Method and system for monitoring state of electric energy metering box based on multi-sensor fusion

Abstract

The application relates to the technical field of electric power metering and intelligent monitoring, and discloses a power metering box state monitoring method and system based on multi-sensor fusion. The method comprises the following steps: firstly, continuous time sequence data of a metering box hardware damage moment is acquired, a box body feature sequence is obtained through preprocessing, and a correlation mapping matrix representing the correlation between physical quantities and electrical quantities is constructed; secondly, historical electrical data before a data fault is extracted, and the metering box electrical features are inversely deduced and reconstructed in combination with the correlation mapping matrix; finally, the box body feature sequence and the reconstructed electrical features are fused in multiple dimensions to obtain a fault panoramic feature, according to which the fault type is identified and the fault occurrence moment is located. The application can restore the complete fault features of the hardware damage moment, realize accurate fault identification and moment positioning, and take into account the real-time performance, accuracy and engineering practicability of the metering box state monitoring.

Description

Technical Field

[0001] This invention relates to the field of power metering and intelligent monitoring technology, and in particular to a method and system for monitoring the status of power metering boxes based on multi-sensor fusion. Background Technology

[0002] With the continuous expansion of the power grid and the increasing demands for electricity metering, smart sensors, as the core carrier of condition monitoring, directly affect the reliability of metering box fault diagnosis due to the integrity of their data.

[0003] In existing technologies, most methods for monitoring the condition of metering boxes rely on continuous and complete electrical data streams. However, when the hardware of an electricity metering box fails, the voltage and current acquisition devices often experience sampling interruptions and data gaps due to the impact, resulting in the loss of key electrical characteristics. The remaining physical characteristics, such as temperature and vibration, are often mixed with a large amount of noise and outliers. This data gap problem that occurs instantly when the hardware fails, in existing technical solutions, because they rely on continuous and complete electrical data streams, leads to a disconnect between physical and electrical quantities, making it prone to misdiagnosis and missed fault detection.

[0004] In summary, existing technologies cannot identify fault types and pinpoint the time of occurrence under multi-dimensional data fusion. Summary of the Invention

[0005] This invention provides a method and system for monitoring the status of power metering boxes based on multi-sensor fusion, in order to solve the problems that existing technologies cannot effectively restore the complete fault characteristics at the moment of hardware failure, nor can they realize the identification of fault types and the location of the time of occurrence under multi-dimensional data fusion, and it is difficult to balance the real-time performance and engineering practicality of metering box status monitoring.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for monitoring the status of an energy metering box based on multi-sensor fusion, comprising:

[0007] Continuous time-series data of the meter box at the moment of hardware failure is acquired and preprocessed to obtain the box feature sequence;

[0008] Based on the aforementioned box feature sequence, an association mapping matrix characterizing the relationship between physical quantities and electrical quantities is constructed;

[0009] Historical electrical data of the metering box before the data gap is obtained, and the electrical characteristics of the metering box are reconstructed by reverse deduction in combination with the aforementioned correlation mapping matrix.

[0010] The box feature sequence is fused with the reconstructed electrical features of the metering box in multiple dimensions to obtain a fault panorama feature, and the fault type and the time of fault occurrence are identified based on the fault panorama feature.

[0011] Secondly, the present invention provides a power metering box status monitoring system based on multi-sensor fusion, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the power metering box status monitoring method based on multi-sensor fusion as described above.

[0012] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for monitoring the status of an energy metering box based on multi-sensor fusion.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] (1) This invention optimizes abnormal electrical features by multi-dimensional spatial mapping and temporal feature reconstruction, and constructs a fault panorama feature that includes physical and electrical dimensions by combining multi-source data fusion, thereby realizing the identification of short-circuit fault types and the location of occurrence time inside the metering box.

[0015] (2) This invention achieves the extraction of effective data and suppression of noise interference under extreme working conditions by performing outlier removal and smoothing denoising on the residual physical features of the meter box at the moment of hardware damage, combined with feature dimensionality reduction reconstruction to generate a high signal-to-noise ratio box feature sequence.

[0016] (3) This invention constructs a correlation mapping matrix between physical features and electrical features by extracting feature change gradients, and completes the reverse deduction of electrical quantities at the time of data discontinuity by combining time-series dependency analysis, thus solving the technical problem that fault features cannot be restored when electrical data is missing. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method for monitoring the status of an energy metering box based on multi-sensor fusion, provided in the first embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 The first embodiment of the present invention provides a method for monitoring the status of an energy metering box based on multi-sensor fusion, comprising the following steps:

[0020] S101, acquire continuous time-series data at the moment of hardware failure of the metering box and perform preprocessing to obtain the box feature sequence;

[0021] S102, Based on the box feature sequence, construct an association mapping matrix representing the relationship between physical quantities and electrical quantities;

[0022] S103, Obtain historical electrical data before the data gap in the metering box, and perform reverse deduction by combining the correlation mapping matrix to reconstruct the electrical characteristics of the metering box;

[0023] S104, the box feature sequence is fused with the reconstructed metering box electrical features in multiple dimensions to obtain a fault panorama feature, and the fault type of the metering box and the time of fault occurrence are identified based on the fault panorama feature.

[0024] In step S101, continuous time-series data of the metering box at the moment of hardware failure is acquired and preprocessed to obtain the box feature sequence, including:

[0025] Continuous time-series data before and after the meter box hardware failure is obtained, and the data is truncated based on a preset voltage trigger threshold to obtain the target time period sequence;

[0026] The target time series is subjected to outlier removal and interpolation to obtain a clean time series;

[0027] Perform Kalman filtering on the pure time series to obtain a smooth time series;

[0028] Principal component analysis was performed on the smoothed time series to reduce its dimensionality, resulting in a box-shaped feature sequence.

[0029] It should be noted that the sampling frequency of continuous time-series data is set to 10,000 times per second. The sampled data includes values ​​for four dimensions: three-phase voltage, three-phase current, enclosure surface temperature, and enclosure vibration acceleration. The sampling frequency of 10,000 times per second was determined by statistically analyzing the temporal characteristics of 100 lightning strike-induced hardware damage events to the metering box. The duration of the damage process ranges from 10ms to 20ms, and at a sampling frequency of 10,000 times per second, at least 100 valid sampling points can be collected for each damage process. The preset voltage trigger threshold is 250V. The 250V preset voltage trigger threshold was determined by statistically analyzing the maximum voltage distribution of 1000 different models of metering boxes under normal operating conditions, covering an ambient temperature range of 10℃ to 40℃ and a load range of 0 to rated load, using the 99th percentile of the maximum voltage distribution. The data truncation process involves truncating 50 sampling points forward and 150 sampling points backward, outputting a target time period sequence with a length of 200 sampling points. The truncation window length was determined through statistical experiments on the distribution of hardware failure characteristics. The experiments showed that more than 95% of the failure characteristics were concentrated in the time range of 5ms before failure to 15ms after failure, corresponding to 50 and 150 sampling points.

[0030] It should be noted that the anomaly score calculation is performed independently for each sampling point. The calculation process first takes the moving average of the 20 valid samples before the current sampling point as the baseline value for each dimension. Then, it calculates the relative deviation of each dimension value of the current sampling point from the corresponding baseline value, and takes the maximum value of the relative deviations of all dimensions as the anomaly score for that sampling point. The preset anomaly score threshold is 0.7. The preset anomaly score threshold of 0.7 is determined by statistically analyzing the anomaly score distribution of 500 sets of normal operation data and 500 sets of data containing electromagnetic interference, taking the median value between the 99th percentile of the anomaly score distribution of normal data and the 1st percentile of the anomaly score distribution of interference data. The linear interpolation completion process uses the values ​​of the two valid sampling points before and after the removal position as endpoints, and calculates the completed values ​​according to the time ratio to maintain the continuity of the time series.

[0031] It should be noted that the preset frequency band is 1000Hz to 5000Hz, and the preset proportion threshold is 5%. The 5% preset proportion threshold was determined by statistically analyzing the high-frequency noise spectrum distribution generated by arc discharge. Experiments show that the noise generated by arc discharge is mainly concentrated in the 1000Hz to 5000Hz frequency band, and its energy proportion exceeds 5% in all cases. The state transition matrix and observation matrix for the state prediction and observation update iteration calculations are set to identity matrices, the process noise covariance is set to 0.01, and the observation noise covariance is set to 0.1. The calculation parameters were determined using 500 sets of metering box operating data containing arc noise. Under these parameter settings, the filtered time series fluctuation amplitude is less than 15% of the original noise amplitude.

[0032] It should be noted that the frequency band energy proportion is calculated using Fast Fourier Transform (FFT). FFT is performed on the four dimensions of the clean time series: three-phase voltage, three-phase current, enclosure surface temperature, and shell vibration acceleration, respectively, to obtain the amplitude spectrum for each dimension. The preset frequency band energy is the sum of squares of the amplitudes at all frequency points within the range of 1000Hz to 5000Hz, and the total energy is the sum of squares of the amplitudes at all frequency points within the range of 0Hz to 5000Hz. The frequency band energy proportion is the ratio of the preset frequency band energy to the total energy. After obtaining the frequency band energy proportions for each of the four dimensions, the maximum value is used as the criterion for whether to trigger Kalman filtering.

[0033]

[0034] in, The energy percentage of the j-th data dimension within the preset frequency band, expressed as % This represents the sum of squares of the amplitude of the j-th data dimension within the frequency band from 1000Hz to 5000Hz. This represents the sum of squared amplitudes of the j-th data dimension within the frequency band from 0Hz to 5000Hz. For the j-th data dimension in frequency Fourier amplitude at the location; is a discrete frequency point; j represents one of the data dimensions: three-phase voltage, three-phase current, enclosure surface temperature, or enclosure vibration acceleration. and Using the same amplitude squared unit, the division of the two yields a dimensionless ratio, which, when multiplied by 100%, gives a percentage; because... for Sub-band energy, The value range is from 0 to 100%. In the example, the energy percentage is 12%, which is greater than the preset percentage threshold of 5%, triggering the Kalman filter calculation process and judgment result.

[0035] It should be noted that the orthogonal transformation processes the four dimensions of the smoothed time series data, generating mutually independent principal components. The preset contribution rate threshold is 90%. This 90% threshold was determined through statistical experiments on information retention rates in principal component analysis. When the cumulative contribution rate reaches 90%, more than 90% of the effective information in the original data is retained, while the data dimensionality is reduced to 2 to 3 dimensions, decreasing computational complexity. The top N principal components with cumulative contribution rates exceeding the preset threshold are extracted and arranged in descending order of contribution rate to form a box-shaped feature sequence.

[0036] It is worth noting that the metering box's characteristic sequence is simultaneously stored on the metering box's local non-volatile memory and the cloud data platform in binary format for a period of 3 years. The stored content includes the metering box's characteristic sequence vector, corresponding timestamp, sampling frequency, and preprocessing parameters. The storage duration is determined by the power industry's regulations on metering box fault tracing cycles, meeting the fault tracing requirements of power grid operation and maintenance.

[0037] For example, in a scenario where a lightning strike damages the metering box hardware, the metering box samples 10,000 times per second. The continuously acquired voltage sequence first exceeds the preset 250V voltage trigger threshold at the 1250th sampling point. Using the time point corresponding to the 1250th sampling point as the center, 50 sampling points are truncated forward and 150 sampling points are truncated backward, resulting in a target time period sequence of 200 sampling points. Anomaly scores are calculated for each sampling point in the target time period sequence. The calculation results show that the anomaly score of the 1272nd sampling point is 0.82, which is higher than the preset anomaly score threshold of 0.7. Therefore, the 1272nd sampling point is identified as interference and removed. Linear interpolation is performed at the removed position to complete the sequence, resulting in a clean time series. Frequency band energy proportion calculation is performed on the clean time series. The energy proportion of the 1000Hz to 5000Hz frequency band is 12%, exceeding the preset proportion threshold of 5%. State prediction and observation update iteration operations are then performed on the clean time series to obtain a smooth time series. An orthogonal transformation is performed on the smooth time series to extract principal components with a cumulative contribution rate exceeding 90%, yielding the bin feature sequence. Preprocessing markers are simultaneously added to the bin feature sequence, and it is stored in both the local non-volatile memory of the metering bin and the cloud data platform.

[0038] In step S102, based on the enclosure feature sequence, an association mapping matrix characterizing the relationship between physical quantities and electrical quantities is constructed, including:

[0039] The box feature sequence is segmented to obtain multiple time slice features. The difference in node split gain between adjacent time slice features is calculated. The difference is divided by the time interval step size to obtain the feature change gradient.

[0040] If the gradient of the feature change exceeds a preset gradient threshold, the correlation coefficient between the corresponding physical feature dimension and the electrical load dimension is extracted, and the core feature pairs whose absolute values ​​of the correlation coefficients exceed the preset correlation coefficient threshold are selected.

[0041] An initial context matrix is ​​constructed with physical feature dimension as rows and electrical load dimension as columns. The matrix elements are the correlation coefficient values ​​of the corresponding core feature pairs, and the elements at the non-core feature pairs are set to zero. Singular value decomposition is performed on the initial context matrix to extract the left singular vectors corresponding to the main singular values, thus obtaining the correlation mapping matrix.

[0042] It should be noted that the fixed window length is set to 20 sampling units, and the fixed sliding step size is set to 5 sampling units. The combination of these parameters—window length and sliding step size—was determined through statistical experiments on the feature distribution of 1000 sets of metering box hardware failure processes. This ensures complete coverage of the time range of physical feature mutations while maintaining temporal resolution suitable for fault location requirements. The number of decision trees is set to 100, forming a random forest model. The random forest model is trained offline using a pre-training method. Training samples are from labeled datasets collected during the metering box's historical operation. The dataset includes physical features and corresponding electrical data under both normal operation and hardware failure conditions, with sample labels corresponding to the operating condition category. The trained model is deployed on the metering box's local processing unit, and the model output node split gain values ​​are directly called during online computation. The number of 100 decision trees was determined through stability experiments on feature importance evaluation. When the number of decision trees reaches 100, the fluctuation range of the node split gain calculation is less than 5%, meeting the numerical stability requirements of subsequent gradient calculations.

[0043] It should be noted that the labeled dataset contains no fewer than 500 normal operation samples and no fewer than 300 hardware failure samples, covering various fault types and severity levels, including internal short circuits, insulation aging, and mechanical damage. Labeling is based on the fault type and occurrence time recorded in the metering box's on-site fault work order as the gold standard, combined with electrical characteristic analysis and fault physical judgment rules, and is completed after independent verification and confirmation by at least two professional maintenance personnel. Samples with inconsistent labels are reviewed and confirmed based on the original fault work order record. Normal operation samples are collected from stable operating periods under different load levels and ambient temperature conditions to ensure that the training data covers the typical operating conditions of the metering box.

[0044] It should be noted that both the training input and online inference input of the random forest model use the box feature sequence output in step S101. The physical features in the training samples are first processed according to the same outlier removal, interpolation completion, Kalman filtering, and principal component analysis process as in step S101, resulting in the same N-dimensional feature space as the online box feature sequence, where N is 2 or 3. During online computation, the classification label output of the random forest is not called. Instead, the impurity reduction of each input feature in each decision tree during node splitting is extracted and averaged across 100 decision trees according to the feature dimension to obtain the node splitting gain for the corresponding time slice feature.

[0045] It should be noted that the node split gain is used to characterize the contribution of the corresponding physical feature dimension to distinguish between normal operation and hardware failure. The correspondence between this contribution and the intensity of physical feature abrupt changes was verified using 500 sets of data with synchronous electrical load abrupt change markers. When the time slice where the peak of the node split gain is located deviates from the time slice of the electrical load abrupt change by no more than one sliding step, it is considered consistent. The consistency rate in the verification data was 92%. Therefore, the difference in node split gain between adjacent time slices is used as the basis for calculating the gradient of physical feature changes.

[0046] It should be noted that the preset gradient threshold is 25. This preset gradient threshold of 25 is determined by statistically analyzing the feature change gradient distributions of 500 sets of normal operating conditions and 500 sets of hardware failure conditions. The midpoint between the 99th percentile of the gradient distribution in the normal state and the 1st percentile in the gradient distribution in the failure state is taken, which can distinguish between fluctuations in normal conditions and feature mutations caused by hardware failure. Synchronous electrical load data includes sampled values ​​of instantaneous active power and reactive power. The data comes from the power metering module of the metering box itself and is synchronously acquired using the same clock source as the box's feature sequence, with the sampling frequency consistent with the sampling frequency of the box's feature sequence. The initial context matrix has the same number of rows as the number of dimensions of the box's feature sequence, and the number of columns is fixed at 2, corresponding to the two electrical load dimensions of active power and reactive power. The core feature pairs in the matrix are filled with calculated correlation coefficient values ​​at their corresponding positions, while the remaining elements are set to zero. A zero value indicates that there is no strong linear correlation between the corresponding physical feature and the electrical load dimension.

[0047] It should be noted that the preset correlation coefficient threshold is 0.8. This 0.8 threshold was determined through 500 sets of physical and electrical correlation experiments under different operating conditions. When the absolute value of the correlation coefficient is greater than 0.8, a strong linear correlation exists between the physical characteristics and the electrical load characteristics, reflecting their intrinsic relationship. Singular value decomposition is performed on the initial context matrix, extracting the left singular vectors corresponding to the first two principal singular values ​​to form the correlation mapping matrix. The number of rows in the correlation mapping matrix matches the dimension of the box feature sequence, the number of columns equals the number of retained singular values, and the dimension values ​​perfectly match the dimension of the feature components in subsequent steps.

[0048] It is worth noting that the association mapping matrix is ​​synchronously stored on the metering box's local non-volatile memory and the cloud data platform in binary format for a period of 3 years. The stored content includes the association mapping matrix vector, corresponding timestamps, node split gain data, and core feature pair information. The storage duration is determined by the power industry's regulations on metering box fault tracing cycles, adapting to the fault tracing needs of power grid operation and maintenance.

[0049] For example, in a scenario where a lightning strike damages the metering box hardware, a 200-character feature sequence is input. A sliding window with a length of 20 sampling units and a step size of 5 sampling units is used to perform segmented extraction, resulting in 37 independent time-slice features. For each time-slice feature, the reduction in impurity at node splits is calculated using 100 decision trees. The average reduction in impurity for the corresponding dimension across all decision trees is taken as the node split gain for that dimension. The difference in node split gains between adjacent time-slice features is calculated, and the difference is divided by the time interval step size to obtain the feature change gradient sequence. The feature change gradient corresponding to the 12th time-slice is 32, exceeding the preset gradient threshold of 25. The instantaneous active and reactive power samples at the corresponding timestamps are extracted and concatenated with the time-slice feature vectors to obtain a joint feature set. The correlation coefficients between each physical feature dimension and each electrical load dimension in the joint feature set were calculated. The correlation coefficient between temperature mutation rate and active power was 0.86, and the correlation coefficient between vibration acceleration amplitude and reactive power was 0.82, both exceeding the preset correlation coefficient threshold of 0.8. Two core feature pairs were thus selected. A two-dimensional initial context matrix was constructed using the two physical feature dimensions as rows and the two electrical load dimensions as columns. The values ​​0.86 and 0.82 were filled into the corresponding matrix positions of the core feature pairs, and the remaining positions were padded with zeros. Singular value decomposition was performed on the initial context matrix, and the left singular vectors corresponding to the first two principal singular values ​​were extracted to form an association mapping matrix. Association construction markers were simultaneously added to the association mapping matrix and stored in the metering box's local non-volatile memory and the cloud data platform.

[0050] In step S103, historical electrical data prior to the data gap in the metering box is obtained, and inverse deduction is performed using the aforementioned correlation mapping matrix to reconstruct the electrical characteristics of the metering box, including:

[0051] Extract three-phase current, three-phase voltage, and shell vibration data for a preset time period before the fault moment, and arrange them in chronological order to obtain the historical electrical operation sequence;

[0052] By analyzing the autocorrelation coefficients and partial autocorrelation coefficients of different time lag terms in the historical electrical operation sequence, the dynamic change pattern of the data is captured, and the equipment state evolution trajectory is obtained.

[0053] Feature components matching the dimension of the association mapping matrix are extracted from the equipment state evolution trajectory, projected onto the feature space of the association mapping matrix, and the magnitude of the projection result is calculated as the load fluctuation benchmark.

[0054] If the load fluctuation benchmark is greater than the preset load fluctuation threshold, the initial reconstructed feature sequence is calculated and derived.

[0055] The feature vectors corresponding to the initial reconstructed feature sequence are projected onto the orthogonal complement space of the association mapping matrix, and the feature dimensions corresponding to the elements whose absolute values ​​exceed a preset projection threshold in the projected vectors are extracted to obtain the electrical features of the metering box.

[0056] It should be noted that the preset duration is set to 60 sampling periods before the data thrombosis. The sampling frequency of this system is 10,000 times / second, with a single sampling period corresponding to 0.1ms, and 60 sampling periods corresponding to 6ms. Experiments show that more than 95% of the equipment state evolution characteristics are concentrated within the 6ms before the thrombosis, covering the core effective information of equipment state changes.

[0057] It should be noted that the preset load fluctuation threshold is 45. This threshold is determined by statistically analyzing the load fluctuation baseline distributions of 500 sets of normal operating conditions and 500 sets of fault conditions. The midpoint between the 99th percentile of the normal load fluctuation baseline distribution and the 1st percentile of the fault load fluctuation baseline distribution is used to distinguish between normal load fluctuations and load surges caused by faults. The order of the autocorrelation coefficient and partial autocorrelation coefficient ranges from order 1 to order 20. The upper limit of the order is determined by statistical analysis of the electrical sequences of 500 metering boxes before faults. An order of 20 can fully cover the time-series dependence characteristics of the data before the fault. The order of the autoregressive model is determined based on the truncation position of the partial autocorrelation coefficient. The model coefficients are obtained by solving the least squares method, and the coefficient vector forms the equipment state evolution trajectory. The initial reconstructed feature sequence is generated using a point-by-point recursive method. The last sample value of the historical electrical operation sequence is used as the initial input, and the sequence values ​​of the fault period are calculated point by point in the autoregressive model. The length of the initial reconstructed feature sequence is 20 sampling points, corresponding to the time range of the data fault period. The length of the 20 sampling points was determined by statistically analyzing the duration of 100 sets of meter box data discontinuity events. The length of the 20 sampling points covers the change process of fault characteristics during the data discontinuity period.

[0058] It should be noted that the association mapping matrix is ​​denoted as... , The number of rows is d, representing the dimension of the box feature sequence, and the number of columns is 2. The autoregressive model in the equipment state evolution trajectory is established according to the corresponding channels of three-phase current, three-phase voltage, and shell vibration data. First, the predicted increment sequence of each channel during the fault period is obtained. Then, the load matrix obtained from principal component analysis in step S101 is used to project the predicted increment of each channel onto the d-dimensional box feature space, forming a state feature vector z with the same number of rows as the correlation mapping matrix. After dimensionality reduction by principal component analysis, each dimension of z is mapped to a unified principal component space, eliminating the dimensional differences between the original three-phase current (A), three-phase voltage (V), and shell vibration acceleration (m / s²) data. Each dimension of z is a dimensionless relative coordinate value within the same feature space. When projecting onto the feature space spanned by the correlation mapping matrix, orthogonal projection of the column space of A is used, and the modulus of the projection result is used as the load fluctuation reference.

[0059]

[0060] in, The correlation mapping matrix has a dimension of d×2; for The transpose of the matrix; This is a d-dimensional state feature vector obtained by mapping the device state evolution trajectory; Let z be the orthogonal projection vector of column A onto column space; d is the L2 norm of the projection result; d is the dimension of the box feature sequence, which is 2 or 3. This is the vector norm 2. In this embodiment, the main singular values ​​retained are non-zero values. Reversible; if there is only one non-zero principal singular value, the number of associated mapping matrices is synchronously adjusted to 1 before the same projection calculation is performed. The dimension is d×d, and After multiplication, a d-dimensional vector is obtained. and Using the same relative amplitude unit, in the example The value is 52, which is greater than the preset load fluctuation threshold of 45. The judgment process is consistent with the formula output.

[0061] It should be noted that when the condition number of the association mapping matrix exceeds 1×10... 6 Direct inversion may lead to numerical instability. In this case, Tikhonov regularization is used, with the regularization parameter λ set to 1 × 10⁻³. The condition number can be determined after singular value decomposition by the ratio of the largest singular value to the smallest non-zero singular value.

[0062] It should be noted that the initial reconstructed feature sequence of the 20 sampling points is not directly used as the orthogonal complement space projection object. It should also be noted that this load matrix mapping method uses the same d×M load matrix obtained from the principal component analysis in step S101, where M is the total number of original sensor channels (corresponding to the three-phase current, three-phase voltage, enclosure surface temperature, and shell vibration acceleration channels), and d is either 2 or 3, completely consistent with the number of rows in the associated mapping matrix and the dimension of the state feature vector z. Instead, it is first mapped according to the aforementioned state feature vector... The same load matrix mapping method is used to convert it into a d-dimensional reconstructed feature vector. During the conversion, the mean increment and peak increment are calculated for each of the 20 sampling points of each channel to form the original reconstructed features, which are then projected onto the d-dimensional box feature space. The resulting d-dimensional reconstructed feature vector and the associated mapping matrix are in the same feature space, and then orthogonal complementary space projection is performed.

[0063]

[0064] in, It is a projection vector of the d-dimensional orthogonal complement space; It is a d-order identity matrix; This is the d-dimensional reconstructed feature vector obtained by mapping the initial reconstructed feature sequence. and Both are d×d matrices, and the difference between them is still a d×d matrix. Multiplication yields a d-dimensional vector. . The elements in the middle and Using the same relative amplitude unit, it can be compared with the preset projection threshold of 10 in the same dimension. In the example, the absolute values ​​of the 3rd and 8th dimensions are 12 and 15 respectively, both exceeding 10. The filtering results are consistent with the calculation logic of the formula.

[0065] It should be noted that the preset projection threshold is 10, which is the relative amplitude of the elements of the orthogonal complement space projection vector. The preset projection threshold is determined by statistically analyzing the absolute value distribution of the projection vector elements of 500 sets of normal reconstruction features and 500 sets of fault reconstruction features. The median value of the 99th percentile of the normal projection value distribution and the 1st percentile of the fault projection value distribution is taken to filter out the abnormal feature dimensions within the fault period. The orthogonal complement space projection operation is constructed based on the column space orthogonal complement space of the correlation mapping matrix. The vector obtained after projection separates the abnormal feature components that are not related to the known correlation pattern. The extracted electrical features of the metering box reflect the electrical anomaly state during the data fault period.

[0066] It is worth noting that the electrical characteristics of the metering box are simultaneously stored on the metering box's local non-volatile memory and the cloud data platform in binary format for a period of 3 years. The stored content includes the metering box's electrical characteristic vector, corresponding timestamps, load fluctuation baseline data, and initial reconstruction characteristic sequence. The storage duration is determined by the power industry's regulations on metering box fault tracing cycles, adapting to the fault tracing needs of power grid operation and maintenance.

[0067] For example, in a scenario where a lightning strike causes a data gap in a metering box, three-phase current, three-phase voltage, and casing vibration data for the 60 sampling periods prior to the data gap are extracted and arranged in timestamp order to obtain a historical electrical operation sequence. The autocorrelation coefficients and partial autocorrelation coefficients of different time lag terms in the historical electrical operation sequence are analyzed to capture the dynamic changes in data over time, resulting in an equipment state evolution trajectory containing 15 feature dimensions. Feature components matching the dimensions of the correlation mapping matrix are extracted from the equipment state evolution trajectory and projected onto the feature space spanned by the correlation mapping matrix. The modulus of the projection result is calculated, yielding a load fluctuation baseline of 52, exceeding a preset load fluctuation threshold of 45. Using the last valid data point of the historical electrical operation sequence as the initial input, the obtained autoregressive model is substituted point by point for recursive extrapolation, resulting in an initial reconstructed feature sequence with a length of 20 sampling points. The feature vectors corresponding to the initial reconstructed feature sequence are projected onto the orthogonal complement space of the correlation mapping matrix to obtain the projection vector. The absolute values ​​of each element in the projection vector are extracted. The absolute values ​​of the 3rd and 8th dimensions are 12 and 15, respectively, which exceed a preset projection threshold of 10. The corresponding feature dimensions are then extracted to obtain the electrical features of the metering box. Reconstruction markers are simultaneously added to the electrical features of the metering box and stored in the metering box's local non-volatile memory and the cloud data platform.

[0068] In one implementation, the method for monitoring the status of an energy metering box based on multi-sensor fusion is characterized by further comprising:

[0069] Extract the current value sequence from the electrical characteristics of the metering box, and calculate the waveform distortion index from the current value sequence;

[0070] The filtering parameters are adjusted according to the waveform distortion index, and the current value sequence is filtered and compensated to obtain a corrected current value sequence. The peak value of the power frequency component in the corrected current value sequence is extracted as the current fundamental amplitude.

[0071] If the amplitude of the current fundamental wave does not exceed the preset current verification threshold, then the corresponding feature dimension is determined to be physically consistent.

[0072] All feature dimensions that conform to the physical consistency are retained and combined to obtain effective electrical feature components.

[0073] It should be noted that the current value sequence extracted from the electrical characteristics of the metering box has a length of 20 sampling points, corresponding to continuous current sampling values ​​during the data discontinuity period. The preset harmonic order range is set to 2nd to 15th. The harmonic order range of 2nd to 15th is determined by statistically analyzing the current harmonic distribution data of 500 sets of metering box hardware failure scenarios. The effective harmonic components caused by the fault are mainly concentrated in the 2nd to 15th order range, which covers more than 95% of the fault harmonic characteristics, while avoiding the introduction of excessively high-order random noise components.

[0074] It should be noted that the waveform distortion index uses the total harmonic distortion rate (THD) calculation method, and the calculation formula is as follows:

[0075]

[0076] Where D represents the degree of waveform distortion, and I n In is the effective value of the nth harmonic current, and I1 is the effective value of the fundamental frequency current. The summation ranges from the 2nd to the 15th harmonics. The larger the waveform distortion value, the more abnormal harmonic components are contained in the waveform. A Butterworth low-pass filter is used for filtering compensation. The filter parameters are dynamically adjusted according to the waveform distortion index. For every 0.05 increase in waveform distortion, the filter cutoff frequency decreases by 30Hz, and the order increases by 1. The minimum cutoff frequency is no less than 100Hz, and the maximum order is no more than 6, filtering out abnormal harmonics while retaining the effective fundamental frequency information.

[0077] It should be noted that the initial cutoff frequency of the Butterworth low-pass filter is set to 500Hz, and the initial order is set to 2. When the waveform distortion D does not exceed 0.10, the initial cutoff frequency and initial order are used; when D exceeds 0.10, for every 0.05 increase in D relative to 0.10, the cutoff frequency decreases by 30Hz and the order increases by 1. The lower limit of the cutoff frequency is set at 100Hz, and the upper limit of the order is set at 6.

[0078] The power frequency component is set to 50Hz, corresponding to the rated power frequency of the domestic power grid. The fundamental amplitude reflects the actual amplitude level of the current, which is suitable for determining the operating parameters of the metering box.

[0079] It should be noted that the preset current verification threshold is 40A, applicable to three-phase energy metering boxes with a rated current of 60A. For metering boxes with other rated currents, a correction is made using I_th = 40 × √(I_rated / 60), where I_rated is the rated current of the metering box in A, and I_th is the corrected current verification threshold in A. The preset current verification threshold is determined by statistically analyzing the current fundamental amplitude distribution under 500 sets of normal operating conditions and 500 sets of data fault reverse inference conditions. The median value of the 99th percentile of the current fundamental amplitude distribution under normal operating conditions and the 1st percentile of the current fundamental amplitude distribution under data fault reverse inference conditions is taken to distinguish between real current values ​​that conform to physical laws and false abnormal current values ​​generated by data fault reverse inference. Physical consistency judgment is used to eliminate false features caused by data fault reverse inference and retain the actual valid features obtained from the inference.

[0080] It should be noted that after the physical consistency judgment is passed, the correlation between the corrected current value sequence and the standard current waveform in the historical normal waveform database is further verified. The historical normal waveform database is composed of typical current waveforms collected from the same metering box under normal operating conditions and stored according to load range. If the Pearson correlation coefficient between the corrected current value sequence and the standard waveform is less than 0.85, the result of this reverse inference reconstruction is deemed unreliable, the total harmonic distortion rate calculation is terminated, and an unreliable data flag is output. The corresponding feature dimension is not included in the effective electrical feature components. The standard waveform is the waveform template corresponding to the load range closest to the current load level.

[0081] It is worth noting that the effective electrical characteristic components are simultaneously stored in the metering box's local non-volatile memory and on the cloud data platform in binary format for a period of 3 years. The stored content includes the effective electrical characteristic component vector, corresponding timestamp, waveform distortion level data, and corrected current value sequence. The storage duration is determined by the power industry's regulations on metering box fault tracing cycles, adapting to the fault tracing needs of power grid operation and maintenance.

[0082] In another implementation, the method for monitoring the status of an energy metering box based on multi-sensor fusion is characterized by further comprising:

[0083] The current fluctuation values ​​of the target metering box and the associated metering boxes adjacent to it in the physical space are obtained at continuous time sampling points before and after the anomaly occurs. A multi-dimensional spatiotemporal matrix is ​​constructed with spatial nodes as rows and time sampling points as columns.

[0084] Singular value decomposition is performed on the multidimensional spatiotemporal matrix, and the time coefficient sequence corresponding to the first principal component is extracted as the spatial mapping feature sequence. The high-dimensional spatiotemporal data is projected onto the low-dimensional feature space, and the spatial mapping feature sequence is expanded in phase space to obtain the temporal reconstruction sequence.

[0085] If the fluctuation amplitude of the time-series reconstructed sequence exceeds a preset fluctuation amplitude threshold, then a nonlinear data transformation is performed on the time-series reconstructed sequence to obtain the transformed feature sequence;

[0086] The numerical values ​​of each dimension of the transformed feature sequence are linearly scaled to map all dimensions to a unified numerical range to complete dimension normalization, thus obtaining standardized electrical features.

[0087] It should be noted that the standard configuration of the multidimensional spatiotemporal matrix includes 5 spatial nodes and 12 continuous time sampling points. The 5 spatial nodes consist of 1 target metering box and 4 physically adjacent associated metering boxes. The selection of associated metering boxes prioritizes the metering boxes with the closest electrical distance within the same distribution room, calculated using a weighted average of distribution topology and line impedance. If the number of available metering boxes within the same distribution room is insufficient, the selection range is expanded to neighboring metering boxes within the power supply range of the transformer in the same area, selected sequentially from smallest to largest electrical distance until the required number is met. The 12 continuous time sampling points cover the complete time period before and after the anomaly. When there are fewer than 4 physically adjacent available associated metering boxes, the number of matrix rows is adjusted according to the actual number of available nodes, and missing node positions are filled with historical data from the same distribution area during the same period, ensuring a minimum of 2 spatial nodes. The number of spatial nodes and time sampling points were determined through feature correlation experiments on 500 sets of metering box cluster fault scenarios. The combination of 5 nodes and 12 sampling points covers more than 90% of the spatial correlation features and temporal evolution information, while controlling data dimensions to avoid computational redundancy. When the number of nodes is reduced to 2, the timing reconstruction error does not exceed 12% of the standard configuration, which meets the accuracy requirements for fault feature extraction.

[0088] It should be noted that the phase space expansion process is set with a delay time parameter of 2 and an embedding dimension parameter of 3. The combination of delay time and embedding dimension parameters is determined through a time series feature reconstruction experiment of 200 sets of different fault types. When the parameter combination is 2 and 3, the reconstructed time series retains the dynamic evolution law of the original features, and the reconstruction error is less than 5%, which meets the accuracy requirements for subsequent fluctuation amplitude determination.

[0089] It should be noted that the time coefficient sequence corresponding to the first principal component is denoted as y. Phase space expansion constructs embedding vectors with a delay time of 2 and an embedding dimension of 3. Each embedding vector consists of the current time coefficient, a time coefficient delayed by 2 sampling points, and a time coefficient delayed by 4 sampling points. For a spatial mapping feature sequence of length 12, expansion yields 8 three-dimensional embedding vectors. To ensure that subsequent fluctuation amplitude determination maintains a scalar form, the L2 norm of each three-dimensional embedding vector is calculated and divided by the square root of the embedding dimension to obtain a scalarized temporal reconstruction sequence; this scalar retains the amplitude units of the original time coefficient sequence.

[0090]

[0091] in, Let i be the i-th 3D embedding vector; , and These are the time coefficients of the first principal component time coefficient sequence at the i-th, i+2-th, and i+4-th sampling points, respectively; Indicates vector transpose; v_i represents the time-series reconstructed value after scaling up the i-th embedding vector; i is the embedding vector index, ranging from 1 to 8. The three components of v_i come from the same time coefficient sequence and have the same units. and Using the same amplitude unit, divide by The units remain the same. A time coefficient sequence of length 12, with a delay time of 2 and an embedding dimension of 3, yields 12-(3-1)×2=8 embedding vectors. The number of 8 time series reconstruction elements in the example is correct.

[0092] It should be noted that the preset fluctuation amplitude threshold is 50A. The preset fluctuation amplitude threshold is determined by statistically analyzing the fluctuation amplitude distribution of 500 sets of time-series reconstructed sequences under normal operating conditions and 500 sets of fault conditions. The median value of the 99th percentile of the fluctuation amplitude distribution under normal conditions and the 1st percentile of the fluctuation amplitude distribution under fault conditions is taken to distinguish between normal time-series fluctuations and drastic amplitude changes caused by faults, and to screen out strong fluctuation characteristics that require nonlinear transformation.

[0093] It should be noted that the nonlinear data transformation employs logarithmic function operations to compress numerical differences within large-amplitude ranges and amplify feature details within small-amplitude ranges. A unified numerical range is set to 0 to 1. The linear scaling process maps values ​​according to the maximum and minimum values ​​of each dimension, unifying all dimensional values ​​to the same range, eliminating dimensional differences, and satisfying the subsequent multi-source feature fusion operation rules. The unified numerical range of 0 to 1 was determined through multi-source feature fusion adaptation experiments, eliminating differences in dimensionality and numerical range across different dimensions.

[0094] It is worth noting that standardized electrical characteristics are simultaneously stored on the metering box's local non-volatile memory and the cloud data platform in binary format for a period of 3 years. The stored content includes standardized electrical characteristic vectors, corresponding timestamps, fluctuation amplitude data, and converted characteristic sequences. The storage duration is determined by the power industry's regulations on metering box fault tracing cycles, adapting to the fault tracing needs of power grid operation and maintenance.

[0095] For example, in a scenario where a lightning strike damages the metering box hardware, the current fluctuation values ​​of the target metering box and four physically adjacent related metering boxes are obtained at 12 consecutive time sampling points before and after the anomaly occurs. A 5-row, 12-column multidimensional spatiotemporal matrix is ​​constructed with spatial nodes as rows and time sampling points as columns. Singular value decomposition is performed on the multidimensional spatiotemporal matrix, and the feature vectors corresponding to the three largest singular values ​​are extracted. The time coefficient sequence corresponding to the first principal component is extracted as a spatial mapping feature sequence with a length of 12. A phase space expansion with a delay time of 2 and an embedding dimension of 3 is performed on the spatial mapping feature sequence to obtain a time-series reconstruction sequence containing 8 elements. The difference between the maximum and minimum values ​​in the time-series reconstruction sequence is calculated, and the fluctuation amplitude is found to be 68A, exceeding the preset fluctuation amplitude threshold of 50A. A nonlinear data transformation in the form of a logarithmic function is performed on the time-series reconstruction sequence to obtain the transformed feature sequence. Linear scaling is performed on the values ​​of each dimension of the transformed feature sequence to map all dimensions to a unified numerical range of 0 to 1, completing dimension normalization and obtaining standardized electrical features. Standardized electrical features are synchronously marked with feature regularization tags and stored in the metering box's local non-volatile memory and cloud data platform.

[0096] In another implementation, the method for monitoring the status of an energy metering box based on multi-sensor fusion is characterized by further comprising:

[0097] Extract the high-frequency transient voltage sequence of the standardized electrical characteristics and the box surface temperature sequence from the box characteristic sequence, and obtain a synchronization state snapshot based on the high-frequency transient voltage sequence and the box surface temperature sequence;

[0098] Based on the synchronous state snapshot, transient stress data and insulation breakdown data are obtained. The correlation coefficient between the transient stress data and the insulation breakdown data is calculated to obtain a cross-correlation sequence.

[0099] The variance of the cross-correlation sequence is calculated as the deviation magnitude. If the deviation magnitude is greater than a preset deviation magnitude threshold, the information entropy of the transient stress data and the insulation breakdown data are calculated respectively.

[0100] Based on the information entropy, a fusion weight is assigned, and the fusion weight is linearly weighted with the transient stress and the insulation breakdown data to obtain a weighted fusion feature.

[0101] The fault panorama features are obtained based on the weighted fusion features.

[0102] In this implementation, forced timestamp alignment employs dynamic time warping to eliminate time offsets between heterogeneous data. The synchronous state snapshot contains 30 discrete data points. The number of 30 discrete data points was determined through 500 sets of multi-source feature fusion experiments, covering the transition period before and after the fault, while ensuring the temporal resolution meets the accuracy requirements of subsequent correlation analysis. Transient stress data is calculated from the box surface temperature sequence in the synchronous state snapshot. A sliding difference method with a window length of 3 is used to calculate the ratio of the temperature difference between adjacent sampling points to the time step, obtaining the temperature mutation rate value corresponding to each sampling point. Insulation breakdown data is calculated from the high-frequency transient voltage sequence in the synchronous state snapshot. The relative deviation between the instantaneous voltage amplitude and the power frequency fundamental amplitude at each sampling point is used as the transient voltage distortion value. The lengths of both sets of data are consistent with the number of sampling points in the synchronous state snapshot, and each point participates in the subsequent correlation coefficient calculation.

[0103] It should be noted that the sliding difference with a window length of 3 is calculated by dividing the difference between the first and last sampling points of the window by the time interval between the first and last sampling points. For the i-th window, the temperature difference between the i-th sampling point and the (i+2)-th sampling point on the box surface is taken and divided by 2 sampling intervals to obtain the temperature abrupt change rate. The transient voltage distortion value is the absolute value of the difference between the instantaneous voltage amplitude and the fundamental frequency amplitude, divided by the fundamental frequency amplitude. Before entering the correlation coefficient, information entropy, and weighted fusion calculations, both the temperature abrupt change rate sequence and the transient voltage distortion sequence are linearly normalized to the interval between 0 and 1 according to their respective maximum and minimum values.

[0104] It should be noted that the temperature abrupt change rate s_i, as a transient stress indicator, plays an auxiliary verification role in millisecond-level fault scenarios such as lightning strikes or arc discharges. In these scenarios, the surface temperature of the enclosure has a thermal time constant on the order of seconds to minutes, and its variation is minimal within the 10ms to 20ms window of the fault duration. The main transient stress characteristics are borne by the high-frequency transient voltage distortion u_i. The temperature abrupt change rate plays a major independent role in fault scenarios lasting longer than seconds, such as contact resistance overheating and long-term overload heating. In actual engineering deployments, the fusion weight allocation range of the two types of transient stress characteristics can be adjusted according to prior knowledge of the fault type or the initial classification results. For lightning strike faults, the transient stress weight can be adjusted to 10% to 20%, and the insulation breakdown weight can be adjusted to 80% to 90%.

[0105]

[0106] in, Let be the temperature abrupt change rate corresponding to the i-th sliding window; Let be the surface temperature of the box at the i-th sampling point, in °C; The temperature of the box surface at the (i+2)th sampling point is expressed in °C. The time interval between adjacent sampling points, in seconds; This represents the transient voltage distortion value corresponding to the i-th sampling point; Let V be the instantaneous voltage amplitude at the i-th sampling point, in V. This represents the amplitude of the fundamental frequency voltage, expressed in volts (V). The unit is ℃ / s. The ratio is dimensionless; both are normalized to the 0 to 1 range before entering the correlation coefficient, information entropy and weighted fusion calculation to avoid the direct addition of data with different dimensions.

[0107] It should be noted that with the sliding window length set to 5 and the sliding step size set to 1, the calculated cross-association sequence length is 26. The parameter combination of window length and step size was determined through 200 sets of feature association analysis experiments. When the parameter combination is 5 and 1, local random fluctuations are smoothed while retaining the detailed changes in association features, and the association analysis error is less than 4%.

[0108] It should be noted that the preset deviation threshold is 0.15. The preset deviation threshold of 0.15 is determined by statistically analyzing the cross-correlation sequence variance distribution of 500 sets of normal operation states and 500 sets of hardware failure states. The midpoint between the 99th percentile of the variance distribution of the normal state and the 1st percentile of the variance distribution of the failure state is taken to distinguish between normal correlation fluctuations and correlation deviations caused by faults, thus triggering the subsequent weight allocation process.

[0109] It should be noted that the fusion weights are calculated separately for the information entropy of the transient stress data sequence and the insulation breakdown data sequence. The lower the information entropy, the higher the dispersion of the feature dimension and the greater the amount of information it carries, and the higher the corresponding weight value. The weight allocation process ensures that the sum of the weights of the two types of features is 100%. The transient stress weight ranges from 30% to 50%, and the insulation breakdown weight ranges from 50% to 70%, with typical operating conditions using 40% and 60%, respectively.

[0110] It should be noted that the information entropy used is the discrete Shannon information entropy. For the normalized transient stress data or insulation breakdown data x, the probability value of each sampling point is first calculated. The probability value is the sum of the value of that sampling point plus a small constant, divided by the sum of the values ​​of the entire sequence plus a small constant. The small constant is set to 0.000001 to avoid zero values ​​that would prevent the logarithm from being calculated. The sequence length is denoted as m, and the information entropy is calculated according to the natural logarithm and normalized using lnm. The normalized information entropy ranges from 0 to 1.

[0111]

[0112] in, This represents the probability value corresponding to the i-th sampling point; This refers to the value of the i-th sampling point in the normalized transient stress data or insulation breakdown data; For a small constant, take the value 0.000001; The length of the sequence used in the information entropy calculation; This is the normalized discrete Shannon information entropy. and All are dimensionless values. It is a dimensionless probability and all The sum is 1; After normalization by ln(m), the value ranges from 0 to 1, which conforms to the theoretical range of discrete information entropy.

[0113] It should be noted that the mapping from entropy values ​​to fusion weights is calculated using discrete scores. The discrete score for transient stress data is 1 minus the transient stress information entropy, and the discrete score for insulation breakdown data is 1 minus the insulation breakdown information entropy. The transient stress weight is allocated according to the proportion of discrete scores within the range of 30% to 50%, while the insulation breakdown weight is 1 minus the transient stress weight. When the sum of the discrete scores of the two types is 0, the transient stress weight is set to 40%, and the insulation breakdown weight is set to 60%. Both types of sequences have been normalized before weighted fusion, and the linear weighted result is a dimensionless value.

[0114]

[0115] in, The fusion weights for transient stress data; d_s represents the fusion weights for insulation breakdown data; d_s represents the dispersion score for transient stress data. The dispersion score for insulation breakdown data; The information entropy of transient stress data; The information entropy of the insulation breakdown data. and The value range of is from 0 to 1, therefore and The value range is also from 0 to 1; when and When the sum is greater than 0, It is between 0.30 and 0.50. Located between 0.50 and 0.70, and The two types of sequences have been normalized before weighting, and the linear weighting result is a dimensionless value.

[0116] It should be noted that the multidimensional feature matrix is ​​set to 3 rows and 8 columns. The three rows correspond to the three time stages before the fault occurs, during the fault, and after the fault occurs, respectively. The eight columns correspond to the eight feature dimensions: mean temperature mutation rate, peak temperature mutation rate, mean transient voltage distortion, peak transient voltage distortion, mean cross-correlation degree, transient stress information entropy, insulation breakdown information entropy, and comprehensive weighted fusion value. The segmentation and recombination process divides the 30 discrete sampling points into three stages according to the time ratio: the first 10 sampling points in the pre-fault stage, the middle 10 sampling points in the during-fault stage, and the last 10 sampling points in the post-fault stage. Within each stage, statistical values ​​are calculated for the eight dimensions of the weighted fusion feature and filled into the corresponding columns of the row to form a complete multidimensional feature matrix. The setting of the three time stages and eight feature dimensions is determined through statistical analysis of the evolution law of fault features, presenting the full-time feature evolution process of the fault from incubation to occurrence and subsequent events.

[0117] It is worth noting that the fault panorama features are simultaneously stored on the metering box's local non-volatile memory and the cloud data platform in binary format for a period of 3 years. The stored content includes a fault panorama feature matrix, corresponding timestamps, cross-correlation sequence data, and weighted fusion features. The storage duration is determined by the power industry's regulations on fault tracing cycles for metering boxes, adapting to the fault tracing needs of power grid operation and maintenance.

[0118] For example, in a scenario where lightning strikes damage the metering box hardware, the high-frequency transient voltage sequence from the standardized electrical features and the box surface temperature sequence from the box feature sequence are extracted. Dynamic time warping is used to calculate the shortest distortion path between the two sequences, achieving forced timestamp alignment and obtaining a synchronized state snapshot containing 30 discrete data points. Temperature mutation rate data is separated from the synchronized state snapshot as transient stress data, and transient voltage distortion data is separated as insulation breakdown data. A fixed-length sliding window with a window length of 5 and a step size of 1 is used to calculate the correlation coefficient between the transient stress data and the insulation breakdown data window by window, resulting in a cross-correlation sequence of length 26. The variance of the cross-correlation sequence is calculated as the deviation magnitude, yielding a deviation magnitude of 0.28, exceeding a preset deviation magnitude threshold of 0.15. The information entropy of the transient stress data sequence and the insulation breakdown data sequence is calculated separately. Fusion weights are dynamically allocated based on the information entropy magnitude, with a weight ratio of 40% for the transient stress dimension and 60% for the insulation breakdown dimension. The transient stress data and insulation breakdown data are linearly weighted according to the assigned fusion weights to obtain weighted fusion features. These weighted fusion features are then divided into three stages—before the fault occurs, during the fault, and after the fault—based on the time proportions of the first 10 sampling points, the middle 10 sampling points, and the last 10 sampling points. For each stage, the corresponding statistical values ​​of the eight feature dimensions are recombined to construct a 3x8 multidimensional feature matrix, resulting in a comprehensive fault profile. Fusion construction markers are simultaneously added to the comprehensive fault profile features, and they are stored in the metering box's local non-volatile memory and the cloud data platform.

[0119] In step S104, the fault panorama features are compared element-by-element with a preset benchmark sample library to obtain fault type labels through classification and identification. High-frequency abrupt change components are extracted to locate the time of internal short circuit occurrence, and the final state monitoring results are output, including:

[0120] The multidimensional feature matrix corresponding to the fault panoramic feature is compared element by element with each standard feature matrix in the preset benchmark sample library to calculate the multidimensional similarity matrix.

[0121] Traverse the multidimensional similarity matrix, extract the global maximum value as the extreme value feature, and classify and determine the initial fault type label based on the sample index corresponding to the extreme value feature.

[0122] By combining the initial fault type label, the high-frequency transient voltage sequence corresponding to the synchronization state snapshot is extracted to determine the time when the internal short circuit occurs;

[0123] Using the initial fault type label and the time of occurrence of the internal short circuit as a joint index, the preset state mapping table is queried to obtain the final state monitoring result and output it.

[0124] It should be noted that the preset benchmark sample library contains standard feature matrices for 500 typical operating conditions, including 200 sets of normal operation conditions, 150 sets of internal short circuit conditions, 100 sets of insulation aging conditions, and 50 sets of mechanical damage conditions. The dimensions of each standard feature matrix are consistent with the fault panorama feature matrix, both being a 3x8 column structure. The sample library was constructed by collecting metering box operating characteristic data under different fault types and severity levels. The data sources include field measured data and semi-physical simulation data. K-means clustering was used to complete feature clustering, and the number of clusters was determined by the elbow rule. The clustering results were manually calibrated after matching the fault type and occurrence time recorded in the field fault work order with each fault physical judgment rule. Clusters with high voltage distortion peaks and significant temperature abrupt change rates correspond to internal short circuit types; clusters with slowly increasing voltage distortion and gradual temperature changes correspond to insulation aging types; clusters with large vibration characteristic abrupt changes and secondary electrical characteristic changes correspond to mechanical damage types; and clusters with fluctuations in all dimensions within the normal range correspond to normal operation types. The calibration process is completed independently by two or more calibration personnel. Samples with inconsistent labels are verified against the original fault records. Each sample corresponds to an independent fault type identifier and severity level. Cosine similarity calculation is performed item by item according to time stage and feature dimension. The similarity is calculated separately for the physical dimension and the electrical dimension, and then the average value is taken to obtain the comprehensive similarity of a single sample. The similarity of all samples is arranged in order to form a multidimensional similarity matrix.

[0125] It should be noted that the preset benchmark sample library stores 500 sets of original standard feature matrices after manual calibration, and does not store K-means cluster centroids. K-means clustering is only used for auxiliary grouping during the sample library construction phase. The elbow rule determines the number of clusters to be 4, corresponding to four operating conditions: normal operation, internal short circuit, insulation aging, and mechanical damage. Each set of standard feature matrices retains a 3-row, 8-column structure and the corresponding fault type, severity level, source identifier, and sampling time.

[0126] It should be noted that in the 3x8 multidimensional feature matrix, columns 1 and 2 are classified as the physical dimension, columns 3 and 4 as the electrical dimension, and columns 5 through 8 as the coupling dimension. When calculating similarity, the cosine similarity of the physical, electrical, and coupling dimensions is calculated separately. The three similarities are then weighted with weights of 0.35, 0.35, and 0.30 respectively to obtain the comprehensive similarity of a single sample. This avoids forcibly classifying the mean of cross-correlation and the comprehensive weighted fusion value into a single physical or electrical dimension.

[0127]

[0128] in, The overall similarity of a single sample; Cosine similarity in physical dimension; Cosine similarity in the electrical dimension; Cosine similarity for the coupling dimension. , and All are cosine similarity values, ranging from 0 to 1; 0.35, 0.35, and 0.30 are the fusion coefficients corresponding to the three dimensions, and the sum of the three is 1. It remains a dimensionless similarity, with a value ranging from 0 to 1.

[0129] It should be noted that the initial fault type label is determined by the sample index corresponding to the extreme value feature. The preset similarity confidence threshold is 0.6. When the global maximum value is lower than the similarity confidence threshold, it is determined as an unknown fault type. When the global maximum value is higher than the similarity confidence threshold, the sample index corresponding to the extreme value feature and the similarity value are used as input to output the fault type classification result. The classification is completed using labeled data from the benchmark sample library, covering common fault types such as normal operation, internal short circuit, insulation aging, and mechanical damage.

[0130] It should be noted that the preset similarity confidence threshold of 0.6 was determined using 300 sets of validation samples, including 240 sets of samples of known types and 60 sets of perturbation samples not covered by the sample library. Whether the overall similarity is higher than the candidate threshold is used as the boundary for determining known and unknown faults. The F1 score is used as the evaluation index, and the search is performed in steps of 0.05 between the candidate thresholds of 0.4 and 0.8. The F1 score corresponding to the threshold of 0.6 is the highest, and therefore determined as the similarity confidence threshold. When the global maximum value is lower than 0.6, the unknown fault type identifier, the maximum similarity value, and the time range of the anomaly occurrence are output, and a manual review is triggered, without proceeding to the state mapping of a specific fault type.

[0131] It should be noted that a three-layer decomposition structure was selected for multi-scale decomposition, which was determined through comparative experiments on the fault feature extraction effect, effectively separating the high-frequency abrupt change component at the moment of fault occurrence. The high-frequency abrupt change component corresponds to the transient feature change when a short-circuit fault occurs. The first modulus maximum point corresponds to the fault initiation time. The time offset is calculated based on the initiation time of the fault panoramic feature, and the time accuracy is consistent with the original sampling frequency.

[0132] It should be noted that the multi-scale decomposition uses the Daubechies 4 wavelet basis function, with a decomposition level of 3. The Daubechies 4 wavelet basis function was determined through comparative experiments of 200 sets of internal short-circuit high-frequency transient voltage sequences. Its error in locating the modulus maxima of the fault initiation point is less than 1 sampling period, which is suitable for the sampling frequency and time location accuracy of this embodiment.

[0133] It is worth noting that the preset status mapping table stores status descriptions, risk levels, and handling suggestions corresponding to different fault types and their occurrence times. Risk level classification is determined by combining fault type attributes and similarity values. Internal short-circuit faults are classified into the high-risk range, while insulation aging and mechanical damage faults are categorized into risk levels based on similarity values; similarity values ​​above 0.8 correspond to higher risk levels, and values ​​below 0.8 correspond to lower risk levels. A combined index quickly locates the corresponding entries and outputs standardized monitoring results. The final status monitoring result includes four items: fault type identifier, fault occurrence time, risk level, and handling suggestions. The results are simultaneously stored locally and in the cloud, with the storage duration consistent with the original feature data.

[0134] It should be noted that the preset status mapping table is indexed by fault type, occurrence time interval, and similarity interval. The fault type field includes normal operation, internal short circuit, insulation aging, mechanical damage, and unknown fault; the occurrence time interval is divided into 0ms to 5ms, 5ms to 10ms, 10ms to 20ms, and more than 20ms after the monitoring start point; the similarity interval is divided into 0.6 to 0.8 and 0.8 to 1.0. The internal short circuit type is matched with the internal short circuit status description, and the risk level increases with the earlier the occurrence time and the similarity interval; the insulation aging and mechanical damage types do not use the internal short circuit occurrence time as an index, but use the identification time and similarity interval as indexes; the unknown fault type outputs the unknown fault status description and manual verification mark.

[0135] It's worth noting that when the initial fault type label does not belong to internal short circuits, the final state monitoring result is directly generated using the fault type label and the corresponding similarity value, without needing to execute the time-based location process. Insulation aging and mechanical damage are slowly changing faults, and their development process lacks clear transient abrupt changes. Therefore, the time-based location accuracy cannot meet the requirements of engineering applications. Using the fault identification time as the monitoring output time is sufficient for maintenance needs. The monitoring results for non-internal short circuit faults include the fault type, similarity value, and risk level. The risk level is determined according to the classification rules in the preset state mapping table; higher similarity corresponds to higher fault reliability, and the risk level increases accordingly.

[0136] It should be noted that for non-internal short-circuit faults, the preset state mapping table is still queried, but the index fields are fault type, identification time, and similarity range, and the occurrence time of internal short circuit is not used. The identification time for insulation aging and mechanical damage is the end time of the current monitoring window, and the risk level is determined by the fault type and similarity range. When no corresponding entry is found in the mapping table, an unknown fault type identifier is output and a manual review flag is triggered.

[0137] For example, in a scenario where a lightning strike causes an internal short circuit in a metering box, a 3x8 multidimensional feature matrix corresponding to the fault panorama features is input. A preset benchmark sample library contains 500 sets of standard feature matrices, covering various operating conditions such as normal operation, internal short circuit, and insulation aging. The fault panorama feature matrix is ​​compared element-by-element with all standard feature matrices in the benchmark sample library. The cosine of the angle between the physical and electrical dimensions is calculated in three stages: before, during, and after the fault, resulting in a 1x500 multidimensional similarity matrix. The multidimensional similarity matrix is ​​traversed, and the global maximum value of 0.92 is extracted as the extreme value feature, corresponding to sample number 145. The initial fault type label is determined as internal short circuit through classification. The high-frequency transient voltage sequence is extracted from the synchronous state snapshot, and multi-scale decomposition is performed to separate the high-frequency abrupt change components. The first modulus maxima of the high-frequency abrupt change components is identified, yielding a corresponding time offset of 12ms. Using the monitoring start time as a benchmark, the internal short circuit occurrence time is determined to be 12ms after the monitoring start time. Using the internal short-circuit fault type and the occurrence time at 12ms as a joint index, a preset status mapping table is queried to obtain the final status monitoring result: the fault type is internal short circuit, the fault occurrence time is 12ms after the monitoring start point, the risk level is high risk, and the handling recommendation is to immediately disconnect power and investigate. The result is synchronously output to the monitoring terminal and simultaneously stored in local non-volatile memory and cloud data platform.

[0138] It should be noted that the overall similarity in this example simultaneously calculates the cosine similarity of the coupling dimensions in columns 5 to 8, and weights the similarities of the physical, electrical, and coupling dimensions with weights of 0.35, 0.35, and 0.30, respectively, resulting in a 1-row, 500-column multidimensional similarity matrix. The overall similarity for sample number 145 is 0.92, which is higher than the preset similarity confidence threshold of 0.6, and the classification result is consistent with the value range of the overall similarity formula.

[0139] In summary, this invention provides a method for monitoring the status of power metering boxes based on multi-sensor fusion, which effectively solves the problems of existing technologies being unable to effectively restore the complete fault characteristics at the moment of hardware failure, and also unable to achieve fault type identification and location of occurrence time under multi-dimensional data fusion, making it difficult to balance the real-time performance and engineering practicality of metering box status monitoring.

[0140] The second embodiment of the present invention provides a power meter box status monitoring system based on multi-sensor fusion, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the power meter box status monitoring method based on multi-sensor fusion as described above.

[0141] It should be noted that the power metering box status monitoring system based on multi-sensor fusion provided in this embodiment of the invention is used to execute all the process steps of the power metering box status monitoring method based on multi-sensor fusion in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0142] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0143] It should be noted that the preset threshold parameters in this embodiment (including the voltage trigger threshold of 250V, the preset gradient threshold of 25, the preset correlation coefficient threshold of 0.8, the preset load fluctuation threshold of 45, the preset projection threshold of 10, the preset current verification threshold of 40A, the preset fluctuation amplitude threshold of 50A, the preset deviation amplitude threshold of 0.15, and other fixed thresholds) are all determined through statistical experiments based on the typical operating conditions of a three-phase power metering box with a rated voltage of 220V to 380V, a rated current of 60A, and an ambient temperature of 10℃ to 40℃. When applied to metering boxes with different rated voltage levels or different rated current specifications, the electrical thresholds (voltage trigger threshold, current verification threshold, fluctuation amplitude threshold, etc.) can be corrected by coefficients proportional to the square root of the rated electrical parameters; the statistical characteristic thresholds (gradient threshold, correlation coefficient threshold, projection threshold, deviation amplitude threshold, etc.) can be recalibrated based on the target operating condition data under the same statistical method.

[0144] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for monitoring the status of an energy metering box based on multi-sensor fusion, characterized in that, include: Continuous time-series data of the meter box at the moment of hardware failure is acquired and preprocessed to obtain the box feature sequence; Based on the aforementioned box feature sequence, an association mapping matrix characterizing the relationship between physical quantities and electrical quantities is constructed; Historical electrical data of the metering box before the data gap is obtained, and the electrical characteristics of the metering box are reconstructed by reverse deduction in combination with the aforementioned correlation mapping matrix. The box feature sequence is fused with the reconstructed electrical features of the metering box in multiple dimensions to obtain a fault panorama feature, and the fault type and the time of fault occurrence are identified based on the fault panorama feature.

2. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 1, characterized in that, The process of acquiring and preprocessing continuous time-series data at the moment of hardware failure in the metering box to obtain a box feature sequence includes: Continuous time-series data before and after the meter box hardware failure is obtained, and the data is truncated based on a preset voltage trigger threshold to obtain the target time period sequence; The target time series is subjected to outlier removal and interpolation to obtain a clean time series. Perform Kalman filtering on the pure time series to obtain a smooth time series; Principal component analysis was performed on the smoothed time series to reduce its dimensionality, resulting in a box-shaped feature sequence.

3. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 1, characterized in that, The construction of the correlation mapping matrix representing the relationship between physical and electrical quantities based on the box feature sequence includes: The box feature sequence is segmented to obtain multiple time slice features. The difference in node split gain between adjacent time slice features is calculated. The difference is divided by the time interval step size to obtain the feature change gradient. If the gradient of the feature change exceeds a preset gradient threshold, the correlation coefficient between the corresponding physical feature dimension and the electrical load dimension is extracted, and the core feature pairs whose absolute values ​​of the correlation coefficients exceed the preset correlation coefficient threshold are selected. An initial context matrix is ​​constructed with physical feature dimension as rows and electrical load dimension as columns. The matrix elements are the correlation coefficient values ​​of the corresponding core feature pairs, and the elements at the non-core feature pairs are set to zero. Singular value decomposition is performed on the initial context matrix to extract the left singular vectors corresponding to the main singular values, thus obtaining the correlation mapping matrix.

4. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 1, characterized in that, The process involves acquiring historical electrical data prior to the data gap in the metering box, combining this data with the correlation mapping matrix to perform reverse deduction, and reconstructing the electrical characteristics of the metering box, including: Extract three-phase current, three-phase voltage, and shell vibration data for a preset time period before the fault moment, and arrange them in chronological order to obtain the historical electrical operation sequence; By analyzing the autocorrelation coefficients and partial autocorrelation coefficients of different time lag terms in the historical electrical operation sequence, the dynamic change pattern of the data is captured, and the equipment state evolution trajectory is obtained. Feature components matching the dimension of the association mapping matrix are extracted from the equipment state evolution trajectory, projected onto the feature space of the association mapping matrix, and the magnitude of the projection result is calculated as the load fluctuation benchmark. If the load fluctuation benchmark is greater than the preset load fluctuation threshold, the initial reconstructed feature sequence is calculated and derived. The feature vectors corresponding to the initial reconstructed feature sequence are projected onto the orthogonal complement space of the association mapping matrix, and the feature dimensions corresponding to the elements whose absolute values ​​exceed a preset projection threshold in the projected vectors are extracted to obtain the electrical features of the metering box.

5. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 1, characterized in that, Also includes: Extract the current value sequence from the electrical characteristics of the metering box, and calculate the waveform distortion index from the current value sequence; The filtering parameters are adjusted according to the waveform distortion index, and the current value sequence is filtered and compensated to obtain a corrected current value sequence. The peak value of the power frequency component in the corrected current value sequence is extracted as the current fundamental amplitude. If the amplitude of the current fundamental wave does not exceed the preset current verification threshold, then the corresponding feature dimension is determined to be physically consistent. All feature dimensions that conform to the physical consistency are retained and combined to obtain effective electrical feature components.

6. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 1, characterized in that, Also includes: The current fluctuation values ​​of the target metering box and the associated metering boxes adjacent to it in the physical space are obtained at continuous time sampling points before and after the anomaly occurs. A multi-dimensional spatiotemporal matrix is ​​constructed with spatial nodes as rows and time sampling points as columns. Singular value decomposition is performed on the multidimensional spatiotemporal matrix, and the time coefficient sequence corresponding to the first principal component is extracted as the spatial mapping feature sequence. The high-dimensional spatiotemporal data is projected onto the low-dimensional feature space, and the spatial mapping feature sequence is expanded in phase space to obtain the temporal reconstruction sequence. If the fluctuation amplitude of the time-series reconstructed sequence exceeds a preset fluctuation amplitude threshold, then a nonlinear data transformation is performed on the time-series reconstructed sequence to obtain the transformed feature sequence; The numerical values ​​of each dimension of the transformed feature sequence are linearly scaled to map all dimensions to a unified numerical range to complete dimension normalization, thus obtaining standardized electrical features.

7. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 1, characterized in that, The obtained fault panorama features include: Extract the high-frequency transient voltage sequence of the standardized electrical characteristics and the box surface temperature sequence from the box characteristic sequence, and obtain a synchronization state snapshot based on the high-frequency transient voltage sequence and the box surface temperature sequence; Based on the synchronous state snapshot, transient stress data and insulation breakdown data are obtained. The correlation coefficient between the transient stress data and the insulation breakdown data is calculated to obtain a cross-correlation sequence. The variance of the cross-correlation sequence is calculated as the deviation magnitude. If the deviation magnitude is greater than a preset deviation magnitude threshold, the information entropy of the transient stress data and the insulation breakdown data are calculated respectively. Based on the information entropy, a fusion weight is assigned, and the fusion weight is linearly weighted with the transient stress and the insulation breakdown data to obtain a weighted fusion feature. The fault panorama features are obtained based on the weighted fusion features.

8. The method for monitoring the status of an energy metering box based on multi-sensor fusion according to claim 7, characterized in that, The method of identifying the fault type of the metering box and locating the time of fault occurrence based on the panoramic fault features includes: The multidimensional feature matrix corresponding to the fault panoramic feature is compared element by element with each standard feature matrix in the preset benchmark sample library to calculate the multidimensional similarity matrix. Traverse the multidimensional similarity matrix, extract the global maximum value as the extreme value feature, and classify and determine the initial fault type label based on the sample index corresponding to the extreme value feature. By combining the initial fault type label, the high-frequency transient voltage sequence corresponding to the synchronization state snapshot is extracted to determine the time when the internal short circuit occurs; Using the initial fault type label and the time of occurrence of the internal short circuit as a joint index, the preset state mapping table is queried to obtain the final state monitoring result and output it.

9. A status monitoring system for an electricity metering box based on multi-sensor fusion, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the power metering box status monitoring method based on multi-sensor fusion as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the power metering box status monitoring method based on multi-sensor fusion as described in any one of claims 1 to 8.

Images