Transformer fault early warning method and system

Abstract

The invention discloses a transformer fault early warning method and system, and relates to the field of electrical equipment fault diagnosis, and the method comprises the steps: collecting multi-source operation data of a transformer through an infrared thermal image, visible light image, vibration voiceprint, electrical quantity and non-electrical quantity data collector, and carrying out the preprocessing; fusing the data into a spatio-temporal feature set by using spatio-temporal feature coding; a specific modal fault early warning expert model and a multi-modal hybrid expert model are initialized and trained, and the multi-modal hybrid expert model is optimized through dynamic weight balance loss; the trained model is applied to real-time data, and a fault early warning task is completed; according to the invention, the transformer fault diagnosis flexibility is improved; iterative parameters are reduced through gating parameters, the calculation complexity is reduced, the generalization ability is improved, and the system robustness is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of electrical equipment fault diagnosis, and in particular to a transformer fault early warning method and system. Background Art

[0002] Transformers are the core electrical equipment for power transmission and conversion in power systems, and their operating status will directly affect the stability and reliability of power system operation. However, long-term service and multi-physics field coupling environment have aggravated the frequency, complexity and uncertainty of transformer failures.

[0003] Early warning and precise positioning of transformer faults are important countermeasures for identifying potential safety risks in power systems. Traditional transformer fault warning methods mostly rely on manual inspections and single-mode sensor data. As the scale of power systems expands and the complexity of equipment structures increases, the workload of manual inspection methods increases sharply, the efficiency is low, and it cannot meet the needs of intelligent fault diagnosis of transformers; relying on single-mode sensor data, it is difficult to fully perceive the operating status of the transformer, and it lacks flexibility and adaptability. At present, multi-modal transformer fault warning methods mostly focus on a single iterative task, and there are problems such as high computational complexity and weak generalization ability in multi-objective optimization scenarios. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to achieve transformer fault warning that fully perceives the transformer operating status and adapts to multi-objective optimization scenarios.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a transformer fault early warning method, comprising:

[0008] Collect and preprocess the transformer multi-source operation data, encode the spatiotemporal features of the pre-processed multi-source operation data, and obtain the spatiotemporal fusion feature set of the transformer multi-source operation data;

[0009] Initializing a specific mode transformer fault warning expert model, wherein the specific mode transformer fault warning expert model is composed of a multi-head self-attention mechanism, an expert network, and a residual connection; using a spatiotemporal fusion feature set of transformer multi-source operation data to train the specific mode transformer fault warning expert model;

[0010] Initialize a multimodal transformer fault warning hybrid expert model, wherein the multimodal transformer fault warning hybrid expert model is composed of a multimodal transformer fault warning hybrid expert module, and the multimodal transformer fault warning hybrid expert module is composed of routing, multi-head attention mechanism, hybrid expert network and residual connection; integrate the trained specific mode transformer fault warning expert model into the hybrid expert network layer of the multimodal transformer fault warning hybrid expert module; use the spatiotemporal fusion feature set of transformer multi-source operation data to train the multimodal transformer fault warning hybrid expert model;

[0011] The trained multi-modal transformer fault warning hybrid expert model is loaded, and the real-time transformer multi-source operation data after data preprocessing and feature encoding is used as input to obtain the fine-grained branch information of the transformer fault warning subtask and complete the fault warning.

[0012] As a preferred solution of the transformer fault early warning method, wherein:

[0013] The collecting of transformer multi-source operation data includes:

[0014] Deploy infrared thermal imaging, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data collectors and sensor measurement points to collect the transformer's full-range infrared thermal image, full-range visible light image, vibration soundprint signals, electrical quantity measurement data, and temperature, humidity, and vibration auxiliary data;

[0015] Among them, the transformer multi-source data operation data acquisition frequency is f, and the vibration soundprint signal acquisition time window is t v , t v ≤f -1 .

[0016] As a preferred solution of the transformer fault early warning method, wherein:

[0017] The preprocessing includes: infrared thermal image and visible light image preprocessing includes data enhancement processing, tensor conversion and normalization processing; the data enhancement processing includes Gaussian filtering processing;

[0018] The vibration voiceprint signal preprocessing includes information gain adjustment, noise reduction, framing, windowing, Fourier transform, Mel filtering, logarithmic transform, Mel cepstrum, data enhancement and normalization processing;

[0019] The preprocessing of electrical quantity data and non-electrical quantity includes data cleaning, numerical processing and normalization processing; among them, per-unit processing is used as the normalization processing method for electrical quantity data.

[0020] As a preferred solution of the transformer fault early warning method, wherein:

[0021] The spatiotemporal feature encoding of the preprocessed multi-source operation data includes:

[0022] Get n Fusion features of multi-source operation data of historical time-step transformer after data preprocessing k=1,2,…,t n ; They represent the infrared thermal image, visible light image, vibration soundprint, electrical quantity data and non-electrical quantity data after data preprocessing at the kth time step respectively; t n Historical time step transformer multi-source operation fusion feature prediction t n+1 timestep transformer state;

[0023] For the fusion feature X after data preprocessing m ,Using image, voiceprint, and time series feature encoders to extract the spatiotemporal data features of each modality and map them to a low-dimensional feature space;

[0024] Infrared thermal imaging data Visible light image data Input to the image feature encoder and convert the vibration voiceprint data Input to the voiceprint feature encoder, convert the electrical quantity data Non-electrical quantity data Input to the temporal feature encoder.

[0025] As a preferred solution of the transformer fault early warning method, wherein:

[0026] The obtained spatiotemporal fusion feature set of transformer multi-source operation data includes:

[0027] After feature encoding, the feature encoding of each mode of transformer multi-source operation data is obtained, that is, the spatiotemporal fusion feature set of transformer multi-source operation data p=1,2,…,t n ; Respectively represent the t p Infrared thermal images, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data after time step encoding by spatiotemporal features.

[0028] As a preferred solution of the transformer fault early warning method, wherein:

[0029] The method of training a transformer fault warning expert model for a specific mode by using a spatiotemporal fusion feature set of transformer multi-source operation data includes:

[0030] Get the spatiotemporal fusion feature set X of transformer multi-source operation data T, based on the feature set, parallel training of N specific mode transformer fault warning expert models;

[0031] The self-attention mechanism in each specific mode transformer fault warning expert model is initialized by Gaussian distribution and the parameters are locked;

[0032] The training process uses dynamic weight balance loss L DWA As the loss function, through the multi-objective optimization method, the specific mode transformer fault warning expert model is used as the shared underlying network, and an independent tower network is assigned to each optimization task to obtain the fine-grained branch information of each subtask of transformer fault warning;

[0033] Each optimization task is based on the focal loss L FL As a loss function, it is expressed as:

[0034]

[0035] Among them, w i represents the weight of each optimization task, T represents the temperature coefficient used to soften the weight distribution, γ i represents the weight of each task after normalization and scaling by the α coefficient factor, and n represents the number of tasks.

[0036] As a preferred solution of the transformer fault early warning method, wherein:

[0037] The multi-mode transformer fault warning hybrid expert module f MoE The expression is:

[0038]

[0039] Among them, X l-1 represents the output of the l-1th multi-mode transformer fault early warning hybrid expert module. In the lth multi-mode transformer fault early warning hybrid expert module, represents the routing gating weight, d represents the feature dimension of the input feature x, M represents the number of expert networks in the hybrid expert network, and f e (·) i represents the i-th expert network in the hybrid expert network, g k,l (·) i represents the probability of the kth task being assigned to the ith expert network; N represents the number of activated expert networks, that is, the N expert networks with the highest allocation ratio are selected to obtain the output result of the multi-modal transformer fault warning hybrid expert module;

[0040] The multi-modal transformer fault warning hybrid expert model is trained using the spatiotemporal fusion feature set of transformer multi-source operation data. The training process uses a dynamic weight balance loss L DWA as a loss function.

[0041] In a second aspect, an embodiment of the present invention provides a transformer fault early warning system, comprising:

[0042] A data feature encoding module is used to collect and preprocess the transformer multi-source operation data, perform spatiotemporal feature encoding on the pre-processed multi-source operation data, and obtain a spatiotemporal fusion feature set of the transformer multi-source operation data;

[0043] A data feature encoding module is used to collect and preprocess the transformer multi-source operation data, perform spatiotemporal feature encoding on the pre-processed multi-source operation data, and obtain a spatiotemporal fusion feature set of the transformer multi-source operation data;

[0044] An offline training module is used to initialize a specific mode transformer fault warning expert model, wherein the specific mode transformer fault warning expert model is composed of a multi-head self-attention mechanism, an expert network and a residual connection; the specific mode transformer fault warning expert model is trained using a spatiotemporal fusion feature set of transformer multi-source operation data; a multi-modal transformer fault warning hybrid expert model is initialized, wherein the multi-modal transformer fault warning hybrid expert model is composed of a multi-modal transformer fault warning hybrid expert module, and the multi-modal transformer fault warning hybrid expert module is composed of routing, a multi-head attention mechanism, a hybrid expert network and a residual connection; the trained specific mode transformer fault warning expert model is integrated into the hybrid expert network layer of the multi-modal transformer fault warning hybrid expert module; the multi-modal transformer fault warning hybrid expert model is trained using a spatiotemporal fusion feature set of transformer multi-source operation data;

[0045] The online fault warning module is used to load the trained multi-modal transformer fault warning hybrid expert model, take the real-time transformer multi-source operation data after data preprocessing and feature encoding as input, obtain the fine-grained branch information of the transformer fault warning subtask, and complete the fault warning.

[0046] In a third aspect, an embodiment of the present invention provides a computing device, including:

[0047] Memory and processor;

[0048] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the transformer fault warning method as described in any embodiment of the present invention.

[0049] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which implement the transformer fault warning method when executed by a processor.

[0050] Beneficial effects of the present invention: In view of the problem that single-modal sensing data is difficult to fully perceive the operating status of the transformer, the present invention adopts a multi-modal fusion method to macroscopically perceive transformer faults through infrared thermal images, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data, thereby improving the flexibility of transformer fault diagnosis; in view of the problem that existing methods are difficult to adapt to multi-objective optimization scenarios, a multi-task learning method and a multi-modal hybrid expert model are adopted to iterate in parallel and dynamically update the transformer fault warning model. Compared with the existing methods, the iteration parameters are reduced by gating parameters to reduce the computational complexity, improve its generalization ability, and enhance the robustness of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0052] Figure 1 is an overall flow chart of the transformer fault early warning method according to the first embodiment of the present invention;

[0053] Figure 2 It is a schematic diagram of an iterative process of a transformer fault early warning method of a transformer fault early warning method according to a first embodiment of the present invention;

[0054] Figure 3 It is a schematic diagram of a multi-modal transformer fault early warning hybrid expert model of the transformer fault early warning method according to the first embodiment of the present invention;

[0055] Figure 4 It is a fine-grained branch diagram of the transformer fault warning subtask of the transformer fault warning method described in the first embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0057] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0059] Example 1

[0060] Reference Figure 1-Figure 4 , which is the first embodiment of the present invention, provides a transformer fault early warning method, comprising:

[0061] S1: Collect and preprocess transformer multi-source operation data, encode the spatiotemporal features of the preprocessed multi-source operation data, and obtain a spatiotemporal fusion feature set of transformer multi-source operation data;

[0062] In the embodiment of the present application, collecting transformer multi-source operation data includes:

[0063] Deploy infrared thermal imaging, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data collectors and sensor measurement points to collect the transformer's full-range infrared thermal image, full-range visible light image, vibration soundprint signals, electrical quantity measurement data, and temperature, humidity, and vibration auxiliary data;

[0064] Among them, the transformer multi-source data operation data acquisition frequency is f, and the vibration soundprint signal acquisition time window is t v , t v ≤f -1 .

[0065] Exemplarily, infrared thermal image acquisition: deploy infrared thermal image collectors at short distances to obtain infrared thermal images of the entire transformer; visible light image acquisition: deploy visible light image collectors at short distances to obtain visible light images of the entire transformer;

[0066] The short-distance limited range means that for a single transformer, the overall image of the transformer can be obtained; vibration soundprint collection: insulation soundprint collection nodes at key locations such as the device body, cooling device, and protection device are used to obtain transformer vibration soundprint signals; electrical quantity data collection: electrical quantity data collection measurement points are deployed in the primary and secondary circuits related to the transformer to obtain transformer electrical quantity data; non-electrical quantity data collection: temperature, humidity, and vibration auxiliary data collection measurement points are deployed at key locations such as the device body, cooling device, and protection device to obtain transformer non-electrical quantity data.

[0067] In the embodiment of the present application, the preprocessing includes: the infrared thermal image and visible light image preprocessing includes data enhancement processing, tensor conversion and normalization processing; the data enhancement processing includes Gaussian filtering processing;

[0068] It should be noted that Gaussian filtering is used to reduce Gaussian noise caused by high temperature, lighting and other electrical components at the sensor measurement point.

[0069] The vibration voiceprint signal preprocessing includes information gain adjustment, noise reduction, framing, windowing, Fourier transform, Mel filtering, logarithmic transform, Mel cepstrum, data enhancement and normalization processing;

[0070] The preprocessing of electrical quantity data and non-electrical quantity includes data cleaning, numerical processing and normalization processing; among them, per-unit processing is used as the normalization processing method for electrical quantity data;

[0071] In an embodiment of the present application, performing spatiotemporal feature encoding on the preprocessed multi-source operation data includes:

[0072] Get n Fusion features of multi-source operation data of historical time-step transformer after data preprocessing k=1,2,…,t n ; They represent the infrared thermal image, visible light image, vibration soundprint, electrical quantity data and non-electrical quantity data after data preprocessing at the kth time step respectively; t n Historical time step transformer multi-source operation fusion feature prediction t n+1 Time step transformer state.

[0073] For the fusion feature X after data preprocessing m ,Image, voiceprint, and time series feature encoders are used to extract the spatiotemporal data features of each modality and map them to a low-dimensional feature space.

[0074] Specifically, each feature encoder has been pre-trained with transformer multi-source operation historical data; the image feature encoder can adopt image pre-training models such as ResNet series, Inception series, and ViT series; the voiceprint feature encoder can adopt pre-training models such as Transformer series; the time series feature encoder can adopt time series pre-training models such as time series CNNs series and CNNs-RNNs series.

[0075] Infrared thermal imaging data Visible light image data Input to the image feature encoder and convert the vibration voiceprint data Input to the voiceprint feature encoder, convert the electrical quantity data Non-electrical quantity data Input to the temporal feature encoder;

[0076] In the embodiment of the present application, the spatiotemporal fusion feature set of transformer multi-source operation data is obtained, including:

[0077] After feature encoding, the feature encoding of each mode of transformer multi-source operation data is obtained, that is, the spatiotemporal fusion feature set of transformer multi-source operation data p=1,2,…,t n . Respectively represent the t p Infrared thermal images, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data after time step encoding by spatiotemporal features.

[0078] S2: Initialize a specific mode transformer fault warning expert model, which is composed of a multi-head self-attention mechanism, an expert network and a residual connection; use a spatiotemporal fusion feature set of transformer multi-source operation data to train the specific mode transformer fault warning expert model;

[0079] In the embodiment of the present application, the training of a specific mode transformer fault warning expert model using a spatiotemporal fusion feature set of transformer multi-source operation data includes:

[0080] Get the spatiotemporal fusion feature set X of transformer multi-source operation data T , based on the feature set, parallel training of N specific mode transformer fault warning expert models;

[0081] It should be noted that N specific modal transformer fault warning expert models are trained in parallel based on the feature set to improve the characterization performance of each modal expert model for specific modal data and the overall performance under multiple tasks.

[0082] The self-attention mechanism in each specific mode transformer fault warning expert model is initialized by Gaussian distribution and the parameters are locked;

[0083] The training process uses dynamic weight balance loss L DWA As the loss function, through the multi-objective optimization method, the specific mode transformer fault warning expert model is used as the shared underlying network, and an independent tower network is assigned to each optimization task to obtain the fine-grained branch information of each subtask of transformer fault warning;

[0084] Each optimization task is based on the focal loss L FL As a loss function, it is expressed as:

[0085]

[0086] Among them, w i represents the weight of each optimization task, T represents the temperature coefficient used to soften the weight distribution, γ irepresents the weight of each task after normalization and scaling by the α coefficient factor, and n represents the number of tasks.

[0087] It should be noted that due to the long tail effect problem derived from the scarcity of transformer fault samples, each optimization task is based on the focus loss L FL as a loss function.

[0088] S3: Initialize a multimodal transformer fault warning hybrid expert model, wherein the multimodal transformer fault warning hybrid expert model is composed of a multimodal transformer fault warning hybrid expert module, and the multimodal transformer fault warning hybrid expert module is composed of routing, multi-head attention mechanism, hybrid expert network and residual connection; integrate the trained specific mode transformer fault warning expert model into the hybrid expert network layer of the multimodal transformer fault warning hybrid expert module; use the spatiotemporal fusion feature set of transformer multi-source operation data to train the multimodal transformer fault warning hybrid expert model;

[0089] In the embodiment of the present application, the multi-mode transformer fault warning hybrid expert module f MoE The expression is:

[0090]

[0091] Among them, X l-1 represents the output of the l-1th multi-mode transformer fault early warning hybrid expert module. In the lth multi-mode transformer fault early warning hybrid expert module, represents the routing gating weight, d represents the feature dimension of the input feature x, M represents the number of expert networks in the hybrid expert network, and f e (·) i represents the i-th expert network in the hybrid expert network, g k,l (·) i represents the probability of the kth task being assigned to the ith expert network. N represents the number of activated expert networks, that is, the N expert networks with the highest allocation ratio are selected to obtain the output results of the multi-modal transformer fault warning hybrid expert module;

[0092] The multi-modal transformer fault warning hybrid expert model is trained using the spatiotemporal fusion feature set of transformer multi-source operation data. The training process uses a dynamic weight balance loss L DWA as a loss function.

[0093] It should be noted that the trained specific mode transformer fault warning expert model is integrated into the hybrid expert network layer of the multi-mode transformer fault warning hybrid expert module to shorten the iteration cycle.

[0094] like Figure 2 As shown in FIG. 1 , it is a schematic diagram of the iterative process of the transformer fault early warning method; Figure 3 As shown, this is a schematic diagram of the multi-modal transformer fault warning hybrid expert model.

[0095] S4: Load the trained multi-modal transformer fault warning hybrid expert model, take the real-time transformer multi-source operation data after data preprocessing and feature encoding as input, obtain the fine-grained branch information of the transformer fault warning subtask, and complete the fault warning.

[0096] In the embodiment of the present application, a multi-modal transformer fault warning hybrid expert model is loaded, and the real-time transformer multi-source operation data after data preprocessing and feature encoding is used as input to obtain the fine-grained branch information of each subtask. Figure 4 As shown, it is a schematic diagram of fine-grained branches of the transformer fault warning subtask. The fine-grained branches of the transformer fault warning subtask in the embodiment of the present invention include but are not limited to the illustrated contents.

[0097] It should be noted that, in practical applications, the above steps S1-S4 may be cycled periodically to iteratively update the multi-modal transformer early warning hybrid expert model.

[0098] The above is a schematic scheme of the transformer fault early warning method of this embodiment. It should be noted that the technical scheme of the transformer fault early warning system and the technical scheme of the above transformer fault early warning method belong to the same concept, and the details of the technical scheme of the transformer fault early warning system in this embodiment that are not described in detail can all be referred to the description of the technical scheme of the above transformer fault early warning method.

[0099] The transformer fault early warning system in this embodiment includes:

[0100] A data feature encoding module is used to collect and preprocess the transformer multi-source operation data, perform spatiotemporal feature encoding on the pre-processed multi-source operation data, and obtain a spatiotemporal fusion feature set of the transformer multi-source operation data;

[0101] A data feature encoding module is used to collect and preprocess the transformer multi-source operation data, perform spatiotemporal feature encoding on the pre-processed multi-source operation data, and obtain a spatiotemporal fusion feature set of the transformer multi-source operation data;

[0102] An offline training module is used to initialize a specific mode transformer fault warning expert model, wherein the specific mode transformer fault warning expert model is composed of a multi-head self-attention mechanism, an expert network and a residual connection; the specific mode transformer fault warning expert model is trained using a spatiotemporal fusion feature set of transformer multi-source operation data; a multi-modal transformer fault warning hybrid expert model is initialized, wherein the multi-modal transformer fault warning hybrid expert model is composed of a multi-modal transformer fault warning hybrid expert module, and the multi-modal transformer fault warning hybrid expert module is composed of routing, a multi-head attention mechanism, a hybrid expert network and a residual connection; the trained specific mode transformer fault warning expert model is integrated into the hybrid expert network layer of the multi-modal transformer fault warning hybrid expert module; the multi-modal transformer fault warning hybrid expert model is trained using a spatiotemporal fusion feature set of transformer multi-source operation data;

[0103] The online fault warning module is used to load the trained multi-modal transformer fault warning hybrid expert model, take the real-time transformer multi-source operation data after data preprocessing and feature encoding as input, obtain the fine-grained branch information of the transformer fault warning subtask, and complete the fault warning.

[0104] This embodiment further provides a computing device, which is applicable to the transformer fault early warning method, and includes:

[0105] Memory and processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the transformer fault early warning method proposed in the above embodiment.

[0106] This embodiment further provides a storage medium on which a computer program is stored. When the program is executed by a processor, the transformer fault early warning method proposed in the above embodiment is implemented.

[0107] The storage medium proposed in this embodiment and the transformer fault warning method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0108] Example 2

[0109] Referring to Table 1-Table 2, an embodiment of the present invention provides a transformer fault early warning method. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0110] This example uses the SC10-200 / 0.38 dry-type transformer simulation data set as the simulation experiment data, and divides it into a training set and a test set at a ratio of 8:2. The sampling period of this data set is 1 day, the acquisition frequency is set to 1Hz, and the vibration soundprint signal acquisition time window is set to 0.2 seconds, covering fault location and fault property data.

[0111] This experiment was conducted in the following experimental environment. Hardware environment: CPU is Intel Xeon(R) W-2255CPU@3.70GHz, GPU is NVIDIA GeForce RTX 3090, and running memory is 128GB. Software environment: operating system is Ubuntu 20.04.3LTS, integrated development environment PyCharm 2021.1.2x64, programming language is Python 3.8, and deep learning framework is Pytorch1.8.0.

[0112] The hyperparameters of the experiment in this embodiment are set as follows: Adam is used as the optimizer, the learning rate of the multi-objective optimization task and each optimization task is set to 0.001, the batch size is set to 1024, and the number of training rounds is set to 1000.

[0113] 1. Data collection and preprocessing:

[0114] Data collection:

[0115] Deploy infrared thermal imaging, visible light imaging, vibration soundprint, electrical quantity data and non-electrical quantity data collectors.

[0116] Preprocessing:

[0117] Infrared thermal images and visible light images: Gaussian filtering data enhancement, tensor conversion and normalization.

[0118] Vibration voiceprint: information gain adjustment, noise reduction, framing, windowing, Fourier transform, Mel filtering, logarithmic transform, Mel cepstrum, data enhancement and normalization.

[0119] Electrical quantities and non-electrical quantities: data cleaning, numerical processing, standardization and normalization.

[0120] 2. Feature encoding:

[0121] Get n = Transformer multi-source operation data of 6 historical time steps.

[0122] Using image, voiceprint, and time series feature encoders:

[0123] Infrared thermal images and visible light images: input to the Inception-v4 image feature encoder.

[0124] Vibration voiceprint: input to the Transformer voiceprint feature encoder.

[0125] Electrical quantity data and non-electrical quantity data: input to the CNN-LSTM time series feature encoder.

[0126] Get the spatiotemporal fusion feature set

[0127] 3. Model training:

[0128] Specific mode transformer fault warning expert model:

[0129] Construct an expert model with multi-head self-attention mechanism, expert network and residual connection.

[0130] Using spatiotemporal fusion feature set X T Conduct training.

[0131] Adopt dynamic weight balancing loss L DWA As the multi-objective optimization loss function, the focal loss L is adopted FL As the loss function of each optimization task.

[0132] Multi-mode transformer fault warning hybrid expert model:

[0133] Initialize the hybrid expert module consisting of routing, multi-head attention mechanism, hybrid expert network and residual connection.

[0134] Integrate the modality-specific expert model into the hybrid expert network and train the hybrid expert model.

[0135] Using X T Perform training and optimize model parameters.

[0136] 4. Fault warning:

[0137] Load the trained multimodal mixture of experts model.

[0138] Input preprocessed and feature-encoded real-time transformer multi-source operation data.

[0139] Obtain fine-grained fault warning information.

[0140] This embodiment uses accuracy (ACC), precision (P), recall rate (TPR), false positive rate (FPR) and F1 score (F1-Score, F1) as evaluation indicators.

[0141] The results of the fault location experiment are shown in Table 1:

[0142] Table 1 Performance evaluation results of transformer fault warning and fault location tasks

[0143]

[0144] Table 2 Transformer fault warning fault nature task performance evaluation results

[0145]

[0146] Method performance: The warning results based on actual faults are compared with those of the warning model. The ACC of the fault location and fault nature tasks of transformer fault warning are 97.49% and 96.12% respectively. The focus loss function enhances the ability to locate rare faults and warn of fault nature categories; the dynamic weight balance loss function strengthens the comprehensive performance of the multimodal warning model; the multi-objective optimization method and multi-modal hybrid expert model are used to iterate in parallel and dynamically update the transformer fault warning model, which effectively reduces the iteration parameters, reduces the computational complexity, and improves the generalization ability of the method and the robustness of the system.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A transformer fault early warning method, characterized in that: include: Collect and preprocess the transformer multi-source operation data, encode the spatiotemporal features of the pre-processed multi-source operation data, and obtain the spatiotemporal fusion feature set of the transformer multi-source operation data; Initializing a specific mode transformer fault warning expert model, wherein the specific mode transformer fault warning expert model is composed of a multi-head self-attention mechanism, an expert network, and a residual connection; using a spatiotemporal fusion feature set of transformer multi-source operation data to train the specific mode transformer fault warning expert model; Initialize a multimodal transformer fault warning hybrid expert model, wherein the multimodal transformer fault warning hybrid expert model is composed of a multimodal transformer fault warning hybrid expert module, and the multimodal transformer fault warning hybrid expert module is composed of routing, multi-head attention mechanism, hybrid expert network and residual connection; integrate the trained specific mode transformer fault warning expert model into the hybrid expert network layer of the multimodal transformer fault warning hybrid expert module; use the spatiotemporal fusion feature set of transformer multi-source operation data to train the multimodal transformer fault warning hybrid expert model; The trained multi-modal transformer fault warning hybrid expert model is loaded, and the real-time transformer multi-source operation data after data preprocessing and feature encoding is used as input to obtain the fine-grained branch information of the transformer fault warning subtask and complete the fault warning.

2. The transformer fault early warning method according to claim 1, characterized in that: The collecting of transformer multi-source operation data includes: Deploy infrared thermal imaging, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data collectors and sensor measurement points to collect the transformer's full-range infrared thermal image, full-range visible light image, vibration soundprint signals, electrical quantity measurement data, and temperature, humidity, and vibration auxiliary data; Among them, the transformer multi-source data operation data acquisition frequency is f, and the vibration soundprint signal acquisition time window is t v , t v ≤f -1 .

3. The transformer fault early warning method according to claim 2, characterized in that: The pre-processing comprises: The infrared thermal image and visible light image preprocessing includes data enhancement processing, tensor conversion and normalization processing; the data enhancement processing includes Gaussian filtering processing; The vibration voiceprint signal preprocessing includes information gain adjustment, noise reduction, framing, windowing, Fourier transform, Mel filtering, logarithmic transform, Mel cepstrum, data enhancement and normalization processing; The preprocessing of electrical quantity data and non-electrical quantity includes data cleaning, numerical processing and normalization processing; among them, per-unit processing is used as the normalization processing method for electrical quantity data.

4. The transformer fault early warning method according to claim 3, characterized in that: The spatiotemporal feature encoding of the preprocessed multi-source operation data includes: Get n Fusion features of multi-source operation data of historical time-step transformer after data preprocessing They represent the infrared thermal image, visible light image, vibration soundprint, electrical quantity data and non-electrical quantity data after data preprocessing at the kth time step respectively; t n Historical time step transformer multi-source operation fusion feature prediction t n+1 timestep transformer state; For the fusion feature X after data preprocessing m ,Using image, voiceprint, and time series feature encoders to extract the spatiotemporal data features of each modality and map them to a low-dimensional feature space; Infrared thermal imaging data Visible light image data Input to the image feature encoder and convert the vibration voiceprint data Input to the voiceprint feature encoder, convert the electrical quantity data Non-electrical quantity data Input to the temporal feature encoder.

5. The transformer fault early warning method according to claim 4, characterized in that: The obtained spatiotemporal fusion feature set of transformer multi-source operation data includes: After feature encoding, the feature encoding of each mode of transformer multi-source operation data is obtained, that is, the spatiotemporal fusion feature set of transformer multi-source operation data Respectively represent the t p Infrared thermal images, visible light images, vibration soundprints, electrical quantity data and non-electrical quantity data after time step encoding by spatiotemporal features.

6. The transformer fault early warning method according to claim 5, characterized in that: The method of training a transformer fault warning expert model for a specific mode by using a spatiotemporal fusion feature set of transformer multi-source operation data includes: Get the spatiotemporal fusion feature set X of transformer multi-source operation data T , based on the feature set, parallel training of N specific mode transformer fault warning expert models; The self-attention mechanism in each specific mode transformer fault warning expert model is initialized by Gaussian distribution and the parameters are locked; The training process uses dynamic weight balance loss L DWA As the loss function, through the multi-objective optimization method, the specific mode transformer fault warning expert model is used as the shared underlying network, and an independent tower network is assigned to each optimization task to obtain the fine-grained branch information of each subtask of transformer fault warning; Each optimization task is based on the focal loss L FL As a loss function, it is expressed as: Among them, w i represents the weight of each optimization task, T represents the temperature coefficient used to soften the weight distribution, γ i represents the weight of each task after normalization and scaling by the α coefficient factor, and n represents the number of tasks.

7. The transformer fault early warning method according to claim 6, characterized in that: The multi-mode transformer fault warning hybrid expert module f MoE The expression is: Among them, X l-1 represents the output of the l-1th multi-mode transformer fault early warning hybrid expert module. In the lth multi-mode transformer fault early warning hybrid expert module, represents the routing gating weight, d represents the feature dimension of the input feature x, M represents the number of expert networks in the hybrid expert network, and f e (·) i represents the i-th expert network in the hybrid expert network, g k,l (·) i represents the probability of the kth task being assigned to the ith expert network; N represents the number of activated expert networks, that is, the N expert networks with the highest allocation ratio are selected to obtain the output result of the multi-modal transformer fault warning hybrid expert module; The multi-modal transformer fault warning hybrid expert model is trained using the spatiotemporal fusion feature set of transformer multi-source operation data. The training process uses a dynamic weight balance loss L DWA as a loss function.

8. A system using the transformer fault early warning method according to any one of claims 1 to 7, characterized in that: include: A data feature encoding module is used to collect and preprocess the transformer multi-source operation data, perform spatiotemporal feature encoding on the pre-processed multi-source operation data, and obtain a spatiotemporal fusion feature set of the transformer multi-source operation data; An offline training module is used to initialize a specific mode transformer fault warning expert model, wherein the specific mode transformer fault warning expert model is composed of a multi-head self-attention mechanism, an expert network and a residual connection; the specific mode transformer fault warning expert model is trained using a spatiotemporal fusion feature set of transformer multi-source operation data; a multi-modal transformer fault warning hybrid expert model is initialized, wherein the multi-modal transformer fault warning hybrid expert model is composed of a multi-modal transformer fault warning hybrid expert module, and the multi-modal transformer fault warning hybrid expert module is composed of routing, a multi-head attention mechanism, a hybrid expert network and a residual connection; the trained specific mode transformer fault warning expert model is integrated into the hybrid expert network layer of the multi-modal transformer fault warning hybrid expert module; the multi-modal transformer fault warning hybrid expert model is trained using a spatiotemporal fusion feature set of transformer multi-source operation data; The online fault warning module is used to load the trained multi-modal transformer fault warning hybrid expert model, take the real-time transformer multi-source operation data after data preprocessing and feature encoding as input, obtain the fine-grained branch information of the transformer fault warning subtask, and complete the fault warning.

9. A computing device comprising: Memory and processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the transformer fault early warning method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the transformer fault early warning method according to any one of claims 1 to 7.

Images