Denoising method for monitoring signal of transformer substation hardware fitting and related device

Abstract

The invention provides a denoising method and related device for a transformer substation hardware monitoring signal, and the method comprises the steps: obtaining the transformer substation hardware monitoring signal, carrying out the signal preprocessing, and obtaining a preprocessed monitoring signal; inputting the monitoring signal into a pre-trained denoising model for denoising processing so as to obtain a denoised transformer substation hardware monitoring signal; wherein the denoising model extracts local features and global features of a signal from a monitoring signal through a dual-channel encoder; then, the local features and the global features are subjected to denoising processing and then spliced, and fusion features are obtained; performing feature reconstruction processing on the fusion features through a dual-channel decoder to obtain local reconstruction features and global reconstruction features, and splicing the local reconstruction features and the global reconstruction features to obtain final optimization features; and reconstructing the monitoring signal so as to realize de-noising processing. According to the method, the local and global features are extracted respectively, and the noise in the monitoring signal of the transformer substation hardware fitting in the complex noise environment can be removed more thoroughly through denoising, feature reconstruction and other processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of denoising of hardware monitoring signals of power system, and in particular relates to a denoising method and related devices for hardware monitoring signals of transformer substations. Background Art

[0002] In power system operations, the stable operation of substations is crucial to ensuring power supply. Hardware, as key components connecting power equipment, directly impacts the reliability and safety of power transmission. Therefore, monitoring the condition of substation hardware is a crucial task in power system maintenance and management.

[0003] Substation hardware generates a variety of monitoring signals during operation, such as vibration and temperature signals. These signals are important indicators of the hardware's operating status, but in actual data collection, due to the complex operating environment of the substation, these signals are easily interfered with by various noises.

[0004] Existing signal denoising methods have the problem of incomplete denoising when processing substation hardware monitoring signals in complex noisy environments, resulting in the processed signals still containing a lot of noise, affecting the subsequent accurate judgment of the hardware status; at the same time, there is also the problem of signal feature loss, resulting in the inability to fully obtain the operating status information of the hardware from the denoised signals, and unable to meet the actual application needs of the power system for accurate monitoring of the hardware status and fault diagnosis. Summary of the Invention

[0005] In view of this, the present invention aims to provide a denoising method and related device for substation hardware monitoring signals, so as to effectively remove various noises in the signals while retaining key detail information to the greatest extent, providing a more reliable basis for power system status monitoring and fault diagnosis, and improving the intelligence and efficiency of power system signal processing.

[0006] In order to achieve the above object, the technical solution provided by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for denoising a substation hardware monitoring signal, comprising the following steps:

[0008] Obtain substation hardware monitoring signals and perform signal preprocessing to obtain preprocessed monitoring signals;

[0009] The monitoring signal is input into the pre-trained denoising model for denoising, thereby obtaining the denoised substation hardware monitoring signal;

[0010] Among them, the denoising process of the denoising model includes:

[0011] The local features and global features of the signal are extracted from the monitoring signal through a dual-channel encoder; the dual-channel encoder includes a local denoising channel and a global denoising channel, which are used to extract local features and global features respectively;

[0012] De-noising the local features and global features and then splicing them together to obtain fused features;

[0013] Through the dual-channel decoder, the fusion features are reconstructed to obtain local reconstruction features and global reconstruction features, and the local reconstruction features and global reconstruction features are spliced ​​to obtain the final optimized features for denoising;

[0014] The monitoring signal is reconstructed according to the final optimized features to achieve denoising.

[0015] Furthermore, in the dual-channel encoder, the local denoising channel is a multi-scale convolution local denoising channel, and the process of extracting local features by the multi-scale convolution local denoising channel includes:

[0016] The monitoring signal is processed in sequence by several Conv1D blocks to obtain the first feature, and the first feature is used as the local feature;

[0017] Among them, each Conv1D block includes a one-dimensional convolution, a batch normalization layer, an activation layer using the LeakyReLU function, and a maximum pooling layer in sequence, and the size of the one-dimensional convolution decreases in sequence, and the number of channels in the downsampling process of the maximum pooling layer increases in sequence.

[0018] Furthermore, in the dual-channel encoder, the global denoising channel is an improved bidirectional long short-term memory network global denoising channel, and the improved bidirectional long short-term memory network global denoising channel extracts global features based on the DM-LSTM network.

[0019] Furthermore, the local features and global features are denoised, including:

[0020] The global periodicity features of the original signal are extracted from the global features through the global average pooling operation, and the local impulse features of the original signal are extracted from the local features through the maximum pooling operation;

[0021] According to the extracted global periodic features and local impulse features, the global denoising threshold and the local denoising threshold are constructed respectively;

[0022] The local features and the global features are denoised according to the local denoising threshold and the global denoising threshold respectively to obtain the local denoising features and the global denoising features.

[0023] Furthermore, the denoised local features and global features are spliced ​​to obtain fused features, including:

[0024] The local denoising features and the global denoising features are spliced ​​to obtain preliminary fusion features; the local denoising features and the global denoising features are obtained based on the local features and the global features after denoising;

[0025] The preliminary fusion features are processed with channel attenuation convolution and identity mapping to obtain the final fusion features.

[0026] Furthermore, the dual-channel decoder includes a local path, which obtains local reconstruction features through the attention-gated SwinTransformer structure, including:

[0027] The features of different scales from the fused features and the query values ​​obtained by the convolution operation are input into the attention-gated SwinTransformer structure, and processed multiple times by the attention unit and SwinTransformer module in sequence to obtain the local reconstructed features;

[0028] The processing flow of the attention unit includes:

[0029] Convolution operations are performed on features and query values ​​of different scales to obtain new multi-scale features and query values;

[0030] Based on the new multi-scale features and query values, the attention gate output is calculated through the attention gate mechanism;

[0031] The multi-scale features and attention gate outputs are concatenated and then convolution is performed to generate the attention unit output;

[0032] The processing flow of the SwinTransformer module includes:

[0033] Perform feature concatenation, block vectorization, and linear embedding operations on the attention unit output;

[0034] Perform layer normalization and multi-head self-attention operations on the linearly embedded features;

[0035] The processed features are output to the next attention unit for processing by subsequent attention units and SwinTransformer modules.

[0036] Furthermore, the dual-channel decoder includes a global path, which obtains global reconstruction features through fine-tuning blocks, including:

[0037] The features from the output of the attention-gated SwinTransformer structure are input into the fine-tuning block and processed through two attention gates respectively to obtain the output of the first attention gate and the output of the second attention gate;

[0038] Input the output of the first attention gate and the output of the second attention gate into the SwinTransformer module for processing and output;

[0039] The features output by the SwinTransformer module are concatenated with the features output by the attention-gated SwinTransformer structure to obtain the global reconstruction features.

[0040] In a second aspect, the present invention further provides a denoising device for substation hardware monitoring signals, comprising:

[0041] The signal preprocessing module is used to obtain the substation hardware monitoring signal and perform signal preprocessing to obtain the preprocessed monitoring signal;

[0042] The signal denoising module is used to input the monitoring signal into the pre-trained denoising model for denoising, thereby obtaining the denoised substation hardware monitoring signal;

[0043] Among them, the signal denoising module includes:

[0044] The feature extraction submodule is used to extract local features and global features of the signal from the monitoring signal through a dual-channel encoder; the dual-channel encoder includes a local denoising channel and a global denoising channel, which are used to extract local features and global features respectively;

[0045] The feature processing submodule is used to denoise local features and global features and then splice them to obtain fused features;

[0046] The feature reconstruction submodule is used to perform feature reconstruction on the fused features through a dual-channel decoder to obtain local reconstruction features and global reconstruction features, and then splice the local reconstruction features and the global reconstruction features to obtain the final optimized features for denoising;

[0047] The signal reconstruction submodule is used to reconstruct the monitoring signal according to the final optimized features, thereby achieving denoising processing.

[0048] In a third aspect, the present invention provides a computer device, comprising a processor and a memory:

[0049] The memory is used to store computer programs and send instructions of the computer programs to the processor;

[0050] The processor executes the method for denoising a substation hardware monitoring signal according to the first aspect according to the instructions of the computer program.

[0051] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, a method for denoising a substation hardware monitoring signal according to the first aspect is implemented.

[0052] In summary, the present invention provides a denoising method for substation hardware monitoring signals, comprising obtaining substation hardware monitoring signals and performing signal preprocessing to obtain preprocessed monitoring signals; inputting the monitoring signals into a pre-trained denoising model for denoising processing, thereby obtaining denoised substation hardware monitoring signals; wherein the denoising processing flow of the denoising model comprises: extracting local features and global features of the signal from the monitoring signals respectively through a dual-channel encoder; the dual-channel encoder comprises a local denoising channel and a global denoising channel, which are respectively used to extract local features and global features; denoising the local features and the global features and then splicing them to obtain fused features; performing feature reconstruction processing on the fused features through a dual-channel decoder to obtain local reconstructed features and global reconstructed features, and splicing the local reconstructed features and the global reconstructed features to obtain the final optimized features for denoising; reconstructing the monitoring signals according to the final optimized features, thereby achieving denoising processing. The present invention can more thoroughly remove the noise in the substation hardware monitoring signals in a complex noise environment by respectively extracting local and global features through a dual-channel encoder, and performing denoising and feature reconstruction processing, thereby reducing the noise residue of the processed signal. At the same time, this method effectively avoids the loss of signal features, so that the denoised signal can fully reflect the operating status information of the hardware, meeting the actual needs of the power system for accurate monitoring of the hardware status and fault diagnosis.

[0053] The present invention also provides a denoising device, computer equipment and computer-readable storage medium for monitoring signals of transformer substation hardware, which, when implemented, have similar effects to the above-mentioned denoising method. Detailed description is omitted here. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A flowchart of a method for denoising substation hardware monitoring signals provided by an embodiment of the present invention;

[0056] Figure 2 This is an overall block diagram of a method for denoising substation hardware monitoring signals based on an improved dual-channel interactive denoising autoencoder provided by an embodiment of the present invention;

[0057] Figure 3 Schematic diagram of the multi-scale convolution local denoising channel provided by an embodiment of the present invention;

[0058] Figure 4DM-LSTM principle diagram provided for an embodiment of the present invention;

[0059] Figure 5 Schematic diagram of the AGDMod attention module provided by an embodiment of the present invention;

[0060] Figure 6 Schematic diagram of AST provided by an embodiment of the present invention;

[0061] Figure 7 A schematic diagram of a fine-tuning block provided in an embodiment of the present invention;

[0062] Figure 8 The internal structure diagram of SwinTransformer provided by the embodiment of the present invention;

[0063] Figure 9 A block diagram of a device for denoising substation hardware monitoring signals provided by an embodiment of the present invention;

[0064] Figure 10 A block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purposes, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0066] See also Figure 1 The embodiment of the present invention provides a method for denoising a substation hardware monitoring signal, comprising the following steps:

[0067] S11: Obtain the substation hardware monitoring signal and perform signal preprocessing to obtain a preprocessed monitoring signal.

[0068] It should be noted that signal preprocessing usually includes operations such as filtering (such as Butterworth filtering to remove high-frequency noise), normalization (scaling the signal amplitude to a specific range), and resampling (unifying the sampling frequency). The purpose is to eliminate obvious interference in the original signal and standardize the format.

[0069] Specifically, before denoising, the collected hardware monitoring signals must first be preprocessed. This step primarily removes DC components, which can interfere with subsequent analysis. Simulated noise is also added to simulate complex, real-world noise environments. This preprocessed signal more realistically reflects the effects of noise in actual operation, providing suitable input for subsequent denoising operations.

[0070] S12: Inputting the monitoring signal into a pre-trained denoising model for denoising, thereby obtaining a denoised substation hardware monitoring signal.

[0071] The denoising process of the denoising model includes:

[0072] S12-1: extracting local features and global features of the signal from the monitoring signal through a dual-channel encoder; the dual-channel encoder includes a local denoising channel and a global denoising channel, which are used to extract the local features and the global features respectively.

[0073] It should be noted that the dual-channel encoder is a deep learning-based neural network structure consisting of two parallel branches (a local denoising channel and a global denoising channel). Local features are detailed features of the signal within a short time window (such as the rising edge slope of a discharge pulse and high-frequency oscillation components). Local neighborhood information can be extracted through convolutional layers. Global features are overall signal trends or long-term dependencies (such as periodic discharge intervals and amplitude envelope changes). Global context can be captured through Transformer encoders or long-short-term memory (LSTM) networks.

[0074] This step can use a multi-layer convolutional neural network (such as ResNet residual blocks) to build a local denoising channel. Using convolution kernels of different scales (such as 3×3 and 5×5), you can extract multi-resolution local features and capture local differences between noise and signal. Alternatively, you can use a self-attention mechanism (such as the Transformer) or a temporal network (such as an LSTM) to build a global denoising channel. This calculates dependencies between signal points and identifies global noise patterns (such as periodic interference).

[0075] This step separates local and global features in parallel, avoiding the entanglement of noise and signal features caused by the traditional single-channel "mixed extraction" of multi-scale features.

[0076] S12-2: De-noising the local features and the global features and then concatenating them to obtain fused features.

[0077] It should be noted that denoising is the process of suppressing noise in extracted features, which can be achieved through soft thresholding, attention mechanism weight adjustment, or noise discrimination using generative adversarial networks. Feature concatenation involves concatenating local and global features in the channel or spatial dimensions (e.g., the Concat operation) to form a fused feature containing multi-dimensional information.

[0078] S12-3: Perform feature reconstruction processing on the fused features through a dual-channel decoder to obtain local reconstruction features and global reconstruction features, and splice the local reconstruction features and the global reconstruction features to obtain final optimized features for denoising.

[0079] It should be noted that the dual-channel decoder includes local and global reconstruction channels, which are used to restore fused features to reconstructed features in the signal space. Feature reconstruction involves mapping abstract features back to the scale of the original signal through deconvolution (transposed convolution) or upsampling operations, restoring local details (such as pulse waveforms) and global trends (such as the signal envelope).

[0080] S12-4: Reconstruct the monitoring signal according to the final optimized feature, thereby achieving denoising processing.

[0081] It should be noted that signal reconstruction involves mapping the final optimized features into a time-domain signal through fully connected or convolutional layers, completing the transformation from feature space to signal space. Denoised signals can be directly generated from the decoder's output features. A loss function is used to constrain the difference between the reconstructed signal and the true clean signal, ensuring that noise is suppressed while retaining useful features.

[0082] This embodiment provides a denoising method for substation hardware monitoring signals, which realizes denoising of substation hardware monitoring signals in complex noise environments through signal preprocessing and a dual-channel encoding and decoding network architecture. First, the original signal is preprocessed to a standardized format, and then a dual-channel encoder (local denoising channel and global denoising channel) is used to extract the local features of the signal (such as high-frequency details or local noise patterns at a specific time point) and global features (such as the overall trend of the signal, periodic laws). After independent denoising processing, the two types of features are spliced ​​together to form a fusion feature, and then the local and global features are reconstructed separately by a dual-channel decoder. Finally, noise suppression is achieved through feature splicing and signal reconstruction. Its core principle is to separate and fuse features of different scales, while retaining the key information of the signal, to eliminate local noise and global interference in a targeted manner, thereby solving the problems of incomplete denoising and feature loss caused by single-scale processing in traditional methods.

[0083] See also Figure 2 , Figure 2This paper illustrates a method for denoising substation hardware monitoring signals using an improved dual-channel interactive denoising autoencoder, based on the aforementioned embodiments. This method first performs signal preprocessing, then performs feature extraction using a multi-scale convolutional local denoising channel and a DM-LSTM (Derivative Memory-Long Short-Term Memory) global denoising channel in parallel. The two types of features are then interactively communicated through an interaction and communication mechanism. The two features are then processed by a dual-channel decoder, followed by model training and optimization, ultimately outputting a denoised signal. This method is further described below in conjunction with other embodiments of the present invention.

[0084] In one embodiment of the present invention, in a dual-channel encoder, the local denoising channel is a multi-scale convolution local denoising channel, and the process of extracting local features by the multi-scale convolution local denoising channel includes:

[0085] The monitoring signal is processed in sequence by several Conv1D blocks to obtain the first feature, and the first feature is used as the local feature;

[0086] Among them, each Conv1D block includes a one-dimensional convolution, a batch normalization layer, an activation layer using the LeakyReLU function, and a maximum pooling layer in sequence, and the size of the one-dimensional convolution decreases in sequence, and the number of channels in the downsampling process of the maximum pooling layer increases in sequence.

[0087] like Figure 3As shown, the local denoising channel of the dual-channel encoder in this embodiment uses a multi-scale convolutional local denoising channel (CL-Path) to extract local features. This channel is designed with multiple Conv1D blocks (six in the specific implementation). Each block consists of a one-dimensional convolution, a batch normalization (BN) layer, an activation layer using the LeakyReLU function, and a max pooling layer. The BN layer helps accelerate model convergence, the LeakyReLU activation function introduces nonlinearity, and the max pooling layer reduces data dimensionality and computational complexity. A variety of convolution kernel sizes, including 13×1, 11×1, 9×1, 7×1, 5×1, and 3×1, are used for different convolution operations. In the initial stages of the network, larger convolution kernels capture more comprehensive local information. As the network depth increases, smaller convolution kernels enable the model to more accurately focus on fine-grained details and extract more complex features. During each downsampling process, the number of feature channels doubles, using 16, 32, 64, 128, 256, and 512 convolution kernels, respectively. Since smaller convolution kernels capture less information per operation, increasing the number of filters compensates for this limitation, ensuring robust feature extraction at each layer and improving the robustness and stability of the model. Finally, through feature concatenation, local features extracted at different scales are fused to obtain a more comprehensive local feature representation. For example, after a recurrent cycle of six Conv1D blocks, a hidden layer representation with dimensions of 16×1, 1 / 64 the size of the original signal, is obtained as the first feature, the local feature. This multi-scale hierarchical design better extracts and preserves local signal features and suppresses local noise.

[0088] In one embodiment of the present invention, in a dual-channel encoder, the global denoising channel is an improved bidirectional long short-term memory network global denoising channel, and the improved bidirectional long short-term memory network global denoising channel extracts global features based on a DM-LSTM network.

[0089] In this embodiment, the improved bidirectional long short-term memory network consists of two LSTM layers, processing forward and backward time series information respectively. Each BiLSTM (Bidirectional Long Short-Term Memory) layer is followed by a max pooling layer to gradually compress the time dimension of the signal. Furthermore, a Derivative Memory Gate (DMGate) is introduced to enhance the ability to extract global dynamic features from monitoring signals.

[0090] This gate can capture the dynamic changes of the input variables. DM-LSTM is calculated by the following formula:

[0091] (1)

[0092] (2)

[0093] (3)

[0094] (4)

[0095] (5)

[0096] (6)

[0097] (7)

[0098] (8)

[0099] in, , , Represents the input gate, forget gate, and output gate at time t. is the differential memory gate. is a new candidate variable. and is the different time steps for the cell state. and are the hidden layer variables for the current and previous time steps. Represents the weight from the input layer to the corresponding gate or unit layer. Represents the weights from the hidden layer to the corresponding gate or unit layer. Represents the bias term of the corresponding layer.

[0100] The formulas described in formulas (5-8) are the proposed DM gates, which are then integrated into the memory cell layer to effectively capture dynamic information.

[0101] The improved BiLSTM network can capture the dynamic changes of the signal through the derivative memory gate and enhance the modeling ability of time series data. Formula (7) is the unit update state of DM-LSTM.

[0102] like Figure 4As shown, in this embodiment, the global denoising channel of the dual-channel encoder uses a modified bidirectional long short-term memory network (DM-LSTM) to extract global features. Its operation process is as follows: after inputting the current data and the hidden layer variables at the previous time, the input gate, forget gate, candidate cell states, and output gate are calculated in parallel. This is the traditional LSTM calculation process, which controls the input of new information, the forgetting of old information, the generation of candidate states, and the output control. Simultaneously, a derivative memory gate (DMGate) is calculated separately, which is a key difference from the traditional BiLSTM and can capture dynamic changes in the input variables. Subsequently, the input gate and forget gate calculate intermediate cell states, which are then combined with the derivative memory gate to calculate the cell state, comprehensively considering traditional information transmission and signal dynamics. The current hidden layer variables are calculated based on the cell states, and the current hidden layer variables and cell states are finally output. This process achieves more accurate extraction of global signal features (such as periodicity and trend), effectively suppresses global noise, and meets the global information processing requirements for denoising substation hardware monitoring signals.

[0103] In one embodiment of the present invention, an interaction and communication mechanism is proposed. The mechanism implements denoising of local and global features through an Adaptive Global-Local Denoising Module (AGDMod), including:

[0104] S21: The global periodic features of the original signal are extracted from the global features through the global average pooling operation, and the local impulse features of the original signal are extracted from the local features through the maximum pooling operation.

[0105] AGDMod receives multi-scale fault features (i.e., local and global features extracted from a dual-channel encoder) and extracts the global periodic features of the original signal from the global features through a global average pooling operation. This feature reflects the arrangement structure of the signal in the time dimension, such as the periodicity of the power frequency signal. At the same time, it uses a maximum pooling operation to extract the local pulse features of the original signal from the local features. These features reflect the fault manifestations of the signal, such as local mutation information such as discharge pulses.

[0106] S22: Based on the extracted global periodic features and local impulse features, construct a global denoising threshold and a local denoising threshold respectively.

[0107] According to the extracted global periodic features and local impulse features, the threshold coefficients are obtained respectively, and then element-wise multiplication is performed with the input features to construct the global denoising threshold and the local denoising threshold.

[0108] S23: performing denoising processing on the local features and the global features according to the local denoising threshold and the global denoising threshold respectively, to obtain local denoising features and global denoising features.

[0109] Using an adaptive smooth soft thresholding function (ASSTF), local and global features are denoised based on local and global denoising thresholds, respectively, to obtain local and global denoised features. This is because traditional soft thresholding functions can mistakenly eliminate near-zero fault features. By adding a control term λ and adaptively adjusting it, ASSTF can simultaneously prevent the mistaken elimination of valid fault features while denoising. Specifically, the traditional soft thresholding function (STF) can mistakenly eliminate near-zero fault features, resulting in the loss of valid features. To address this issue, the control term λ is added to the traditional STF, converting it into a smooth curve function, resulting in the SSTF (SmoothSoftThresholdingFunction). By adjusting the value of λ, the slope of the curve can be changed, preserving near-zero fault features. To enable the SSTF to adapt to different noise levels and types, an attention-like module is used to adapt λ, ultimately resulting in the ASSTF. The advantage of the ASSTF is that it can adaptively eliminate noise while preventing the mistaken elimination of valid fault features near zero. The calculation formula of SSTF is as follows:

[0110] (9)

[0111] Where: and Represent the input and output of SSTF respectively. Indicates the threshold value. Represents the scale factor. =0 when SSTF represents the traditional STF.

[0112] Although SSTF can alleviate the false elimination of nearly zero effective faults, The setting is also the key task of effective noise elimination nowadays. In order to facilitate the SSTF to make adaptive adjustments to the module according to different noise levels and types, Adaptation of λ. Unlike the attention used to construct the denoising threshold, the attention used for λ adaptation has an additional scaling factor layer. Its purpose is to limit the scaling range of λ, thereby ensuring the denoising ability of SSTF. Finally, SSTF with adaptive λ is called ASSTF. In this embodiment, the scaling factor is set to 0.005.

[0113] like Figure 5As shown, the adaptive global-local denoising module AGDMod proposed in this embodiment processes the input multi-scale fault features in parallel. On the one hand, global periodic features are obtained through global average pooling, and then the threshold coefficient is obtained and the global denoising threshold is obtained by operation with the input features. Then, global denoising is implemented through ASSTF to obtain global denoising features; on the other hand, local pulse features are obtained through maximum pooling, and the local denoising threshold is obtained through similar operations and local denoising is completed through ASSTF to obtain local denoising features. Subsequently, the two types of denoising features are spliced, and then processed by channel attenuation convolution and identity mapping, and finally the denoised features are output. This process uses different pooling operations and adaptive threshold functions to effectively remove noise from global and local dimensions and retain key features.

[0114] In one embodiment of the present invention, the denoised local features and global features are spliced ​​to obtain fused features, including:

[0115] S31: splicing the local denoising features and the global denoising features to obtain preliminary fusion features; the local denoising features and the global denoising features are obtained based on the local features and the global features after denoising;

[0116] S32: Perform channel attenuation convolution and identity mapping on the preliminary fusion features to obtain the final fusion features.

[0117] Please refer again Figure 5 In this embodiment, the denoised global features and local features are concatenated. Then, the cascaded features are processed using channel-attenuated convolution and identity mapping to ensure the consistency of feature dimensions and prevent overfitting. Finally, the output feature Y0 of AGDMod is obtained and expressed as:

[0118] (10)

[0119] (11)

[0120] (12)

[0121] Where: and They represent the feature learning operation with GAP (Global Average Pooling) as the branch of the first network layer and the feature learning operation with MAP (Maximum Pooling) as the branch of the first network layer, respectively. and Represent the global and local features of denoising respectively. represents channel-attenuated convolution.

[0122] In one embodiment of the present invention, the dual-channel decoder includes a local path, which obtains local reconstruction features through an attention-gated SwinTransformer structure, including:

[0123] S41: The features of different scales from the fused features and the query values ​​obtained by the convolution operation are input into the attention-gated SwinTransformer structure, and are processed multiple times by the attention unit and the SwinTransformer module in sequence to obtain the local reconstructed features.

[0124] This embodiment constructs an attention-gated feature fusion module based on the attention-gated SwinTransformer structure. This module is an attention-gated SwinTransformer (AST) structure, which aims to fuse monitoring signal features of different scales and enhance key features through the attention mechanism. It contains multiple parallel attention units, each of which receives and enhances monitoring signal features from different scales and sends them to the SwinTransformer arranged in a hierarchical manner. Each attention unit has two inputs, which come from features of different scales. Assume that after the monitoring signal is pre-processed, the features of different scales are represented as Skn (k represents different feature extraction paths, and n represents the scale level of the feature), and the query value Gkn obtained by the convolution operation.

[0125] See also Figure 6 , the processing flow of the attention unit includes:

[0126] S41-1: Perform convolution operations on the features Skn and query values ​​Gkn of different scales to obtain new multi-scale features Rkn and query values ​​Gkn;

[0127] S41-2: Based on the new multi-scale features and query value, the attention gate output Attkn is calculated through the attention gate mechanism;

[0128] S41-3: Concatenate the multi-scale features Rkn and the attention gate output Attkn, and then generate the attention unit output through convolution operation;

[0129] The processing flow of the SwinTransformer module includes:

[0130] S41-4: Perform feature concatenation, block vectorization, and linear embedding operations on the attention unit output.

[0131] The SwinTransformer module is a model component used to process features. In theory, it mainly involves steps such as feature splicing, block vectorization, linear embedding, feature processing and output. Its structure is as follows Figure 7As shown. Specifically, the input features of different scales are first spliced ​​and converted into a form suitable for subsequent processing. Then a block operation is performed to divide the spliced ​​features into multiple small blocks, each of which is regarded as an independent feature unit. These small blocks are then linearly embedded and mapped to a new feature space through a specific weight matrix to obtain feature representations with different dimensions. Vectorization and Linear Embedding Each block is vectorized and arranged into a matrix PP. Then, the matrix PP is linearly embedded in the linear embedding block using the following formula to obtain the output LE.

[0132] (13)

[0133] Where: W is the weight matrix. LE is the output of the linear embedding module.

[0134] S41-5: Perform layer normalization and multi-head self-attention operations on the features after linear embedding;

[0135] The linearly embedded features are subjected to layer normalization and multi-head self-attention operations. When the SwinTransformer block is processed, LE is input into it. The block first focuses on the local features within each block, and then considers the interactions between different blocks to capture short-range and long-range dependencies. Its operations include: Layer Normalization (LN), which normalizes the input features to accelerate model training and improve stability; Shifted Window Multi-Head Self-Attention (WMSA), which performs multi-head self-attention calculations within each window to capture the correlation of features within the local area; Shifted Window Multi-Head Self-Attention (SWMSA), which realizes the interaction of information between different windows through window shifting operations and expands the range of feature interaction; Residual connection, which adds the input directly to the output to solve the gradient vanishing problem in deep networks and makes the network easier to train; Multi-Layer Perceptron (MLP), with GELU nonlinear activation function, further transforms and extracts features. The operations are expressed as follows:

[0136] (14)

[0137] (15)

[0138] (16)

[0139] (17)

[0140] Where: . T is the output of SwinTransformer.

[0141] S41-6: Output the processed features to the next attention unit for processing by subsequent attention units and SwinTransformer modules.

[0142] Each attention unit consists of two convolutional blocks, an attention gate (AG), and a connection layer. The AG in each attention unit processes mixed input features. Through a series of convolution and activation function operations, an attention weight is calculated. This weight is used to highlight important features and suppress unimportant ones, thereby extracting the feature information that is critical for decoding. Finally, the features processed by the attention gate are concatenated with other features and convolved to obtain the output of the attention unit.

[0143] In one embodiment of the present invention, the dual-channel decoder includes a global path, which obtains global reconstruction features through a fine-tuning block, including:

[0144] S42: Input the features from the output of the attention-gated SwinTransformer structure into the fine-tuning block, and process them through two attention gates respectively to obtain the output of the first attention gate and the output of the second attention gate;

[0145] S43: Input the output of the first attention gate and the output of the second attention gate into the SwinTransformer module for processing and output;

[0146] S44: Connect the features output by the SwinTransformer module with the features output by the attention-gated SwinTransformer structure to obtain the global reconstruction features.

[0147] The main function of the fine-tuning block is to further optimize the features processed by the attention-gated feature fusion module to improve the accuracy and reliability of decoding. The fine-tuning block receives the output features from the attention-gated feature fusion module. The fine-tuning block consists of two attention gates (AG1 and AG2) and a SwinTransformer module. It receives the outputs of two attention units (CL14 and CL24). The two attention gates process them separately to identify significant signal features and prune away irrelevant feature responses in the signal obtained at one scale. The important features at different scales processed by the attention gates are combined in the SwinTransformer module, which further optimizes the features by capturing global and contextual information from the two different scale feature sets to generate the final optimized features for denoising.

[0148] like Figure 8As shown, in this embodiment, the global path of the dual-channel decoder uses a fine-tuning block to obtain global reconstruction features. It first inputs the features CL14 and CL24 output by the attention-gated SwinTransformer structure into the fine-tuning block. In the fine-tuning block, AG1 processes CL14 and outputs Att1, and AG2 processes CL24 and outputs Att2. Then Att1 and Att2 are input to the SwinTransformer module and output after processing. At the same time, CL24 is also directly connected to the output path. Finally, the features output by the SwinTransformer module are connected to CL24 to obtain global reconstruction features. In this process, the two attention gates AG1 and AG2 process the input features respectively, identify significant signal features and prune irrelevant feature responses. The SwinTransformer module further captures global and contextual information to optimize features, realizes effective reconstruction of global features, and provides strong support for the final signal denoising.

[0149] For the denoising model constructed based on the above embodiment, a large amount of hardware monitoring signals and data are collected during the training and optimization process, and are divided into a ratio of 70% training set, 20% test set, and 10% validation set.

[0150] Evaluation metrics used to assess model performance include root mean square error (RMSE), root mean square percentage difference (PRD), signal-to-noise ratio (SNR), and cosine similarity (Cosine Similarity). RMSE measures the error between the original and denoised signals, PRD reflects the relative error, SNR reflects the ratio of signal strength to noise strength, and Cosine Similarity measures the similarity between the denoised signal and the original signal. These metrics evaluate the denoising effect of the model from different perspectives, ensuring a comprehensive and accurate analysis.

[0151] Next, set the training environment and parameters. For example, use the Adam optimizer, set the batch size to 32, the initial learning rate to 0.001, and the number of iterations to 65. These parameters are determined after comprehensively considering model training efficiency and performance, and can effectively optimize the model training process.

[0152] Constructing a loss function to effectively train model parameters and achieve optimal denoising performance. This paper combines the mean absolute error (MAE) and PRD to construct a loss function as follows.

[0153] (18)

[0154] (19)

[0155] (20)

[0156] Where: x and Represents the original signal and the reconstructed signal. Represents the adjustable coefficient.

[0157] Finally, the pre-processed hardware monitoring signal is input into the trained model to output the denoised signal to obtain the denoised signal.

[0158] Based on the same inventive concept, an embodiment of the present application further provides a device for denoising substation hardware monitoring signals, which is used to implement the aforementioned method for denoising substation hardware monitoring signals. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of the embodiments of the device for denoising substation hardware monitoring signals provided below can be found in the aforementioned definition of the method for denoising substation hardware monitoring signals, and will not be further elaborated here.

[0159] See also Figure 9 The embodiment of the present invention further provides a denoising device for monitoring signals of transformer substation hardware, comprising:

[0160] The signal preprocessing module is used to obtain the substation hardware monitoring signal and perform signal preprocessing to obtain the preprocessed monitoring signal;

[0161] The signal denoising module is used to input the monitoring signal into the pre-trained denoising model for denoising, thereby obtaining the denoised substation hardware monitoring signal;

[0162] Among them, the signal denoising module includes:

[0163] The feature extraction submodule is used to extract local features and global features of the signal from the monitoring signal through a dual-channel encoder; the dual-channel encoder includes a local denoising channel and a global denoising channel, which are used to extract local features and global features respectively;

[0164] The feature processing submodule is used to denoise local features and global features and then splice them to obtain fused features;

[0165] The feature reconstruction submodule is used to perform feature reconstruction on the fused features through a dual-channel decoder to obtain local reconstruction features and global reconstruction features, and then splice the local reconstruction features and the global reconstruction features to obtain the final optimized features for denoising;

[0166] The signal reconstruction submodule is used to reconstruct the monitoring signal according to the final optimized features, thereby achieving denoising processing.

[0167] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0168] Reference Figure 10 An embodiment of the present invention further provides a computer device, comprising: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, the denoising method for substation hardware monitoring signals as described in any one of the above methods is implemented.

[0169] The computer device may be a desktop computer, notebook computer, PDA, cloud server or other computing device. The computer device may include, but is not limited to, a processor and a memory. It will be understood by those skilled in the art that Figure 10 The computer device is merely an example and does not constitute a limitation on the computer device. The computer device may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the computer device may also include input and output devices, network access devices, etc.

[0170] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0171] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk equipped with the computer device, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Furthermore, the memory may include both an internal storage unit of the computer device and an external storage device. The memory is used to store an operating system, application programs, a boot loader, data, and other programs, such as the program code of the computer program. The memory may also be used to temporarily store data that has been output or is about to be output.

[0172] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for denoising a substation hardware monitoring signal as described in any one of the above methods is implemented.

[0173] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the process steps in the above-mentioned method embodiments by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, removable hard drives, magnetic disks, or optical disks. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0174] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0175] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0176] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely schematic. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0177] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for denoising substation hardware monitoring signals, characterized in that: The steps include: Obtain substation hardware monitoring signals and perform signal preprocessing to obtain preprocessed monitoring signals; Inputting the monitoring signal into a pre-trained denoising model for denoising, thereby obtaining a denoised substation hardware monitoring signal; The denoising process of the denoising model includes: Extracting local features and global features of the signal from the monitoring signal through a dual-channel encoder; the dual-channel encoder includes a local denoising channel and a global denoising channel, respectively used to extract the local features and the global features; Denoising the local features and the global features and then concatenating them to obtain fused features; Performing feature reconstruction processing on the fused features through a dual-channel decoder to obtain local reconstruction features and global reconstruction features, and splicing the local reconstruction features and the global reconstruction features to obtain final optimized features for denoising; The monitoring signal is reconstructed according to the final optimized feature, thereby achieving denoising processing.

2. The method for denoising substation hardware monitoring signals according to claim 1, characterized in that: In the dual-channel encoder, the local denoising channel is a multi-scale convolution local denoising channel, and the process of extracting the local features by the multi-scale convolution local denoising channel includes: Processing the monitoring signal in sequence through a plurality of Conv1D blocks to obtain a first feature, and using the first feature as the local feature; Among them, each of the Conv1D blocks includes a one-dimensional convolution, a batch normalization layer, an activation layer using the LeakyReLU function, and a maximum pooling layer in sequence, and the size of the one-dimensional convolution decreases in sequence, and the number of channels in the downsampling process of the maximum pooling layer increases in sequence.

3. The method for denoising substation hardware monitoring signals according to claim 1, characterized in that: In the dual-channel encoder, the global denoising channel is an improved bidirectional long short-term memory network global denoising channel, and the improved bidirectional long short-term memory network global denoising channel extracts the global features based on the DM-LSTM network.

4. The method for denoising substation hardware monitoring signals according to claim 1, characterized in that: The local features and the global features are subjected to denoising processing, comprising: Extracting the global periodicity feature of the original signal from the global feature through a global average pooling operation, and extracting the local impulse feature of the original signal from the local feature through a maximum pooling operation; Constructing a global denoising threshold and a local denoising threshold respectively according to the extracted global periodic feature and the local impulse feature; The local features and the global features are denoised according to the local denoising threshold and the global denoising threshold respectively, to obtain local denoising features and global denoising features.

5. The method for denoising substation hardware monitoring signals according to claim 1, characterized in that: The local features and the global features after denoising are combined to obtain fused features, including: Splicing the local denoising features and the global denoising features to obtain preliminary fusion features; the local denoising features and the global denoising features are obtained based on the local features and the global features after denoising; The preliminary fusion features are subjected to channel attenuation convolution and identity mapping processing to obtain the final fusion features.

6. The method for denoising substation hardware monitoring signals according to claim 1, characterized in that: The dual-channel decoder includes a local path, and the local path obtains the local reconstruction features through an attention-gated SwinTransformer structure, including: Input the features of different scales from the fused features and the query value obtained by the convolution operation into the attention-gated SwinTransformer structure, and process them through the attention unit and SwinTransformer module multiple times in sequence to obtain the local reconstructed features; The processing flow of the attention unit includes: Performing convolution operations on the features of different scales and the query value respectively to obtain new multi-scale features and query values; Based on the new multi-scale features and the query value, calculating the attention gate output through the attention gate mechanism; Concatenate the multi-scale features and the attention gate output, and then generate the attention unit output through a convolution operation; The processing flow of the SwinTransformer module includes: Perform feature concatenation, block vectorization, and linear embedding operations on the attention unit output; Perform layer normalization and multi-head self-attention operations on the linearly embedded features; The processed features are output to the next attention unit for subsequent processing by the attention unit and the SwinTransformer module.

7. The method for denoising substation hardware monitoring signals according to claim 6, characterized in that: The dual-channel decoder includes a global path, and the global path obtains the global reconstruction feature through a fine-tuning block, including: Input the features output from the attention-gated SwinTransformer structure into the fine-tuning block, process them through two attention gates respectively, and obtain the first attention gate output and the second attention gate output; Input the first attention gate output and the second attention gate output into the SwinTransformer module for processing and output; The features output by the SwinTransformer module are connected with the features output by the attention-gated SwinTransformer structure to obtain the global reconstruction features.

8. A noise removal device for monitoring signals of transformer substation hardware, characterized in that: include: The signal preprocessing module is used to obtain the substation hardware monitoring signal and perform signal preprocessing to obtain the preprocessed monitoring signal; A signal denoising module is used to input the monitoring signal into a pre-trained denoising model for denoising, thereby obtaining a denoised substation hardware monitoring signal; Wherein, the signal denoising module includes: A feature extraction submodule is used to extract local features and global features of the signal from the monitoring signal through a dual-channel encoder; the dual-channel encoder includes a local denoising channel and a global denoising channel, which are used to extract the local features and the global features respectively; A feature processing submodule, configured to perform denoising on the local features and the global features and then combine them to obtain fused features; A feature reconstruction submodule is used to perform feature reconstruction processing on the fused features through a dual-channel decoder to obtain local reconstruction features and global reconstruction features, and to splice the local reconstruction features and the global reconstruction features to obtain the final optimized features for denoising; The signal reconstruction submodule is used to reconstruct the monitoring signal according to the final optimized feature, thereby achieving denoising processing.

9. A computer device, characterized in that: The device includes a processor and a memory: The memory is used to store the computer program and send instructions of the computer program to the processor; The processor executes the method for denoising a substation hardware monitoring signal according to any one of claims 1 to 7 according to instructions of the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for denoising a substation hardware monitoring signal according to any one of claims 1 to 7 is implemented.

Images