Voltage transformer error prediction method based on modal decomposition and gating circulation unit

Abstract

The invention discloses a voltage transformer error prediction method based on modal decomposition and a gating circulation unit. The method comprises the following steps: 1, collecting historical error data of a capacitor voltage transformer to construct a time sequence data set; 2, dynamically adjusting a noise amplitude coefficient based on signal frequency domain characteristics by adopting an improved modal decomposition method, and performing multiple decomposition to obtain a mean modal component; 3, calculating an information entropy value of each component, classifying the components according to the entropy values and the time-frequency graph, and carrying out noise reduction on the high-entropy components by adopting a self-adaptive wavelet packet threshold value; 4, designing a non-local attention mechanism enhanced error signal prediction model, and dynamically fusing local time sequence features and global context information through a gating coefficient; 5, establishing a sub-modal parallel training framework, performing modeling prediction on each component, and performing fusion output; and 6, performing model performance evaluation and verification to verify the effectiveness of the method. The method effectively improves the prediction precision of the output signal of the voltage transformer, and can effectively meet the signal prediction requirements under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of online monitoring of voltage transformers, and particularly discloses a voltage transformer error prediction method based on modal decomposition and a gated cyclic unit. Background Art

[0002] Voltage transformers (CVTs) are core measurement devices in power systems. Their core function is to achieve precise conversion and electrical isolation between the primary high-voltage system and secondary equipment, providing the necessary voltage signals for key power system functions such as energy metering, relay protection, and system monitoring. With the construction of ultra-high voltage (UHV) power grids and the large-scale integration of energy sources, capacitive voltage transformers (CVTs) face increasingly severe operating environment challenges. On the one hand, harmonic pollution and DC bias caused by AC / DC hybrid systems significantly increase the complexity of the electromagnetic environment. On the other hand, during long-term operation, their key internal components (such as the capacitive voltage divider and electromagnetic unit) are subject to multi-physics coupling effects such as temperature fluctuations and mechanical disturbances, resulting in insulation aging and deterioration of the core's magnetic properties. This leads to a gradual increase in measurement errors, seriously affecting the measurement accuracy and operational reliability of the power system.

[0003] Currently, fault diagnosis for voltage transformers (VTs) primarily relies on a passive approach, involving periodic calibration or post-inspection maintenance, lacking online monitoring and early warning for early-stage performance degradation. Because CVTs typically operate in environments with high voltage, strong electromagnetic interference, and extreme temperatures, the performance degradation of their internal components is a gradual process. If not promptly detected and addressed, this can lead to serious consequences such as metering inaccuracies and protection misoperation, directly impacting the safe and stable operation of the power system. Therefore, studying the error evolution patterns of voltage transformers and developing high-precision signal state prediction methods are crucial for achieving intelligent equipment operation and maintenance, reducing sudden failures, and improving grid reliability.

[0004] Existing methods for predicting voltage transformer status have certain limitations and defects, including the following:

[0005] 1) Mechanism modeling method: Based on the law of electromagnetic induction and equivalent circuit theory, a differential equation or state-space model is established that includes elements such as primary / secondary winding parameters, core excitation impedance, and distributed capacitance to quantitatively analyze the error characteristics of the voltage transformer. However, this method relies too much on idealized assumptions and fails to consider actual factors such as core nonlinearity and insulation aging, resulting in limited model accuracy. At the same time, this method completely ignores transient interference factors such as lightning surges and switching operations, resulting in the model being unable to truly reflect the dynamic characteristics changes of the device in the actual complex operating environment, resulting in significant deviations between the error prediction results and the actual situation.

[0006] 2) Statistical regression: This method uses historical data to establish a linear relationship model between a single variable, such as temperature or load, and the error. While this method is easy to implement, it only reflects the static relationship between a single environmental variable and the error and cannot account for the coupled effects of multiple factors. Furthermore, the model's prediction accuracy decreases dramatically when the operating environment changes significantly. When faced with new operating conditions, the model requires extensive data collection and modeling, resulting in poor adaptability.

[0007] 3) Shallow machine learning methods: These employ algorithms such as BP neural networks and support vector machines to establish error prediction models through training. However, the limitations of this approach's network structure lead to insufficient modeling of instantaneous state-owned enterprise dependencies, requiring repeated training under different operating conditions. Furthermore, the cumulative effect of errors is significant in multi-step predictions, and prediction accuracy decreases significantly after the prediction compensation exceeds a certain unit limit.

[0008] Compared with the prior art, the differences are as follows:

[0009] Technical comparison with patent CN115438576A "Electronic voltage transformer error prediction method based on Prophet, self-attention mechanism and time series convolutional network"

[0010] Patent CN115438576A proposes an error prediction method for electronic voltage transformers based on Prophet, a self-attention mechanism, and a time series convolutional network. This method focuses on the problem of predicting the ratio error of electronic voltage transformers. By combining the periodicity analysis capabilities of the Prophet model, the local feature extraction capabilities of the time series convolutional network, and the feature enhancement function of the self-attention mechanism, it predicts the medium- and long-term error trends of electronic voltage transformers. This technical solution is suitable for condition monitoring and maintenance of electronic voltage transformers in power systems, and is particularly suitable for processing error data with significant periodic characteristics. This patent proposes a method for error prediction for capacitor voltage transformers based on modal decomposition and gated recurrent units. This method addresses the difficulty of error prediction for capacitor voltage transformers under complex operating conditions. It decomposes non-stationary signals into multimodal components through an improved modal decomposition technique and combines gated recurrent units with a non-local attention mechanism to achieve dynamic feature fusion. This solution is particularly suitable for processing voltage transformer output signals with transient responses and noise interference, and can effectively solve the error accumulation problem in multi-step predictions caused by traditional methods. There are essential differences in the application scenarios and objects of the two.

[0011] Patent CN115438576A standardizes the ratio difference data of electronic voltage transformers and decomposes it into three basic components: trend, period, and noise. It uses the dilated causal convolution and residual connection of a time series convolutional network to extract local time series features and weightedly fuses them with the periodic term of the Prophet. It calculates feature weights through a self-attention mechanism to enhance key information, and finally outputs the prediction results through dimensionality reduction using a fully connected layer. By adopting a cascaded structure of Prophet and a time series convolutional network, it focuses on combining the periodicity of time series with local features. This patent uses an improved modal decomposition method combined with frequency-domain adaptive noise adjustment technology to decompose the output signal of a capacitive voltage transformer into multiple intrinsic modal function components; constructs a feature evaluation system based on multi-scale information entropy theory to classify the components; performs adaptive wavelet packet threshold denoising on high-entropy noise modes, retaining effective signal features while reducing noise; designs a hybrid architecture of gated recurrent units and non-local attention, dynamically balances local temporal features with global contextual information through a gating mechanism, performs parallel predictions on each modal component, and superimposes the prediction results of each component to obtain the final output signal prediction value. This patent splits the signal into multi-modal components through modal decomposition and adopts a parallel modeling strategy, focusing more on multi-scale decomposition and global feature fusion. There are essential differences between the two technical solutions.

[0012] Patent CN115438576A utilizes a Prophet model, using piecewise linear functions to fit trend terms and Fourier series to fit periodic terms. This model is robust to missing values and outliers, and can identify and adapt to periodic changes at different time scales. In a time series convolutional network, dilated causal convolutions are used to expand the receptive field, combined with residual connections to prevent gradient vanishing and enhance local feature extraction. Feature weights are calculated within a self-attention mechanism to enhance the model's focus on key information, primarily using fixed-structure convolution and attention mechanisms. The improved modal decomposition method in this patent dynamically adjusts the noise amplitude coefficient through frequency domain analysis, achieving adaptive optimization of signal decomposition and significantly improving decomposition quality. Wavelet packet transforms combined with a dynamic thresholding strategy are used for noise reduction, effectively suppressing noise while fully preserving the transient characteristics of the signal. When building the prediction model, gating coefficients are calculated in real time to dynamically adjust the fusion weights of local time series features (GRU outputs) and global contextual information (attention outputs), enabling the model to automatically adjust its processing strategy based on the characteristics of the input signal. The two technologies differ fundamentally in their technical approaches.

[0013] The Prophet model in patent CN115438576A shows excellent fitting ability for error data with obvious periodic characteristics and can capture the law of periodic changes. The time series convolutional network effectively extracts the nonlinear characteristics in the error signal through multi-level feature abstraction, which is suitable for medium- and long-term predictions, but its processing ability for high-frequency noise and transient interference signals is relatively limited. In this patent, multimodal signal decomposition is used to maintain stable performance under complex working conditions such as voltage fluctuations and harmonic interference; wavelet packet denoising technology can accurately suppress noise in different frequency bands, while retaining the key features of the signal; the non-local attention mechanism breaks through the "memory length: limitation" of the traditional time series model, establishes feature associations across time periods, and significantly improves the accuracy of multi-step predictions. There is an essential difference between the two in terms of technical effects.

[0014] As summarized above, common problems with existing methods include poor dynamic adaptability and significant prediction error accumulation. Especially when dealing with complex nonlinear systems like CVTs, the prediction reliability of existing methods cannot meet the accuracy requirements of smart grids. Therefore, it is urgent to develop a data-driven voltage transformer error prediction method that can deeply exploit the time-dependent characteristics of the signal, effectively mitigate the error accumulation problem during the prediction process, and achieve accurate prediction of the error signal for multiple time steps into the future. Summary of the Invention

[0015] In order to solve the problem of prediction error accumulation of CVT under actual variable operating conditions, the present invention proposes a voltage transformer error prediction method based on modal decomposition and gated cyclic units. This method decomposes the transformer output signal into multi-modal variables through an adaptive empirical mode decomposition method, performs modal classification according to signal entropy, establishes multiple prediction models for different modes, and superimposes the multi-component prediction values on the prediction results, thereby realizing intelligent prediction of voltage transformer signals.

[0016] To achieve the above object, the technical solution adopted by the present invention is:

[0017] A voltage transformer error prediction method based on modal decomposition and gated cyclic units is characterized by comprising the following steps:

[0018] Step 1: Read the original output signal of the voltage transformer through equal-interval sampling to establish a historical data set;

[0019] Step 2: Using the improved adaptive noise fully integrated empirical mode decomposition method, the non-stationary signal is decomposed into several intrinsic mode function components through the white noise spectrum adaptive adjustment mechanism;

[0020] Step 3: Construct a feature evaluation system based on multi-scale information entropy, achieve accurate separation of eigenmodes through modal entropy-frequency band joint analysis, and use adaptive wavelet packet threshold denoising for high-entropy noise modes to preserve the transient response characteristics of the eigenmodes;

[0021] Step 4: Perform attention-enhanced signal prediction and design a hybrid architecture of gated recurrent units and attention mechanisms. This architecture captures local dynamics through spatiotemporal gated units and long-term dependencies through non-local attention mechanisms, achieving adaptive weight allocation for key features.

[0022] Step 5: Input each modal component into the signal prediction model in parallel to train and predict the future trend data of each component. The prediction results are superimposed as the output signal prediction value;

[0023] Step 6: Validate the model on the validation set and use multiple evaluation indicators to comprehensively evaluate the model performance.

[0024] As a further improvement of the present invention, the improved adaptive noise fully integrated empirical mode decomposition method in step 2 includes the following process:

[0025] (2-1) In the process of improving modal decomposition, adaptive noise is added to the signal, and the noise amplitude adjustment coefficient is set to ξ, which is used to add noise to generate white noise w i (n) satisfies the mean of 0 and the variance of 1, where i = 1, 2, ..., M;

[0026] Add noise to generate a new signal: Add noise for the i-th time to generate a disturbance signal x i (n):

[0027] x (i) (n) = x(n) + ξ·w (i) (n)

[0028] The selection of noise amplitude ξ directly affects the purity and decomposition stability of the modal components. Frequency domain adaptive noise adjustment is used for measurement:

[0029]

[0030] Where f represents the signal frequency, σ x represents the signal standard deviation;

[0031] (2-2) Perform EMD decomposition on each noise disturbance signal to obtain a set of IMF components and residuals;

[0032]

[0033] where K i represents the number of IMFs decomposed for the i-th time, represents the kth order eigenmode component, r (i) (n) represents the residual component. Take the first IMF component of all noise disturbance signals, calculate the average as IMF1(n), update the signal, repeat the iteration, and extract the remaining modal components.

[0034] As a further improvement of the present invention, in step 3, accurate separation of characteristic modes is achieved through modal entropy value-frequency band joint analysis, and adaptive wavelet packet threshold denoising is used for high entropy noise modes, including the following process:

[0035] (3-1) For the extracted multiple IMF components, calculate the signal entropy of each component and use the Gaussian kernel function to perform kernel density estimation:

[0036]

[0037] in Represents the density estimate at position x, n is the sample size, h is the bandwidth, and K is the kernel function;

[0038] (3-2) For high entropy modal components, wavelet packet transform combined with adaptive threshold method is used for denoising. The adaptive threshold is used to maintain the detailed characteristics of the signal during denoising, thereby achieving a better balance between maintaining signal continuity and effective denoising. The calculation formula of the threshold η is:

[0039]

[0040] Where: w is the wavelet packet coefficient; λ is the global threshold calculated based on the noise level; m is the adaptive factor, m∈(0,1];

[0041] Each wavelet packet coefficient w is compared with the calculated threshold η: if |w|>η, the coefficient is retained; if |w|<η, the coefficient is set to zero to remove noise.

[0042] As a further improvement of the present invention, the attention-enhanced signal prediction model in step 4 includes the following process:

[0043] (4-1) The application of non-local attention mechanism can help the model better understand and capture dependencies over long time intervals. The non-local attention operation calculates the non-local block response based on the relationship between nodes at different positions, namely:

[0044]

[0045] Where x is the input timing signal; y i is the enhanced feature at position i; f(x i ,x j) calculates the correlation strength between positions i and j to quantify the dependency between two time points; g(x j ) converts the input signal into a feature representation that is more suitable for attention aggregation; C(x) is a normalization factor to prevent the attention weight factor from being invalid due to different scales;

[0046] (4-2) Dynamically allocate weights through gated fusion, adaptively balance non-local attention and GRU weights, and decide in real time whether to rely on global context or local state based on the input features of the current time step. During the modeling process, smooth transition is achieved through Sigmoid gating. The gated fusion unit adaptively mixes local and global features by calculating the gate value. The gate value calculation splices the local features and global features into a matrix space, which is mapped to the gate space through the weight matrix and bias vector.

[0047] Beneficial effects:

[0048] Through an improved modal decomposition method, multi-scale feature extraction of the signal is achieved. Through signal entropy-guided modal components and noise reduction processing, the influence of noise in strong electromagnetic interference environments is suppressed, improving the signal-to-noise ratio while retaining the effective signal characteristics. The gated fusion mechanism dynamically balances local temporal characteristics and global contextual information, solving the error accumulation problem of traditional methods in long-term series prediction. Multimodal parallel processing can adapt to different operating conditions and maintain stable prediction performance under complex conditions. This method is applicable to the field of online voltage transformer monitoring technology and solves the problem of error prediction of voltage transformers under complex operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the voltage transformer signal prediction method provided by the present invention;

[0050] Figure 2 It is the GRU prediction model based on the non-local attention mechanism proposed in this invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0052] The present invention discloses a voltage transformer error prediction method based on modal decomposition and gated cyclic unit. The method flow of the present invention is as follows: Figure 1 As shown in , the GRU prediction model based on the non-local attention mechanism is as follows Figure 2 As shown, the specific steps include:

[0053] Step 1: Historical output signal acquisition. Collect the voltage transformer output signals. Label and classify these signals for subsequent analysis and diagnosis.

[0054] Step 2: Adaptive signal decomposition. Adaptive noise is introduced based on the signal standard deviation, and the complete noise set empirical mode decomposition method is improved to decompose the noisy signal and obtain multiple intrinsic mode components. The specific steps include:

[0055] (2-1) During the CEEMDAN decomposition process, adaptive noise is added to the signal. The noise amplitude adjustment coefficient is set to ξ, which is used to add noise to generate white noise w i (n) satisfies the mean of 0 and the variance of 1, where i = 1, 2, ..., M.

[0056] Add noise to generate a new signal: Add noise for the i-th time to generate a disturbance signal x i (n):

[0057] x (i) (n) = x(n) + ξ·w (i) (n)

[0058] The selection of noise amplitude ξ directly affects the purity and decomposition stability of the modal components. This paper adopts frequency domain adaptive noise adjustment measurement:

[0059]

[0060] Where f represents the signal frequency, σ x Indicates the signal standard deviation.

[0061] (2-2) Perform EMD decomposition on each noise disturbance signal to obtain a set of IMF components and residuals.

[0062]

[0063] where K i represents the number of IMFs decomposed for the i-th time, represents the kth order eigenmode component, r (i) (n) represents the residual component. Take the first IMF component of all noise disturbance signals, calculate the average as IMF1(n), update the signal, repeat the iteration, and extract the remaining modal components.

[0064] Step 3: Multimodal feature extraction and classification. Based on the information entropy theory, a feature evaluation system is constructed to calculate the information entropy of each component, identify the high entropy component of the noise-dominant mode, the medium entropy component of the characteristic information mode, and the low entropy component of the trend component, and use wavelet thresholding to reduce noise on the high entropy component. The specific steps include:

[0065] (3-1) For the extracted multiple IMF components, calculate the signal entropy of each component and use the Gaussian kernel function to perform kernel density estimation:

[0066]

[0067] in Represents the density estimate at position x, n is the sample size, h is the bandwidth, and K is the kernel function.

[0068] (3-2) For high entropy modal components, wavelet packet transform combined with adaptive threshold method is used for noise reduction. The adaptive threshold is used to maintain the detailed characteristics of the signal during denoising, thereby achieving a better balance between maintaining signal continuity and effective denoising. The calculation formula of the threshold η is:

[0069]

[0070] Where: w is the wavelet packet coefficient; λ is the global threshold calculated based on the noise level; m is the adaptive factor, m∈(0,1].

[0071] Each wavelet packet coefficient w is compared with the calculated threshold η: if |w|>η, the coefficient is retained; if |w|<η, the coefficient is set to zero to remove noise.

[0072] Step 4: Build an attention-enhanced signal prediction model. Use a gated recurrent unit network and introduce a non-local attention mechanism to capture dependencies over long time intervals, improving the model's ability to focus on important features and its long-term memory. The specific steps are as follows:

[0073] (4-1) The application of non-local attention mechanism can help the model better understand and capture dependencies over long time intervals, thereby improving the accuracy of prediction. The non-local attention operation calculates the non-local block response based on the relationship between nodes at different positions, that is:

[0074]

[0075] Where x is the input timing signal; y i is the enhanced feature at position i; f(x i ,x j ) calculates the correlation strength between positions i and j to quantify the dependency between two time points; g(x j ) converts the input signal into a feature representation that is more suitable for attention aggregation; C(x) is a normalization factor to prevent the attention weight factor from becoming invalid due to different scales.

[0076] (4-2) The Gated Recurrent Unit (GRU) is a simplified variant of the recurrent neural network based on the LSTM. The basic network architecture of the improved GRU prediction model dynamically allocates weights through gated fusion, adaptively balancing the weights of non-local attention (global features) and GRU (local temporal features). Based on the input features of the current time step, it decides in real time whether to rely on the global context or the local state. During the modeling process, a smooth transition is achieved through Sigmoid gating to avoid the instability caused by hard switching. The gated fusion unit adaptively mixes local and global features by calculating the gate value. The gate value calculation splices the local features and global features into a matrix space, which is mapped to the gate space through the weight matrix and bias vector.

[0077] Step 5: Multimodal model training and optimization. Each modal component is input into the signal prediction model for training and the future trend data of each component is predicted. The prediction results are superimposed as the output signal prediction value.

[0078] Step 6: Performance evaluation and verification. Verify the model on the validation set and use multiple evaluation indicators to comprehensively evaluate the model performance to ensure its accuracy and stability. The specific steps are as follows:

[0079] (6-1) The mean absolute error (MAE) and root mean square error (RMSE) indicators are used to evaluate the reliability of the model.

[0080] Mean Absolute Error (MAE) is used to measure the average absolute deviation between the predicted value and the true value. It can intuitively reflect the size of the prediction error and is suitable for evaluating the overall prediction stability of the prediction model. The calculation formula is:

[0081]

[0082] where y i Represents the true value of the i-th sample; represents the predicted value of the i-th sample; n is the number of samples.

[0083] The root mean square error (RMSE) is used to measure the deviation between the predicted value and the true value, more strictly evaluate the stability of the model, and avoid extreme errors. The calculation formula is:

[0084]

[0085] The improved NL-GRU model can effectively enhance the model's dependence on long time series, the prediction curve has less deviation, and the prediction gradually tends to be stable in the later stage.

[0086] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A voltage transformer error prediction method based on modal decomposition and gated cyclic units, characterized in that: The steps include: Step 1: Read the original output signal of the voltage transformer through equal-interval sampling to establish a historical data set; Step 2: Using the improved adaptive noise fully integrated empirical mode decomposition method, the non-stationary signal is decomposed into several intrinsic mode function components through the white noise spectrum adaptive adjustment mechanism; Step 3: Construct a feature evaluation system based on multi-scale information entropy, achieve accurate separation of eigenmodes through modal entropy-frequency band joint analysis, and use adaptive wavelet packet threshold denoising for high-entropy noise modes to preserve the transient response characteristics of the eigenmodes; Step 4: Perform attention-enhanced signal prediction and design a hybrid architecture of gated recurrent units and attention mechanisms. This architecture captures local dynamics through spatiotemporal gated units and long-term dependencies through non-local attention mechanisms, achieving adaptive weight allocation for key features. Step 5: Input each modal component into the signal prediction model in parallel to train and predict the future trend data of each component. The prediction results are superimposed as the output signal prediction value; Step 6: Validate the model on the validation set and use multiple evaluation indicators to comprehensively evaluate the model performance.

2. The voltage transformer error prediction method based on modal decomposition and gated cyclic unit according to claim 1, characterized in that: The improved adaptive noise fully integrated empirical mode decomposition method in step 2 includes the following steps: (2-1) In the process of improving modal decomposition, adaptive noise is added to the signal, and the noise amplitude adjustment coefficient is set to ξ, which is used to add noise to generate white noise w i (n) satisfies the mean of 0 and the variance of 1, where i = 1, 2, ..., M; Add noise to generate a new signal: Add noise for the i-th time to generate a disturbance signal x i (n): x (i) (n)=x(n)+ξ·w (i) (n) The selection of noise amplitude ξ directly affects the purity and decomposition stability of the modal components. Frequency domain adaptive noise adjustment is used for measurement: Where f represents the signal frequency, σ x represents the signal standard deviation; (2-2) Perform EMD decomposition on each noise disturbance signal to obtain a set of IMF components and residuals; where K i represents the number of IMFs decomposed for the i-th time, represents the kth order eigenmode component, r (i) (n) represents the residual component. Take the first IMF component of all noise disturbance signals, calculate the average as IMF1(n), update the signal, repeat the iteration, and extract the remaining modal components.

3. The voltage transformer error prediction method based on modal decomposition and gated cyclic unit according to claim 1, characterized in that: In step 3, the precise separation of characteristic modes is achieved through the modal entropy value-frequency band joint analysis. The adaptive wavelet packet threshold denoising is used for the high entropy noise mode, which includes the following process: (3-1) For the extracted multiple IMF components, calculate the signal entropy of each component and use the Gaussian kernel function to perform kernel density estimation: in Represents the density estimate at position x, n is the sample size, h is the bandwidth, and K is the kernel function; (3-2) For high entropy modal components, wavelet packet transform combined with adaptive threshold method is used for denoising. The adaptive threshold is used to maintain the detailed characteristics of the signal during denoising, thereby achieving a better balance between maintaining signal continuity and effective denoising. The calculation formula of the threshold η is: Where: w is the wavelet packet coefficient; λ is the global threshold calculated based on the noise level; m is the adaptive factor, m∈(0,1]; Each wavelet packet coefficient w is compared with the calculated threshold η: if |w|>η, the coefficient is retained; if |w|<η, the coefficient is set to zero to remove noise.

4. The voltage transformer error prediction method based on modal decomposition and gated cyclic unit according to claim 1, characterized in that: The signal prediction model for attention enhancement in step 4 includes the following steps: (4-1) The application of non-local attention mechanism can help the model better understand and capture dependencies over long time intervals. The non-local attention operation calculates the non-local block response based on the relationship between nodes at different positions, namely: Where x is the input timing signal; y i is the enhanced feature at position i; f(x i ,x j ) calculates the correlation strength between positions i and j to quantify the dependency between two time points; g(x j ) converts the input signal into a feature representation that is more suitable for attention aggregation; C(x) is a normalization factor to prevent the attention weight factor from being invalid due to different scales; (4-2) Dynamically allocate weights through gated fusion, adaptively balance non-local attention and GRU weights, and decide in real time whether to rely on global context or local state based on the input features of the current time step. During the modeling process, smooth transition is achieved through Sigmoid gating. The gated fusion unit adaptively mixes local and global features by calculating the gate value. The gate value calculation splices the local features and global features into a matrix space, which is mapped to the gate space through the weight matrix and bias vector.

Images