Method and system for identifying communication interference source of electric energy meter under multi-scale features and medium

Abstract

The application discloses a method and system for identifying communication interference sources of electric energy meters under multi-scale features and a medium, and relates to the technical field of signal identification.The method comprises the following steps: collecting and extracting electric energy meter communication signals, constructing a time-frequency spectrum matrix through multi-scale time-frequency transformation; performing time-frequency domain feature analysis based on the multi-scale time-frequency spectrum matrix to generate a multi-scale feature set; performing multi-scale fusion identification based on the multi-scale feature set to obtain a fusion feature vector; and performing interference source classification identification according to the fusion feature vector to determine the identification result of the communication interference sources of the electric energy meters.The application solves the technical problem that the existing method for identifying communication interference sources of electric energy meters cannot comprehensively and accurately capture the features of interference signals under different scales, resulting in low identification accuracy, and achieves the technical effects of complementary fusion of multi-scale time-frequency features, accurate classification and identification of multiple types of communication interference sources, and improvement of the communication anti-interference capability and identification reliability of electric energy meters in complex electromagnetic environments.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, and in particular to a method, system and medium for identifying communication interference sources of electricity meters under multi-scale characteristics. Background Technology

[0002] With the deepening of smart grid construction, electricity meters, as the core terminal of the electricity consumption information collection system, directly affect the normal operation of remote meter reading, fee control management, load monitoring, and other services due to their communication reliability. Currently, electricity meter communication modules widely adopt communication methods such as power line carrier, low-power wireless, and RS-485, facing increasingly complex electromagnetic interference problems in actual deployment environments. Various interference sources, including frequency converters, switching power supplies, arc discharge, and co-channel wireless interference, exhibit significant differences in time scale and spectral characteristics: transient pulse interference is characterized by strong suddenness and short duration; periodic sudden interference exhibits regular pulse group characteristics; and continuous narrowband interference shows stable spectral occupancy. These interference signals highly overlap with normal communication signals in the time and frequency domains, leading to increased communication error rates and decreased transmission success rates, severely affecting the accuracy and real-time performance of electricity consumption information collection.

[0003] Existing interference identification technologies typically employ single-resolution time-frequency analysis methods, such as fixed-window-length short-time Fourier transforms or fixed-scale power spectrum estimation, which struggle to simultaneously capture interference features at different time scales. Fine-grained analysis windows can capture transient details but lose long-term periodic features, while coarse-grained analysis windows can present continuous spectrum structures but smooth out transient abrupt changes. Furthermore, existing methods often employ simple splicing during feature fusion, lacking the ability to adaptively utilize the complementary value of features at different scales. This results in insufficient accuracy and poor robustness in identifying multi-type mixed interference in complex electromagnetic environments, failing to meet the practical needs of electricity meters for accurate identification of interference source types. Summary of the Invention

[0004] This invention provides a method, system, and medium for identifying communication interference sources in electricity meters under multi-scale features. It addresses the technical problem that existing methods for identifying communication interference sources in electricity meters are unable to comprehensively and accurately capture the characteristics of interference signals at different scales, resulting in low identification accuracy. The invention achieves the technical effect of using complementary fusion of multi-scale time and frequency features to realize accurate classification and identification of multiple types of communication interference sources, thereby improving the anti-interference capability and identification reliability of electricity meter communication in complex electromagnetic environments.

[0005] In a first aspect, the present invention provides a method for identifying communication interference sources of electricity meters under multi-scale features, wherein the method for identifying communication interference sources of electricity meters under multi-scale features includes:

[0006] The communication signals from the electricity meter's communication module are collected and extracted to obtain the signal data to be identified. Multi-scale time-frequency transformation is then performed to construct a multi-scale time-frequency spectrum matrix. Based on this matrix, time-frequency domain feature analysis is conducted to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. Multi-scale fusion identification is then performed based on these features to obtain a fused feature vector. Finally, interference source classification and identification are performed based on the fused feature vector to determine the communication interference source identification result for the electricity meter.

[0007] Secondly, the present invention also provides a multi-scale feature-based system for identifying communication interference sources in electricity meters, wherein the multi-scale feature-based system for identifying communication interference sources in electricity meters includes: Signal Processing Unit: Collects communication signals from the electricity meter's communication module, extracts the signals, obtains the signal data to be identified, performs multi-scale time-frequency transformation, and constructs a multi-scale time-spectrum matrix; Feature Analysis Unit: Performs time-frequency domain feature analysis based on the multi-scale time-spectrum matrix to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features; Multi-scale Fusion Unit: Performs multi-scale fusion identification based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features to obtain a fused feature vector; Interference Source Identification Unit: Classifies and identifies interference sources based on the fused feature vector to determine the communication interference source identification result of the electricity meter.

[0008] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for identifying communication interference sources of electricity meters under multi-scale features provided by the present invention.

[0009] This invention discloses a method, system, and medium for identifying communication interference sources in electricity meters under multi-scale characteristics, including: The invention involves collecting communication signals from the communication module of an energy meter, extracting the signal data to be identified, performing multi-scale time-frequency transformation, and constructing a multi-scale time-frequency spectrum matrix. Based on this matrix, time-frequency domain feature analysis is performed to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. Multi-scale fusion identification is then performed based on these features to obtain a fused feature vector. Interference source classification and identification are performed according to the fused feature vector to determine the energy meter's communication interference source identification result. This invention, through its multi-scale feature-based energy meter communication interference source identification method, system, and medium, solves the technical problem of existing energy meter communication interference source identification methods' inability to comprehensively and accurately capture interference signal features at different scales, resulting in low identification accuracy. It achieves the technical effect of utilizing complementary fusion of multi-scale time-frequency features to realize accurate classification and identification of multiple types of communication interference sources, improving the energy meter's anti-interference capability and identification reliability in complex electromagnetic environments. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the method for identifying communication interference sources in electricity meters under multi-scale features according to the present invention.

[0011] Figure 2 This is a schematic diagram of the structure of the electricity meter communication interference source identification system under multi-scale features according to the present invention.

[0012] Figure labeling: Signal processing unit 11, Feature analysis unit 12, Multi-scale fusion unit 13, Interference source identification unit 14. Detailed Implementation

[0013] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0014] Example 1, as Figure 1 This is a flowchart illustrating the method for identifying communication interference sources in energy meters under multi-scale features according to the present invention. The method includes: The communication signals from the electricity meter's communication module are collected and extracted to obtain the signal data to be identified. Multi-scale time-frequency transformation is then performed to construct a multi-scale time-frequency matrix.

[0015] Specifically, the communication signals received by the electricity meter's communication module are first collected. These signals are typically transmitted during data exchange between the electricity meter and the backend system and may be subject to various noises and interferences. Collecting these signals yields the raw digital signal sequence. Subsequently, signal extraction is performed on the collected raw digital signal sequence to extract valuable information, forming the signal data to be identified. This signal data contains the characteristics of signals that may be affected by interference. Next, multi-scale time-frequency transformation is used to process the signal to be identified. That is, time-frequency analysis is performed on the signal under the first, second, and third transformation scale parameters to reveal fine-grained changes, medium-scale pulse characteristics, and coarse-scale continuous spectrum characteristics, thereby more comprehensively capturing the complex changes in the signal. Finally, using the feature data obtained from these multi-scale time-frequency transformations, a multi-scale time-spectrum matrix is ​​constructed to provide data support for subsequent analysis and interference source identification, thereby improving the accuracy and reliability of electricity meter communication interference source identification.

[0016] In some embodiments, the communication signals of the electricity meter's communication module are collected and signal extraction is performed to obtain the signal data to be identified. The method includes: The front-end receiving link of the positioning energy meter communication module is equipped with multiple signal acquisition nodes; multiple communication signals are acquired through the multiple signal acquisition nodes, and the multiple communication signals are converted from analog to digital according to a preset sampling rate to generate an original digital signal sequence; the original digital signal sequence is divided into equal length segments according to a sliding time window to generate multiple signal frames; the multiple signal frames are subjected to amplitude normalization processing to generate a signal frame combination, and the signal frame combination is used as the signal data to be identified.

[0017] Specifically, firstly, based on the actual receiving structure of the electricity meter's communication module, the locations in the front-end receiving link that reflect the interference status of the communication signal are determined, and multiple signal acquisition nodes are set up at these locations. The front-end receiving link includes at least two of the following: a communication signal receiving end, a coupling input end, a filtering and amplification output end, a demodulation input end, or a communication chip signal input end. After setting up the acquisition nodes, multiple communication signals received by the electricity meter's communication module are acquired synchronously or according to a unified time reference through each signal acquisition node. The acquired communication signals undergo anti-aliasing filtering to remove high-frequency components exceeding the target analysis frequency band. Then, the multiple communication signals are converted from analog to digital according to a preset sampling rate to obtain a digital sampling sequence corresponding to each signal acquisition node. This preset sampling rate is set based on the operating frequency band and bandwidth of the electricity meter's communication signal and the frequency range of the interference source to be identified, ensuring that the sampling rate is not less than twice the highest frequency of the target signal, so as to guarantee that the sampled digital signal can completely retain the time-domain variation characteristics of both the communication signal and the interference signal. Subsequently, the digital sampling sequences corresponding to each acquisition node are numbered and aligned in chronological order to form the original digital signal sequence. When there are multiple acquisition nodes, the multiple digital sampling sequences are combined in a channelized manner according to the acquisition node number, so that the sampling values ​​of different acquisition nodes at the same sampling time are stored accordingly, thereby forming a multi-channel original digital signal sequence.

[0018] Next, the sliding time window length and sliding step size are set. The sliding time window length determines the number of sampling points contained in each signal frame, and the sliding step size determines the starting position interval between two adjacent signal frames. The system extracts a fixed length of sampled data from the starting sampling point of the original digital signal sequence in chronological order to form the first signal frame. Then, the sliding time window is moved forward according to the sliding step size to continue extracting the next fixed length of sampled data until the entire original digital signal sequence has been traversed, thereby generating multiple signal frames of equal length. In addition, a preset overlapping area can be set between adjacent signal frames. For example, some repeated sampling points are retained between adjacent signal frames to avoid truncation when transient pulse interference or short-term burst interference happens to be located at the segmentation boundary, thereby improving the completeness and stability of subsequent interference source identification. Then, amplitude normalization is performed on the obtained signal frames. Specifically, for each signal frame, the amplitude value of each sampling point is divided by the maximum absolute amplitude value of that signal frame, ensuring the normalized amplitude value falls within a preset range. Alternatively, the mean of the signal frame can be subtracted first, followed by the standard deviation, to ensure the signal frame satisfies a zero-mean and unit-variance amplitude distribution. Amplitude normalization reduces the impact of gain differences between different acquisition nodes, communication distance variations, signal attenuation, and amplitude drift of the acquisition equipment on subsequent feature extraction results. This allows the subsequent identification process to focus more on the time-frequency characteristics and interference features of the signal, rather than simply being affected by amplitude magnitude. Finally, the multiple signal frames after amplitude normalization are combined according to acquisition time order, acquisition node number, or channel number to generate a signal frame combination. This signal frame combination is then used as the signal data to be identified in the subsequent multi-scale time-frequency transformation step. In this signal data to be identified, each signal frame corresponds to a communication signal segment to be analyzed. Multiple signal frames collectively characterize the signal changes of the energy meter communication module during continuous communication, providing a standardized data foundation for the subsequent construction of a multi-scale time-frequency matrix.

[0019] In some embodiments, the method of obtaining the signal data to be identified, performing multi-scale time-frequency transformation, and constructing a multi-scale time-spectrum matrix includes: A first transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the first transform scale parameter to generate a first time-frequency spectrum. Transient pulse interference signals are captured based on the first time-frequency spectrum to determine a first time-frequency distribution feature. A second transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the second transform scale parameter to generate a second time-frequency spectrum. Periodic burst interference signals are captured based on the second time-frequency spectrum to determine a second time-frequency distribution feature. A third transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the third transform scale parameter to generate a third time-frequency spectrum. Continuous narrowband interference signals are captured based on the third time-frequency spectrum to determine a third time-frequency distribution feature. The first time-frequency spectrum, the second time-frequency spectrum, and the third time-frequency spectrum are stacked according to the first time-frequency distribution feature, the second time-frequency distribution feature, and the third time-frequency distribution feature to construct the multi-scale time-frequency matrix.

[0020] Specifically, after obtaining the signal data to be identified, multiple continuous wavelet transform scales are first set according to the interference type and analysis requirements of the electricity meter communication signal. The continuous wavelet transform is used to convert a one-dimensional time series signal into a two-dimensional time-frequency expression that simultaneously contains time and frequency information. For each signal frame in the signal data to be identified, a preset mother wavelet function is selected as the analysis basis function. This mother wavelet function can be a Morlet wavelet or a complex Gaussian wavelet. By changing the wavelet scale parameter, the wavelet function is scaled on the time axis, thereby obtaining the signal time-frequency distribution at different time and frequency resolutions.

[0021] In multi-scale time-frequency transformation, a first transformation scale parameter is set. This first transformation scale parameter is a relatively small scale parameter, such as 1 to 8, to obtain a higher time resolution. The system performs continuous wavelet transform on the signal data to be identified according to the first transformation scale parameter to obtain the wavelet coefficient matrix at the first scale. Then, the energy value is taken from the wavelet coefficient matrix, that is, the modulus of the wavelet coefficients is taken and then squared to generate the first time spectrum. Since the time window corresponding to the first transformation scale is narrow, it can more sensitively reflect transient components in the signal with short duration and rapid energy change. Therefore, transient pulse interference signals are captured based on the first time spectrum. Specifically, the energy distribution in the first time spectrum is scanned along the time direction to detect time-frequency regions where the energy suddenly increases and the duration is short within a local time interval. When the energy amplitude or energy change rate of a certain time-frequency region exceeds the preset transient judgment threshold, the time-frequency region is judged as a transient pulse interference candidate region, and its occurrence time, center frequency, peak energy, duration, and frequency band coverage are recorded. This determines the first time-frequency distribution characteristics, which are used to characterize the location and energy concentration of transient pulse interference in time and frequency.

[0022] A second transform scale parameter is set, which is a medium-scale parameter, such as 9-32, to achieve a balance between time resolution and frequency resolution. The system performs a continuous wavelet transform on the signal data to be identified according to the second transform scale parameter, obtaining the wavelet coefficient matrix at the second scale, and generates a second time-frequency spectrum based on the wavelet coefficient matrix. Since the second transform scale can take into account both pulse duration and frequency range, it is suitable for capturing periodic burst interference or intermittent pulse interference. Specifically, in the second time-frequency spectrum, recurring high-energy regions are searched along the time direction, and the time interval, repetition count, duration, and frequency position between adjacent high-energy regions are statistically analyzed. When multiple high-energy regions exhibit an approximately periodic or regular distribution on the time axis, they are identified as candidate regions for periodic burst interference, and parameters such as burst period, pulse interval, pulse duration, main energy band, and energy change intensity are extracted to determine the second time-frequency distribution characteristics, which characterize the temporal repetition pattern and frequency distribution range of periodic burst interference.

[0023] A third transform scale parameter is set, which is a relatively large scale parameter, such as 33 to 128, to obtain higher frequency resolution. The system performs continuous wavelet transform on the signal data to be identified according to the third transform scale parameter to obtain the wavelet coefficient matrix at the third scale, and generates a third time-spectrum based on the wavelet coefficient matrix. Since the time window corresponding to the third transform scale is relatively wide, it can highlight stable interference components with long duration and narrow frequency range. Therefore, continuous narrowband interference signals are captured based on the third time-spectrum. Specifically, the energy distribution in the third time-spectrum is analyzed along the frequency direction to detect high-energy regions that persist for a long time and are concentrated in a certain narrow frequency band. When the energy of a certain frequency interval is higher than the preset narrowband judgment threshold at multiple consecutive time points, and the bandwidth is less than the preset bandwidth threshold, the region is judged as a candidate region for continuous narrowband interference, and its center frequency, bandwidth, duration, average energy, spectral stability, and energy ratio are extracted to determine the third time-frequency distribution characteristics, which are used to characterize the frequency position, duration, and spectral concentration of continuous narrowband interference.

[0024] After obtaining the first, second, and third time-frequency spectrograms, a unified processing method was applied to the spectrograms at the three scales. Specifically, the three spectrograms were interpolated and resampled using the same time sampling interval and frequency resolution to ensure consistent time and frequency dimensions. Then, the energy amplitude of each spectrogram was normalized by subtracting the maximum and minimum values, bringing the energy distribution across different scales within a uniform numerical range. Subsequently, based on the first, second, and third time-frequency distribution characteristics, the effective interference regions in the spectrograms at each scale were determined, and the time-frequency energy values ​​and interference location markers of the corresponding regions were added to the spectral representation at the corresponding scale. Finally, the processed first, second, and third time-spectrum maps are stacked in scale order to form a three-dimensional matrix structure, constructing a multi-scale time-spectrum matrix. In this matrix, the first dimension represents the time position, the second the frequency position, and the third the transformation scale. The first scale layer characterizes the fine-grained time-frequency variations of transient impulse interference, the second scale layer characterizes the mesoscale time-frequency patterns of periodic burst interference, and the third scale layer characterizes the coarse-scale spectral distribution of continuous narrowband interference. Through this processing, interference information with different durations, frequency distributions, and energy variations in the signal data to be identified can be uniformly expressed in the same multi-scale time-spectrum matrix, providing a complete data foundation for subsequent time-frequency domain feature analysis and multi-scale feature fusion.

[0025] Based on the multi-scale time-frequency spectrum matrix, time-frequency domain feature analysis is performed to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features.

[0026] Specifically, after obtaining the multi-scale time-frequency spectrum matrix, time-frequency domain feature analysis of the matrix can reveal the performance of communication signals in terms of short-duration abrupt changes, medium-interval pulse variations, and long-term continuous spectrum distribution. For fine-grained analysis, the focus is on whether the signal exhibits sudden energy increases, frequency jumps, or instantaneous impacts within a very short timeframe, thereby extracting fine-grained transient features to reflect transient interference with strong suddenness and short duration. For medium-scale analysis, the focus is on whether the signal exhibits periodic bursts, intermittent pulses, or recurring energy enhancement phenomena within a certain time range, thereby extracting medium-scale pulse features to reflect the changing patterns of periodic or phased interference. For coarse-scale analysis, the focus is on the spectral distribution of the signal over a longer time range, thereby extracting coarse-scale continuous spectrum features to describe interference signals with strong persistence and stability. Finally, the features obtained at different scales are organized and combined to form a multi-scale feature set, which enables the subsequent identification process to focus on both instantaneous changes in the signal and periodic pulse features and continuous spectrum features. This provides a more complete feature basis for accurately determining the type of communication interference source in the electricity meter, improves the ability to distinguish different forms of communication interference signals, and enhances the accuracy and stability of subsequent interference source classification and identification.

[0027] In some embodiments, time-frequency domain feature analysis is performed based on the multi-scale time-spectrum matrix to generate a multi-scale feature set, wherein the multi-scale feature set includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. The method includes: Fine-grained transient feature analysis is performed on the first time-spectrum to generate transient change feature parameters. Based on these transient change feature parameters, a first-scale feature vector is constructed. Mesoscale pulse feature analysis is performed on the second time-spectrum to generate periodic burst feature parameters. Based on these periodic burst feature parameters, a second-scale feature vector is constructed. Coarse-scale continuous spectrum feature analysis is performed on the third time-spectrum to generate continuous spectrum distribution feature parameters. Based on these continuous spectrum distribution feature parameters, a third-scale feature vector is constructed. The first-scale feature vector, the second-scale feature vector, and the third-scale feature vector are combined to generate a multi-scale feature set, which simultaneously includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features.

[0028] Specifically, after obtaining the first, second, and third time-frequency spectrograms, feature analysis is performed on the time-frequency energy distribution at different scales. First, fine-grained transient feature analysis is conducted on the first time-frequency spectrogram. Since the first time-frequency spectrogram is generated from a smaller transformation scale and has high temporal resolution, it is used to capture transient interference components in communication signals that are short-duration and rapidly changing in energy. In this process, the first time-frequency spectrogram is scanned frame by frame along the time direction, extracting the frequency domain energy distribution corresponding to each time position and calculating multiple jump amplitude parameters for adjacent time points. When the jump amplitude parameter near a certain time position is greater than the energy change threshold and its energy amplitude is higher than a preset energy threshold, that position is determined as a transient change location. Subsequently, the corresponding instantaneous energy change rate, pulse width estimate, target time coordinate, and target frequency coordinate are determined around the transient change location, thereby generating transient change feature parameters. These transient change feature parameters are then arranged in a preset order and their values ​​are normalized to construct a first-scale feature vector, used to characterize the time-frequency change features of short-duration interference such as spike pulses, switching transients, and electromagnetic shocks at a fine scale.

[0029] Mesoscale pulse characteristic analysis is performed on the second time-spectrum map. Since the second time-spectrum map is generated at a medium transformation scale, it possesses both temporal and frequency resolution, making it suitable for analyzing burst interference with intermittent, repetitive, or periodic distributions. In this process, time-frequency regions with energy exceeding a preset pulse energy threshold are searched within the second time-spectrum map. High-energy points that are temporally adjacent and have similar frequency distributions are merged into a single pulse candidate region. The occurrence time, end time, duration, center frequency, peak energy, average energy, and bandwidth of multiple pulse candidate regions are then statistically analyzed. Next, the time interval between adjacent pulse candidate regions is calculated, and the consistency of multiple time intervals is used to determine whether they exhibit periodic or regular burst characteristics. If the time interval between adjacent burst regions meets a preset periodic fluctuation range, periodic burst characteristic parameters such as pulse repetition period, pulse interval, number of bursts, duty cycle, and main frequency band position are extracted. These periodic burst characteristic parameters are then standardized and arranged in a preset order to generate a second-scale feature vector, used to characterize the recurrence patterns and energy distribution characteristics of periodic burst interference, intermittent equipment start-stop interference, and periodic communication collision interference at the mesoscale.

[0030] Coarse-scale continuous spectrum feature analysis is performed on the third time-spectrum map. Since the third time-spectrum map is generated by a larger transformation scale and has high frequency resolution, it is used to identify continuous spectrum interference with long duration and relatively stable frequency distribution. In this process, energy statistics are performed on the third time-spectrum map along the frequency direction, calculating the average energy, spectral energy percentage, and duration of each frequency point or band within a continuous time range. When a frequency interval maintains a high energy state continuously within multiple continuous time windows, and its bandwidth is less than or meets a preset bandwidth range, this frequency interval is identified as a candidate region for continuous spectrum interference. Next, continuous spectrum distribution feature parameters such as center frequency, occupied bandwidth, duration, average energy, and peak energy are extracted based on the candidate regions. These parameters are then standardized and arranged in a preset order to form a third-scale feature vector, used to characterize the frequency occupancy and continuous distribution characteristics of continuous narrowband interference, persistent carrier interference, adjacent channel equipment interference, or background electromagnetic noise at a coarse scale.

[0031] Finally, it is determined whether the dimensions of the three scale feature vectors meet the input requirements of the subsequent fusion recognition network. If the dimensions of the feature vectors at different scales are inconsistent, they are converted to the preset dimensions by padding with zeros and truncating. Then, scale identifiers are added to the three scale feature vectors respectively, so that the subsequent recognition process can distinguish different feature sources. During combination, the first-scale feature vector, the second-scale feature vector, and the third-scale feature vector are concatenated into a comprehensive feature matrix in order from fine scale to coarse scale. The resulting multi-scale feature set simultaneously contains fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. It can describe the interference features in the communication signal of the energy meter from three perspectives: short-term mutation, periodic burst, and continuous spectrum occupancy. This provides a complete feature foundation for the subsequent fusion feature vector generation and interference source classification and identification, thereby improving the accuracy and stability of subsequent interference source classification and identification.

[0032] In some embodiments, fine-grained transient feature analysis is performed on the first time-spectrum to generate transient change feature parameters. Based on the transient change feature parameters, a first-scale feature vector is constructed. The method includes: The time dimension of the first time-spectrum is extracted, and the first time-spectrum is scanned frame by frame according to the time dimension to obtain multiple time-spectrum energy data. Based on the multiple time-spectrum energy data, adjacent time points are analyzed to generate multiple transition amplitude parameters. A preset energy change threshold is set, and the multiple transition amplitude parameters are compared with the energy change threshold. Transition amplitude parameters exceeding the energy change threshold are extracted, and the instantaneous energy change rate of the signal is calculated, where the instantaneous energy change rate includes instantaneous abrupt change characteristics. An energy threshold is set based on the first time-spectrum, and the first time-spectrum is traversed according to the energy threshold for searching. Extract the continuous time intervals where the energy amplitude of the first time-spectrum exceeds the energy threshold; locate the start and end times based on the continuous time intervals, calculate the time span based on the start and end times, and generate a pulse width estimate; locate the maximum energy amplitude of the first time-spectrum based on the instantaneous change characteristics, determine the target time-frequency point, perform time-frequency coordinate analysis according to the target time-frequency point, and determine the target time coordinate and target frequency coordinate; integrate the instantaneous energy change rate of the signal, the pulse width estimate, the target time coordinate, and the target frequency coordinate to construct the first scale feature vector.

[0033] Specifically, after obtaining the first time-frequency spectrum, it is represented as a time-frequency energy matrix with time and frequency axes as two-dimensional coordinates. Each element in the matrix represents the energy amplitude at the corresponding time and frequency point. Then, the time dimension information of the first time-frequency spectrum is extracted, and the first time-frequency spectrum is scanned in chronological order. At each time point, the energy amplitude values ​​of all frequency points corresponding to that time point are read, and the maximum value of all frequency point energy amplitude values ​​is processed to obtain the time-frequency energy data corresponding to that time point. This process is repeated for all time points in the first time-frequency spectrum to obtain multiple time-frequency energy data points, which are used to characterize the overall energy change of the first time-frequency spectrum at different time positions. Next, the time-frequency energy data corresponding to two adjacent time points are selected sequentially according to time order, and the absolute value of the difference between the energy data at the later time point and the energy data at the previous time point is calculated, or the relative change ratio between the two is calculated, to obtain the corresponding jump amplitude parameter. The above calculation is repeated for all adjacent time points to generate multiple jump amplitude parameters, which represent the degree of energy change in the first time spectrum at adjacent time positions. When transient pulse interference exists in the communication signal, the jump amplitude parameters will increase significantly near the corresponding time. Then, a preset energy change threshold is read. This energy change threshold is determined based on the energy fluctuation range and noise background level of historical normal communication signals. Multiple jump amplitude parameters are compared with the energy change threshold. If a jump amplitude parameter is greater than the energy change threshold, it is considered that there is a significant energy change between adjacent time points corresponding to that jump amplitude parameter, and this is used as a candidate point for transient change. All jump amplitude parameters exceeding the energy change threshold are then extracted, and the instantaneous energy change rate of the signal containing transient change characteristics is calculated based on the jump amplitude parameters and the corresponding time interval. This rate characterizes the speed at which the energy of the communication signal rises or falls in a short period of time; the larger the value, the higher the probability of a transient change in the signal.

[0034] Furthermore, an energy threshold is set based on the first time-frequency spectrum. This energy threshold is used to filter high-energy regions in the first time-frequency spectrum. The system traverses and searches the first time-frequency spectrum according to the energy threshold, checking whether the energy amplitude value at each time point and frequency point in the first time-frequency spectrum exceeds the energy threshold. If the energy amplitude value at a certain time-frequency point exceeds the energy threshold, the time-frequency point is marked as a valid high-energy point. Then, consecutive valid high-energy points in the time direction are merged to obtain the time interval in which the energy amplitude continuously exceeds the energy threshold, which is used to represent the duration range of transient pulse interference in the first time-frequency spectrum. Next, the first time point in the continuous time interval that exceeds the energy threshold is taken as the start time, and the last time point in the continuous time interval that exceeds the energy threshold is taken as the end time. Then, the time span is calculated based on the time difference between the end time and the start time to generate a pulse width estimate, which is used to represent the duration of transient pulse interference on the time axis and can reflect whether the interference signal is a short-time spike-type interference or a relatively wide pulse-type interference. Then, using the transient change candidate point or continuous time interval as the search range, the time-frequency point with the largest energy amplitude value is found within this range, and this time-frequency point with the largest energy amplitude value is determined as the target time-frequency point. Next, time-frequency coordinate analysis is performed on the target time-frequency point, and the time coordinate and frequency coordinate of the target time-frequency point are read from the first time-frequency spectrum. The time coordinate is determined as the target time coordinate, used to represent the time position of the energy peak of the transient pulse interference, and the frequency coordinate is determined as the target frequency coordinate, used to represent the frequency position where the transient pulse interference energy is most concentrated. Finally, the instantaneous energy change rate of the signal, the pulse width estimate, the target time coordinate, and the target frequency coordinate are integrated in a preset order, and each parameter is normalized to construct the first-scale feature vector. Through the above processing, the energy change intensity, duration, and peak time-frequency position of the transient pulse interference can be accurately extracted from the first time-frequency spectrum, enabling the first-scale feature vector to fully characterize the fine-grained transient interference features and providing reliable fine-scale feature input for subsequent multi-scale fusion identification.

[0035] Multi-scale fusion recognition is performed based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features to obtain a fused feature vector.

[0036] Specifically, after obtaining fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features, these three types of features are encoded to convert them into a unified dimension. Then, based on the importance of different scale features to the current interference identification task, corresponding attention weights are assigned to each type of feature. Scale features that are more obvious and discriminative in the current signal are given higher attention weights; features with weak or redundant interference representations are given lower influence. Subsequently, the weighted fine-grained, mesoscale, and coarse-scale features are concatenated or fused to form a unified fused feature vector. This fused feature vector simultaneously contains transient change information, periodic pulse information, and continuous spectrum distribution information, enabling a more comprehensive description of the interference source characteristics in the electricity meter communication signal. This enhances the complementary expressive ability between different scale features, improving the accuracy and adaptability of subsequent interference source classification and identification.

[0037] In some embodiments, multi-scale fusion recognition is performed based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features to obtain a fused feature vector. The method includes: An adaptive fusion recognition network is constructed, comprising a first feature encoding branch, a second feature encoding branch, and a third feature encoding branch, which operate in parallel. The fine-grained transient feature vector, the mesoscale pulse feature vector, and the coarse-scale continuous spectrum feature vector are synchronized to the first, second, and third feature encoding branches of the adaptive scale fusion recognition network for encoding, generating first, second, and third encoded features. The first, second, and third encoded features are then... The first encoded feature is subjected to scale attention weight calculation to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector. Based on the first attention weight vector, the first encoded feature is weighted and modulated to generate a first weighted modulation encoded feature. Based on the second attention weight vector, the second encoded feature is weighted and modulated to generate a second weighted modulation encoded feature. Based on the third attention weight vector, each channel of the third encoded feature is weighted and modulated to generate a third weighted modulation encoded feature. The first weighted modulation encoded feature, the second weighted modulation encoded feature, and the third weighted modulation encoded feature are concatenated to generate the fused feature vector.

[0038] Specifically, after obtaining the fine-grained transient feature vector, the mesoscale pulse feature vector, and the coarse-scale continuous spectrum feature vector, an adaptive fusion recognition network for multi-scale fusion processing is constructed. This adaptive fusion recognition network includes a first feature encoding branch, a second feature encoding branch, and a third feature encoding branch. The three feature encoding branches are in parallel structure and are used to process interference features from different scales. Specifically, the first feature encoding branch is used to receive and encode the fine-grained transient feature vector, the second feature encoding branch is used to receive and encode the mesoscale pulse feature vector, and the third feature encoding branch is used to receive and encode the coarse-scale continuous spectrum feature vector. Each branch employs an encoding structure consisting of an input layer, a first fully connected layer, a batch normalization layer, a ReLU activation layer, a second fully connected layer, and a feature compression output layer. Let the dimension of the input scale feature vector be d. The input layer receives scale feature vectors of dimension d. The first fully connected layer maps the input feature vector from d dimensions to 128 dimensions, obtaining the initial mapped features. The batch normalization layer normalizes the mean and variance of the initial mapped features. The ReLU activation layer performs a non-linear transformation on the normalized features. The second fully connected layer further maps the 128-dimensional features to 64 dimensions. The feature compression output layer normalizes the 64-dimensional features, converting features of different scales into encoded features of a uniform dimension. The first feature encoding branch uses fine-grained transient feature vectors as training and input objects, mainly learning the mapping relationship between features such as instantaneous energy change rate, pulse width, peak time coordinate, and peak frequency coordinate, to highlight the local time-frequency characteristics of short-term abrupt interference. The second feature encoding branch uses mesoscale pulse feature vectors as training and input objects, mainly learning the combination relationship between features such as pulse repetition period, burst interval, duty cycle, burst count, and burst energy, to highlight the temporal regularity characteristics of periodic or intermittent burst interference. The third feature encoding branch uses coarse-scale continuous spectrum feature vectors as training and input objects, mainly learning the distribution relationship between features such as center frequency, occupied bandwidth, duration, average energy, and spectral stability, to highlight the frequency occupancy characteristics of continuous narrowband interference or persistent spectral interference. Therefore, the first feature encoding branch outputs a 64-dimensional first-dimensional encoding feature, the second feature encoding branch outputs a 64-dimensional second-dimensional encoding feature, and the third feature encoding branch outputs a 64-dimensional third-dimensional encoding feature. Although the three branches have the same output dimension, their parameters are updated according to the corresponding scale feature samples during training. Therefore, they can form a special coding capability for different forms of interference, providing input data with differentiated discriminative information for subsequent scale attention weight calculation and multi-scale feature fusion.

[0039] Subsequently, the fine-grained transient feature vector is simultaneously input into the first feature encoding branch, the mesoscale pulse feature vector is simultaneously input into the second feature encoding branch, and the coarse-scale continuous spectrum feature vector is simultaneously input into the third feature encoding branch, thereby generating the first, second, and third encoded features. Encoding through three parallel branches avoids feature interference caused by early direct mixing of features at different scales, ensuring that each scale feature retains its own discriminative information. Next, the first, second, and third encoded features are input into the scale attention weight calculation module, where global statistical processing is performed on each feature to obtain corresponding descriptors. These descriptors are then subjected to multi-layer shared perception processing to generate the first, second, and third attention weight vectors corresponding to each encoded feature. The first attention weight vector characterizes the importance of the fine-grained transient feature in the current recognition task; the second attention weight vector characterizes the importance of the mesoscale pulse feature; and the third attention weight vector characterizes the importance of the coarse-scale continuous spectrum feature. Next, the first attention weight vector is multiplied element-wise with the feature values ​​of the corresponding dimensions in the first coding feature to enhance the effective feature response related to transient impulse interference and suppress the low-correlation feature response caused by noise or irrelevant abrupt changes, generating the first weighted modulation and coding feature. The second attention weight vector is matched and multiplied with the corresponding feature dimensions in the second coding feature to highlight the discriminative features such as the repetition period, burst interval, and energy distribution of periodic burst interference, generating the second weighted modulation and coding feature. The third attention weight vector is applied to each dimension of the third coding feature to enhance the center frequency, bandwidth occupancy, and spectral stability features corresponding to continuous narrowband interference or persistent spectral interference, generating the third weighted modulation and coding feature. Finally, the first weighted modulation and coding features, the second weighted modulation and coding features, and the third weighted modulation and coding features are concatenated in a preset order to obtain a fused feature vector. This fused feature vector contains not only the coding information of fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features, but also the adaptive contribution relationship of features at different scales to the current interference source identification task. This enables the subsequent classification and identification module to determine the type of interference source based on a more complete and discriminative fused expression, thereby improving the accuracy and robustness of the identification results in complex communication interference environments.

[0040] In some embodiments, scale attention weights are calculated on the first encoded feature, the second encoded feature, and the third encoded feature to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector. The method includes: Calculate the global average value based on the first encoding feature to generate a first channel descriptor; calculate the global average value based on the second encoding feature to generate a second channel descriptor; calculate the global average value based on the third encoding feature to generate a third channel descriptor; perform multi-layer shared perception based on the first channel descriptor, the second channel descriptor, and the third channel descriptor to generate a first-dimensional perception parameter, a second-dimensional perception parameter, and a third-dimensional perception parameter; normalize the first-dimensional perception parameter, the second-dimensional perception parameter, and the third-dimensional perception parameter to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector.

[0041] Specifically, after obtaining the first, second, and third coded features, the three coded features are input into the scale attention weight calculation module to determine the contribution of different scale features to the current communication signal to be identified. First, a global average is calculated based on the first coded feature. If the first coded feature is a one-dimensional feature vector, the average of all feature elements in the feature vector is calculated to obtain the first channel descriptor. If the first coded feature is a feature matrix containing multiple channels, global average pooling is performed on all feature values ​​in each channel to obtain the average response value corresponding to each channel. The average response values ​​are then used to form the first channel descriptor, which characterizes the overall response intensity of fine-grained transient features in each feature channel. Similarly, a global average is calculated based on the second coded feature to generate a second channel descriptor, which characterizes the importance of periodic burst interference features in different channels. A global average is calculated based on the third coded feature to generate a third channel descriptor, which characterizes the overall energy distribution and stable response of continuous narrowband interference or persistent spectral interference in different feature channels. Subsequently, the first, second, and third channel descriptors are input into a multilayer shared perceptron for weight mapping. This multilayer shared perceptron consists of at least two fully connected layers, a nonlinear activation function, and an output mapping layer. The channel descriptors at all three scales share the same perceptron network parameters, enabling the importance of features at different scales to be evaluated under the same weight calculation rules. Specifically, the first channel descriptor is input into the shared perceptron, undergoes dimensionality compression through the first fully connected layer, is then enhanced with a nonlinear activation function to improve its nonlinear expressive power, and is subsequently restored to the dimension corresponding to the first encoded feature through the second fully connected layer, yielding the first-dimensional perceptron parameter. The second channel descriptor is input into the same shared perceptron and undergoes the same mapping process to obtain the second-dimensional perceptron parameter. The third channel descriptor is input into the same shared perceptron to obtain the third-dimensional perceptron parameter. The first, second, and third-dimensional perceptron parameters represent the initial importance scores of fine-scale, mesoscale, and coarse-scale features in each dimension, respectively. Subsequently, to ensure that the importance scores corresponding to the three scales are within a unified comparison range, the first, second, and third dimension perception parameters are input into the Softmax normalization function, so that the sum of the weights of the three scales under the same feature dimension is 1. After normalization, the first, second, and third attention weight vectors are obtained. Through the above processing, the corresponding attention weight vectors can be adaptively generated according to the response strength of different scale interference features in the current signal to be identified. This allows the subsequent fusion process to highlight the scale features with higher contribution and suppress redundant or interference features, thereby improving the discriminative ability of the fused feature vectors and the accuracy of interference source identification.

[0042] Interference sources are classified and identified based on the fused feature vectors to determine the communication interference source identification result of the electricity meter.

[0043] Specifically, after obtaining the fused feature vector, this vector is used as input data to the classification and recognition module for interference source classification and identification. This module categorizes the interference sources based on patterns learned during training and information from the input data, classifying them into types such as radio frequency interference, power line noise, and interference caused by equipment malfunctions. Then, by selecting the category corresponding to the highest probability value, the type of interference source in the electricity meter communication is finally determined. This process effectively identifies various interference sources in the electricity meter communication signal and provides specific interference source identification results, helping to improve the stability of the power monitoring system. It also provides a basis for locating and eliminating interference sources, ensuring the normal operation of the electricity meter and the accurate transmission of data.

[0044] In some embodiments, the method for classifying and identifying interference sources based on the fused feature vector to determine the communication interference source identification result of the electricity meter includes: A classification and recognition module is constructed, comprising N fully connected layers and a classification output layer, where N is an integer greater than or equal to 1. The fused feature vector is synchronized to the N fully connected layers of the classification and recognition module for nonlinear mapping to generate a discriminative feature vector. The discriminative feature vector is synchronized to the classification output layer to calculate the interference source type attribution, obtaining a probability distribution vector. The probability distribution vector is traversed to perform maximum value filtering, generating a maximum attribution probability to determine the interference source type, where the interference source type includes an interference source type identifier. The maximum attribution probability is used as the confidence level of the interference source type identifier for association, thereby determining the communication interference source identification result of the electricity meter.

[0045] Specifically, after obtaining the fused feature vector, a classification and recognition module is constructed to output the category of the interference source. This classification and recognition module can adopt a multilayer perceptron structure, including an input layer, N fully connected layers and a classification output layer, where N is an integer greater than or equal to 1.

[0046] For example, during construction, the input dimension of the fused feature vector is first determined to be D, and the classification output dimension is determined to be K based on the preset number of interference source types. Then, the number of input nodes in the first fully connected layer is set to D, and the number of output nodes is set to 128, used for initial feature mapping of the fused feature vector. The number of input nodes in the second fully connected layer is set to 128, and the number of output nodes is set to 64, used to further compress the feature dimension and enhance class discrimination ability. The number of input nodes in the third fully connected layer is set to 64, and the number of output nodes is set to 32, used to form a more compact discriminative feature representation. A ReLU activation function is set after each fully connected layer to enhance the model's ability to fit complex interference feature boundaries. A batch normalization layer and a Dropout layer are also set after the fully connected layers, with the Dropout ratio set to 0.2 to 0.5 to reduce the risk of overfitting during training. The feature vector output after the last fully connected layer is the discriminative feature vector. The number of input nodes in the classification output layer is the same as the number of output nodes in the last fully connected layer, and the number of output nodes is set to K, with each output node corresponding to an interference source type identifier. The classification output layer performs a linear transformation on the discriminant feature vector to obtain a classification score vector of length K. Then, the classification score vector is converted into a probability distribution vector through the Softmax function, so that the probability values ​​corresponding to each interference source type are between 0 and 1, and the sum of all probability values ​​is 1.

[0047] During the module training phase, communication signal samples from electricity meters labeled with interference source types are collected. Following the aforementioned multi-scale time-frequency transformation, feature extraction, and multi-scale fusion steps, a fused feature vector for training is generated. Each fused feature vector is then paired with its corresponding interference source type label to form a training sample pair. During training, the training samples are input into the classification and recognition module, which outputs a predicted probability distribution vector. This vector is then compared with the actual interference source type labels, and the classification loss is calculated using the cross-entropy loss function. Next, the Adam optimization algorithm is used to update the weight parameters of the fully connected layer and the classification output layer. In this process, the initial learning rate is set to 0.001, the batch size is set to 32 or 64, and the number of training epochs is set to 50 to 200. Training is stopped or the learning rate is reduced when the validation set accuracy no longer improves. After training, the weight parameters of the classification and recognition module are saved, and the trained module is used for actual electricity meter communication interference source identification.

[0048] In the actual recognition phase, the system inputs the fused feature vector into the first fully connected layer of the classification and recognition module. This first fully connected layer performs weighted summation and bias operations on each feature component of the fused feature vector, and then performs a nonlinear transformation using a nonlinear activation function to obtain the first layer of mapped features. When N is greater than 1, the first layer of mapped features continues to be input into subsequent fully connected layers, undergoing weighted mapping, nonlinear activation, normalization, and other processing sequentially, until the mapping calculation for the Nth fully connected layer is completed. Through layer-by-layer processing across N fully connected layers, the fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features contained in the fused feature vector can be further transformed into deep feature expressions with class discrimination capabilities, generating a discriminative feature vector to characterize the degree of matching between the current communication signal to be identified and various interference sources. Next, the discriminant feature vector is input into the classification output layer for interference source type attribution calculation. The classification output layer calculates the classification score of each output node based on the discriminant feature vector and converts multiple classification scores into a probability distribution vector using the Softmax function. Each element in this probability distribution vector represents the probability value of the communication signal to be identified belonging to the corresponding interference source type, and each probability value reflects the likelihood of attribution to different interference source types. Then, the probability distribution vector is traversed, and each probability value is compared. The probability value with the largest value is selected as the maximum attribution probability, and the output node number corresponding to the maximum attribution probability is obtained. Based on this output node number, a preset interference source type mapping relationship is queried to determine the interference source type identifier corresponding to the maximum attribution probability. For example, when the probability value corresponding to continuous narrowband interference in the probability distribution vector is the largest, the type identifier corresponding to continuous narrowband interference is determined as the interference source type identifier of the current energy meter communication signal. Finally, the maximum value of the attribution probability is used as the confidence level for the interference source type identifier, and the interference source type identifier is correlated with the confidence level to form the communication interference source identification result of the electricity meter. This communication interference source identification result includes the interference source type identifier, the interference source type name, the confidence level value, and the corresponding identification time. Through the above process, the fused feature vector can be transformed into a clear interference source category result, while simultaneously providing the identification confidence level, thereby improving the interpretability, reliability, and engineering application value of the communication interference source identification result of the electricity meter.

[0049] In summary, the method for identifying communication interference sources in energy meters under multi-scale features provided by this invention has the following technical effects: The communication signals from the electricity meter's communication module are collected and extracted to obtain the signal data to be identified. Multi-scale time-frequency transformation is then performed to construct a multi-scale time-frequency spectrum matrix. Based on this matrix, time-frequency domain feature analysis is conducted to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. Multi-scale fusion identification is then performed based on these features to obtain a fused feature vector. Interference source classification and identification are then performed based on the fused feature vector to determine the communication interference source identification result of the electricity meter. This achieves the technical effect of using complementary fusion of multi-scale time-frequency features to accurately classify and identify multiple types of communication interference sources, thereby improving the anti-interference capability and identification reliability of electricity meter communication in complex electromagnetic environments.

[0050] Example 2, as Figure 2 This is a schematic diagram of the structure of the electricity meter communication interference source identification system under multi-scale features according to the present invention. For example, Figure 1 The flowchart of the method for identifying communication interference sources of electricity meters under multi-scale features of the present invention can be illustrated as follows: Figure 2 The structure shown is implemented.

[0051] Based on the same concept as the method for identifying electricity meter communication interference sources under multi-scale features in the above embodiments, the present invention also provides an electricity meter communication interference source identification system under multi-scale features, comprising: Signal processing unit 11: Collects communication signals from the electricity meter's communication module, extracts the signals, obtains the signal data to be identified, performs multi-scale time-frequency transformation, and constructs a multi-scale time-frequency spectrum matrix; Feature analysis unit 12: Performs time-frequency domain feature analysis based on the multi-scale time-frequency spectrum matrix to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features; Multi-scale fusion unit 13: Performs multi-scale fusion identification based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features to obtain a fused feature vector; Interference source identification unit 14: Classifies and identifies interference sources based on the fused feature vector to determine the communication interference source identification result of the electricity meter.

[0052] In some embodiments, the signal processing unit 11 includes: The front-end receiving link of the positioning energy meter communication module is equipped with multiple signal acquisition nodes; multiple communication signals are acquired through the multiple signal acquisition nodes, and the multiple communication signals are converted from analog to digital according to a preset sampling rate to generate an original digital signal sequence; the original digital signal sequence is divided into equal length segments according to a sliding time window to generate multiple signal frames; the multiple signal frames are subjected to amplitude normalization processing to generate a signal frame combination, and the signal frame combination is used as the signal data to be identified.

[0053] In some embodiments, the signal processing unit 11 includes: A first transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the first transform scale parameter to generate a first time-frequency spectrum. Transient pulse interference signals are captured based on the first time-frequency spectrum to determine a first time-frequency distribution feature. A second transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the second transform scale parameter to generate a second time-frequency spectrum. Periodic burst interference signals are captured based on the second time-frequency spectrum to determine a second time-frequency distribution feature. A third transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the third transform scale parameter to generate a third time-frequency spectrum. Continuous narrowband interference signals are captured based on the third time-frequency spectrum to determine a third time-frequency distribution feature. The first time-frequency spectrum, the second time-frequency spectrum, and the third time-frequency spectrum are stacked according to the first time-frequency distribution feature, the second time-frequency distribution feature, and the third time-frequency distribution feature to construct the multi-scale time-frequency matrix.

[0054] In some embodiments, the feature analysis unit 12 includes: Fine-grained transient feature analysis is performed on the first time-spectrum to generate transient change feature parameters. Based on these transient change feature parameters, a first-scale feature vector is constructed. Mesoscale pulse feature analysis is performed on the second time-spectrum to generate periodic burst feature parameters. Based on these periodic burst feature parameters, a second-scale feature vector is constructed. Coarse-scale continuous spectrum feature analysis is performed on the third time-spectrum to generate continuous spectrum distribution feature parameters. Based on these continuous spectrum distribution feature parameters, a third-scale feature vector is constructed. The first-scale feature vector, the second-scale feature vector, and the third-scale feature vector are combined to generate a multi-scale feature set, which simultaneously includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features.

[0055] In some embodiments, the feature analysis unit 12 includes: The time dimension of the first time-spectrum is extracted, and the first time-spectrum is scanned frame by frame according to the time dimension to obtain multiple time-spectrum energy data. Based on the multiple time-spectrum energy data, adjacent time points are analyzed to generate multiple transition amplitude parameters. A preset energy change threshold is set, and the multiple transition amplitude parameters are compared with the energy change threshold. Transition amplitude parameters exceeding the energy change threshold are extracted, and the instantaneous energy change rate of the signal is calculated, where the instantaneous energy change rate includes instantaneous abrupt change characteristics. An energy threshold is set based on the first time-spectrum, and the first time-spectrum is traversed according to the energy threshold for searching. Extract the continuous time intervals where the energy amplitude of the first time-spectrum exceeds the energy threshold; locate the start and end times based on the continuous time intervals, calculate the time span based on the start and end times, and generate a pulse width estimate; locate the maximum energy amplitude of the first time-spectrum based on the instantaneous change characteristics, determine the target time-frequency point, perform time-frequency coordinate analysis according to the target time-frequency point, and determine the target time coordinate and target frequency coordinate; integrate the instantaneous energy change rate of the signal, the pulse width estimate, the target time coordinate, and the target frequency coordinate to construct the first scale feature vector.

[0056] In some embodiments, the multi-scale fusion unit 13 includes: An adaptive fusion recognition network is constructed, comprising a first feature encoding branch, a second feature encoding branch, and a third feature encoding branch, which operate in parallel. The fine-grained transient feature vector, the mesoscale pulse feature vector, and the coarse-scale continuous spectrum feature vector are synchronized to the first, second, and third feature encoding branches of the adaptive scale fusion recognition network for encoding, generating first, second, and third encoded features. The first, second, and third encoded features are then... The first encoded feature is subjected to scale attention weight calculation to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector. Based on the first attention weight vector, the first encoded feature is weighted and modulated to generate a first weighted modulation encoded feature. Based on the second attention weight vector, the second encoded feature is weighted and modulated to generate a second weighted modulation encoded feature. Based on the third attention weight vector, each channel of the third encoded feature is weighted and modulated to generate a third weighted modulation encoded feature. The first weighted modulation encoded feature, the second weighted modulation encoded feature, and the third weighted modulation encoded feature are concatenated to generate the fused feature vector.

[0057] In some embodiments, the multi-scale fusion unit 13 includes: Calculate the global average value based on the first encoding feature to generate a first channel descriptor; calculate the global average value based on the second encoding feature to generate a second channel descriptor; calculate the global average value based on the third encoding feature to generate a third channel descriptor; perform multi-layer shared perception based on the first channel descriptor, the second channel descriptor, and the third channel descriptor to generate a first-dimensional perception parameter, a second-dimensional perception parameter, and a third-dimensional perception parameter; normalize the first-dimensional perception parameter, the second-dimensional perception parameter, and the third-dimensional perception parameter to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector.

[0058] In some embodiments, the interference source identification unit 14 includes: A classification and recognition module is constructed, comprising N fully connected layers and a classification output layer, where N is an integer greater than or equal to 1. The fused feature vector is synchronized to the N fully connected layers of the classification and recognition module for nonlinear mapping to generate a discriminative feature vector. The discriminative feature vector is synchronized to the classification output layer to calculate the interference source type attribution, obtaining a probability distribution vector. The probability distribution vector is traversed to perform maximum value filtering, generating a maximum attribution probability to determine the interference source type, where the interference source type includes an interference source type identifier. The maximum attribution probability is used as the confidence level of the interference source type identifier for association, thereby determining the communication interference source identification result of the electricity meter.

[0059] In embodiment three, the present invention also provides a computer-readable storage medium that can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the multi-scale feature-based electricity meter communication interference source identification method in the embodiments of the present invention, thereby realizing the above-mentioned multi-scale feature-based electricity meter communication interference source identification method.

[0060] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A method for identifying communication interference sources in electricity meters under multi-scale characteristics, characterized in that, The method includes: The communication signals from the communication module of the electricity meter are collected and extracted to obtain the signal data to be identified. Multi-scale time-frequency transformation is then performed to construct a multi-scale time-frequency matrix. Based on the multi-scale time-frequency spectrum matrix, time-frequency domain feature analysis is performed to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. Multi-scale fusion recognition is performed based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features to obtain a fused feature vector; The interference source is classified and identified based on the fused feature vector to determine the communication interference source identification result of the electricity meter; Multi-scale fusion recognition is performed based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features to obtain a fused feature vector. The method includes: An adaptive fusion recognition network is constructed, which includes a first feature coding branch, a second feature coding branch, and a third feature coding branch, wherein the first feature coding branch, the second feature coding branch, and the third feature coding branch are in parallel. The fine-grained transient feature vector, the mesoscale pulse feature vector, and the coarse-scale continuous spectrum feature vector are respectively synchronized to the first feature coding branch, the second feature coding branch, and the third feature coding branch of the adaptive scale fusion recognition network for encoding to generate the first coding feature, the second coding feature, and the third coding feature; Scale attention weights are calculated on the first encoded feature, the second encoded feature, and the third encoded feature to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector. The first encoded feature is weighted and modulated based on the first attention weight vector to generate the first weighted modulation encoded feature; The second encoded feature is weighted and modulated based on the second attention weight vector to generate the second weighted modulation encoded feature; Based on the third attention weight vector, each channel of the third coding feature is weighted and modulated to generate the third weighted modulation coding feature; The first weighted modulation and coding feature, the second weighted modulation and coding feature, and the third weighted modulation and coding feature are concatenated to generate the fused feature vector.

2. The method for identifying communication interference sources of electricity meters under multi-scale features as described in claim 1, characterized in that, The method involves collecting communication signals from the communication module of the electricity meter, extracting the signals, and obtaining the signal data to be identified. The front-end receiving link of the positioning energy meter communication module is equipped with multiple signal acquisition nodes; Multiple communication signals are acquired through the multiple signal acquisition nodes, and analog-to-digital conversion is performed on the multiple communication signals according to a preset sampling rate to generate an original digital signal sequence; The original digital signal sequence is divided into equal-length segments according to a sliding time window to generate multiple signal frames; The multiple signal frames are subjected to amplitude normalization processing to generate a signal frame combination, which is then used as the signal data to be identified.

3. The method for identifying communication interference sources of electricity meters under multi-scale features as described in claim 1, characterized in that, The method for obtaining the signal data to be identified, performing multi-scale time-frequency transformation, and constructing a multi-scale time-spectrum matrix includes: A first transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the first transform scale parameter to generate a first time spectrum. Based on the first time-frequency spectrum, transient pulse interference signals are captured, and the first time-frequency distribution characteristics are determined. A second transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the second transform scale parameter to generate a second time-spectrum diagram. Based on the second time-frequency spectrum, periodic burst interference signals are captured, and the second time-frequency distribution characteristics are determined. A third transform scale parameter is set, and a continuous wavelet transform is performed on the signal data to be identified according to the third transform scale parameter to generate a third time-spectrum diagram. Based on the third time-frequency spectrum, continuous narrowband interference signals are captured, and the third time-frequency distribution characteristics are determined. The first time-frequency spectrum, the second time-frequency spectrum, and the third time-frequency spectrum are stacked according to the first time-frequency distribution characteristics, the second time-frequency distribution characteristics, and the third time-frequency distribution characteristics to construct the multi-scale time-frequency matrix.

4. The method for identifying communication interference sources of energy meters under multi-scale features as described in claim 3, characterized in that, Based on the multi-scale time-frequency spectrum matrix, time-frequency domain feature analysis is performed to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features. The method includes: Fine-grained transient feature analysis is performed on the first time-frequency spectrum to generate transient mutation feature parameters. Based on the transient mutation feature parameters, a first-scale feature vector is constructed. Mesoscale pulse feature analysis is performed on the second time-spectrum to generate periodic burst feature parameters. Based on the periodic burst feature parameters, a second-scale feature vector is constructed. A coarse-scale continuous spectrum feature analysis is performed on the third time-frequency spectrum to generate continuous spectrum distribution feature parameters. Based on the continuous spectrum distribution feature parameters, a third-scale feature vector is constructed. The first-scale feature vector, the second-scale feature vector, and the third-scale feature vector are combined to generate a multi-scale feature set, which simultaneously includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features.

5. The method for identifying communication interference sources of electricity meters under multi-scale features as described in claim 4, characterized in that, Fine-grained transient feature analysis is performed on the first time-frequency spectrum to generate transient change feature parameters. Based on the transient change feature parameters, a first-scale feature vector is constructed. The method includes: Extract the time dimension of the first time spectrum, and scan the first time spectrum frame by frame according to the time dimension to obtain multiple time spectrum energy data; Based on the multiple time-spectrum energy data, adjacent time points are analyzed to generate multiple jump amplitude parameters; A preset energy change threshold is set, the multiple jump amplitude parameters are compared with the energy change threshold, the jump amplitude parameters that exceed the energy change threshold are extracted, and the instantaneous energy change rate of the signal is calculated. The instantaneous energy change rate of the signal includes instantaneous change characteristics. Based on the first time spectrum, an energy threshold is set, and the first time spectrum is searched traversed according to the energy threshold to extract the continuous time intervals in which the energy amplitude of the first time spectrum exceeds the energy threshold. Based on the continuous time interval, the start and end times are located, and the time span is calculated according to the start and end times to generate a pulse width estimate. Based on the instantaneous mutation characteristics, the maximum energy amplitude of the first time-frequency graph is located, the target time-frequency point is determined, and time-frequency coordinate analysis is performed according to the target time-frequency point to determine the target time coordinate and target frequency coordinate. The instantaneous energy change rate of the signal, the estimated pulse width, the target time coordinate, and the target frequency coordinate are integrated to construct the first scale feature vector.

6. The method for identifying communication interference sources of energy meters under multi-scale features as described in claim 1, characterized in that, The method involves calculating scale attention weights on the first encoded feature, the second encoded feature, and the third encoded feature to generate a first attention weight vector, a second attention weight vector, and a third attention weight vector. Calculate the global average value based on the first encoded feature to generate the first channel descriptor; Calculate the global average value based on the second encoded feature to generate the second channel descriptor; Calculate the global average value based on the third coding feature to generate the third channel descriptor; Multi-layer shared perception is performed based on the first channel descriptor, the second channel descriptor, and the third channel descriptor to generate first-dimensional perception parameters, second-dimensional perception parameters, and third-dimensional perception parameters; The first dimension perception parameter, the second dimension perception parameter, and the third dimension perception parameter are normalized to generate the first attention weight vector, the second attention weight vector, and the third attention weight vector.

7. The method for identifying communication interference sources of electricity meters under multi-scale features as described in claim 1, characterized in that, The method for classifying and identifying interference sources based on the fused feature vector to determine the communication interference source identification result of the electricity meter includes: A classification and recognition module is constructed, which includes N fully connected layers and a classification output layer, where N is an integer greater than or equal to 1; The fused feature vector is synchronized to the N fully connected layers of the classification and recognition module for nonlinear mapping to generate a discriminative feature vector; The discriminative feature vector is synchronized to the classification output layer to calculate the interference source type attribution and obtain the probability distribution vector. The probability distribution vector is traversed to filter for maximum values, and the maximum value of the belonging probability is generated to determine the type of interference source. The type of interference source includes an interference source type identifier. The maximum value of the attribution probability is used as the confidence level of the interference source type identifier to determine the communication interference source identification result of the electricity meter.

8. A multi-scale characteristic-based system for identifying communication interference sources in electricity meters, characterized in that, The system is used to implement the method for identifying communication interference sources of energy meters under multi-scale features as described in any one of claims 1-7, the system comprising: Signal processing unit: Collects communication signals from the electricity meter's communication module, extracts the signals, obtains the signal data to be identified, performs multi-scale time-frequency transformation, and constructs a multi-scale time-frequency matrix; Feature analysis unit: Performs time-frequency domain feature analysis based on the multi-scale time-frequency spectrum matrix to generate a multi-scale feature set, which includes fine-grained transient features, mesoscale pulse features, and coarse-scale continuous spectrum features; Multi-scale fusion unit: Based on the fine-grained transient features, the mesoscale pulse features, and the coarse-scale continuous spectrum features, multi-scale fusion recognition is performed to obtain a fused feature vector; Interference source identification unit: Classifies and identifies interference sources based on the fused feature vector to determine the communication interference source identification result of the electricity meter.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the method for identifying communication interference sources of electricity meters under multi-scale features as described in any one of claims 1 to 7.

Images