Equipment state anomaly detection method and system under unsupervised learning framework

Abstract

The invention belongs to the technical field of transformer state detection, and particularly relates to an equipment state anomaly detection method and system under an unsupervised learning framework. Comprising the following steps: inputting vibration data into an anomaly detection model, and performing feature extraction by using an encoder in a time sequence feature enhancement module to obtain a first potential feature representation; inputting the first potential feature representation into an attention-guided potential diffusion module, and learning to obtain a second potential feature representation through forward noise addition and reverse denoising in combination with multi-scale feature extraction and attention optimization expression; inputting the second potential feature representation into a decoder of a time sequence feature enhancement module to obtain reconstructed data; and calculating a reconstruction error between the vibration data and the reconstruction data to realize online state detection of the transformer. Model training can be completed without depending on abnormal samples, meanwhile, the model reconstruction precision is improved, and the abnormal condition of the transformer is efficiently judged.

Description

Technical Field

[0001] This invention belongs to the field of transformer condition detection technology, and particularly relates to a method and system for detecting equipment condition anomalies under an unsupervised learning framework. Background Technology

[0002] As the core equipment for power conversion and transmission within the power system, the stable operation of power transformers is a crucial foundation for ensuring the safe and reliable operation of the power grid. However, due to the long-term effects of harsh operating environments, coupled with continuous load operation, transformers are highly susceptible to various anomalies. If these anomalies are not detected and addressed in a timely manner, they can not only affect the safe and stable operation of the power grid but also potentially lead to significant personal injury and property damage. Therefore, to effectively improve the safety and reliability of power grid operation and reduce the maintenance costs throughout the transformer's lifecycle, research on anomaly detection methods for transformers, focusing on the easily manifested anomalies during operation, has extremely important practical significance and application value.

[0003] Currently, in the field of transformer condition anomaly detection, commonly used anomaly detection methods mainly include condition scoring methods, data analysis methods, and intelligent detection methods. Among them: (1) The status scoring method uses pre-defined status scoring standards to score the transformer's operating status from different perspectives, and uses this as a basis to determine whether the equipment is normal. However, this method has the problem that the threshold setting is too absolute, so it cannot flexibly cope with the impact of changes in the field environment on the monitoring data.

[0004] (2) Data analysis method uses data mining to extract valuable information from transformer monitoring data and judges the equipment status by analyzing the statistical characteristics, trend changes or correlation patterns of the data. However, this method is limited by the generalization ability of the mining algorithm. When faced with nonlinear and non-stationary monitoring data under complex transformer operating conditions, it is easy to have insufficient feature extraction and insufficient ability to identify early weak anomalies. Therefore, its application effect is limited in actual complex scenarios.

[0005] (3) Intelligent detection method uses deep learning and other technologies to automatically learn deep features in monitoring data through the model. It has stronger nonlinear fitting and adaptive capabilities, and is therefore more suitable for application in real engineering scenarios. It has significant research value and practical application significance.

[0006] Currently, transformer condition anomaly detection based on intelligent detection methods has received widespread attention and some progress has been made. However, many traditional condition anomaly detection techniques based on intelligent detection methods still have the following limitations: 1) Traditional anomaly detection methods generally rely on sufficient anomaly samples for model training. However, transformers in real-world scenarios are mainly in stable operation, and abnormal states are low-probability and sporadic, resulting in a scarcity of high-quality anomaly samples. This severely limits the effective learning of models from anomaly samples and restricts their deployment in engineering practice.

[0007] 2) Traditional anomaly detection methods have difficulty effectively capturing long-term time-series dependencies and multi-scale features when processing time-series signals such as transformer vibration. Furthermore, they are prone to computational redundancy when the feature dimension is high, resulting in insufficient sensitivity to identify weak anomalies under complex operating conditions, which in turn affects the accuracy and robustness of anomaly detection. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a method and system for detecting equipment state anomalies under an unsupervised learning framework. The anomaly detection model constructed based on temporal feature enhancement and attention-guided potential diffusion can complete model training without relying on abnormal samples. The temporal feature enhancement module compresses feature dimensions while retaining key information, and the attention-guided potential diffusion module accurately models the distribution of normal samples, improving the model reconstruction accuracy. This enables efficient identification of transformer anomalies and adapts to actual engineering needs.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for detecting equipment state anomalies within an unsupervised learning framework.

[0010] An equipment state anomaly detection method under an unsupervised learning framework includes the following steps: Obtain vibration data of the transformer under normal operating conditions; Vibration data is input into the anomaly detection model, and the encoder in the temporal feature enhancement module is used to extract features to obtain the first latent feature representation. The first latent feature representation is input into the attention-guided latent diffusion module. Through forward noise addition and backward noise removal, and by combining multi-scale feature extraction and attention-optimized expression, the feature distribution of vibration data under normal operating conditions is learned to obtain the second latent feature representation. The second latent feature representation is input into the decoder of the temporal feature enhancement module to obtain the reconstructed data; Calculate the reconstruction error between vibration data and reconstructed data, and train the anomaly detection model accordingly; The online vibration data of the transformer is input into the trained anomaly detection model to achieve online condition detection of the transformer.

[0011] The second aspect of this invention provides an equipment status anomaly detection system under an unsupervised learning framework.

[0012] An equipment status anomaly detection system based on an unsupervised learning framework includes: The data acquisition module is configured to acquire vibration data of the transformer under normal operating conditions. The first latent feature representation extraction module is configured to: input vibration data into the anomaly detection model, and use the encoder in the temporal feature enhancement module to extract features to obtain the first latent feature representation; The second latent feature representation estimation module is configured to: input the first latent feature representation into the attention-guided latent diffusion module, and learn the feature distribution of vibration data under normal operating conditions by forward noise addition and backward noise removal, and by combining multi-scale feature extraction and attention-optimized expression, to obtain the second latent feature representation; The data reconstruction module is configured to input the second latent feature representation into the decoder of the temporal feature enhancement module to obtain reconstructed data; The model training module is configured to: calculate the reconstruction error between vibration data and reconstructed data, and train the anomaly detection model; The online detection module is configured to input the online vibration data of the transformer into the trained anomaly detection model to achieve online condition detection of the transformer. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the equipment state anomaly detection method under the unsupervised learning framework described in the first aspect of the present invention.

[0013] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the equipment state anomaly detection method under the unsupervised learning framework as described in the first aspect of the present invention.

[0014] The above one or more technical solutions have the following beneficial effects: This invention provides a method and system for detecting equipment status anomalies under an unsupervised learning framework. The anomaly detection model constructed based on temporal feature enhancement and attention-guided potential diffusion can complete model training without relying on abnormal samples. The temporal feature enhancement module compresses feature dimensions while retaining key information, and the attention-guided potential diffusion module accurately models the distribution of normal samples, improving the model reconstruction accuracy. This enables efficient identification of abnormal conditions of transformers and adapts to actual engineering needs.

[0015] The temporal feature enhancement module proposed in this invention for feature dimension compression and accurate signal reconstruction effectively captures the long temporal feature expression of the original temporal signal by using dilated causal convolution and compresses the high-dimensional temporal signal into a low-dimensional latent space, which greatly reduces the input dimension of the subsequent diffusion model, directly reduces the number of parameters and computational overhead, and achieves the preservation of key features while compressing the signal dimension.

[0016] The attention-guided latent diffusion module proposed in this invention for accurately modeling the feature distribution of normal samples optimizes computational efficiency by leveraging the advantages of low-dimensional space. Through multi-scale feature extraction and attention-optimized feature representation, it effectively learns the distribution of normal samples and achieves high-quality generation. This enables the model to be effectively trained and efficiently identify anomalies even in scenarios without abnormal samples, significantly improving detection robustness and accuracy.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0020] Figure 2 This is a schematic diagram of the measurement point arrangement of 12 sensors on the transformer in Embodiment 1 of the present invention.

[0021] Figure 3 This is a network structure diagram of an anomaly detection model based on temporal feature enhancement and attention-guided potential diffusion, according to Embodiment 1 of the present invention.

[0022] Figure 4 This is a structural diagram of the timing feature enhancement module in Embodiment 1 of the present invention.

[0023] Figure 5 This is a schematic diagram of the timing residual block in Embodiment 1 of the present invention.

[0024] Figure 6 This is a structural diagram of the attention-guided potential diffusion module according to Embodiment 1 of the present invention.

[0025] Figure 7 This is a diagram of the multi-scale UNet network structure for attention optimization in Embodiment 1 of the present invention.

[0026] Figure 8 This is a structural diagram of the multi-scale attention block according to Embodiment 1 of the present invention. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0029] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0030] Example 1 This embodiment discloses an equipment state anomaly detection method under an unsupervised learning framework. The constructed temporal feature enhancement module for feature dimension compression and accurate signal reconstruction retains key features while compressing signal dimensions, providing high-quality feature support for subsequent anomaly detection. The constructed attention-guided latent diffusion module for accurate modeling of sample feature distribution enables efficient training and anomaly discrimination of the model in scenarios without abnormal samples, significantly improving detection robustness and accuracy.

[0031] like Figure 1 As shown, the equipment state anomaly detection method under the unsupervised learning framework includes the following steps: Obtain vibration data of the transformer under normal operating conditions; Vibration data is input into the anomaly detection model, and the encoder in the temporal feature enhancement module is used to extract features to obtain the first latent feature representation. The first latent feature representation is input into the attention-guided latent diffusion module. Through forward noise addition and backward noise removal, and by combining multi-scale feature extraction and attention-optimized expression, the feature distribution of vibration data under normal operating conditions is learned to obtain the second latent feature representation. The second latent feature representation is input into the decoder of the temporal feature enhancement module to obtain the reconstructed data; Calculate the reconstruction error between vibration data and reconstructed data, and train the anomaly detection model accordingly; The online vibration data of the transformer is input into the trained anomaly detection model to achieve online condition detection of the transformer.

[0032] Overall, the embodiments of the present invention first use multiple accelerometers to collect vibration data at different measuring points on the surface of the transformer under normal operating conditions, then establish an anomaly detection model based on time-series feature enhancement and attention-guided potential diffusion, and then use the collected vibration data to train the model, ultimately realizing online anomaly detection of the transformer.

[0033] The equipment state anomaly detection method under the unsupervised learning framework proposed in this embodiment will be explained in detail below. Its basic steps are as follows: Figure 1 As shown, it mainly includes: (1) Data acquisition: Multiple accelerometers are placed at different measuring points on the surface of the transformer tank to synchronously collect vibration data generated during normal operation of the transformer; (2) Anomaly detection model construction: Based on the unsupervised learning framework, an anomaly detection model based on temporal feature enhancement and attention-guided potential diffusion is established, so that the model can be trained using only normal samples, avoiding dependence on abnormal samples, and is more in line with real engineering scenarios. (3) Anomaly detection model training: For the anomaly detection model, a two-stage model training method is adopted. In the first stage, only the temporal feature enhancement module in the anomaly detection model is pre-trained, and in the second stage, only the attention-guided potential diffusion module in the anomaly detection model is trained. (4) Online anomaly detection: The collected vibration data of the transformer in operation is input into the trained anomaly detection model. The model will automatically perform anomaly detection on the input data and complete the online anomaly detection of the transformer.

[0034] More specifically, the implementation steps are as follows: (1) Data collection.

[0035] Twelve accelerometers were used to synchronously collect vibration data of the transformer at different measuring points during normal operation at a sampling frequency of 10 kHz. . This represents a set of vibration data collected. This refers to the number of channels, which is 12 in this case. (This is the sequence length, set to 1024 here), a total of 1000 normal samples were collected. The schematic diagram of the measurement point arrangement of the 12 sensors on the transformer is shown below. Figure 2 As shown.

[0036] (2) Anomaly detection model construction.

[0037] This invention proposes a transformer state anomaly detection model based on temporal feature enhancement and attention-guided latent diffusion. This model is an unsupervised network structure, mainly composed of a temporal feature enhancement module and an attention-guided latent diffusion module. Wherein: The temporal feature enhancement module captures long temporal dependencies and dynamic features in vibration data through dilated causal convolution, and achieves dimensional compression and accurate reconstruction of high-dimensional vibration data through encoding and decoding processes; The attention-guided latent diffusion module learns the feature distribution of normal samples in a low-dimensional latent space, and combines multi-scale feature extraction with attention-optimized feature representation to accurately model the sample distribution and achieve high-quality generation.

[0038] The final model relies on the difference between the reconstructed signal and the original vibration data to detect anomalies, effectively adapting to scenarios without anomaly samples. The overall model structure is as follows: Figure 3 As shown.

[0039] (2.1) Temporal feature enhancement module.

[0040] The temporal feature enhancement module is a network model specifically designed for processing temporal signals, built upon a temporal convolutional network. It consists of an encoder and a decoder. Its core idea is to utilize the encoder to effectively capture local features and long-range dependencies in the input temporal vibration data, mapping the high-dimensional temporal data to a low-dimensional latent space. The decoder then reconstructs the original input from the latent representation generated by the attention-guided latent diffusion module. The module's network structure diagram is shown below. Figure 4 As shown.

[0041] The main function of the encoder is to extract features and compress dimensions from the input time-series vibration data. This process can be represented as: (1) in, This indicates the encoder's mapping operation; This represents the latent features of the output.

[0042] The encoder consists of multiple stacked temporal residual blocks, each containing two layers of dilated causal convolutions, such as... Figure 5 As shown. Dilated causal convolution comprises two aspects: dilated convolution and causal convolution. Dilated convolution can be used to expand the receptive field when the dilation rate is... At that time, the effective receptive field size of this layer for: (2) in, This represents the kernel size.

[0043] Causal convolution strictly guarantees temporal causality, meaning the model at time step... The output depends only on the time in the input sequence. It incorporates information from previous moments without being affected by information from future moments. Specifically, for a one-dimensional time series... and convolution kernel Causal convolution at time Output for: (3) Among them, when When (i.e., beyond the start range of the sequence), It is then defined as 0 (achieved through zero padding). To avoid tail redundancy caused by padding, excess length is pruned after each convolutional layer to ensure that the output sequence is the same length as the input.

[0044] Combining dilated convolution and causal convolution, the dilated causal convolution at time... The output can be represented as: (4) In this embodiment, the encoder consists of three layers of temporal residual blocks, with dilation rates of 1, 2, and 4 for each layer, respectively. This reduces computational cost while capturing long-term temporal feature dependencies. Finally, average pooling is used to further compress the sequence length, yielding the latent feature representation. .

[0045] The main function of the decoder is to attempt to guide the latent feature representations generated by the latent diffusion module from the attention. The process of reconstructing the original signal can be represented as follows: (5) in, For the decoder mapping; For reconstructing the signal.

[0046] The decoder employs a nearly symmetrical structure to the encoder, first restoring the feature length to its original value through upsampling. Then, the temporal details are reconstructed using the inverse dilation rate of the encoder (i.e., 4, 2, 1), and finally, the number of channels is mapped back to the original input dimension through a 1×1 convolution. The signal reconstruction is completed.

[0047] After obtaining the reconstructed signal Then, the original signal can be calculated. With reconstructed signal The reconstruction error between the two is used as the outlier score of the data. Defined as: (6) in, Describe the square of the Euclidean norm (i.e. (Loss), used to quantify the original samples With reconstructed samples The element-wise difference. This score reflects the degree of deviation between the test sample and the normal pattern learned by the model.

[0048] This invention is based on " "Criteria for determining the anomaly score threshold, typically" It can be set to 3, which is the mean of the reconstruction error using normal training samples. Add 3 times the standard deviation As an abnormal score threshold, if the abnormal score of the sample to be tested exceeds the threshold, it is judged as abnormal; otherwise, it is normal.

[0049] (2.2) Attention-guided potential diffusion module.

[0050] The attention-guided latent diffusion module aims to perform refined distribution modeling of the latent features output by the encoder in the temporal feature enhancement module. By introducing noise perturbation and denoising learning mechanisms, and combining multi-scale feature extraction and attention-optimized representation, it effectively learns the feature distribution of normal samples. The model primarily models the data distribution through two complementary processes: forward noise addition and backward denoising. Figure 6 As shown.

[0051] The forward noise addition process starts from the original latent features. In the beginning, Gaussian noise is gradually added within each time step to ultimately generate noisy features. The noise addition result at any step. It can be calculated in one step using the following formula: (7) in, Indicates the cumulative coefficient; Indicates the number of noise-adding steps; Indicates standard Gaussian noise. Represents a standard Gaussian distribution. It is the zero vector. It is an identity matrix.

[0052] The reverse denoising process aims to learn the inverse of forward denoising through the model, that is, from noisy features... Starting with a noise prediction network Obtain the noise prediction results And gradually derived .feature It can be done It is calculated by the following formula: (8) The input to the noise prediction network is noisy features. and time step Output relative to the original noise The estimate; This is the random noise coefficient, which controls the randomness of the reverse process. Here it is set to 0, which means that deterministic sampling is used to avoid randomness and achieve accurate generation of the original latent features. Is with Irrelevant standard Gaussian noise.

[0053] Furthermore, the reverse denoising process can accelerate the sampling process by using a skip-step update method, the core of which is to directly update the subsequence. The above derivation does not require iterating through all the... Step. Among them, yes medium length is The increasing subsequence. At this point, formula (8) can be slightly modified to obtain the skip-step update formula (9): (9) in The first step in the skip sequence One time step; This refers to random noise during the skipping process; The step noise figure is derived from the parameters. Control is defined as: (10) when At that time, there was no random noise, and it was a deterministic skip sampling; while At this time, random step sampling is performed. To obtain accurate generation results, it is set to 0 here.

[0054] For noise prediction networks This invention combines the UNet structure to propose an attention-optimized multi-scale UNet network, the overall network structure diagram of which is shown below. Figure 7 As shown. This network features noisy characteristics. and time step As input, the output is sensitive to noise. The prediction, that is The network also employs a symmetrical encoder-decoder structure: The encoder part of the network captures local patterns and global dependencies of the input latent features through multi-scale attention blocks, and gradually compresses the feature dimension by downsampling; The decoder part of the network gradually restores the sequence dimensions through upsampling. Each upsampling step first performs linear interpolation on the high-dimensional features, and then adjusts the number of channels through a 1×1 convolution. Then, it is fused with the same-level features from the encoder, and the fused features are input into a multi-scale attention block for further feature extraction, finally outputting the prediction result for the noise.

[0055] To selectively enhance key information and suppress redundant information in encoder features during fusion, a gated attention fusion method is proposed, which solves the noise interference problem that may be caused by indiscriminate feature splicing in traditional skip connections. The specific process can be represented as follows: (11) in, These are the characteristics after fusion; Indicates a splicing operation; These are the same-level features output after multi-scale attention block feature extraction in the encoder; It is element-wise multiplication; This represents the Sigmoid activation function; This represents a 1×1 convolution operation; These are features obtained after upsampling.

[0056] For multi-scale attention blocks, the structure diagram is as follows: Figure 8 As shown. To reduce the number of model parameters, the multi-scale attention block adopts a lightweight structural design, mainly composed of depthwise separable convolutions, efficient channel attention modules, etc.

[0057] exist Figure 8 In this context, the multi-scale attention block mainly consists of two structurally identical sub-modules, which, for a given input and time step... The multi-scale attention block is processed through feature extraction from these two sub-modules to obtain the output. Specifically: Time step The scale parameters are obtained by MLP transformation and projection. and offset parameters Furthermore, by enhancing temporal correlation through channel-wise affine transformation, the time step is effectively embedded into the noisy features, enabling the multi-scale attention block to know the current state of the diffusion process in order to better predict noise. The features embedded at each time step are then processed by multi-scale depthwise separable convolutional blocks. This module captures multi-scale features through depthwise separable convolutions with three different kernel sizes (3, 5, and 7). It combines a high-efficiency channel attention module to perform weighted fusion of features from different channels, and finally concatenates the fusion results and restores them to the original channel dimension through 1×1 convolution. And layer normalization is used to stabilize the training process.

[0058] Depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations independently on each channel of the input feature map, while pointwise convolution uses a 1×1 convolution kernel to perform convolution operations on the input feature map, achieving cross-channel feature fusion and adjusting the number of channels. In this way, the feature extraction capability of the model can be preserved while reducing the computational cost.

[0059] The efficient channel attention module first performs global average pooling on the temporal dimension of each channel to compress spatial information into channel-level statistics. Then, it performs one-dimensional convolution on the pooling results to learn the interaction relationships between channels and generate channel attention weights. Since the kernel size of the one-dimensional convolution is dynamically calculated based on the number of channels, no dimensionality reduction operation is required. The specific calculation formula is shown in Equation (12).

[0060] (12) In the formula, Indicates the kernel size; and These are hyperparameters, set to 1 and 2 respectively. This indicates that the odd number closest to the calculated result is selected. Then, the channel attention weights are mapped to... using the Sigmoid function. The interval is used to obtain the normalized weight values. Finally, the attention weights are... Multiply the original input features channel by channel to enhance the features of important channels.

[0061] (3) Training of anomaly detection model.

[0062] When training the anomaly detection model, a two-stage model training method is adopted, which can be mainly divided into two stages: pre-training of the temporal feature enhancement module and training of the attention-guided potential diffusion module. Through step-by-step optimization, accurate modeling of the potential feature distribution of normal samples is achieved.

[0063] (3.1) Pre-training of the temporal feature enhancement module.

[0064] The temporal feature enhancement module achieves dimensionality compression and accurate reconstruction of high-dimensional vibration data through encoding and decoding processes. To achieve good feature extraction capabilities and reconstruction results, and to learn effective low-dimensional latent feature representations, the following training objectives are used to optimize it during the pre-training stage: (13) in, This represents taking the mathematical expectation. This represents the parameters used in model training.

[0065] (3.2) Attention-guided training of potential diffusion modules.

[0066] The attention-guided latent diffusion module is trained on the latent space based on a pre-trained temporal feature enhancement module. It learns the feature distribution of normal samples by simulating a "noise addition-denoising" process. To achieve good noise prediction capabilities, the module is optimized using the following training objective: (14) in, For the denoising network at time step and noisy samples Predicted noise below, For time steps Injected real Gaussian noise.

[0067] (4) Detection of abnormal online status.

[0068] Once the transformer condition anomaly detection model is trained, it can be used for online anomaly detection of transformers. Vibration data collected from operating transformers is input into the detection model, which automatically calculates the anomaly score of these vibration data points and compares it to an anomaly score threshold. If the score exceeds the threshold, an anomaly is output; otherwise, no output is made, thus achieving online anomaly detection of transformers.

[0069] Example 2 This embodiment discloses an equipment status anomaly detection system under an unsupervised learning framework.

[0070] An equipment status anomaly detection system based on an unsupervised learning framework includes: The data acquisition module is configured to acquire vibration data of the transformer under normal operating conditions. The first latent feature representation extraction module is configured to: input vibration data into the anomaly detection model, and use the encoder in the temporal feature enhancement module to extract features to obtain the first latent feature representation; The second latent feature representation estimation module is configured to: input the first latent feature representation into the attention-guided latent diffusion module, and learn the feature distribution of vibration data under normal operating conditions by forward noise addition and backward noise removal, and by combining multi-scale feature extraction and attention-optimized expression, to obtain the second latent feature representation; The data reconstruction module is configured to input the second latent feature representation into the decoder of the temporal feature enhancement module to obtain reconstructed data; The model training module is configured to: calculate the reconstruction error between vibration data and reconstructed data, and train the anomaly detection model; The online detection module is configured to input the online vibration data of the transformer into the trained anomaly detection model to achieve online condition detection of the transformer. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0071] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the equipment state anomaly detection method under the unsupervised learning framework as described in Embodiment 1 of this disclosure.

[0072] Example 4 The purpose of this embodiment is to provide an electronic device.

[0073] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the equipment state anomaly detection method under the unsupervised learning framework as described in Embodiment 1 of this disclosure.

[0074] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0075] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0076] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting equipment state anomalies within an unsupervised learning framework, characterized in that, Includes the following steps: Obtain vibration data of the transformer under normal operating conditions; Vibration data is input into the anomaly detection model, and the encoder in the temporal feature enhancement module is used to extract features to obtain the first latent feature representation. The first latent feature representation is input into the attention-guided latent diffusion module. Through forward noise addition and backward noise removal, and by combining multi-scale feature extraction and attention-optimized expression, the feature distribution of vibration data under normal operating conditions is learned to obtain the second latent feature representation. The second latent feature representation is input into the decoder of the temporal feature enhancement module to obtain the reconstructed data; Calculate the reconstruction error between vibration data and reconstructed data, and train the anomaly detection model accordingly; The online vibration data of the transformer is input into the trained anomaly detection model to achieve online condition detection of the transformer.

2. The equipment state anomaly detection method under the unsupervised learning framework as described in claim 1, characterized in that, Feature extraction is performed using the encoder in the temporal feature enhancement module to obtain the first latent feature representation, specifically as follows: ； in, This indicates the encoder's mapping operation; This represents the first latent feature representation; Represents vibration data; The encoder in the temporal feature enhancement module consists of multiple stacked temporal residual blocks. Each residual block contains two layers of dilated causal convolutions, which are applied at time 10:

00. Output Represented as: ； in, Represents the convolution kernel sequence The first in One convolutional kernel; Represents a one-dimensional time series. The first in A point in time; Represents the convolution kernel sequence The sequence index of the last convolutional kernel in the process; Indicates an index in the convolution kernel sequence; Indicates the expansion rate.

3. The equipment state anomaly detection method under the unsupervised learning framework as described in claim 1, characterized in that, The forward noise addition process specifically includes: The first latent feature is represented as the original latent feature. ; From original latent features In the beginning, Gaussian noise is gradually added within each time step to ultimately generate noisy features. ; No. Noise addition result of the step The following formula is used to calculate: ； in, Indicates the cumulative coefficient; Indicates the number of noise-adding steps; Indicates standard Gaussian noise; Represents a standard Gaussian distribution. It is the zero vector. It is the identity matrix; The reverse denoising process specifically includes: From noisy features Starting from this point, the noise prediction result is obtained through a noise prediction network, and the second latent feature representation is gradually derived. feature pass It is calculated by the following formula: ； in, and These represent the time steps. and The noisy features at that location, and yes The result after one round of noise reduction; Indicates at time step Cumulative coefficient over time; Indicates noise prediction network noise For the prediction results of the original noise, the input of the noise prediction network is the noisy features. and time step ,Right now For noisy features and time step The corresponding noise prediction and estimation; It is the random noise figure; Is with Irrelevant standard Gaussian noise.

4. The equipment state anomaly detection method under the unsupervised learning framework as described in claim 3, characterized in that, The noise prediction results are obtained through a noise prediction network. The specific data processing steps include: The noise prediction network employs an attention-optimized multi-scale UNet network with an encoder-decoder symmetric structure: The input to the noise prediction network is and time step In the encoder part of the noise prediction network, local patterns and global dependencies of input latent features are captured by multi-scale attention blocks, and feature dimensions are progressively compressed by downsampling. In the decoder part of the noise prediction network, the sequence dimension is gradually recovered by upsampling and fused with the same level features from the encoder. The fused features are then input into a multi-scale attention block for further feature extraction, and finally the prediction result for the noise is output.

5. The equipment state anomaly detection method under the unsupervised learning framework as described in claim 4, characterized in that, A gated attention fusion approach is used to fuse features from the decoder and encoder at the same level in the noise prediction network, as shown below: ； in, These are the characteristics after fusion; Indicates a splicing operation; These are the same-level features output after multi-scale attention block feature extraction in the encoder; It is element-wise multiplication; This represents the Sigmoid activation function; This represents a 1×1 convolution operation; These are features obtained after upsampling.

6. The equipment state anomaly detection method under the unsupervised learning framework as described in claim 4, characterized in that, In multi-scale attention blocks: Time step The scale parameters are obtained by MLP transformation and projection. and offset parameters Furthermore, time correlation is enhanced through channel-wise affine transformation, embedding time steps into noisy features; The features embedded at time steps are then processed by multi-scale deep separable convolutional blocks to capture multi-scale features. The features of different channels are weighted and fused by combining an efficient channel attention module. Finally, the fusion results are spliced ​​together and then restored to the original channel dimension by 1×1 convolution. The efficient channel attention module first performs global average pooling on the temporal dimension of each channel to compress spatial information into channel-level statistics. Then, it performs one-dimensional convolution on the pooling results to learn the interaction relationship between channels and generate channel attention weights.

7. The equipment state anomaly detection method under the unsupervised learning framework as described in claim 1, characterized in that, The second latent feature representation is input into the decoder of the temporal feature enhancement module to obtain the reconstructed data, specifically represented as follows: ； in, This is the mapping of the decoder in the temporal feature enhancement module; To reconstruct the data; This represents the second latent feature. or, The reconstruction error between the vibration data and the reconstructed data is calculated to obtain the anomaly score of the data. Specifically: ； in, Represents the square of the Euclidean norm; Represents vibration data; This indicates data reconstruction; For the number of channels, The sequence length; or, A two-stage model training approach is used when training the anomaly detection model: The first stage only pre-trains the temporal feature enhancement module in the anomaly detection model; The second stage only trains the attention-guided potential diffusion module in the anomaly detection model.

8. An equipment status anomaly detection system under an unsupervised learning framework, characterized in that, include: The data acquisition module is configured to acquire vibration data of the transformer under normal operating conditions. The first latent feature representation extraction module is configured to: input vibration data into the anomaly detection model, and use the encoder in the temporal feature enhancement module to extract features to obtain the first latent feature representation; The second latent feature representation estimation module is configured to: input the first latent feature representation into the attention-guided latent diffusion module, and learn the feature distribution of vibration data under normal operating conditions by forward noise addition and backward noise removal, and by combining multi-scale feature extraction and attention-optimized expression, to obtain the second latent feature representation; The data reconstruction module is configured to input the second latent feature representation into the decoder of the temporal feature enhancement module to obtain reconstructed data; The model training module is configured to: calculate the reconstruction error between vibration data and reconstructed data, and train the anomaly detection model; The online detection module is configured to input the online vibration data of the transformer into the trained anomaly detection model to achieve online condition detection of the transformer.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the equipment state anomaly detection method under the unsupervised learning framework as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the equipment state anomaly detection method under the unsupervised learning framework as described in any one of claims 1-7.

Images