Low-voltage switch abnormal breaking detection method based on sparse coding arc identification

Abstract

The invention discloses a low-voltage switch abnormal breaking detection method based on sparse coding arc identification. In the training stage, low-sampling current and high-frequency reference signals of the same breaking event are synchronously collected, a breaking action window is positioned and aligned, a multi-scale low-sampling dictionary and a multi-scale high-frequency dictionary which meet shared sparse coefficient constraints are established, and a reconstruction residual threshold value and a high-frequency compensation amount threshold value are obtained. In the detection stage, only low-sampling current is collected, an action window is intercepted and aligned, a shared sparse coefficient is solved on a low-sampling dictionary, a low-sampling reconstruction signal and a high-frequency compensation signal are generated, a residual index and a compensation amount index are calculated, and dual-threshold joint judgment is carried out to realize abnormal breaking detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault diagnosis of low-voltage switchgear, and in particular to a low-voltage switch abnormal breaking detection method based on sparse coding arc recognition. BACKGROUND

[0002] During the breaking of load current, low-voltage switchgear such as circuit breakers and contactors may produce persistent or repeated arc phenomena due to factors such as contact ablation, contact rebound, mechanism wear, and load fluctuation, which may cause abnormal breaking, contact damage, temperature rise, and safety hazards. Therefore, online detection and alarm of arc and abnormal state during the breaking process have become an important research direction and engineering requirement in the state monitoring and operation of low-voltage distribution equipment.

[0003] In the prior art, abnormal breaking detection is mostly based on feature extraction and discrimination of electrical quantity signals. For example, high-frequency components, harmonic content, and time-frequency energy distribution features of current signals or voltage signals are used to construct threshold criteria using time-frequency analysis methods such as short-time Fourier transform and wavelet transform; or further combined with template matching, statistical learning methods, and neural network models for classification and recognition. To enhance the ability to capture arc transients, some schemes increase the sampling frequency and expand the front-end analog bandwidth, or introduce voltage sensors, radio frequency sensors, acoustic sensors, and optical sensors for multi-source information fusion judgment, thereby improving detection reliability.

[0004] However, the above prior art still has deficiencies in field deployment, mainly in the following aspects. First, many schemes rely on high sampling frequency and wide bandwidth acquisition links, or rely on multi-sensor configuration; in low-cost, single-current-sensor, limited sampling frequency, and limited front-end bandwidth application scenarios, arc sparse transient information is easily weakened or even lost, leading to decreased detection accuracy. Second, the discrimination method based on fixed time-frequency feature threshold is sensitive to load type, operating conditions, and noise interference, and is prone to false positives or false negatives, and has insufficient adaptability across devices and operating conditions. Third, the breaking action occurrence time has jitter, and the window start point and alignment error between different breaking events affect feature statistics and model matching, thereby reducing algorithm robustness and repeatability.

[0005] Therefore, a detection method is needed that can effectively separate and characterize arc transient information under the conditions of single-current sensing, low sampling frequency, and limited bandwidth, and achieve reliable abnormal breaking judgment. SUMMARY

[0006] One objective of this invention is to propose a low-voltage switchgear abnormal disconnection detection method based on sparse coding arc identification. Addressing the limitations of existing technologies in effectively acquiring sparse transient arc information under conditions of single current sensing, low sampling frequency, and limited front-end bandwidth, and the reliance on fixed time-frequency characteristic thresholds which are susceptible to false alarms and missed alarms due to changes in operating conditions and jitter at the moment of action, this invention proposes a technical solution based on disconnection action window truncation and alignment, multi-scale coupled dictionary sparse representation, and joint determination of dual-index thresholds. Specifically, during the training phase, low-sampling current training signals and high-frequency reference training signals are acquired simultaneously to establish a multi-scale low-sampling dictionary and a multi-scale high-frequency dictionary that satisfy shared sparse coefficient constraints, and to determine the reconstruction residual threshold and the high-frequency compensation threshold. During the detection phase, the low-sampling current detection signal is located and the disconnection action window is aligned. The shared sparse coefficients are solved on the multi-scale low-sampling dictionary to generate a low-sampling reconstruction signal and a high-frequency compensation signal. The reconstruction residual index and the high-frequency compensation index are calculated and jointly determined. This invention achieves the technical effects of enhancing arc transient characterization and improving the accuracy and robustness of abnormal disconnection detection under limited sampling information.

[0007] This invention provides a method for detecting abnormal disconnection of low-voltage switches based on sparse coded arc identification, comprising: S1. Acquire the low-sampling current training signal and the corresponding high-frequency reference training signal synchronously obtained for the same interruption event. Determine the interruption action window based on the low-sampling current training signal and capture the interruption action window training signal. Perform interruption action window alignment processing on the interruption action window training signal according to the preset alignment benchmark. Establish a multi-scale low-sampling dictionary based on the aligned interruption action window training signal and a multi-scale high-frequency dictionary based on the high-frequency reference training signal. Ensure that the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary satisfy the shared sparse coefficient constraint to form a coupling relationship. Obtain training samples for the reconstructed residual index and training samples for the high-frequency compensation index based on the training samples, and determine the first and second judgment thresholds. S2. Acquire the low-sampling current detection signal of the low-voltage switch to be detected. S3. Determine the interruption action based on the low-sampling current detection signal. S4. The break action window detection signal is captured by the window; S5. The break action window detection signal is aligned according to the preset alignment benchmark to obtain the aligned break action window detection signal; S6. The aligned break action window detection signal is sparsely encoded and solved on the multi-scale low-sampling dictionary to obtain shared sparse coefficients. Based on the shared sparse coefficients, a low-sampling reconstruction signal is generated on the multi-scale low-sampling dictionary, and a high-frequency compensation signal is generated on the multi-scale high-frequency dictionary; S7. The reconstruction residual index is calculated based on the aligned break action window detection signal and the low-sampling reconstruction signal, and the high-frequency compensation amount index is calculated based on the high-frequency compensation signal; S8. When the reconstruction residual index is not less than the first judgment threshold and the high-frequency compensation amount index is not less than the second judgment threshold, it is determined that the low-voltage switch has abnormally broken; otherwise, it is determined that the low-voltage switch has not abnormally broken.

[0008] Optionally, S1 includes: For the same interruption event, low-sampling current training signal and high-frequency reference training signal are simultaneously acquired; The current change characteristics are calculated based on the low-sampling current training signal, and the break action window is determined when the current change characteristics meet the preset window triggering conditions. The break action window training signal is extracted from the low-sampling current training signal. Using the extreme moment of the current change characteristics within the training signal of the interrupted action window as a preset alignment reference, the training signal of the interrupted action window is subjected to interrupted action window alignment processing to obtain the aligned training signal of the interrupted action window. The aligned segmented action window training signal is decomposed into a multi-scale component set including slow-changing components and peak components. The multi-scale component set is used as training samples to establish a multi-scale low-sampling dictionary, and the high-frequency reference training signal is used as training samples to establish a multi-scale high-frequency dictionary. By jointly optimizing and simultaneously updating the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary and solving for the shared sparse coefficients for training, the shared sparse coefficients for training of the aligned segmented action window training signal on the multi-scale low-sampling dictionary are kept consistent with the shared sparse coefficients for training of the high-frequency reference training signal on the multi-scale high-frequency dictionary. This applies a shared sparse coefficient constraint and couples the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary. Based on the aligned segmentation action window training signal, the multi-scale low-sampling dictionary, and the multi-scale high-frequency dictionary, a low-sampling reconstruction signal and a high-frequency compensation signal for training are generated. Training samples for the reconstruction residual index and training samples for the high-frequency compensation index are calculated. A first judgment threshold and a second judgment threshold are determined based on the training samples for the reconstruction residual index and the training samples for the high-frequency compensation index, respectively.

[0009] Optionally, S2 includes: A single current sensor is installed on the main circuit conductor of the low-voltage switch to be tested to collect an analog current signal that is proportional to the main circuit current. The analog current signal is subjected to analog filtering so that the bandwidth of the filtered signal is less than or equal to half of the preset sampling frequency; The filtered analog current signal is converted from analog to digital at the preset sampling frequency to obtain a discrete current sample value sequence arranged in time order. The discrete current sample value sequence constitutes the low-sampling current detection signal.

[0010] Optionally, S3 includes: The current change characteristics are calculated for the low-sampling current detection signal, and the current change characteristics include the absolute value of the current difference between adjacent sampling points; When the current change characteristics meet the preset window triggering conditions, the starting point of the disconnection action window is determined, and the ending point of the disconnection action window is determined by the preset window duration. The disconnection action window detection signal is extracted from the low-sampling current detection signal according to the start and end points of the disconnection action window.

[0011] Optionally, S4 includes: The extreme value moment of the current change characteristic is determined from the detection signal of the disconnection action window as the alignment moment; The time difference between the alignment time and the preset alignment reference is used as the shift amount to perform time shift on the segmentation action window detection signal, and the missing part exceeding the segmentation action window after time shift is padded with zeros to obtain the aligned segmentation action window detection signal.

[0012] Optionally, S5 includes: The aligned segmentation action window detection signal is input into the multi-scale low-sampling dictionary, and the shared sparse coefficients are solved under the shared sparse coefficient constraint. The solution includes minimizing the error between the aligned segmentation action window detection signal and the estimated signal obtained by multiplying the multi-scale low-sampling dictionary by the shared sparse coefficients and applying sparsity constraints to the shared sparse coefficients. The shared sparse coefficients are input into the multi-scale low-sampling dictionary to generate a low-sampling reconstruction signal, and the shared sparse coefficients are input into the multi-scale high-frequency dictionary to generate a high-frequency compensation signal, wherein the low-sampling reconstruction signal is the product of the multi-scale low-sampling dictionary and the shared sparse coefficients, and the high-frequency compensation signal is the product of the multi-scale high-frequency dictionary and the shared sparse coefficients.

[0013] Optionally, S6 includes: The aligned segmentation action window detection signal is subtracted point by point from the low-sampled reconstructed signal to obtain the residual signal, and the L2 norm of the residual signal is calculated as the reconstruction residual index. The high-frequency compensation signal is squared point by point, and the squared results are summed to obtain the high-frequency energy value, which is then used as the high-frequency compensation quantity index.

[0014] Optionally, the S7 includes: When the reconstructed residual index is not less than the first judgment threshold, a first judgment result is generated as abnormal; when the reconstructed residual index is less than the first judgment threshold, a first judgment result is generated as normal. When the high-frequency compensation amount index is not less than the second determination threshold, a second determination result is generated as abnormal; when the high-frequency compensation amount index is less than the second determination threshold, a second determination result is generated as normal. When both the first and second determination results are abnormal, it is determined that the low-voltage switch has experienced an abnormal disconnection; otherwise, it is determined that the low-voltage switch has not experienced an abnormal disconnection.

[0015] Optionally, the step of simultaneously updating the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary through joint optimization and solving for the shared sparse coefficients used in training includes: The aligned segmented action window training signal is used as the low-sampled training signal. The corresponding high-frequency reference training signal is used as the high-frequency training signal. Using a multi-scale low-sampling dictionary as Using a multi-scale high-frequency dictionary as Construct a joint optimization problem: minimizing and Reconstruction error and and The weighted sum of reconstruction errors is used as the objective, and the sparse coefficients are... Apply sparsity constraints; Wherein, the shared sparse coefficient constraint is: for the same segmentation event, the At the same time as exist sparse representation coefficients on and exist Sparse representation coefficients on; The joint optimization problem is solved using an alternating iterative approach, which includes: [the following steps are taken when a fixed number of nodes are in use]. and Time-based solution and in fixed Updated separately and ; And the updated and Normalization constraints are applied to the dictionary atoms.

[0016] Optionally, the step of performing multi-scale decomposition on the aligned segmented action window training signal to obtain a multi-scale component set including slowly changing components and spike components includes: Perform a preset number of decomposition layers on the aligned segmentation action window training signal. The discrete wavelet decomposition yields the th Layer approximation components and layers 1 to 1 Layer detail components; The first The layer approximation component is determined as the slowly changing component, and the first layer to the second layer are... At least one layer of detail components is determined as the peak component, and the slow-changing component and the peak component together constitute the multi-scale component set. Based on the multi-scale component set, low-sampling sub-dictionaries of the corresponding scales are trained respectively, and the low-sampling sub-dictionaries of each scale are concatenated column by column to form the multi-scale low-sampling dictionary.

[0017] The beneficial effects of this invention are: 1. Under the conditions of single current sensor, low sampling frequency and limited front-end bandwidth, high-frequency compensation information can be inferred from low-sampling current signal by coupling modeling of multi-scale low-sampling dictionary and multi-scale high-frequency dictionary and applying shared sparse coefficient constraints. This enhances the ability to characterize the sparse transients of electric arc and improves the availability and accuracy of abnormal breakage detection.

[0018] 2. By automatically truncating the break action window and aligning the window with the extreme moments of current change characteristics, the impact of timing jitter of different break events on sparse coding solution and index calculation is reduced, thereby improving the consistency and robustness of the detection results.

[0019] 3. A dual-threshold joint judgment mechanism of reconstruction residual index and high-frequency compensation index is adopted to ensure that the anomaly judgment simultaneously satisfies two types of evidence: low-sampling domain reconstruction mismatch and high-frequency compensation enhancement, thereby reducing the risk of false alarms and false negatives caused by relying solely on a single time-frequency feature or a single threshold. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the training phase of a low-voltage switch abnormal disconnection detection method based on sparse coding arc identification proposed in this invention. Figure 2 This is a flowchart of the operational phase of a low-voltage switch abnormal disconnection detection method based on sparse coding arc identification proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 and Figure 2 A method for detecting abnormal disconnection of low-voltage switches based on sparse coded arc identification, comprising: S1. Acquire the low-sampling current training signal and the corresponding high-frequency reference training signal synchronously obtained for the same interruption event. Determine the interruption action window based on the low-sampling current training signal and capture the interruption action window training signal. Perform interruption action window alignment processing on the interruption action window training signal according to the preset alignment benchmark. Establish a multi-scale low-sampling dictionary based on the aligned interruption action window training signal and a multi-scale high-frequency dictionary based on the high-frequency reference training signal. Ensure that the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary satisfy the shared sparse coefficient constraint to form a coupling relationship. Obtain training samples for the reconstructed residual index and training samples for the high-frequency compensation index based on the training samples, and determine the first and second judgment thresholds. S2. Acquire the low-sampling current detection signal of the low-voltage switch to be detected. S3. Determine the interruption action based on the low-sampling current detection signal. S4. The break action window detection signal is captured by the window; S5. The break action window detection signal is aligned according to the preset alignment benchmark to obtain the aligned break action window detection signal; S6. The aligned break action window detection signal is sparsely encoded and solved on the multi-scale low-sampling dictionary to obtain shared sparse coefficients. Based on the shared sparse coefficients, a low-sampling reconstruction signal is generated on the multi-scale low-sampling dictionary, and a high-frequency compensation signal is generated on the multi-scale high-frequency dictionary; S7. The reconstruction residual index is calculated based on the aligned break action window detection signal and the low-sampling reconstruction signal, and the high-frequency compensation amount index is calculated based on the high-frequency compensation signal; S8. When the reconstruction residual index is not less than the first judgment threshold and the high-frequency compensation amount index is not less than the second judgment threshold, it is determined that the low-voltage switch has abnormally broken; otherwise, it is determined that the low-voltage switch has not abnormally broken.

[0023] In this specific embodiment, S1 includes: Two synchronous acquisition channels are set up for the same interruption event, and sampling is initiated by the same hardware trigger signal. The sampling frequency of the low sampling current training signal is set to... The sampling frequency of the high-frequency reference training signal is set to The low-sampled current training signal of a single interruption event is denoted as... ,in For low-sample discrete time index and The unit is A, and the corresponding high-frequency reference training signal is denoted as ,in For high-frequency discrete time index and The unit is A; based on Calculate the characteristics of current variation ,in Represents the absolute value of the current difference between adjacent sampling points, with units of A. continuous Each sampling point satisfies Determine the trigger time index ,in , The rated current of the low-voltage switch under test is expressed in amperes (A). by Backtracking Each sampling point determines the starting point of the segmentation action window. and based on window length Determine the endpoint of the break action window ,from Intercept the training signal of the segmentation action window ,in Arranged in chronological order A column vector consisting of low-sampled current values; Calculate within this window Extreme moment index Set the preset alignment reference to a fixed index within the window. Calculate the shift amount And on Perform time shift alignment processing, with the time shift performed as a "global translation". The rule of "sampling points and padding with zeros for out-of-bounds portions" is implemented to obtain the aligned segmented action window training signal. ; The high-frequency reference training signal is truncated based on the start and end times of the low-sampling window. ,in and will Converted to high-frequency shift amount Later Perform the same "overall translation" The alignment process, which involves "sampling points and padding with zeros for out-of-bounds portions", yields... ; right The number of decomposition layers to be executed is Discrete wavelet decomposition, using wavelet basis functions set as , obtained the Layer approximate components and the first floor to the second floor Layer detail components Define the slowly changing component as Define the peak component as Three layers of detail, and for The inverse wavelet reconstruction, which retains only the component of this layer and sets the rest to zero, is performed to obtain four scale signals of the same length as the segmentation action window. The four scale signals constitute a multi-scale component set and are used to train the low-sampling sub-dictionary. Establish training sample set ,in Number the break event and The number of normal segmentation events was [number]. And the number of abnormal segmentation event samples is For each event The aligned low-sampling window is obtained through the above process. Aligned high-frequency window ; The low-sampling domain is trained with low-sampling sub-dictionaries at four different scales, and then concatenated column-wise to form a multi-scale low-sampling dictionary. The number of atoms in the sub-dictionary corresponding to the slowly changing component is set to In the peak component The number of atoms in the sub-dictionaries corresponding to the three levels are respectively set as follows: Total number of atoms Furthermore, each sub-dictionary is trained using... The sparsity upper limit and with The iteration is the number of iterations for dictionary learning; The same wavelet basis is used in the high-frequency domain as in the low-sampling domain. With the number of decomposition layers right Perform decomposition and equal-length reconstruction, and configure the atomic number consistent with the low-sampling domain. Training yields a multi-scale high-frequency dictionary This ensures a low-sampling dictionary High-frequency dictionary They correspond one-to-one in the column direction; During the coupled training phase, a shared sparse coefficient constraint is imposed on all training samples, and updates are performed simultaneously through joint optimization. and And solve for the shared sparse coefficients used in training. The joint optimization problem can be written as: ; in Indicates the first Alignment of the segmentation action window with low-sampled training signal for each segmentation event Indicates the first The high-frequency reference training signal of the aligned segmentation action window for each segmentation event. This represents a multi-scale low-sampling dictionary whose column vectors are low-sampling dictionary atoms. This represents a multi-scale high-frequency dictionary whose column vectors are high-frequency dictionary atoms. Indicates the first The shared sparse coefficient vector of each discontinuity event and simultaneously serves as... exist sparse representation coefficients on and exist sparse representation coefficients on, This represents the weighting coefficient of the high-frequency domain reconstruction error in the objective function. This represents the sparsity constraint strength coefficient. This represents the L2 norm and is used to measure reconstruction error. Denotes a norm and is used to promote Sparsely distributed, Indicates the total number of training samples; The above joint optimization problem is solved using an alternating iterative approach, with a fixed... and The FISTA algorithm is used to solve each Furthermore, the number of iterations within FISTA is set to 100, and the relative change in the objective function between two adjacent iterations is less than [a certain value]. As a condition for early termination, in all fixed Time and The dictionary atoms are updated column-wise, and joint normalization is performed on the corresponding coupled atom pairs in each column. The joint normalization is performed by "combining low-sampled atoms with high-frequency atoms". After stacking, normalize the whole and then split it back. and The rule implementation, in which Indexed by dictionary column and and They are respectively and The Atoms; After completing the coupled dictionary training, the low-sampled reconstructed signal for training is calculated for each training sample. High-frequency compensation signal for training ,pass and The residual signal is obtained by point-by-point subtraction, and the reconstructed residual index is obtained by taking the L2 norm of the residual signal. ,right The high-frequency compensation index is obtained by squaring and summing the points. ; normal sample set Maximum value and the set of outliers The arithmetic mean of the minimum values ​​is determined as the first judgment threshold. , the normal sample set Maximum value and the set of outliers The arithmetic mean of the minimum values ​​is used as the second judgment threshold. This completes the multi-scale low-sampling dictionary in step S1. Multi-scale high-frequency dictionary Coupling modeling and the first decision threshold With the second judgment threshold The determination.

[0024] In this specific embodiment, S2 includes: A single current sensor is fitted around the outer periphery of the main circuit conductor of the low-voltage switch to be tested. The current sensor uses an open-type current transformer and converts the secondary current into an analog voltage signal through a load resistor. The analog voltage signal is denoted as... ,in For continuous time and The unit is V; Will The input signal conditioning circuit includes a bias circuit, an amplifier circuit, and an analog filter circuit. The bias circuit sets the signal reference level to... To match the single-supply input range of the analog-to-digital converter, the voltage gain of the amplifier circuit is set to... To ensure that the output voltage corresponding to the full-scale current does not exceed The analog filter circuit adopts a second-order Butterworth low-pass structure and the cutoff frequency is set to... ,in satisfy and The preset sampling frequency ensures that the bandwidth of the filtered signal does not exceed the Nyquist frequency and suppresses aliasing. The filtered analog voltage signal is input to an analog-to-digital converter for sampling. The analog-to-digital converter uses... The sampling period is a successive approximation structure. ,in The sampling time interval is in seconds, and the discrete voltage sampling sequence is output sequentially in time-triggered mode. ,in For discrete sampling point index and The unit is V; In the digital domain Perform current calibration conversion to obtain low-sampling current detection signal The conversion relationship is as follows: ; in These are discrete current samples arranged in chronological order, with units of Ampere (A). This is the bias voltage of the signal conditioning circuit, and its unit is V. The total scaling factor from the main circuit current to the analog-to-digital converter input voltage, in units of... The The current transformer ratio, load resistance value, and amplifier gain are all factors influencing the current transformer ratio. The values ​​are jointly determined and calibrated to a constant at the rated current point when the device leaves the factory; All sampling points The discrete current sample values ​​are arranged in chronological order to form a sequence.

[0025] In this specific embodiment, S3 includes: Low sampling current detection signal Perform point-by-point processing and calculate current variation characteristics ,in Indicates the first The discrete current values ​​at each sampling point are expressed in units of... The index of the discrete sampling point is a positive integer, and the current variation characteristic is calculated using the following formula: ; in Indicates the first The current variation characteristics corresponding to each sampling point are given in A. This indicates the absolute value operation. Indicates the first The discrete current values ​​at each sampling point are expressed in amperes (A). Set the preset window trigger condition to continuous. Each sampling point satisfies When a breakout action is detected, the index of the sampling point that first satisfies the continuity condition is determined as the trigger time index. ,in The rated current of the low-voltage switch to be tested is expressed in amperes (A). To ensure the tripping action window includes current change information before triggering, the number of pre-set points for the window is set to [value]. Set the window length corresponding to the preset window duration to . Each sampling point, and in the implementation... Maintenance at least A circular cache of length and only when the cache has been written to consecutively. Trigger determination is allowed after each sampling point, thus ensuring that the trigger time index satisfies the condition. ; The starting point of the break action window is determined as... and based on window length Determine the endpoint of the break action window as ,in The index of the window's starting point is a positive integer. The index of the window endpoint is a positive integer; According to the starting point and the end point from Intercepting the detection signal of the break action window ,in For length is The column vector and its first Each element is defined as The index of the sample within the window and its value ranges from 1 to... .

[0026] In this specific embodiment, S4 includes: The break action window detection signal is denoted as ,in The index of the sampling point within the window, with a value range from 1 to... The length of the segmentation action window is expressed in units of sampling points. For the first The discrete current values ​​at each sampling point are expressed in amperes (A). Recalculate the current variation characteristics within this window ,in Defined as the absolute value of the current difference between adjacent sampling points, with units of Ampere (A), and in to Search within the range The location of the maximum value is determined, and the index at which the maximum value is obtained is used as the alignment time. ,in It represents the moment within the window where the current change is most drastic and is used to characterize the critical transient position of the disconnection action; Set the preset alignment reference to a fixed index. ,in The number of points in front of the window and Calculate the shift amount at the expected trigger position within the corresponding window. ,in The integer is represented by the number of sampling points. This indicates that the entire signal has shifted backward. This indicates that the entire signal has shifted forward. According to the shift amount right The execution time shift yields the aligned segmentation action window detection signal. Time shifting press " Overall translation Sample points are collected, and the window length is maintained at [number]. The rule of "unchanged" is implemented specifically when an index shift occurs. or When Set to 0 to complete the zero-padding process, which satisfies the condition after index shift. season ,in The unit is A and it is still a length. Window signals; Will As input to step S5 and in accordance with the above As a preset alignment benchmark consistent with the training phase, the key transient positions of different segmentation events are aligned on a unified time benchmark.

[0027] In this specific embodiment, S5 includes: The aligned segmentation action window detection signal is denoted as... ,in This indicates the length of the segmentation action window, expressed in sample points. It is a column vector of low-sample discrete current values ​​arranged in chronological order, with units of A; Read the multi-scale low-sampling dictionary obtained from training in step S1 from memory. Multi-scale high-frequency dictionary Where M=160 represents the total number of atoms in the dictionary and is a positive integer. This represents the length of the high-frequency window, expressed in units of sampling points. Each column is a low-sampled dictionary atom and the column vector is normalized to the first norm. Each column is with The high-frequency dictionary atoms corresponding to the same column index coupling are normalized to 1 for the 2-norm of the column vector; Under the constraint of shared sparse coefficients, Solving sparse coding with shared sparse coefficients ,in At the same time as exist The sparse representation coefficients on and the subsequent high-frequency compensation signal in The generation coefficients on the above, the solution is achieved by minimizing the reconstruction error and imposing sparsity constraints, and is written as: ; in Let be the sparse coefficient vector to be determined. Let the variable be the one that minimizes the objective function. It represents the L2 norm and is used to measure vector reconstruction error. Denotes the norm and is used to promote sparse coefficients. This represents the sparsity constraint strength coefficient and is a fixed parameter shared by both the training and detection phases; The FISTA algorithm is used to iteratively solve the above problem, with the iteration initialization set to... and express A zero-dimensional vector, with a maximum number of iterations set to The step size uses the Lipschitz constant of the low-sampling dictionary, which is pre-computed and stored during the training phase. The reciprocal of and Defined as a matrix The maximum eigenvalue is obtained by offline calculation during the training phase, where This represents the transpose operation, and the convergence criterion is set to the relative change of the coefficient vector between two adjacent iterations being less than 1. Termination in advance; After obtaining the shared sparsity coefficient The low-sampled reconstructed signal is then generated. With high-frequency compensation signal ,in Through matrix multiplication The value is obtained and the unit is A. Through matrix multiplication Obtain and the unit is A, and and The results are output to step S6 for the calculation of the reconstructed residual index and the high-frequency compensation index.

[0028] In this specific embodiment, S6 includes: The aligned segmentation action window detection signal is denoted as... ,in The index of the sampling point within the window, with a value range from 1 to... For low-sampling segmentation action window length and in units of sampling points, For the first The discrete current values ​​at each sampling point are expressed in amperes (A). Let the low-sampled reconstructed signal be denoted as ,in For the first The low-sampling reconstructed current value of each sampling point is given in A, and the high-frequency compensation signal is denoted as... ,in This is the index of the high-frequency sampling point, and its value ranges from 1 to... This is the length of the high-frequency window, expressed in units of sampling points. For the first The high-frequency compensation current value at each sampling point is expressed in A. right and The residual signal is obtained by performing point-by-point subtraction. ,in For the first The reconstructed residual current values ​​at each sampling point are in A, and are based on the same interruption action window. Calculate the reconstruction residual index At the same time, based on Calculate the high-frequency compensation index The two indicators are calculated using the following formula: ; in Indicates that it is composed of all The residual column vector is composed in chronological order. This represents the L2 norm, which is equal to the square root of the sum of the squares of the elements of the vector and is used to measure the degree of reconstruction mismatch in the low-sampling domain. To reconstruct the residual index and the unit is This represents the summation operation on all sampling points within the high-frequency window. This is a high-frequency compensation quantity indicator, and the unit is... And it is used to measure the intensity of high-frequency compensation generated by shared sparse coefficients; The above pointwise subtraction, squaring, and summation operations are performed under the same data type and using floating-point accumulation to avoid fixed-point overflow. and The stable output is sent to step S7 for joint determination of the two thresholds.

[0029] In this specific embodiment, S7 includes: Read the reconstructed residual metrics from step S6. High-frequency compensation index ,in This represents the norm 2 error between the aligned segmentation action window detection signal and the low-sampled reconstructed signal, expressed in units of 1. This represents the sum of squares of the energy of the high-frequency compensated signal within the high-frequency window, with units of . ; Read the first judgment threshold determined in the offline training step S1 from the non-volatile memory. With the second judgment threshold ,in To reconstruct the threshold of the residual index and the unit is The threshold value for the high-frequency compensation quantity is expressed in units of 1. and will and Set as fixed parameters that remain unchanged during equipment operation; Generate the first judgment result Compared with the second judgment result ,in This indicates a binary judgment based on the reconstructed residual index, and Indicates an anomaly, This indicates that it is normal. This indicates a binary determination based on the high-frequency compensation quantity index, and Indicates an anomaly, This indicates normal operation, and the determination rule is executed according to the following formula: ; in The reconstructed residual index is calculated in real time. To reconstruct the residual threshold, This refers to the high-frequency compensation quantity index obtained through real-time calculation. The high-frequency compensation threshold, symbol and These represent comparisons of "greater than or equal to" and "less than", respectively. If and only if and Output the abnormal segmentation judgment result in real time. ,in and This indicates that the low-voltage switch has tripped abnormally. This indicates that the low-voltage switch has not experienced an abnormal disconnection. or Time output and will The timestamp of this interruption event is written to the event log area to support subsequent tracing. It outputs alarm flags to the host computer or protection unit in a timely manner to complete the linkage processing of abnormal disconnection.

[0030] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0031] This invention addresses the problem that only a single current can be acquired on-site, and the sampling frequency is low, while the front-end bandwidth is limited, making it difficult to directly observe sparse transient information of the arc. It combines "super-resolution reconstruction of sparse representation using a coupled dictionary" with "reconstruction residual discrimination": During the training phase, a multi-scale low-sampling dictionary and a multi-scale high-frequency dictionary with shared sparse coefficient constraints are established using the low-sampling current training signal and the high-frequency reference training signal of the same breaking event. This allows the shared sparse coefficients to be obtained from the low-sampling current during the detection phase, resulting in a low-sampling reconstruction signal from the low-sampling dictionary and a high-frequency compensation signal from the high-frequency dictionary. Subsequently, the reconstruction residual index and the high-frequency compensation index in the low-sampling domain are calculated simultaneously, and anomalous breaking is determined using a dual-threshold joint method. This achieves separation and enhancement of arc transients with limited sampling information, improving the accuracy and anti-interference capability of anomalous breaking detection.

[0032] This invention incorporates practical improvements: First, it introduces window truncation for the interruption action and aligns it with the window based on the extreme moments of current change characteristics, ensuring consistency in the time reference for different interruption events and reducing the impact of action moment jitter on sparse coding and threshold determination. Second, it employs multi-scale decomposition to construct a multi-scale low-sampling dictionary, modeling slow-changing components and peak components separately to improve the representation of sparse transients of the arc. Furthermore, it establishes a coupling relationship with the multi-scale high-frequency dictionary through joint optimization, making high-frequency compensation more closely aligned with arc characteristics. Third, it uses the high-frequency compensation amount as an independent criterion, jointly constraining anomaly determination with the reconstructed residual, avoiding the problem of false alarms and missed alarms easily caused by relying solely on a single time domain or time-frequency threshold under load changes and noise conditions, thus further ensuring detection robustness from a structural perspective.

Claims

1. A method for detecting abnormal disconnection of low-voltage switches based on sparse coded arc identification, characterized in that, include: S1. Obtain the low-sampling current training signal and the corresponding high-frequency reference training signal synchronously obtained for the same interruption event. Determine the interruption action window based on the low-sampling current training signal and intercept the interruption action window training signal. Perform interruption action window alignment processing on the interruption action window training signal according to the preset alignment benchmark. Establish a multi-scale low-sampling dictionary based on the aligned interruption action window training signal. Establish a multi-scale high-frequency dictionary based on the high-frequency reference training signal. Make the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary satisfy the shared sparse coefficient constraint to form a coupling relationship. Obtain the training samples of the reconstructed residual index and the training samples of the high-frequency compensation index based on the training samples. Determine the first judgment threshold and the second judgment threshold. S2. Obtain the low-sampling current detection signal of the low-voltage switch to be tested; S3. Determine the tripping action window based on the low-sampling current detection signal and intercept the tripping action window detection signal; S4. Perform break action window alignment processing on the break action window detection signal according to the preset alignment benchmark to obtain the aligned break action window detection signal; S5. Sparsely encode and solve the aligned break action window detection signal on the multi-scale low-sampling dictionary to obtain shared sparse coefficients. Based on the shared sparse coefficients, generate a low-sampling reconstruction signal on the multi-scale low-sampling dictionary and a high-frequency compensation signal on the multi-scale high-frequency dictionary. S6. Calculate the reconstruction residual index based on the aligned break action window detection signal and the low-sampling reconstruction signal, and calculate the high-frequency compensation amount index based on the high-frequency compensation signal. S7. When the reconstruction residual index is not less than the first judgment threshold and the high-frequency compensation amount index is not less than the second judgment threshold, determine that the low-voltage switch has experienced abnormal breakage; otherwise, determine that the low-voltage switch has not experienced abnormal breakage.

2. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S1 includes: For the same interruption event, low-sampling current training signal and high-frequency reference training signal are simultaneously acquired; The current change characteristics are calculated based on the low-sampling current training signal, and the break action window is determined when the current change characteristics meet the preset window triggering conditions. The break action window training signal is extracted from the low-sampling current training signal. Using the extreme moment of the current change characteristics within the training signal of the interrupted action window as a preset alignment reference, the training signal of the interrupted action window is subjected to interrupted action window alignment processing to obtain the aligned training signal of the interrupted action window. The aligned segmented action window training signal is decomposed into a multi-scale component set including slow-changing components and peak components. The multi-scale component set is used as training samples to establish a multi-scale low-sampling dictionary, and the high-frequency reference training signal is used as training samples to establish a multi-scale high-frequency dictionary. By jointly optimizing and simultaneously updating the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary and solving for the shared sparse coefficients for training, the shared sparse coefficients for training of the aligned segmented action window training signal on the multi-scale low-sampling dictionary are kept consistent with the shared sparse coefficients for training of the high-frequency reference training signal on the multi-scale high-frequency dictionary. This applies a shared sparse coefficient constraint and couples the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary. Based on the aligned segmentation action window training signal, the multi-scale low-sampling dictionary, and the multi-scale high-frequency dictionary, a low-sampling reconstruction signal and a high-frequency compensation signal for training are generated. Training samples for the reconstruction residual index and training samples for the high-frequency compensation index are calculated. A first judgment threshold and a second judgment threshold are determined based on the training samples for the reconstruction residual index and the training samples for the high-frequency compensation index, respectively.

3. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S2 include: A single current sensor is installed on the main circuit conductor of the low-voltage switch to be tested to collect an analog current signal that is proportional to the main circuit current. The analog current signal is subjected to analog filtering so that the bandwidth of the filtered signal is less than or equal to half of the preset sampling frequency; The filtered analog current signal is converted from analog to digital at the preset sampling frequency to obtain a discrete current sample value sequence arranged in time order. The discrete current sample value sequence constitutes the low-sampling current detection signal.

4. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S3 includes: The current change characteristics are calculated for the low-sampling current detection signal, and the current change characteristics include the absolute value of the current difference between adjacent sampling points; When the current change characteristics meet the preset window triggering conditions, the starting point of the disconnection action window is determined, and the ending point of the disconnection action window is determined by the preset window duration. The disconnection action window detection signal is extracted from the low-sampling current detection signal according to the start and end points of the disconnection action window.

5. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S4 includes: The extreme value moment of the current change characteristic is determined from the detection signal of the disconnection action window as the alignment moment; The time difference between the alignment time and the preset alignment reference is used as the shift amount to perform time shift on the segmentation action window detection signal, and the missing part exceeding the segmentation action window after time shift is padded with zeros to obtain the aligned segmentation action window detection signal.

6. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S5 include: The aligned segmentation action window detection signal is input into the multi-scale low-sampling dictionary, and the shared sparse coefficients are solved under the shared sparse coefficient constraint. The solution includes minimizing the error between the aligned segmentation action window detection signal and the estimated signal obtained by multiplying the multi-scale low-sampling dictionary by the shared sparse coefficients and applying sparsity constraints to the shared sparse coefficients. The shared sparse coefficients are input into the multi-scale low-sampling dictionary to generate a low-sampling reconstruction signal, and the shared sparse coefficients are input into the multi-scale high-frequency dictionary to generate a high-frequency compensation signal, wherein the low-sampling reconstruction signal is the product of the multi-scale low-sampling dictionary and the shared sparse coefficients, and the high-frequency compensation signal is the product of the multi-scale high-frequency dictionary and the shared sparse coefficients.

7. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S6 include: The aligned segmentation action window detection signal is subtracted point by point from the low-sampled reconstructed signal to obtain the residual signal, and the L2 norm of the residual signal is calculated as the reconstruction residual index. The high-frequency compensation signal is squared point by point, and the squared results are summed to obtain the high-frequency energy value, which is then used as the high-frequency compensation quantity index.

8. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 1, characterized in that, S7 includes: When the reconstructed residual index is not less than the first judgment threshold, a first judgment result is generated as abnormal; when the reconstructed residual index is less than the first judgment threshold, a first judgment result is generated as normal. When the high-frequency compensation amount index is not less than the second determination threshold, a second determination result is generated as abnormal; when the high-frequency compensation amount index is less than the second determination threshold, a second determination result is generated as normal. When both the first and second determination results are abnormal, it is determined that the low-voltage switch has experienced an abnormal disconnection; otherwise, it is determined that the low-voltage switch has not experienced an abnormal disconnection.

9. The method for detecting abnormal disconnection of low-voltage switches based on sparse coded arc identification according to claim 2, characterized in that, The step of simultaneously updating the multi-scale low-sampling dictionary and the multi-scale high-frequency dictionary through joint optimization and solving for the shared sparse coefficients used in training includes: The aligned segmented action window training signal is used as the low-sampled training signal. The corresponding high-frequency reference training signal is used as the high-frequency training signal. Using a multi-scale low-sampling dictionary as Using a multi-scale high-frequency dictionary as Construct a joint optimization problem: minimizing and Reconstruction error and and The weighted sum of reconstruction errors is used as the objective, and the sparse coefficients are... Apply sparsity constraints; Wherein, the shared sparse coefficient constraint is: for the same segmentation event, the At the same time as exist sparse representation coefficients on and exist Sparse representation coefficients on; The joint optimization problem is solved using an alternating iterative approach, wherein the alternating iterative approach includes: [the following steps are taken when a fixed number of nodes are in use]. and Time-based solution and in fixed Updated separately and ; And the updated and Normalization constraints are applied to the dictionary atoms.

10. The low-voltage switch abnormal disconnection detection method based on sparse coded arc identification according to claim 2, characterized in that, The step of performing multi-scale decomposition on the aligned segmented action window training signal to obtain a multi-scale component set including slowly changing components and spike components includes: Perform a preset number of decomposition layers on the aligned segmentation action window training signal. The discrete wavelet decomposition yields the th Layer approximation components and layers 1 to 1 Layer detail components; The first The layer approximation component is determined as the slowly changing component, and the first layer to the second layer are... At least one layer of detail components is determined as the peak component, and the slow-changing component and the peak component together constitute the multi-scale component set. Based on the multi-scale component set, low-sampling sub-dictionaries of the corresponding scales are trained respectively, and the low-sampling sub-dictionaries of each scale are concatenated column by column to form the multi-scale low-sampling dictionary.

Images