A method, apparatus, and electronic device for equipment condition assessment and fault diagnosis.

Abstract

This application provides a method, apparatus, and electronic device for equipment condition assessment and fault diagnosis, relating to the field of equipment condition assessment technology. The method includes: acquiring multi-source heterogeneous data associated with a target device; performing online dynamic calibration of a pre-constructed parameterized failure physics model based on the multi-source heterogeneous data, and performing condition inference of the target device's health parameters based on the calibrated parameterized failure physics model; inputting extracted real-time symptom data into a pre-constructed fault knowledge graph for inference, and using a pre-constructed health benchmark cloud model to perform time-series pattern recognition of the target device to obtain a fault diagnosis and early warning strategy. This method can deeply integrate failure mechanisms with real-time operational monitoring, overcoming the shortcomings of fixed parameters in degradation models. Furthermore, through the linkage of data benchmarks and knowledge graphs, it can achieve early fault diagnosis with small samples and zero samples that possess physical interpretability.

Description

Technical Field

[0001] This application relates to the field of equipment condition assessment technology, and in particular to a method, apparatus and electronic equipment for equipment condition assessment and fault diagnosis. Background Technology

[0002] In related technologies, the reliability requirements for core electrical equipment such as transformers are increasingly stringent in power systems and specific industrial scenarios. Traditional fault diagnosis methods based on threshold alarms or statistical models have significant limitations: on the one hand, fixed threshold methods cannot capture the continuous, minute evolution of equipment degradation and are insensitive to early, subtle signs; on the other hand, while pure data-driven models (such as conventional deep learning) perform well in big data scenarios, they heavily rely on a large number of labeled fault samples. In actual engineering, faults in high-reliability equipment are often extremely rare, leading to significant "small sample" or even "zero sample" diagnostic problems. Furthermore, existing parametric physical models are usually set based on factory design values ​​or empirical values, and their model parameters remain fixed after the equipment is put into operation. However, during actual operation, the actual degradation parameters of the equipment drift over time due to various factors such as load, environment, and aging, causing the fixed physical model to gradually deviate from the actual state of the equipment, making it difficult to accurately assess the health status of the equipment and predict its remaining lifespan. At the same time, although knowledge graphs are used to assist in fault reasoning, they are often isolated from physical models and data-driven models, failing to be deeply integrated with real-time monitoring data to form a complementary relationship. Summary of the Invention

[0003] This application provides a method, apparatus, and electronic device for equipment condition assessment and fault diagnosis.

[0004] Firstly, this application provides a method for equipment condition assessment and fault diagnosis, including: Acquire multi-source heterogeneous data associated with a target device; wherein the target device includes electrical equipment; Based on the multi-source heterogeneous data, the pre-constructed parameterized failure physics model is dynamically calibrated online, and the health parameters of the target device are deduced based on the calibrated parameterized failure physics model. The extracted real-time symptom data is input into a pre-built fault knowledge graph for inference, and a pre-built health benchmark cloud model is used to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy.

[0005] In one possible implementation, the online dynamic calibration of the pre-constructed parameterized failure physics model based on the multi-source heterogeneous data includes: Using extended Kalman filtering or Bayesian online learning algorithms, and taking the multi-source heterogeneous data as observations, the key degradation parameters to be calibrated in the parameterized failure physics model are recursively estimated; the key degradation parameters include activation energy and pre-exponential factor. The obtained parameter estimates are then updated into the parameterized failure physics model.

[0006] In one possible implementation, the step of performing state deduction of the health parameters of the target device based on the calibrated parameterized failure physics model includes: Based on the calibrated parameterized failure physics model, the evolution path of equipment health status under various preset operating conditions is deduced; With minimizing the total lifecycle cost as the optimization objective, the long-term benefits of each preset maintenance strategy are quantitatively evaluated based on the evolution path of the health status of the target device, and the corresponding target maintenance strategy is determined. The total lifecycle cost includes maintenance cost and failure risk cost, and the failure risk cost is determined by multiplying the failure probability by the failure loss.

[0007] In one possible implementation, the step of inputting the extracted real-time symptom data into a pre-constructed fault knowledge graph for inference includes: The real-time symptom data is input as observational evidence into the pre-constructed fault knowledge graph, wherein the knowledge graph contains fault feature representations stored in the form of triples. The belief propagation algorithm is used to perform probability propagation calculations within the pre-constructed fault knowledge graph to obtain inference results, including potential fault types and confidence features, derived from the observed evidence.

[0008] In one possible implementation, prior to performing temporal pattern recognition on the target device using a pre-built health benchmark cloud model, the process includes: Acquire multi-dimensional time-series data of the target device in a healthy operating state; The multivariate time-series data is used as training sample data, and latent space representation is calculated using a variational autoencoder or a Gaussian process latent variable model to construct the health benchmark cloud model corresponding to the target device; wherein, the health benchmark cloud model includes the benchmark mean vector and benchmark covariance matrix in the observation space.

[0009] In one possible implementation, the step of using a pre-built health benchmark cloud model to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy includes: Calculate the joint deviation of the data points of the monitored target device relative to the pre-built health baseline cloud model; For the time trajectory obtained based on the joint deviation tracking, temporal pattern recognition is performed through a temporal convolutional network or a Transformer encoder to obtain the identification result of whether the target device exhibits early degradation characteristics of a specific fault type; When the early degradation characteristics are identified, a fault diagnosis and early warning strategy is generated for the target device based on the identification results.

[0010] In one possible implementation, calculating the joint deviation of the monitored target device's data points relative to the pre-built health baseline cloud model includes: Obtain the baseline mean vector and baseline covariance matrix of the pre-constructed health baseline cloud model in the observation space; By combining the benchmark mean vector and the benchmark covariance matrix, the spatial distance between the data points of the monitored target device in the current space is calculated using Mahalanobis distance, and this distance is used as the corresponding joint deviation; wherein, the joint deviation is the spatial distance calculation result.

[0011] Secondly, this application provides an equipment condition assessment and fault diagnosis device, comprising: An acquisition module is used to acquire multi-source heterogeneous data associated with a target device; wherein the target device includes electrical equipment; The deduction module is used to perform online dynamic calibration of the pre-constructed parameterized failure physics model based on the multi-source heterogeneous data, and to perform state deduction of the health parameters of the target device based on the calibrated parameterized failure physics model. The early warning module is used to input the extracted real-time symptom data into a pre-built fault knowledge graph for inference, and to use a pre-built health benchmark cloud model to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy.

[0012] Thirdly, this application provides an electronic device, comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above.

[0013] Fourthly, this application provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method described in the first aspect above.

[0014] The technical solution disclosed in this application brings at least the following beneficial effects: In the embodiments of this application, multi-source heterogeneous data associated with the target device is acquired; a pre-constructed parameterized failure physical model is dynamically calibrated online based on the multi-source heterogeneous data; and the health parameters of the target device are inferred based on the calibrated parameterized failure physical model. The extracted real-time symptom data is input into a pre-constructed fault knowledge graph for inference, and a pre-constructed health benchmark cloud model is used to perform temporal pattern recognition on the target device to obtain a fault diagnosis and early warning strategy. Thus, on the one hand, by dynamically calibrating the parameterized failure physical model online based on multi-source heterogeneous data, the model can evolve synchronously with the aging state of the physical device, overcoming the shortcomings of traditional physical models that are detached from reality; on the other hand, by using the health benchmark cloud model to perform temporal pattern recognition on the target device, it can keenly capture degradation trends in the early stages of minor drifts in multi-dimensional joint features, avoiding the risk of missed detection by the single-dimensional threshold method; furthermore, by inputting real-time symptom data into the fault knowledge graph for inference and fusing it with the recognition results of the health benchmark cloud model, high-reliability early warning can be achieved without relying on a large number of labeled fault samples, solving the technical problems of insufficient diagnostic capabilities and isolated knowledge graphs.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application, and do not constitute an undue limitation of this application.

[0017] Figure 1 A flowchart of a first specific embodiment of a device condition assessment and fault diagnosis method provided in this application; Figure 2 A flowchart of a second specific embodiment of a device condition assessment and fault diagnosis method provided in this application; Figure 3 A flowchart of a third specific embodiment of a device condition assessment and fault diagnosis method provided in this application; Figure 4 This application provides an overall system architecture diagram for a method of equipment condition assessment and fault diagnosis. Figure 5 This application provides a schematic diagram of a transformer fault knowledge graph. Figure 6 This is a schematic diagram of the structure of an equipment condition assessment and fault diagnosis device provided in this application. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0019] The following description, with reference to the accompanying drawings, describes a method, apparatus, and electronic device for equipment status assessment and fault diagnosis according to embodiments of this application.

[0020] Figure 1 This is a schematic flowchart illustrating a device condition assessment and fault diagnosis method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101, acquire multi-source heterogeneous data associated with the target device.

[0021] The target equipment includes electrical equipment, such as transformers.

[0022] In the embodiments of this application, the target device can be various core electrical equipment in industrial scenarios, such as transformers. Multi-source heterogeneous data refers to a data set from different sources with different structures, which can contain multiple information dimensions that can reflect the operating status, external conditions, and physical characteristics of the equipment.

[0023] S102, based on multi-source heterogeneous data, perform online dynamic calibration of the pre-constructed parameterized failure physics model, and perform state deduction of the health parameters of the target device according to the calibrated parameterized failure physics model.

[0024] In the embodiments of this application, the parameterized failure physical model is a pre-constructed mathematical model describing the degradation patterns of the equipment. This step utilizes real-time acquired multi-source heterogeneous data to dynamically correct unknown or time-varying parameters in the model, enabling the model to closely follow the actual aging state of the equipment. Subsequently, the calibrated model is used to perform forward-looking simulations and projections of the equipment's future health status, providing a basis for maintenance decisions.

[0025] S103, the extracted real-time symptom data is input into the pre-built fault knowledge graph for inference, and the pre-built health benchmark cloud model is used to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy.

[0026] In the embodiments of this application, on the one hand, features representing abnormalities (i.e., real-time symptom data) are extracted from monitoring data and input as evidence into a structured fault knowledge graph. The logical association capabilities of the graph are used to infer possible causes of the fault. On the other hand, a probabilistic model (i.e., a health baseline cloud model) constructed based on data under healthy conditions is used to perform time-series pattern analysis on the current monitoring data to identify whether it exhibits a degradation trajectory deviating from the normal pattern. Combining the results of the above two aspects, a final diagnostic conclusion and early warning strategy are generated as a fault diagnosis and early warning strategy. The above steps together solve the technical problems mentioned in the background art, such as the fixed physical model parameters and the difficulty of diagnosis under small sample conditions.

[0027] In the embodiments of this application, multi-source heterogeneous data associated with the target device is acquired; a pre-constructed parameterized failure physical model is dynamically calibrated online based on the multi-source heterogeneous data; and the health parameters of the target device are inferred based on the calibrated parameterized failure physical model. The extracted real-time symptom data is input into a pre-constructed fault knowledge graph for inference, and a pre-constructed health benchmark cloud model is used to perform temporal pattern recognition on the target device to obtain a fault diagnosis and early warning strategy. Thus, on the one hand, by dynamically calibrating the parameterized failure physical model online based on multi-source heterogeneous data, the model can evolve synchronously with the aging state of the physical device, overcoming the shortcomings of traditional physical models that are detached from reality; on the other hand, by using the health benchmark cloud model to perform temporal pattern recognition on the target device, it can keenly capture degradation trends in the early stages of minor drifts in multi-dimensional joint features, avoiding the risk of missed detection by the single-dimensional threshold method; furthermore, by inputting real-time symptom data into the fault knowledge graph for inference and fusing it with the recognition results of the health benchmark cloud model, high-reliability early warning can be achieved without relying on a large number of labeled fault samples, solving the technical problems of insufficient diagnostic capabilities and isolated knowledge graphs.

[0028] In some possible implementations, online dynamic calibration of the pre-constructed parameterized failure physics model based on the multi-source heterogeneous data includes: Using extended Kalman filtering or Bayesian online learning algorithms with preset filtering or online learning algorithms, and taking the multi-source heterogeneous data as observations, the key degradation parameters to be calibrated in the parameterized failure physics model are recursively estimated; the key degradation parameters include activation energy and pre-exponential factor. The obtained parameter estimates are then updated into the parameterized failure physics model.

[0029] In this embodiment, the extended Kalman filter uses a two-stage prediction-update loop to optimally estimate the model state using real-time observations. The preset online learning algorithm can be implemented as Bayesian online learning, which treats the parameters as random variables and continuously updates the posterior probability distribution of the parameters using observation data. The key degradation parameters can be implemented as the activation energy Ea and the pre-exponential factor A, which determine the sensitivity of the aging rate to temperature in the physical model.

[0030] In some possible implementations, the health parameters of the target device are deduced based on the calibrated parameterized failure physics model, including: Based on the calibrated parameterized failure physics model, the evolution path of equipment health status under various preset operating conditions is deduced; With minimizing the total lifecycle cost as the optimization objective, the long-term benefits of each preset maintenance strategy are quantitatively evaluated based on the evolution path of the health status of the target equipment, and the corresponding target maintenance strategy is determined. The total lifecycle cost includes maintenance cost and failure risk cost, and the failure risk cost is determined by multiplying the failure probability by the failure loss.

[0031] In this embodiment, multiple preset operating conditions include conditions that maintain the status quo and conditions that involve active intervention with a 10% derating reduction. Taking an oil-immersed transformer in a coal preparation plant as an example, after the equipment model is synchronized with the latest aging parameters, the evolution path over the next 30 days is simulated under Scheme 1 (maintaining the status quo, with the aggregation degree DP decreasing to 200) and Scheme 2 (active intervention with a 10% derating reduction, with DP remaining at 250). The objective function for optimization is:

[0032] in, For maintenance actions, A set of feasible strategies. This represents the probability of failure. For the loss due to malfunction, To minimize maintenance costs, the optimal maintenance strategy is calculated and output to the user, suggesting a reduction in spending to minimize overall impact.

[0033] In some possible implementations, the extracted real-time symptom data is input into a pre-built fault knowledge graph for inference, including: Real-time symptom data is input as observational evidence into a pre-constructed fault knowledge graph, which contains fault feature representations stored in the form of triples. By using the belief propagation algorithm to perform probability propagation calculations within a pre-constructed fault knowledge graph, inference results including potential fault types and confidence features derived from observational evidence are obtained.

[0034] In this embodiment, the real-time symptom data is a sudden increase in the concentration of acetylene (C2H2) in the dissolved gas analysis data of oil. Taking an oil-immersed transformer in a coal preparation plant as an example, the edges in the knowledge graph are stored in the form of triples and assigned transfer weights, including "insulation dampness" causing "H2↑, CH4↑" with a probability of 0.8, and "overheating" causing "C2H4↑, CO↑" with a probability of 0.9. When a significant sudden increase in C2H2 is detected at the current instant, this symptom is input into the knowledge graph as observation evidence, activating the corresponding "discharge-type fault" local subgraph. Using the belief propagation algorithm, the confidence level of "floating discharge" is calculated to be 0.72, and the confidence level of "internal flashover" is calculated to be 0.15. The diagnostic prediction conclusion and the corresponding causal chain are then output.

[0035] In some possible implementations, prior to performing temporal pattern recognition on the target device using a pre-built health benchmark cloud model, the following steps are included: Acquire multi-dimensional time-series data indicating that the target device is in a healthy operating state; Multivariate time-series data are used as training sample data, and latent space representation is calculated through variational autoencoder or Gaussian process latent variable model to construct the health benchmark cloud model corresponding to the target device; wherein, the health benchmark cloud model includes the benchmark mean vector and benchmark covariance matrix in the observation space.

[0036] In this embodiment, taking an oil-immersed transformer in a coal preparation plant as an example, a large amount of operational data from healthy, similar machines during normal operating periods is collected as multivariate time-series data. Specifically, this includes data from 100 machines of the same model spanning six months, forming a 10-dimensional high-level data feature, including 3-dimensional temperature features, 6 types of channel gases, and 1-dimensional vibration feature parameters. Reconstruction operations are performed using a variational autoencoder, with the latent space dimension set to 3. The overall reconstruction error is controlled within 5%. Finally, the system distribution followed by the overall "cloud" is extracted and output, and the baseline mean vector μ and baseline covariance matrix Σ are stored as the health baseline for this type of equipment. The health baseline "cloud" model is constructed using a variational autoencoder (VAE) or a Gaussian process latent variable model (GPLVM), and its loss function is:

[0037] in, For multivariate health data, Represented as a latent space. To control the balance between reconstruction and regularization, Let KL divergence be the KL divergence. For the true posterior distribution, For approximate posterior distribution, To approximate the posterior distribution The expected operation below, Let N(0,I) be the prior distribution of the latent variable z, which is usually taken as the standard normal distribution.

[0038] In some possible implementations, a pre-built health benchmark cloud model is used to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy, including: Calculate the joint deviation of the data points of the monitored target device relative to the pre-built health baseline cloud model; For the time trajectory obtained based on joint deviation tracking, temporal pattern recognition is performed through a temporal convolutional network or a Transformer encoder to obtain the identification result of whether the target device exhibits early degradation characteristics of a specific fault type; When early degradation characteristics are identified, a fault diagnosis and early warning strategy is generated for the target device based on the identification results.

[0039] In this embodiment, taking an oil-immersed transformer in a coal preparation plant as an example, a deviation sequence is formed from the distance points corresponding to the time axis and pushed into a temporal convolutional network or a Transformer encoder via a sliding window. During the real-time monitoring process, once it is identified that the deviation of the monitored target equipment slowly increases from the normal 0.8 to 1.5 within the past 7 days, with the deviation threshold set to 2.0, the temporal similarity between the trajectory and the "early partial discharge" training sample is identified by the temporal convolutional network model as reaching 85%, thereby determining that the non-stationary sliding area is a pre-fault anomaly and issuing an early warning of "potential insulation defects" 14 days in advance.

[0040] In some possible implementations, the joint deviation of the data points of the monitored target device relative to a pre-built health baseline cloud model is calculated, including: Obtain the baseline mean vector and baseline covariance matrix of the pre-constructed health baseline cloud model in the observation space; By combining the benchmark mean vector and the benchmark covariance matrix, the spatial distance between the data points of the monitored target device in the current space is calculated using Mahalanobis distance, and this distance is used as the corresponding joint deviation; wherein, the joint deviation is the spatial distance calculation result.

[0041] In this embodiment, firstly, the baseline mean vector and baseline covariance matrix of the pre-constructed health baseline cloud model are obtained in the observation space. The baseline mean vector represents the central value of each monitored parameter under healthy conditions, and the baseline covariance matrix represents the correlation and fluctuation range between the parameters. Then, combining the baseline mean vector and baseline covariance matrix, the spatial distance value of the currently monitored transformer data point in the current observation space is calculated. This spatial distance value is the joint deviation. The joint deviation is used to quantify the overall deviation between the current equipment state and the health baseline. Its characteristic is that it can eliminate covariance interference and dimensional influence among multiple variables, and can keenly capture degradation trends in the early stages of small drifts in multidimensional joint features. The calculated joint deviation is recorded in chronological order to form a deviation time series, which serves as the input for subsequent time-series pattern recognition steps. The joint deviation is defined as Mahalanobis distance:

[0042] in, and The mean vector and covariance matrix of the health benchmark "cloud" in the observation space. Let t be the data point of the target device being monitored. Let be the joint deviation at time t.

[0043] To make the methods provided in the embodiments of this application clearer, the equipment status assessment and fault diagnosis methods will be described in detail below with reference to the following examples, such as... Figure 2 As shown, the following processes are included: S201, acquire multi-source heterogeneous data associated with the target device.

[0044] For example, the target equipment is implemented as an oil-immersed power transformer in a coal preparation plant. The multi-source heterogeneous data is specifically obtained by combining online monitoring data, operating condition data, and offline test data. The online monitoring data may include: dissolved gas analysis data in the oil (such as the concentration values ​​of various components like hydrogen (H2), methane (CH4), and acetylene (C2H2), top oil temperature, winding hot spot temperature, and load operating current; the operating condition data includes the plant's daily power load curve and the environmental temperature and humidity variation sequence; the offline test data involves periodic oil quality sampling and testing values ​​(such as furfural content) and frequency domain dielectric spectrum data during shutdown inspections. This application does not limit the specific types of heterogeneous dimensions included in the embodiments.

[0045] S202 uses a preset filtering algorithm and multi-source heterogeneous data as observations to recursively estimate the key degradation parameters to be calibrated in the parameterized failure physical model, and updates the obtained parameter estimates to the parameterized failure physical model.

[0046] For example, in this transformer scenario, a thermal aging model of the insulating paper is constructed as a parametric failure physical model. The preset filtering algorithm is either an extended Kalman filter (EKF) or a Bayesian online learning algorithm. The mathematical form of this model is, for example, a function of the remaining lifetime percentage with respect to the hotspot temperature. The model includes key degradation parameters such as activation energy and pre-exponential factor. To more accurately track equipment aging, the system uses a preset extended Kalman filter (EKF) as an online learning mechanism, employing measured data such as furfural concentration as observations to continuously back-calculate and update these key parameters. The parameters are updated daily to ensure that the model output strictly matches the measured trends.

[0047] S203, based on the calibrated parameterized failure physics model, deduces the evolution path of equipment health status under various preset operating conditions.

[0048] For example, after the device model is synchronized with the latest life parameters, it simulates the state under different operating conditions in a future preset interval (e.g., 30 days later). For example, it simulates different evolution paths under scheme one (maintaining the status quo, which will cause the aggregation degree DP to drop to the critical value of 200) and scheme two (actively intervening to reduce the DP by 10%, so that the DP can be kept at a safe 250).

[0049] S204, with the goal of minimizing the total lifecycle cost, quantitatively evaluates the long-term benefits of each preset maintenance strategy based on the evolution path of the equipment's health status, and determines the corresponding target maintenance strategy.

[0050] For example, the objective function for optimization in the decision-making process comprehensively considers maintenance costs and failure risk costs, where failure risk costs are obtained by multiplying the equipment failure probability by the failure loss. Through calculation, the optimal conclusion of "recommending depreciation to minimize the overall impact" is output to the user, achieving proactive maintenance guidance.

[0051] S205. The extracted real-time symptom data is input as observational evidence into the pre-constructed fault knowledge graph. The probability propagation calculation is performed within the pre-constructed fault knowledge graph using the belief propagation algorithm to determine the inference result.

[0052] For example, the knowledge graph stores the relationships between fault modes, symptoms, and causes in the form of triples, and assigns probability weights to the edges. For instance, inputting "a sudden and significant increase in C2H2 at the current instant" as real-time symptom data into the graph activates the corresponding "discharge-type fault" subgraph. Using a belief propagation algorithm as the graph inference engine, the confidence level of "floating discharge" is calculated to be 0.72, and the confidence level of "internal flashover" is calculated to be 0.15, thereby outputting diagnostic inference conclusions and corresponding causal chains.

[0053] S206: Obtain multivariate time-series data of the target device in a healthy operating state, and construct a healthy benchmark cloud model containing the benchmark mean vector and benchmark covariance matrix through latent space representation calculation and processing.

[0054] For example, to reduce reliance on fault samples, a large amount of operational data of healthy transformers of the same model during normal periods is collected (e.g., data from a total of 100 machines over 6 months, forming a 10-dimensional high-level data feature). A variational autoencoder model is used to calculate the latent space representation, training a high-dimensional probabilistic model describing the health status distribution, i.e., the health baseline cloud model, from which the baseline mean vector and baseline covariance matrix of the observation space are extracted.

[0055] S207, Combining the benchmark mean vector and the benchmark covariance matrix, calculate the spatial distance value of the data points of the monitored target device in the current space, as the joint deviation.

[0056] For example, based on the established health baseline, the Mahalanobis distance of newly acquired data points is calculated in real time as a joint deviation, which quantifies the degree of systematic deviation between the current state and the health baseline.

[0057] S208 inputs the deviation sequence corresponding to the recorded time trajectory into a temporal convolutional network or Transformer encoder for classification pattern recognition processing to determine whether the monitored target device exhibits early degradation characteristics of a specific fault type, and generates a fault diagnosis and early warning strategy based on the recognition results.

[0058] For example, the deviation values ​​calculated over a period of time are arranged chronologically to form a deviation sequence. This sequence is then input into a temporal convolutional network or a Transformer encoder for pattern recognition. For instance, when the deviation sequence over the past 7 days is identified as slowly increasing from the normal 0.8 to 1.5, and this trajectory has an 85% similarity to the typical temporal pattern of "early partial discharge," the system determines that the device exhibits early degradation characteristics, thus issuing an early warning of "potential insulation defects."

[0059] To make the equipment condition assessment and fault diagnosis method provided in this application embodiment clearer, the following is combined with... Figure 3 Please provide an explanation. For example... Figure 3 As shown, the specific steps are as follows: S1. Construct a parameterized failure physical model library. The model is based on the equipment type and establishes mathematical expressions that reflect the core degradation mechanism, including calibrable key degradation parameters. S2. Access to multi-source heterogeneous data streams, including online monitoring data, operating condition data, and offline test data; S3. Through the online adaptive calibration engine, the key degradation parameters in the failure physical model are dynamically inverted and updated using real-time data, so that the digital twin and the physical equipment aging state evolve synchronously. For example, a parameterized failure physics model includes a transformer insulation thermal aging model, the calculation formula of which is:

[0060]

[0061] in, for t Percentage of remaining lifespan at any given moment This is the aging rate coefficient at temperature T. , For the parameters to be calibrated, This represents the initial lifespan percentage. R The gas constant is... t For runtime, Absolute temperature (unit: Kelvin, K). 。

[0062] The online adaptive calibration engine employs either an extended Kalman filter (EKF) or a Bayesian online learning algorithm to recursively estimate the degradation parameters. The calculation formula is as follows:

[0063] in, for t Time parameter estimates, For the observed values, For adaptive gain, Output function for the physical model, for t- Parameter estimates at time 1.

[0064] S4. Based on the calibrated model, conduct forward-looking simulations to deduce the evolution path of equipment health status under different load plans, environmental conditions, or maintenance interventions, and quantify and compare the long-term benefits of each strategy. At the same time, construct a knowledge graph of electrical equipment failure domain, integrate design specifications, failure physics principles, and maintenance records to form a structured causal network. For example, the fault knowledge graph is stored in the form of triples: (fault mode, cause, root cause) and (root cause, trigger, observed symptom), and edge weights are assigned to represent conditional probability or causal strength.

[0065] S5. Real-time abnormal signs are used as observation evidence and input into the knowledge graph. Probability propagation is performed through the graph reasoning engine to output potential fault hypotheses and their confidence levels. This supports small sample or zero sample fault diagnosis. Furthermore, multi-dimensional time-series data on the health status of the same type of equipment are collected to build a high-dimensional health benchmark "cloud" model. For example, the health benchmark "cloud" model is constructed using a variational autoencoder (VAE) or a Gaussian process latent variable model (GPLVM), with the following loss function:

[0066] in, For multivariate health data, Represented as a latent space. Balancing control refactoring and regularization Let KL divergence be the KL divergence. For the true posterior distribution, For approximate posterior distribution, To approximate the posterior distribution The expected operation below, Let z be the prior probability distribution of the latent variable z, which is usually set as the standard normal distribution.

[0067] S6. Calculate the joint deviation of the monitored device data points relative to the health baseline "cloud" in real time and track its time trajectory; For example, the joint deviation is defined as the Mahalanobis distance:

[0068] in, and The mean vector and covariance matrix of the health benchmark "cloud" in the observation space. Let t be the data point of the target device being monitored. Let be the joint deviation at time t.

[0069] S7. Perform time-series pattern recognition on deviations from the trajectory to determine whether they exhibit early degradation characteristics of a specific fault type, thereby achieving unsupervised early warning.

[0070] For example, the multi-scenario simulation and decision-making module outputs the optimal maintenance strategy, whose optimization objective is to minimize the total lifecycle cost, and its calculation formula is as follows:

[0071] in, For maintenance actions, A set of feasible strategies. This represents the probability of failure.

[0072] Based on the above embodiments, taking an oil-immersed power transformer in a coal preparation plant as an example, such as Figure 4 The implementation process of this method is explained in detail below. Figure 4 HSI (Health Status Index), RUL (Remaining Useful Life): 1. Construct a parameterized failure physics model library; for transformers, establish a thermal aging model for insulating paper, the calculation formula of which is:

[0073] in, for t The degree of polymerization of the insulating paper at any given time. The initial degree of polymerization of the insulating paper. The hotspot temperature is estimated using a thermal circuit model based on the load current and ambient temperature. This is the aging rate coefficient. This refers to the cumulative operating time of the equipment.

[0074] The initial parameters A = 1.5 × 10¹² and Ea = 110 kJ / mol were taken from the IEEE standard as prior values.

[0075] 2. Multi-source data access and feature extraction; Online data: DGA (H2, CH4, C2H2, etc.), top oil temperature, winding hot spot temperature (estimated by fiber optic or model), load current; Operating data: daily load curve, ambient temperature and humidity; Offline data: furfural content in periodic oil samples, FDS frequency domain dielectric spectrum.

[0076] 3. Online adaptive calibration; Use the extended Kalman filter (EKF) to apply A and Perform a recursive update; State vector:

[0077] Observation equation:

[0078] in, The observed value at time t (the estimated degree of aggregation after taking the logarithm). Let be the estimated degree of aggregation at time t. The initial degree of polymerization of the insulating paper. Pre-exponential factor, For activation energy, R Let be the ideal gas constant. T t Let be the hotspot temperature at time t. This represents the sampling time interval.

[0079] The parameters are updated daily to ensure that the model output matches the measured furfural concentration trend.

[0080] 4. Forward-looking simulation and maintenance deduction; Assuming a high-temperature warning is issued for the next 30 days, the system simulates two strategies: Strategy 1: Maintain current load → DP is expected to drop to 200 (critical value); Strategy 2: Reduce the amount by 10% → Keep DP above 250.

[0081] Considering maintenance costs and failure losses, strategy 2 is recommended.

[0082] 5. Construct a fault knowledge graph; like Figure 5 As shown, the nodes include: "inter-turn short circuit in winding", "insulation dampness", "tap switch jamming", etc. Example of edge relationship: "Insulation dampness" — [Initiates] → "H2↑, CH4↑" (probability 0.8); "Overheating" leads to "C2H4↑, CO↑" (probability 0.9).

[0083] 6. Diagnostic based on diagrams; The current DGA (Dissolved Gas Analysis) shows a sudden increase in C2H2, and the system activates the "discharge-related fault" subgraph in the spectrum. The confidence level for "levitational discharge" was calculated to be 0.72 and for "internal flashover" to be 0.15 using the belief propagation algorithm. The diagnostic conclusion and causal chain are output as follows: "C2H2↑ → Possible suspension discharge → Recommended combined infrared and ultrasound detection".

[0084] 7. Cloud-based modeling of health benchmarks; Collect 6 months of operating data from 100 identical health transformers (10 dimensions: temperature ×3, gas ×6, vibration ×1). When training a VAE model with a latent space dimension of 3, the reconstruction error is less than 5%. Store μ and Σ as health benchmarks.

[0085] 8. Early warning; Real-time calculation of Mahalanobis distance dt Discovered in the last 7 days dt The value gradually increased from 0.8 to 1.5 (threshold = 2.0). The TCN model identified that the trajectory had an 85% similarity to the "early partial discharge" training sample. The system issues a warning of "potential insulation defects" 14 days in advance, earlier than the DGA exceeding the standard.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0087] According to embodiments of this application, this application also provides a device for equipment condition assessment and fault diagnosis. For example, Figure 6 This is a schematic diagram of a device for assessing equipment condition and diagnosing faults, provided in an embodiment of this application. The device 600 includes: The acquisition module 610 is used to acquire multi-source heterogeneous data associated with a target device; wherein the target device includes electrical equipment; The deduction module 620 is used to perform online dynamic calibration of the pre-constructed parameterized failure physics model based on the multi-source heterogeneous data, and to perform state deduction of the health parameters of the target device based on the calibrated parameterized failure physics model. The early warning module 630 is used to input the extracted real-time symptom data into a pre-built fault knowledge graph for inference, and to use a pre-built health benchmark cloud model to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy.

[0088] The specific implementation methods and technical effects of each module in this embodiment are similar to those in the above-described method embodiments, and will not be repeated here.

[0089] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0090] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when it is run.

[0091] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0092] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0093] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0094] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] The method provided in this application has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A device state evaluation and failure diagnosis method characterized by, include: Acquire multi-source heterogeneous data associated with a target device; wherein the target device includes electrical equipment; Based on the multi-source heterogeneous data, the pre-constructed parameterized failure physics model is dynamically calibrated online, and the health parameters of the target device are deduced based on the calibrated parameterized failure physics model. The extracted real-time symptom data is input into a pre-built fault knowledge graph for inference, and a pre-built health benchmark cloud model is used to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy.

2. The equipment condition assessment and fault diagnosis method according to claim 1, characterized in that, The online dynamic calibration of the pre-constructed parameterized failure physics model based on the multi-source heterogeneous data includes: Using extended Kalman filtering or Bayesian online learning algorithms, and taking the multi-source heterogeneous data as observations, the key degradation parameters to be calibrated in the parameterized failure physics model are recursively estimated; the key degradation parameters include activation energy and pre-exponential factor. The obtained parameter estimates are then updated into the parameterized failure physics model.

3. The equipment condition assessment and fault diagnosis method according to claim 1, characterized in that, The process of performing state deduction of the health parameters of the target device based on the calibrated parameterized failure physics model includes: Based on the calibrated parameterized failure physics model, the evolution path of equipment health status under various preset operating conditions is deduced; With minimizing the total lifecycle cost as the optimization objective, the long-term benefits of each preset maintenance strategy are quantitatively evaluated based on the evolution path of the health status of the target device, and the corresponding target maintenance strategy is determined; the total lifecycle cost includes maintenance cost and failure risk cost.

4. The equipment condition assessment and fault diagnosis method according to claim 1, characterized in that, The step of inputting the extracted real-time symptom data into a pre-constructed fault knowledge graph for inference includes: The real-time symptom data is input as observational evidence into the pre-constructed fault knowledge graph, wherein the knowledge graph contains fault feature representations stored in the form of triples. The belief propagation algorithm is used to perform probability propagation calculations within the pre-constructed fault knowledge graph to obtain inference results, including potential fault types and confidence features, derived from the observed evidence.

5. The equipment condition assessment and fault diagnosis method according to claim 1, characterized in that, Before performing time-series pattern recognition on the target device using a pre-built health benchmark cloud model, the following steps are included: Acquire multi-dimensional time-series data of the target device in a healthy operating state; The multivariate time-series data is used as training sample data, and latent space representation is calculated using a variational autoencoder or a Gaussian process latent variable model to construct the health benchmark cloud model corresponding to the target device; wherein, the health benchmark cloud model includes the benchmark mean vector and benchmark covariance matrix in the observation space.

6. The equipment condition assessment and fault diagnosis method according to claim 1, characterized in that, The method of using a pre-built health benchmark cloud model to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy includes: Calculate the joint deviation of the data points of the monitored target device relative to the pre-built health baseline cloud model; For the time trajectory obtained based on the joint deviation tracking, temporal pattern recognition is performed through a temporal convolutional network or a Transformer encoder to obtain the identification result of whether the target device exhibits early degradation characteristics of a specific fault type; When the early degradation characteristics are identified, a fault diagnosis and early warning strategy is generated for the target device based on the identification results.

7. The equipment condition assessment and fault diagnosis method according to claim 6, characterized in that, The calculation of the joint deviation of the monitored target device's data points relative to the pre-built health baseline cloud model includes: Obtain the baseline mean vector and baseline covariance matrix of the pre-constructed health baseline cloud model in the observation space; By combining the benchmark mean vector and the benchmark covariance matrix, the spatial distance between the data points of the monitored target device in the current space is calculated using Mahalanobis distance, and this distance is used as the corresponding joint deviation; wherein, the joint deviation is the spatial distance calculation result.

8. A device for equipment condition assessment and fault diagnosis, characterized in that, include: An acquisition module is used to acquire multi-source heterogeneous data associated with a target device; wherein the target device includes electrical equipment; The deduction module is used to perform online dynamic calibration of the pre-constructed parameterized failure physics model based on the multi-source heterogeneous data, and to perform state deduction of the health parameters of the target device based on the calibrated parameterized failure physics model. The early warning module is used to input the extracted real-time symptom data into a pre-built fault knowledge graph for inference, and to use a pre-built health benchmark cloud model to perform time-series pattern recognition on the target device to obtain a fault diagnosis and early warning strategy.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed, enables the at least one processor to perform the device condition assessment and fault diagnosis method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the equipment status assessment and fault diagnosis method according to any one of claims 1-7.

Images