Transformer health state monitoring method and system based on unsupervised online learning

Abstract

The invention discloses a transformer health state monitoring method and system based on unsupervised online learning, and belongs to the technical field of power system monitoring, and the method comprises the steps: obtaining transformer monitoring type time sequence data and partial discharge map data; performing multi-modal feature extraction and fusion on the acquired transformer monitoring type time sequence data and partial discharge spectrum data by using a pre-established multi-modal model to obtain a current monitoring result type; according to the category of the current monitoring result, storing a normal result or an abnormal result of which the confidence exceeds a set threshold value in a data cache queue at a set time interval; and when the data cache queue reaches a set length, performing unsupervised online learning on a pre-established multi-modal model based on the current scene, replacing the original multi-modal model with the multi-modal model after the unsupervised online learning, and emptying the data cache queue. The method adapts to continuous change of field data distribution through unification of data features, cross-modal feature extraction and unsupervised updating.

Description

Technical Field

[0001] This invention belongs to the field of power system monitoring technology, specifically relating to a method and system for monitoring the health status of transformers based on unsupervised online learning. Background Technology

[0002] In the field of artificial intelligence and adaptive systems, unsupervised online learning has always been an important research direction. Unsupervised online learning enables machines to learn the inherent patterns and structures of data directly from continuously arriving real-time data streams, without relying on pre-labeled training datasets. This characteristic makes it particularly suitable for dynamic, real-world environments where data distributions may drift over time, laying the foundation for building continuously evolving intelligent agents.

[0003] Among various mechanisms for achieving unsupervised learning, Hebbian learning, as a classic theory inspired by neuroscience, provides a concise yet powerful theoretical foundation. Its core idea, "synchronous firing of neurons enhances connections," simulates the fundamental law of synaptic plasticity in biological neural networks. This mechanism provides a theoretical basis for constructing highly adaptive computational models, enabling the network to autonomously adjust its internal connection weights based on the correlation of input signals, thereby achieving pattern discovery and feature extraction.

[0004] To overcome the limitations of single data sources and achieve more comprehensive and robust environmental perception capabilities, multimodal fusion technology has been introduced and become a key component. This technology aims to collaboratively integrate data from different types of sensors or information sources (such as vision, hearing, touch, and text), forming a unified and deeper understanding of a scene or object through complementarity and verification. When unsupervised online learning frameworks process multimodal data streams, effectively fusing this heterogeneous information becomes a core issue. Information entropy technology plays a crucial role in this process. As a mathematical tool derived from information theory, entropy provides a standard for quantifying the uncertainty, randomness, or value of information. In adaptive learning systems, entropy values ​​can be used to assess the novelty of data, the confidence of model predictions, or the degree of disorder in the current state of the system, thus providing an intrinsic, computable guiding signal for the learning process. For example, a high entropy value may indicate the emergence of new patterns that the model has not learned before, suggesting that the system should allocate more attentional resources.

[0005] As a critical piece of equipment in the power system, the operating status of power transformers directly affects the safe operation of the entire power system. A fault in a transformer can cause not only severe grid outages and huge economic losses, but also potential threats to personal safety. Existing methods collect time-series data of various observed state variables of the transformer and train predictive models to identify the transformer health status represented by different time-series state variables. However, due to significant differences in the inherent properties of transformers and their operating environments, as well as variations in monitoring equipment, the distribution of recorded monitoring data varies greatly. Existing supervised training predictive models struggle to achieve accurate predictions for time-series state variables with such large data distribution differences. Summary of the Invention

[0006] The purpose of this invention is to address the problems in the prior art by providing a transformer health status monitoring method and system based on unsupervised online learning. This method and system can unify multimodal data features, extract cross-modal features, and perform unsupervised updates for time-series data and partial discharge spectrum data generated by transformer monitoring, thus adapting to the continuous changes in the distribution of field data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for monitoring the health status of transformers based on unsupervised online learning is provided, including: Acquire time-series data and partial discharge spectrum data for transformer monitoring; By using a pre-established multimodal model, multimodal features are extracted and fused from the acquired transformer monitoring time series data and partial discharge spectrum data to obtain the current monitoring result category; Based on the current monitoring result category, normal results or abnormal results with confidence levels exceeding a set threshold are saved in the data cache queue at set intervals. When the data cache queue reaches the set length, the pre-established multimodal model is subjected to unsupervised online learning based on the current scenario. The multimodal model after unsupervised online learning is used to replace the original multimodal model, and the data cache queue is cleared.

[0008] As a preferred embodiment, in the step of extracting and fusing multimodal features from the acquired transformer monitoring time-series data and partial discharge map data using a pre-established multimodal model to obtain the current monitoring result category, the multimodal model includes a time-series data processing branch and a map data processing branch. The time-series data processing branch adopts a combination of an attention-based deep learning model architecture Transformer and a Long Short-Term Memory (LSTM) network, while the map data processing branch adopts a convolutional neural network VGG16. Furthermore, a Heb layer is added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network. Feature extraction of time-series data for transformer monitoring is performed using a deep learning model architecture based on attention mechanism, Transformer and Long Short-Term Memory (LSTM), to obtain time-series data feature vectors. Feature extraction of partial discharge spectral data is performed using a convolutional neural network, VGG16, to obtain spectral data feature vectors. The time-series data feature vectors and spectral data feature vectors are then fused in a weighted manner and classified using a classifier to obtain the current monitoring result category.

[0009] As a preferred approach, the step of performing unsupervised online learning on the pre-established multimodal model based on the current scenario includes the following two stages: In the first stage, Heblin layers, added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, respectively traverse the temporal modality data and perform unsupervised fine-tuning of parameters based on the local plasticity mechanism. In the second stage, the Heblin layer parameters are frozen, and the unsupervised samples are used to obtain the predicted distribution of all categories through the multimodal model. The non-Heblin layer parameters are then fine-tuned using the information entropy loss function.

[0010] As a preferred embodiment, the Heb layer traverses both temporal modality data and graphical modality data, and performs unsupervised parameter fine-tuning based on the local plasticity mechanism. The unsupervised parameter fine-tuning depends on the response intensity of adjacent neurons in the following expression:

[0011] In the formula, the parameter and These are fixed constants, taking values ​​of 0.1 and 1 respectively; related variables These correspond to the output layer neurons, input layer neurons, inactive output signals, and connection weights between neurons in two layers, respectively. The connection weights are adaptively adjusted based on the response strength of neurons in the previous and next layers.

[0012] As a preferred approach, the step of performing unsupervised online learning on a pre-established multimodal model based on the current scenario utilizes the reduction of the information entropy loss function value to update the model parameters through the backpropagation algorithm, so that the probability distribution of all categories output by the model is biased towards a certain category. The information entropy loss function is established as follows: For a system with K categories or states, the probability distribution is as follows: ,in, Let represent the probability of the i-th category occurring, and satisfy the following formula:

[0013] Then, the information entropy of the corresponding distribution Calculated using the following expression:

[0014] In the unsupervised online learning phase of the non-Heb layer, the predicted probability distribution of the model is ordered by the information entropy loss function, and the parameters of the non-Heb layer are updated by the backpropagation algorithm to adapt the model to the current data distribution and complete the alignment with the Heb layer.

[0015] As a preferred approach, in the step of acquiring transformer monitoring time-series data and partial discharge spectrum data, a supervised training multimodal model is used to label historical transformer monitoring time-series data and historical partial discharge spectrum data. The label content is the classification result of the corresponding transformer monitoring time-series data and partial discharge spectrum data, and the corresponding numerical code is used.

[0016] As a preferred embodiment, in the pre-training phase of the supervised multimodal model, the non-Heb layer updates its parameters using a cross-entropy loss function combined with the backpropagation algorithm. The expression for the cross-entropy loss function is as follows:

[0017] In the formula, K represents the total number of categories. The sample belongs to the category The true probability, The model predicts that the sample belongs to a certain category. The probability is calculated using the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true probability distribution. The greater the difference, the higher the function value. When the function value is 0, the probability distribution predicted by the model is the same as the true probability distribution.

[0018] As a preferred approach, based on the current monitoring result category, the abnormal results, the confidence level of the abnormal results, and the data at the corresponding time are returned for manual secondary confirmation to reassess whether there are any abnormalities in the current transformer health status.

[0019] As a preferred embodiment, the step of storing normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals involves storing one frame of normal result or abnormal result data with confidence levels exceeding the set threshold every half hour.

[0020] As a preferred embodiment, the data cache queue is set to a length of 500, corresponding to a maximum storage of 500 sets of transformer health status monitoring data; The maximum number of normal samples is 300, and the maximum number of abnormal samples is 200. When the number of normal or abnormal samples exceeds the maximum, the oldest sample will be overwritten by the newest sample.

[0021] Secondly, a transformer health status monitoring system based on unsupervised online learning is provided, including: The data acquisition module is used to acquire time-series data and partial discharge spectrum data for transformer monitoring. The feature extraction and fusion module is used to extract and fuse multimodal features from the acquired transformer monitoring time series data and partial discharge spectrum data using a pre-established multimodal model to obtain the current monitoring result category. The monitoring result storage module is used to save normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals, according to the current monitoring result category. The unsupervised online learning module is used to perform unsupervised online learning on a pre-built multimodal model based on the current scenario when the data cache queue reaches a set length. The unsupervised online learned multimodal model replaces the original multimodal model, and the data cache queue is cleared.

[0022] As a preferred embodiment, the feature extraction and fusion module utilizes a pre-established multimodal model to perform multimodal feature extraction and fusion on the acquired transformer monitoring time-series data and partial discharge map data to obtain the current monitoring result category. When obtaining the current monitoring result category, the multimodal model includes a time-series data processing branch and a map data processing branch. The time-series data processing branch adopts a combination of the attention-based deep learning model architecture Transformer and the Long Short-Term Memory network LSTM, while the map data processing branch adopts the convolutional neural network VGG16. Furthermore, a Heb layer is added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the convolutional neural network VGG16. Feature extraction of time-series data for transformer monitoring is performed using a deep learning model architecture based on attention mechanism, Transformer and Long Short-Term Memory (LSTM), to obtain time-series data feature vectors. Feature extraction of partial discharge spectral data is performed using a convolutional neural network, VGG16, to obtain spectral data feature vectors. The time-series data feature vectors and spectral data feature vectors are then fused in a weighted manner and classified using a classifier to obtain the current monitoring result category.

[0023] As a preferred embodiment, the unsupervised online learning module for the model includes the following two stages when performing unsupervised online learning on a pre-established multimodal model based on the current scenario: In the first stage, Heblin layers, added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, respectively traverse the temporal modality data and perform unsupervised fine-tuning of parameters based on the local plasticity mechanism. In the second stage, the Heblin layer parameters are frozen, and the unsupervised samples are used to obtain the predicted distribution of all categories through the multimodal model. The non-Heblin layer parameters are then fine-tuned using the information entropy loss function.

[0024] As a preferred embodiment, the unsupervised online learning module of the model performs unsupervised parameter fine-tuning based on the response strengths of adjacent neurons in the first stage, according to the following expression:

[0025] In the formula, the parameter and These are fixed constants, taking values ​​of 0.1 and 1 respectively; related variables These correspond to the output layer neurons, input layer neurons, inactive output signals, and connection weights between neurons in two layers, respectively. The connection weights are adaptively adjusted based on the response strength of neurons in the previous and next layers.

[0026] As a preferred approach, the unsupervised online learning module of the model utilizes the reduction of the information entropy loss function value to update the model parameters through the backpropagation algorithm, thereby biasing the probability distribution of all categories output by the model towards a certain category. The information entropy loss function is established as follows: For a system with K categories or states, the probability distribution is as follows: ,in, Let represent the probability of the i-th category occurring, and satisfy the following formula:

[0027] Then, the information entropy of the corresponding distribution Calculated using the following expression:

[0028] In the unsupervised online learning phase of the non-Heb layer, the predicted probability distribution of the model is ordered by the information entropy loss function, and the parameters of the non-Heb layer are updated by the backpropagation algorithm to adapt the model to the current data distribution and complete the alignment with the Heb layer.

[0029] As a preferred embodiment, when acquiring transformer monitoring time-series data and partial discharge spectrum data, the data acquisition module uses a supervised-trained multimodal model to label historical transformer monitoring time-series data and historical partial discharge spectrum data. The label content is the classification result of the corresponding transformer monitoring time-series data and partial discharge spectrum data, and uses the corresponding numerical code.

[0030] As a preferred embodiment, when the data acquisition module labels historical transformer monitoring time-series data and historical partial discharge map data using a supervised-trained multimodal model, the non-Heb layer of the supervised-trained multimodal model updates parameters through a cross-entropy loss function combined with a backpropagation algorithm during the model pre-training phase. The expression for the cross-entropy loss function is as follows:

[0031] In the formula, K represents the total number of categories. The sample belongs to the category The true probability, The model predicts that the sample belongs to a certain category. The probability is calculated using the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true probability distribution. The greater the difference, the higher the function value. When the function value is 0, the probability distribution predicted by the model is the same as the true probability distribution.

[0032] As a preferred solution, the monitoring result storage module returns the abnormal results, the confidence level of the abnormal results, and the data at the corresponding time according to the current monitoring result category, and performs a second manual confirmation to reassess whether there is any abnormality in the current health status of the transformer.

[0033] As a preferred embodiment, the monitoring result storage module stores one frame of normal result data or abnormal result data with a confidence level exceeding a set threshold every half hour.

[0034] As a preferred embodiment, when the monitoring result storage module saves normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals, the set length of the data cache queue is 500, corresponding to a maximum storage of 500 sets of transformer health status monitoring data. The maximum number of normal samples is 300, and the maximum number of abnormal samples is 200. When the number of normal or abnormal samples exceeds the maximum, the oldest sample will be overwritten by the newest sample.

[0035] Thirdly, an electronic device is provided, including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the transformer health status monitoring method based on unsupervised online learning as described in the first aspect.

[0036] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the transformer health status monitoring method based on unsupervised online learning as described in the first aspect.

[0037] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects: For transformer monitoring time-series data and partial discharge spectrum data, this invention proposes an unsupervised online learning-based transformer health status monitoring method. This method utilizes a pre-established multimodal model for multimodal feature extraction and fusion, unifying multimodal data features to obtain the current monitoring result category. Based on this category, normal results or abnormal results with confidence levels exceeding a set threshold are stored in a data cache queue at set intervals. When the data cache queue reaches a set length, the pre-established multimodal model undergoes unsupervised online learning based on the current scenario. The unsupervised online-learned multimodal model replaces the original multimodal model, and the data cache queue is cleared. This enables end-to-end unsupervised online updates, allowing for fine-tuning of the transformer health status prediction model without human intervention. This adapts the multimodal model to the differences in data distribution across various scenarios, achieving accurate predictions for time-series state quantities with significant data distribution variations.

[0038] Furthermore, this invention proposes a multimodal feature extraction method based on Heb learning. This method combines Heb learning with the Transformer architecture for time-series data related to transformer monitoring, extracting feature vectors from the time-series data. It also combines Heb learning with the VGG16 convolutional neural network for partial discharge spectral data, extracting feature vectors from the spectral data. By designing the multimodal model architecture, this invention adds a Heb layer before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, achieving unsupervised multimodal feature extraction capabilities while retaining the global feature analysis capabilities of traditional deep learning architectures.

[0039] Furthermore, this invention proposes an unsupervised alignment method for non-Herb layers based on information entropy. The method uses the information entropy algorithm to complete the unsupervised update alignment of non-Herb layers to Heb layers, enabling unsupervised adaptation of non-Herb layers without human intervention.

[0040] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of the transformer health status monitoring method based on unsupervised online learning according to an embodiment of the present invention; Figure 2 A schematic diagram of the distribution of Heb layers and non-Heb layers in the multimodal model architecture of this invention embodiment; Figure 3 A schematic diagram illustrating the principle of a pre-established multimodal model in an embodiment of the present invention. Detailed Implementation

[0043] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail. Flowcharts are used in the embodiments of this application to illustrate the operations performed by the apparatus according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps may be processed in reverse order or simultaneously as needed. Furthermore, other operations may be added to these processes, or one or more steps may be removed from these processes.

[0044] Please see Figure 1 The prediction model needs to be able to fine-tune the model for time-series state quantities with different data distributions without human intervention. To this end, this invention proposes a transformer health status monitoring method based on unsupervised online learning, which mainly includes the following steps: S101. Obtain time-series data and partial discharge spectrum data for transformer monitoring; S102. Using a pre-established multimodal model, multimodal feature extraction and fusion are performed on the acquired transformer monitoring time series data and partial discharge spectrum data to obtain the current monitoring result category; S103. Based on the current monitoring result category, normal results or abnormal results with confidence levels exceeding the set threshold are saved in the data cache queue at set intervals. S104. When the data cache queue reaches the set length, perform unsupervised online learning on the pre-established multimodal model based on the current scenario, replace the original multimodal model with the multimodal model after unsupervised online learning, and clear the data cache queue.

[0045] In one possible implementation, the transformer monitoring time series data in step S101 is mainly obtained by field chromatography, grounding current and other acquisition devices, while the partial discharge spectrum data is mainly acquired by ultra-high frequency and high frequency sensors.

[0046] Furthermore, transformer fault types include, but are not limited to, high-temperature overheating, overheating defects, partial discharge, moisture-induced faults, floating discharge, and normal conditions.

[0047] In one possible implementation, this embodiment uses an initially supervised training multimodal model to obtain corresponding labels based on historical transformer monitoring time series data and historical partial discharge spectrum data through manual annotation. The label content is the classification result of the corresponding transformer monitoring time series data and partial discharge spectrum data, as shown in Table 1. This embodiment is divided into six categories, including one normal category and five abnormal categories, and uses corresponding numerical codes, with the numerical numbers ranging from 0 to 5.

[0048] Table 1

[0049] Please see Figure 3 In one possible implementation, the multimodal model described in step S102 of this embodiment includes a temporal data processing branch and a graph data processing branch. The temporal data processing branch uses a combination of an attention-based deep learning model architecture, Transformer, and a Long Short-Term Memory (LSTM) network. The graph data processing branch uses a convolutional neural network, VGG16. Furthermore, a Hebbe layer is added before both the attention-based deep learning model architecture, Transformer, and the first convolutional layer of the VGG16 convolutional neural network. The Hebbe layer added before the Transformer is a convolutional layer with 1 input and 1 output channel, a kernel size of 7, a stride of 1, and padding of 3. The Hebbe layer added before the first convolutional layer of the VGG16 is also a convolutional layer with 3 input channels, 64 output channels, a stride of 1, and padding of 1.

[0050] Feature extraction of time-series data for transformer monitoring is performed using a deep learning model architecture based on attention mechanism, Transformer and Long Short-Term Memory (LSTM), to obtain time-series data feature vectors. Feature extraction of partial discharge spectral data is performed using a convolutional neural network, VGG16, to obtain spectral data feature vectors. The time-series data feature vectors and spectral data feature vectors are then fused in a weighted manner and classified using a classifier to obtain the current monitoring result category.

[0051] The parameter adjustment mechanism of the Heb layer adopts an unsupervised learning strategy, and its update process depends only on the response strength of neurons in adjacent layers. The specific calculation expression is as follows:

[0052] Among them, parameters and This is a fixed constant, taking values ​​of 0.1 and 1 respectively. For example... Figure 2 The Hebbian learning principle is shown, with relevant variables. These represent the output layer neurons, input layer neurons, inactive output signals, and connection weights between neurons in two layers, respectively. Based on the above expressions, the network can adaptively adjust the connection weights according to the response strength of neurons in the preceding and following layers.

[0053] During the supervised pre-training phase, the non-Heb layer updates its parameters using a cross-entropy loss function combined with the backpropagation algorithm. The expression for the cross-entropy loss function is as follows:

[0054] Where K represents the total number of categories, The sample belongs to the category The true probability (usually one-hot encoded, i.e., the true class is 1 and the other classes are 0). The model predicts that the sample belongs to a certain category. The probability (usually the output of the Softmax function) is the probability of the model's predictions. The cross-entropy loss function measures the difference between the model's predicted probability distribution and the true probability distribution. The greater the difference, the higher the function value. When the function value is 0, it means that the model's predicted probability distribution is the same as the true probability distribution.

[0055] In the unsupervised online learning phase, the model parameters are updated via backpropagation by reducing the information entropy loss function value, so that the probability distribution of all categories in the model output is biased towards a certain category. The information entropy loss function is established as follows: For a system with K categories or states, the probability distribution is as follows: ,in, Let represent the probability of the i-th category occurring, and satisfy the following formula:

[0056] Then, the information entropy of the corresponding distribution Calculated using the following expression:

[0057] In the unsupervised online learning phase of the non-Heb layer, the prediction probability distribution of the model is made more ordered by the information entropy loss function, and the parameters of the non-Heb layer are updated by the backpropagation algorithm to adapt the model to the current data distribution and complete the alignment with the Heb layer.

[0058] According to the unsupervised online learning of the transformer health status monitoring model in this embodiment, a pre-trained model on historical data is first obtained. Then, the Heb layer completes the traversal of the data in the data cache and performs unsupervised fine-tuning. Next, the non-Heb layer completes the unsupervised fine-tuning of the data in the data cache and alignment with the Heb layer through the information entropy loss function. The entire process model adapts to the field monitoring data in an unsupervised manner, effectively improving the model's adaptability to field data.

[0059] In one possible implementation, step S104 involves unsupervised online learning of the pre-established multimodal model based on the current scenario, comprising the following two stages: In the first stage, Heblin layers, added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, respectively traverse the temporal modality data and perform unsupervised fine-tuning of parameters based on the local plasticity mechanism. In the second stage, the Heblin layer parameters are frozen, and the unsupervised samples are used to obtain the predicted distribution of all categories through the multimodal model. The non-Heblin layer parameters are then fine-tuned using the information entropy loss function.

[0060] In one possible implementation, this embodiment outputs the current monitoring result category through a multimodal model and performs the next operation based on whether the result is abnormal.

[0061] The multimodal model mentioned in step S102 is an initial model trained on an offline dataset. Its characteristic is that the initial model performs well on the offline dataset, but its performance deteriorates significantly when the distribution of real-time monitoring data differs greatly from that of the offline dataset. The subsequent operation steps are determined by judging the output category and a category confidence threshold. In this embodiment, the confidence threshold is set to 0.5; exceeding the threshold indicates the model's confidence in correctly classifying the monitoring data.

[0062] Furthermore, this embodiment also includes returning the abnormal results, the confidence level of the abnormal results, and the data at the corresponding time according to the current monitoring result category, for manual secondary confirmation, to re-evaluate whether there is any abnormality in the current transformer health status.

[0063] When the platform receives abnormal results and corresponding monitoring data, staff can quickly view the abnormality category, confidence level, and data at the corresponding time from the platform to further assess whether the current equipment is abnormal and what subsequent measures can be taken.

[0064] In one possible implementation, step S103 stores a frame of normal result data or abnormal result data with a confidence level exceeding a set threshold every half hour.

[0065] In one possible implementation, the data cache queue in this embodiment is set to a length of 500, corresponding to a maximum storage of 500 sets of transformer health status monitoring data; The maximum number of normal samples is 300, and the maximum number of abnormal samples is 200. When the number of normal or abnormal samples exceeds the maximum, the oldest sample will be overwritten by the newest sample.

[0066] In step S104 of this embodiment, when the length of the data cache queue reaches 500, the pre-established multimodal model will undergo unsupervised online learning based on the current scenario. The model will be updated unsupervised based on monitoring data without manual annotation, so that the model can adapt to the monitoring data of the current scenario. The updated model will replace the original model and the data cache queue will be cleared.

[0067] In terms of testing, this embodiment uses 100 samples of time-series and partial discharge spectrum data collected from monitoring equipment with significant differences as the test set. A total of 7820 samples of historical cumulative monitoring time-series data and partial discharge spectrum data are used as the training set. Both the training and test sets include the defect categories in Table 1. This embodiment conducts performance comparison experiments on the test set before and after online learning, as well as comparison experiments on the original model without added Heb layers before and after online learning. The experimental results are shown in Table 2. Table 2

[0068] As shown in Table 2, the accuracy of the original multimodal model without the addition of the Heb layer can be improved by 4%, from 89% to 92%, using the online learning framework provided by this invention. With the addition of the Heb layer, the accuracy of the multimodal model using the online learning framework can be improved by 11%, from 87% to 98%. These two sets of comparative experiments fully demonstrate that the online learning process and Heb layer structure of this invention can effectively improve the unsupervised online learning capability of the model in the transformer health status monitoring scenario.

[0069] Another embodiment of the present invention proposes a transformer health status monitoring system based on unsupervised online learning, comprising: The data acquisition module is used to acquire time-series data and partial discharge spectrum data for transformer monitoring. The feature extraction and fusion module is used to extract and fuse multimodal features from the acquired transformer monitoring time series data and partial discharge spectrum data using a pre-established multimodal model to obtain the current monitoring result category. The monitoring result storage module is used to save normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals, according to the current monitoring result category. The unsupervised online learning module is used to perform unsupervised online learning on a pre-built multimodal model based on the current scenario when the data cache queue reaches a set length. The unsupervised online learned multimodal model replaces the original multimodal model, and the data cache queue is cleared.

[0070] In one possible implementation, the feature extraction and fusion module uses a pre-established multimodal model to perform multimodal feature extraction and fusion on the acquired transformer monitoring time-series data and partial discharge map data to obtain the current monitoring result category. The multimodal model includes a time-series data processing branch and a map data processing branch. The time-series data processing branch adopts a combination of the attention-based deep learning model architecture Transformer and the Long Short-Term Memory network LSTM, and the map data processing branch adopts the convolutional neural network VGG16. Furthermore, a Heb layer is added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the convolutional neural network VGG16. Feature extraction of time-series data for transformer monitoring is performed using a deep learning model architecture based on attention mechanism, Transformer and Long Short-Term Memory (LSTM), to obtain time-series data feature vectors. Feature extraction of partial discharge spectral data is performed using a convolutional neural network, VGG16, to obtain spectral data feature vectors. The time-series data feature vectors and spectral data feature vectors are then fused in a weighted manner and classified using a classifier to obtain the current monitoring result category.

[0071] In one possible implementation, the unsupervised online learning module for the model includes the following two stages when performing unsupervised online learning on a pre-built multimodal model based on the current scenario: In the first stage, Heblin layers, added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, respectively traverse the temporal modality data and perform unsupervised fine-tuning of parameters based on the local plasticity mechanism. In the second stage, the Heblin layer parameters are frozen, and the unsupervised samples are used to obtain the predicted distribution of all categories through the multimodal model. The non-Heblin layer parameters are then fine-tuned using the information entropy loss function.

[0072] In one possible implementation, during the first phase of the unsupervised online learning module, the parameters are fine-tuned unsupervised based on the response strengths of neurons in adjacent layers, according to the following expression:

[0073] In the formula, the parameter and These are fixed constants, taking values ​​of 0.1 and 1 respectively; related variables These correspond to the output layer neurons, input layer neurons, inactive output signals, and connection weights between neurons in two layers, respectively. The connection weights are adaptively adjusted based on the response strength of neurons in the previous and next layers.

[0074] In one possible implementation, the unsupervised online learning module of the model updates the model parameters through the backpropagation algorithm by reducing the value of the information entropy loss function, so that the probability distribution of all categories output by the model is biased towards a certain category. The information entropy loss function is established as follows: For a system with K categories or states, the probability distribution is as follows: ,in, Let represent the probability of the i-th category occurring, and satisfy the following formula:

[0075] Then, the information entropy of the corresponding distribution Calculated using the following expression:

[0076] In the unsupervised online learning phase of the non-Heb layer, the predicted probability distribution of the model is ordered by the information entropy loss function, and the parameters of the non-Heb layer are updated by the backpropagation algorithm to adapt the model to the current data distribution and complete the alignment with the Heb layer.

[0077] In one possible implementation, when the data acquisition module acquires transformer monitoring time-series data and partial discharge spectrum data, it uses a supervised-trained multimodal model to label historical transformer monitoring time-series data and historical partial discharge spectrum data. The label content is the classification result of the corresponding transformer monitoring time-series data and partial discharge spectrum data, and uses the corresponding numerical code.

[0078] In one possible implementation, when the data acquisition module labels historical transformer monitoring time-series data and historical partial discharge map data using a supervised-trained multimodal model, the non-Heb layer of the supervised-trained multimodal model updates its parameters during the model pre-training phase using a cross-entropy loss function combined with a backpropagation algorithm. The expression for the cross-entropy loss function is as follows:

[0079] In the formula, K represents the total number of categories. The sample belongs to the category The true probability, The model predicts that the sample belongs to a certain category. The probability is calculated using the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true probability distribution. The greater the difference, the higher the function value. When the function value is 0, the probability distribution predicted by the model is the same as the true probability distribution.

[0080] In one possible implementation, the monitoring result storage module returns the abnormal result, the confidence level of the abnormal result, and the data at the corresponding time according to the current monitoring result category, and performs a second manual confirmation to reassess whether there is an abnormality in the current transformer health status.

[0081] In one possible implementation, the monitoring result storage module stores one frame of normal result or abnormal result data with a confidence level exceeding a set threshold every half hour.

[0082] In one possible implementation, when the monitoring result storage module saves normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals, the set length of the data cache queue is 500, corresponding to a maximum of 500 sets of transformer health status monitoring data to be stored. The maximum number of normal samples is 300, and the maximum number of abnormal samples is 200. When the number of normal or abnormal samples exceeds the maximum, the oldest sample will be overwritten by the newest sample.

[0083] Another embodiment of the present invention also proposes an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the transformer health status monitoring method based on unsupervised online learning.

[0084] Another embodiment of the present invention provides a computer-readable storage medium storing at least one instruction that, when executed by a processor, implements the transformer health status monitoring method based on unsupervised online learning.

[0085] The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals. For ease of explanation, the above content only shows the parts related to the embodiments of the present invention; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. This computer-readable storage medium is non-transitory and can be stored in storage devices formed by various electronic devices, enabling the execution process described in the method of the embodiments of the present invention.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for monitoring the health status of transformers based on unsupervised online learning, characterized in that, include: Acquire time-series data and partial discharge spectrum data for transformer monitoring; By using a pre-established multimodal model, multimodal features are extracted and fused from the acquired transformer monitoring time series data and partial discharge spectrum data to obtain the current monitoring result category; Based on the current monitoring result category, normal results or abnormal results with confidence levels exceeding a set threshold are saved in the data cache queue at set intervals. When the data cache queue reaches the set length, the pre-established multimodal model is subjected to unsupervised online learning based on the current scenario. The multimodal model after unsupervised online learning is used to replace the original multimodal model, and the data cache queue is cleared.

2. The transformer health status monitoring method based on unsupervised online learning according to claim 1, characterized in that, In the step of extracting and fusing multimodal features from the acquired transformer monitoring time-series data and partial discharge map data using a pre-established multimodal model to obtain the current monitoring result category, the multimodal model includes a time-series data processing branch and a map data processing branch. The time-series data processing branch adopts a combination of the attention-based deep learning model architecture Transformer and the Long Short-Term Memory network LSTM, while the map data processing branch adopts the convolutional neural network VGG16. Furthermore, a Heb layer is added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the convolutional neural network VGG16. Feature extraction of time-series data for transformer monitoring is performed using a deep learning model architecture based on attention mechanism, Transformer and Long Short-Term Memory (LSTM), to obtain time-series data feature vectors. Feature extraction of partial discharge spectral data is performed using a convolutional neural network, VGG16, to obtain spectral data feature vectors. The time-series data feature vectors and spectral data feature vectors are then fused in a weighted manner and classified using a classifier to obtain the current monitoring result category.

3. The transformer health status monitoring method based on unsupervised online learning according to claim 2, characterized in that, The steps for performing unsupervised online learning of the pre-established multimodal model based on the current scenario include the following two stages: In the first stage, Heblin layers, added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, respectively traverse the temporal modality data and perform unsupervised fine-tuning of parameters based on the local plasticity mechanism. In the second stage, the Heblin layer parameters are frozen, and the unsupervised samples are used to obtain the predicted distribution of all categories through the multimodal model. The non-Heblin layer parameters are then fine-tuned using the information entropy loss function.

4. The transformer health status monitoring method based on unsupervised online learning according to claim 3, characterized in that, The Heb layer traverses both temporal modal data and graphical modal data, and performs unsupervised parameter fine-tuning based on the local plasticity mechanism. The unsupervised parameter fine-tuning depends on the response intensity of adjacent neurons in the following expression: In the formula, the parameters and These are fixed constants, taking values ​​of 0.1 and 1 respectively; Related variables These correspond to the output layer neurons, input layer neurons, inactive output signals, and connection weights between neurons in two layers, respectively. The connection weights are adaptively adjusted based on the response strength of neurons in the previous and next layers.

5. The transformer health status monitoring method based on unsupervised online learning according to claim 4, characterized in that, The step of performing unsupervised online learning on the pre-established multimodal model based on the current scenario involves updating the model parameters through backpropagation by reducing the value of the information entropy loss function, so that the probability distribution of all categories output by the model is biased towards a certain category. The information entropy loss function is established as follows: For a system with K categories or states, the probability distribution is as follows: ,in, Let represent the probability of the i-th category occurring, and satisfy the following formula: Then, the information entropy of the corresponding distribution Calculated using the following expression: In the unsupervised online learning phase of the non-Heb layer, the predicted probability distribution of the model is ordered by the information entropy loss function, and the parameters of the non-Heb layer are updated by the backpropagation algorithm to adapt the model to the current data distribution and complete the alignment with the Heb layer.

6. The transformer health status monitoring method based on unsupervised online learning according to claim 1, characterized in that, In the step of acquiring transformer monitoring time-series data and partial discharge map data, a supervised training multimodal model is used to label historical transformer monitoring time-series data and historical partial discharge map data. The label content is the classification result of the corresponding transformer monitoring time-series data and partial discharge map data, and the corresponding numerical code is used.

7. The transformer health status monitoring method based on unsupervised online learning according to claim 6, characterized in that, In the pre-training phase of the supervised multimodal model, the non-Heb layer updates its parameters using a cross-entropy loss function combined with the backpropagation algorithm. The expression for the cross-entropy loss function is as follows: In the formula, K represents the total number of categories. The sample belongs to the category The true probability, The model predicts that the sample belongs to a certain category. The probability is calculated using the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true probability distribution. The greater the difference, the higher the function value. When the function value is 0, the probability distribution predicted by the model is the same as the true probability distribution.

8. The transformer health status monitoring method based on unsupervised online learning according to claim 1, characterized in that, Based on the current monitoring result category, the abnormal results, the confidence level of the abnormal results, and the data at the corresponding time are returned for manual secondary confirmation to reassess whether there are any abnormalities in the current transformer health status.

9. The transformer health status monitoring method based on unsupervised online learning according to claim 1, characterized in that, The step of storing normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals involves storing one frame of normal results or abnormal results with confidence levels exceeding a set threshold every half hour.

10. The transformer health status monitoring method based on unsupervised online learning according to claim 1, characterized in that, The data cache queue is set to a length of 500, which corresponds to storing a maximum of 500 sets of transformer health status monitoring data. The maximum number of normal samples is 300, and the maximum number of abnormal samples is 200. When the number of normal or abnormal samples exceeds the maximum, the oldest sample will be overwritten by the newest sample.

11. A transformer health status monitoring system based on unsupervised online learning, characterized in that, include: The data acquisition module is used to acquire time-series data and partial discharge spectrum data for transformer monitoring. The feature extraction and fusion module is used to extract and fuse multimodal features from the acquired transformer monitoring time series data and partial discharge spectrum data using a pre-established multimodal model to obtain the current monitoring result category. The monitoring result storage module is used to save normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals, according to the current monitoring result category. The unsupervised online learning module is used to perform unsupervised online learning on a pre-built multimodal model based on the current scenario when the data cache queue reaches a set length. The module then replaces the original multimodal model with the unsupervised online learning model and clears the data cache queue.

12. The transformer health status monitoring system based on unsupervised online learning according to claim 11, characterized in that, The feature extraction and fusion module uses a pre-established multimodal model to extract and fuse multimodal features from the acquired transformer monitoring time-series data and partial discharge map data to obtain the current monitoring result category. The multimodal model includes a time-series data processing branch and a map data processing branch. The time-series data processing branch uses a combination of the attention-based deep learning model architecture Transformer and the Long Short-Term Memory network LSTM, while the map data processing branch uses a convolutional neural network VGG16. Furthermore, a Heb layer is added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the convolutional neural network VGG16. Feature extraction of time-series data for transformer monitoring is performed using a deep learning model architecture based on attention mechanism, Transformer and Long Short-Term Memory (LSTM), to obtain time-series data feature vectors. Feature extraction of partial discharge spectral data is performed using a convolutional neural network, VGG16, to obtain spectral data feature vectors. The time-series data feature vectors and spectral data feature vectors are then fused in a weighted manner and classified using a classifier to obtain the current monitoring result category.

13. The transformer health status monitoring system based on unsupervised online learning according to claim 12, characterized in that, The unsupervised online learning module for the model includes the following two stages when performing unsupervised online learning on a pre-established multimodal model based on the current scenario: In the first stage, Heblin layers, added before the attention-based deep learning model architecture Transformer and before the first convolutional layer of the VGG16 convolutional neural network, respectively traverse the temporal modality data and perform unsupervised fine-tuning of parameters based on the local plasticity mechanism. In the second stage, the Heblin layer parameters are frozen, and the unsupervised samples are used to obtain the predicted distribution of all categories through the multimodal model. The non-Heblin layer parameters are then fine-tuned using the information entropy loss function.

14. The transformer health status monitoring system based on unsupervised online learning according to claim 13, characterized in that, During the first stage of the unsupervised online learning module, the parameters are fine-tuned in an unsupervised manner based on the response strengths of neurons in adjacent layers, according to the following expression: In the formula, the parameters and These are fixed constants, taking values ​​of 0.1 and 1 respectively; Related variables These correspond to the output layer neurons, input layer neurons, inactive output signals, and connection weights between neurons in two layers, respectively. The connection weights are adaptively adjusted based on the response strength of neurons in the previous and next layers.

15. The transformer health status monitoring system based on unsupervised online learning according to claim 14, characterized in that, The unsupervised online learning module of the model uses the reduction of the information entropy loss function value to update the model parameters through the backpropagation algorithm, so that the probability distribution of all categories output by the model is biased towards a certain category. The information entropy loss function is established as follows: For a system with K categories or states, the probability distribution is as follows: ,in, Let represent the probability of the i-th category occurring, and satisfy the following formula: Then, the information entropy of the corresponding distribution Calculated using the following expression: In the unsupervised online learning phase of the non-Heb layer, the predicted probability distribution of the model is ordered by the information entropy loss function, and the parameters of the non-Heb layer are updated by the backpropagation algorithm to adapt the model to the current data distribution and complete the alignment with the Heb layer.

16. The transformer health status monitoring system based on unsupervised online learning according to claim 11, characterized in that, When acquiring transformer monitoring time-series data and partial discharge spectrum data, the data acquisition module uses a supervised-trained multimodal model to label historical transformer monitoring time-series data and historical partial discharge spectrum data. The label content is the classification result of the corresponding transformer monitoring time-series data and partial discharge spectrum data, and uses the corresponding numerical code.

17. The transformer health status monitoring system based on unsupervised online learning according to claim 16, characterized in that, When the data acquisition module labels historical transformer monitoring time-series data and historical partial discharge map data using a supervised-trained multimodal model, the non-Heb layer of the supervised-trained multimodal model updates parameters through a cross-entropy loss function combined with a backpropagation algorithm during the model pre-training phase. The expression for the cross-entropy loss function is as follows: In the formula, K represents the total number of categories. The sample belongs to the category The true probability, The model predicts that the sample belongs to a certain category. The probability is calculated using the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true probability distribution. The greater the difference, the higher the function value. When the function value is 0, the probability distribution predicted by the model is the same as the true probability distribution.

18. The transformer health status monitoring system based on unsupervised online learning according to claim 11, characterized in that, The monitoring result storage module returns the abnormal results, the confidence level of the abnormal results, and the data at the corresponding time according to the current monitoring result category, and performs a second manual confirmation to reassess whether there is any abnormality in the current health status of the transformer.

19. The transformer health status monitoring system based on unsupervised online learning according to claim 11, characterized in that, The monitoring result storage module stores one frame of normal result or abnormal result data with a confidence level exceeding a set threshold every half hour.

20. The transformer health status monitoring system based on unsupervised online learning according to claim 11, characterized in that, When the monitoring result storage module saves normal results or abnormal results with confidence levels exceeding a set threshold in the data cache queue at set intervals, the set length of the data cache queue is 500, corresponding to a maximum of 500 sets of transformer health status monitoring data to be stored. The maximum number of normal samples is 300, and the maximum number of abnormal samples is 200. When the number of normal or abnormal samples exceeds the maximum, the oldest sample will be overwritten by the newest sample.

21. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the transformer health status monitoring method based on unsupervised online learning as described in any one of claims 1 to 10.

22. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the transformer health status monitoring method based on unsupervised online learning as described in any one of claims 1 to 10.

Images