Power equipment multi-source fault detection method based on data fusion

Abstract

The invention discloses a power equipment multi-source fault detection method based on data fusion, and relates to the technical field of power equipment faults, and the method comprises the steps: synchronously collecting ultrahigh frequency, high frequency, ultrasonic, transient ground voltage and infrared / voiceprint data under a unified time reference, and generating a unified timestamp frame stream; checking and splicing frame streams, slicing a fixed window, carrying out frequency band preprocessing, carrying out noise bottom quantity normalization to obtain an anomaly degree, determining an event window, and outputting an event data packet; a minimum feature set is extracted to calculate a channel quality coefficient, a suspected fault event is screened out through binary classification gating, and a random forest outputs a fault type probability and forms an evidence packet in combination with channel quality; related consistency items are calculated, rule consistency items are obtained in combination with association rules, and D-S combination is executed to output fusion fault distribution, conflict degree and fusion uncertainty after a consistency index and a discount evidence are formed through fusion; and calculating a growth trend quantity so as to realize non-fault event pre-filtering and fault type probabilistic output.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault technology, and in particular to a method for detecting multi-source faults in power equipment based on data fusion. Background Technology

[0002] Power equipment fault detection technology has been evolving alongside the development of online monitoring and digital operation and maintenance. Focusing on objects such as high-voltage switchgear, transformers, and cable accessories, a multi-parameter collaborative acquisition system for partial discharge detection methods such as UHF / HF / ultrasound / transient ground voltage has been gradually formed. At the same time, the improvement of mobile terminal and edge computing capabilities has promoted the field deployment of event windowing processing, feature extraction, and intelligent recognition algorithms. Related research has developed from single-channel threshold discrimination to multi-source fusion, probabilistic diagnosis, and comprehensive assessment for fault prediction and health management.

[0003] However, existing methods still have two limitations: First, in engineering implementations for multi-source data frame streams, common practices only involve simple data reception and buffering, lacking frame-by-frame consistency verification and frame sequence continuity verification. This makes it difficult to identify and trace abnormal frames, lost frames, and out-of-order frames caused by transmission disturbances or link jitter in a timely manner, thus creating hidden gaps in subsequent window slicing and channel alignment, affecting the stability of fault detection. Second, partial discharge signals are prone to out-of-band noise superposition and front-end saturation distortion under strong interference and range-limited conditions. If existing methods do not perform explicit frequency band preprocessing on the relevant channels of partial discharge and threshold judgment isolation of the saturation ratio, they often mistake high-amplitude segments of distortion for valid abnormal features. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-source fault detection method for power equipment based on data fusion to solve the problems of existing technologies, such as lack of consistency verification and traceable marking for abnormal data frames, lost frames and out-of-order data, and lack of effective isolation for out-of-band interference and saturation distortion of partial discharge channels, which leads to feature distortion.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for multi-source fault detection of power equipment based on data fusion, which includes establishing a detection task and synchronously collecting monitoring data under the same time reference, and generating a multi-source data frame stream with a unified timestamp; Verification and splicing of multi-source data frame streams are performed, and candidate windows are formed by slicing them into fixed windows. Frequency band preprocessing is performed on each channel. The anomaly degree of each channel window is calculated by normalizing the noise floor, the event window is determined, and the event data packet is output. Extract the minimum feature set from the event data packet and calculate the channel quality coefficient. Use a binary classification model to gating the event window and screen out suspected fault events. Use a random forest classifier to output the probability distribution of the suspected fault events and combine the probability distribution with the channel quality coefficient to obtain the evidence packet. The relevant consistency items are calculated by preprocessing the waveform in the event window, and the rule consistency items are calculated based on the preset set of association rules and evidence packages. The relevant consistency items and rule consistency items are fused to form a consistency index. The consistency index and the channel quality coefficient are used to discount the evidence of each channel and then DS combination is performed to output the fusion fault distribution, conflict degree and fusion uncertainty. The determination of whether to trigger a retest is made by integrating the fault distribution, conflict degree, and integration uncertainty, and the growth trend quantity is calculated and written into the test result record.

[0007] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the specific steps for establishing a detection task and synchronously collecting monitoring data under the same time reference to generate a multi-source data frame stream with a unified timestamp are as follows. Establish a detection task and synchronously collect monitoring data under the same time reference, output a synchronous trigger pulse to the partial discharge related channel, and complete the alignment using a two-trigger calibration method; The system continuously samples the ultra-high frequency, high frequency, ultrasonic, and transient voltage channels, and writes a master clock timestamp to each sampling point. For the infrared channel, it records the master clock timestamp corresponding to each acquired image frame and writes it to the image frame header. For the acoustic pattern array channel, it acquires array data with a fixed frame length to obtain a multi-source data frame stream.

[0008] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the steps of verifying and splicing multi-source data frame streams and forming candidate windows by slicing them into fixed window segments, and performing frequency band preprocessing on each channel, are as follows: Each frame in the multi-source data frame stream is checked for consistency using CRC. If two adjacent frames in the same channel are not consecutive or not increasing, a missing flag is written. With a fixed window length and sliding step size, a candidate window sequence is generated based on a unified timestamp. The missing proportion of the candidate windows is calculated by missing markers. When the missing proportion is greater than the missing threshold, the candidate window is marked as having too many missing elements. Frequency band preprocessing is performed on the partial discharge related channels within the candidate window. Convolution operation is performed on the original segments within the candidate window to obtain the bandpass output. Saturation detection is performed on each partial discharge related channel. If the proportion of samples whose sampled values ​​reach the upper limit of the range exceeds the saturation threshold, they are marked as saturated.

[0009] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the steps of calculating the anomaly degree of each channel window by normalizing the noise floor, determining the event window, and outputting event data packets are as follows: For each partial discharge-related channel, maintain a noise floor. Take the median of the window energy of a fixed number of windows before the start of the task. For each candidate window and each partial discharge-related channel, calculate the anomaly. For each local amplifier channel, maintain an anomaly queue. When the anomaly of a channel in a candidate window is greater than the dynamic trigger threshold and there are no candidate windows with too many missing or saturated markings, mark the channel as an anomaly and valid in the current window, determine the candidate window as an event window, and encapsulate it into an event data packet.

[0010] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the specific steps for extracting the minimum feature set from event data packets and calculating the channel quality coefficient are as follows: Pulse detection is performed on the bandpass output of the event window. The number of pulse segments exceeding the pulse detection threshold within the event window is counted as the pulse counting feature. The ratio of the maximum phase bucket count to the total count within the channel window is used as the degree of aggregation to obtain the phase aggregation feature. The ratio of the energy detected as a pulse segment within the window to the total energy of the window is used as the pulse energy ratio feature. The normalized energy characteristics, pulse count characteristics, phase aggregation characteristics, and pulse energy ratio characteristics are sequentially concatenated into a minimum feature set, and the channel quality coefficient is calculated for each partial discharge-related channel within the event window.

[0011] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the steps include: using a binary classification model to gate the event window, filtering out suspected fault events, using a random forest classifier to output a probability distribution for the suspected fault events, and combining the probability distribution with the channel quality coefficient to obtain an evidence package. The specific steps are as follows. The minimum feature set is input into the binary classification model. When the binary classification model outputs a suspected fault, the current event window is marked as a suspected fault event. Based on suspected fault events, multi-class inference is performed on each partial discharge-related channel using a random forest classifier to obtain probability distributions. The probability distributions are then combined with the channel quality coefficients to obtain an evidence package.

[0012] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the specific steps for calculating relevant consistency items through preprocessed waveforms of event windows and calculating rule consistency items based on a preset set of association rules and evidence packages are as follows: Calculate relevant consistency terms for preprocessed waveforms within the same event window; Set up a set of association rules, and judge each rule of the minimum feature set based on the evidence package, and count the number of "supporting rule hits" and "conflicting rule hits" respectively, and calculate the rule consistency item.

[0013] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the method involves fusing relevant consistency items and rule consistency items to form a consistency index. After discounting the evidence from each channel using the consistency index and channel quality coefficient, a Data Synthesis (DS) combination is performed to output the fused fault distribution, conflict degree, and fusion uncertainty. The specific steps are as follows: The relevant consistency items are merged with the rule consistency items to form a consistency index. The consistency index and the channel quality coefficient are used to discount the evidence of each channel to obtain the quality allocation. Perform DS combination on the discounted evidence for each channel, calculate the conflict degree, and use the conflict degree obtained from the last combination as the overall conflict degree of the event window; The mass distribution of each fault type is extracted and normalized to obtain the fused fault distribution, and the fused uncertainty is calculated at the same time.

[0014] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the specific steps for determining whether to trigger a retest by fusing fault distribution, conflict degree, and fusion uncertainty are as follows: For each suspected fault event window, a retest trigger judgment is executed. When the fusion uncertainty is higher than the fusion uncertainty judgment threshold, or the overall conflict degree is higher than the conflict judgment threshold, or the consistency index is lower than the consistency judgment threshold, the credibility of the current fusion result is determined to be insufficient and a retest is required. When it is determined that a retest is required, select a unique retest action and execute it, calculate the unit energy consumption fusion uncertainty reduction index, and select the retest action with the largest unit energy consumption fusion uncertainty reduction index as the unique retest action.

[0015] As a preferred embodiment of the data fusion-based multi-source fault detection method for power equipment described in this invention, the specific steps for calculating the growth trend and writing it into the detection result record are as follows: After the retest is completed, the corresponding retest data is received, and the fusion uncertainty, overall conflict degree and consistency index are recalculated. When the fusion uncertainty, overall conflict degree and consistency index are all lower than their respective judgment thresholds, the fault growth trend quantity is calculated, and the fault growth trend quantity and fusion fault distribution are written into the test result record.

[0016] The beneficial effects of this invention are as follows: By inputting the minimum feature set into a binary classification model for gating and screening out suspected fault events, and then using a random forest to output a probability distribution and combine it with the channel quality coefficients to generate an evidence package, pre-filtering of a large number of non-fault events and probabilistic expression of fault type output are achieved; By performing FIR bandpass convolution on the partial discharge related channels within a window to obtain bandpass output and performing saturation detection, and setting a saturation mark when the saturation ratio exceeds the saturation threshold, the enhancement of target frequency band features and isolation of saturated distortion data are achieved; By performing CRC consistency verification frame by frame on the multi-source data frame stream and writing missing marks for discontinuous / non-incremental frame order, the automatic removal of abnormal frames and traceable labeling of missing intervals are achieved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-source fault detection method for power equipment based on data fusion.

[0019] Figure 2 A flowchart for verification, splicing, and event window construction.

[0020] Figure 3 This is a flowchart for gating recognition, evidence generation, and consistency discount fusion.

[0021] Figure 4 A flowchart for CRC consistency verification, missing flags, and candidate window construction. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for multi-source fault detection of power equipment based on data fusion, including the following steps: S1. Establish a detection task and synchronously collect monitoring data under the same time reference to generate a multi-source data frame stream with a unified timestamp.

[0026] Enter the type of equipment under test, site / interval information, test point number, tester number, and environmental information to establish the test task; collect the sampling frequency / frame rate of each channel, the activation status of each channel, the frequency band parameters of the partial discharge related channels, the number and sampling frequency of the acoustic pattern array, the infrared image resolution, and the flag bits for whether to generate phase / real-time / continuous / pulse spectra; the acquisition host internally starts a high-stability time base counter as a unified time reference, sets the time base counter as the master clock for this task, and outputs a synchronization trigger pulse to the partial discharge related channels (UHF, HF, ultrasound, transient ground voltage). Alignment is completed using a two-stage trigger calibration method to ensure that data from different channels can be mapped to the same time reference. Specifically, during the first synchronization trigger, the acquisition host records the trigger time of the master clock and records the sampling index position corresponding to the trigger pulse in each partial discharge related channel; during the second synchronization trigger, the acquisition host again records the trigger time of the master clock and the sampling index position corresponding to each channel; during the first and second synchronization triggers, the acquisition host records the trigger time under the master clock and the sampling index position of each channel, respectively. The sampling point number corresponding to the trigger pulse in the sample sequence is compared with the actual time length experienced by the master clock between two triggers and the cumulative increase in the number of sampling points between the two triggers of the channel. If there is a deviation between the cumulative number of sampling points per unit time and the nominal sampling rate of the channel, the deviation is fixed as a slight drift of the channel relative to the master clock. Taking the first trigger as the alignment anchor point, the sampling point number of the channel at the first trigger is converted to the master clock time axis and compared with the difference recorded by the master clock for the first trigger time. The difference is used as the time offset of the channel relative to the master clock. For any sampling point in the channel, the difference between the sampling index of the sampling point and the sampling index corresponding to the first trigger is used as the index increment. The slight drift is used to proportionally correct the elapsed time corresponding to the index increment to eliminate the cumulative error caused by the difference in the speed of the channel sampling clock relative to the master clock. The time offset of the channel is added on the elapsed time after proportional correction so that the alignment time of the first trigger on the channel is consistent with the first trigger time recorded by the master clock. The sampling time of the sampling point under the master clock is obtained, realizing the alignment of multiple channels under the same time base.

[0027] After alignment and calibration, the ultra-high frequency, high frequency, ultrasonic, and transient ground voltage channels are continuously sampled, and a master clock timestamp is written to each sampling point. For the infrared channel, the master clock timestamp corresponding to each acquired image frame is recorded and written to the image frame header. For the acoustic pattern array channel, array data is acquired with a fixed frame length, and the master clock timestamp corresponding to the starting sampling point of each frame is recorded. The data from each channel are encapsulated into data frames with a fixed frame length and continuously output to generate a multi-source data frame stream with a unified timestamp.

[0028] S2. Based on the multi-source data frame stream, perform verification, splicing and slicing according to a fixed window to form candidate windows, perform frequency band preprocessing on each channel, calculate the anomaly degree of each channel window by normalizing the noise floor, determine the event window, and output the event data packet.

[0029] Each frame in the multi-source data frame stream is verified. Specifically, the payload byte sequence of the frame is extracted and concatenated with the fixed fields of the frame header in order to form a byte stream to be verified. The CRC16-CCITT polynomial operation is performed on the byte stream to be verified to obtain the calculated check value. The check value is compared bit by bit with the check code carried in the frame header. If they do not match, the frame verification is determined to be failed and the frame is discarded directly. If they match, the frame verification is determined to be passed and the frame is written into the concatenation buffer of the channel. For frames that pass the CRC check, the order consistency check is performed. If two adjacent frames in the same channel are not consecutive or not increasing, a missing marker is written into the concatenation buffer for the corresponding time interval.

[0030] With a fixed window length and sliding step size, a candidate window sequence is generated based on a unified timestamp. For each candidate window, the corresponding data segment is extracted from the continuous sequence of each channel to form the original window fragment. The missing ratio of the window is calculated by missing markers. When the missing ratio is greater than the missing threshold, the window is marked as having too many missing data.

[0031] It should be noted that the missing threshold is determined by the empirical distribution of the proportion of missing items in the continuous background window at the beginning of the task, and the 0.99 quantile of the empirical distribution is taken as the missing threshold.

[0032] For the partial discharge-related channels, frequency band preprocessing is performed within the window. Specifically, a set of fixed-order FIR bandpass filter coefficients is preset for each type of partial discharge-related channel. The original segments of the window are convolved within the window to obtain the bandpass output. The infrared and acoustic pattern array channels are not bandpassed. Only the image frames and array frames after time stamp alignment are retained as auxiliary fields. Saturation detection is performed for each partial discharge-related channel within the window. If the proportion of samples whose sampled values ​​reach the upper limit of the range exceeds the saturation threshold, they are marked as saturated.

[0033] It should be noted that the preset set of FIR bandpass filter coefficients with fixed order is determined by using the window function method to determine the filter order and obtain the impulse response coefficients based on the predetermined passband and stopband parameters of each channel; the saturation threshold is determined by statistically analyzing the empirical distribution of the proportion of samples near the upper / lower limits within a continuous background window, and the 0.99 quantile of the empirical distribution is taken as the saturation threshold.

[0034] For each partial discharge-related channel, maintain a noise floor. Take the median window energy of a fixed number (e.g., 120) windows before the start of the task. For each candidate window and each partial discharge-related channel, calculate the anomaly, expressed as:

[0035]

[0036] in, Indicates channel In the Window energy value within a window, Indicates channel In the Anomaly level within a window, Indicates the first The start timestamp of each window. Indicates the window length threshold. Indicates the bandpass output sequence. Indicates channel The noise floor threshold is denoted by ∇, which represents a small constant.

[0037] It should be noted that, The method involves continuously reading several verified data frames from each partial discharge-related channel, taking the difference between the start timestamps of the headers of two adjacent frames as the single-frame time span of the channel, and using the number of data frames included in the reading window as the cumulative result corresponding to the maximum single-frame time span calculated in the partial discharge-related channel as the window length threshold. The process involves calculating the corresponding window energy for each window and forming an energy sample set. The energy sample set is then sorted from smallest to largest, and the energy value corresponding to its 0.50 quantile is taken as the noise floor threshold.

[0038] For each local amplifier (FAA) related channel, an anomaly queue of fixed length (e.g., 240) is maintained. The anomaly at the 0.99 quantile of the anomaly queue is used as the current dynamic trigger threshold for the channel. Anomalies are generated for the current candidate window. Specifically, if the anomaly of a channel within a window is greater than the dynamic trigger threshold and there are no windows with excessive missing values ​​or saturation markings, the channel is marked as anomaly and valid in the current window. Otherwise, it is marked as non-anomaly, thus eliminating windows with excessive missing values ​​or saturation distortion. When at least two types of channels in the AFA related channel are marked as anomaly and valid in a window, the candidate window is determined as an event window. For each time segment determined as an event window, an event data packet is encapsulated. The event data packet includes the task number, window sequence number, window start timestamp, window length, bandpass output sequence, anomaly of each channel, noise floor threshold of each channel, missing value ratio, saturation marking, and auxiliary fields.

[0039] S3. Extract the minimum feature set from the event data packet and calculate the channel quality coefficient. Use a binary classification model to gate the event window, screen out suspected fault events, use a random forest classifier to output the probability distribution of the suspected fault events, and combine the probability distribution with the channel quality coefficient to obtain the evidence packet.

[0040] For each partial discharge-related channel, a minimum feature set is extracted within the event window. Specifically, the anomaly degree within the window is used as the normalized energy feature. Pulse detection is performed on the bandpass output of the event window, and a background amplitude sample set is obtained using background data. The amplitude corresponding to the 0.99 quantile of the background amplitude sample set is taken as the pulse detection threshold. Then, the number of pulse segments exceeding the pulse detection threshold within the event window is counted as the pulse counting feature. Within the channel window, the ratio of the maximum phase bucket count to the total count is used as the aggregation degree to obtain the phase aggregation feature. The ratio of the energy detected as a pulse segment within the window to the total energy of the window is used as the pulse energy proportion feature. The normalized energy feature, pulse counting feature, phase aggregation feature, and pulse energy proportion feature are concatenated in sequence to form a minimum feature set, and the aligned frames of infrared and acoustic text within the window are extracted as auxiliary features.

[0041] It should be noted that the background data is the bandpass output of the corresponding channel within a window that is not marked as an event window.

[0042] For each partial discharge-related channel, calculate the channel quality factor within the event window. The expression is:

[0043] in, Indicates the channel quality factor. Indicates saturation marker, Indicates the percentage of missing items. Indicates the missing threshold. Indicates the degree of abnormality.

[0044] The minimum feature set is input into the binary classification model. When the binary classification model outputs a suspected fault, the current event window is marked as a suspected fault event. Based on the screened suspected fault events, multi-class inference is performed on each partial discharge related channel. Specifically, the feature vector of the channel is input into the random forest classifier. The random forest classifier outputs the voting ratio of the channel for each fault type in the fault type set. The voting ratio is used as the basic probability output of the channel for each fault type to obtain the probability distribution. When the binary classification model outputs a non-fault event, the event window is archived as a non-fault event and multi-class inference for that window is stopped.

[0045] It should be noted that the training process of the binary classification model involves extracting event data packets from historical inspections, extracting the minimum feature set for each event window, labeling samples as "suspected faults" and "non-faults" based on manual review, cleaning all samples (removing samples with excessive missing values ​​or saturated distortion), and dividing the data into training and validation sets. In the training set, the minimum feature set is standardized, and a support vector machine binary classifier is trained. Kernel function parameters and penalty parameters are selected in the validation set with the goal of minimizing both false positive and false negative rates. Finally, the optimal model parameters and standardized parameters are solidified. The training steps of the random forest classifier are as follows: based on the minimum feature set and the inspection conclusions, each sample is labeled as a fault type from a predefined set of fault types. All samples are cleaned and divided into training and validation sets. During the training phase, samples are extracted from the training set for each decision tree using a bootstrap sampling method. At each node, the optimal splitting feature is selected from a random subset of the minimum feature set to grow the tree structure. After training, the validation set is used to evaluate the classification accuracy and class confusion, and the number of trees, node splitting rules, and output method of the random forest are solidified.

[0046] To avoid the random forest classifier outputting false high-confidence evidence when there is uncertainty or the channels are unreliable, evidence quality assignment is calculated, expressed as follows:

[0047] in, Indicates that the random forest classifier is applied to the first... The base probability of each fault type is output. Represents information entropy. Represents the degree of certainty coefficient. Indicates the channel to the first Evidence quality allocation for each fault type This indicates the uncertain quality assignment of the channel to the entire set of fault types.

[0048] For each suspected fault event, an evidence package is constructed. The evidence package includes the channel quality coefficient of each channel, the evidence quality allocation of each channel for each fault type, the uncertainty quality allocation of each channel, and auxiliary features.

[0049] S4. Calculate relevant consistency items through the preprocessed waveform of the event window, calculate rule consistency items based on the preset set of association rules and evidence packages, merge relevant consistency items and rule consistency items to form a consistency index, and perform DS combination after discounting the evidence of each channel with the consistency index and channel quality coefficient, and output the fusion fault distribution, conflict degree and fusion uncertainty.

[0050] The set of channels participating in the fusion is fixed as the partial discharge related channel set, and the fusion order is fixed as "UHF → transient ground voltage → HF → ultrasound". For the preprocessed waveforms within the same event window, a relevant consistency term is calculated, expressed as:

[0051] in, Indicates channel With channel Relevant consistency scores within the event window Indicates channel The continuous-time equivalent representation of the bandpass output sequence within the event window. Indicates channel The continuous-time equivalent representation of the bandpass output sequence within the event window. Indicates the start timestamp of the event window. Indicates the end timestamp of the event window. This indicates the consistency items related to the event window. Indicates the number of channel pairs.

[0052] A set of association rules is defined. Based on channel anomaly-related features, pulse count-related features, phase aggregation-related features, evidence quality allocation, channel quality coefficient, and auxiliary features, each rule is judged, and the "number of hits for supporting rules" and "number of hits for conflicting rules" are counted separately. The rule consistency term is calculated, and the expression is as follows:

[0053] in, This indicates the rule consistency items for the event window. This indicates the number of rules that the event window supports. This indicates the number of conflict rules that the event window satisfies.

[0054] It should be noted that the association rule set is generated by collecting event data packets from historical inspections, producing the minimum features and labels (confirmed fault type or confirmed non-fault) for each event window, discretizing continuous features into a finite number of states (e.g., "high / medium / low", with the boundary point taken as the background sample quantile), converting each event window into an "item set" (e.g., "UHF (ultra-high frequency detection channel) energy = high, TEV (transient ground voltage detection channel) pulse count = high, infrared hotspot = yes, channel quality = high"). The FP-Growth algorithm is used to mine frequent itemsets from these item sets and generate association rules. Rules consistent with the confirmed fault mode are marked as supporting rules, and rules contradicting the confirmed mode are marked as conflicting rules, thus obtaining the association rule set.

[0055] The consistency index is formed by merging relevant consistency terms with rule consistency terms. This consistency index, along with the channel quality coefficient, is used to discount evidence from each channel, automatically transferring the quality of "inconsistent or unreliable" evidence to the overall uncertain quality. The expression is:

[0056] in, Indicators of consistency for event windows Indicates the discount factor. This indicates the allocation of channel evidence quality after the discount. This indicates that the channel quality allocation is uncertain after the discount.

[0057] DS combination is performed on the discounted evidence from each channel in the fusion order, and the degree of conflict is calculated. The expression is:

[0058] in, This indicates the degree of conflict in a single evidence combination process. Indicates the first The intermediate evidence obtained from the sub-combination is for the set Quality allocation, This represents the discounted evidence set of the current channel to be merged. Quality allocation, Represents the empty set. This represents the set corresponding to the combination result. This represents the new intermediate evidence set after completing this combination. Quality allocation, It should be noted that for the current channels to be merged, the discounted quality allocation meets the requirements when... For the first When a single-element set corresponds to a fault type ,when For the complete set of fault types hour, The rest of the set The mass allocation is 0.

[0059] The conflict level obtained from the last combination is used as the overall conflict level of the event window.

[0060] The mass distribution of each fault type is extracted and normalized to obtain the fused fault distribution. Simultaneously, the fused uncertainty is calculated, expressed as:

[0061] in, Indicates the effect of fusion on the first Normalized probabilities for each fault type This represents the fusion uncertainty.

[0062] S5. Determine whether to trigger a retest by integrating the fault distribution, conflict degree, and integration uncertainty, calculate the growth trend quantity, and write it into the test result record.

[0063] For each suspected fault event window, a retest trigger judgment is executed. When the fusion uncertainty is higher than the fusion uncertainty judgment threshold, or the overall conflict degree is higher than the conflict judgment threshold, or the consistency index is lower than the consistency judgment threshold, the credibility of the current fusion result is determined to be insufficient, and a retest is required.

[0064] It should be noted that the fusion uncertainty judgment threshold, conflict judgment threshold, and consistency judgment threshold are determined by selecting event samples that have been confirmed by maintenance conclusions in historical inspections, calculating the fusion uncertainty, overall conflict degree, and consistency index for each sample, and summarizing the fusion uncertainty, overall conflict degree, and consistency index for similar equipment and similar scenarios to form an empirical distribution. The selection criterion is "to minimize the false alarm rate while ensuring that the false alarm rate does not deteriorate". On the verification samples, the candidate quantiles are evaluated one by one, and the corresponding quantile values ​​are solidified as the fusion uncertainty judgment threshold, conflict judgment threshold, and consistency judgment threshold.

[0065] When it is determined that a retest is required, a unique retest action is selected from the preset retest action set and executed. The retest action set includes, but is not limited to, extending the sampling time, increasing the sampling rate of a specific channel, enabling backup or redundant sensors, and adjusting the trigger sensitivity. For each candidate retest action, the amount of reduction in fusion uncertainty that the retest action can bring in similar scenarios is evaluated based on historical retest samples. Combined with the energy consumption cost corresponding to the action, the ratio of the expected reduction in fusion uncertainty after the retest action to the energy consumption cost required for the retest action is used as the unit energy consumption fusion uncertainty reduction index of the retest action. In the same event, only the retest action with the largest unit energy consumption fusion uncertainty reduction index is selected as the unique retest action and immediately issued for execution.

[0066] It should be noted that the expected reduction in fusion uncertainty is obtained from the statistical median of the difference in fusion uncertainty before and after retesting in similar historical events; the energy cost is obtained by recording the start and end times of the retesting action at the same time as issuing the retesting action, and reading the instantaneous power consumption data reported by the acquisition host at fixed intervals within the time period, and accumulating the instantaneous power consumption data within the time period to obtain the energy cost of the retesting action.

[0067] After the retest is completed, the corresponding retest data is received, and the fusion uncertainty, overall conflict degree, and consistency index are recalculated. When the fusion uncertainty, overall conflict degree, and consistency index are all below their respective judgment thresholds, the fault growth trend quantity is calculated, expressed as:

[0068] in, This indicates the quantity representing the growth trend of faults. express and The total duration between.

[0069] The fault growth trend and the fused fault distribution are written into the detection result record for operation and maintenance personnel to judge the degree of fault evolution.

[0070] In summary, this invention achieves pre-filtering of a large number of non-fault events and probabilistic representation of fault type output by inputting the minimum feature set into a binary classification model for gating and screening out suspected fault events, and then generating an evidence package by combining the output probability distribution of a random forest with the channel quality coefficients. By performing FIR bandpass convolution on the partial discharge related channels within a window to obtain bandpass output and performing saturation detection, and setting a saturation mark when the saturation ratio exceeds the saturation threshold, this invention achieves enhancement of target frequency band features and isolation of saturated distortion data. By performing CRC consistency verification frame by frame on multi-source data frame streams and writing missing markers for discontinuous / non-incremental frame order, this invention achieves automatic removal of abnormal frames and traceable labeling of missing intervals.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for multi-source fault detection in power equipment based on data fusion, characterized in that: include, Establish a detection task and synchronously collect monitoring data under the same time reference to generate a multi-source data frame stream with a unified timestamp; Verification and splicing of multi-source data frame streams are performed, and candidate windows are formed by slicing them into fixed windows. Frequency band preprocessing is performed on each channel. The anomaly degree of each channel window is calculated by normalizing the noise floor, the event window is determined, and the event data packet is output. Extract the minimum feature set from the event data packet and calculate the channel quality coefficient. Use a binary classification model to gating the event window and screen out suspected fault events. Use a random forest classifier to output the probability distribution of the suspected fault events and combine the probability distribution with the channel quality coefficient to obtain the evidence packet. The relevant consistency items are calculated by preprocessing the waveform in the event window, and the rule consistency items are calculated based on the preset set of association rules and evidence packages. The relevant consistency items and rule consistency items are fused to form a consistency index. The consistency index and the channel quality coefficient are used to discount the evidence of each channel and then DS combination is performed to output the fusion fault distribution, conflict degree and fusion uncertainty. The determination of whether to trigger a retest is made by integrating the fault distribution, conflict degree, and integration uncertainty, and the growth trend quantity is calculated and written into the test result record.

2. The method for multi-source fault detection of power equipment based on data fusion as described in claim 1, characterized in that: The specific steps for establishing a detection task and synchronously collecting monitoring data under the same time reference to generate a multi-source data frame stream with a unified timestamp are as follows: Establish a detection task and synchronously collect monitoring data under the same time reference, output a synchronous trigger pulse to the partial discharge related channel, and complete the alignment using a two-trigger calibration method; The system continuously samples the ultra-high frequency, high frequency, ultrasonic, and transient voltage channels, and writes a master clock timestamp to each sampling point. For the infrared channel, it records the master clock timestamp corresponding to each acquired image frame and writes it to the image frame header. For the acoustic pattern array channel, it acquires array data with a fixed frame length to obtain a multi-source data frame stream.

3. The method for multi-source fault detection of power equipment based on data fusion as described in claim 2, characterized in that: The process involves verifying and splicing multi-source data frame streams, slicing them into candidate windows using a fixed window structure, and performing frequency band preprocessing on each channel. The specific steps are as follows: Each frame in the multi-source data frame stream is checked for consistency using CRC. If two adjacent frames in the same channel are not consecutive or not increasing, a missing flag is written. With a fixed window length and sliding step size, a candidate window sequence is generated based on a unified timestamp. The missing proportion of the candidate windows is calculated by missing markers. When the missing proportion is greater than the missing threshold, the candidate window is marked as having too many missing elements. Frequency band preprocessing is performed on the partial discharge related channels within the candidate window. Convolution operation is performed on the original segments within the candidate window to obtain the bandpass output. Saturation detection is performed on each partial discharge related channel. If the proportion of samples whose sampled values ​​reach the upper limit of the range exceeds the saturation threshold, they are marked as saturated.

4. The method for multi-source fault detection of power equipment based on data fusion as described in claim 3, characterized in that: The steps involve calculating the anomaly degree of each channel window through noise floor normalization, determining the event window, and outputting event data packets. For each partial discharge-related channel, maintain a noise floor. Take the median of the window energy of a fixed number of windows before the start of the task. For each candidate window and each partial discharge-related channel, calculate the anomaly. For each local amplifier channel, maintain an anomaly queue. When the anomaly of a channel in a candidate window is greater than the dynamic trigger threshold and there are no candidate windows with too many missing or saturated markings, mark the channel as an anomaly and valid in the current window, determine the candidate window as an event window, and encapsulate it into an event data packet.

5. The multi-source fault detection method for power equipment based on data fusion as described in claim 4, characterized in that: The specific steps for extracting the minimum feature set from event data packets and calculating the channel quality coefficient are as follows: Pulse detection is performed on the bandpass output of the event window. The number of pulse segments exceeding the pulse detection threshold within the event window is counted as the pulse counting feature. The ratio of the maximum phase bucket count to the total count within the channel window is used as the degree of aggregation to obtain the phase aggregation feature. The ratio of the energy detected as a pulse segment within the window to the total energy of the window is used as the pulse energy ratio feature. The normalized energy characteristics, pulse count characteristics, phase aggregation characteristics, and pulse energy ratio characteristics are sequentially concatenated into a minimum feature set, and the channel quality coefficient is calculated for each partial discharge-related channel within the event window.

6. The method for multi-source fault detection of power equipment based on data fusion as described in claim 5, characterized in that: The process involves using a binary classification model to gate the event window, filtering out suspected fault events, using a random forest classifier to output a probability distribution for these suspected fault events, and combining this probability distribution with channel quality coefficients to obtain an evidence package. The specific steps are as follows: The minimum feature set is input into the binary classification model. When the binary classification model outputs a suspected fault, the current event window is marked as a suspected fault event. Based on suspected fault events, multi-class inference is performed on each partial discharge-related channel using a random forest classifier to obtain probability distributions. The probability distributions are then combined with the channel quality coefficients to obtain an evidence package.

7. The method for multi-source fault detection of power equipment based on data fusion as described in claim 6, characterized in that: The specific steps for calculating relevant consistency terms using the preprocessed waveform of the event window, and calculating rule consistency terms based on a preset set of association rules and evidence packages, are as follows: Calculate relevant consistency terms for preprocessed waveforms within the same event window; Set up a set of association rules, and judge each rule of the minimum feature set based on the evidence package, and count the number of "supporting rule hits" and "conflicting rule hits" respectively, and calculate the rule consistency item.

8. The method for multi-source fault detection of power equipment based on data fusion as described in claim 7, characterized in that: The process involves fusing relevant consistency items with rule consistency items to form a consistency index. Then, after discounting the evidence from each channel using the consistency index and channel quality coefficient, a Data Synthesis (DS) combination is performed to output the fused fault distribution, conflict degree, and fusion uncertainty. The specific steps are as follows: The relevant consistency items are merged with the rule consistency items to form a consistency index. The consistency index and the channel quality coefficient are used to discount the evidence of each channel to obtain the quality allocation. Perform DS combination on the discounted evidence for each channel, calculate the conflict degree, and use the conflict degree obtained from the last combination as the overall conflict degree of the event window; The mass distribution of each fault type is extracted and normalized to obtain the fused fault distribution, and the fused uncertainty is calculated at the same time.

9. The method for multi-source fault detection of power equipment based on data fusion as described in claim 8, characterized in that: The specific steps for determining whether to trigger a retest by fusing fault distribution, conflict degree, and fusion uncertainty are as follows: For each suspected fault event window, a retest trigger judgment is executed. When the fusion uncertainty is higher than the fusion uncertainty judgment threshold, or the overall conflict degree is higher than the conflict judgment threshold, or the consistency index is lower than the consistency judgment threshold, the credibility of the current fusion result is determined to be insufficient and a retest is required. When it is determined that a retest is required, select a unique retest action and execute it, calculate the unit energy consumption fusion uncertainty reduction index, and select the retest action with the largest unit energy consumption fusion uncertainty reduction index as the unique retest action.

10. The method for multi-source fault detection of power equipment based on data fusion as described in claim 9, characterized in that: The specific steps for calculating the growth trend and writing it into the detection result record are as follows: After the retest is completed, the corresponding retest data is received, and the fusion uncertainty, overall conflict degree and consistency index are recalculated. When the fusion uncertainty, overall conflict degree and consistency index are all lower than their respective judgment thresholds, the fault growth trend quantity is calculated, and the fault growth trend quantity and fusion fault distribution are written into the test result record.

Images