Method, device, equipment, storage medium and program product for optimizing fault detection model

Abstract

The application relates to an optimization method and device of a fault detection model, equipment, a storage medium and a program product. The method comprises the following steps: obtaining an initial fault detection model and multi-modal training samples; the initial fault detection model is used for fault detection of power equipment; feature extraction is performed on the multi-modal training samples by an encoder to obtain multiple hidden variables; a causal relationship of the multiple hidden variables is constructed by a hidden space module; feature reconstruction is performed on the hidden variables and the causal relationship by a decoder to obtain updated multi-modal training samples; a log-likelihood function is constructed according to the updated multi-modal training samples by an optimization module, and the initial fault detection model is optimized to obtain a fault detection model, with the maximum of the log-likelihood function as the target. The initial fault detection model comprises the encoder, the hidden space module, the decoder and the optimization module. The hidden variables and the causal relationship of the multi-modal training samples are extracted, and the fault detection precision is improved.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to an optimization method, apparatus, equipment, storage medium, and program product for a fault detection model. Background Technology

[0002] Fault diagnosis of power equipment is a core part of ensuring power grid safety. Currently, with the continuous improvement of the intelligence level of industrial systems, equipment fault detection technology is gradually developing from traditional threshold alarms and statistical analysis to data-driven intelligent diagnosis.

[0003] Current mainstream fault detection methods mainly rely on deep learning models, which extract features from multimodal monitoring data generated during equipment operation to identify and classify abnormal states.

[0004] However, the above methods suffer from low accuracy in fault identification. Summary of the Invention

[0005] Therefore, it is necessary to provide an optimized method, apparatus, equipment, storage medium, and program product for a fault detection model that can improve the accuracy of fault identification, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides an optimization method for a fault detection model, including:

[0007] An initial fault detection model and multimodal training samples are obtained; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module.

[0008] The encoder is used to extract features from the multimodal training samples to obtain multiple latent variables;

[0009] The causal relationships among the multiple latent variables are constructed using the latent space module;

[0010] The decoder reconstructs features from the latent variables and the causal relationships to obtain updated multimodal training samples;

[0011] The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function to obtain the fault detection model.

[0012] In one embodiment, obtaining multimodal training samples includes:

[0013] Determine the equipment type of the power equipment, and obtain multiple fault types that have occurred in multiple power equipment belonging to the equipment type;

[0014] Determine the representative signal type of each fault type under the device type;

[0015] Configure weighting factors for the characterization signal type according to the fault type;

[0016] The operating data of each power device under the characterizing signal type are obtained according to the weighting factor;

[0017] The running data under each of the aforementioned characterization signal types are combined to obtain the multimodal training samples.

[0018] In one embodiment, configuring a weighting factor for the characterization signal type according to the fault type includes:

[0019] Obtain the operation logs of each of the aforementioned power devices, and determine the maintenance logs corresponding to the fault types of each of the aforementioned power devices from the operation logs; the maintenance logs include the fault occurrence time, fault handling time, and signal anomaly time under each of the aforementioned characteristic signal types;

[0020] Determine the temporal dependencies of each maintenance log and calculate the matching degree of the temporal dependencies of each maintenance log;

[0021] The target dependency is determined based on the matching degree, and a weight factor is configured for the characterization signal type based on the target dependency.

[0022] In one embodiment, configuring weight factors for the representation signal type according to the target dependency includes:

[0023] For each maintenance log, identify the abnormal time points between the fault occurrence time and the fault handling time;

[0024] If the characteristic signal type of the abnormal time point is only one, then the characteristic signal type of the abnormal time point is determined to be the first key type, and weight factors are configured for the first key type and other characteristic signal types according to the first preset weight rule.

[0025] If the abnormal time point belongs to at least two types of characterization signals, then the abnormal time point closest to the fault handling time is determined, and a weighting factor is configured for the characterization signal type based on the time interval between the abnormal time point and the fault handling time.

[0026] In one embodiment, the feature extraction of the multimodal training samples by the encoder to obtain multiple latent variables includes:

[0027] The multimodal training samples are input into the encoder of the initial fault detection model, and feature data of the multimodal training samples under each type of characterization signal are extracted respectively.

[0028] The feature data of each of the aforementioned characterization signal types are concatenated to obtain a feature sequence;

[0029] Extract the hidden state vector of the feature sequence, and calculate the posterior distribution and the latent variables corresponding to the posterior distribution based on the hidden state vector.

[0030] In one embodiment, the step of constructing a log-likelihood function based on the updated multimodal training samples by the optimization module, and optimizing the initial fault detection model with the objective of maximizing the log-likelihood function to obtain a fault detection model, includes:

[0031] Obtain the prior and posterior distributions of the multimodal training samples;

[0032] The causal loss is determined based on the prior distribution and the posterior distribution, and the reconstruction loss is determined based on the multimodal training samples and the updated multimodal training samples.

[0033] Construct a log-likelihood function based on the causal loss and the reconstruction loss;

[0034] With the goal of maximizing the log-likelihood function, the encoder, the latent space module, and the decoder are optimized to obtain a fault detection model.

[0035] Secondly, this application also provides an optimization device for a fault detection model, comprising:

[0036] An acquisition module is used to acquire an initial fault detection model and multimodal training samples; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module.

[0037] The extraction module is used to extract features from the multimodal training samples through the encoder to obtain multiple latent variables;

[0038] A construction module is used to construct the causal relationships of the multiple latent variables through the latent space module;

[0039] The reconstruction module is used to reconstruct features of the latent variables and the causal relationship through the decoder to obtain updated multimodal training samples;

[0040] An optimization module is used to construct a log-likelihood function based on the updated multimodal training samples, and optimize the initial fault detection model with the goal of maximizing the log-likelihood function to obtain a fault detection model.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0042] An initial fault detection model and multimodal training samples are obtained; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module.

[0043] The encoder is used to extract features from the multimodal training samples to obtain multiple latent variables;

[0044] The causal relationships among the multiple latent variables are constructed using the latent space module;

[0045] The decoder reconstructs features from the latent variables and the causal relationships to obtain updated multimodal training samples;

[0046] The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function to obtain the fault detection model.

[0047] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0048] An initial fault detection model and multimodal training samples are obtained; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module.

[0049] The encoder is used to extract features from the multimodal training samples to obtain multiple latent variables;

[0050] The causal relationships among the multiple latent variables are constructed using the latent space module;

[0051] The decoder reconstructs features from the latent variables and the causal relationships to obtain updated multimodal training samples;

[0052] The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function to obtain the fault detection model.

[0053] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0054] An initial fault detection model and multimodal training samples are obtained; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module.

[0055] The encoder is used to extract features from the multimodal training samples to obtain multiple latent variables;

[0056] The causal relationships among the multiple latent variables are constructed using the latent space module;

[0057] The decoder reconstructs features from the latent variables and the causal relationships to obtain updated multimodal training samples;

[0058] The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function to obtain the fault detection model.

[0059] The aforementioned fault detection model optimization method, apparatus, equipment, storage medium, and program product acquire an initial fault detection model and multimodal training samples. The initial fault detection model is used to detect faults in power equipment. An encoder extracts features from the multimodal training samples to obtain multiple latent variables. A latent space module constructs causal relationships among these latent variables. A decoder reconstructs features from the latent variables and causal relationships to obtain updated multimodal training samples. An optimization module constructs a log-likelihood function based on the updated multimodal training samples and optimizes the initial fault detection model by maximizing the log-likelihood function, thus obtaining the final fault detection model. The initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module. This method, by extracting latent variables and causal relationships from the multimodal training samples, can uncover the dependencies between physical factors, providing a basis for reverse reasoning in the subsequent application of the fault detection model and improving fault detection accuracy. Simultaneously, constructing log-likelihood enables the encoder and decoder to compress and recover normal data losslessly, allowing the latent space module to learn the physical state of power equipment and the initial fault detection model to learn the fine features of multimodal sample data as a whole, ensuring that the latent variables contain sufficient reconstruction information. Compared to the fault identification accuracy problem of traditional deep learning models, this application focuses on the hidden causal relationships in multimodal sample data, which is beneficial to improving fault detection accuracy. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is an internal structural diagram of a computer device in one embodiment.

[0062] Figure 2 This is a flowchart illustrating the optimization method for a fault detection model in one embodiment;

[0063] Figure 3 This is a schematic diagram of the process for obtaining multimodal training samples in one embodiment;

[0064] Figure 4 This is a flowchart illustrating the process of configuring weighting factors for characterizing signal types in one embodiment.

[0065] Figure 5 This is a flowchart illustrating the process of configuring weighting factors for characterizing signal types based on abnormal time points in one embodiment.

[0066] Figure 6 This is a flowchart illustrating the process of determining latent variables in one embodiment;

[0067] Figure 7 This is a flowchart illustrating the process of generating a fault detection model in one embodiment;

[0068] Figure 8 This is a structural block diagram of an optimization device for a fault detection model in one embodiment. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0070] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0071] Fault diagnosis of power equipment is a core component of ensuring power grid safety. Currently, with the continuous improvement of the intelligence level of industrial systems, equipment fault detection technology is gradually evolving from traditional threshold alarms and statistical analysis towards data-driven intelligent diagnosis. Current mainstream fault detection methods primarily rely on deep learning models, extracting features from multimodal monitoring data generated during equipment operation to identify and classify abnormal states. However, these methods suffer from low fault identification accuracy.

[0072] In view of the above-mentioned technical problems, this application provides an optimization method for a fault detection model that can improve the accuracy of the output results of the fault detection model. The following embodiments will specifically illustrate the optimization method for the fault detection model.

[0073] The optimization method for the fault detection model provided in this application embodiment can be applied to, for example, Figure 1 The computer device shown can be a server, and its internal structure diagram can be as follows: Figure 1As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores multimodal training samples of power equipment. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an optimization method for a fault detection model.

[0074] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0075] In one exemplary embodiment, such as Figure 2 As shown, an optimization method for a fault detection model is provided. This embodiment illustrates the method by applying it to a computer device. In this embodiment, the method includes:

[0076] S201, Obtain the initial fault detection model and multimodal training samples.

[0077] The initial fault detection model is used to detect faults in power equipment. This model can be any of the following: deep learning network, machine learning model, or neural network model. The initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module. Multimodal training samples can include data such as thermal imager video frames, audio waveform spectrograms, and visible light video frames, as well as other types of power equipment operating data, depending on the specific scenario; no restrictions are placed here. Each power device corresponds to a set of multimodal training samples.

[0078] In the embodiments of this application, historical operation logs of various types of power equipment can be pre-stored in a database. These historical operation logs include the type and model of the power equipment, and multimodal sample data (e.g., operation information and fault information). When it is necessary to optimize the initial fault detection model, the power equipment to which the initial fault detection model to be optimized should be applied is first determined. Then, using the type or model of the power equipment as an index, a search is performed in the database, and the searched multimodal sample data is used as the multimodal training samples required for training the current initial fault detection model.

[0079] Optionally, sensors for collecting various types of operational data, such as optical sensors, frequency sensors, and thermal imagers, can be pre-installed on various power equipment. When it is necessary to optimize the initial fault detection model, the power equipment to which the initial fault detection model to be optimized should be applied is first determined, and the operational data collected by the sensors configured on the power equipment is collected according to a preset period. The various types of operational data are then integrated to obtain multimodal training samples.

[0080] S202 extracts features from multimodal training samples using an encoder to obtain multiple latent variables.

[0081] The encoder can be a recurrent neural network, such as a Long Short-Term Memory (LSTM) network or a gated recurrent unit (RGU). The latent variable is a set of independent, physically meaningful factors that can represent the probabilistic representation of the physical state of electrical equipment on a low-dimensional manifold. For example, a latent variable can contain any combination of physical factors such as the degree of contamination of the insulation surface, the intensity of partial discharge, the contact resistance state of the contacts, the actual temperature of the contacts, and the influence of ambient humidity. Each multimodal training sample yields one latent variable.

[0082] In the embodiments of this application, the multimodal training samples of each power device can be sequentially input into the encoder for feature extraction to obtain multiple latent variables; alternatively, the multimodal training samples of each power device can be input into the encoder in parallel for feature extraction to obtain multiple latent variables.

[0083] S203 constructs causal relationships among multiple latent variables through latent space modules.

[0084] The latent space module is a latent space containing a structured causal model. The latent space module can be pre-configured with various causal discovery algorithms, such as the Peter-Clark algorithm and the LiNGAM algorithm (Linear Non-Gaussian Acyclic Model).

[0085] In the embodiments of this application, for a multimodal training sample, one type can be selected from the preset causal discovery algorithms, and the latent variables can be associated with features based on the selected causal discovery algorithm to obtain the causal relationship of the latent variables.

[0086] Optionally, feature association can be performed on latent variables based on each preset causal discovery algorithm to obtain multiple candidate causal relationships; the number of features involved in each candidate causal relationship can be determined, and the candidate causal relationship with the largest number of features can be taken as the causal relationship of the latent variable.

[0087] S204 uses a decoder to reconstruct features from latent variables and causal relationships to obtain updated multimodal training samples.

[0088] The decoder has the opposite structure to the encoder. The decoder consists of LSTM units (or units built using the Transformer architecture) and convolutional units.

[0089] In the embodiments of this application, after obtaining the latent variables and their causal relationships, the latent variables and their causal relationships are input into the decoder. First, the latent variables are decoded using an LSTM unit to obtain a feature sequence; then, the feature sequence is decomposed using a convolutional unit to obtain updated multimodal training samples. Optionally, after obtaining the updated multimodal training samples, they can be matched with the causal relationships to determine whether the updated multimodal training samples conform to the physical laws corresponding to the causal relationships. If they do not conform, it indicates that the currently reconstructed sample data still has a large deviation from the operating state of the power equipment and is not suitable for fault detection; it needs to be filtered out. If it conforms, it indicates that the currently reconstructed sample data has no deviation or only a small deviation from the operating state of the power equipment, and the currently reconstructed sample data can be used as updated multimodal training samples in the initial fault training model training process.

[0090] S205, the optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function, thus obtaining the fault detection model.

[0091] In the embodiments of this application, the optimization module obtains the posterior distribution based on the encoder output and configures the prior distribution according to the multimodal training samples. Then, it calculates the causal loss based on the posterior and prior distributions. It also calculates the reconstruction loss based on the multimodal training samples and the updated multimodal training samples. The log-likelihood function is constructed through the causal loss and the reconstruction loss. With the goal of maximizing the log-likelihood, the initial fault detection model is trained to obtain the trained fault detection model.

[0092] Optionally, in the application stage of the fault detection model, the posterior distribution and latent variables are first obtained by the optimized encoder, and new real-time multimodal data are obtained by the optimized decoder. Then, the log-likelihood is calculated. If the log-likelihood is lower than the preset threshold, it indicates that the target equipment has a fault. In this case, fault intervention can be performed on the power equipment to determine which candidate fault type the fault type is concentrated in. This candidate fault type is the fault type that actually occurs in the power equipment. In this case, the causal relationship conditional likelihood of each fault type is calculated through the latent space module, which determines the candidate fault type corresponding to the maximum causal relationship conditional likelihood.

[0093] The optimization method for the aforementioned fault detection model involves obtaining an initial fault detection model and multimodal training samples. The initial fault detection model is used to detect faults in power equipment. An encoder extracts features from the multimodal training samples to obtain multiple latent variables. A latent space module constructs causal relationships among these latent variables. A decoder reconstructs features from the latent variables and causal relationships to obtain updated multimodal training samples. An optimization module constructs a log-likelihood function based on the updated multimodal training samples and optimizes the initial fault detection model by maximizing the log-likelihood function, resulting in the final fault detection model. The initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module. This method, by extracting latent variables and causal relationships from the multimodal training samples, can uncover the dependencies between physical factors, providing a basis for reverse reasoning in the subsequent application of the fault detection model and improving fault detection accuracy. Simultaneously, constructing the log-likelihood function enables the encoder and decoder to compress and recover normal data losslessly, allowing the latent space module to learn the physical state of the power equipment and the initial fault detection model to learn the fine features of the multimodal sample data, ensuring that the latent variables contain sufficient reconstruction information. Compared to traditional deep learning models that suffer from accuracy issues in fault identification, this application focuses on the hidden causal relationships in multimodal sample data, which is beneficial to improving fault detection accuracy.

[0094] In one exemplary embodiment, such as Figure 3 As shown, obtaining multimodal training samples includes:

[0095] S301, determine the equipment type of the power equipment, and obtain multiple fault types that have occurred for multiple power equipment belonging to the equipment type.

[0096] When the power equipment is a transformer, the fault type can be winding deformation or partial discharge fault.

[0097] In the embodiments of this application, the various fault types of the power equipment can be obtained through public databases, government open data (such as data released by the National Bureau of Statistics), and official enterprise websites. Alternatively, they can be automatically crawled from data competition websites or data trading platforms using web crawling technology, and the obtained data is stored in a database for later retrieval. There is a correspondence between equipment type and fault type. It is understood that different types of power equipment will have different fault types, and may even have unique fault types. Therefore, considering the optimization accuracy of the fault detection model, this application uses equipment type as a starting point when acquiring multimodal training samples for targeted acquisition and optimization training. This ensures that the optimized fault detection model can quickly and accurately identify the fault types of power equipment in the real-world application stage.

[0098] Optionally, fault types can be retrieved from a database based on scenario requirements. This involves traversing the database, matching the device type against pre-stored device types, and if the similarity exceeds a preset similarity threshold, then the fault type corresponding to that device type is retrieved. For example, when detecting faults in a transformer, the fault type for the transformer type is determined from the database. Since the fault type corresponding to the transformer type is winding deformation, the retrieved fault type is winding deformation.

[0099] S302, determine the characteristic signal type of each fault type under the equipment type.

[0100] In the embodiments of this application, each fault type corresponds to a set of characteristic signal types. For example, for a transformer, the fault type is winding deformation, and the corresponding characteristic signal types are vibration and noise signals. Similarly, for a partial discharge fault type in a transformer, the corresponding characteristic signal types are high-frequency current and ultrasonic waves. The characteristic signal types for each fault type are pre-stored in a database. When it is necessary to determine the characteristic signal type of a power device, the fault type is matched with the pre-stored fault types, and the characteristic signal type of the successfully matched fault type is taken as the characteristic signal type of the current power device.

[0101] It should be noted that when determining the type of characterizing signal based on the fault type, the historical input data of the initial fault detection model is normal operation data. That is, for each type of characterizing signal, the historical period collects the normal operation data of the power equipment belonging to that type of characterizing signal.

[0102] S303, configure weighting factors for the characterization signal type according to the fault type.

[0103] In the embodiments of this application, different characterization signal types will play different roles in fault diagnosis or fault analysis of power equipment. For example, some characterization signal types are the signal types to which the root cause of the fault belongs, while other characterization signal types are only the signal types to which other effects caused by the root cause belong. The importance is different. Weighting factors can be configured for each characterization signal type according to its importance, and data can be collected in a differentiated manner according to the weighting factors, specifically by differentiating the amount of data.

[0104] S304, obtains the operating data of each power device under the characteristic signal type according to the weighting factor.

[0105] In the embodiments of this application, after determining the weighting factor for each characterizing signal type, the operating data of each characterizing signal type of the power equipment can be collected uniformly, and the quantity of operating data of each characterizing signal type can be determined from the collected data based on the weighting factor.

[0106] Optionally, the total number of power equipment operation data to be collected can be preset, and the total number can be multiplied by the weighting factor of each characteristic signal type to obtain the number of each characteristic signal type. Finally, the corresponding data can be collected based on the number of each characteristic signal type to obtain the power equipment operation data under each characteristic signal type.

[0107] S305 combines the running data under each type of characterizing signal to obtain multimodal training samples.

[0108] In the embodiments of this application, the operating data of the same power equipment under the same characteristic signal type are combined according to the equipment type to obtain the multimodal training samples of the power equipment under the corresponding equipment type.

[0109] The above method acquires operational data in a targeted manner by characterizing signal types, which is beneficial for targeted training of the initial fault detection model and improvement of fault detection accuracy.

[0110] In one exemplary embodiment, such as Figure 4 As shown, weighting factors are configured for the characterization signal type according to the fault type, including:

[0111] S401, obtain the operation logs of each power equipment, and determine the maintenance logs corresponding to the fault types of each power equipment from the operation logs.

[0112] The maintenance log includes the time of the fault occurrence, the time of the fault handling, and the time of signal anomaly under each type of characteristic signal.

[0113] In the embodiments of this application, the database pre-stores the operation logs of power equipment under each equipment type. The operation logs are filtered by whether they contain fault data, that is, whether the operation log contains any data or combination of fault occurrence time, fault handling time, and signal abnormality time under each characteristic signal type. If it contains any of these, the operation log is regarded as a maintenance log. This operation is performed on the operation logs of each power equipment to obtain the maintenance logs of each power equipment. The maintenance logs are then divided according to the fault type to obtain the maintenance logs for each fault type.

[0114] S402, determine the temporal dependencies of each maintenance log and calculate the matching degree of the temporal dependencies of each maintenance log.

[0115] Among them, the time-series dependency reflects which type of representation signal first causes data change after a fault occurs, and whether the representation signal that first causes data change triggers fault handling. For example, if maintenance personnel discover that data belonging to the corresponding representation signal type is abnormal, they determine that the power equipment has failed and therefore handle the fault.

[0116] In the embodiments of this application, for each maintenance log, the fault occurrence time, fault handling time, and abnormal time of each characteristic signal are sorted to obtain the temporal dependency relationship of each maintenance log. The similarity matching degree of the temporal dependency relationship of any two maintenance logs is calculated to obtain the matching degree of the temporal dependency relationship between each maintenance log and other maintenance logs.

[0117] S403, determine the target dependency based on the matching degree, and configure weight factors for the characterization signal type based on the target dependency.

[0118] In the embodiments of this application, if the matching degree of the temporal dependencies of all maintenance logs is greater than a preset similarity threshold, then all maintenance logs are determined to have the same temporal dependencies. If at least one of the matching degrees of the temporal dependencies of all maintenance logs is less than the preset similarity threshold, then the number of temporal dependencies of each type is counted, and the temporal dependency with the largest number is determined as the target dependency. After determining the target dependency, the weighting factor representing the signal type can be configured based on the abnormal time points existing in the target dependency.

[0119] The above method can measure the importance of characterizing signal types by determining target dependencies, thereby improving the accuracy of weight allocation and the accuracy of training samples for fault detection modes.

[0120] In one exemplary embodiment, such as Figure 5 As shown, weighting factors are configured for the characterization signal type based on the target dependency, including:

[0121] S501, for each maintenance log, identifies the abnormal time points included between the fault occurrence time and the fault handling time.

[0122] In the embodiments of this application, an abnormal time point can be determined by whether there is a data mutation between the fault occurrence time and the fault handling time. That is, if the running data at a certain time point has a mutation compared with the previous time point (the difference exceeds the preset difference threshold), then the time point is determined to be an abnormal time point.

[0123] S502, if the characteristic signal type of the abnormal time point is only one type, then the characteristic signal type of the abnormal time point is determined as the first key type, and weight factors are configured for the first key type and other characteristic signal types according to the first preset weight rule.

[0124] Optionally, Table 1 shows the specific details of the first preset weighting rule.

[0125] Table 1 First Preset Weighting Rules

[0126]

[0127] In the embodiments of this application, for cases where only one type of characteristic signal corresponds to an abnormal time point, the characteristic signal type of the abnormal time point is determined, and the highest weight factor is assigned to the characteristic signal type corresponding to this abnormal time point. All other characteristic signal types are assigned the same, smaller weight factor. The reason is that the abnormal time points corresponding to other characteristic signal types occur after fault handling, which is relatively delayed. That is, these abnormal time points only appear in response to the occurrence of the fault and have not yet responded to the fault handling operation. Therefore, assigning the same weight factor to these characteristic signal types is sufficient.

[0128] S503 If the characteristic signal type of the abnormal time point includes at least two types, then determine the abnormal time point closest to the fault handling time, and configure a weight factor for the characteristic signal type according to the time interval between the abnormal time point and the fault handling time.

[0129] In the embodiments of this application, for cases where the characteristic signal type of an abnormal time point includes at least two types, the abnormal time point closest to the fault handling time and the characteristic signal type of that abnormal time point can be determined, and the time interval between the abnormal time point and the fault handling time can be recorded. In practical scenarios, the time interval between the abnormal time point and the fault handling time is the criterion for determining whether its corresponding characteristic signal type is the most important type; the shorter the time interval, the greater the probability of it being the most important; the longer the time interval, the less likely it is to be important. This time interval is compared with a preset interval threshold, and a weighting factor is configured for the characteristic signal type based on the comparison result.

[0130] Optionally, if the time interval is greater than a preset interval threshold, it indicates that the characterization signal type is equally important as the characterization signal types corresponding to other time points to be analyzed. The reason is that the joint response of these characterization signal types triggered the fault handling. These characterization signal types are then designated as the second key type and assigned a larger weight factor. For other characterization signal types, the same smaller weight factor is assigned. For example, the weight factors for the second key type and other characterization signal types are configured according to the second preset weight rule. The second preset weight rule is set according to Table 2 below.

[0131] Table 2 Second Preset Weighting Rules

[0132]

[0133] Optionally, if the time interval is not greater than a preset time threshold, it indicates that the most recent characterization signal type is the most likely to trigger fault repair. This type is designated as the third key type, and the characterization signal types corresponding to other time points to be analyzed are designated as secondary key types. The third key type is assigned the highest weight factor, the secondary key types are assigned relatively high weight factors, and the other characterization signal types are assigned the lowest weight factors. For example, weight factors are configured for the third key type, secondary key types, and other characterization signal types according to the third preset weight rule. The third preset weight rule is set according to Table 3 below.

[0134] Table 3 Third Preset Weighting Rules

[0135]

[0136] The above method sets the weights of the representative signal types according to different types of abnormal time points, which can effectively measure the importance of the representative signal types, thereby improving the accuracy of weight allocation and the accuracy of training samples for fault detection modes.

[0137] In one exemplary embodiment, such as Figure 6 As shown, feature extraction is performed on the multimodal training samples using an encoder to obtain multiple latent variables, including:

[0138] S601, input the multimodal training samples into the encoder of the initial fault detection model, and extract the feature data of the multimodal training samples under each type of characterization signal.

[0139] The encoder includes convolutional units and LSTM units (or units built using the Transformer architecture). LSTM units or units built using the Transformer architecture can capture temporal dependencies. If units built using the Transformer architecture are used, their self-attention mechanism can better capture long-distance temporal dependencies between feature sequences.

[0140] In the embodiments of this application, after the multimodal training samples are input into the encoder, the convolutional unit decomposes the multimodal training samples according to the size of the convolutional kernel, and multiplies the decomposed small blocks with the convolutional kernel through matrix dot product operation to output the feature data of the multimodal training samples under each representation signal type.

[0141] S602, the feature data under each type of characterizing signal are spliced ​​together to obtain the feature sequence.

[0142] In the embodiments of this application, the convolutional unit concatenates the feature data under each characterization signal type to obtain a feature sequence, and outputs it to the LSTM unit.

[0143] S603, extract the hidden state vector of the feature sequence, and calculate the posterior distribution and the corresponding latent variables based on the hidden state vector.

[0144] In the embodiments of this application, after receiving the feature sequence, the LSTM unit extracts the hidden state vector of the feature sequence by forgetting, and performs mean processing or variance processing on the hidden state vector to obtain the posterior distribution, and samples the posterior distribution to obtain the latent variable.

[0145] In one exemplary embodiment, such as Figure 7 As shown, the optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function, resulting in a fault detection model, including:

[0146] S701, obtain the prior and posterior distributions of the multimodal training samples.

[0147] The prior distribution can be a standard normal distribution. The posterior distribution can be the posterior distribution of the latent variables, including the mean and variance.

[0148] In embodiments of this application, the computer device can configure a prior distribution based on multimodal training samples. Optionally, the parameter estimates of the multimodal training samples are used as the mean or median of the prior distribution, and the variance or standard deviation of the prior distribution is set according to its degree of variability, thereby obtaining the prior distribution. The computer device can also extract the posterior distribution of the multimodal training samples through an encoder.

[0149] S702 determines the causal loss based on the prior and posterior distributions, and the reconstruction loss based on the multimodal training samples and the updated multimodal training samples.

[0150] In the embodiments of this application, the causal loss is determined based on the prior and posterior distributions. The causal loss makes the posterior distribution approximate the prior distribution, that is, makes the posterior distribution approximate the standard normal distribution. The computer can determine the reconstruction loss through the difference between the multimodal training samples and the updated multimodal training samples.

[0151] S703, construct the log-likelihood function based on causal loss and reconstruction loss.

[0152] In the embodiments of this application, the log-likelihood is calculated based on causal loss and reconstruction loss. The purpose is to enable the encoder and decoder to compress and recover normal data without loss, to enable the latent space module to truly learn the physical state of the device, and to enable the fault detection model to learn the fine features of multimodal data samples, ensuring that the latent variables contain sufficient reconstruction information.

[0153] S704 optimizes the encoder, latent space module, and decoder with the goal of maximizing the log-likelihood function to obtain a fault detection model.

[0154] In the embodiments of this application, the initial fault detection model is trained with the goal of maximizing the log-likelihood. Optionally, maximizing the lower bound of evidence is used as the objective function. The log-likelihood can be optimized by optimizing a computable lower bound. The objective function is expressed by the following relation (1):

[0155] (1);

[0156] in, Describe the objective function. Represents multimodal data samples. Represents latent variables. Indicates the parameters of the encoder, Indicates the parameters of the decoder. Denotes the posterior distribution. Represents the likelihood function. Describe the prior distribution, Denotes the expectation of the posterior distribution. This represents the difference between the posterior and prior distributions.

[0157] Alternatively, the objective function can be optimized by introducing additional loss terms to force the decoupling of the various physical factors in the latent variables and to satisfy causal relationships.

[0158] In one exemplary embodiment, a fault detection method is also provided, which is implemented based on the fault detection model trained in the foregoing embodiments, including:

[0159] The system acquires real-time multimodal data of the target device and its preset prior distribution; inputs the real-time multimodal data into the encoder of the fault detection model to obtain the posterior distribution and its latent variables; inputs the latent variables into the decoder of the fault detection model to obtain new real-time multimodal data; calculates the log-likelihood based on the prior distribution, posterior distribution, real-time multimodal data, and new real-time multimodal data; when the log-likelihood is lower than a preset threshold, it is determined that the target device has a fault; inputs the latent variables and each candidate fault type from the preset fault type set into the latent space module of the fault detection model to calculate the causal conditional likelihood corresponding to each candidate fault type; and selects the candidate fault type corresponding to the largest causal conditional likelihood as the fault type of the target device.

[0160] Specifically, the latent variables input into the latent space module may include multiple latent variables. After determining that the target device has a fault, it is possible to analyze which latent variable corresponds to the real-time multimodal data that caused the fault. The latent variable corresponding to the fault can then be input into the latent space module in the fault detection model. This can speed up the calculation of the causal conditional likelihood of the model.

[0161] Alternatively, the calculation of the causal conditional likelihood can be expressed by the following relation (2):

[0162] (2);

[0163] in, Indicates the fault type of the target device. For the set of fault types, This indicates that we are searching for candidate fault types that maximize the conditional likelihood of the causal relationship within the set of fault types. Indicates the candidate fault type The probability of observing real-time multimodal data under intervention, Represents real-time multimodal data. This indicates fault intervention for candidate fault types within a fault type set, which implies forcibly setting the candidate fault types. occur.

[0164] The above embodiments will now be explained with an example, which does not limit the technical solutions of this application.

[0165] The target power equipment is a 220V mains-grid oil-immersed power transformer, equipped with multiple sensors, including acoustic sensors, a thermal imager, and an online dissolved gas analyzer (DGA). Multiple acoustic sensors are mounted on the transformer tank wall to capture acoustic signals generated inside the transformer. The thermal imager continuously monitors the temperature distribution of the high- and low-voltage bushings, on-load tap changers, and cooling system. The DGA monitors and uploads real-time time-series data on the concentrations of key fault gases in the oil, such as hydrogen, methane, and acetylene. Based on the multimodal data samples collected from these sensors, an optimized fault detection model is trained to learn the implicit causal relationships.

[0166] Then, these sensors are used to acquire real-time multimodal data of the transformer. When anomalies are found in the real-time multimodal data, it is determined whether there is a fault in the target power equipment. If so, the fault hypotheses in the fault type set are tested to find the root cause that best explains the anomalies in the real-time multimodal data. The transformer includes three candidate fault types: severe overload, cooler failure, and loose internal bushing connection. For severe overload, the causal relationship condition likelihood corresponding to this candidate fault type is very low; the causal relationship condition likelihood corresponding to cooler failure is also very low; while the causal relationship condition likelihood corresponding to loose internal bushing connection is the highest. Therefore, the fault type of the transformer is determined to be loose internal bushing connection.

[0167] Ultimately, the transformer can be diagnosed based on the fault type. The diagnosis indicates that the internal conductive rod connection of the A-phase high-voltage bushing is loose or has poor contact, causing persistent partial discharge. This is a high-level fault, and a power outage for maintenance is recommended as soon as possible. Furthermore, a maintenance procedure corresponding to this transformer bushing can be generated from the equipment knowledge base, dynamically adapted to the switching operation sequence under the current load conditions, and pushed to maintenance personnel. This ensures the normal operation of the transformer afterward, achieving predictive maintenance for target power equipment faults.

[0168] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0169] Based on the same inventive concept, this application also provides an optimization apparatus for a fault detection model to implement the optimization method for the fault detection model described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more embodiments of the optimization apparatus for the fault detection model provided below can be found in the limitations of the optimization method for the fault detection model described above, and will not be repeated here.

[0170] In one exemplary embodiment, such as Figure 8 As shown, an optimization device for a fault detection model is provided, comprising: an acquisition module 81, an extraction module 82, a construction module 83, a reconstruction module 84, and an optimization module 85, wherein:

[0171] The acquisition module 81 is used to acquire the initial fault detection model and multimodal training samples; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder and an optimization module;

[0172] The extraction module 82 is used to extract features from multimodal training samples through the encoder to obtain multiple latent variables;

[0173] Module 83 is used to construct causal relationships between multiple latent variables through the latent space module;

[0174] Reconstruction module 84 is used to reconstruct features of latent variables and causal relationships through the decoder to obtain updated multimodal training samples;

[0175] The optimization module 85 is used to construct a log-likelihood function based on the updated multimodal training samples, and optimize the initial fault detection model with the goal of maximizing the log-likelihood function to obtain the fault detection model.

[0176] In an exemplary embodiment, the acquisition module 81 described above includes:

[0177] The first acquisition unit is used to determine the equipment type of the power equipment and acquire multiple fault types that have occurred in multiple power equipment belonging to the equipment type.

[0178] The first determining unit is used to determine the representative signal type of each fault type under the equipment type;

[0179] A configuration unit is used to configure weighting factors for characterization signal types according to fault types;

[0180] The acquisition unit is used to acquire the operating data of each power device under the characteristic signal type according to the weighting factor;

[0181] The combination unit is used to combine the running data under various representative signal types to obtain multimodal training samples.

[0182] In an exemplary embodiment, the above configuration unit includes:

[0183] The acquisition subunit is used to acquire the operation logs of each power device and determine the maintenance logs corresponding to the fault types of each power device from the operation logs; the maintenance logs include the fault occurrence time, fault handling time, and signal anomaly time under each characteristic signal type;

[0184] Determine the sub-units to determine the temporal dependencies of each maintenance log and calculate the matching degree of the temporal dependencies of each maintenance log;

[0185] The configuration subunit is used to determine the target dependency based on the matching degree and to configure the weight factor for the characterization signal type based on the target dependency.

[0186] In an exemplary embodiment, the above-described configuration subunit is specifically used for:

[0187] For each maintenance log, identify the abnormal time points between the time of the fault occurrence and the time of the fault handling;

[0188] If the characteristic signal type of the abnormal time point contains only one type, then the characteristic signal type of the abnormal time point is determined as the first key type, and weight factors are configured for the first key type and other characteristic signal types according to the first preset weight rule.

[0189] If the abnormal time point belongs to at least two types of characterization signals, then the abnormal time point closest to the fault handling time is determined, and a weighting factor is configured for the characterization signal type based on the time interval between the abnormal time point and the fault handling time.

[0190] In an exemplary embodiment, the extraction module 82 includes:

[0191] The extraction unit is used to input multimodal training samples into the encoder of the initial fault detection model and extract the feature data of the multimodal training samples under each type of representation signal.

[0192] The splicing unit is used to splice feature data from various representative signal types to obtain a feature sequence;

[0193] The computational unit is used to extract the hidden state vector of the feature sequence and calculate the posterior distribution and the corresponding latent variables based on the hidden state vector.

[0194] In an exemplary embodiment, the optimization module 85 includes:

[0195] The second acquisition unit is used to acquire the prior and posterior distributions of the multimodal training samples;

[0196] The second determining unit is used to determine the causal loss based on the prior distribution and the posterior distribution, and to determine the reconstruction loss based on the multimodal training samples and the updated multimodal training samples.

[0197] Building units are used to construct log-likelihood functions based on causal loss and reconstruction loss;

[0198] The optimization unit is used to optimize the encoder, latent space module, and decoder with the goal of maximizing the log-likelihood function, so as to obtain the fault detection model.

[0199] Each module in the optimization device of the aforementioned fault detection model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0200] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, performs the following steps:

[0201] Obtain the initial fault detection model and multimodal training samples; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module;

[0202] Multiple latent variables are obtained by extracting features from multimodal training samples using an encoder;

[0203] Causal relationships among multiple latent variables are constructed using latent space modules;

[0204] The latent variables and causal relationships are reconstructed using a decoder to obtain updated multimodal training samples;

[0205] The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function, thus obtaining the fault detection model.

[0206] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0207] Determine the equipment type of the power equipment and obtain multiple fault types that have occurred for multiple power equipment belonging to the equipment type;

[0208] Determine the representative signal type for each fault type under the device type;

[0209] Configure weighting factors for the characterization signal type according to the fault type;

[0210] The operating data of each power device under the characteristic signal type are obtained based on the weighting factors;

[0211] The running data under each type of representative signal are combined to obtain multimodal training samples.

[0212] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0213] Obtain the operation logs of each power device, and determine the maintenance logs corresponding to the fault types of each power device from the operation logs; the maintenance logs include the fault occurrence time, fault handling time, and signal anomaly time under each characteristic signal type;

[0214] Determine the temporal dependencies of each maintenance log and calculate the matching degree of the temporal dependencies of each maintenance log;

[0215] The target dependency is determined based on the matching degree, and weighting factors are configured for the characterization signal type based on the target dependency.

[0216] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0217] For each maintenance log, identify the abnormal time points between the time of the fault occurrence and the time of the fault handling;

[0218] If the characteristic signal type of the abnormal time point contains only one type, then the characteristic signal type of the abnormal time point is determined as the first key type, and weight factors are configured for the first key type and other characteristic signal types according to the first preset weight rule.

[0219] If the abnormal time point belongs to at least two types of characterization signals, then the abnormal time point closest to the fault handling time is determined, and a weighting factor is configured for the characterization signal type based on the time interval between the abnormal time point and the fault handling time.

[0220] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0221] The multimodal training samples are input into the encoder of the initial fault detection model, and the feature data of the multimodal training samples under each type of representation signal are extracted respectively.

[0222] The feature data of each type of characterizing signal are concatenated to obtain a feature sequence;

[0223] Extract the hidden state vector of the feature sequence, and calculate the posterior distribution and the corresponding latent variables based on the hidden state vector.

[0224] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0225] Obtain the prior and posterior distributions of the multimodal training samples;

[0226] The causal loss is determined based on the prior and posterior distributions, and the reconstruction loss is determined based on the multimodal training samples and the updated multimodal training samples.

[0227] Construct a log-likelihood function based on causal loss and reconstruction loss;

[0228] By optimizing the encoder, latent space module, and decoder with the goal of maximizing the log-likelihood function, a fault detection model is obtained.

[0229] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0230] Obtain the initial fault detection model and multimodal training samples; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module;

[0231] Multiple latent variables are obtained by extracting features from multimodal training samples using an encoder;

[0232] Causal relationships among multiple latent variables are constructed using latent space modules;

[0233] The latent variables and causal relationships are reconstructed using a decoder to obtain updated multimodal training samples;

[0234] The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function, thus obtaining the fault detection model.

[0235] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0236] Determine the equipment type of the power equipment and obtain multiple fault types that have occurred for multiple power equipment belonging to the equipment type;

[0237] Determine the representative signal type for each fault type under the device type;

[0238] Configure weighting factors for the characterization signal type according to the fault type;

[0239] The operating data of each power device under the characteristic signal type are obtained based on the weighting factors;

[0240] The running data under each type of representative signal are combined to obtain multimodal training samples.

[0241] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0242] Obtain the operation logs of each power device, and determine the maintenance logs corresponding to the fault types of each power device from the operation logs; the maintenance logs include the fault occurrence time, fault handling time, and signal anomaly time under each characteristic signal type;

[0243] Determine the temporal dependencies of each maintenance log and calculate the matching degree of the temporal dependencies of each maintenance log;

[0244] The target dependency is determined based on the matching degree, and weighting factors are configured for the characterization signal type based on the target dependency.

[0245] In one embodiment, when the computer program is executed by a processor, it further performs the following steps:

[0246] For each maintenance log, identify the abnormal time points between the time of the fault occurrence and the time of the fault handling;

[0247] If the characteristic signal type of the abnormal time point contains only one type, then the characteristic signal type of the abnormal time point is determined as the first key type, and weight factors are configured for the first key type and other characteristic signal types according to the first preset weight rule.

[0248] If the abnormal time point belongs to at least two types of characterization signals, then the abnormal time point closest to the fault handling time is determined, and a weighting factor is configured for the characterization signal type based on the time interval between the abnormal time point and the fault handling time.

[0249] In one embodiment, when the computer program is executed by a processor, it further performs the following steps:

[0250] The multimodal training samples are input into the encoder of the initial fault detection model, and the feature data of the multimodal training samples under each type of representation signal are extracted respectively.

[0251] The feature data of each type of characterizing signal are concatenated to obtain a feature sequence;

[0252] Extract the hidden state vector of the feature sequence, and calculate the posterior distribution and the corresponding latent variables based on the hidden state vector.

[0253] In one embodiment, when the computer program is executed by a processor, it further performs the following steps:

[0254] Obtain the prior and posterior distributions of the multimodal training samples;

[0255] The causal loss is determined based on the prior and posterior distributions, and the reconstruction loss is determined based on the multimodal training samples and the updated multimodal training samples.

[0256] Construct a log-likelihood function based on causal loss and reconstruction loss;

[0257] By optimizing the encoder, latent space module, and decoder with the goal of maximizing the log-likelihood function, a fault detection model is obtained.

[0258] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0259] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0260] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0261] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application's patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method of optimizing a failure detection model, the method comprising: The method includes: An initial fault detection model and multimodal training samples are obtained; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module. The encoder is used to extract features from the multimodal training samples to obtain multiple latent variables; The causal relationships among the multiple latent variables are constructed using the latent space module; The decoder reconstructs features from the latent variables and the causal relationships to obtain updated multimodal training samples; The optimization module constructs a log-likelihood function based on the updated multimodal training samples, and optimizes the initial fault detection model with the goal of maximizing the log-likelihood function to obtain the fault detection model.

2. The method of claim 1, wherein, Obtaining multimodal training samples includes: Determine the equipment type of the power equipment, and obtain multiple fault types that have occurred in multiple power equipment belonging to the equipment type; Determine the representative signal type of each fault type under the device type; Configure weighting factors for the characterization signal type according to the fault type; The operating data of each power device under the characterizing signal type are obtained according to the weighting factor; The running data under each of the aforementioned characterization signal types are combined to obtain the multimodal training samples.

3. The method of claim 2, wherein, The step of configuring a weighting factor for the characterization signal type according to the fault type includes: Obtain the operation logs of each of the aforementioned power devices, and determine the maintenance logs corresponding to the fault types of each of the aforementioned power devices from the operation logs; the maintenance logs include the fault occurrence time, fault handling time, and signal anomaly time under each of the aforementioned characteristic signal types; Determine the temporal dependencies of each maintenance log and calculate the matching degree of the temporal dependencies of each maintenance log; The target dependency is determined based on the matching degree, and a weight factor is configured for the characterization signal type based on the target dependency.

4. The method of claim 3, wherein, The step of configuring weight factors for the representation signal type according to the target dependency includes: For each maintenance log, identify the abnormal time points between the fault occurrence time and the fault handling time; If the characteristic signal type of the abnormal time point is only one, then the characteristic signal type of the abnormal time point is determined to be the first key type, and weight factors are configured for the first key type and other characteristic signal types according to the first preset weight rule. If the abnormal time point belongs to at least two types of characterization signals, then the abnormal time point closest to the fault handling time is determined, and a weighting factor is configured for the characterization signal type based on the time interval between the abnormal time point and the fault handling time.

5. The method according to any one of claims 1 to 4, characterized in that, The encoder extracts features from the multimodal training samples to obtain multiple latent variables, including: The multimodal training samples are input into the encoder of the initial fault detection model, and feature data of the multimodal training samples under each type of characterization signal are extracted respectively. The feature data of each of the aforementioned characterization signal types are concatenated to obtain a feature sequence; Extract the hidden state vector of the feature sequence, and calculate the posterior distribution and the latent variables corresponding to the posterior distribution based on the hidden state vector.

6. The method according to any one of claims 1 to 4, characterized in that, The step involves constructing a log-likelihood function based on the updated multimodal training samples using the optimization module, and optimizing the initial fault detection model with the objective of maximizing the log-likelihood function to obtain the fault detection model, including: Obtain the prior and posterior distributions of the multimodal training samples; The causal loss is determined based on the prior distribution and the posterior distribution, and the reconstruction loss is determined based on the multimodal training samples and the updated multimodal training samples. Construct a log-likelihood function based on the causal loss and the reconstruction loss; With the goal of maximizing the log-likelihood function, the encoder, the latent space module, and the decoder are optimized to obtain a fault detection model.

7. An optimization device for a fault detection model, characterized in that, The device includes: An acquisition module is used to acquire an initial fault detection model and multimodal training samples; the initial fault detection model is used to detect faults in power equipment; the initial fault detection model includes an encoder, a latent space module, a decoder, and an optimization module. The extraction module is used to extract features from the multimodal training samples through the encoder to obtain multiple latent variables; A construction module is used to construct the causal relationships of the multiple latent variables through the latent space module; The reconstruction module is used to reconstruct features of the latent variables and the causal relationship through the decoder to obtain updated multimodal training samples; An optimization module is used to construct a log-likelihood function based on the updated multimodal training samples, and optimize the initial fault detection model with the goal of maximizing the log-likelihood function to obtain a fault detection model.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images