A transformer dga data dynamic denoising and fault classification method fusing environmental multi-modal features

Abstract

The present application relates to the field of transformer fault detection, and particularly relates to a kind of transformer DGA data dynamic denoising and fault classification method fusing environmental multi-modal features. In order to solve the technical problem that existing algorithm cannot accurately identify and inhibit sensor false features caused by external strong electromagnetic transient, leading to model false alarm, the method first synchronously collects and time window alignment pretreatment to multi-source heterogeneous data;Cross-modal feature fusion network based on dynamic gating mechanism is constructed;Fault classification mapping and diagnosis output are carried out based on cleaning features;And offline training and forced noise coupling enhancement are carried out. The present application constructs dynamic gating weight by introducing electromagnetic auxiliary modal data, actively shields DGA data pollution caused by strong electromagnetic pulse, and when determining that the current measurement value is contaminated, it is switched to the predicted value derived based on historical features, effectively solving the frequent false alarm problem caused by severe electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of transformer fault detection, and specifically to a method for dynamic denoising and fault classification of transformer DGA data that integrates environmental multimodal features. Background Technology

[0002] Dissolved gas analysis (DGA) is the most effective means of monitoring latent faults inside large power transformers. In recent years, the introduction of machine learning and pattern recognition technologies (such as neural networks and support vector machines) for automatic fault classification of DGA data has become the mainstream development direction in this field.

[0003] However, existing pure data-driven algorithms face an extremely serious and unresolved physical challenge when deployed in actual substation environments: they cannot accurately identify and suppress sensor spurious features (such as data spikes or baseline drift) caused by external strong electromagnetic transients.

[0004] In real substations (especially 330kV and above high-voltage substations), the operation of disconnecting switches and the action of circuit breakers generate extremely strong transient electromagnetic radiation. The weak signal processing circuits inside existing online chromatography monitoring sensors are highly susceptible to this interference, causing instantaneous spikes or baseline drifts in the measured concentration of certain gases (such as hydrogen and carbon monoxide) even when no actual physical fault has occurred. Existing pattern recognition algorithms rely solely on a single DGA data input and cannot mathematically distinguish whether such data abrupt changes are due to genuine high-energy arc discharge inside the transformer or sensor distortion caused by external electromagnetic interference. If traditional mathematical smoothing filtering is used, it may miss real serious faults; if the data is fed directly into the model without processing, it will lead to frequent false alarms from the system.

[0005] Therefore, accurately identifying and suppressing sensor spurious features (data spikes or baseline drift) caused by external strong electromagnetic transients in transformer DGA online monitoring, thereby dynamically assessing the authenticity and reliability of DGA sampling data at a single moment and reducing false alarms has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0006] To address the technical problem of false alarms in classification models caused by abrupt changes (pseudo-features) in sensor data acquired due to strong electromagnetic interference in complex substation environments, this invention aims to provide a dynamic denoising and fault classification method for transformer DGA data that integrates multimodal environmental features. The specific technical solution adopted is as follows:

[0007] This invention proposes a dynamic denoising and fault classification method for transformer DGA data that integrates environmental multimodal features. The method includes:

[0008] The concentrations of various characteristic gases and electromagnetic signals in the transformer are synchronously acquired and preprocessed with time window alignment to obtain gas phase sequences and electromagnetic feature vectors.

[0009] A cross-modal feature fusion network based on a dynamic gating mechanism is constructed. Through the cross-modal feature fusion network, actual gas phase feature vectors and predicted gas phase feature vectors are extracted based on the gas phase sequence. At the same time, gating weights are calculated based on the electromagnetic feature vectors. The actual gas phase feature vectors and the predicted gas phase feature vectors are then weighted and fused based on the gating weights to obtain clean gas phase feature vectors.

[0010] The clean gas phase feature vector is fed into a classifier for classification and mapping to obtain the fault diagnosis results of the transformer.

[0011] Furthermore, the acquisition of the gas phase sequence and electromagnetic feature vector includes:

[0012] According to the preset sampling period, the concentration values ​​of various characteristic gases at the current sampling time and multiple previous historical times are obtained, multiple gas phase concentration vectors are constructed, and the multiple gas phase concentration vectors are spliced ​​together in chronological order to form the gas phase sequence.

[0013] Continuously record the transient electromagnetic waveform signals around the transformer;

[0014] Using the current sampling time as a reference, a time window of a preset time length is extracted in the direction of historical time. Within the time window, the key statistical features of the transient electromagnetic waveform signal are calculated, and the key statistical features are normalized and then concatenated to obtain the electromagnetic feature vector aligned to the current sampling time.

[0015] Furthermore, the key statistical features include the maximum pulse amplitude, the number of pulses exceeding a preset threshold, and the high-frequency energy integral. The electromagnetic feature vector is composed of the normalized maximum pulse amplitude, the pulse frequency, and the high-frequency energy integral.

[0016] Furthermore, the extraction of the actual gas phase feature vector and the predicted gas phase feature vector includes:

[0017] The gas phase sequence is input into the first long short-term memory network in the main branch network, and the hidden state of the last time step is taken as the actual gas phase feature vector representing the true measurement value.

[0018] The remaining historical sequence after removing the elements collected at the current sampling time from the gas phase sequence is input into a parallel second long short-term memory network, and the theoretical prediction value for the current sampling time is output as the predicted gas phase feature vector.

[0019] Furthermore, the calculation of the gating weights includes:

[0020] The electromagnetic feature vector is input into the multilayer perceptron in the auxiliary branch network, and the gating weights representing the confidence of data contamination are output by the activation function in the last layer.

[0021] The gating weights are calculated as follows: the electromagnetic feature vectors are weighted and summed and biased using the multilayer perceptron to obtain intermediate feature values, and the intermediate feature values ​​are mapped using an activation function to obtain the gating weights.

[0022] Furthermore, obtaining the clean gas phase feature vector includes:

[0023] Calculate the difference between the value 1 and the gate weight, and multiply the difference by the actual gas phase feature vector to obtain the first weighted feature;

[0024] The gating weights are multiplied by the predicted gas phase feature vector to obtain the second weighted feature;

[0025] The clean gas phase feature vector is obtained by performing a vector addition operation between the first weighted feature and the second weighted feature.

[0026] Furthermore, the obtained fault diagnosis results for the transformer include:

[0027] The clean gas phase feature vector is fed into the classifier consisting of a hidden layer and an output layer;

[0028] The output layer features are converted into a relative probability distribution of the current sample belonging to each preset transformer fault type through the Softmax layer;

[0029] The fault category corresponding to the highest probability value in the relative probability distribution is selected as the final fault diagnosis result.

[0030] Furthermore, the method also includes an offline training phase and a forced noise coupling enhancement step, wherein in the offline training phase, the overall loss function of the cross-modal feature fusion network is:

[0031] Calculate the standard cross-entropy loss of the classification output;

[0032] Calculate the mean squared error loss between the predicted features output by the main branch network and the noise-free true features, and weight the mean squared error loss using a first preset weight parameter;

[0033] Calculate the regularization penalty value applied to the gate weight, and weight the regularization penalty value using the second preset weight parameter;

[0034] The standard cross-entropy loss, the weighted mean squared error loss, and the weighted regularization penalty are summed to obtain the overall loss function used for model gradient update.

[0035] Furthermore, the forced noise coupling enhancement step includes:

[0036] A normal continuous gas phase sequence and its corresponding stable electromagnetic waveform signal are extracted from the historical database as the initial training sample.

[0037] At random moments in the continuous gas phase sequence, Gaussian pulses with a preset amplitude multiple are forcibly injected into the concentration data of the characteristic gas to simulate sudden interference to the sensor circuit.

[0038] In the electromagnetic feature vector corresponding to the same random moment, the pulse frequency and high-frequency energy integral are forcibly replaced with the corresponding historical maximum values ​​of a preset multiple, generating strong noise training samples with causal correspondence and inputting them into the network for training.

[0039] Furthermore, the method also includes a step of interpretable log display based on the gating weights:

[0040] While outputting the final fault diagnosis result, the gating weight and its associated electromagnetic feature vector are extracted and recorded simultaneously when judging the current sample.

[0041] When the gating weight reaches or exceeds the interception threshold, a prompt message is output on the system console, indicating that the current chromatographic sample is judged as dirty data due to strong electromagnetic interference, and the system has automatically switched to using the predicted gas phase feature vector for smoothing.

[0042] The present invention has the following beneficial effects:

[0043] 1. Traditional pure mathematical denoising algorithms cannot distinguish between genuine fault gas generation and false sensor noise when faced with high-frequency abrupt changes in data. This invention introduces electromagnetic / environmental auxiliary mode data and constructs a dynamic gating weight. When the auxiliary mode detects a strong external electromagnetic pulse, the gating mechanism actively assigns extremely low confidence to the concurrent DGA data, thereby physically shielding the data glitches caused by sensor induced current and solving the problem of frequent false alarms in transformer condition monitoring systems caused by harsh electromagnetic environments.

[0044] 2. Existing anti-interference methods typically employ crude processing approaches such as direct rejection or blind smoothing when anomaly data is detected, which can easily lead to data loss or feature distortion at critical time points. This invention, through an internal fusion mechanism, allows the neural network to smoothly switch to a predicted value derived from historical time-series stationary features when the current measurement value is determined to be contaminated. Even if the online sensor experiences strong electromagnetic blindness lasting for several seconds or even minutes, the system can still maintain a reasonable baseline output through an autoregressive mechanism, ensuring the continuity of the transformer condition assessment curve and the extremely robustness of the final fault diagnosis conclusion. Attached Figure Description

[0045] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart of a method for dynamic denoising and fault classification of transformer DGA data that integrates environmental multimodal features, provided in one embodiment of the present invention;

[0047] Figure 2 A schematic diagram of a cross-modal feature fusion network topology based on a dynamic gating mechanism provided in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram comparing the anti-disturbance smoothing effect of the dynamic gating mechanism provided in one embodiment of the present invention. Detailed Implementation

[0049] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a method for dynamic denoising and fault classification of transformer DGA data that integrates environmental multimodal characteristics, proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0050] Unless otherwise defined, 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.

[0051] Example 1

[0052] The following description, in conjunction with the accompanying drawings, details the specific scheme of the transformer DGA data dynamic denoising and fault classification method that integrates environmental multimodal features provided by the present invention.

[0053] The implementation architecture of this embodiment relies on a hardware acquisition system deployed at the transformer site, specifically including: an online oil chromatography monitoring device (for acquiring primary mode gas phase data) and an ultra-high frequency (UHF) sensor (for acquiring secondary mode electromagnetic environment data).

[0054] Please see Figure 1 The diagram illustrates a flowchart of a method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features, according to an embodiment of the present invention. The method includes:

[0055] Step S1: Synchronous acquisition and time window alignment preprocessing of multi-source heterogeneous data.

[0056] In real power grid environments, chemical parameters (gas concentration) and physical parameters (electromagnetic waves) differ significantly in data dimension, sampling frequency, and physical units. Without cross-modal spatiotemporal alignment preprocessing, subsequent neural networks will face convergence problems. Therefore, this invention first constructs a standardized multimodal temporal feature matrix to compensate for sampling differences between different sensor hardware, laying a feasible data foundation for subsequent cross-modal fusion.

[0057] This invention first employs an online transformer oil chromatography monitoring device to perform discrete sampling according to a preset sampling period T (preferably T = 4 hours in this embodiment), obtaining the concentration values ​​of multiple characteristic gases (unit: The characteristic gases include at least, for example, hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide. For the current sampling time t, construct the gas phase concentration vector for a single sampling. Where n represents the number of gas types, indicating that the dimension of the gas phase concentration vector is n, and the elements in the gas phase concentration vector are the concentration values ​​of each sampled gas. In one embodiment of the present invention, n can be set to 7, that is, the 7 types of gases mentioned above. Taking the current sampling time t and L−1 historical data points prior to it, a gas phase sequence of length L is constructed as the main mode input. Since the absolute concentration of characteristic gases at a single moment cannot accurately reflect the degradation trend of insulating materials and is easily affected by short-term fluctuations in ambient temperature and load, this embodiment of the invention employs a gas phase sequence of length L. As a subsequent input, it is to introduce the time dimension of the internal physicochemical reaction of the transformer, thereby transforming the static concentration value into a dynamic gas production rate. This allows the subsequent network model to not only learn the static threshold features of the absolute value of gas exceeding the limit, but also to capture the gas micro-climbing features unique to the early stage of latent faults (i.e., the time-series slow change features), which greatly improves the detection sensitivity of early low-energy discharge and local overheating.

[0058] Electromagnetic waves within a substation capable of penetrating the metal shielding of a chromatograph and causing transient latch-up or breakdown interference to the internal high-precision analog-to-digital converter primarily originate from ultra-fast transient overvoltages generated by the operation of high-voltage disconnect switches. The spectral energy of this type of interference is mainly concentrated in the ultra-high frequency (UHF) band and its duration is extremely short (microseconds to nanoseconds). Therefore, this embodiment of the invention also requires the use of an ultra-high frequency (UHF) sensor with a bandwidth of 300MHz to 1.5GHz to continuously record transient electromagnetic waveform signals around the transformer at a microsecond-level sampling rate. This allows for the filtering out of conventional 50Hz power frequency magnetic field interference and low-frequency motor noise within the substation, ensuring a strong physical causal relationship between the acquired electromagnetic characteristics and the hardware distortion of the chromatographic sensor. This provides high signal-to-noise ratio data for the accurate evaluation of the subsequent gated network.

[0059] Since microsecond-level waveforms cannot be directly spliced ​​with 4-hour-level gas phase data, this embodiment of the invention designs a downsampling mapping based on a time window. First, using the current sampling time t of the gas phase as a reference, a length of [missing information] is extracted towards the historical direction. Time window (preferred in this embodiment) The physical environment within 10 minutes prior to the current sampling time t has the greatest impact on the analog-to-digital conversion circuit of the sensor.

[0060] Calculate the UHF waveform signal within this time window. Three key statistical characteristics:

[0061] 1. Maximum pulse amplitude : Characterizes the instantaneous intensity limit of electromagnetic interference.

[0062] 2. Pulse frequency The number of pulses exceeding a preset threshold is counted to characterize the density of interference. The preset threshold is a known value that can be manually calibrated on-site. This can be done by measuring the background high-frequency electromagnetic noise during normal and stable operation of the substation during the equipment installation and commissioning phase, taking its maximum envelope or average value, and then multiplying it by a certain redundancy coefficient (such as 1.5 times). This value is then directly burned into the system as the preset threshold for interference judgment. Alternatively, it can be an inherent parameter of the sensor, i.e., the broadband (300MHz~1.5GHz) ultra-high frequency sensor hardware or its preamplifier circuit itself has a fixed trigger threshold.

[0063] 3. High-frequency energy integration The calculation formula is: , which represents the total energy accumulation of the disturbance.

[0064] After normalizing the above three physical quantities, they are concatenated to form the electromagnetic eigenvector at time t. This serves as the input for subsequent auxiliary modes.

[0065] By analyzing electromagnetic waveform signals Feature extraction was performed to reduce the dimensionality of the massive and chaotic high-frequency electromagnetic waveform signal into a low-dimensional vector containing clear physical meaning (intensity, frequency, energy), and it was strictly aligned with the time t of the gas phase data. This eliminated the modal differences between different sensors and significantly reduced the computational load and overfitting risk of the subsequent auxiliary branch network.

[0066] Step S2: Construct a cross-modal feature fusion network based on a dynamic gating mechanism.

[0067] Traditional deep learning diagnostic models (such as simple backpropagation or CNN) will rigidly attempt to classify any data sent by the sensor as a certain fault, lacking a self-checking mechanism for input quality. The embodiments of this invention, through the design of a dual-branch structure, enable the network to dynamically evaluate the reliability of its own input data, shielding the fatal misleading effect of hardware limitations from the algorithm's underlying layer.

[0068] First, a main branch network is constructed for the gas phase sequence of the gas. Autoregressive feature extraction is performed, and the main branch network uses a first long short-term memory (LSTM) network with 128 hidden layer units. Then, the gas phase sequence obtained in step S1 is processed. Input the LSTM and take the hidden state of the last time step as the actual gas phase feature vector at the current sampling time. and the gas phase sequence The last element in (i.e., the gas phase concentration vector at the current sampling time) Remove, and extract the sequence. Inputting into another parallel second long short-term memory network, the output is a theoretical prediction of the features at the current sampling time, i.e., a predicted gas phase feature vector. Transformer insulating oil pyrolysis is an extremely slow thermodynamic and chemical reaction process, conforming to the Markov chain property (i.e., future states are highly dependent on historical states). This invention proposes a dual-output LSTM structure. On the one hand, it extracts the actual gas phase feature vector to ensure sensitivity to real sudden faults. On the other hand, it simultaneously constructs a stable predictive feature that does not depend on the gas phase concentration vector at the current sampling time, i.e., predicts the gas phase feature vector, which serves as a backup reference for the system under extreme disturbances. Even if the online chromatographic sensor encounters an extreme electromagnetic flash that contaminates the single sampling data, the model still has the self-healing ability to maintain stable output based on this backup reference (historical inertial features), preventing partial paralysis of the diagnostic system.

[0069] Because the electromagnetic environment at transformer sites is extremely complex and the interference intensity varies greatly between different substations, rigidly setting a fixed electromagnetic threshold to determine whether the data is distorted is prone to misjudgment. Therefore, this embodiment of the invention further constructs an auxiliary branch network to evaluate the confidence level of the generated electromagnetic interference. The auxiliary branch network uses a multilayer perceptron (MLP) combined with a sigmoid function for affine transformation and nonlinear mapping, allowing the network to automatically learn the implicit correlation between specific electromagnetic pulse energy and the probability of chromatographic hardware failure through backpropagation.

[0070] Then the electromagnetic eigenvectors The input is fed into a multilayer perceptron (MLP) containing two fully connected layers (16 nodes and 1 node respectively). The last layer is connected to a sigmoid activation function and outputs gating weights. :

[0071]

[0072] in, The weight matrix represents the parameters learned by the network; it is a 1x3 row vector. The bias term, also a parameter learned by the network, is a scalar value.

[0073] Among them, gating weights Used to dynamically and quantitatively assess the contamination level of gas phase data collected by the chromatograph, the value range of the gate weight is strictly limited to 0~1. When the key statistical features of the electromagnetic waveform signals in each dimension of the input electromagnetic feature vector increase sharply, the neuron is activated and the gate weight quickly approaches the value of 1. Then the system determines that the current external electromagnetic environment is extremely bad and the gas chromatographic data of the same period is likely to be pseudo data generated by circuit interference.

[0074] When the online transformer oil chromatography monitoring device generates dirty data due to electromagnetic interference, conventional feature splicing cannot remove the dirty data, and dirty features will still penetrate deep into the network and destroy the classification boundary. In this embodiment of the invention, the following controlled fusion calculation is performed in the feature fusion layer of the network:

[0075]

[0076] in, The fused clean gas phase feature vector is given when there is no electromagnetic interference in the environment. At this time This indicates that the model largely trusts the current measurements, even when subjected to strong electromagnetic interference. At this time This indicates that the model automatically filters out abnormally high dirty data and replaces it with historical predictions. This mechanism fundamentally gives the network robustness against local hardware failures and avoids major false alarm incidents caused by electromagnetic interference.

[0077] Please see Figure 2 This diagram illustrates a cross-modal feature fusion network topology based on a dynamic gating mechanism, according to an embodiment of the present invention. Traditional networks are single-line straight-through, while the network in this embodiment introduces a bypass safety switching mechanism similar to that used in industrial control. The auxiliary branch network on the right acts as a quality judge, calculating gating weights between 0 and 1 by learning electromagnetic features. At the central fusion node, when When strong electromagnetic interference occurs, the network will automatically truncate the actual gas phase feature vector on the left side that is contaminated. The signal path smoothly switches to the predicted gas phase feature vector. Alternative pathways are used to block the spread of pseudo-mutation signatures.

[0078] Please see Figure 3 The diagram illustrates a comparison of the anti-disturbance smoothing effect of the dynamic gating mechanism provided in an embodiment of the present invention. As can be seen from the diagram, the measured gas phase characteristic curve exhibits an extremely high peak at a certain moment of contamination (sudden dirty data), while the predicted gas phase characteristic curve is a dotted line that maintains a steady upward trend or remains stable based on historical trends. At the same time, when the gas phase data is contaminated and the peak appears, the value of the gating weight spikes sharply to close to 1. By using a weighted fusion method, the predicted gas phase characteristics are adopted and the interfered measured gas phase characteristics are automatically shielded. As can be seen from the diagram, the fused clean gas phase characteristic curve coincides with the measured curve most of the time, but at the peak moment, it smoothly switches and fits the predicted curve.

[0079] Step S3: Fault classification mapping and diagnostic output based on cleaning features.

[0080] High-dimensional feature vectors extracted from deep network layers This is unreadable gibberish to human engineers; it must be reverse-mapped back into the standard discourse of the electrical engineering field to generate engineering application value. This invention embodiment uses clean gas phase feature vectors... The sample is fed into a classifier consisting of a hidden layer (64 nodes, ReLU activation) and an output layer (the number of nodes is consistent with the preset number of transformer fault types; in this embodiment, the transformer fault types are set to 6 categories: normal, low-energy discharge, high-energy discharge, partial discharge, medium-low temperature overheating, and high-temperature overheating). Finally, the probability that the current sample belongs to the k-th type of fault is calculated using the Softmax function. Since different types of transformer faults often exhibit highly nonlinear coupling in the feature space (for example, the gas production ratios of high-energy discharge and high-temperature overheating are extremely similar at certain stages), this embodiment of the invention uses the ReLU activation function to provide sufficient nonlinear segmentation capability and avoid gradient vanishing. At the same time, the Softmax layer is used to convert the absolute feature distance into a relative probability distribution with a sum of 1, thereby effectively stripping away the boundaries of coupled fault modes and transforming the internal numerical calculation into a clear normalized probability of occurrence. This provides a direct basis in mathematical statistics for on-site dispatchers to set alarm confidence thresholds (e.g., shutdown is only allowed if the probability is >80%).

[0081] After obtaining the probability of the gas phase data at the current sampling time belonging to various transformer faults through the above process, the category corresponding to the maximum probability can be selected as the final classification result. For example, if the probability of belonging to the low-energy discharge fault is the maximum, then the transformer is finally determined to be in a low-energy discharge state. While outputting the fault diagnosis result, the system console can also synchronously display the gating weight at the current sampling time through the log. and its corresponding electromagnetic eigenvectors Staff members displayed the gate control weights and its corresponding electromagnetic eigenvectors This allows for a direct assessment of the degree of electromagnetic interference affecting the transformer, thus addressing the lack of interpretability in deep learning. For example, when the model intercepts an abnormal spike in data, operations and maintenance experts can clearly determine by reviewing the logs that "the alarm was intercepted because the electromagnetic probe captured an electromagnetic shock at the same time ( This mechanism greatly enhances the production teams' trust in the intelligent algorithm and its interpretability.

[0082] Step S4: Offline training phase coupled with forced noise enhancement.

[0083] In order for the above cross-modal feature fusion network to truly have gating assignment capability, it is also necessary to pair it with a loss function with strong constraints and targeted training data to avoid feature collapse during network iteration, which would cause the fusion mechanism to lose its expected effect.

[0084] First, the loss function is defined in this embodiment of the invention. for:

[0085]

[0086] in, The standard cross-entropy loss; This represents the mean squared error loss of the autoregressive branch (i.e., the main branch network). Defined as (The predicted features should approximate the true noise-free features as closely as possible), where The ground truth values ​​in the training set, i.e., noise-free true features, can be obtained by selecting absolutely clean data from historical databases that have been confirmed by experts, are in a stable operating period, and have no external operation records. This set of clean data is then directly processed through the first long short-term memory network of the main branch network to extract features. The extracted feature vector is the noise-free true feature. (i.e., truth value), and It only exists during the training and validation phases of the network. The predicted gas phase feature vector, i.e., the predicted feature, is extracted through the second long short-term memory network of the main branch network; The L1 regularization penalty applied to the gating weights aims to force the network to be lazy, that is, to force the gating weights to remain at a value of 0 as much as possible under normal circumstances, preventing the network from lazily using predicted values ​​instead of actual measured values. In this embodiment of the invention, a first preset weight parameter is preferred. The preferred second preset weight parameter .

[0087] The loss function precisely balances the algorithm's "robustness against external interference" and "sensitivity to tracking actual internal gas production," ensuring that the false alarm prevention mechanism is only activated during real environmental crises. During the transformer's stable service period, the model still maintains the highest sensitivity in tracking early trace fault gases, eliminating the underreporting of real faults caused by excessive smoothing.

[0088] Meanwhile, in the substation historical database, there are very few training samples that simultaneously meet the criteria of "real transformer fault" and "exactly strong electromagnetic interference". To avoid network underfitting, artificial data augmentation is necessary. The specific process is as follows: 1. Extract a normal continuous DGA sequence (gas phase sequence) and the corresponding stationary UHF background sequence (electromagnetic waveform signal) from the historical database; 2. At random moments in the gas phase sequence... Forcibly injecting a Gaussian pulse with a preset amplitude multiple (e.g., a Gaussian pulse with an amplitude 50 times the normal value) into the hydrogen and carbon monoxide concentration data causes electromagnetic interference to the analog circuit; 3. At the same time The electromagnetic eigenvector of the corresponding auxiliary mode input Force the pulse frequency and energy integral Replace with the historical maximum value corresponding to the preset multiple (e.g., the historical maximum value at 1.5 times).

[0089] This strongly coupled training method forces the neural network to understand the only loss reduction path when updating gradients through backpropagation: as long as the electromagnetic eigenvectors... If the value is extremely large, the gating weight must be increased. The value is used to block the actual gas phase eigenvector with abrupt noise. The propagation mechanism enables the implementation of the anti-disturbance technology mechanism set in the embodiments of the present invention, while significantly improving the generalization immunity of the model when facing completely unknown cross-power station background noise in the future.

[0090] Example 2

[0091] To verify the effectiveness of the method described in Embodiment 1 and its advancement compared to the prior art, this embodiment extracts 12 hours of continuous monitoring data of a main transformer in a 500kV hub substation under actual operating conditions as a test set.

[0092] Test subjects and scenario settings:

[0093] Comparison Method (Traditional Method): This method employs the widely used "moving average filter preprocessing + BP neural network classifier". It utilizes only historical gas phase data for averaging and smoothing, then directly inputs the data into the BP network for fault identification.

[0094] The method of this invention is the method described in Example 1.

[0095] Real Events Timeline:

[0096] At time T1 (08:00): The transformer is operating normally, with no internal faults and a quiet external electromagnetic environment.

[0097] At time T2 (12:00): The transformer was operating normally, but the 500kV disconnector next to the transformer performed a switching operation, which generated intense high-frequency electromagnetic radiation. This radiation directly penetrated the chromatograph's shielding box, causing a millisecond-level latching abnormality in the analog-to-digital conversion circuit.

[0098] At time T3 (16:00): the disconnecting switch operation ended and the environment returned to quiet, but an early low-energy floating potential discharge actually occurred inside the transformer.

[0099] Comparison process between measured data and diagnosis:

[0100] 1. T1 Normal Operating Time (Verifying Basic Diagnostic Capabilities)

[0101] Physical data acquisition:

[0102] Dominant Mode (DGA): The hydrogen concentration was 12.1 μL / L and the total hydrocarbon concentration was 25.4 μL / L, both within the absolute normal baseline.

[0103] UHF (Upper High Frequency): Background noise level. Extremely low.

[0104] Traditional method results: The BP network input features are normal, and the output classification result is: Normal operation.

[0105] This invention demonstrates that, due to the extremely low input of the auxiliary modality, the confidence-gated weights are calculated from the auxiliary branch of the MLP. Fusion features The final network output classification result is: Normal operation, both are correct.

[0106] 2. T2 Electromagnetic Interference Time (Verification of Core Disturbance Immunity and False Alarm Prevention Capabilities)

[0107] Physical data acquisition:

[0108] Dominant Mode (DGA): Due to the sensor circuit being subjected to strong electromagnetic wave impact, the raw data output... The concentration suddenly jumped to 358 μL / L. Acetylene suddenly appeared and reached 15 μL / L.

[0109] UHF modality: Within a 10-minute time window prior to the chromatographic sampling, the UHF sensor captures a high-frequency pulse group, characterized by... and energy integral It reached more than 400 times the normal baseline value.

[0110] Traditional methods, despite the existence of moving average filtering, are still insufficient when faced with data exhibiting sudden changes of tens of times. Feature vectors containing high concentrations of hydrogen and acetylene are fed into a backpropagation (BP) neural network. Based on learned chemical rules, the BP network immediately identifies an extremely dangerous arc fault, outputting a classification result of "high-energy arc discharge," and triggering the highest-level red alarm at the substation.

[0111] Actual result: Serious false alarms, which will lead to emergency response by the maintenance team, and may even cause unnecessary power outages for transformer maintenance, resulting in huge economic losses.

[0112] This invention is characterized by:

[0113] The secondary branch receives a sharply increased UHF signature. Then, calculate the gate control weights. .

[0114] During fusion computing, The network forcibly blocked the propagation of mutation data, and the feature representation regressed to a stable state based on the trend prediction at time T1.

[0115] The final output classification result is: Normal operation. The background log also includes a note: Strong electromagnetic interference detected. The current chromatographic sample is contaminated, and a predictive model is used for smoothing.

[0116] Actual result: This major false alarm incident caused by limitations in sensor hardware was avoided.

[0117] 3. The actual time of T3 failure (verifying physical sensitivity and the ability to prevent false negatives)

[0118] Physical data acquisition:

[0119] Dominant Mode (DGA): Actual discharge inside the transformer, leading to The concentration climbed to 85 μL / L within 4 hours, and The actual output was 2.5 μL / L.

[0120] UHF (Underground High Frequency): The external environment is quiet, and the switching operation has ended. However, because the discharge is located deep within the transformer tank, the external UHF probe only detects a very weak signal spike (not reaching the interference threshold).

[0121] Traditional method performance: Due to the huge data spikes experienced in the previous time step (T2), the moving average filtering algorithm in the traditional method produces a severe tailing effect while smoothing the T2 data. This causes the real fault data features at time step T3, which are not extremely concentrated, to be masked by the previous filtering residuals. The input features of the BP network are blurred, and the final output classification result is: normal operation (or anomaly with extremely low confidence).

[0122] Actual result: Missed detection occurred, latent faults were not discovered, and there is a risk of breakdown.

[0123] This invention is characterized by:

[0124] At this point, the secondary modality features revert to normal levels, and the MLP branch calculates the gate control weights. .

[0125] The network determines that the current data collection environment is safe. Reintroducing a large proportion of current actual chromatographic measurements .

[0126] The deep learning classifier accurately captured the abnormal characteristics of hydrogen and acetylene, outputting the classification result as: low-energy discharge, and prompting personnel to retest.

[0127] Actual results: Accurate alarms. The solution of this invention does not employ blind mathematical filtering, thus avoiding the tailing effect of historical contaminated data and ensuring the model's high sensitivity to real faults.

[0128] The table below visually compares the performance of the two methods during the testing period:

[0129]

[0130] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0131] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for dynamic denoising and fault classification of transformer DGA data that integrates multimodal environmental features, characterized in that, The method includes: The concentrations of various characteristic gases and electromagnetic signals in the transformer are synchronously acquired and preprocessed with time window alignment to obtain gas phase sequences and electromagnetic feature vectors. A cross-modal feature fusion network based on a dynamic gating mechanism is constructed. Through the cross-modal feature fusion network, actual gas phase feature vectors and predicted gas phase feature vectors are extracted based on the gas phase sequence. At the same time, gating weights are calculated based on the electromagnetic feature vectors. The actual gas phase feature vectors and the predicted gas phase feature vectors are then weighted and fused based on the gating weights to obtain clean gas phase feature vectors. The clean gas phase feature vector is fed into a classifier for classification and mapping to obtain the fault diagnosis results of the transformer.

2. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The acquisition of the gas phase sequence and electromagnetic feature vector includes: According to the preset sampling period, the concentration values ​​of various characteristic gases at the current sampling time and multiple previous historical times are obtained, multiple gas phase concentration vectors are constructed, and the multiple gas phase concentration vectors are spliced ​​together in chronological order to form the gas phase sequence. Continuously record the transient electromagnetic waveform signals around the transformer; Using the current sampling time as a reference, a time window of a preset time length is extracted in the direction of historical time. Within the time window, the key statistical features of the transient electromagnetic waveform signal are calculated, and the key statistical features are normalized and then concatenated to obtain the electromagnetic feature vector aligned to the current sampling time.

3. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 2, characterized in that, The key statistical features include the maximum pulse amplitude, the number of pulses exceeding a preset threshold, and the high-frequency energy integral. The electromagnetic feature vector is composed of the normalized maximum pulse amplitude, the pulse frequency, and the high-frequency energy integral.

4. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The extraction of actual gas phase feature vectors and predicted gas phase feature vectors includes: The gas phase sequence is input into the first long short-term memory network in the main branch network, and the hidden state of the last time step is taken as the actual gas phase feature vector representing the true measurement value. The remaining historical sequence after removing the elements collected at the current sampling time from the gas phase sequence is input into a parallel second long short-term memory network, and the theoretical prediction value for the current sampling time is output as the predicted gas phase feature vector.

5. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The calculation of the gating weights includes: The electromagnetic feature vector is input into the multilayer perceptron in the auxiliary branch network, and the gating weights representing the confidence of data contamination are output by the activation function in the last layer. The gating weights are calculated as follows: the electromagnetic feature vectors are weighted and summed and biased using the multilayer perceptron to obtain intermediate feature values, and the intermediate feature values ​​are mapped using an activation function to obtain the gating weights.

6. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The obtained clean gas phase feature vector includes: Calculate the difference between the value 1 and the gate weight, and multiply the difference by the actual gas phase feature vector to obtain the first weighted feature; The gating weights are multiplied by the predicted gas phase feature vector to obtain the second weighted feature; The clean gas phase feature vector is obtained by performing a vector addition operation between the first weighted feature and the second weighted feature.

7. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The obtained fault diagnosis results for the transformer include: The clean gas phase feature vector is fed into the classifier consisting of a hidden layer and an output layer; The output layer features are converted into a relative probability distribution of the current sample belonging to each preset transformer fault type through the Softmax layer; The fault category corresponding to the highest probability value in the relative probability distribution is selected as the final fault diagnosis result.

8. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The method further includes an offline training phase and a forced noise coupling enhancement step. In the offline training phase, the overall loss function of the cross-modal feature fusion network is: Calculate the standard cross-entropy loss of the classification output; Calculate the mean squared error loss between the predicted features output by the main branch network and the noise-free true features, and weight the mean squared error loss using a first preset weight parameter; Calculate the regularization penalty value applied to the gate weight, and weight the regularization penalty value using the second preset weight parameter; The standard cross-entropy loss, the weighted mean squared error loss, and the weighted regularization penalty are summed to obtain the overall loss function used for model gradient update.

9. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 8, characterized in that, The forced noise coupling enhancement step includes: A normal continuous gas phase sequence and its corresponding stable electromagnetic waveform signal are extracted from the historical database as the initial training sample. At random moments in the continuous gas phase sequence, Gaussian pulses with a preset amplitude multiple are forcibly injected into the concentration data of the characteristic gas to simulate sudden interference to the sensor circuit. In the electromagnetic feature vector corresponding to the same random moment, the pulse frequency and high-frequency energy integral are forcibly replaced with the corresponding historical maximum values ​​by a preset multiple, generating strong noise training samples with causal correspondence and inputting them into the network for training.

10. The method for dynamic denoising and fault classification of transformer DGA data based on the fusion of environmental multimodal features as described in claim 1, characterized in that, The method further includes an interpretable log display step based on the gating weight: While outputting the final fault diagnosis result, the gating weight and its associated electromagnetic feature vector are extracted and recorded simultaneously when judging the current sample. When the gating weight reaches or exceeds the interception threshold, a prompt message is output on the system console, indicating that the current chromatographic sample is judged as dirty data due to strong electromagnetic interference, and the system has automatically switched to using the predicted gas phase feature vector for smoothing.

Images