions of power equipment, the parameter learning algorithm based on the Bayesian network is used to obtain the conditional probability and prior probability of each node in the Bayesian network through the maximum likelihood estimation method; Using the abnormal monitoring data samples under the historical operating conditions of power equipment, the Bayesian network-based structural learning algorithm is used to score the directed edges through the minimum description length criterion MDL, and the Bayesian network structure with the highest score is selected as the causal relationship diagram between the abnormal monitoring quantities. According to the Bayesian network node conditional probability and node prior probability obtained by parameter learning, and the causal relationship graph obtained by structure learning, an abnormal causal relationship graph model is constructed for causal reasoning analysis under abnormal operating conditions of power equipment.

7. The collection reminder and life cycle abnormal data analysis method according to claim 6 is characterized by: The Bayesian network structure with the highest score is selected as the causal relationship diagram between abnormal monitoring quantities, including: The abnormal monitoring data samples under the historical operating conditions of the power equipment are divided into a training set and a validation set; Based on the training set, the directed edges are combined and enumerated using the Bayesian network structure learning algorithm to obtain a candidate Bayesian network structure; For each candidate Bayesian network structure, the corresponding structural complexity and fitting error are calculated based on the validation set using the minimum description length criterion MDL; where structural complexity refers to the shortest encoding length of the Bayesian network structure; error fitting refers to the prediction error of the Bayesian network structure on the validation set; Perform an MDL score on each candidate Bayesian network structure according to the calculated structural complexity and fitting error, and select the Bayesian network structure with the highest MDL score as the causal relationship diagram between abnormal monitoring quantities; Among them, when the corresponding structural complexity is calculated by the minimum description length criterion MDL, the penalty term coefficient is set to control the sparsity of the Bayesian network structure; when the corresponding fitting error is calculated by the minimum description length criterion MDL, the sample quantity penalty term is set to control the generalization ability of the Bayesian network structure.

8. The method for collecting reminders and analyzing abnormal life cycle data according to claim 7, characterized in that: According to the abnormal causal relationship diagram, the Markov chain Monte Carlo probability inference algorithm is used to trace the abnormal propagation path layer by layer to obtain the most probable abnormal cause, including: The abnormal monitoring nodes in the constructed abnormal causal relationship graph model are divided into known nodes and unknown nodes; the known nodes represent the monitoring quantities with abnormalities detected in the abnormal monitoring results; the unknown nodes represent the monitoring quantities connected to the known nodes but not detected in the causal relationship graph between the abnormal monitoring quantities; Based on the abnormal causal relationship graph model, the abnormal state of unknown nodes is probabilistically predicted through the Markov chain Monte Carlo probability inference algorithm to obtain the unknown nodes with predicted abnormalities; Add the unknown nodes predicted to be abnormal to the known node set to form a new known node set; According to the abnormal causal relationship graph model, the new unknown nodes connected to the new known node set are determined, and the Markov chain Monte Carlo probability inference process is recursively executed to trace the abnormal propagation path layer by layer until the abnormal states of all nodes in the abnormal causal relationship graph model are known; According to the abnormal propagation path obtained by tracing, the joint probability corresponding to the abnormal state combination of the initial known nodes in the abnormal propagation path is counted, and the node with the largest joint probability is selected as the abnormal root cause of the power equipment failure; wherein, the abnormal state of the initial known node is the abnormal state of the corresponding monitoring quantity in the abnormal detection result, and the joint probability between nodes is calculated according to the obtained node conditional probability and node prior probability; In the abnormal propagation path, the physical monitoring quantity corresponding to the abnormal root node with the highest probability is output as the fault cause of the power equipment.

9. The collection reminder and life cycle abnormal data analysis method according to claim 8 is characterized by: Get the unknown nodes with predicted abnormalities, including: The abnormal state of the known nodes in the abnormal causal relationship graph model is used as the initial state of the Markov chain; The node conditional probability and node prior probability obtained by parameter learning are used as the Markov chain state transition probability; The Markov chain is randomly sampled by Gibbs sampling method to obtain random sampling samples of the abnormal state of unknown nodes; The posterior probability of the abnormal state of the unknown node is statistically estimated based on the randomly sampled samples, and the abnormal state with the largest posterior probability is taken as the predicted state of the unknown node.

Images