A converter transformer fault diagnosis method and system based on future perception

Abstract

The application discloses a converter transformer fault diagnosis method and system based on future perception. Multi-source monitoring signals are collected to form multi-channel time series, and a sliding window is used to construct a training sample pair including a current observation segment and a future prediction segment; trend discrimination and segment reconstruction subtasks are constructed respectively, and a shared encoder is jointly trained based on the training sample pair; the shared encoder is migrated to a fault diagnosis network, features of the current observation segment are extracted, and different monitoring channels and time steps are weighted through channel attention and time attention mechanisms; the weighted time series features are input into an LSTM to obtain global time series features; and then, a classification output head is used to obtain prediction probabilities corresponding to different operating states or fault types, and a fault diagnosis result is output. The application can learn the state evolution law of the converter transformer by using unlabeled operation data, reduce the dependence on fault labeled samples, and improve the identification ability of early weak faults under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of power system equipment condition monitoring and intelligent operation and maintenance technology, and particularly relates to a method and system for fault diagnosis of converter transformers based on future perception. Background Technology

[0002] Converter transformers are core equipment in high-voltage direct current (HVDC) transmission systems, enabling voltage matching, electrical isolation, and energy transfer between the AC system and converter valves. Compared to conventional AC transformers, the valve-side windings of converter transformers endure complex electrical stresses caused by the combined effects of AC, DC, and harmonic voltages over extended periods. They may also experience complex operating processes such as commutation failure, DC bias, polarity reversal, inrush current, and rapid load changes. These processes subject the internal insulation, windings, core, and mechanical support structures of the converter transformer to continuous or intermittent electrical, magnetic, thermal, and mechanical stresses, resulting in fault evolution characterized by long duration, numerous influencing factors, and weak early-stage symptoms.

[0003] Current diagnostic methods for converter transformers typically employ dissolved gas analysis, oil temperature or winding temperature threshold alarms, partial discharge detection, and vibration analysis. While single physicochemical indicators or fixed thresholds have relatively clear physical meanings, fixed thresholds are difficult to adapt to dynamic changes in converter transformer load rate, ambient temperature, and system operating conditions. When the baseline of the monitoring signal changes with operating conditions, fixed thresholds may misinterpret normal load fluctuations as abnormalities or fail to identify early deterioration processes that have not yet reached the fixed alarm value.

[0004] Other technologies extract time-frequency features from monitoring signals using wavelet analysis and empirical mode decomposition, and utilize support vector machines, random forests, convolutional neural networks, or long short-term memory networks (LSTM) to identify converter transformer faults. However, converter transformers operate normally for most of the time in actual power grids, and faults such as severe winding deformation, inter-turn short circuits, and multi-point grounding of the core occur with a low probability, resulting in a limited number of fault samples that can be obtained and accurately labeled. Conventional supervised models rely on massive amounts of labeled samples. When the number of fault samples is small and the class distribution is unbalanced, they are prone to memorizing accidental features from the limited training samples, making it difficult to learn the general laws governing the long-term state evolution of converter transformers.

[0005] Furthermore, different faults in the converter transformer have varying impacts on different monitoring signals. For example, DC bias may be more pronounced in vibration and specific harmonic channels, overload or localized overheating may initially cause changes in temperature and load-related channels, while partial discharge or early inter-turn insulation abnormalities may only form pulses or sudden changes within a short period. Existing models typically directly splice together monitoring channels or use fixed weights for fusion, failing to dynamically adjust the importance of different channels according to the current operating state. This easily leads to weak anomalies in a few key channels being diluted by a large amount of steady-state information.

[0006] Over time, changes in conditions such as insulation aging and structural loosening may last for months, while partial discharge pulses, vibration shocks, or current surges may only last for a few sampling periods. When processing long sequences containing a large amount of normal operation data, ordinary time-series networks struggle to simultaneously maintain attention to both long-term degradation trends and short-term abrupt changes, and early, weak fault characteristics are easily overwhelmed by normal periodic fluctuations and AC / DC transient noise. Summary of the Invention

[0007] The main technical problem to be solved by this invention is: how to learn the evolution law of equipment state by using a large amount of unlabeled operating data under the condition of a limited number of fault tags, so as to improve the diagnostic capability of early weak faults of converter transformers under complex operating conditions.

[0008] Therefore, the present invention provides a fault diagnosis method for converter transformers based on future perception, which includes an offline model construction stage and an online fault diagnosis stage; The offline model construction phase includes: S1. Acquire multi-source monitoring signals of the converter transformer to form a multi-channel time series observation matrix; S2. Extract training sample pairs from the multi-channel time series observation matrix using a sliding window method. Each training sample pair includes the current observation segment and a future predicted segment that immediately follows the current observation segment in time. S3. Based on the training samples, pre-train the shared encoder; S4. Transfer the pre-trained shared encoder to the fault diagnosis network, and train the fault diagnosis network using training samples with fault category labels. The online fault diagnosis stage includes: acquiring the current observation segment of the converter transformer to be diagnosed, inputting the current observation segment into the trained fault diagnosis network to obtain a fault category probability vector, and determining the operating state or fault type corresponding to the maximum predicted probability in the fault category probability vector as the fault diagnosis result of the converter transformer to be diagnosed.

[0009] Furthermore, the pre-training includes: The trend discrimination subtask involves constructing trend pseudo-labels based on trend features in the future prediction segment, inputting the current observation segment into the shared encoder, and using the trend prediction head to predict the trend pseudo-labels based on the output of the shared encoder. The trend discrimination loss is then determined based on the prediction results and the trend pseudo-labels. The segment reconstruction subtask involves partially masking the data in the future prediction segment, concatenating the current observation segment with the masked future prediction segment to obtain the reconstruction input sequence; using the shared encoder to extract features from the reconstruction input sequence, and using the decoder to reconstruct the masked future prediction segment based on the features; and determining the segment reconstruction loss based on the reconstruction result and the future prediction segment before masking. The shared encoder is jointly trained based on the trend discrimination loss and the segment reconstruction loss.

[0010] Furthermore, the fault diagnosis network includes a pre-trained shared encoder, a channel attention module, a temporal attention module, a long short-term memory network, and a classification output head. In S4, the current observation segment of the training sample is input into the shared encoder to obtain a high-dimensional temporal feature tensor. The channel attention mechanism is used to recalibrate different monitoring channels in the high-dimensional temporal feature tensor. The temporal attention mechanism is used to weight the recalibrated features at different time steps. The weighted temporal feature sequence that retains the temporal order is input into the long short-term memory network to obtain global temporal features. The global temporal features are then input into the classification output head to obtain the prediction probabilities corresponding to different operating states or fault types.

[0011] Furthermore, let the total number of monitoring channels be... The number of sampling time steps is The time series observation matrix is ​​then represented as:

[0012] in, For the first Multi-channel monitoring signal vectors acquired at time steps; For the initial time step The current observation segment is represented as:

[0013] The future prediction segment is represented as:

[0014] in, The number of time steps contained in the current observation segment. The number of time steps contained in the future prediction segment.

[0015] Furthermore, the construction of the trend pseudo-label includes the following steps: For the first future prediction segment c For each monitoring channel, perform max pooling along the time dimension to obtain the trend extreme value of that monitoring channel:

[0016] in, The first segment representing a future prediction c Signal sequences of each monitoring channel This represents the maximum pooling operation on the signal sequence. For the first c Trend extremes of each monitoring channel; According to the first c The dynamic threshold corresponding to a monitoring channel is determined by the historical normal operation data of each monitoring channel. The trend extreme value is compared with the dynamic threshold to obtain the trend pseudo-label of the monitoring channel:

[0017] in, Indicates the first c Trend pseudo-labels for each monitoring channel; Constructing a multi-channel fusion trend indicator: ; Combine the fusion trend indicator with a unified threshold By comparing, we can obtain the overall trend pseudo-label. As the final pseudo-trend label; ,

[0018] in, For the first c The contribution weight of each monitoring channel to trend judgment.

[0019] Furthermore, partial masking is applied to the data in the future prediction segment, including: Construct a masking matrix with the same dimensions as the predicted future fragment:

[0020] In the occlusion matrix, elements with a value of 1 indicate that the corresponding data in the future prediction segment is visible, and elements with a value of 0 indicate that the corresponding data in the future prediction segment is occluded. The masking methods include time-dimensional masking and / or channel-dimensional masking. The time-dimensional masking involves randomly selecting a continuous time interval in the future prediction segment and masking all or some monitoring channels within the continuous time interval. The channel-dimensional masking involves randomly selecting at least one monitoring channel and masking all or some time positions of the selected monitoring channel within the future prediction segment. The masked future prediction fragment is represented as follows: ⊙ ⊙

[0021] Where ⊙ represents element-wise multiplication. This is the masking fill value.

[0022] Furthermore, the different monitoring channels in the high-dimensional temporal feature tensor are recalibrated using a channel attention mechanism, including the following steps: Global max pooling and global average pooling are performed on the high-dimensional temporal feature tensor output by the shared encoder along the time dimension to obtain max pooled features. and average pooling characteristics ; The max-pooling feature and the average-pooling feature are respectively input into a weight-shared multilayer perceptron. The two outputs of the multilayer perceptron are summed element-wise and then input into a sigmoid activation function to obtain the channel sensitivity weight vector.

[0023] in, Represents the channel sensitivity weight vector. This represents the Sigmoid activation function. This indicates the processing operations of a multilayer perceptron; The high-dimensional temporal feature tensor is recalibrated channel-by-channel using the channel sensitivity weight vector to obtain channel-enhanced features:

[0024] in, Indicates channel enhancement features, For the high-dimensional temporal feature tensor, This indicates multiplying each element in the high-dimensional temporal feature tensor by... The corresponding channel sensitivity weights.

[0025] Furthermore, a temporal attention mechanism is used to weight the recalibrated features at different time steps, including: The channel enhancement features are expanded in the channel dimension and feature embedding dimension, and the first... The comprehensive features corresponding to each time step are denoted as follows: ; Calculate the first based on a learnable global context query vector. Fault correlation score at each time step:

[0026] in, For the global context query vector, It is a linear mapping matrix. For bias terms; The fault correlation scores at each time step are normalized using the Softmax function to obtain the... Temporal attention weights for each time step:

[0027] The comprehensive features at the corresponding time steps are weighted using the aforementioned temporal attention weights to obtain weighted features:

[0028] By arranging the weighted features of each time step in chronological order, a time-series feature sequence that preserves the time sequence is obtained:

[0029] The temporal feature sequence is sequentially input into a long short-term memory network to obtain the temporal hidden features corresponding to each time step, and the temporal hidden features corresponding to the last time step are taken as the global temporal features.

[0030] Preferably, the classification output head includes a single fully connected output layer and a Softmax probability normalization layer; The single-layer fully connected output layer maps the global temporal features to a K-dimensional class score space:

[0031] in, For the global timing characteristics, K is the total number of pre-defined categories of operating states and fault types. The weight matrix for the classification output layer. The bias vector for the classification output layer. This is a category score vector; The Softmax probability normalization layer uses the Softmax function to normalize the category score vectors.

[0032] in, This is the probability vector for the fault category. This indicates that the currently observed segment belongs to the first... The predicted probability of a certain operating state or fault type.

[0033] The present invention also provides a converter transformer fault diagnosis system based on the above method, the system comprising: The multi-source monitoring and sample construction module is used to acquire multi-source monitoring signals during the operation of the converter transformer, form a multi-channel time series based on the multi-source monitoring signals, and construct training sample pairs from the multi-channel time series in a sliding window manner. Each training sample pair includes the current observation segment and the future predicted segment that is located after the current observation segment in time. The future-aware self-supervised pre-training module includes a shared encoder, a trend discrimination unit, a segment reconstruction unit, and a joint optimization unit. The trend discrimination unit constructs trend pseudo-labels based on the trend features of the future prediction segment and a dynamic threshold determined based on historical normal operation data, and uses the shared encoder to predict the trend pseudo-labels based on the currently observed segment. The segment reconstruction unit partially masks the data in the future prediction segment and reconstructs the masked portion using the shared encoder and its connected decoder based on the currently observed segment and the masked future prediction segment. The joint optimization unit jointly optimizes and trains the shared encoder based on the trend discrimination result and the segment reconstruction result. The spatiotemporal attention feature decoupling module is used to extract the temporal features corresponding to the current observation segment using the shared encoder trained by the future perception self-supervised pre-training module, and to recalibrate the channel dimension weights and assign the time dimension weights to the temporal features to obtain a weighted temporal feature sequence that retains the time order. The LSTM time series modeling module is used to input the weighted time series feature sequence into the long short-term memory network, extract the state evolution relationship between different time positions within the current observation segment, and output global time series features characterizing the operating state of the converter transformer. The fault classification module is used to map the global time-series features to the category prediction probabilities corresponding to the preset operating state and / or fault type, and to determine the diagnostic results of the converter transformer based on the category prediction probabilities. During the model training phase, the future prediction segments are used to construct the trend pseudo-labels and segment reconstruction targets; during the online diagnosis phase, the system obtains the diagnostic results based on the current observation segments of the converter transformer to be diagnosed through the shared encoder, the spatiotemporal attention feature decoupling module, the LSTM time series modeling module, and the fault classification module.

[0034] Compared with the prior art, the present invention has at least the following beneficial effects: This invention constructs a future-aware self-supervised learning mechanism, utilizing current observation segments and subsequent future prediction segments to form trend discrimination and segment reconstruction tasks. Without requiring manual fault labeling of large amounts of historical operational data, this enables a shared encoder to learn the intrinsic correlation between the current state and subsequent evolution trend of a converter transformer from massive amounts of normal operation data and a small amount of fault data. Specifically, the trend discrimination task generates trend pseudo-labels based on the trend extremes of future prediction segments and dynamic thresholds determined by historical normal operation data, guiding the model to extract predictive features characterizing subsequent state changes from the current observation segments. The segment reconstruction task, by masking parts of the time locations or monitoring channels in the future prediction segments and using current observation information and unmasked information to recover the masked signals, prompts the model to learn the temporal continuity and physical coupling relationships between multi-source monitoring signals. Through the collaborative training of these two self-supervised tasks, this invention reduces the dependence of the fault diagnosis model on the number of severe fault samples and the quality of manual labels, improving the model's applicability and generalization ability under conditions of small samples, class imbalance, and scarcity of actual fault data.

[0035] To address the shortcomings of existing methods that typically statically stitch together multi-source monitoring signals, making it difficult to highlight sensitive channels and critical anomaly moments corresponding to different faults, this invention sequentially implements a channel attention mechanism and a time attention mechanism after pre-training a shared encoder. The channel attention mechanism comprehensively utilizes the overall steady-state characteristics and extreme abrupt change characteristics of each monitoring channel in the current observation segment, adaptively adjusting the importance of different monitoring channels and their embedded features. This enhances monitoring information relevant to the current fault state and suppresses redundant channels and background noise. The time attention mechanism further calculates the correlation between different time steps and the fault state, allocating higher weights to critical moments where partial discharge pulses, vibration shocks, current surges, or other short-term anomalies occur, preventing early, weak fault features from being averaged or submerged in long-term sequences and periodic fluctuations.

[0036] Building upon this foundation, this invention utilizes an LSTM network to perform time-series modeling on feature sequences that retain temporal order after channel attention and temporal attention processing. This enables the model to combine state information before and after an anomaly to determine whether changes at different time points are continuous and to extract long-term evolution characteristics of the converter transformer's operating state. Thus, this invention can simultaneously address slow-developing processes such as insulation aging and overheating, as well as short-term abrupt changes such as partial discharge and inter-turn anomalies, achieving collaborative diagnosis of long-term degradation trend tracking and short-term anomaly localization. The global time-series features output by the LSTM network are mapped to predicted probabilities of normal operating state and different fault types via a classification output head, allowing the system to directly output the converter transformer's operating state or fault category.

[0037] Furthermore, the future prediction segments in this invention are only used to construct self-supervised learning objectives during the offline training phase. In the actual online diagnosis process, only the monitoring data collected at the current moment and before the current moment needs to be input, without waiting for the formation of real future monitoring signals. Therefore, this invention does not have the problem of using data that has not yet occurred for current diagnosis, and can be deployed on converter station servers, edge computing devices, or remote operation and maintenance platforms to perform continuous online status identification and fault diagnosis of converter transformers.

[0038] In summary, this invention is beneficial for improving the early fault identification capability, diagnostic stability, and anti-interference capability of converter transformers under complex AC / DC operating conditions, strong noise, multi-source heterogeneous monitoring, and scarce fault samples. Attached Figure Description

[0039] The accompanying drawings, together with the embodiments of the present invention, are used to explain the present invention and do not constitute a limitation thereof.

[0040] Figure 1 This is a schematic diagram of the technical framework provided in the embodiments of the present invention; Figure 2 This is a schematic diagram illustrating the construction of training sample pairs provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the pre-training of the shared encoder provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a segment reconstruction subtask based on a partial occlusion strategy provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0042] Example 1 This embodiment provides a method for fault diagnosis of converter transformers based on future sensing. For example... Figure 1 As shown, this method mainly includes the following steps: I. Acquisition of Multi-Source Monitoring Signals and Sample Construction Monitoring devices are installed or connected at key locations such as the converter transformer body, bushings, tank, grid-side leads, valve-side leads, and cooling system to acquire multi-source monitoring signals related to the converter transformer's operating status. These multi-source monitoring signals may include grid-side three-phase voltage, grid-side three-phase current, valve-side voltage, valve-side current, DC voltage, DC current, converter transformer load rate, top oil temperature, winding hot spot temperature, partial discharge amplitude, partial discharge pulse count, tank vibration, core vibration, sound signals, dissolved gas concentration, and ambient temperature.

[0043] Assume the total number of signal acquisition channels is The sampling time step is Then it forms Time series observation matrix:

[0044] in, The total length of the time series corresponds to the number of time steps in signal acquisition, and its size is determined by both the sampling frequency and the monitoring duration. For a moment Collected The monitoring signal vector represents the observation value of each monitoring channel at that moment.

[0045] In time series modeling tasks, extracting representative fragment samples from the original long sequence is fundamental to model learning. Considering that converter transformer faults typically exhibit characteristics such as localized suddenness and short-term drastic fluctuations, this embodiment employs a sliding window mechanism on the time series observation matrix to fragment the original data and construct training sample pairs.

[0046] like Figure 2 As shown, each sample pair consists of the currently observed segment and the immediately following future predicted segment. For the starting time step... The current observation segment is represented as:

[0047] Future prediction segments are represented as follows:

[0048] in, This represents the number of time steps in the current observation window. The number of time steps for the future forecast window.

[0049] The future prediction segment is sequentially connected to the current observation segment in time, so a sample pair can simultaneously describe both the state changes that have already occurred and the subsequent state development of the device. It should be noted that the future prediction segment is only used to construct the self-supervised objective during the offline training phase. During the online execution phase, the model only receives the current observation segment and does not need to wait for the future prediction segment to form. The length of the future prediction window... The preferred size is the current observation window length. 1 / 2 to 2 / 3. The preferred sliding step size is... The step size is set to 1 / 4 to 1 / 2 to ensure sample diversity while avoiding sample redundancy caused by an excessively small step size.

[0050] II. Future Perception Self-Supervised Pre-training like Figure 1 and Figure 3 As shown, the future-aware self-supervised pre-training includes a trend discrimination subtask and a segment reconstruction subtask. The two subtasks share encoder parameters, enabling the shared encoder to simultaneously learn predictable features of future abnormal trends and structural correlations of multi-channel signals.

[0051] 1. Trend Judgment Subtask The task aims to simulate the trend prediction capability of converter transformers during operation: in the offline training phase, a supervision signal is constructed using a future prediction window, forcing the encoder to perform feature inference based solely on the current observation window, thereby learning the physical mapping relationship between the current state and future evolution.

[0052] Most converter transformer faults typically manifest as a sustained increase or a sudden surge in amplitude before they occur. Therefore, the first step is to target future prediction segments. The peak value of abrupt changes in the monitoring signal is extracted using channel maximum pooling. For the first segment of the future prediction... c For each monitoring channel, perform max pooling along the time dimension to obtain the trend extreme values:

[0053] in, The first segment representing a future prediction c Signal sequences of each monitoring channel This represents the maximum pooling operation on a signal sequence, i.e., finding the maximum value in the sequence.

[0054] Compare the extreme value of the trend with the corresponding threshold of the channel. Compare to generate a trend pseudo-label for that channel. :

[0055] Given the dynamic variability of signal amplitude under different operating conditions (such as equipment load changes and ambient temperature fluctuations), this embodiment introduces an adaptive dynamic threshold mechanism to avoid the problem of fixed thresholds failing to adapt to load changes. The 90th percentile is calculated from the historical normal operation data of the monitoring channel as the trend classification threshold. Historical normal operation data can be determined based on equipment maintenance records, protection action records, and confirmation results from operating personnel.

[0056] To reduce the possibility of false labels being generated due to accidental noise interference in a single channel, a multi-channel fusion trend index is constructed: ,

[0057]

[0058] in, Indicates the first c The channel's contribution weight to trend judgment can be set based on historical fault sensitivity experience. The sample pair corresponds to a trend pseudo-label. According to the integration trend indicator With uniform threshold The judgment is made to improve the stability and generalization ability of the overall judgment while avoiding interference from channel fluctuations.

[0059] The current observation segment of the sample pair The corresponding trend pseudo-labels are used as training samples for training the shared encoder. During training, the current observation segment is used... The input is a shared encoder. The output of the shared encoder is passed through a trend prediction head (which can be a conventional binary classification network in existing technology) to output a trend classification prediction result. The trend discrimination loss is calculated based on the trend classification prediction result and the corresponding trend pseudo-label, and is used to update the encoder parameters. Since the trend discrimination subtask does not see future prediction segments in the input, it is necessary to extract features that can characterize future state changes from the current observation segment.

[0060] 2. Fragment Reconstruction Subtask like Figure 4 As shown, this task aims to simulate scenarios where critical signals are missing or data is incomplete in actual operation, guiding the encoder to recover obscured future segments from historical observations and enhancing its understanding of fault evolution structures. It forms a complementary closed loop with the trend discrimination subtask, jointly improving the encoder's ability to model future abnormal trends.

[0061] To simulate the incompleteness of future signals, a local masking strategy is adopted to partially mask the future prediction segments in the sample pairs, generating sequences with missing information.

[0062] First, construct a masking matrix of the same dimension as the future prediction segment:

[0063] In the occlusion matrix, an element of 1 indicates that the corresponding position in the future prediction segment is visible, while an element of 0 indicates that the corresponding position is occluded. Occlusion methods include time-dimensional occlusion and channel-dimensional occlusion. Time-dimensional occlusion involves randomly selecting a continuous time interval within the future prediction segment and occluding all or some channels within that interval to simulate a period of unavailable monitoring data. Channel-dimensional occlusion involves randomly selecting one or more monitoring channels and occluding all or some of the selected channels' time positions within the future prediction segment to simulate sensor failure or communication interruption.

[0064] The future prediction fragment after occlusion processing can be represented as: ⊙ ⊙

[0065] ⊙ represents element-wise multiplication. The masking fill value is usually set to 0, but it can also be designed according to the channel type.

[0066] The current observation segment and the masked future prediction segment are concatenated along the time dimension to form a complete input sequence:

[0067] Extracting the input sequence using a shared encoder The global features are then used by the decoder to output the reconstruction results of future predicted segments based on the global features. The goal is to recover occluded future prediction segments. A shared encoder and decoder can employ existing encoder-decoder network architectures, where global features are extracted by encoding the input sequence, and the decoder reconstructs the occluded regions of the future prediction segments.

[0068] The optimization objective of the fragment reconstruction subtask is to minimize the reconstruction error:

[0069] This loss function guides the shared encoder to learn the trend continuity of the time-series signal and the implicit structural features between channels, enabling the extracted latent representation to predict future evolution patterns.

[0070] 3. Self-supervised joint optimization The trend discrimination subtask and the segment reconstruction subtask are performed simultaneously and collaboratively during the training phase: the former provides trend discrimination signals, enhancing the model's sensitivity to whether the future is "abnormal"; the latter provides structural completion signals, strengthening the model's representation of "how anomalies evolve". Both work together on the encoder, resulting in extracted representation features that have significant advantages in both trend predictability and structural reconstruction.

[0071] The total loss of the network during the self-supervised pre-training phase is:

[0072] in, This indicates the loss of the trend-judgment subtask.

[0073] The parameters of the shared encoder, trend prediction head, and decoder are updated using mini-batch gradient descent. After pre-training is complete, the shared encoding parameters are saved for later use.

[0074] III. Spatiotemporal attention feature decoupling and LSTM temporal modeling The pre-trained shared encoder is migrated to the fault diagnosis network and fine-tuned using a small number of samples with clearly defined fault labels. Fault labels can be determined based on protection action records, power outage maintenance results, partial discharge tests, winding deformation detection results, oil chromatography analysis results, and equipment defect records. Fault categories can be set according to actual operation and maintenance needs. For example, they can be set to normal state, DC bias, overload or overheating, partial discharge, early inter-turn short circuit, winding deformation, and core multi-point grounding. The number of different categories does not constitute a limitation on the invention.

[0075] 1. Channel attention mechanism The operating signals of a converter transformer include multiple heterogeneous channels (such as three-phase AC current, DC voltage, vibration, etc.), and the fault sensitivity of each channel changes dynamically under different fault modes. To address this, this embodiment introduces a channel attention mechanism to assign importance weights to the physical characteristics of each channel.

[0076] Let the high-dimensional temporal feature tensor extracted by the self-supervised pre-trained encoder be... , This defines the feature embedding dimension for a single channel. The specific feature purification process is as follows: First, along the time dimension Global max pooling (MaxPool) and global average pooling (AvgPool) are performed separately to capture the extreme transient abrupt changes and the overall steady-state trend of each channel, resulting in two spatially compressed representations. and : ,

[0077] in, .

[0078] Next, the two compressed features mentioned above are input into a weight-shared multilayer perceptron network (MLP, containing one nonlinear ReLU activation layer) to calculate the coupling correlation between channels. The outputs of both are then summed element-wise and passed through a sigmoid activation function. Normalization generates sensitivity weight vectors for each channel. :

[0079] Finally, using the obtained weight vector For the original feature tensor Channel-by-channel recalibration is performed to obtain channel enhancement features. : .

[0080] This step achieves adaptive focusing on key fault measurement points at the physical level.

[0081] 2. Time-based attention mechanism Features after channel enhancement It still contains lengthy timing information. Since converter transformer faults are often accompanied by signal abrupt changes within a very short time, the time attention mechanism aims to guide the model to actively focus on key anomalous moments on the time axis.

[0082] First, enhance the channel features. Fusion is carried out across the channel and embedded feature dimensions to record time points. The comprehensive feature vector is Secondly, learnable global context query vectors are introduced. (The abstract memory standard representing the network's "critical failure moment") is calculated at each time step through a feedforward network. Initial score of fault correlation :

[0083] in, and Let be the linear mapping matrix and the bias term.

[0084] Subsequently, the scores of all time steps are normalized using the Softmax function to obtain the temporal attention weight distribution:

[0085] Weight The peak value directly maps to the physical moment when the fault occurs.

[0086] Finally, using The features at each time step are dynamically weighted to obtain a refined time series feature sequence. :

[0087] Preserve the weighted temporal structure (the shape remains inclusive). The feature sequence of each time step is filtered out to remove redundant interference from regular periodic fluctuations, and then input into the downstream LSTM network for deep modeling of long-range evolution dependence.

[0088] 3. LSTM timing modeling After the feature sequences are enhanced with channel and temporal attention, the dynamic evolution of the signal needs to be captured through time-series modeling, especially the state correlation before and after the fault occurs. LSTM, as a recurrent neural network specifically designed for processing time-series data, uses three gated units (input gate, forget gate, and output gate) to control the updating and retrieval of long-term memory units. It can effectively solve the gradient vanishing problem of traditional RNNs and is suitable for modeling long-range dependencies in the operating signals of converter transformers.

[0089] This embodiment directly uses the existing LSTM as the temporal feature extraction network without improving the gating structure and internal operation method of the LSTM itself. This embodiment uses the temporal feature sequence, which retains its temporal order after processing by the temporal attention mechanism, as the input to the LSTM network. This allows the LSTM network to read the features of each time step sequentially from earliest to latest, establishing the correlation between the preceding and following states within the current observation segment, and extracting the evolutionary features of the converter transformer's operating state over time.

[0090] After channel attention recalibration and temporal attention weighting, a temporal feature sequence arranged in chronological order is obtained. ,in, The number of time steps contained in the current observation segment. For the first The weighted feature vectors corresponding to each time step. The weighted feature vector corresponding to the first time step... Input the LSTM network, then input the following in sequence: to When processing the features of the current time step, the LSTM network also incorporates the state information formed in previous time steps to determine whether the abnormal changes at different time positions are continuous.

[0091] In one implementation, the LSTM network employs a unidirectional LSTM structure, ensuring that features are transmitted sequentially within the current observation segment, thus avoiding the use of uncollected data outside the current observation window during online diagnosis. The LSTM network can be configured as a single-layer or multi-layer stacked structure, with the hidden feature dimension of each layer determined based on the number of monitoring channels and the complexity of the classification task, for example, set to 64, 128, or 256.

[0092] After temporally encoding the input sequence, the LSTM network outputs the hidden feature sequence corresponding to each time step:

[0093] in, This indicates that the LSTM network is reading the first... The time-series hidden features are formed after considering the features of each time step. The hidden features not only include the weighted monitoring information of the current time step, but also incorporate the running state information before the current time step.

[0094] This embodiment will use the hidden features corresponding to the last time step. Global temporal features of the current observation segment The global time-series features serve as input to the subsequent classification output head, used to calculate the predicted probability that the converter transformer belongs to a normal state or different fault types. The hidden features of the last time step have gathered the state change information from the first time step to the last time step within the current observation window, which can reflect the overall operating state and its evolution trend corresponding to the current observation segment.

[0095] IV. Fault Classification Training After obtaining the global temporal features corresponding to the current observation segment through an LSTM network, this invention uses a classification output head to map the global temporal features to the predicted probabilities corresponding to the normal operating state or fault type of the converter transformer, generating the final fault diagnosis result. The classification output head adopts a multi-class classification structure, mainly including a single-layer fully connected output layer and a Softmax probability normalization layer.

[0096] 1. Single-layer fully connected output layer A single fully connected output layer with global temporal features As input, the output dimension equals the pre-defined total number K of operating states or fault categories. Its mathematical expression is:

[0097] in, The weight matrix for the classification output layer. The bias vector for the classification output layer; The category score vector generated for the classification output layer. Category Score Used to characterize the current global temporal features and the first The degree of matching between different operating states or fault types. The higher the category score, the closer the temporal evolution characteristics of the current observation segment are to the feature patterns corresponding to that category.

[0098] The main function of the fully connected output layer is to learn the mapping relationship between global temporal features and different operating states or fault types, mapping the global temporal features to a K-dimensional class score space. Since the LSTM network has already extracted the temporal dependencies, the classification output head does not need to perform temporal modeling again, but directly completes the class discrimination based on the global temporal features.

[0099] The total number of fault categories K described in this embodiment can be set according to actual monitoring needs. For example, the operating status of the converter transformer can be divided into categories such as normal operation, DC bias, overload or overheating, partial discharge, inter-turn short circuit, winding deformation, and core multi-point grounding. In this case, the classification output layer is set with output nodes corresponding to the number of the above categories, and each output node corresponds to a predetermined operating status or fault type.

[0100] 2. Softmax probability normalization layer The category scores obtained from a single fully connected output layer do not directly represent probabilities. To obtain the predicted probabilities of the current observation segment belonging to each operating state or fault type, the Softmax function is used to normalize the category score vector.

[0101] in, This is the probability vector for the fault category. This indicates that the currently observed segment belongs to the first... The predicted probability of a certain operating state or fault type.

[0102] The Softmax function performs an exponential transformation and normalization on the scores of each category, making the sum of the predicted probabilities of all categories equal to 1. This facilitates a direct comparison of the probabilities of different fault categories. For example, when the predicted probability of a local discharge category is significantly higher than that of other categories, it indicates that the fault-sensitive channel features, critical time-period abrupt change features, and long-term temporal evolution features extracted in the current observation segment are closest to the local discharge feature patterns learned during the model training phase.

[0103] 3. Model Training During the training phase of the fault classification model, corresponding fault category labels are configured for the training samples based on the maintenance records, protection action records, partial discharge detection results, oil chromatography analysis results, winding deformation detection results, or equipment status confirmed by maintenance personnel of the converter transformer.

[0104] For the For each labeled training sample, its true class label can be represented as a one-hot encoded vector:

[0105] When the The training sample belongs to the first... When setting a category, The label values ​​for all other categories are set to 0. The fault classification cross-entropy loss is calculated based on the true category labels and the predicted probabilities generated by the classification output head.

[0106] Where N is the number of labeled training samples. For the first The training sample belongs to the first... Predicted probability of each category.

[0107] During training, the parameters of the classification output layer, LSTM network, temporal attention mechanism, and channel attention mechanism are updated in reverse according to the fault classification cross-entropy loss, so that the global temporal features of the same fault type are gradually mapped to the same category, and the category scores corresponding to different operating states or fault types are more clearly distinguished.

[0108] After the model training is completed, the trained shared encoder, channel attention mechanism, temporal attention mechanism, LSTM network and classification output head are deployed on the converter station server, edge computing device or remote operation and maintenance platform.

[0109] V. Online Fault Diagnosis During online diagnosis, the current observation segment is first extracted from the continuous monitoring data of the converter transformer and input into a shared encoder to obtain the corresponding high-dimensional spatiotemporal features. Then, the features of each monitoring channel are recalibrated using a channel attention mechanism, and the features at different time steps are weighted using a time attention mechanism. The weighted time-series feature sequence obtained after dual attention processing is input into an LSTM network, which extracts the evolution relationship between the previous and subsequent states within the current observation segment and outputs a global time-series feature vector. Finally, the global time-series feature vector is input into the classification output head to obtain a fault category probability vector. The prediction probability with the largest value is selected from all predicted probabilities, and the corresponding category is used as the diagnosis result of the current converter transformer.

[0110] Example 2 Based on the above method, this embodiment provides a converter transformer fault diagnosis system based on future perception. The system constructs current observation segments and future prediction segments through multi-source monitoring signals of the converter transformer, performs self-supervised pre-training of the shared encoder using future trend discrimination tasks and segment occlusion reconstruction tasks, and sequentially uses the shared encoder, channel attention mechanism, time attention mechanism, LSTM network and classification output head to identify the operating status or fault type of the converter transformer during the online diagnosis process.

[0111] The system mainly includes a multi-source monitoring and sample construction module, a future perception self-supervised pre-training module, a spatiotemporal attention feature decoupling module, an LSTM temporal modeling module, a fault classification module, and an online diagnostic output module.

[0112] 1. Multi-source monitoring and sample construction module The multi-source monitoring and sample construction module is used to acquire multi-source monitoring signals during the operation of the converter transformer and synchronize the signals from different monitoring channels according to the acquisition time to form a multi-channel time series observation matrix.

[0113] During the model training phase, the multi-source monitoring and sample construction module uses a sliding window approach to truncate the multi-channel time-series observation matrix and construct training sample pairs. Each training sample pair includes a current observation segment and a future prediction segment that immediately follows the current observation segment in time. The current observation segment is used to characterize the state changes that have already occurred in the converter transformer, and the future prediction segment is used to construct future-aware self-supervised targets during the offline training phase.

[0114] During the online diagnostic phase, the multi-source monitoring and sample construction module only constructs the current observation segment and does not need to wait for or obtain future predicted segments after the current observation segment.

[0115] 2. Future-aware self-supervised pre-training module This module is used to perform self-supervised pre-training on the shared encoder using training samples, and specifically includes a trend discrimination unit, a segment reconstruction unit, and a joint optimization unit.

[0116] The trend discrimination unit is used to extract the trend extreme values ​​of each monitoring channel in the future prediction segment, and compare the trend extreme values ​​with the dynamic threshold determined based on historical normal operation data to generate trend pseudo-labels; it is also used to input the current observation segment into the shared encoder, and output the trend prediction result through the trend prediction head, and calculate the trend discrimination loss based on the trend prediction result and the trend pseudo-labels.

[0117] The segment reconstruction unit is used to mask part of the time location and / or part of the monitoring channel in the future prediction segment, stitch the current observation segment and the masked future prediction segment along the time dimension, and reconstruct the masked future prediction segment using a shared encoder and decoder. The segment reconstruction loss is calculated based on the reconstruction result and the future prediction segment before masking.

[0118] The joint optimization unit updates the network parameters of the shared encoder, trend prediction head, and decoder based on the trend discrimination loss and segment reconstruction loss, and saves the network parameters of the shared encoder after self-supervised pre-training. This enables the shared encoder to simultaneously learn the correlation between the current operating state and future trends, as well as the temporal coupling relationship between different monitoring channels.

[0119] 3. Spatiotemporal attention feature decoupling module The spatiotemporal attention feature decoupling module is used to load the pre-trained shared encoder and encode the current observation segment using the shared encoder to obtain a high-dimensional temporal feature tensor. This module further includes a channel attention unit and a temporal attention unit.

[0120] The channel attention unit is used to perform max pooling and average pooling on the high-dimensional temporal feature tensor along the time dimension, and inputs the pooling results into a weight-sharing multilayer perceptron. The attention weights corresponding to different monitoring channels and their embedded features are obtained through the sigmoid activation function. The high-dimensional temporal feature tensor is then recalibrated using the attention weights to enhance the monitoring channel features related to the current fault state and suppress the monitoring channel features with low fault correlation.

[0121] The time attention unit is used to expand the features after channel recalibration according to time steps, calculate the correlation score between each time step and the fault state of the converter transformer, and obtain the time attention weight through the Softmax function; the features of each time step are weighted according to the time attention weight to obtain a weighted time series feature sequence that still retains the time sequence.

[0122] 4. LSTM timing modeling module This module receives the weighted temporal feature sequence output by the temporal attention unit. In this embodiment, an existing unidirectional LSTM network is used to sequentially read the weighted features corresponding to each time step from earliest to latest, establishing state associations between different time positions within the current observation segment and extracting the evolutionary features of the converter transformer's operating state over time. The LSTM network outputs temporal hidden features corresponding to each time step, and uses the temporal hidden features corresponding to the last time step as the global temporal feature of the current observation segment. The global temporal feature aggregates abnormal fluctuations of each monitoring channel within the current observation segment, feature mutations at key time positions, and long-term dependencies between preceding and following time steps.

[0123] 5. Fault Classification Module This module is connected to the LSTM time series modeling module and is used to convert the acquired global time series features into predicted probabilities corresponding to different operating states or fault types of the converter transformer.

[0124] The fault classification module includes a single-layer fully connected output layer and a Softmax probability normalization layer. The single-layer fully connected output layer takes global temporal features as input and outputs a category score that corresponds one-to-one with a pre-defined operating state or fault category; the Softmax probability normalization layer is used to convert each category score into a fault category probability vector.

[0125] During the model training phase, the fault classification module calculates the fault classification loss based on the true fault categories and predicted probabilities of the labeled training samples, and updates the parameters of the classification output layer, LSTM network, temporal attention unit, and channel attention unit based on the fault classification loss.

[0126] 6. Online diagnostic output module This module is used to call the trained shared encoder, channel attention unit, time attention unit, LSTM network and fault classification module to perform online fault diagnosis on the converter transformer.

[0127] During online diagnosis, the online diagnosis output module inputs the current observation segment into the shared encoder to obtain a high-dimensional temporal feature tensor; then, it sequentially uses the channel attention unit and the temporal attention unit to perform feature recalibration and temporal weighting; the obtained weighted temporal feature sequence is input into the LSTM network to obtain global temporal features; and then the global temporal features are input into the fault classification module to obtain the predicted probability corresponding to each operating state or fault type.

[0128] The online diagnostic output module selects the maximum predicted probability from the fault category probability vector and determines the operating status or fault type corresponding to the maximum predicted probability as the diagnostic result of the converter transformer.

[0129] The above modules can be implemented by the processor in the converter station server, edge computing device or remote operation and maintenance platform executing the corresponding computer program, or they can be implemented by hardware units and software units with corresponding data acquisition, storage and computing functions.

[0130] The above system can execute the converter transformer fault diagnosis method described in Embodiment 1, and has the corresponding functional modules and beneficial effects of the method. For technical details not described in detail in this embodiment, please refer to Embodiment 1 of the present invention.

[0131] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0132] 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 them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for fault diagnosis of converter transformers based on future sensing, characterized in that, This includes the offline model building phase and the online fault diagnosis phase; The offline model construction phase includes: S1. Acquire multi-source monitoring signals of the converter transformer to form a multi-channel time series observation matrix; S2. Extract training sample pairs from the multi-channel time series observation matrix using a sliding window method. Each training sample pair includes the current observation segment and a future predicted segment that immediately follows the current observation segment in time. S3. Based on the training samples, pre-train the shared encoder; S4. Transfer the pre-trained shared encoder to the fault diagnosis network, and train the fault diagnosis network using training samples with fault category labels. The online fault diagnosis stage includes: acquiring the current observation segment of the converter transformer to be diagnosed, inputting the current observation segment into the trained fault diagnosis network to obtain a fault category probability vector, and determining the operating state or fault type corresponding to the maximum predicted probability in the fault category probability vector as the fault diagnosis result of the converter transformer to be diagnosed.

2. The converter transformer fault diagnosis method as described in claim 1, characterized in that: Assume the total number of monitoring channels is The number of sampling time steps is The time series observation matrix is ​​then represented as: in, For the first Multi-channel monitoring signal vectors acquired at time steps; For the initial time step The current observation segment is represented as: The future prediction segment is represented as: in, The number of time steps contained in the current observation segment. The number of time steps contained in the future prediction segment.

3. The converter transformer fault diagnosis method as described in claim 2, characterized in that, The pre-training includes: The trend discrimination subtask involves constructing trend pseudo-labels based on trend features in the future prediction segment, inputting the current observation segment into the shared encoder, and using the trend prediction head to predict the trend pseudo-labels based on the output of the shared encoder. The trend discrimination loss is then determined based on the prediction results and the trend pseudo-labels. The segment reconstruction subtask involves partially masking the data in the future prediction segment, concatenating the current observation segment with the masked future prediction segment to obtain the reconstruction input sequence; using the shared encoder to extract features from the reconstruction input sequence, and using the decoder to reconstruct the masked future prediction segment based on the features; and determining the segment reconstruction loss based on the reconstruction result and the future prediction segment before masking. The shared encoder is jointly trained based on the trend discrimination loss and the segment reconstruction loss.

4. The converter transformer fault diagnosis method as described in claim 3, characterized in that, The construction of the trend pseudo-label includes the following steps: For the first future prediction segment c For each monitoring channel, perform max pooling along the time dimension to obtain the trend extreme value of that monitoring channel: in, The first segment representing a future prediction c Signal sequences of each monitoring channel This represents the maximum pooling operation on the signal sequence. For the first c Trend extremes of each monitoring channel; According to the first c The dynamic threshold corresponding to a monitoring channel is determined by the historical normal operation data of each monitoring channel. The trend extreme value is compared with the dynamic threshold to obtain the trend pseudo-label of the monitoring channel: in, Indicates the first c Trend pseudo-labels for each monitoring channel; Constructing a multi-channel fusion trend indicator: ; Combine the fusion trend indicator with a unified threshold By comparing, we can obtain the overall trend pseudo-label. As the final pseudo-trend label; ， in, For the first c The contribution weight of each monitoring channel to trend judgment.

5. The converter transformer fault diagnosis method as described in claim 3, characterized in that, Partial masking of data in the predicted future segments includes: Construct a masking matrix with the same dimensions as the predicted future fragment: In the occlusion matrix, elements with a value of 1 indicate that the corresponding data in the future prediction segment is visible, and elements with a value of 0 indicate that the corresponding data in the future prediction segment is occluded. The masking methods include time-dimensional masking and / or channel-dimensional masking. The time-dimensional masking involves randomly selecting a continuous time interval in the future prediction segment and masking all or some monitoring channels within the continuous time interval. The channel-dimensional masking involves randomly selecting at least one monitoring channel and masking all or some time positions of the selected monitoring channel within the future prediction segment. The masked future prediction fragment is represented as follows: ⊙ ⊙ Where ⊙ represents element-wise multiplication. This is the masking fill value.

6. The converter transformer fault diagnosis method as described in claim 1, characterized in that, The fault diagnosis network includes a pre-trained shared encoder, a channel attention module, a temporal attention module, a long short-term memory network, and a classification output head. In S4, the current observation segment of the training sample is input into the shared encoder to obtain a high-dimensional temporal feature tensor. The channel attention mechanism is used to recalibrate different monitoring channels in the high-dimensional temporal feature tensor. The temporal attention mechanism is used to weight the recalibrated features at different time steps. The weighted temporal feature sequence that retains the temporal order is input into the long short-term memory network to obtain global temporal features. The global temporal features are then input into the classification output head to obtain the prediction probabilities corresponding to different operating states or fault types.

7. The converter transformer fault diagnosis method as described in claim 6, characterized in that, The channel attention mechanism is used to recalibrate different monitoring channels in the high-dimensional temporal feature tensor, including: Global max pooling and global average pooling are performed on the high-dimensional temporal feature tensor output by the shared encoder along the time dimension to obtain max pooled features. and average pooling characteristics ; The max-pooling feature and the average-pooling feature are respectively input into a weight-shared multilayer perceptron. The two outputs of the multilayer perceptron are summed element-wise and then input into a sigmoid activation function to obtain the channel sensitivity weight vector. in, Represents the channel sensitivity weight vector. This represents the Sigmoid activation function. This indicates the processing operations of the multilayer perceptron; The high-dimensional temporal feature tensor is recalibrated channel-by-channel using the channel sensitivity weight vector to obtain channel-enhanced features: in, Indicates channel enhancement features, For the high-dimensional temporal feature tensor, This indicates multiplying each element in the high-dimensional temporal feature tensor by... The corresponding channel sensitivity weights.

8. The converter transformer fault diagnosis method as described in claim 7, characterized in that, The recalibrated features are weighted at different time steps using a temporal attention mechanism, including: The channel enhancement features are expanded in the channel dimension and feature embedding dimension, and the first... The comprehensive features corresponding to each time step are denoted as follows: ; Calculate the first based on a learnable global context query vector. Fault correlation score at each time step: in, This is the global context query vector. It is a linear mapping matrix. For bias terms; The fault correlation scores at each time step are normalized using the Softmax function to obtain the... Temporal attention weights for each time step: The comprehensive features at the corresponding time steps are weighted using the aforementioned temporal attention weights to obtain weighted features: By arranging the weighted features of each time step in chronological order, a time-series feature sequence that preserves the time sequence is obtained: The temporal feature sequence is sequentially input into a long short-term memory network to obtain the temporal hidden features corresponding to each time step, and the temporal hidden features corresponding to the last time step are taken as the global temporal features.

9. The converter transformer fault diagnosis method as described in claim 6, characterized in that, The classification output head includes a single fully connected output layer and a Softmax probability normalization layer; The single-layer fully connected output layer maps the global temporal features to a K-dimensional class score space: in, For the global timing characteristics, K is the total number of pre-defined categories of operating states and fault types. The weight matrix for the classification output layer. The bias vector for the classification output layer. This is a category score vector; The Softmax probability normalization layer uses the Softmax function to normalize the category score vectors. in, This is a probability vector for the fault category. This indicates that the currently observed segment belongs to the first... The predicted probability of a certain operating state or fault type.

10. A converter transformer fault diagnosis system based on the method of any one of claims 1 to 9, characterized in that, Includes the following modules: The multi-source monitoring and sample construction module is used to acquire multi-source monitoring signals during the operation of the converter transformer, form a multi-channel time series based on the multi-source monitoring signals, and construct training sample pairs from the multi-channel time series in a sliding window manner. Each training sample pair includes the current observation segment and the future predicted segment that is located after the current observation segment in time. The future-aware self-supervised pre-training module includes a shared encoder, a trend discrimination unit, a segment reconstruction unit, and a joint optimization unit. The trend discrimination unit constructs trend pseudo-labels based on the trend features of the future prediction segment and a dynamic threshold determined based on historical normal operation data, and uses the shared encoder to predict the trend pseudo-labels based on the currently observed segment. The segment reconstruction unit partially masks the data in the future prediction segment and reconstructs the masked portion using the shared encoder and its connected decoder based on the currently observed segment and the masked future prediction segment. The joint optimization unit jointly optimizes and trains the shared encoder based on the trend discrimination result and the segment reconstruction result. The spatiotemporal attention feature decoupling module is used to extract the temporal features corresponding to the current observation segment using the shared encoder trained by the future perception self-supervised pre-training module, and to recalibrate the channel dimension weights and assign the time dimension weights to the temporal features to obtain a weighted temporal feature sequence that retains the time order. The LSTM time series modeling module is used to input the weighted time series feature sequence into the long short-term memory network, extract the state evolution relationship between different time positions within the current observation segment, and output global time series features characterizing the operating state of the converter transformer. The fault classification module is used to map the global time-series features to the category prediction probabilities corresponding to the preset operating state and / or fault type, and to determine the diagnostic results of the converter transformer based on the category prediction probabilities. In the model training phase, the future prediction segments are used to construct the trend pseudo-labels and segment reconstruction targets; in the online diagnosis phase, the system obtains the diagnostic results based on the current observation segments of the converter transformer to be diagnosed through the shared encoder, the spatiotemporal attention feature decoupling module, the LSTM time series modeling module, and the fault classification module.

Images