A data processing method for quality monitoring of power distribution equipment production line

Abstract

The present invention relates to the field of electrical digital data processing technology, specifically a data processing method for quality monitoring of distribution equipment production lines, comprising the following steps: based on the collected continuous distribution equipment sensor data stream, calculating the time-attenuation dynamic fluctuation index, recording the data points and timestamps whose changes exceed the threshold, and obtaining a time-attenuation sampling point sequence. The present invention implements a dynamic perception and time-attenuation modeling mechanism for the continuous distribution equipment sensor data stream, constructs a time-attenuation dynamic fluctuation index to capture numerical mutations within a short period, and can timely locate key change points where quality anomalies may exist and record timestamps, thereby improving the response sensitivity to data behavior before and after event triggering. By actively intercepting dense raw data segments after anomaly identification and performing multi-level discrete wavelet decomposition of time domain signals, the smoothing trend and multi-scale mutation features are refined and extracted, thereby enhancing the ability to characterize detailed changes in key process details of circuit breakers.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical digital data processing, and in particular to a data processing method for quality monitoring of a power distribution equipment production line. Background Art

[0002] The data processing method for quality monitoring of distribution equipment production lines is an online monitoring and intelligent evaluation technology for distribution equipment manufacturing scenarios. It mainly analyzes sensor data during the operation of key processes of distribution equipment to achieve timely identification and judgment of potential quality problems in the production process.

[0003] Existing technologies primarily analyze the overall trends in distribution equipment sensor data, but lack the ability to independently identify and track specific periods of sudden changes. This results in delayed responses to short-term disturbances or localized quality anomalies. Data processing generally employs full-process averaging or interval statistics, making it difficult to accurately capture the high-frequency energy concentrations that occur in certain processes, and prone to misjudgments or omissions. For example, during the closing process of a switchgear, if the local current fluctuates dramatically but the overall trend remains unchanged, conventional monitoring will not be able to effectively capture this anomaly. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a data processing method for quality monitoring of a power distribution equipment production line.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a data processing method for quality monitoring of a power distribution equipment production line, comprising the following steps:

[0006] Based on the collected continuous distribution equipment sensor data stream, a time-decay dynamic fluctuation index is calculated, and data points and timestamps with changes exceeding a threshold are recorded to obtain a time-decay sampling point sequence. Based on the time-decay sampling point sequence, preset circuit breaker key process events are monitored to generate a hybrid quality data summary;

[0007] Based on the hybrid quality data summary, a dense raw data stream segment generated by event triggering is extracted to obtain an event-triggered raw data segment; based on the event-triggered raw data segment, the data segment is decomposed into a low-frequency approximate coefficient sequence representing a smooth trend of the signal and a high-frequency detail coefficient sequence representing a sudden change of the signal at different time scales, and a multi-scale decomposition coefficient set is established;

[0008] Based on the multi-scale decomposition coefficient set, at each wavelet decomposition scale and each divided time window, the energy value is calculated, all energy values ​​are summarized to obtain a piecewise scale energy value set, and based on the piecewise scale energy value set, each energy value in the set is arranged according to the rule that the corresponding time window is a column and the corresponding decomposition scale is a row, to construct a time-frequency energy feature matrix;

[0009] The Frobenius norm between the time-frequency energy feature matrix and the standard energy matrix template is calculated, and the numerical value is compared with the preset quality judgment threshold to generate a production quality status mark.

[0010] Preferably, the step of acquiring the time decay sampling point sequence is:

[0011] Based on the continuously recorded sensor values ​​and corresponding timestamps in the distribution equipment sensor data stream, a time window range is set and all sensor value and timestamp combinations within each time window are intercepted to establish a sensor value sequence and timestamp sequence of the distribution equipment within the time window;

[0012] Calculate the time decay dynamic fluctuation index of the time window based on the sensor value sequence and timestamp sequence of the power distribution equipment within the time window;

[0013] According to the difference between the time decay dynamic fluctuation index and the standard reference index, it is determined whether the difference exceeds the preset alarm threshold. If the condition is met, all sensor values ​​and timestamps in the time window are recorded as a time decay sampling point sequence.

[0014] Preferably, the steps of obtaining the hybrid quality data summary are:

[0015] Based on the time-attenuated sampling point sequence, all distribution equipment sensor values ​​and corresponding timestamps in the sequence are parsed, preset circuit breaker key process event trigger conditions are retrieved and matched, and the distribution equipment sensor values ​​are compared one by one to see whether they meet the event trigger conditions, the event trigger position and corresponding timestamp are determined, and the circuit breaker key process event trigger time information is generated;

[0016] Based on the triggering time information of the circuit breaker key process event, a target time period before and after the event is defined, all distribution equipment sensor values ​​and corresponding timestamps in the original distribution equipment sensor data stream with timestamps within the target time period are intercepted, and the target time period data of the original distribution equipment sensor data stream is extracted and recorded to form an event triggering original data stream segment;

[0017] According to the event-triggered original data stream fragment, the timestamp in the event-triggered original data stream fragment is used as the matching benchmark, the event-triggered original data stream fragment and the time-attenuated sampling point sequence are integrated one by one, and the integrated distribution equipment sensor values ​​and the corresponding timestamps are merged to generate a hybrid quality data summary.

[0018] Preferably, the steps of obtaining the event-triggered original data segment are:

[0019] Based on the hybrid quality data summary, all entries recording the triggering moments of key process events of the circuit breaker are extracted, the timestamp corresponding to each event trigger is located, and the sampling length is traced back and extended backward by a set sampling length before and after each timestamp. The sensor values ​​and timestamps of all distribution equipment within the time interval are intercepted to form event trigger data segments;

[0020] Calculating an event-triggered weighted curvature index according to the event-triggered data segment;

[0021] According to the event-triggered weighted curvature index, a comparison is performed item by item with the preset curvature alarm threshold, and all event-triggered data segments that meet the condition that the event-triggered weighted curvature index is greater than the preset curvature alarm threshold are screened, and all distribution equipment sensor values ​​and timestamps in the data segments are extracted and recorded to generate the event-triggered original data segment.

[0022] Preferably, the steps of obtaining the multi-scale decomposition coefficient set are:

[0023] Based on the event-triggered raw data segments, extract the continuously arranged distribution equipment sensor value sequences in each event-triggered raw data segment one by one, determine the starting and ending data point positions of each value sequence, record the position index information and the corresponding value sequence, and form a continuous event-triggered raw value sequence;

[0024] According to the original numerical sequence triggered by the continuous event, each numerical sequence is subjected to discrete wavelet decomposition step by step, and the sequence is separated into low-frequency components and smoothing trends layer by layer. The sequence after each layer of decomposition is recorded as the corresponding low-frequency approximate coefficient sequence and high-frequency detail coefficient sequence, and a multi-scale decomposition coefficient set is established.

[0025] Preferably, the steps of obtaining the segmented scale energy value set are:

[0026] Based on the multi-scale decomposition coefficient set, a low-frequency approximate coefficient sequence and a high-frequency detail coefficient sequence at each wavelet decomposition scale in the multi-scale decomposition coefficient set are extracted one by one; according to the start and end timestamps of the uniformly divided time windows, a coefficient segment range corresponding to each time window is determined; all coefficient values ​​within each coefficient segment range are intercepted to form a segmented scale coefficient segment;

[0027] According to the piecewise scale coefficient segments, the coefficient values ​​in each piecewise scale coefficient segment are operated one by one, and the energy values ​​under each wavelet decomposition scale and the corresponding time window are calculated in turn, and the energy values ​​are recorded as piecewise scale energy values ​​to generate a piecewise scale energy value set.

[0028] Preferably, the step of obtaining the time-frequency energy feature matrix is:

[0029] Based on the segmented scale energy value set, the energy values ​​are extracted one by one, with each time window as the column of the matrix and each wavelet decomposition scale as the row of the matrix. The energy values ​​are then filled into a two-dimensional matrix with the time window and the wavelet decomposition scale as the rows and columns to form a time-frequency energy feature matrix.

[0030] Preferably, the steps for obtaining the production quality status mark are:

[0031] Based on the time-frequency energy characteristic matrix, the real-time operating conditions of the current transformer winding are retrieved and matched from a pre-built standard energy matrix template library, the corresponding standard energy matrix template is retrieved, and the values ​​of each element in the standard energy matrix template are obtained to form standard energy matrix template data;

[0032] According to the standard energy matrix template data and the time-frequency energy feature matrix, the energy values ​​at the same position of the two matrices are subtracted from each other and then squared one by one, the sum of the squares of the difference values ​​at all positions of the matrices is summarized, and the result is recorded after taking the square root of the sum to generate the Frobenius norm value of the degree of matrix difference;

[0033] Based on the Frobenius norm value of the matrix difference degree, it is compared with the preset quality judgment threshold item by item. If the Frobenius norm value exceeds the quality judgment threshold, the operating condition corresponding to the current transformer winding is marked as abnormal, otherwise it is marked as normal, and a production quality status mark is generated.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are:

[0035] This invention implements a dynamic perception and time-attenuation modeling mechanism for sensor data streams from continuous power distribution equipment. It constructs a time-attenuation dynamic fluctuation index to capture short-term numerical mutations. This allows for timely location of key change points that may indicate quality anomalies and records timestamps, improving sensitivity to data behavior before and after event triggering. After anomaly identification, the method proactively intercepts dense raw data segments and performs multi-level discrete wavelet decomposition of the time-domain signals to refine and extract smoothing trends and multi-scale mutation features, enhancing the ability to characterize detailed changes in key circuit breaker processes. The decomposition coefficients are segmented by time window and scale range to generate energy values. A time-frequency energy feature matrix with horizontal and vertical scale information dimensions is then constructed, mapping dynamic behaviors in the time and frequency domains uniformly into a well-structured two-dimensional feature space. This feature matrix is ​​then compared with a standard energy matrix template for structural similarity, using the Frobenius norm as a difference measure. This effectively determines whether a product deviates from normal operating conditions, enabling accurate assessment of quality status. This improves the granularity of detecting abnormal quality segments, the dimensional integrity of feature identification, and the consistency of the judgment output. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] See also Figure 1 The present invention provides a technical solution, a data processing method for quality monitoring of a power distribution equipment production line, comprising the following steps:

[0039] Based on the collected continuous distribution equipment sensor data stream, the time-decay dynamic fluctuation index is calculated, and the data points and timestamps where the change exceeds the threshold are recorded to obtain a time-decay sampling point sequence. Based on the time-decay sampling point sequence, the preset circuit breaker key process events are monitored to generate a hybrid quality data summary.

[0040] Based on the hybrid quality data summary, the dense raw data stream segments generated by event triggering are extracted to obtain the event-triggered raw data segments. Based on the event-triggered raw data segments, the data segments are decomposed into low-frequency approximate coefficient sequences representing the smooth trend of the signal and high-frequency detail coefficient sequences representing the sudden changes of the signal at different time scales, and a multi-scale decomposition coefficient set is established.

[0041] Based on the multi-scale decomposition coefficient set, the energy value is calculated at each wavelet decomposition scale and each divided time window, and all energy values ​​are summarized to obtain a piecewise scale energy value set. Based on the piecewise scale energy value set, each energy value in the set is arranged according to the rule of corresponding time windows as columns and corresponding decomposition scales as rows to construct a time-frequency energy feature matrix;

[0042] The Frobenius norm between the time-frequency energy feature matrix and the standard energy matrix template is calculated, and the numerical value is compared with the preset quality judgment threshold to generate a production quality status mark.

[0043] The steps for obtaining the time decay sampling point sequence are:

[0044] Based on the continuously recorded sensor values ​​and corresponding timestamps in the distribution equipment sensor data stream, a time window range is set and all sensor value and timestamp combinations within each time window are intercepted to establish a sensor value sequence and timestamp sequence of the distribution equipment within the time window;

[0045] According to the value sequence and timestamp sequence of the distribution equipment sensor in the time window, the time decay dynamic fluctuation index of the time window is calculated. The calculation formula is:

[0046] ;

[0047] in, Indicates the The time-decayed dynamic volatility index of the time window, Indicates the The distribution equipment sensor values ​​of data points, Indicates the The timestamp of each data point, represents the total number of data points in the time window, Indicates the end timestamp of the current time window. represents the time decay constant, represents the time-decayed weighted average of the window, Represents the dynamic volatility factor, which adjusts the impact of volatility on the total index;

[0048] Based on the difference between the time-attenuated dynamic fluctuation index and the standard reference index, it is determined whether the difference exceeds the preset alarm threshold. If the condition is met, all sensor values ​​and timestamps within the time window are recorded as a time-attenuated sampling point sequence.

[0049] Specifically, based on the continuously recorded sensor values ​​and corresponding timestamps in the distribution equipment sensor data stream, it is first necessary to determine the time window range for data segmentation. The setting of this range is directly related to the sensitivity and completeness of event capture, and its size is determined by the typical duration of the key process of the production line. For example, for the automatic contact assembly process on the circuit breaker production line, its core actions, such as grasping, positioning, and pressing, are usually completed within 2 to 5 seconds. Therefore, a fixed time window of 5 seconds can be set. To ensure that instantaneous changes occurring at the window boundary are not missed, a sliding step of 1 second is adopted. That is, a new data window with a 4-second overlap with the previous window is generated every 1 second. This setting is obtained by synchronously analyzing video records and sensor data of more than 100 complete process cycles in historical production data, counting the start and end time points of key actions, and selecting the minimum time length that can cover more than 99% of the key action intervals. The specific operation is that the system continuously reads the torque value and corresponding timestamp pairs collected by the sensor from the real-time data stream. For example, (1.1Nm, 1677721600.5s), (1.2Nm, 1677721601.0s), …, (4.8Nm, 1677721605.5s). When the first data point enters, the first 5-second window is opened. From the timestamp 1677721600.5s to 1677721605.5s, all sensor values ​​and timestamps in this interval are intercepted to form a value list and a timestamp list respectively. These two lists together constitute the first time window. Then, the time window slides forward 1 second, and the new window range becomes 1677721601.5s to 1677721606.5s. The system again intercepts all data points in this range to form a second set of value lists and timestamp lists. This process is repeated as the data stream continues to input, providing continuous data slices for subsequent dynamic fluctuation analysis, and finally establishing a series of distribution equipment sensor value sequences and timestamp sequences arranged in chronological order, containing specific sensor readings and precise time records.

[0050] formula: The usefulness of the formula is that it combines the time-decayed weighted mean and the time-decayed weighted standard deviation to form a composite index , used to comprehensively evaluate the status of power distribution equipment within a certain time window during the production process. The first Represents the weighted central trend of the signal, and the second term represents the weighted dispersion or volatility of the signal, by introducing the time decay factor , so that the data points closer to the current time point have higher weights, which enables the index to immediately and sensitively reflect the latest dynamic changes in the production process, which is crucial for timely detection of sudden quality problems such as tool wear, part jamming, etc. In addition, the dynamic fluctuation factor The introduction of allows the flexible adjustment of the proportion of central trend and volatility in the final index according to different monitoring scenarios and signal characteristics. For example, for processes that require high stability, the This design enables the index to identify potential quality risks earlier and more accurately than traditional moving averages or standard deviations.

[0051] Indicates the The distribution equipment sensor value of each data point is directly obtained from the distribution equipment sensor value sequence within the time window established in the previous step. It represents the real-time measurement value of the physical quantity in the production process. Taking the torque sensor used to monitor the screw tightening process during circuit breaker assembly as an example, its unit is Newton meter (Nm). The sensor is installed on the automatic tightening shaft and collects torque data at a frequency of 100Hz to ensure that every subtle change in the tightening process can be captured. For example, within a time window, the torque sensor value sequence obtained is [1.2, 1.3, 1.5, 4.5, 4.8].

[0052] Indicates the The timestamp of the data point is the same as the sensor value. One-to-one correspondence is also obtained from the timestamp sequence of the distribution equipment sensors within the time window established in the previous step. The timestamp records the moment when each sensor value is collected. For example, the timestamp sequence corresponding to the above torque values ​​may be [1677721603.1, 1677721603.8, 1677721604.2, 1677721604.5, 1677721604.9], in seconds.

[0053] Indicates the total number of data points in the time window. This value is determined by the total number of sensor data points contained in the time window intercepted in the previous step. It is an integer that changes with the data acquisition frequency and window length in the time window. For example, if the time window size is 2 seconds and the sensor sampling frequency is 100Hz, then theoretically The value of is 200. In actual operation, this value is obtained by directly counting the sensor value sequence in the time window. For example, for the sequence [1.2, 1.3, 1.5, 4.5, 4.8], The value of is 5.

[0054] Indicates the end timestamp of the current time window. This value defines the reference end point for calculating the time decay weight and is the right boundary of the current analysis window. In the sliding window mechanism, The end timestamp of a window is obtained by adding the fixed window duration to its start timestamp. For example, a window starting at 1677721603.0s and lasting 2 seconds has an end timestamp of That is 1677721605.0s.

[0055] Represents the time decay constant, which is a key adjustment parameter used to control the speed at which the weight of historical data decays over time. The larger the value, the higher the weight of recent data points, and the more sensitive the index is to immediate changes, but it may ignore slow and persistent changes. The setting of the value should be based on the analysis of the dynamic characteristics of the production process. The specific acquisition steps are: first, collect the historical data set containing known faults (such as insufficient tightening torque) and normal processes, and apply different The value (for example, from 0.5 to 5.0, with a step size of 0.1) is calculated sequence, then, evaluate each Fault samples under The value of normal sample The discrimination of the values ​​can be calculated by The divergence of the distribution (such as KL divergence) or the accuracy of the classification model is used to quantify the difference. For example, by testing the data of 100 cases of insufficient torque failure and 500 cases of normal tightening process, it was found that When the fault sample The difference between the value and the normal value is the most significant, so the .

[0056] Represents the dynamic volatility factor, which is used to adjust the signal volatility (weighted standard deviation) in the total index The proportion of the value reflects the sensitivity of the monitoring system to the change of process stability. The value setting depends on the key quality points of the monitored process. If the quality problem is mainly manifested in the violent fluctuation of the signal value (such as equipment jitter), a higher value should be set. If the main manifestation is the deviation of the signal mean (such as insufficient pressure), then The value can be appropriately reduced. The setting process is as follows: Build a simulation or historical data set containing a variety of typical failure modes (such as mean shift, variance increase, and instantaneous spikes). For each failure type, set an optimization goal, that is, maximize the time it takes for the failure to occur. The rate of change of is searched within a certain range (for example, 0.1 to 3.0) through grid search or optimization algorithm. value to find the most balanced response to various critical faults For example, in circuit breaker assembly, the mean and fluctuation of torque are equally important. By evaluating the F1 score of historical data, the When the torque is too low, the comprehensive recognition effect of the two types of faults, "insufficient torque" and "excessive torque fluctuation", is the best.

[0057] Calculation process:

[0058] A specific time window , its end timestamp s.

[0059] The data points within the window are as follows:

[0060] ;

[0061] Timestamp sequence :

[0062] Torque value sequence : Nm;

[0063] The parameters to be set are: , .

[0064] Calculate the time difference for each data point :

[0065] s;

[0066] s;

[0067] s;

[0068] s;

[0069] s;

[0070] Calculate the time decay weight for each data point :

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] Calculate the weight and :

[0077] ;

[0078] Calculate the time-decaying weighted average :

[0079] ;

[0080] ;

[0081] Nm

[0082] Calculate the numerator of the time-decayed weighted variance :

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] Calculate the final time-decay dynamic fluctuation index :

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] The result shows that within the current time window, after comprehensively considering the latest trend and volatility of the torque value, the dynamic fluctuation index is 6.0017.

[0097] The time-decayed dynamic fluctuation index obtained from the previous calculation step The system compares it with a dynamically generated standard reference index. This standard reference index is not a single fixed value, but is established after in-depth analysis of the same production process data of at least 100 batches of qualified products in historical data. The specific establishment process is: the sensor data stream of each batch of qualified products is divided into exactly the same windows and Calculation method, get a series of time-varying Benchmark sequence, then align these benchmark sequences according to process time and calculate all The average value of the values ​​forms a standard reference index mean curve , and calculate the standard deviation of the corresponding position to form a standard deviation curve , alarm threshold The setting is also based on this historical data, by calculating the value of each window in all historical normal samples. Value and its corresponding reference mean The absolute difference of , forms a difference distribution, and takes the 98.5 percentile of the distribution as the alarm threshold to balance the missed alarm rate and the false alarm rate. For example, after statistical analysis, the value corresponding to the 98.5 percentile of the distribution is 0.85, then Set to 0.85, in real-time monitoring, for the time window, the system obtains the corresponding reference value from the standard reference index mean curve , for example, 4.9, and then calculate the absolute difference between the current index and the reference index, that is, Then, this difference is compared with the preset alarm threshold, that is, to determine whether 1.1017 is greater than 0.85. In this case, the condition is met, indicating that the production status in the current time window has shown signs of significant deviation from the normal range. Once this condition is met, the system will immediately perform a recording operation, extracting all the raw data contained in the time window that triggered the alarm, that is, the complete sensor value sequence and the corresponding timestamp sequence, intact, and storing them in a sequence set specifically used to store potential abnormal data fragments. This set is the time-attenuated sampling point sequence.

[0098] The steps to obtain the hybrid quality data summary are:

[0099] Based on the time-decay sampling point sequence, the sensor values ​​of all distribution equipment in the sequence and the corresponding timestamps are parsed, the preset trigger conditions of the circuit breaker key process events are retrieved and matched, and the sensor values ​​of the distribution equipment are compared one by one to see if they meet the event trigger conditions. The event trigger position and the corresponding timestamp are determined, and the trigger time information of the circuit breaker key process event is generated;

[0100] Based on the triggering time information of the key process event of the circuit breaker, a target time period before and after the event is defined. All the sensor values ​​and corresponding timestamps of the distribution equipment in the original distribution equipment sensor data stream whose timestamps are within the target time period are intercepted. The data of the original distribution equipment sensor data stream in the target time period is extracted and recorded to form a fragment of the event-triggered original data stream;

[0101] According to the event-triggered original data stream fragments, the timestamps in the event-triggered original data stream fragments are used as the matching benchmark. The event-triggered original data stream fragments are integrated one by one with the time-attenuated sampling point sequence. The integrated distribution equipment sensor values ​​and the corresponding timestamps are merged to generate a hybrid quality data summary.

[0102] Specifically, based on the time-attenuated sampling point sequence, the system first parses it and extracts all the distribution equipment sensor values ​​and corresponding timestamps contained therein. These data points are fragments that were previously marked as having significant dynamic fluctuations. Next, the system loads a preset circuit breaker key process event trigger condition rule base, which is established based on the analysis of historical sensor data and synchronized video recordings of more than 500 standard production cycles. It decomposes each key process, such as "main contact closure" and "arc extinguishing chamber assembly pressing", into a series of sub-stages with clear parameterized characteristics. For example, for the "main contact closure" event, its trigger condition is defined as: the torque sensor value quickly climbs from a baseline value of less than 0.5 Newton meters to more than 3.0 Newton meters within 0.2 seconds, and reaches a peak value when it reaches the peak value. Within 0.1 seconds after the value is reached, the change in the displacement sensor reading is less than 0.1 mm. The data points in the time-attenuated sampling point sequence are traversed one by one, and the continuous data points are input into the state machine. The state machine performs state transfer according to the conditions defined in the rule base. When the appearance of a data point or a sequence of data points causes the state machine to successfully transition from the initial state to the final "event confirmation" state, the system considers that a key process event has been triggered. At this time, the system will record the timestamp of the most critical data point that triggered the event. For example, the timestamp of the data point that first exceeds 3.0 Newton meters during the torque ramp-up process is used as the precise trigger position of the event. The system summarizes all events successfully matched in the time-attenuated sampling point sequence and their corresponding trigger positions and timestamps to generate the triggering moment information of the circuit breaker key process event.

[0103] According to the triggering time information of the circuit breaker key process event generated in the previous step, which includes a series of event types and corresponding precise timestamps, the system will define a dedicated target time period for each identified event. The range of this time period is not fixed, but is dynamically determined by an "event-time window" configuration table. This configuration table is based on the contextual analysis of various types of events in historical data, and aims to capture the complete data waveform that is most relevant to event quality assessment. For example, by analyzing the "main contact closure" event data of historical qualified and defective products, it is determined that the pre-preparation stage and post-stabilization stage of the torque signal are crucial for quality judgment. Therefore, the target time period set for the "main contact closure" event in the configuration table is 2.5 seconds before the event triggering moment to 1.5 seconds after the event triggering moment. 1.5 seconds, total duration 4 seconds. When the system processes a "main contact closed" event with a trigger time of 1677721630.5s, it will query the configuration table and calculate the target time period as [1677721628.0s, 1677721632.0s]. Subsequently, the system uses this time range to perform a time range query on the most original and highest sampling frequency distribution equipment sensor data stream stored in the database, and extracts all sensor values ​​and corresponding timestamps within this interval. This process will not perform any form of deletion or downsampling on the data, retaining the most complete process details. The extracted high-density, high-fidelity data set is recorded as an independent unit to form an event-triggered original data stream fragment.

[0104] Based on the event-triggered raw data stream segment formed in the previous step, the system starts the integration process. The core of this process is to fuse the high-resolution raw process data with the sparse abnormal fluctuation information identified previously. The system uses the timestamp in the event-triggered raw data stream segment as the only matching benchmark and traverses each data point in the segment. For each data point, for example (4.5 Newton meters, 1677721630.85s), the system queries the time-attenuated sampling point sequence generated in the previous step to check whether there are data points with exactly the same timestamp in the sequence. In order to handle slight time deviations, an extremely small time tolerance is used during the query. The tolerance value is set according to the sensor sampling frequency. For example, for a sampling rate of 100Hz, the tolerance can be set to 0.005 seconds. If a timestamp in [1677721630.845s, 1677721630.85s] is found in the time-attenuated sampling point sequence, 55s] range, the data point currently being processed from the event-triggered raw data stream fragment is considered to be "double-confirmed", that is, it is both part of the key event and within a time window of overall fluctuation anomaly. When merging and integrating, the system will attach one or more source identifiers to each data point. For example, a data point will be represented as {timestamp: 1677721630.85, value: 4.5, source: ['raw_event', 'decay_sampled']} in the new data structure. If no corresponding point is found in the time decay sampling point sequence, its source identifier is {source: ['raw_event']}. By performing this corresponding integration operation on all data points in the event-triggered raw data stream fragment, all merged data points are rearranged in chronological order to generate a hybrid quality data summary.

[0105] The steps for obtaining the event-triggered raw data segment are:

[0106] Based on the hybrid quality data summary, all entries recording the triggering moments of key circuit breaker process events are extracted. The timestamp corresponding to each event trigger is located. The sampling length is then traced back and extended backward by a set sampling length before and after each timestamp. The sensor values ​​and timestamps of all distribution equipment within this time interval are intercepted to form event trigger data fragments.

[0107] According to the event trigger data fragment, the event trigger weighted curvature index is calculated, and the calculation formula is:

[0108] ;

[0109] in, Indicates the The event triggering weighted curvature index corresponding to each circuit breaker key process event; Indicates the The sensor values ​​of the power distribution equipment, 、 Represent the sensor values ​​of the adjacent front and back data points respectively, Indicates the total number of data points in the event-triggered data segment. Indicates the The timestamp of each data point, Indicates the timestamp corresponding to the event trigger, represents the time decay coefficient, represents the standard deviation of all sensor values ​​in the event-triggered data segment, Represents a very small constant used to prevent the denominator from being zero;

[0110] According to the event trigger weighted curvature index and the preset curvature alarm threshold, each item is compared, and all event trigger data segments that meet the condition that the event trigger weighted curvature index is greater than the preset curvature alarm threshold are screened, and all distribution equipment sensor values ​​and timestamps in the data segments are extracted and recorded to generate the event trigger original data segment.

[0111] Specifically, based on the hybrid quality data summary, a screening operation is first performed to specifically extract those entries that clearly record the triggering time information of the circuit breaker key process event in their source identifier. This process locates the entry with the 'raw_event' identifier and associated with the specific event type (such as "main contact closure") by scanning the metadata of each data point in the summary. After locating the timestamp corresponding to each event trigger, the system will query a preset "event sampling length configuration table", which defines in detail the data window size required to be intercepted for different key process events. This configuration table is based on signal analysis of historical event data of more than 200 marked quality states (qualified and unqualified), and aims to accurately capture the signal form most relevant to quality defects. For example, for the high-speed transient event of "energy storage spring release", it was found through analysis. Its characteristic signals are mainly concentrated within 50 sampling points before and after the event. Therefore, the forward and backward sampling length set for the event in the configuration table is 50 points, and the backward sampling length is also 50 points. For example, if the sensor sampling frequency is 500Hz, the corresponding time interval is 100 milliseconds before and after the event stamp. When processing an "energy storage spring release" event with a trigger timestamp of 1677721645.120s, the system will calculate the time interval that needs to be intercepted, that is, from 1677721645.020s to 1677721645.220s. Subsequently, the system uses this time interval to intercept all distribution equipment sensor values ​​and corresponding timestamps whose timestamps fall within this range in the hybrid quality data summary, forming an independent data sequence containing 101 data points. This sequence constitutes an event trigger data fragment.

[0112] formula: The benefit of the formula is that it can achieve a precise quantification of the instantaneous signal morphology of key process events by calculating a standardized, time-weighted curvature index. This item approximates the second-order derivative of the signal, i.e., the curvature, in a discrete time series. It can sensitively capture nonlinear changes in the signal, such as the sharpness of the inflection point or the oscillation of the signal. This is crucial for identifying abnormal signal mutations caused by mechanical impact, material breakage, or poor electrical contact. By introducing the time attenuation weight , so that the calculation focus is on the core moment of the event The closer to the event core, the greater its contribution to the curvature index, which effectively filters out irrelevant fluctuation interference far away from the event core. Finally, the root mean square value of the weighted curvature is divided by the standard deviation of the entire data segment. , normalized, which makes The index is free from the influence of the absolute amplitude of the signal and only focuses on the "shape" or "smoothness" of the signal, enabling fair comparison between different working conditions or different batches of products.

[0113] , , Respectively represent , and The sensor values ​​of the power distribution equipment are directly derived from the sensor value sequence in the event trigger data segment generated in the previous step. They represent the physical quantities collected continuously before and after the event. For example, the acceleration sensor of the operating mechanism during the closing process of the circuit breaker is used as an example. Its unit is , the sensor collects data at a frequency of 1000 Hz. For example, the acceleration sequence in an event-triggered data segment is selected as [..., 5.2, 5.8, 15.3, 7.1, 6.5, ...], where , , .

[0114] Indicates the total number of data points in the event-triggered data segment. This value is determined by the sampling length intercepted in the previous step. For example, if you look back 50 points, extend 50 points backward, and add the event point itself, then The total number is .

[0115] Indicates the The timestamp of each data point and the sensor value One-to-one correspondence, also obtained from the event trigger data segment, provides the time information required to calculate the time decay weight. For example, the timestamp sequence corresponding to the above acceleration sequence excerpt is [..., 1677721645.119, 1677721645.120, 1677721645.121, ...], in seconds.

[0116] Indicates the timestamp corresponding to the event trigger. This is the precise moment determined by matching the key process event trigger conditions in the previous step. It is the center point of the entire event trigger data segment and the reference point for calculating the time decay weight. In this example, The value is 1677721645.120s.

[0117] represents the time decay coefficient, which controls the rate at which the time weight decays as the distance from the core moment of the event increases. The larger the value, the sharper the weight function and the more focused the analysis. The process of determining its value is as follows: prepare a sample set of event-triggered data segments containing at least 200 labeled (for example, confirmed as "normal impact" or "secondary impact" through high-frame-rate camera video analysis). For this sample set, Calculate the values ​​of all samples under the candidate value (for example, from 10 to 200, with a step size of 10) Index, then, for each value, using the calculated The index constructs a simple logistic regression classifier to distinguish normal and abnormal samples, and calculates its AUC (Area Under Curve) value on the cross-validation set, and selects the class that maximizes the AUC value. The value is taken as the optimal parameter. For example, through this method test, it is found that when When the normal and abnormal samples The index has the highest discrimination, and AUC reaches 0.96, so it is set .

[0118] It represents the standard deviation of all sensor values ​​in the event-triggered data segment. This is a statistic used to measure the degree of dispersion of the entire data segment. Before that, you need to first indivual Calculate the standard deviation of the values.

[0119] Represents a very small constant used to prevent the standard deviation In the case of extremely small values ​​(for example, the signal is a straight line), the error of the denominator being zero occurs. Its value does not affect the magnitude of the calculation result and is usually set to a very small positive number. For example, .

[0120] Calculation process:

[0121] With a As an example, the simplified event trigger data fragment Calculation.

[0122] Event trigger timestamp s.

[0123] The parameters to be set are: , .

[0124] The data snippet is as follows:

[0125] :1,2,3,4,5;

[0126] (s): [1677721645.118, 1677721645.119, 1677721645.120, 1677721645.121, 1677721645.122];

[0127] ( ): [5.2, 5.8, 15.3, 7.1, 6.5];

[0128] Calculating standard deviation :

[0129] average value ;

[0130] ;

[0131] The loop calculates the summation term (from arrive ):

[0132] for :

[0133] ;

[0134] Weight ;

[0135] Curvature term ;

[0136] Numerator ;

[0137] for :

[0138] ;

[0139] Weight ;

[0140] Curvature term ;

[0141] Numerator ;

[0142] for :

[0143] ;

[0144] Weight ;

[0145] Curvature term ;

[0146] Numerator ;

[0147] Compute the sum of the numerator and denominator:

[0148] ;

[0149] ;

[0150] Calculate the final event-triggered weighted curvature index :

[0151] ;

[0152] ;

[0153] The results show that the weighted curvature index of this event-triggered data segment is 1.6727. This dimensionless value quantifies the sharpness and roughness of the signal waveform near the core moment of the event. The higher the value, the steeper and more irregular the signal shape.

[0154] The calculated event trigger weighted curvature index The system compares it with a preset curvature alarm threshold item by item. The setting of this threshold is based on the statistical learning results of a large amount of historical data. The specific setting process is as follows: First, a database containing more than 1,000 event-triggered data fragments is collected. These data have been accurately marked as "qualified" or "defective" through subsequent physical inspection or expert review. For each data fragment in the database, the corresponding Value, thus obtaining two groups of "qualified" and "defective" The distribution of values ​​is obtained by plotting the receiver operating characteristic (ROC) curve to analyze the trade-off between correctly identifying "defects" (true positive rate) and incorrectly identifying "qualified" (false positive rate) at different thresholds. The threshold closest to the upper left corner (0, 1) on the ROC curve is selected. This point represents the threshold that can achieve the highest recall rate with the lowest false alarm rate among all possible thresholds, thus achieving the best classification performance. For example, it is concluded that when the threshold is set to 0.95, the system's recognition accuracy for defects can reach 98%, while the false alarm rate is less than 3%. Therefore, the curvature alarm threshold is set to 0.95. In the real-time processing process, the system will calculate the value of the curvature alarm threshold at the previous step. The value 1.6727 is compared with the threshold value 0.95. Since 1.6727 is greater than 0.95, the alarm condition is met. Therefore, the system completely filters out the event trigger data segment that is determined to be abnormal, extracts all the distribution equipment sensor values ​​and corresponding timestamps it contains, and records them in a new set. All data segments retained after this screening step together constitute the event trigger original data segment.

[0155] The steps to obtain the multi-scale decomposition coefficient set are:

[0156] Based on the event-triggered raw data segments, extract the continuously arranged distribution equipment sensor value sequences in each event-triggered raw data segment one by one, determine the starting and ending data point positions of each value sequence, record the position index information and the corresponding value sequence, and form a continuous event-triggered raw value sequence;

[0157] According to the original numerical sequence triggered by continuous events, each numerical sequence is decomposed by discrete wavelet decomposition layer by layer, and the low-frequency components and smooth trends of the sequence are separated layer by layer. The decomposed sequence of each layer is recorded as the corresponding low-frequency approximate coefficient sequence and high-frequency detail coefficient sequence, and a multi-scale decomposition coefficient set is established.

[0158] Specifically, based on the event-triggered original data segment, the system iteratively processes each record therein. For each record, that is, an independent event-triggered original data segment, the system first extracts the distribution equipment sensor values ​​arranged continuously in chronological order to form a one-dimensional array of pure values. At the same time, the system determines the starting data point position and the ending data point position of the value sequence in the original data segment. For example, a data segment may contain 101 data points, and the system records its starting index as 0 and the ending index as 100, and packages this value sequence with the corresponding index information to form a structured data object, which is a continuous event-triggered original value sequence. The purpose of this operation is to simplify the data segment containing metadata such as timestamps into a pure value sequence to prepare for subsequent wavelet decomposition calculations. This process is executed in a loop until all entries in the event-triggered original data segment are processed, and finally a set consisting of multiple continuous event-triggered original value sequences is formed.

[0159] According to the continuous event formed in the previous step, the original numerical sequence set is triggered. The system independently performs discrete wavelet decomposition (DWT) on each numerical sequence. The selection of wavelet basis function is the key and needs to be determined according to the characteristics of the signal. For the impact and vibration signals commonly seen in the production process of distribution equipment, they usually have non-stationary and transient characteristics. Therefore, the Daubechies series wavelets with tight support and good symmetry are selected, such as the "db4" wavelet. The number of decomposition layers also needs to be set in advance. The maximum number of layers is limited by the sequence length and is usually determined according to the frequency band range where the main characteristics of the signal are located. Through spectral analysis of historical fault data, it is found that the frequencies of most key defect characteristics (such as secondary impacts and ultrasonic waves generated by tiny cracks) are concentrated in the medium and high frequency bands. Setting the number of decomposition layers to 5 can effectively cover these frequency bands while retaining sufficient time resolution. The decomposition process is carried out step by step using the MALLAT algorithm. For a continuous event of length N that triggers the original numerical sequence, The first layer decomposition passes it through a low-pass filter and a high-pass filter to obtain a low-frequency approximate coefficient sequence of length N / 2. and high-frequency detail coefficient sequence , represents the smooth trend of the signal, and Capture the highest frequency mutation information, then the low frequency approximate coefficient sequence of the first layer As the new input, the same filtering and downsampling operation is performed again to obtain the low-frequency approximate coefficient sequence of the second layer. and high-frequency detail coefficient sequence , this process is repeated 5 times, and finally a sequence of 5 high-frequency detail coefficients is obtained ( ) and a final low-frequency approximation coefficient sequence ( ), which together constitute the multi-scale representation of the original numerical sequence. The system stores all coefficient sequences obtained after the decomposition of the original numerical sequence triggered by each continuous event, together with their corresponding decomposition level information, to establish a multi-scale decomposition coefficient set.

[0160] The steps for obtaining the piecewise scale energy value set are:

[0161] Based on the multi-scale decomposition coefficient set, the low-frequency approximate coefficient sequence and the high-frequency detail coefficient sequence at each wavelet decomposition scale in the multi-scale decomposition coefficient set are extracted one by one. According to the start and end timestamps of the uniformly divided time window, the coefficient segment range corresponding to each time window is determined, and all coefficient values ​​within each coefficient segment are intercepted to form a segmented scale coefficient segment;

[0162] According to the piecewise scale coefficient segments, the coefficient values ​​in each piecewise scale coefficient segment are operated one by one, and the energy values ​​under each wavelet decomposition scale and the corresponding time window are calculated in turn. The energy values ​​are recorded as piecewise scale energy values ​​to generate a piecewise scale energy value set.

[0163] Specifically, based on the multi-scale decomposition coefficient set, the system processes each set of decomposition coefficients. First, the system loads a unified time window division scheme, which aims to standardize the time axis of the original event. The division is based on the analysis of a large amount of historical event data to identify the typical stages of event development. For example, for an event with a total duration of 200 milliseconds, it can be divided into 10 continuous, non-overlapping time windows, each with a duration of 20 milliseconds. The system traverses each set of coefficients in the multi-scale decomposition coefficient set, that is, all decomposition coefficient sequences corresponding to an original event, including 5 high-frequency detail coefficient sequences ( arrive ) and a low-frequency approximation coefficient sequence ( ), due to the downsampling characteristics of discrete wavelet transform, the length of coefficient sequence under different decomposition scales is different. For example, if the original sequence length is 101, then The length is about 51, The length is about 26, and so on, the system is based on each decomposition scale The coefficient points The number of original signal points The proportional relationship ( ), the start and end timestamps of the uniformly divided time windows are mapped to the corresponding index range on each coefficient sequence. For example, the first 20 millisecond time window corresponds to indexes 0 to 20 on the original signal, and the coefficients are decomposed in the first layer. The above corresponds to index 0 to 10, and the coefficients are decomposed in the second layer The above corresponds to indexes 0 to 5. Based on these calculated index ranges, the system accurately extracts the coefficient fragments corresponding to each time window from the low-frequency approximate coefficient sequence and high-frequency detail coefficient sequence at each decomposition scale, forming a set of segmented scale coefficient fragments organized by decomposition scale and time window.

[0164] According to the set of segmented scale coefficient segments formed in the previous step, the system performs energy calculation on each segment. The energy value is calculated by summing the squares of all coefficient values ​​in the segment. Decomposition scale, coefficient segments under a time window , its energy Calculated by the following formula: ,in This is the first coefficient values, is the length of the segment, and the calculation is applied to all the segment coefficient segments one by one. For example, for the first decomposition scale ( ) under the first time window ( ), if the coefficients it contains are [-0.5, 1.2, -0.8], then its energy value is The system calculates a unique energy value for each combination of wavelet decomposition scale (a total of 6 scales, including 5 high-frequency and 1 low-frequency) and each divided time window (a total of 10 windows). These calculated energy values ​​are recorded one by one and associated with their corresponding decomposition scale and time window index to form a data record containing 60 energy values. This record is a piecewise-scale energy value. By summarizing all the piecewise-scale energy values ​​obtained after processing an original event, a piecewise-scale energy value set is generated.

[0165] The steps to obtain the time-frequency energy feature matrix are:

[0166] Based on the piecewise scale energy value set, the energy values ​​are extracted one by one, with each time window as the column of the matrix and each wavelet decomposition scale as the row of the matrix. The energy values ​​are then filled into a two-dimensional matrix with the time window and wavelet decomposition scale as the rows and columns to form a time-frequency energy feature matrix.

[0167] Specifically, based on the piecewise scale energy value set, the system creates an initially empty two-dimensional matrix for each event to be analyzed. The dimension of the matrix is ​​predefined, the number of rows is equal to the total number of scales of the wavelet decomposition, that is, 6 (5 high-frequency detail scales and 1 low-frequency approximation scale), and the number of columns is equal to the number of uniformly divided time windows, that is, 10. Then, the system extracts each energy value in the piecewise scale energy value set one by one, and each energy value is accompanied by its corresponding decomposition scale index. and time window index Based on these two indices, the system accurately fills the energy value into the first Row, No. The position of the column. For example, if an energy value is 2.33, its decomposition scale index is 1, and the time window index is 1, then the value is filled in the (1, 1) position of the matrix. This process is performed automatically and cyclically until all 60 energy values ​​corresponding to the current event in the segmented scale energy value set are accurately filled in the corresponding positions of the matrix. After the filling is completed, the 6x10 two-dimensional matrix is ​​constructed, which intuitively shows the distribution of event energy at different frequencies (represented by the decomposition scale) and different times (represented by the time window), forming a time-frequency energy feature matrix.

[0168] The steps to obtain the production quality status mark are:

[0169] Based on the time-frequency energy feature matrix, the real-time operating conditions of the current transformer winding are retrieved and matched from the pre-built standard energy matrix template library, the corresponding standard energy matrix template is retrieved, and the values ​​of each element in the standard energy matrix template are obtained to form the standard energy matrix template data;

[0170] Based on the standard energy matrix template data and the time-frequency energy feature matrix, the energy values ​​at the same position in the two matrices are subtracted one by one and then squared one by one. The sum of the squares of the difference values ​​at all positions in the matrix is ​​summarized, and the result is recorded after taking the square root of the sum to generate the Frobenius norm value of the degree of matrix difference;

[0171] Based on the Frobenius norm value of the matrix difference degree, it is compared with the preset quality judgment threshold item by item. If the Frobenius norm value exceeds the quality judgment threshold, the operating condition corresponding to the current transformer winding is marked as abnormal, otherwise it is marked as normal, and a production quality status mark is generated.

[0172] Specifically, based on the time-frequency energy characteristic matrix, the system first needs to identify the specific model of the distribution equipment currently being processed and the ongoing production process. This information is usually obtained from the work order data of the production execution system (MES). For example, the circuit breaker currently being processed is model "DZ47-63C20" and the "closing and tripping characteristic test" process is being carried out. The system uses these working conditions as search keywords to perform precise searches in the pre-built standard energy matrix template library. This template library is generated by extracting the complete time-frequency energy characteristic matrix under the same working conditions for more than 1,000 qualified products verified as "gold standards", and then taking the average value of all elements in the matrix at the same position. Each template represents an ideal energy distribution pattern under a specific working condition. After a successful retrieval, the system retrieves the standard energy matrix template that fully matches the current working condition and reads the values ​​of all elements in the template matrix to form the standard energy matrix template data.

[0173] Based on the standard energy matrix template data and the time-frequency energy feature matrix generated in the previous step, the system performs Frobenius norm calculations. This calculation process strictly follows its mathematical definition. First, a difference matrix of the same size as the time-frequency energy feature matrix is ​​created. Each element in the difference matrix is ​​obtained by subtracting the element values ​​at the same row and column positions in the time-frequency energy feature matrix and the standard energy matrix template data. For example, the value at position (1, 1) of the time-frequency energy feature matrix is ​​2.33, and the corresponding value at position (1, 1) of the standard energy matrix template data is 2.10. Then, the value at position (1, 1) of the difference matrix is ​​0.23. After completing the subtraction operation on all 60 positions, the system squares each element in the difference matrix and accumulates the sum of all squared results. Finally, the square root of this sum is taken. The result is the Frobenius norm value between the two matrices. This value quantifies the overall degree of difference between the energy distribution pattern of the current production event and the standard pattern, generating a Frobenius norm value for the degree of matrix difference.

[0174] Based on the Frobenius norm value of the matrix difference degree generated in the previous step, the system compares it with a preset quality judgment threshold. The setting of this threshold is based on the statistical analysis of a large amount of historical data. The specific method is: collect a time-frequency energy feature matrix database containing "qualified products" and various known "defective products", calculate the Frobenius norm between each matrix in the library and the standard energy matrix template of its corresponding working condition, and thus obtain two sets of norm value distributions, namely "qualified product norm distribution" and "defective product norm distribution". Ideally, the norm value of defective products will be significantly greater than that of qualified products. The goal of setting the quality judgment threshold is to maximize the recognition of defective products. At the same time, the probability of misjudging qualified products as defective products is controlled within an acceptable range, for example, less than 1%. By analyzing the "qualified product norm distribution", its 99% quantile is taken as the threshold. For example, if the value is calculated to be 2.58, the quality judgment threshold is set to 2.58. During real-time judgment, if the calculated Frobenius norm value, for example, 3.12, is greater than 2.58, the system will mark the status of the circuit breaker currently being produced under this working condition as "abnormal". If the value is less than or equal to 2.58, it will be marked as "normal". This mark is eventually recorded in the quality traceability file of the product to generate a production quality status mark.

[0175] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A data processing method for quality monitoring of a power distribution equipment production line, characterized in that: The following steps are involved: Based on the collected continuous distribution equipment sensor data stream, a time-decay dynamic fluctuation index is calculated, and data points and timestamps with changes exceeding a threshold are recorded to obtain a time-decay sampling point sequence. Based on the time-decay sampling point sequence, preset circuit breaker key process events are monitored to generate a hybrid quality data summary; The calculation formula of the time-decay dynamic fluctuation index is: ； in, Indicates the The time-decayed dynamic volatility index of the time window, Indicates the The distribution equipment sensor values ​​of data points, Indicates the The timestamp of each data point, represents the total number of data points in the time window, Indicates the end timestamp of the current time window. represents the time decay constant, represents the time-decayed weighted average of the window, Represents the dynamic volatility factor, which adjusts the impact of volatility on the total index; Based on the hybrid quality data summary, a dense raw data stream segment generated by event triggering is extracted to obtain an event-triggered raw data segment; based on the event-triggered raw data segment, the data segment is decomposed into a low-frequency approximate coefficient sequence representing a smooth trend of the signal and a high-frequency detail coefficient sequence representing a sudden change of the signal at different time scales, and a multi-scale decomposition coefficient set is established; Based on the multi-scale decomposition coefficient set, at each wavelet decomposition scale and each divided time window, the energy value is calculated, all energy values ​​are summarized to obtain a piecewise scale energy value set, and based on the piecewise scale energy value set, each energy value in the set is arranged according to the rule that the corresponding time window is a column and the corresponding decomposition scale is a row, to construct a time-frequency energy feature matrix; Calculating the Frobenius norm between the time-frequency energy feature matrix and the standard energy matrix template, comparing the numerical value with a preset quality judgment threshold, and generating a production quality status mark; The steps for obtaining the hybrid quality data summary are: Based on the time-attenuated sampling point sequence, all distribution equipment sensor values ​​and corresponding timestamps in the sequence are parsed, preset circuit breaker key process event trigger conditions are retrieved and matched, and the distribution equipment sensor values ​​are compared one by one to see whether they meet the event trigger conditions, the event trigger position and corresponding timestamp are determined, and the circuit breaker key process event trigger time information is generated; Based on the triggering time information of the circuit breaker key process event, a target time period before and after the event is defined, all distribution equipment sensor values ​​and corresponding timestamps in the original distribution equipment sensor data stream with timestamps within the target time period are intercepted, and the target time period data of the original distribution equipment sensor data stream is extracted and recorded to form an event triggering original data stream segment; According to the event-triggered original data stream fragment, the timestamp in the event-triggered original data stream fragment is used as the matching benchmark, the event-triggered original data stream fragment and the time-attenuated sampling point sequence are integrated one by one, and the integrated distribution equipment sensor values ​​and the corresponding timestamps are merged to generate a hybrid quality data summary.

2. The data processing method for quality monitoring of a power distribution equipment production line according to claim 1, characterized in that: The steps for obtaining the time decay sampling point sequence are: Based on the continuously recorded sensor values ​​and corresponding timestamps in the distribution equipment sensor data stream, a time window range is set and all sensor value and timestamp combinations within each time window are intercepted to establish a sensor value sequence and timestamp sequence of the distribution equipment within the time window; Calculate the time decay dynamic fluctuation index of the time window based on the sensor value sequence and timestamp sequence of the power distribution equipment within the time window; According to the difference between the time decay dynamic fluctuation index and the standard reference index, it is determined whether the difference exceeds the preset alarm threshold. If the condition is met, all sensor values ​​and timestamps in the time window are recorded as a time decay sampling point sequence.

3. The data processing method for quality monitoring of a power distribution equipment production line according to claim 1, characterized in that: The steps for obtaining the event-triggered original data segment are: Based on the hybrid quality data summary, all entries recording the triggering moments of key process events of the circuit breaker are extracted, the timestamp corresponding to each event trigger is located, and the sampling length is traced back and extended backward by a set sampling length before and after each timestamp. The sensor values ​​and timestamps of all distribution equipment within the time interval are intercepted to form event trigger data segments; Calculating an event-triggered weighted curvature index according to the event-triggered data segment; According to the event-triggered weighted curvature index, a comparison is performed item by item with the preset curvature alarm threshold, and all event-triggered data segments that meet the condition that the event-triggered weighted curvature index is greater than the preset curvature alarm threshold are screened, and all distribution equipment sensor values ​​and timestamps in the data segments are extracted and recorded to generate the event-triggered original data segment.

4. The data processing method for quality monitoring of a power distribution equipment production line according to claim 1, characterized in that: The steps for obtaining the multi-scale decomposition coefficient set are: Based on the event-triggered raw data segments, extract the continuously arranged distribution equipment sensor value sequences in each event-triggered raw data segment one by one, determine the starting and ending data point positions of each value sequence, record the position index information and the corresponding value sequence, and form a continuous event-triggered raw value sequence; According to the original numerical sequence triggered by the continuous event, each numerical sequence is subjected to discrete wavelet decomposition step by step, and the sequence is separated into low-frequency components and smoothing trends layer by layer. The sequence after each layer of decomposition is recorded as the corresponding low-frequency approximate coefficient sequence and high-frequency detail coefficient sequence, and a multi-scale decomposition coefficient set is established.

5. The data processing method for quality monitoring of a power distribution equipment production line according to claim 1, characterized in that: The steps for obtaining the segmented scale energy value set are: Based on the multi-scale decomposition coefficient set, a low-frequency approximate coefficient sequence and a high-frequency detail coefficient sequence at each wavelet decomposition scale in the multi-scale decomposition coefficient set are extracted one by one; according to the start and end timestamps of the uniformly divided time windows, a coefficient segment range corresponding to each time window is determined; all coefficient values ​​within each coefficient segment range are intercepted to form a segmented scale coefficient segment; According to the piecewise scale coefficient segments, the coefficient values ​​in each piecewise scale coefficient segment are operated one by one, and the energy values ​​under each wavelet decomposition scale and the corresponding time window are calculated in turn, and the energy values ​​are recorded as piecewise scale energy values ​​to generate a piecewise scale energy value set.

6. The data processing method for quality monitoring of a power distribution equipment production line according to claim 1, characterized in that: The steps for obtaining the time-frequency energy feature matrix are: Based on the segmented scale energy value set, the energy values ​​are extracted one by one, with each time window as the column of the matrix and each wavelet decomposition scale as the row of the matrix. The energy values ​​are correspondingly filled into a two-dimensional matrix with the time window and the wavelet decomposition scale as the rows and columns to form a time-frequency energy feature matrix.

7. The data processing method for quality monitoring of a power distribution equipment production line according to claim 1, characterized in that: The steps for obtaining the production quality status mark are: Based on the time-frequency energy characteristic matrix, the real-time operating conditions of the current transformer winding are retrieved and matched from a pre-built standard energy matrix template library, the corresponding standard energy matrix template is retrieved, and the values ​​of each element in the standard energy matrix template are obtained to form standard energy matrix template data; According to the standard energy matrix template data and the time-frequency energy feature matrix, the energy values ​​at the same position of the two matrices are subtracted from each other and then squared one by one, the sum of the squares of the difference values ​​at all positions of the matrices is summarized, and the result is recorded after taking the square root of the sum to generate the Frobenius norm value of the degree of matrix difference; Based on the Frobenius norm value of the matrix difference degree, it is compared with the preset quality judgment threshold item by item. If the Frobenius norm value exceeds the quality judgment threshold, the operating condition corresponding to the current transformer winding is marked as abnormal, otherwise it is marked as normal, and a production quality status mark is generated.

Images