A voltage fluctuation sensing energy meter and determination method integrating deep discriminant networks

Abstract

This invention discloses a voltage fluctuation sensing energy meter and judgment method integrating a deep discriminant network. The method obtains raw voltage waveform data under a synchronous clock reference; outputs voltage time window data segments; performs adaptive window-length short-time Fourier transform on these segments to generate time-frequency graph data; obtains disturbance gating flag data; triggers the main model inference process when the disturbance gating flag data indicates the presence of a disturbance; otherwise, maintains a low-power, lightweight operating state; outputs initial voltage disturbance judgment results and confidence data; generates voltage disturbance event-level judgment data; and outputs backoff judgment data. This invention effectively improves the online judgment capability for flicker, sag / boost, and harmonic group sudden disturbance events. In field tests, it significantly reduces the false alarm and missed alarm probabilities under conditions of rapid load switching and grid interference, achieving high-sensitivity, low-latency real-time inference while maintaining controllable resource consumption within the SoC.

Description

Technical Field

[0001] This invention relates to the field of electricity meter technology, and in particular to a voltage fluctuation sensing electricity meter and determination method that integrates a deep discriminant network. Background Technology

[0002] With the widespread deployment of electricity meters on the power distribution side, online power quality sensing and event determination have become crucial for improving the reliability of power grid operation and enhancing user experience. Existing technologies generally employ two main methods for identifying voltage fluctuation events from electricity meters: one is the traditional criterion method based on time-domain statistics or single frequency-domain features; the other utilizes end-to-end deep learning models to directly extract and classify features from the original waveform or basic characteristics.

[0003] In practical applications, traditional methods based on single features or rule thresholds, while simple to implement and low in resource consumption, are prone to false alarms and missed alarms under conditions of rapid load switching, inverter grid-connection interference, and operating noise. They struggle to handle non-stationary and non-Gaussian voltage fluctuation signals and often fail to meet the real-time consistency requirements of the electricity metering link and the integration requirements of the power quality assessment system. Meanwhile, end-to-end methods based purely on deep learning typically rely on high computing power and large-capacity storage, making them difficult to deploy in resource-constrained environments such as embedded electricity meter SoCs and MCUs. Furthermore, the black-box nature of model inference in practical engineering applications makes it difficult to map the results to traditional power quality criteria, lacking traceability and interpretability, and failing to meet the rigid requirements of industry compliance and audit traceability.

[0004] Existing technologies lack sufficient exploration of the multi-scale and structured characteristics of voltage fluctuation events, and single statistical quantities or spectral indicators are insufficient to reflect the full picture of complex disturbance processes. At the same time, existing AI discrimination methods have low coupling with the electricity metering business process and cannot achieve seamless integration with national standard power quality levels or on-site operation and maintenance scenarios. Summary of the Invention

[0005] One objective of this invention is to propose a voltage fluctuation sensing energy meter and judgment method that integrates a deep discriminant network. This invention effectively improves the online discrimination capability for flicker, sag / boost, and harmonic group sudden disturbance events. Under field testing, it significantly reduces the probability of false alarms and missed alarms under conditions of rapid load switching and grid interference. It achieves high sensitivity and low latency real-time inference under the premise of controllable resource consumption within the SoC.

[0006] A method for determining voltage fluctuation sensing energy meters based on a fused deep discriminant network according to an embodiment of the present invention includes:

[0007] The voltage input signal inside the energy meter is read and preprocessed to obtain the original voltage waveform data under the synchronous clock reference.

[0008] Based on the power frequency phase-locked loop results, the original voltage waveform data is divided into continuous frames according to a fixed number of sampling points, and voltage time window data segments are output. Adaptive window length short-time Fourier transform is applied to the voltage time window data segments to generate time-frequency graph data.

[0009] The amplitude spectrum first-order difference, phase spectrum first-order difference, and transient disturbance significance index are calculated on the time-frequency graph data and fused to construct the disturbance graph data. The disturbance graph data is filtered according to the disturbance gating threshold to obtain disturbance gating flag data. When the disturbance gating flag data indicates the presence of a disturbance, the main model inference process is triggered; otherwise, the low-power lightweight operation state is maintained.

[0010] The perturbation map data is input into the improved DenseNet deep discriminant network, which consists of reconfigured dense connection blocks, bottleneck layers, depth-separable convolutions, and group convolutions. Inference is performed on the embedded SoC platform, and the initial voltage perturbation discrimination result data and confidence data are output.

[0011] Determine whether the confidence data is higher than the early termination threshold. If it is higher than the early termination threshold, directly output the early termination discrimination result data. Otherwise, continue to execute the remaining dense connection blocks of the improved DenseNet deep discrimination network to obtain the final discrimination result data. Perform a monotonic mapping between the final discrimination result data and the preset power quality threshold mapping table to generate voltage disturbance event-level judgment data.

[0012] When the confidence data is lower than the confidence threshold or the improved DenseNet deep discriminant network fails to perform self-testing, the frequency domain threshold rule determination process is executed, and the rollback determination data is output.

[0013] Optionally, the preprocessing includes:

[0014] At a preset sampling frequency, the voltage input signal inside the energy meter is read through an analog-to-digital converter to obtain the original voltage waveform data;

[0015] A digital anti-aliasing filter is applied to the raw voltage waveform data to perform filtering calculations with predetermined numerator and denominator coefficients, and the anti-aliasing filtered output data is output.

[0016] The adaptive gain factor is calculated by the ratio of the reference voltage amplitude to the maximum absolute value of the anti-aliasing filter output data in the current frame. The adaptive gain factor is then used to amplify or reduce each sampling point in the anti-aliasing filter output data to obtain gain correction data.

[0017] The DC bias is obtained by taking the arithmetic mean of the values ​​of all sample points in the current frame of the gain correction data. The DC bias is then subtracted from each sample point in the gain correction data to obtain the bias correction data.

[0018] The sampling time corresponding to each sampling point is determined according to the sampling synchronization condition. All sampling points in the bias correction data are aligned with the detected power frequency reference phase and time-marked, and the original voltage waveform data under the synchronous clock reference is output.

[0019] Optionally, the adaptive window length short-time Fourier transform of the voltage time window data segment includes:

[0020] The original voltage waveform data under the synchronous clock reference is divided according to the number of sampling points in each frame to obtain voltage time window data segments;

[0021] Based on the reference phase obtained from the power frequency phase-locked loop result, phase alignment is performed on each voltage time window data segment to obtain the phase-aligned voltage time window data segment.

[0022] Short-time Fourier transform is performed on the phase-aligned voltage time window data segments using an adaptive window function to obtain the time-frequency plot data for each frame;

[0023] The time-frequency graph data is concatenated according to the frame index order to output the complete time-frequency graph data.

[0024] Optionally, the step of filtering the disturbance map data based on the disturbance gating threshold to obtain disturbance gating flag data includes:

[0025] Based on the time-frequency graph data, calculate the magnitude of the time-frequency graph data under each frequency index and each frame index to obtain the amplitude spectrum data, and calculate the phase angle of the time-frequency graph data under each frequency index and each frame index to obtain the phase spectrum data;

[0026] For the amplitude spectrum data, the difference between the amplitude spectrum data of two adjacent frequency indices is calculated under each frame index according to the frequency index direction to obtain the first-order difference data of the amplitude spectrum;

[0027] For each frame index, the difference between the phase spectrum data of two adjacent frequency indices is calculated according to the frequency index direction to obtain the first-order difference data of the phase spectrum.

[0028] The absolute value of the difference between the time-frequency plot data under each frame index and the time-frequency plot data of the corresponding frequency index under the previous frame index is calculated and accumulated to obtain the transient disturbance significance index.

[0029] The first-order difference data of amplitude spectrum, the first-order difference data of phase spectrum, and the transient perturbation significance index are weighted and summed to obtain the perturbation map data;

[0030] Set a perturbation gating threshold to distinguish perturbation map data for all frequency indices under each frame index:

[0031] When the perturbation map data at any sampling point is greater than or equal to the perturbation gating threshold, a perturbation gating flag of one is generated, triggering the main model inference process;

[0032] If the disturbance map data of all sampling points is less than the disturbance gating threshold, the disturbance gating flag data is generated as zero, maintaining a low-power and lightweight operating state.

[0033] Optionally, the improved DenseNet deep discriminative network includes:

[0034] The perturbation map data is rearranged into input tensors according to three feature channels: first-order difference of amplitude spectrum, first-order difference of phase spectrum, and transient perturbation significance. The frequency band weighting vector is calculated according to the ratio of the cumulative value of the perturbation map data at each frequency index in all frames to the cumulative value of the overall perturbation map data. The weighted tensor is obtained by multiplying the perturbation map data at each frequency index with the frequency band weighting vector corresponding to the frequency index.

[0035] The weighted tensor is input into the first bandwidth-sensing dense block, which comprises multiple bandwidth-sensing dense units, and the output features are... ;

[0036] For output features Channel compression is achieved through a bottleneck compression layer, which uses a one-dimensional convolution operation to reduce the number of channels in the output features, resulting in channel-compressed features.

[0037] Based on the variance of each channel under all frequency indices and all frame indices according to the channel compression features, a sparsification mask is generated. The sparsification mask marks channels with variance greater than or equal to the sparsity threshold as one, and otherwise as zero. The sparsification mask is used to dynamically filter and prune the convolution kernel to output sparsification features.

[0038] The sparse features are input into the depthwise separable convolutional unit to obtain the depthwise separable convolutional features, which are then fed into the group convolutional module. The group convolutional module divides all channels into multiple groups and performs convolution operations on each group independently, outputting the group convolutional features of the first stage.

[0039] Repeat the operation of the first bandwidth-aware dense block and the depth-separable convolutional unit to form multiple sets of convolutional features;

[0040] Configure an early termination discrimination head at several specified stage indices, calculate the classification score vector for the current group's convolutional features, and obtain the probability vector through Softmax. When the maximum value in the probability vector is greater than or equal to the early termination threshold, directly output the early termination voltage perturbation discrimination result and confidence level for the current stage.

[0041] When the maximum probability of all early termination discriminators does not reach the early termination threshold, the same discrimination process is performed on the final stage features, and the initial voltage disturbance discrimination result data and confidence data are output. All calculations are completed and output within one inference cycle of the embedded SoC platform.

[0042] Optionally, the generation of voltage disturbance event-level determination data includes:

[0043] Based on the discrimination results of each stage, the probability vector of the current stage is obtained. The component with the largest value in the probability vector of the current stage is taken as the confidence data of the current stage, and the voltage disturbance category corresponding to the component is taken as the voltage disturbance category discrimination result of the current stage.

[0044] The confidence data of the current stage is compared with the early termination threshold. When the confidence data of the current stage is greater than or equal to the early termination threshold, the early termination discrimination result data and confidence data of the current stage are directly output, and the calculation of the remaining dense connection blocks of the improved DenseNet deep discrimination network is stopped.

[0045] When the confidence data of the current stage is less than the early termination threshold, the output features of the current stage are continued to be input into the remaining dense connection blocks of the improved DenseNet deep discriminant network to obtain the probability vector of the final stage. The component with the largest value in the probability vector of the final stage is taken as the confidence data of the final stage, and the voltage disturbance category corresponding to the component is taken as the voltage disturbance category discrimination result of the final stage.

[0046] The early termination judgment result data is monotonically mapped to the preset power quality threshold mapping table to obtain voltage disturbance event-level judgment data.

[0047] Voltage disturbance event-level determination data includes disturbance type, disturbance level, and disturbance duration. The disturbance type is equal to the voltage disturbance category determination result of the current stage or the final stage.

[0048] Optionally, the execution frequency domain threshold rule determination process includes:

[0049] Obtain the confidence data output from the final discrimination stage, compare the confidence data with the preset confidence threshold, and start the frequency domain threshold rule determination process when the confidence data is less than the preset confidence threshold, or when the improved DenseNet deep discrimination network self-test module returns an error flag.

[0050] In the frequency domain threshold rule determination process, the first-order difference data of amplitude spectrum and transient disturbance significance index are extracted from the disturbance map data, and the average disturbance intensity index is calculated.

[0051] Set the amplitude spectrum perturbation threshold, transient perturbation threshold, and average perturbation intensity threshold, and perform the following judgment logic for each frame index:

[0052] If there are sampling points in the first-order difference data of the amplitude spectrum that are greater than or equal to the amplitude spectrum disturbance threshold, or the transient disturbance significance index is greater than or equal to the transient disturbance threshold, or the average disturbance intensity index is greater than or equal to the average disturbance intensity threshold, then the current frame is determined to be a voltage disturbance frame; otherwise, it is determined to be a normal frame.

[0053] All frame determination results are merged to obtain a set of disturbance event frames. The disturbance type is then mapped according to the main channel of the disturbance triggering characteristics.

[0054] If the average value of the first-order difference data of the amplitude spectrum in the current frame is greater than or equal to the average absolute value of the first-order difference data of the phase spectrum, then the current disturbance type is determined to be an amplitude disturbance; otherwise, it is determined to be a phase disturbance.

[0055] The set of disturbance event frames is a set of frame indices that are all identified as voltage disturbance frames;

[0056] The duration of the disturbance is obtained by multiplying the number of frames in the set of disturbance event frames and the frame duration for each frame.

[0057] Within the set of disturbance event frames, find the maximum disturbance amplitude value in the disturbance map data, and input the disturbance type, disturbance duration, and maximum disturbance amplitude value into the preset backoff level mapping function to obtain the disturbance level;

[0058] Output backoff determination data, which includes disturbance type, disturbance duration, and disturbance level.

[0059] A voltage fluctuation sensing energy meter incorporating a deep discriminant network is equipped with a storage system for executing a determination method for the voltage fluctuation sensing energy meter incorporating a deep discriminant network.

[0060] The beneficial effects of this invention are:

[0061] (1) This invention is based on the fusion of multi-scale frequency domain disturbance analysis and bandwidth-aware dense discrimination to improve the accuracy and robustness of on-site identification of complex disturbance events. For application scenarios with limited electricity meter resources, a multi-channel disturbance map input mechanism with first-order difference of amplitude spectrum, first-order difference of phase spectrum and transient disturbance significance is proposed. The feature resolution is dynamically adjusted and the bandwidth-aware multi-scale feature reuse is realized through the frequency band weighted adaptive DenseNet structure. The bandwidth-aware dense unit dynamically adjusts the growth rate according to the actual disturbance energy distribution of each frequency band. With the help of sparse channel pruning, depth separable convolution and group convolution, the parameter quantity and computing power resources are optimized and allocated. This effectively improves the online discrimination capability of flicker, sag / boost and harmonic group sudden disturbance events. Under the actual field test, the false alarm and missed alarm probability under the conditions of rapid load switching and grid interference is significantly reduced. Under the premise of controllable resource consumption in SoC, high sensitivity and low latency real-time inference are realized.

[0062] (2) This invention integrates an event-triggered gating mechanism on the electricity meter body side, uses a perturbation graph layered threshold for rapid event screening, dynamically activates the main DenseNet discrimination process only when a significant perturbation event occurs, and sets up a multi-layer early termination discrimination head in the backbone network. The calculation is intelligently terminated in advance according to the discrimination confidence, and outputs the perturbation category and level results with high confidence. For scenarios with insufficient confidence or model self-check failure, the system can automatically fall back to the traditional criteria based on frequency domain rules, so as to achieve uninterrupted judgment continuity. The entire discrimination and fallback process supports interpretable output of perturbation type, duration and level. The entire judgment link is traceable, realizing seamless collaboration and audit interface docking with the electricity meter metering link, effectively ensuring the consistency of criteria, industry compliance and traceability. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a voltage fluctuation sensing energy meter and determination method that integrates a deep discriminant network, as proposed in this invention. Detailed Implementation

[0065] Example 1: Reference Figure 1 A method for determining voltage fluctuation sensing energy meters that integrates deep discriminant networks, comprising:

[0066] The voltage input signal inside the energy meter is read and preprocessed to obtain the original voltage waveform data under the synchronous clock reference.

[0067] In this embodiment, the preprocessing includes:

[0068] At a preset sampling frequency, the voltage input signal inside the energy meter is read through an analog-to-digital converter to obtain the original voltage waveform data;

[0069] The raw voltage waveform data consists of the voltage sample value corresponding to each discrete sampling number.

[0070] A digital anti-aliasing filter is applied to the raw voltage waveform data to perform filtering calculations with predetermined numerator and denominator coefficients, and the anti-aliasing filtered output data is output.

[0071] Each sampling point in the anti-aliasing filter output data is obtained by convolving the original voltage waveform data with the filter coefficients.

[0072] The adaptive gain factor is calculated by the ratio of the reference voltage amplitude to the maximum absolute value of the anti-aliasing filter output data in the current frame. The adaptive gain factor is then used to amplify or reduce each sampling point in the anti-aliasing filter output data to obtain gain correction data.

[0073] Each sampling point in the gain correction data is equal to the product of the adaptive gain factor and the anti-aliasing filter output data.

[0074] The DC bias is obtained by taking the arithmetic mean of the values ​​of all sample points in the current frame of the gain correction data. The DC bias is then subtracted from each sample point in the gain correction data to obtain the bias correction data.

[0075] Each sampling point in the bias correction data is equal to the corresponding sampling point in the gain correction data minus the DC bias.

[0076] The sampling time corresponding to each sampling point is determined according to the sampling synchronization condition. All sampling points in the bias correction data are aligned with the detected power frequency reference phase and time-marked, and the original voltage waveform data under the synchronous clock reference is output.

[0077] Based on the power frequency phase-locked loop results, the original voltage waveform data is divided into continuous frames according to a fixed number of sampling points, and voltage time window data segments are output. Adaptive window length short-time Fourier transform is applied to the voltage time window data segments to generate time-frequency graph data.

[0078] In this embodiment, an adaptive window length short-time Fourier transform is performed on the voltage time window data segment, including:

[0079] The original voltage waveform data under the synchronous clock reference is divided according to the number of sampling points in each frame to obtain voltage time window data segments;

[0080] In the voltage time window data segment, each sampling point within a frame corresponds to a sampling point of the original voltage waveform data. All sampling points in each frame cover a continuous time interval of the original voltage waveform data under the synchronous clock reference.

[0081] Based on the reference phase obtained from the power frequency phase-locked loop result, phase alignment is performed on each voltage time window data segment to obtain the phase-aligned voltage time window data segment.

[0082] The phase alignment process ensures that the starting sampling point of each frame is kept in the same position on the power frequency cycle. Each sampling point in the voltage time window data segment after phase alignment is equal to the linear combination of the corresponding sampling point value of the voltage time window data segment and the cosine and sine components of the power frequency reference phase.

[0083] Short-time Fourier transform is performed on the phase-aligned voltage time window data segments using an adaptive window function to obtain the time-frequency plot data for each frame;

[0084] The value of each time-frequency point in the time-frequency graph data is equal to the sum of the product of all the sampled points in that frame and the corresponding adaptive window function weights, and then the result of a linear combination with the sine and cosine modulation factors. The adaptive window function can adjust its length and shape in different frames to adapt to voltage fluctuation characteristics.

[0085] The time-frequency graph data is concatenated according to the frame index order to output the complete time-frequency graph data.

[0086] In the time-frequency graph data, the data corresponding to each frame maintains the same sampling time, reference phase, and number of sampling points as the original voltage waveform data under the synchronous clock reference.

[0087] The amplitude spectrum first-order difference, phase spectrum first-order difference, and transient disturbance significance index are calculated on the time-frequency graph data and fused to construct the disturbance graph data. The disturbance graph data is filtered according to the disturbance gating threshold to obtain disturbance gating flag data. When the disturbance gating flag data indicates the presence of a disturbance, the main model inference process is triggered; otherwise, the low-power lightweight operation state is maintained.

[0088] In this embodiment, the disturbance map data is filtered according to the disturbance gating threshold to obtain disturbance gating flag data, including:

[0089] Based on the time-frequency graph data, calculate the magnitude of the time-frequency graph data under each frequency index and each frame index to obtain the amplitude spectrum data, and calculate the phase angle of the time-frequency graph data under each frequency index and each frame index to obtain the phase spectrum data;

[0090] In Example 1, the modulus of the time-frequency graph data under each frequency index and each frame index is calculated based on the time-frequency graph data. The modulus calculation is to treat the time-frequency graph data under the frequency index and the corresponding frame index as complex numbers, calculate the square root of the sum of the squares of the real part and the imaginary part respectively, and then obtain the amplitude spectrum data. The amplitude spectrum data represents the intensity distribution of each frequency component in different frames of the electricity meter voltage input signal.

[0091] Based on the time-frequency graph data, the phase angle is calculated for the time-frequency graph data under each frequency index and each frame index. The phase angle calculation is to treat the time-frequency graph data under the frequency index and the corresponding frame index as complex numbers, and then calculate the arctangent values ​​of the real part and the imaginary part respectively to obtain the phase spectrum data. The phase spectrum data represents the phase information of each frequency component in different frames of the energy meter voltage input signal.

[0092] For the amplitude spectrum data, the difference between the amplitude spectrum data of two adjacent frequency indices is calculated under each frame index according to the frequency index direction to obtain the first-order difference data of the amplitude spectrum;

[0093] The first-order difference data of the amplitude spectrum is used to characterize the energy change of the voltage input signal of the energy meter between different frequency components. If the current frequency index is the starting frequency, the amplitude spectrum data corresponding to the starting frequency is used to calculate the difference between itself and the data.

[0094] For each frame index, the difference between the phase spectrum data of two adjacent frequency indices is calculated according to the frequency index direction to obtain the first-order difference data of the phase spectrum.

[0095] The first-order differential phase spectrum data is used to characterize the phase change of the voltage input signal of the energy meter between different frequency components. If the current frequency index is the starting frequency, the phase spectrum data corresponding to the starting frequency is subtracted from itself.

[0096] The absolute value of the difference between the time-frequency plot data under each frame index and the time-frequency plot data of the corresponding frequency index under the previous frame index is calculated and accumulated to obtain the transient disturbance significance index.

[0097] The transient disturbance significance index is used to characterize the overall dynamic change of the electricity meter voltage input signal between adjacent frames. If the current frame index is the starting frame, the time-frequency graph data corresponding to the starting frame index is used to calculate the difference between itself and the data.

[0098] The first-order difference data of amplitude spectrum, the first-order difference data of phase spectrum, and the transient perturbation significance index are weighted and summed to obtain the perturbation map data;

[0099] The disturbance plot data is used to comprehensively represent the disturbance characteristics of the electricity meter's voltage input signal in the frequency domain and time-frequency domain.

[0100] Set a perturbation gating threshold to distinguish perturbation map data for all frequency indices under each frame index:

[0101] When the perturbation map data at any sampling point is greater than or equal to the perturbation gating threshold, a perturbation gating flag of one is generated, triggering the main model inference process;

[0102] If the disturbance map data of all sampling points is less than the disturbance gating threshold, the disturbance gating flag data is generated as zero, maintaining a low-power and lightweight operating state.

[0103] The perturbation map data is input into the improved DenseNet deep discriminant network, which consists of reconfigured dense connection blocks, bottleneck layers, depth-separable convolutions, and group convolutions. Inference is performed on the embedded SoC platform, and the initial voltage perturbation discrimination result data and confidence data are output.

[0104] In this embodiment, the DenseNet depth discriminant network is improved, including:

[0105] The perturbation map data is rearranged into input tensors according to three feature channels: first-order difference of amplitude spectrum, first-order difference of phase spectrum, and transient perturbation significance. The frequency band weighting vector is calculated according to the ratio of the cumulative value of the perturbation map data at each frequency index in all frames to the cumulative value of the overall perturbation map data. The weighted tensor is obtained by multiplying the perturbation map data at each frequency index with the frequency band weighting vector corresponding to the frequency index.

[0106] In Example 1, three feature channels—first-order difference of amplitude spectrum, first-order difference of phase spectrum, and transient perturbation significance—are extracted from the perturbation map data. These three feature channels are arranged in a channel-first manner to form a three-channel input tensor. For each frequency index in the three-channel input tensor, the cumulative value of the perturbation map data at that frequency index under all frame indices is calculated to obtain the cumulative value of each frequency index. The cumulative value of the perturbation map data at all frequency indices and all frame indices of the three-channel input tensor is calculated to obtain the overall cumulative value. The cumulative value of each frequency index is compared with the overall cumulative value to obtain the frequency band weighting vector corresponding to each frequency index. The perturbation map data at each frequency index in the three-channel input tensor is multiplied with the frequency band weighting vector corresponding to the frequency index. The product result replaces the original value at the original frequency index to obtain the frequency band weighted weighted tensor.

[0107] The weighted tensor is input into the first bandwidth-sensing dense block, which comprises multiple bandwidth-sensing dense units, and the output features are... ;

[0108] Each bandwidth-aware dense cell dynamically sets its growth rate based on the weighted result of the base growth rate and the frequency band weighting vector corresponding to the current frequency index. :

[0109] ;

[0110] in, Based on the growth rate, The amplification factor is adjustable. This is the frequency band weighting vector.

[0111] For output features Channel compression is achieved through a bottleneck compression layer, which uses a one-dimensional convolution operation to reduce the number of channels in the output features, resulting in channel-compressed features.

[0112] Based on the variance of each channel under all frequency indices and all frame indices according to the channel compression features, a sparsification mask is generated. The sparsification mask marks channels with variance greater than or equal to the sparsity threshold as one, and otherwise as zero. The sparsification mask is used to dynamically filter and prune the convolution kernel to output sparsification features.

[0113] In Example 1, the disturbance gating flag value and disturbance map data can also be jointly input into the improved DenseNet network, and a logical "OR" triggering mechanism can be used to execute the disturbance determination activation process. The disturbance activation signal is triggered when any of the following conditions are met:

[0114] (1) The disturbance gating flag is in the triggered state;

[0115] (2) There are sampling points in the first-order difference data of the amplitude spectrum of any frame in the disturbance map data that exceed the preset significance threshold, or the transient disturbance significance index exceeds the disturbance significance threshold;

[0116] The logical "OR" triggering mechanism ensures that the disturbance judgment process can be triggered independently when any judgment condition is met, thereby improving the ability to quickly identify abnormal disturbances.

[0117] The sparse features are input into the depthwise separable convolutional unit to obtain the depthwise separable convolutional features, which are then fed into the group convolutional module. The group convolutional module divides all channels into multiple groups and performs convolution operations on each group independently, outputting the group convolutional features of the first stage.

[0118] In Example 1, sparse features are input into a depthwise separable convolutional unit. A spatial convolution operation is performed on each channel individually to obtain the spatial structure features of each channel. This operation does not change the number of channels. Then, the spatial structure features of all channels are weighted and summed with the pointwise convolution kernels according to the channels to achieve information fusion between all channels and obtain depthwise separable convolutional features. The depthwise separable convolutional features are input into a group convolutional module. The group convolutional module divides all channels into multiple groups. All channels in each group are independently convolved using the same convolution kernel parameters. The convolution results of each group are concatenated in the original channel order to output the group convolutional features of the first stage.

[0119] Repeat the operation of the first bandwidth-aware dense block and the depth-separable convolutional unit to form multiple sets of convolutional features;

[0120] Configure an early termination discrimination head at several specified stage indices, calculate the classification score vector for the current group's convolutional features, and obtain the probability vector through Softmax. When the maximum value in the probability vector is greater than or equal to the early termination threshold, directly output the early termination voltage perturbation discrimination result and confidence level for the current stage.

[0121] In Example 1, an early termination detection head is configured at several specified stage indices. The early termination detection head performs a global average pooling operation on the convolutional features of the current stage group, averaging the values ​​at all spatial locations along the channel dimension to obtain the feature vector of the current stage. The early termination detection head uses a set of fully connected weights and biases to perform a linear transformation on the current feature vector to obtain the classification score vector of the current stage. The classification score vector of the current stage is input into the Softmax function. The Softmax function maps all classification scores into probability vectors. Each component in the probability vector corresponds to the predicted probability of the voltage disturbance category. When the maximum probability value in the current stage probability vector is greater than or equal to the early termination threshold, the category corresponding to the maximum probability in the probability vector is taken as the early termination voltage disturbance detection result, and the maximum probability value is taken as the early termination confidence. The early termination voltage disturbance detection result and confidence of the current stage are directly output.

[0122] When the maximum probability of all early termination discriminators does not reach the early termination threshold, the same discrimination process is performed on the final stage features, and the initial voltage disturbance discrimination result data and confidence data are output. All calculations are completed and output within one inference cycle of the embedded SoC platform.

[0123] Determine whether the confidence data is higher than the early termination threshold. If it is higher than the early termination threshold, directly output the early termination discrimination result data. Otherwise, continue to execute the remaining dense connection blocks of the improved DenseNet deep discrimination network to obtain the final discrimination result data. Perform a monotonic mapping between the final discrimination result data and the preset power quality threshold mapping table to generate voltage disturbance event-level judgment data.

[0124] In this embodiment, the generation of voltage disturbance event-level determination data includes:

[0125] Based on the discrimination results of each stage, the probability vector of the current stage is obtained. The component with the largest value in the probability vector of the current stage is taken as the confidence data of the current stage, and the voltage disturbance category corresponding to the component is taken as the voltage disturbance category discrimination result of the current stage.

[0126] In the current stage probability vector, each component corresponds to the predicted probability of each voltage disturbance category. The component with the largest value in the current stage probability vector is taken as the confidence level data for the current stage.

[0127] The confidence data of the current stage is compared with the early termination threshold. When the confidence data of the current stage is greater than or equal to the early termination threshold, the early termination discrimination result data and confidence data of the current stage are directly output, and the calculation of the remaining dense connection blocks of the improved DenseNet deep discrimination network is stopped.

[0128] When the confidence data of the current stage is less than the early termination threshold, the output features of the current stage are continued to be input into the remaining dense connection blocks of the improved DenseNet deep discriminant network to obtain the probability vector of the final stage. The component with the largest value in the probability vector of the final stage is taken as the confidence data of the final stage, and the voltage disturbance category corresponding to the component is taken as the voltage disturbance category discrimination result of the final stage.

[0129] In the final stage probability vector, each component corresponds to the predicted probability of each voltage disturbance category.

[0130] The early termination judgment result data is monotonically mapped to the preset power quality threshold mapping table to obtain voltage disturbance event-level judgment data.

[0131] Voltage disturbance event-level determination data includes disturbance type, disturbance level, and disturbance duration. The disturbance type is equal to the voltage disturbance category determination result of the current stage or the final stage.

[0132] In Example 1, all frame indices can be traversed based on the disturbance map data. The voltage disturbance category discrimination result corresponding to each frame index is compared with the disturbance type. For all frame indices whose discrimination category and disturbance type are consistent, their corresponding time lengths are accumulated. The time lengths of all consistent frames are summed to obtain the disturbance duration. The time length of a single frame is equal to the product of the number of sampling points in each frame and the sampling period.

[0133] The maximum amplitude of the disturbance map data under all frequency indices and all frame indices is counted in the disturbance map data. The maximum amplitude is compared with the disturbance gating threshold, and the disturbance level is determined according to the preset disturbance level mapping relationship.

[0134] When the confidence data is lower than the confidence threshold or the improved DenseNet deep discriminant network fails to perform self-testing, the frequency domain threshold rule determination process is executed, and the rollback determination data is output.

[0135] In this embodiment, the frequency domain threshold rule determination process includes:

[0136] Obtain the confidence data output from the final discrimination stage, compare the confidence data with the preset confidence threshold, and start the frequency domain threshold rule determination process when the confidence data is less than the preset confidence threshold, or when the improved DenseNet deep discrimination network self-test module returns an error flag.

[0137] In Example 1, when the improved DenseNet depth discriminative network self-test module returns an error flag, it includes:

[0138] When the self-testing module of the improved DenseNet deep discriminant network triggers an error flag due to any of the following conditions: model quantization overflow, structural integrity verification failure, perturbation graph dimension mismatch, intermediate inference anomaly, or self-testing process timeout, it is considered a self-test failure, and the frequency domain backoff judgment process is executed.

[0139] In the frequency domain threshold rule determination process, the first-order difference data of amplitude spectrum and transient disturbance significance index are extracted from the disturbance map data, and the average disturbance intensity index is calculated.

[0140] The average perturbation intensity index is equal to the sum of the first-order difference data of the amplitude spectrum corresponding to all frequency indices in the current frame, divided by the total number of frequency indices. The average perturbation intensity index is used to represent the overall level of amplitude perturbation in the current frame.

[0141] Set the amplitude spectrum perturbation threshold, transient perturbation threshold, and average perturbation intensity threshold, and perform the following judgment logic for each frame index:

[0142] If there are sampling points in the first-order difference data of the amplitude spectrum that are greater than or equal to the amplitude spectrum disturbance threshold, or the transient disturbance significance index is greater than or equal to the transient disturbance threshold, or the average disturbance intensity index is greater than or equal to the average disturbance intensity threshold, then the current frame is determined to be a voltage disturbance frame; otherwise, it is determined to be a normal frame.

[0143] All frame determination results are merged to obtain a set of disturbance event frames. The disturbance type is then mapped according to the main channel of the disturbance triggering characteristics.

[0144] If the average value of the first-order difference data of the amplitude spectrum in the current frame is greater than or equal to the average absolute value of the first-order difference data of the phase spectrum, then the current disturbance type is determined to be an amplitude disturbance; otherwise, it is determined to be a phase disturbance.

[0145] The set of disturbance event frames is a set of frame indices that are all identified as voltage disturbance frames;

[0146] The duration of the disturbance is obtained by multiplying the number of frames in the set of disturbance event frames and the frame duration for each frame.

[0147] The disturbance duration is used to represent the time span of a voltage disturbance event under the backoff decision.

[0148] Within the set of disturbance event frames, find the maximum disturbance amplitude value in the disturbance map data, and input the disturbance type, disturbance duration, and maximum disturbance amplitude value into the preset backoff level mapping function to obtain the disturbance level;

[0149] In Example 1, within the set of disturbance event frames, each disturbance event frame index is traversed. For each disturbance event frame index, all frequency indices are traversed to find the value of the disturbance map data corresponding to all frequency indices under the current frame. The maximum value is selected from all disturbance event frame indices and all frequency indices, and the obtained maximum value is taken as the maximum disturbance amplitude value.

[0150] The disturbance type, disturbance duration, and maximum disturbance amplitude are sequentially input into a preset backoff level mapping function. The preset backoff level mapping function searches for or calculates the corresponding disturbance level based on the combination of disturbance type, disturbance duration, and maximum disturbance amplitude. The disturbance level is used to reflect the severity of the voltage disturbance event under the backoff judgment.

[0151] The backoff level mapping function is used to match and search the disturbance type, disturbance duration, and maximum disturbance amplitude value according to a preset leveling threshold table and leveling rules to obtain the corresponding disturbance level. The disturbance level reflects the severity of the voltage disturbance event under the backoff judgment.

[0152] Output backoff determination data, which includes disturbance type, disturbance duration, and disturbance level.

[0153] The rollback decision data is used to replace the final output when the DenseNet discrimination process fails, and ensures that the rollback decision data has interpretability, mappability and audit traceability capabilities while being aligned with the preset power quality rule table structure.

[0154] A voltage fluctuation sensing energy meter that integrates a deep discriminant network is provided, which has a built-in storage system for executing a determination method for the voltage fluctuation sensing energy meter that integrates a deep discriminant network.

[0155] Example 2: At a power distribution network user site, the electricity meter has long been responsible for real-time monitoring and metering. One day, the electricity meter received a 220V AC voltage signal. The signal had a total of 5000 sampling points within a certain time period, with a sampling frequency of 5kHz. After signal preprocessing, it underwent digital anti-aliasing filtering, automatic gain adjustment, and DC bias correction to ensure that the signal amplitude of all sampling points fell within the normalized range of [-1, 1]. At this time, the electricity meter system received a set of raw voltage waveform data, with an average amplitude of 218.7V, a maximum amplitude of 233.4V, and a minimum amplitude of 207.1V.

[0156] The electricity meter divides the raw waveform data into 256 sampling points per frame, obtaining 19 data segments. All data undergoes power frequency phase-locked loop (PLL) and phase alignment, followed by short-time Fourier transform (SFT) to generate time-frequency graph data with a length of 128 frames per segment. Taking frame 7 as an example, the variation characteristics of the main frequency band of its time-frequency graph amplitude spectrum are: [9.2, 11.6, 13.3, 15.7, 12.1, 9.8]V (corresponding to different frequency indices), the first-order difference of the phase spectrum is: [0.09, 0.13, 0.18, 0.11, 0.07, 0.08], and the transient disturbance significance index is 0.41. The disturbance gating threshold of frame 7 is 0.36, which significantly exceeds the preset gating threshold (0.28).

[0157] After the disturbance gating flag data is triggered, the electricity meter system automatically activates the improved DenseNet deep discrimination network. At this time, the three-channel disturbance map data are input respectively. The peak value of channel 1 (first-order difference of amplitude spectrum) reaches 6.7 in the main frequency band, the average value of channel 2 (first-order difference of phase spectrum) is 0.13, and the value of channel 3 (transient disturbance significance) is 0.41. The bandwidth-aware dense block automatically increases the growth rate of the current frequency band, the group convolution module allocates more computing resources to the high-energy channels, the network is quantized into an INT8 structure, and the single-frame inference time is about 4.1ms.

[0158] The DenseNet network's discriminant head outputs an early termination confidence score of 0.995, directly returning a "flicker + transient disturbance" discrimination result with a confidence interval of [0.970, 1.000]. The disturbance type is classified as Level III, and the disturbance duration is 0.23s. The system also generates traceable evidence: the main disturbance frequency band is indexed as 3-6, corresponding to the actual frequency range of 115Hz-230Hz, and the start and end frame numbers of the disturbance event are 6-9.

[0159] For comparison, a traditional threshold method was used to judge the same set of waveform data, with a preset effective value threshold of ±10% and a harmonic amplitude threshold of 10V. The system only detected an effective value change of 7.6V and a 5th harmonic amplitude of 8.3V in the 5th frame, neither of which exceeded the threshold, and no alarm was triggered. Subsequently, from the 9th to the 11th frame, although the peak amplitude briefly reached 11.4V, the duration was less than 0.2s, and it was not judged as a disturbance event by the traditional rule method. Ultimately, the traditional method failed to record the entire process of this actual sudden disturbance, missing the opportunity for early warning of the on-site anomaly.

[0160] Another set of comparative data shows that when the electricity meter encounters a short-term high-power load (air conditioner, water heater) that starts simultaneously, the highest amplitude of the acquired signal is 231.5V, and the lowest amplitude is 190.2V. Traditional methods mainly focus on the effective value offset, detecting an instantaneous offset of 15.6V, only indicating "short-term fluctuation," with a judgment level of I, without detailed information such as frequency band and duration. In contrast, in the method of this invention, the first-order differential channel of the amplitude spectrum reaches a peak of 10.2 in the 12th frame, the transient disturbance significance is 0.67, and the phase spectrum change is 0.15. After DenseNet inference, the output disturbance type is "temporary sag," the disturbance level is II, the duration is 0.36s, the judgment time is 4.2ms, and the main frequency band of the disturbance is clearly marked as 25Hz~75Hz. The relevant judgment information is archived synchronously, meeting the needs of subsequent on-site operation and maintenance and audit traceability.

[0161] In more extreme scenarios, the electricity meter acquires signal segments superimposed with high-frequency noise. Traditional methods suffer from two misjudgments due to noise interference, misidentifying normal fluctuations as harmonic disturbances. Actual statistical samples show that in 5000 fluctuation samples, the traditional method has a false alarm rate of 8.3%, a false negative rate of 5.9%, and a judgment time of approximately 1.2ms. Using the method of this invention, the false alarm rate is reduced to 0.7% and the false negative rate is reduced to 0.2% in the same data set, with an average judgment time of 4.4ms. All disturbance events can be accurately output with disturbance type, dominant frequency band, level, and duration, without any missing information.

[0162] In the fallback mechanism-triggered scenario, due to a storage self-test failure in the embedded SoC, the deep network self-test module returns an error flag. At this point, the system automatically switches to the frequency domain threshold rule-based judgment process, extracting three indicators: the first-order difference of the amplitude spectrum, the significance of transient disturbances, and the average disturbance intensity. After setting a threshold, it is found that the disturbance amplitude in frames 15-17 exceeds the threshold, which the system classifies as an "amplitude-type disturbance." The disturbance duration is 0.11s, and the maximum disturbance amplitude is 9.5. The disturbance level is determined as Level II by looking up a table using a preset function. The system automatically outputs the fallback judgment result, ensuring no event is missed.

[0163] Based on a comprehensive analysis of various scenarios and extensive field comparisons, the data demonstrates that the method of this invention achieves high accuracy, low false negatives, interpretable and auditable outputs for complex voltage disturbance events while maintaining controllable inference speed and system energy consumption. This significantly enhances the intelligent sensing capabilities of electricity meters and improves industry compliance.

[0164] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for determining voltage fluctuation sensing energy meters that integrates deep discriminant networks, characterized in that, include: The voltage input signal inside the energy meter is read and preprocessed to obtain the original voltage waveform data under the synchronous clock reference. Based on the power frequency phase-locked loop results, the original voltage waveform data is divided into continuous frames according to a fixed number of sampling points, and voltage time window data segments are output. Adaptive window length short-time Fourier transform is applied to the voltage time window data segments to generate time-frequency graph data. The amplitude spectrum first-order difference, phase spectrum first-order difference, and transient disturbance significance index are calculated on the time-frequency graph data and fused to construct the disturbance graph data. The disturbance graph data is filtered according to the disturbance gating threshold to obtain disturbance gating flag data. When the disturbance gating flag data indicates the presence of a disturbance, the main model inference process is triggered; otherwise, the low-power lightweight operation state is maintained. The perturbation map data is input into the improved DenseNet deep discriminant network, which consists of reconfigured dense connection blocks, bottleneck layers, depth-separable convolutions, and group convolutions. Inference is performed on the embedded SoC platform, and the initial voltage perturbation discrimination result data and confidence data are output. Determine whether the confidence data is higher than the early termination threshold. If it is higher than the early termination threshold, directly output the early termination discrimination result data. Otherwise, continue to execute the remaining dense connection blocks of the improved DenseNet deep discrimination network to obtain the final discrimination result data. Perform a monotonic mapping between the final discrimination result data and the preset power quality threshold mapping table to generate voltage disturbance event-level judgment data. When the confidence data is lower than the confidence threshold or the improved DenseNet deep discriminant network fails to self-test, the frequency domain threshold rule judgment process is executed, and the backoff judgment data is output. The step of filtering the disturbance map data based on the disturbance gating threshold to obtain disturbance gating flag data includes: Based on the time-frequency graph data, the magnitude values ​​of the time-frequency graph data under each frequency index and each frame index are calculated to obtain the amplitude spectrum data. The phase angle of the time-frequency graph data under each frequency index and each frame index is also calculated to obtain the phase spectrum data. For the amplitude spectrum data, the difference between the amplitude spectrum data of two adjacent frequency indices is calculated under each frame index according to the frequency index direction to obtain the first-order difference data of the amplitude spectrum; For each frame index, the difference between the phase spectrum data of two adjacent frequency indices is calculated according to the frequency index direction to obtain the first-order difference data of the phase spectrum. The absolute value of the difference between the time-frequency plot data under each frame index and the time-frequency plot data of the corresponding frequency index under the previous frame index is calculated and accumulated to obtain the transient disturbance significance index. The first-order difference data of amplitude spectrum, the first-order difference data of phase spectrum, and the transient perturbation significance index are weighted and summed to obtain the perturbation map data; Set a perturbation gating threshold to distinguish perturbation map data for all frequency indices under each frame index: When the perturbation map data at any sampling point is greater than or equal to the perturbation gating threshold, a perturbation gating flag of one is generated, triggering the main model inference process; If the disturbance map data of all sampling points is less than the disturbance gating threshold, the disturbance gating flag data is generated as zero, maintaining a low-power and lightweight operating state.

2. The method for determining voltage fluctuation sensing energy meters based on a fused deep discriminant network according to claim 1, characterized in that, The preprocessing includes: At a preset sampling frequency, the voltage input signal inside the energy meter is read through an analog-to-digital converter to obtain the original voltage waveform data; A digital anti-aliasing filter is applied to the raw voltage waveform data to perform filtering calculations with predetermined numerator and denominator coefficients, and the anti-aliasing filtered output data is output. The adaptive gain factor is calculated by the ratio of the reference voltage amplitude to the maximum absolute value of the anti-aliasing filter output data in the current frame. The adaptive gain factor is then used to amplify or reduce each sampling point in the anti-aliasing filter output data to obtain gain correction data. The DC bias is obtained by taking the arithmetic mean of the values ​​of all sample points in the current frame of the gain correction data. The DC bias is then subtracted from each sample point in the gain correction data to obtain the bias correction data. The sampling time corresponding to each sampling point is determined according to the sampling synchronization condition. All sampling points in the bias correction data are aligned with the detected power frequency reference phase and time-marked, and the original voltage waveform data under the synchronous clock reference is output.

3. The method for determining voltage fluctuation sensing energy meters based on a fused deep discriminant network according to claim 2, characterized in that, The adaptive window length short-time Fourier transform of the voltage time window data segment includes: The original voltage waveform data under the synchronous clock reference is divided according to the number of sampling points in each frame to obtain voltage time window data segments; Based on the reference phase obtained from the power frequency phase-locked loop result, phase alignment is performed on each voltage time window data segment to obtain the phase-aligned voltage time window data segment. Short-time Fourier transform is performed on the phase-aligned voltage time window data segments using an adaptive window function to obtain the time-frequency plot data for each frame; The time-frequency graph data is concatenated according to the frame index order to output the complete time-frequency graph data.

4. The method for determining voltage fluctuation sensing energy meters based on a fused deep discriminant network according to claim 1, characterized in that, The improved DenseNet deep discriminative network includes: The perturbation map data is rearranged into input tensors according to three feature channels: first-order difference of amplitude spectrum, first-order difference of phase spectrum, and transient perturbation significance. The frequency band weighting vector is calculated according to the ratio of the cumulative value of the perturbation map data at each frequency index in all frames to the cumulative value of the overall perturbation map data. The weighted tensor is obtained by multiplying the perturbation map data at each frequency index with the frequency band weighting vector corresponding to the frequency index. The weighted tensor is input into the first bandwidth-aware dense block, which includes multiple bandwidth-aware dense units, and the output features are generated. The output features are compressed through a bottleneck compression layer. The bottleneck compression layer uses a one-dimensional convolution operation to reduce the number of channels in the output features, resulting in channel-compressed features. Based on the variance of each channel under all frequency indices and all frame indices according to the channel compression features, a sparsification mask is generated. The sparsification mask marks channels with variance greater than or equal to the sparsity threshold as one, and otherwise as zero. The sparsification mask is used to dynamically filter and prune the convolution kernel to output sparsification features. The sparse features are input into the depthwise separable convolutional unit to obtain the depthwise separable convolutional features, which are then fed into the group convolutional module. The group convolutional module divides all channels into multiple groups and performs convolution operations on each group independently, outputting the group convolutional features of the first stage. Repeat the operation of the first bandwidth-aware dense block and the depth-separable convolutional unit to form multiple sets of convolutional features; Configure an early termination discrimination head at several specified stage indices, calculate the classification score vector for the current group's convolutional features, and obtain the probability vector through Softmax. When the maximum value in the probability vector is greater than or equal to the early termination threshold, directly output the early termination voltage perturbation discrimination result and confidence level for the current stage. When the maximum probability of all early termination discriminators does not reach the early termination threshold, the same discrimination process is performed on the final stage features, and the initial voltage disturbance discrimination result data and confidence data are output. All calculations are completed and output within one inference cycle of the embedded SoC platform.

5. The method for determining voltage fluctuation sensing energy meters based on a fused deep discriminant network according to claim 4, characterized in that, The generated voltage disturbance event-level determination data includes: Based on the discrimination results of each stage, the probability vector of the current stage is obtained. The component with the largest value in the probability vector of the current stage is taken as the confidence data of the current stage, and the voltage disturbance category corresponding to the component is taken as the voltage disturbance category discrimination result of the current stage. The confidence data of the current stage is compared with the early termination threshold. When the confidence data of the current stage is greater than or equal to the early termination threshold, the early termination discrimination result data and confidence data of the current stage are directly output, and the calculation of the remaining dense connection blocks of the improved DenseNet deep discrimination network is stopped. When the confidence data of the current stage is less than the early termination threshold, the output features of the current stage are continued to be input into the remaining dense connection blocks of the improved DenseNet deep discriminant network to obtain the probability vector of the final stage. The component with the largest value in the probability vector of the final stage is taken as the confidence data of the final stage, and the voltage disturbance category corresponding to the component is taken as the voltage disturbance category discrimination result of the final stage. The early termination judgment result data is monotonically mapped to the preset power quality threshold mapping table to obtain voltage disturbance event-level judgment data. Voltage disturbance event-level determination data includes disturbance type, disturbance level, and disturbance duration. The disturbance type is equal to the voltage disturbance category determination result of the current stage or the final stage.

6. The method for determining voltage fluctuation sensing energy meters based on a fused deep discriminant network according to claim 5, characterized in that, The execution frequency domain threshold rule determination process includes: Obtain the confidence data output from the final discrimination stage, compare the confidence data with the preset confidence threshold, and start the frequency domain threshold rule determination process when the confidence data is less than the preset confidence threshold, or when the improved DenseNet deep discrimination network self-test module returns an error flag. In the frequency domain threshold rule determination process, the first-order difference data of amplitude spectrum and transient disturbance significance index are extracted from the disturbance map data, and the average disturbance intensity index is calculated. Set the amplitude spectrum perturbation threshold, transient perturbation threshold, and average perturbation intensity threshold, and perform the following judgment logic for each frame index: If there are sampling points in the first-order difference data of the amplitude spectrum that are greater than or equal to the amplitude spectrum disturbance threshold, or the transient disturbance significance index is greater than or equal to the transient disturbance threshold, or the average disturbance intensity index is greater than or equal to the average disturbance intensity threshold, then the current frame is determined to be a voltage disturbance frame; otherwise, it is determined to be a normal frame. All frame determination results are merged to obtain a set of disturbance event frames. The disturbance type is then mapped according to the main channel of the disturbance triggering characteristics. If the average value of the first-order difference data of the amplitude spectrum in the current frame is greater than or equal to the average absolute value of the first-order difference data of the phase spectrum, then the current disturbance type is determined to be an amplitude disturbance; otherwise, it is determined to be a phase disturbance. The set of disturbance event frames is a set of frame indices that are all identified as voltage disturbance frames; The duration of the disturbance is obtained by multiplying the number of frames in the set of disturbance event frames and the frame duration for each frame. Within the set of disturbance event frames, find the maximum disturbance amplitude value in the disturbance map data, and input the disturbance type, disturbance duration, and maximum disturbance amplitude value into the preset backoff level mapping function to obtain the disturbance level; Output backoff determination data, which includes disturbance type, disturbance duration, and disturbance level.

7. A voltage fluctuation sensing energy meter incorporating a deep discriminant network, characterized in that, It is equipped with a storage system for executing a method for determining voltage fluctuation sensing energy meters that integrates a deep discriminant network as described in any one of claims 1-6.

Images