LED short circuit fault waveform semantic primitive feature extraction and identification method

Abstract

The application discloses an LED short-circuit fault waveform semantic primitive feature extraction and identification method and concretely relates to the field of fault monitoring, which comprises the following steps: firstly, extracting three-dimensional features of a current waveform offline, generating semantic primitives and setting primitive distance, combining a fault evolution link to build a template library; collecting current signals at the edge, performing hardware fault bottom protection, extracting features to generate time series data and transmitting to the cloud; the cloud relies on semantic primitives to complete sequence conversion and smoothing, realizes adaptive anomaly detection through posteriori entropy, completes template matching by combining a dynamic time warping algorithm after detecting anomalies, divides three operation modes and processes them separately; for different modes, report generation, parameter updating, fault tracing and other operations are carried out, and new samples and operation and maintenance feedback are combined to merge similar templates, clean up inefficient templates, and iteratively optimize the knowledge base combined with cross-scene migration; the scheme realizes intelligent identification and closed-loop optimization of the whole process of faults, has high identification precision and strong practicability.

Description

Technical Field

[0001] This invention relates to the field of fault monitoring technology, and more specifically, to a method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms. Background Technology

[0002] When a short circuit fault occurs in an LED driver circuit, its current waveform will usually produce specific time-domain distortion, and the frequency-domain energy distribution will also shift accordingly. Acquisition and waveform analysis based on current signals are common technical approaches for fault detection.

[0003] In field applications, current signals are generally digitally acquired through high-speed analog-to-digital converters. A sliding window is used to continuously capture the waveform and extract basic measurement parameters such as amplitude changes and frequency band energy. Some systems also support local storage and offline retesting of waveform data. Existing methods often extract time-frequency features such as short-time energy, bandpass filter energy ratio, and peak factor from current waveforms, setting fixed thresholds for each feature; exceeding these thresholds indicates an anomaly. Other approaches pre-establish a fault waveform template library, aligning the measured waveform with the templates in the library in terms of amplitude and matching key points, relying on similarity calculations to determine the fault type. To reduce the interference of random noise on measurement accuracy, some systems introduce multi-cycle averaging or synchronous sampling strategies to improve the stability of feature extraction.

[0004] However, it still has some drawbacks in practical use, such as: 1. Currently, the extracted feature types are limited and the information dimensions are relatively simple. Many subtle changes in current during the development of short-circuit faults are difficult to record completely. In particular, the early fault signals are relatively weak, and the measurement sensitivity often cannot keep up, making it easy to miss the early signs. 2. The front end lacks the hardware-level rapid discrimination capability based on instantaneous current amplitude and rate of change. When a fault occurs, the data can only be sent back to the back end for analysis. The circuit cannot be cut off on the spot immediately. The slow response speed is one aspect, and the data acquisition sensor is also at risk of damage from being subjected to overload current for a long time. 3. Once the template library is built, it is fixed and lacks a mechanism for dynamic calibration based on actual operating data. Aging of the drive circuit or changes in ambient temperature will cause the waveform baseline to gradually shift. As the fixed template is used for a long time, the matching accuracy will naturally decrease, and the number of misjudgments will increase. Moreover, similar redundant templates are not cleaned up for a long time, which also affects the matching efficiency. 4. When performing root cause analysis on abnormal waveforms, the analysis did not first distinguish whether there were obvious short circuit characteristics detected in the current data. Even if the deviation was not caused by a short circuit, the logic of short circuit location was applied. The analysis results often led the cause astray, and the measured data could not play its due role when it was really necessary to trace the source. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides a method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms, which solves the problems mentioned in the background art through the following scheme.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms, comprising: S1. Semantic primitive learning and evolution template library construction: Three-dimensional features are extracted offline from LED current waveforms, semantic primitives are generated through Gaussian mixture model and primitive distance is set, and a template library is built based on the fault evolution link; S2, Edge Computing and Upload: The edge end collects LED current signals and implements hardware fault fallback protection. It extracts features in real time with reference to the feature form of S1, generates time-series data, and uploads it to the cloud. S3, Cloud-based Lightweight Processing and Matching: Receives time-series data uploaded by S2, combines it with S1 semantic primitives to complete sequence transformation and smoothing, and performs adaptive anomaly detection based on posterior entropy; after detecting anomalies, it calls the S1 template library to complete dynamic time warping matching, classifies the patterns into three categories—known, fuzzy, and unknown—and distributes them accordingly. S4, Result Output, Calibration and Quantile Smoothing: Generate a structured report based on the known patterns output by S3, update the template weights and remaining time quantiles based on operation and maintenance feedback, and optimize the template library matching effect; S5. Root Cause Analysis and Backup: Root cause analysis is performed on the fuzzy pattern output by S3. When a short circuit event exists, mutual information is used to locate the abnormal primitives. When there is no short circuit event, the KL divergence comparison is used to complete the fault source tracing. S6. Incremental Learning and Template Lifecycle Management: New templates are added using unknown pattern samples from S3 and operational feedback. Similar templates are merged, inefficient templates are cleaned up, and cross-scenario migration is achieved by adjusting parameters. The S1 knowledge base is iteratively optimized to form a closed loop throughout the entire process.

[0007] The technical effects and advantages of this invention are as follows: 1. This scheme extracts the short-time energy, bandpass filter energy ratio, and peak factor of the current waveform to form three-dimensional features. Semantic primitives are generated by Gaussian mixture model clustering, which effectively enriches the representation dimension of fault information, can capture the subtle distortion of current during short-circuit evolution, and significantly improves the detection sensitivity of early weak fault signals.

[0008] 2. This solution simultaneously completes current signal acquisition and hardware fault fallback protection at the edge. It enables rapid disconnection of faulty branches based on instantaneous current amplitude and rate of change, without waiting for data feedback and analysis. It provides timely response and effectively avoids the problem of damage to acquisition sensors due to prolonged overload.

[0009] 3. This solution establishes a full lifecycle management mechanism for the template library, which can automatically merge similar templates and clean up inefficient templates based on on-site test data and operation and maintenance feedback, thereby realizing dynamic calibration and iterative optimization of the template library and effectively overcoming the matching inaccuracy problem caused by driver aging or environmental changes in fixed templates.

[0010] 4. This solution sets up differentiated tracing logic for different abnormal scenarios. When a short-circuit event exists, mutual information is used to locate the abnormal primitive. When there is no short-circuit event, the KL divergence comparison is used to complete the fault tracing. The two types of scenarios have clear division of labor, which greatly improves the accuracy of root cause location. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0012] Figure 2 This is a schematic diagram of the L1 edge real-time protection and feature upload process of the present invention.

[0013] Figure 3 This is a schematic diagram of the L2 first-level anomaly detection process of the present invention.

[0014] Figure 4 This is a schematic diagram of the second-level template matching and result branching process of the present invention.

[0015] Figure 5 This is a schematic diagram of the root cause analysis and incremental learning closed-loop process of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] refer to Figure 1 - Figure 5 The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms, as shown, includes: This solution consists of an L1 edge real-time protection layer and an L2 cloud analysis layer, asynchronously decoupled. L1 is responsible for high-frequency data acquisition, fallback protection, and feature uploading, while L2 is responsible for semantic primitive modeling, anomaly detection, template matching, and root cause analysis. The collaboration logic between the two layers is as follows: In the offline phase (S1), semantic primitives and fault template libraries are first trained using simulation and real data to provide a knowledge base for subsequent online identification.

[0018] During operation, L1 continuously collects current (S2), calculates fixed-dimensional features, and reliably uploads them to L2.

[0019] After receiving the feature stream, L2 calculates the maximum a posteriori primitive and filters it in real time (S3), and performs the first-level anomaly detection based on the entropy change rate; once an anomaly is triggered, the second-level DTW template matching is immediately executed, and the results such as the fault type and remaining time quantile are output (S4).

[0020] When a fuzzy pattern appears, proceed to root cause analysis (S5) and use mutual information or KL divergence to locate problem primitives.

[0021] The entire system continuously optimizes the template library through operation and maintenance feedback and incremental learning (S6) to achieve closed-loop iteration.

[0022] The specific steps are as follows: S1: Semantic Primitive Learning and Evolution Template Library Construction (Offline): First, unified three-dimensional features are extracted through simulation and real data acquisition. Then, GMM clustering is used to discretize the continuous waveform space into a finite number of semantic primitives with physical meaning, and the regularized Mahalanobis distance between primitives is defined. Finally, the complete fault evolution chain is stored as a template, with fault type, typical duration, primitive sequence and progress-remaining time quantile table. The output of this step (GMM parameters and template library) will be used as fixed knowledge in L2 processing in S3 and will not be updated during runtime, but can be extended through incremental learning in S6.

[0023] S101: Multi-source data acquisition and feature extraction: Simulation data is generated by building a typical LED driver circuit using SPICE or MATLAB / Simulink, and faults such as wire bending, connector oxidation, and insulation aging are injected. The sampling rate is 1MHz and the total duration is ≥1000 hours.

[0024] Real data was collected from pilot projects, with ≥10 data points for each type of fault, for subsequent fine-tuning and verification.

[0025] Fixed feature set ,in: To normalize the short-time energy, an exponentially weighted moving average (EWMA) is used for normalization, with the denominator being the total energy of the original signal, and the dimension being 1.

[0026] The bandpass filter energy ratio is defined as follows: Similarly, EWMA normalization is used. , Rated current, in amperes (A), to prevent division by zero; As the peak factor, , where i is the current sample (in A). The window length is 1 (1000 points in normal mode, 500 points in downsampling mode).

[0027] The bandpass filter adopts a second-order Butterworth bandpass (passband 1kHz-10kHz), direct type II structure, and sampling rate of 1MHz (the coefficients need to be recalculated for the downsampling mode corresponding to 500kHz); the filtering is run once per window, and the estimated time after fixed-pointing on the Cortex-M4F is 0.1-0.2ms.

[0028] The 3D feature points generated in this step will be input into S102 for clustering.

[0029] S102: GMM Clustering and Semantic Primitives: For a large number of feature points Gaussian Mixture Model (GMM) clustering is performed, and the number of clusters K (typically 8-16) is selected using the Bayesian Information Criterion (BIC) to obtain the mixing coefficients. Mean vector (as semantic embedding), covariance matrix .

[0030] The distance between primitives is defined as the regularized Mahalanobis distance: ,in , This is a regularization parameter to enhance numerical stability, and its dimensions are consistent with the covariance. , It represents any two distinct semantic primitives.

[0031] MAP fine-tuning (maximum a posteriori estimation): introducing an equivalent sample size =1000, update parameters using a small amount of real data: ;in, This represents the number of samples belonging to cluster j in the actual data. The mean of these samples, It is the initial mean vector obtained by training the j-th semantic primitive on pure simulation data; the covariance matrix is ​​updated similarly.

[0032] Output GMM parameters and the distance function between primitives It is used for building template libraries in S103 and for real-time decoding in S3.

[0033] S103: Construction of the Evolutionary Template Library: Store each complete fault evolution chain (a continuous sequence of semantic primitives) as a template T; each template contains: Fault type (String identifier); Nominal duration The unit is seconds, where The feature point interval is 1ms in normal mode and 2ms in downsampling mode; primitive sequence Weight (Initial 1.0, dimensionless); Progress-remaining time quantile table (using t-digest data structure, progress is bucketed in 5% increments, for a total of 20 intervals); Last match time (Unix timestamp, updated each time the template is output as a known pattern).

[0034] The time interval of the template library must be consistent with the feature extraction interval of L1; if the downsampling mode (2ms interval) is used, the templates of the corresponding intervals need to be regenerated from the simulation data, or the 1ms templates are uniformly downsampled (taking the mode of every 2 points); the normal mode and the downsampling mode can be chosen and cannot be mixed.

[0035] The output template library serves as the baseline set for DTW matching in S3, and will also be incrementally updated and cleaned up in S6.

[0036] S2: Edge Computing and Upload (L1): The current is detected under three conditions (current peak, adaptive di / dt, lower limit calibration) to achieve 5μs-level hardware cutoff; then, three-dimensional features are calculated at fixed intervals and reliably uploaded to L2 via UDP; at the same time, a circular buffer is maintained to support data backtracking and retransmission; the output of this step is a continuous stream of timestamped feature points, which is the only data source for all analyses in L2.

[0037] S201: Safety Net: Triggering Condition: Both of the following conditions must be met simultaneously: Condition 1: Peak current of the sliding window ; Condition 2: Where di / dt is in A / s, This represents the upper limit of the normal operating current change rate. It is the minimum rate of change of current at which a short-circuit fault can be identified.

[0038] di / dt calculation: First, perform a first-order low-pass filter on the current (cutoff frequency 10kHz), and then use a 5-point average slope to suppress noise.

[0039] Update: The 99.9th percentile of the absolute value of di / dt over the past 10 seconds was estimated online using t-digest (compression parameter 100 centroids).

[0040] L1 maintains an "exemption list" (recording up to the most recent 100 triggers), excluding data points within 50ms before and after the trigger time, and discarding old records in FIFO order when the capacity is exceeded.

[0041] For each product, metal short-circuit and high-resistance short-circuit tests were performed separately, and 0.5 times the minimum di / dt of each type was taken as the lower limit, in A / s.

[0042] The fallback protection is independent of the subsequent upload logic, ensuring that the faulty branch can be cut off within 5μs under any circumstances without affecting feature acquisition.

[0043] S202: Feature Extraction (Compatible Downsampling): Normal Mode: Sampling Rate Window length =1000 points, interval =1ms, EWMA coefficient =0.999 (time constant approximately 1s).

[0044] Downsampling mode: , =500 points, =2ms, =0.998, and the filter coefficients are adjusted accordingly.

[0045] All timing parameters (sliding window length, threshold time window, etc.) vary with Scaling proportionally.

[0046] Each feature point generated in this step carries a microsecond-level acquisition timestamp and sequence number, providing a basis for subsequent L2 time alignment and error tolerance.

[0047] S203: Upload and Reliable Transmission: Transmission Protocol: UDP, one message per feature point, containing a sequence number (16-bit circular), acquisition timestamp (microsecond level), and three feature values.

[0048] Circular buffer: capacity for 2 seconds of data (2000 points in normal mode, 1000 points in downsampling mode).

[0049] Batch ACK: L2 sends an acknowledgment interval every 100ms; L1 maintains the range of confirmed sequence numbers, and cancels the retransmission if an ACK has been received when a timeout occurs.

[0050] Dynamic timeout: L1 estimated round-trip time (RTT), retransmission timeout = 2 * RTT (minimum 200ms), maximum of 3 retransmissions.

[0051] Loss tolerance error: After L2 detects a missing sequence number, it does not actively send a retransmission request, but relies on L1's timeout retransmission. L2 starts a waiting timer, and the waiting time = min(L1 maximum retransmission time + 100ms, 500ms), where L1 maximum retransmission time = 2*RTT*3 (if RTT<50ms, it is counted as 50ms).

[0052] If the data is not received after the timeout, it is considered lost. Missing rate = number of missing points / total number of points in the current 1-second window; if 5 consecutive points are missing or the missing rate is 20%, the data segment is discarded; otherwise, linear interpolation is used to complete the feature values ​​(boundary missing: the next valid point is used at the beginning, and the previous valid point is used at the end), and the confidence level is reduced (multiplied by 0.9).

[0053] Deduplication: L2 maintains a sliding window of the most recent 2000 sequence numbers for each module and discards duplicate messages.

[0054] L1 restart process: After L1 restarts, it sends three special messages (sequence number = -1) consecutively, with an interval of 10ms. Upon receiving any one of these messages, L2 clears the module's buffer and state, and restarts.

[0055] State persistence: L2 stores the cold start phase flag and collected statistics for each module every 10 seconds. baseline and sensitivity multiplier Write to Redis. When the master node fails, the standby node takes over and first reads the most recent state from Redis. If it is missing, it resets the state (only losing a small amount of history).

[0056] L1 does not calculate additional changes: L1 is only responsible for data collection, feature calculation and uploading, and does not perform local change detection, so as to reduce the computational burden on the edge side.

[0057] Through the aforementioned reliable transmission mechanism, L2 can obtain nearly complete feature time series, providing high-quality data for subsequent anomaly detection and matching.

[0058] S3: Lightweight processing and matching in the cloud (L2): It receives the feature stream uploaded by S2 in real time and performs online analysis using the knowledge base built by S1. First, it achieves module-level parallel processing through horizontal sharding and high availability design (S301). Then, it calculates the maximum a posteriori primitive for each feature point and applies mode filtering to obtain a smooth primitive sequence (S302). Next, it performs first-level anomaly detection based on a sliding window of a posteriori entropy, with dynamic thresholds adaptively adjusted according to statistics and including a cold start rollback mechanism (S303). Once an anomaly is detected, it immediately performs band-limited DTW template matching on the primitive sequence of the most recent second (S304). Finally, it classifies the results into three modes—known, fuzzy, or unknown—based on the matching score and score ratio, and then enters the subsequent output, root cause analysis, or fingerprint deduplication process respectively (S305).

[0059] S301: Horizontal sharding and high availability (stateful persistence): Module IDs are assigned to different shards via hashing, and each shard is processed independently with a capacity of ≤2000 modules.

[0060] In master-slave mode, the state is persisted to Redis (snapshot every 10 seconds); the state is restored from Redis when the standby node takes over; if there is no snapshot, it is reset (short-term accuracy degradation is acceptable); this design ensures the system's horizontal scalability as the number of modules increases, as well as rapid recovery in case of master node failure.

[0061] S302: Maximum a posteriori primitive at each point and sliding window mode filtering (no filtering at the boundaries): For each feature point Calculate the posterior probability: ;in Let be the Gaussian probability density function, j be the index of the semantic primitive currently being computed (the j-th Gaussian component of the GMM), and k be the summation dummy variable (index) for all semantic primitives. The maximum a posteriori primitive is taken. .

[0062] Apply sliding window mode filtering (window length 3 points) to the output primitive sequence: the primitive with the highest frequency in the window is taken as the output, and if they are the same frequency, the most recently appearing primitive is taken.

[0063] Boundary processing: The first two points and the last two points of the sequence are not filtered, and the original maximum a posteriori primitives are directly retained.

[0064] This step transforms the continuous feature stream into a discrete primitive sequence, preparing a symbolic time series for subsequent entropy calculation and DTW matching.

[0065] S303: Level 1 Anomaly Detection (based on posterior entropy, dynamic threshold, cold start, and anomaly freeze): For each feature point... Calculate the posterior entropy The dimension is nat.

[0066] Maintain the mean of the sliding window entropy (Window length W corresponds to 200ms: 200 points in normal mode, 100 points in downsampling mode).

[0067] Maintain historical baseline (Exponentially weighted moving average, time constant 60 seconds), initial value is taken from the first window. .

[0068] Define the relative rate of change of entropy at time t. Dimensionless (Dimensionless).

[0069] Dynamic threshold update: maintenance mean and variance Using EWMA (time constant 60 seconds), only... Update when no exception is triggered, freeze during the exception period; initial. =0, =1 (based on simulation priors).

[0070] Sensitivity multiplier Initially set to 1, dynamic threshold The lower limit is 1.5.

[0071] Cold start and smooth transition (including rollback mechanism): Fixed threshold =2.0, used for the first M=1000 feature points, while collecting statistics. , and baseline .

[0072] If detected within the first M points This will completely reset the cold start: clear collected statistics, reset... The counter is reset to zero, and the collection of 1000 normal points will begin again (continued). Set a maximum number of resets to 3. If this number is exceeded, the current statistic (even if it is less than 1000 points) will be used to enter the transition period, and the system will be marked as "incomplete cold start".

[0073] Starting from point M+1, a linear interpolation transition is used: mixing coefficients The threshold decreases linearly from 1 to 0, and after M points, a dynamic threshold is used entirely; transition period threshold. .

[0074] when When this occurs, it is marked as an anomaly, triggering a second-level matching process.

[0075] This first-level detection filters out suspicious time periods with extremely low computational cost (only posterior entropy and sliding window calculations are required), avoiding the time-consuming DTW matching for each time point.

[0076] S304: Second-level template matching (DTW, with Sakoe-Chiba band restriction): matching the primitive sequence most recent 1 second after the trigger time. (Length N, 1000 points in normal mode, 500 points in downsampling mode) and each template T (primitive sequence) in the template library ,length DTW matching is performed, using regularized Mahalanobis distance as the distance function. (See definition in S102). With the Sakoe-Chiba band constraint added, the bandwidth r = 0.15 * max(N, ).

[0077] Match score: ;in =2.0 (dimensionless attenuation factor). The average Mahalanobis distance between different primitive pairs (during training, all primitive pairs are traversed). After calculation, take the average value; if the result is less than 0.1, take 0.1 to prevent division by zero. Use the data collection timestamp (in seconds). If it is unavailable, roll back to the previous time. ; ;t first t last These are the start time and end time of the exception, respectively. The time interval for a single feature point; the template library size is 50-200, and the DTW of all templates is calculated directly (complexity O(N*r), which is acceptable); the 3 templates with the highest scores are selected.

[0078] This step is only executed when an anomaly is triggered, avoiding continuous high-load computation, while further reducing DTW overhead through band limiting.

[0079] S305: Branching (Quantization Definition): Known Pattern: Highest score ≥ 0.8 and (second highest score does not exist or highest score / second highest score ≥ 1.2); Output fault type and observation progress. Retrieve the remaining time quantiles from the quantile table of the template; extrapolation rules: If the value is less than 5%, a warning will be issued and a 1.5 times 5% interval quantile will be provided as a reference. Use the 95% range value when >95%; switch to S4.

[0080] Fuzzy mode: The known mode conditions are not met, and the highest score is ≥0.6 or the ratio of the highest score to the second highest score is <1.2; when the cumulative number of fuzzy occurrences reaches 3 (or 5 consecutive occurrences), root cause analysis (S5) is triggered, and the sensitivity multiplier of this module is increased. Multiply by 0.8 (corresponding to the lower limit) >1.2), reset the recovery timer and fuzzy count counter; if no fuzzy pattern appears for one consecutive hour, gradually restore. (Multiply by 1.05 every 10 minutes until 1); If a blurry pattern reappears during recovery, immediately stop the recovery and reduce the speed again. .

[0081] Unknown mode: Does not meet the above conditions (highest score < 0.6 and highest score / second highest score ≥ 1.2); Precise fingerprint deduplication: Calculate MD5 (primitive sequence + statistical feature encoding), if the hash has not appeared before, push it to the operation and maintenance interface and record it; Optional approximate deduplication (normalized DTW distance threshold 0.2), which is off by default.

[0082] Through three levels of branches, the system can handle situations with high determinism, high uncertainty, and complete unknownness in different ways, guide fuzzy patterns to root cause analysis, and push unknown patterns to manual annotation, forming a closed loop.

[0083] S4: Result Output, Calibration and Quantile Smoothing: This step transforms the recognition results generated in S3 into a deliverable structured report, and uses operational feedback to self-calibrate the template weights while updating the remaining time quantile statistics; by persisting matching records, providing feedback interfaces and incremental quantile updates, the template library can adapt to the evolution of actual working conditions.

[0084] S401: Structured Report Generation: Outputs a JSON-formatted report containing: Module ID, timestamp, recognition result (known / fuzzy / unknown), fault type, confidence level (highest score), remaining time quantile (or warning information), and feature trajectory summary.

[0085] Feature Trajectory Summary: Outputs the turning points of the primitive sequence changes in the most recent second, in the format: (time offset, original primitive, new primitive).

[0086] S402: Confidence self-calibration and operation and maintenance feedback interface: persistent matching records: L2 storage for each matching event (retained for at least one week), including: matching ID, module ID, timestamp, observation sequence hash, the first three templates and their scores, and the final judgment result.

[0087] Operations and maintenance feedback interface: Provides API or UI for operations and maintenance personnel to annotate matching records (correct / incorrect / partially correct) and specify the actual fault type; the system updates the TP / FP count of the template based on the feedback.

[0088] Weight update (weekly offline execution): For each template T, count the number of times it was marked as correct (TP) and the number of times it was marked as incorrect (FP) within the week, and use an exponentially weighted moving average: ;in =0.1 (learning rate), upper limit of weight 1.0, lower limit 0.3, The weights before template T is updated are historical weights retained from the previous iteration. The new weight value after template T is updated is the result of this round of calculation, which will overwrite the old weight and participate in subsequent matching.

[0089] The final score during matching is .

[0090] S403: Online Update of Remaining Time Quantiles: Each template maintains 20 progress intervals (every 5%), and each interval uses a t-digest data structure to store the 25%, 50%, and 75% quantiles. When the number of samples in a certain interval is less than 20, it reverts to the global quantile of that template (i.e., without considering progress).

[0091] Through the above mechanism, the system can continuously learn from human feedback, improve the rationality of template weights, and make the remaining time estimation more accurate.

[0092] S5: Root Cause Analysis and Backup Plan When a fuzzy pattern appears in S3 (accumulated to the threshold) or when maintenance actively requests it, this step initiates root cause analysis. If a short-circuit event exists, mutual information (MI) is used to analyze the correlation strength between primitives and healthy primitives in the short window before the short circuit, and to locate the primitive most likely to cause the short circuit. If there is no short-circuit event, the difference between the module's recent primitive distribution and the healthy baseline distribution is compared using KL divergence, and the three primitives with the largest deviations are output as abnormal signs.

[0093] S501: Triggering condition: The cumulative number of fuzzy patterns reaches 3 times, or the operation and maintenance personnel actively request it.

[0094] S502: Mutual information calculation (when a short-circuit event exists): The short-circuit moment is defined as the sampling point preceding the trigger moment of the L1 fallback protection; Positive samples: Only primitives within 10ms before the short circuit occurs are taken (10 points in normal mode); the number of positive samples must be ≥5; if 3≤positive samples<5, calculate mutual information but mark "low confidence"; if <3, output "insufficient data"; Negative samples: Select healthy time periods (quantification conditions: <0.1*lnK and The coefficient of variation is <0.1 (based on a 200ms sliding window) and Random primitives in (mean > 0.5); if there are insufficient healthy periods resulting in negative samples < 1000, then use all non-short-circuit periods and mark them as low confidence.

[0095] Smoothing: Calculating mutual information Laplace smoothing is used (add 1 to both the numerator and denominator) to avoid zero probability.

[0096] S503: Root cause analysis without short-circuit events: Each module independently maintains the frequency distribution of healthy primitives. During deployment, a normal operating time period (e.g., 10 minutes after power-on) is manually calibrated by operations and maintenance personnel first; if not calibrated, it is automatically extracted: Priority condition: current within ±5% of rated current, and The coefficient of variation is <0.1 and The mean value is greater than 0.5 for 30 consecutive seconds.

[0097] If the above conditions cannot be met (such as in a dimming scenario), then instead extract 30 seconds of data at the current steady-state operating point (current change coefficient < 0.05 for 10 seconds) as a health baseline and mark the corresponding average current percentage.

[0098] Calculate the KL divergence between the primitive frequency distribution of this module over the past 7 days and its own health distribution, and output the 3 primitives with the largest deviation; if there is no historical health baseline, skip the root cause analysis.

[0099] S6: Incremental Learning and Template Lifecycle Management This step is responsible for the dynamic evolution and size control of the template library, enabling the system to adapt to newly emerging failure modes while preventing templates from expanding indefinitely. New templates are generated based on the operations team's annotations of unknown modes (from the unknown mode branches in S3 and feedback from S5). Highly similar templates are periodically merged to reduce redundancy. Templates that have not been matched for a long time and have low accuracy are cleaned up. When the total number of templates exceeds the capacity limit, the least active templates are eliminated based on a comprehensive score. Furthermore, it supports cross-scenario knowledge transfer by shifting the GMM mean. The output of this step is a continuously optimized template library, directly affecting the quality and efficiency of DTW matching in S3.

[0100] S601: New template generation and merging: The initial weight of the new template is 0.6, and the most recent matching time is set to the creation time.

[0101] The merging conditions are as follows (normalized DTW distance is used uniformly). Merge condition 1 <0.2; Merging condition 2: Same fault type; Merging condition 3: Length difference .

[0102] When merging, templates with higher weights are retained, and a weighted average is calculated using the quantile table. The more recent matching time is taken as the newer of the two.

[0103] Template cleanup and capacity control: Clean up templates that meet the following conditions (archived) every month: have not been matched as a known pattern for more than 6 months, that is, have not entered the S305 known pattern branch, have feedback records, and have an accuracy of <0.8.

[0104] Templates with no feedback records (never been matched) are not automatically cleaned up, but the total capacity of the template library is set to a maximum of 200; when this limit is exceeded, the template with the lowest score is eliminated according to the comprehensive score = weight × (normalized value of the most recent matching time) (manual review or automatic archiving by the system is required).

[0105] "Matched with a known pattern" means that the template is selected as the highest score and meets the known pattern conditions and is output to the user. At this time, its most recent matching time is updated to the current time.

[0106] S602: Scene migration (translation mean): Applicable scope: LED driver topologies of source and target domains are similar (e.g., both are buck-constant voltage type).

[0107] The target domain acquires 30 minutes of normal waveform data and calculates ( The mean vector of ) .

[0108] The source domain already has a normal mean. ; Mean shift for each cluster in the GMM: Preserving covariance No change. If the topological difference is large, use MAP fine-tuning (S102) instead of translation.

[0109] Through incremental learning and scenario transfer, the system can quickly adapt to new environments and new failure modes without retraining the entire model, significantly reducing deployment costs.

[0110] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms, characterized in that, include: S1. Semantic primitive learning and evolution template library construction: Three-dimensional features are extracted offline from LED current waveforms, semantic primitives are generated through Gaussian mixture model and primitive distance is set, and a template library is built based on the fault evolution link; S2, Edge Computing and Upload: The edge end collects LED current signals and implements hardware fault fallback protection. It extracts features in real time with reference to the feature form of S1, generates time-series data, and uploads it to the cloud. S3, Cloud-based Lightweight Processing and Matching: Receives time-series data uploaded by S2, combines it with S1 semantic primitives to complete sequence transformation and smoothing, and performs adaptive anomaly detection based on posterior entropy; after detecting anomalies, it calls the S1 template library to complete dynamic time warping matching, classifies the patterns into three categories—known, fuzzy, and unknown—and distributes them accordingly. S4, Result Output, Calibration and Quantile Smoothing: Generate a structured report based on the known patterns output by S3, update the template weights and remaining time quantiles based on operation and maintenance feedback, and optimize the template library matching effect; S5. Root Cause Analysis and Backup: Root cause analysis is performed on the fuzzy pattern output by S3. When a short circuit event exists, mutual information is used to locate the abnormal primitives. When there is no short circuit event, the KL divergence comparison is used to complete the fault source tracing. S6. Incremental Learning and Template Lifecycle Management: New templates are added using unknown pattern samples from S3 and operational feedback. Similar templates are merged, inefficient templates are cleaned up, and cross-scenario migration is achieved by adjusting parameters. The S1 knowledge base is iteratively optimized to form a closed loop throughout the entire process.

2. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: S1 includes: Simulation data and field-measured LED current waveform data are collected, and three types of features—normalized short-time energy, bandpass filter energy ratio, and peak factor—are extracted to form three-dimensional features. All three-dimensional features are input into a Gaussian mixture model to complete clustering. The number of clusters is determined according to the Bayesian information criterion to obtain the semantic primitives of the corresponding clusters. The regularized Mahalanobis distance between different semantic primitives is calculated to complete the primitive distance setting. Finally, a fault template library is built by combining the complete fault evolution link.

3. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: S2 includes: The LED current signal is collected in real time, and the hardware fault fallback protection is triggered based on the current peak value and current change rate judgment conditions to realize the rapid disconnection of the faulty branch; the feature data is extracted window by window according to the feature form of S1, and continuous time sequence data is generated at fixed time intervals and the time sequence data is transmitted to the cloud; the specific content of triggering the hardware fault fallback protection includes: detecting the current peak value and current change rate within the sliding window, and when both indicators exceed the corresponding set threshold, the faulty branch disconnection operation is performed.

4. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: The sequence transformation and smoothing process is completed by combining S1 semantic primitives, and adaptive anomaly detection is performed based on posterior entropy, including: For each point in the received time-series data, the posterior probability is calculated, and the semantic primitive corresponding to the maximum posterior probability is selected to complete the sequence conversion. The converted primitive sequence is smoothed using a sliding window mode method. The posterior entropy corresponding to the primitive is calculated, and the entropy change rate is obtained by combining the sliding window statistics with the historical baseline. Anomalies are determined based on dynamic thresholds and cold start rules to complete adaptive anomaly detection.

5. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: After detecting an anomalies, the S1 template library is invoked to perform dynamic time warping matching, classifying patterns into three categories: known, fuzzy, and unknown, and then routing them accordingly. Extract the primitive sequence corresponding to the abnormal period, call the template library built by S1, and use the dynamic time warping algorithm with set band limit to perform matching one by one and calculate the matching score; compare the matching scores of each template, and divide the results into known mode, fuzzy mode and unknown mode according to the preset judgment rules; send the different modes to the corresponding subsequent processing flow respectively. The calculation of the matching score specifically includes: combining the overall sequence distance, the average distance between primitives, the sequence duration deviation, and the duration distribution parameters obtained by dynamic time warping, performing decay operations through exponential functions respectively, and multiplying the results of the two operations to obtain the final matching score.

6. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: The S4 includes: Generate structured recognition reports for known patterns; statistically analyze the operation and maintenance feedback annotation results, count the number of correct matches and the number of incorrect matches for each template, and iteratively update the template weights using an exponential weighting method combined with the preset learning rate; divide the intervals according to the fault progress, continuously update the remaining time quantiles for each interval using sample data within the interval, and call the global quantiles of the template to supplement when the number of samples is insufficient.

7. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: The method of using mutual information to locate abnormal primitives when a short-circuit event occurs includes: The primitive sequence during the short-circuit event period is used as the observation sequence, and the primitive sequence during the normal period is selected as the reference sequence. The occurrence probability and joint occurrence probability of each primitive in the observation sequence and the reference sequence are statistically analyzed. The mutual information value between each primitive in the observation sequence and the short-circuit event is calculated according to the mutual information formula. The primitives are arranged in descending order of mutual information value, and the primitives that exceed the preset threshold are selected as abnormal primitives to complete the localization.

8. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: When no short-circuit event occurs, fault source tracing is performed using KL divergence comparison distribution differences, including: The frequency distribution of primitives under normal operating conditions is collected and calibrated in advance as a healthy baseline distribution; the frequency distribution of primitives during recent working periods is collected as the distribution to be tested; the KL divergence between the distribution to be tested and the healthy baseline distribution is calculated to quantify the degree of deviation between the two distributions; the primitives are sorted according to their contribution to the distribution deviation, and the primitives with the largest deviation are selected as abnormal symptom primitives to complete the fault tracing in the absence of short circuits.

9. The method for extracting and recognizing semantic primitive features of LED short-circuit fault waveforms according to claim 1, characterized in that: The S6 includes: New templates are generated based on unknown pattern samples output by S3 and operation and maintenance feedback, and these new templates are merged into the existing template library. The primitive sequences and matching features of each template in the template library are compared, and templates with similarity reaching a set threshold are identified and merged. The call frequency and matching accuracy of each template are statistically analyzed, and templates that do not meet the preset standards are identified as inefficient templates and cleaned up. At the same time, parameter adjustments are carried out to achieve cross-scenario migration, and the S1 knowledge base is iteratively optimized to form a closed loop of the entire process.

Images