Vibration feature fusion analysis and artificial labeling closed-loop retraining abnormal early warning method and system

Abstract

This invention relates to an anomaly early warning method and system based on vibration feature fusion analysis and manual annotation closed-loop retraining. The aim is to recover complete waveforms and calculate multi-dimensional time-domain vibration features using multi-loop packet framing technology. Based on this, statistical baseline detection, unsupervised / weakly supervised machine learning detection, and manually annotated supervised learning detection are performed in parallel. Through an innovative fusion rule, the three detection results are integrated to output more robust and interpretable anomaly levels and scores, while establishing a complete closed loop of "collection-analysis-alarm-annotation-retraining". The system saves all intermediate features and provides a manual annotation interface. When historical samples are manually annotated, the system automatically triggers incremental model retraining for the corresponding loop, enabling the model to continuously improve with the accumulation of field knowledge.

Description

Technical Field

[0001] This invention relates to the technical field of industrial equipment condition monitoring and intelligent diagnosis, specifically to an anomaly early warning method and system based on vibration feature fusion analysis and manual annotation closed-loop retraining. Background Technology

[0002] In the field of online monitoring and intelligent operation and maintenance of industrial equipment, vibration analysis is one of the core methods for fault diagnosis. Existing vibration monitoring solutions typically use fixed thresholds or a single model for anomaly detection.

[0003] However, fixed threshold methods struggle to adapt to vibration variations across different devices, loads, and installation locations, resulting in poor generalization. Meanwhile, single-model methods (such as those relying solely on unsupervised learning) are susceptible to insufficient training samples, changes in operating conditions, and environmental noise pollution, leading to high false alarm or false negative rates and severely impacting the on-site maintenance experience.

[0004] Furthermore, most existing systems only focus on real-time alarms and lack the ability to persist waveform data and intermediate features, and to drive continuous model optimization through manual annotation and closed-loop feedback. In industrial scenarios, equipment status will drift over long periods due to component wear and operating condition adjustments. If the model cannot be updated based on new data, its detection performance will inevitably degrade gradually.

[0005] Therefore, there is an urgent need for a vibration anomaly detection scheme that takes into account real-time performance, accuracy, interpretability, and continuous self-evolution. Summary of the Invention

[0006] The present invention proposes a method and system for anomaly early warning based on vibration feature fusion analysis and manual annotation closed-loop retraining, which can at least solve one of the technical problems in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: Based on the vibration feature fusion analysis and manual annotation closed-loop retraining anomaly early warning method, the following steps are performed using computer equipment. S1: Receive a message from the vibration acquisition terminal. The message includes at least the loop number, total number of packets, sub-packet number, sub-packet waveform load, and packet tail. S2: Establish a buffer area based on the circuit number, sort, deduplicate and assemble each sub-packet, and splice together the sub-packets after collecting all sub-packets of the same circuit to obtain a complete waveform load; S3: Analyze the complete waveform load to obtain a vibration sampling value sequence, and extract multiple time-domain vibration features from the sequence; S4: Perform statistical baseline anomaly detection, unsupervised machine learning anomaly detection, and supervised learning anomaly detection respectively, and obtain statistical anomaly score and statistical anomaly level, machine learning confidence and binary label, and supervised learning anomaly probability respectively; S5: According to the preset fusion rules, the multiple detection results obtained in step S4 are fused to obtain the final abnormality level and fusion abnormality score of the target loop. S6: Write the complete waveform load, multiple time-domain vibration characteristics, and the final anomaly level and fused anomaly score into the database and log medium; S7: In response to manual annotation, generate and add to the retraining task queue, and the background task processing module triggers the corresponding loop model to perform incremental retraining according to the task queue.

[0008] An anomaly early warning system based on vibration feature fusion analysis and manual annotation closed-loop retraining includes: The data acquisition and framing module is used to receive the original vibration message. The data acquisition and framing module establishes an independent buffer area according to the circuit number, completes packet sorting, deduplication, and assembly, and outputs a complete waveform load. The waveform analysis and feature extraction module is used to receive the complete waveform load, and its output is connected to the input of the multi-model fusion detection module and the database respectively; the waveform analysis and feature extraction module analyzes the load into a vibration acceleration sequence, calculates multiple time-domain vibration features, and outputs a feature vector; The multi-model fusion detection module is used to receive the feature vector, and its output is connected to the alarm output module and the input of the database respectively. The multi-model fusion detection module internally calls the statistical baseline detection unit, the unsupervised machine learning detection unit and the supervised learning detection unit in parallel, and outputs the final anomaly level and fusion anomaly score according to the preset fusion rules. The alarm output module is used to receive the final anomaly level and the fused anomaly score. Its output end is connected to the database and log medium to persistently write the original waveform, features and anomaly results according to different table structures. The manual annotation interface module is used to receive external annotation instructions. Its output is connected to the input of the retraining task queue module to write the manual annotation results into the database and trigger the generation of retraining tasks. The retraining task queue module is used to manage the status and queuing of retraining tasks, and its output is connected to the input of the model retraining execution module. The model retraining execution module is used to periodically poll the retraining task queue module and read labeled feature data from the database according to the task information. It performs incremental retraining on the supervised learning model and unsupervised learning model of the specified loop, and performs retraining on the unsupervised anomaly detection model and supervised binary classification model of the specified loop respectively. The supervised branch loads samples from the database association table when the configuration is enabled.

[0009] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0010] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0011] As can be seen from the above technical solution, the anomaly early warning method and system based on vibration feature fusion analysis and manual annotation closed-loop retraining of the present invention aims to recover the complete waveform through multi-loop packet framing technology and calculate multi-dimensional time-domain vibration features. Based on this, statistical baseline detection, unsupervised / weakly supervised machine learning detection, and manually labeled supervised learning detection are performed in parallel. Through an innovative fusion rule, the three detection results are integrated to output a more robust and interpretable anomaly level and score.

[0012] Furthermore, this invention establishes a complete closed loop of "collection-analysis-alarm-labeling-retraining". The system saves all intermediate features and provides a manual labeling interface. When historical samples are manually labeled, the system automatically triggers incremental retraining of the corresponding loop, enabling the model to continuously improve as on-site knowledge accumulates. Attached Figure Description

[0013] Figure 1 This is a flowchart of the method of the present invention; Figure 1 This is a framework diagram of the system of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.

[0015] like Figure 1 As shown in the figure, the anomaly early warning method based on vibration feature fusion analysis and manual annotation closed-loop retraining described in this embodiment performs the following steps through a computer device: S1: Receive a message from the vibration acquisition terminal. The message includes at least the loop number, total number of packets, sub-packet number, sub-packet waveform load, and packet tail. S2: Establish a buffer area based on the circuit number, sort, deduplicate and assemble each sub-packet, and splice together the sub-packets after collecting all sub-packets of the same circuit to obtain a complete waveform load; S3: Analyze the complete waveform load to obtain a vibration sampling value sequence, and extract multiple time-domain vibration features from the sequence; S4: Perform statistical baseline anomaly detection, unsupervised machine learning anomaly detection, and supervised learning anomaly detection respectively, and obtain statistical anomaly score and statistical anomaly level, machine learning confidence and binary label, and supervised learning anomaly probability respectively; S5: According to the preset fusion rules, the multiple detection results obtained in step S4 are fused to obtain the final abnormality level and fusion abnormality score of the target loop. S6: Write the complete waveform load, multiple time-domain vibration characteristics, and the final anomaly level and fused anomaly score into the database and log medium; S7: In response to manual annotation, generate and add to the retraining task queue, and the background task processing module triggers the corresponding loop model to perform incremental retraining according to the task queue.

[0016] The specific steps are as follows; Example 1: System Overall Architecture like Figure 2 As shown, the anomaly early warning system based on vibration feature fusion analysis and manual annotation closed-loop retraining provided by the present invention mainly includes: a data acquisition and framing module, a waveform analysis and feature extraction module, a multi-model fusion detection module, an alarm output module, a manual annotation interface module, a retraining task queue module, and a model retraining execution module.

[0017] The data acquisition and framing module is used to execute steps S1 and S2. It is responsible for receiving the original vibration message and completing the reassembly and assembly of the sub-packets according to the circuit.

[0018] The waveform analysis and feature extraction module is used to execute step S3, which is responsible for parsing the assembled byte stream into an acceleration sequence with clear physical meaning and calculating multi-dimensional time-domain features.

[0019] The multi-model fusion detection module is used to execute steps S4 and S5. It is responsible for calling the statistical baseline model, unsupervised machine learning model and supervised learning model in parallel, and outputting the final anomaly level and score according to the preset fusion rules.

[0020] The alarm output module is used to execute step S6 and is responsible for persisting the original waveform, features and abnormal results to different media.

[0021] The manual annotation interface module, the retraining task queue module, and the model retraining execution module work together to execute step S7, which is responsible for receiving external annotation feedback and triggering and executing the model's background incremental retraining.

[0022] The modules work together to form a complete technical solution from data acquisition, real-time analysis, integrated alarms to closed-loop optimization.

[0023] Example 2: Vibration data reception and framing (corresponding to steps S1 and S2) This embodiment describes in detail the specific implementation of steps S1 and S2.

[0024] S1: Message Reception After the data acquisition service starts, it listens on the specified IP address and port. It receives UDP or TCP packets from the vibration acquisition terminal. Each packet follows a predetermined protocol and includes a header, trailer, loop number (LoopID), total number of packets (TotalPackets), packet sequence number (SeqID), and packet waveform payload (Payload). The system first verifies the packet length and CRC to confirm packet integrity.

[0025] S2: Sub-packet assembly and cache management The system uses LoopID as the key to establish a dedicated buffer for each individual monitoring loop. This buffer maintains the following metadata: 1. An ordered dictionary or list for indexing received packet payloads by SeqID.

[0026] 2. ExpectedTotal: The total number of packets expected to be received, extracted from the first arriving message or packet.

[0027] 3. StartTime: The timestamp of the first time this batch of messages was received, used for timeout detection.

[0028] When the new subcontract arrives: 1. Sorting and Deduplication: Place the payload into the corresponding SeqID position in the buffer. If data already exists at that position, discard any duplicate packets that arrive later.

[0029] 2. Timeout Cleanup: The background polling thread checks all caches. If the current time minus StartTime exceeds a preset threshold (e.g., 10 seconds), the batch reception is considered to have timed out, and all caches corresponding to that LoopID are forcibly cleared to prevent memory leaks.

[0030] 3. Integrity Check: After each new packet is inserted, check if the number of received packets equals ExpectedTotal. If they are equal, it means that the batch of packets has been received.

[0031] Once all packets from the same loop have been received, the system concatenates the payloads of all packets sequentially according to their SeqID values, from smallest to largest, to form a complete waveform byte stream. Afterward, the loop buffer is reset, ready to receive the next batch of data.

[0032] If, during the reception process, a change in the TotalPackets field value is detected within the same loop (for example, the first packet indicates the need for 10 packets, while the fifth packet indicates the need for 12 packets), the system determines this as batch disorder and immediately performs a cache reset to prevent cross-batch packet mixing.

[0033] Example 3: Waveform Analysis and Feature Extraction (corresponding to step S3) This embodiment describes in detail the specific implementation of step S3.

[0034] S3-1: Waveform Load Analysis The complete waveform byte stream obtained in Example 2 is parsed in units of 4 bytes each. First, a 32-bit integer is read in little-endian order and interpreted as a single-precision floating-point number according to IEEE 754. Then, combined with the sensor's sensitivity coefficient (e.g., 100mV / g, where g is the gravitational acceleration, taken as 9.8m / s² or 9.81m / s²), it is converted into an engineering value of vibration acceleration using the formula a = value × 1000 / Sensitivity (the default sensitivity coefficient in this example is Sensitivity = 100). The unit can be g or m / s². For invalid or NaN values, the system filters them, and only the valid sampling sequence {x_i} (where i = 1, ..., N, and N is the valid sample length) is processed further.

[0035] S3-2: Calculation of Temporal Feature Vectors To comprehensively characterize the statistical and impact properties of the vibration signal, the following set of time-domain features are calculated for the effective sampling sequence {x_i}: First, calculate the auxiliary statistic: Sample mean: ; Standard deviation: ; Further calculation of core features: effective value This reflects the average energy of the signal; peak This reflects the maximum instantaneous amplitude of the signal; Peak-to-peak value This reflects the fluctuation range of the signal; cliff ( =0 (takes 0 when =0), sensitive to impact pulses, used to detect early faults; Skewness ( =0 is taken as 0), reflecting the asymmetry of signal distribution; Rectified mean ; Pulse factor ; Waveform factor ; margin factor (When the denominator is 0, the corresponding factor is 0). Zero crossing rate (Z represents the number of sign changes, N≥2), reflecting the frequency of signal changes. The above features constitute a multidimensional feature vector, which serves as the unified input for all subsequent detection models.

[0036] Example 4: Multi-model fusion anomaly detection (corresponding to steps S4 and S5) This embodiment is the core of the invention, and it elaborates on the implementation of steps S4 and S5 in detail; S4: Parallel execution of three anomaly detection methods S4-1: Statistical Baseline Anomaly Detection (Statistical Branch) This method maintains a historical baseline statistic for each vibration feature of each loop. For new samples... The eigenvalue v and the baseline establishment and update process are as follows: Baseline establishment: When the system is initialized or a loop is first connected, collect the first M samples (e.g., M=100) that are confirmed as normal by the system or manually, and calculate the baseline mean. and baseline standard deviation The calculation formula is as follows: ; .

[0037] Deviation detection: A unique dual-threshold strategy is used to quantify the degree of deviation. Define moderate deviation from upper and lower limits ; Significant deviation from upper and lower limits ; Judgment logic: like or This characteristic is then recorded as a severe baseline deviation (e.g., 10 points). Otherwise if or This is recorded as moderate baseline deviation (e.g., 5 points).

[0038] Scoring and Ranking: The statistical anomaly score is obtained by summing the fixed bonus for exceeding the hard threshold and the bonus for deviation from the baseline of each feature. =min(100, Then, based on rules such as the hard threshold severity item, the number of moderate / severe baseline deviation features, and the ratio of effective values ​​to the preset severity threshold, they are mapped to levels such as normal, mild, moderate, severe, and extreme.

[0039] Baseline Update: An exponentially weighted moving average is used for updating only when the current period's statistics are determined to be "normal" and the baseline has been initialized. ; variance .

[0040] in, The learning rate is set to 0.05 in this embodiment.

[0041] S4-2: Machine Learning Anomaly Detection (Unsupervised Branch) This method trains an unsupervised anomaly detection model based on random principal component analysis independently for each loop.

[0042] Model Training: Model training is automatically triggered when the number of historical feature samples for a loop reaches a preset threshold (e.g., 100). The training process is label-independent and uses the ML.NET Randomized PCA anomaly detector to learn the main structure of normal samples. After training, the model is persisted as files differentiated by loop.

[0043] Online detection: Input the feature vector of the current sample into the model, and the model outputs an anomaly score. (Between 0 and 1). To integrate with the statistical branch on the same scale, it is mapped to a confidence level. : =max(0, min(100, 100· )).

[0044] At the same time, the model also outputs binary predicted labels (normal / abnormal).

[0045] S4-3: Anomaly Detection in Supervised Learning (Supervised Branch) Prerequisites: This function requires manually labeled data. It will automatically start when the system detects that the number of labeled samples for the current loop has reached the minimum requirement (e.g., 50 samples).

[0046] Model training: Labeled samples are jointly loaded from the feature table and the manually labeled table to construct feature vector-label pairs. After concatenating the features and performing Min-Max normalization, a binary classification model is trained. Online detection: Input the feature vector of the current sample into the trained supervised model, and the model outputs an anomaly probability P∈[0,1]; when P≥τ (0.7 in this embodiment), it is determined to be an anomaly.

[0047] S5: Multi-model fusion and level correction (fusion rules) This invention proposes an innovative, hierarchical fusion rule to comprehensively evaluate the detection results of different branches.

[0048] Rule 1: Dual-channel consistent enhancement If both the statistical branch and the unsupervised branch are judged as "abnormal", it indicates that the anomalous signals are strong and consistent.

[0049] Abnormal level: Based on the highest level of both, increase by one level (e.g., from moderate to severe).

[0050] The fusion score is calculated using a weighted average method, reflecting the design philosophy of "primarily interpretable statistical methods, supplemented by models." = × + × , In this embodiment ; Rule 2: Potential Risk Warning for Single Channels If only the unsupervised branch is judged as "abnormal," while the statistical branch is judged as normal, this usually means that the device may have unknown patterns or early, subtle signs.

[0051] Anomaly Level: Set to "Minor Anomaly" to serve as a notification.

[0052] Fusion score: directly using the confidence level of the unsupervised branch: =C ML Rule 3: Single-channel false alarm suppression If only the statistical branch is judged as "abnormal", while the unsupervised branch is judged as normal, and the confidence level C of the unsupervised branch is... ML Lower.

[0053] Judgment: When C ML When the value is less than 30, the system considers the statistical anomaly to be a false alarm caused by normal fluctuations or noise momentarily hitting the threshold.

[0054] Action: Downgrade the anomaly level by one level (e.g., from severe to moderate) to suppress the interference of false alarms on on-site operations and maintenance.

[0055] Rule 4: Supervised Learning and Secondary Calibration When a supervised learning model is available, its output anomaly probability P is used as a high-level correction signal to perform a secondary correction on the aforementioned fusion result.

[0056] Score Correction: Adjustment of existing fusion scores Weighting: ,and =1; In this embodiment ,in This is the preset second weighting coefficient; The fusion anomaly score before the second correction; Level Correction: If P≥τ (0.7), then the current level is adjusted upward by one level to enhance the sensitivity to complex faults.

[0057] If P≤0.2 and the current level is higher than “slight”, then the level is adjusted down by one level to further reduce false alarms.

[0058] Through the layered refinement of the above four rules, the system outputs a final robust anomaly level and a fused anomaly score. The description of the abnormal features is accompanied by a description of the abnormal features (e.g., "severe baseline deviation due to kurtosis").

[0059] Example 5: Manual annotation and closed-loop retraining (corresponding to step S7) This embodiment elaborates on step S7 and the implementation of the system's closed-loop optimization capability; S7-1: Manually labeled interface The system provides a Web API or graphical interface for operations and maintenance personnel to query unlabeled samples (displaying their waveforms and characteristics) and annotate them. Annotation content includes: annotation result (normal / abnormal / fault type A / B / C), annotator, annotation time, etc.

[0060] S7-2: Injecting Markers into Libraries and Triggering Tasks After a user submits a label, the API service performs the following atomic operations: Write the annotation information into t_annotation (annotation table).

[0061] Within the database transaction, query the t_retrain_task table (retraining task table) to check if there are any tasks for the same loop with a status of PENDING (pending) or RUNNING (in progress).

[0062] Deduplication: If a task in the above-mentioned state does not exist, a new retraining task record will be automatically inserted, including the loop ID, task creation time, and task status as PENDING. If the task already exists, it will not be inserted again to avoid unnecessary consumption of computing resources.

[0063] S7-3: Background Task Processing A separate background service (model retraining execution module) periodically (e.g., every 5 minutes) polls the t_retrain_task table. When it retrieves a task with a status of PENDING: Lock the task and update its status to RUNNING.

[0064] Based on the loop ID in the task, perform a joint query from t_feature (feature table) and t_annotation (annotation table) to obtain the feature vector data of the loop with labels for the most recent period (e.g., the most recent 7 days) or the most recent N (e.g., 1000) times.

[0065] Using this new data, incrementally retrain or fully retrain both the supervised and unsupervised learning models of this loop. The training process is consistent with that described in S4-2 and S4-3.

[0066] After training is complete, the model files will be overwritten in the configuration directory, and the model version number will be updated.

[0067] Update the task status to SUCCESS and record the end time.

[0068] Anomaly Handling: If any anomaly occurs during training (such as insufficient data or model training failure), the task status will be updated to FAILED and an error log will be recorded for manual intervention or subsequent retry.

[0069] Through the above mechanism, the wisdom of each manual annotation is accumulated into the nourishment for model iteration. The system's ability to identify anomalies in specific working conditions and specific equipment will continue to improve as the running time increases, realizing a true self-adaptive and self-optimizing closed loop.

[0070] Example 6: Data Persistence and Alarm Output (corresponding to step S6) S6: Unified Data Writing To ensure data consistency and traceability, the system performs the following write operations within the same database transaction: Write the original vibration acceleration sequence into t_waveform (waveform data table).

[0071] The calculated multidimensional feature vectors are written to t_feature (feature data table) in JSON format and as independent fields.

[0072] Based on the final fusion result, if an anomaly is determined, information such as the anomaly time, loop number, anomaly level, fusion anomaly score, anomaly feature description, and detailed diagnostic JSON will be written into t_vib_alarm (vibration alarm table) for in-depth analysis on the algorithm side.

[0073] Meanwhile, to ensure seamless compatibility with existing monitoring interfaces, the system converts alarm information into the format required by the target system and writes it into t_ui_alarm (interface alarm table).

[0074] In addition, all alarms can be synchronously output to the rolling log file and system console as configured, facilitating real-time debugging and monitoring by on-site engineers.

[0075] Thus, through the above six embodiments, this invention fully elucidates the entire process from original vibration signal input to multi-model fusion diagnosis, and then to alarm output and closed-loop self-optimization. Those skilled in the art should understand that the parameters, thresholds, and model algorithms used herein can be equivalently replaced and adaptively adjusted according to specific application scenarios; however, any modifications based on the core concept of this invention fall within the protection scope of this invention.

[0076] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0077] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0078] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the above embodiments of the abnormal early warning method based on vibration feature fusion analysis and manual annotation closed-loop retraining.

[0079] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0080] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0082] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0083] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vibration feature fusion analysis and artificial labeling closed-loop retraining anomaly early warning method, characterized in that, Execute the following steps through a computer device: S1: Receive a message from a vibration acquisition terminal, where the message at least includes a loop number, total number of packets, sub-packet sequence number, sub-packet waveform load, and packet tail; S2: Based on the loop number, establish a buffer area, sort, de-duplicate, and assemble each sub-packet. When all sub-packets of the same loop are received, splice them to obtain a complete waveform load; S3: Analyze the complete waveform load to obtain a vibration sampling value sequence, and extract multiple time-domain vibration characteristics from the sequence; S4: Respectively perform statistical baseline anomaly detection, unsupervised machine learning anomaly detection when the system configuration is enabled, and supervised learning anomaly detection, and obtain statistical anomaly scores and statistical anomaly levels, machine learning confidence levels and binary labels, and supervised learning anomaly probabilities respectively; S5: According to a preset fusion rule, perform fusion processing on the multiple detection results obtained in step S4 to obtain the final anomaly level and fusion anomaly score of the target loop; S6: Write the complete waveform load, multiple time-domain vibration characteristics, and the final anomaly level and fusion anomaly score into a database and a log medium; S7: In response to a manual annotation operation, generate and add a retraining task queue, and trigger the corresponding loop model for incremental retraining by a background task processing module according to the task queue.

2. The method of claim 1, wherein the method further comprises: In step S2, independent sub-packet caches and metadata are maintained for each loop. The metadata includes the expected total number of sub-packets and the start reception time; when it is detected that the total number of packets in the same loop message changes or the cache times out, cache reset is performed.

3. The method according to claim 1, characterized in that, The time-domain vibration characteristics extracted in step S3 at least include effective value, peak value, peak-to-peak value, variance, standard deviation, kurtosis, skewness, impulse factor, margin factor, waveform factor, and zero-crossing rate.

4. The method of claim 1, wherein the method further comprises: The statistical baseline anomaly detection in step S4 includes: Maintain historical baseline mean for each vibration feature of each loop and baseline standard deviation ; Calculating on M samples in history buffer ; Definition of moderate deviation band boundaries U2 = μ + 2.5 L3 = μ 3.5 U3 = μ + 3.5 ; When the current feature value v < L3 or v > U3, it is determined as a serious baseline deviation and the first score is accumulated; When L3 ≤ v ≤ U3 and v < L2 or v > U2, it is determined as a moderate baseline deviation and the second score is accumulated, and the first score is greater than the second score; In addition, compare the effective value, peak value, and kurtosis with their respective preset hard thresholds and accumulate fixed scores; Let the above accumulation be denoted as , the statistical abnormal score ; and according to the hard threshold exceeding items, the number of moderate / severe baseline deviation characteristics, and the relationship with the preset effective value multiple, the statistical abnormality is divided into normal, slight, moderate, severe, and extreme levels.

5. The method of claim 1, wherein the method further comprises: The fusion rule in step S5 includes: Rule 1: When both statistical baseline anomaly detection and unsupervised machine learning anomaly detection are identified as anomalies, the fused anomaly score is calculated as follows: Calculation, where To analyze outlier scores, =max(0,min(100,100·s_ML)), where s_ML is the anomaly score output by the unsupervised model. , For the preset weights, where, If the current statistical anomaly level has not reached the highest level, the anomaly level will be raised by one level. Rule 2: When an anomaly is determined solely by unsupervised machine learning anomaly detection, the anomaly level is set to "minor anomaly," and the fused anomaly score is set accordingly. ; Rule three: when only baseline anomaly detection determines an anomaly and the anomaly level is downgraded when the anomaly score is below a preset first threshold.

6. The method of claim 5, wherein the method further comprises: The preset first weight coefficient The preset first threshold value is 30.

7. The method of claim 5, wherein the method further comprises: The fusion rule in step S5 further includes rule four: When supervised learning anomaly detection is available, utilizing its output, the anomaly probability P, on the fusion anomaly score Performing a quadratic correction: ； in The second weighting coefficient is preset, and =1; The fusion anomaly score before the second correction; And according to the relationship between the anomaly probability P and the preset second threshold and third threshold, the anomaly level is corrected upward or downward.

8. The method of claim 1, wherein the method further comprises: In step S6, write the waveform data into a waveform data table, write the feature data into a feature data table, and write the anomaly alarms into a vibration alarm table and an interface alarm table respectively. The interface alarm table is used to be compatible with the existing alarm display system.

9. The method of claim 1, wherein the method further comprises: In step S7, after saving the manual annotation result, first detect whether there is a pending retraining task for the same loop; if not, automatically insert a new retraining task record to avoid duplicate queuing.

10. An abnormality early warning system based on vibration feature fusion analysis and artificial marking closed-loop retraining, used for executing the abnormality early warning method based on vibration feature fusion analysis and artificial marking closed-loop retraining according to any one of claims 1 to 9, characterized in that, Include: A data acquisition framing module for receiving original vibration messages. The data acquisition framing module establishes an independent buffer area according to the loop number, completes sub-packet sorting, de-duplication, and assembly, and outputs a complete waveform load; The waveform analysis and feature extraction module is used to receive the complete waveform load, and its output is connected to the input of the multi-model fusion detection module and the database respectively; the waveform analysis and feature extraction module analyzes the load into a vibration acceleration sequence, calculates multiple time-domain vibration features, and outputs a feature vector; The multi-model fusion detection module is used to receive the feature vector, and its output is connected to the alarm output module and the input of the database respectively. The multi-model fusion detection module internally calls the statistical baseline detection unit, the unsupervised machine learning detection unit and the supervised learning detection unit in parallel, and outputs the final anomaly level and fusion anomaly score according to the preset fusion rules. The alarm output module is used to receive the final anomaly level and the fused anomaly score. Its output end is connected to the database and log medium to persistently write the original waveform, features and anomaly results according to different table structures. The manual annotation interface module is used to receive external annotation instructions. Its output is connected to the input of the retraining task queue module to write the manual annotation results into the database and trigger the generation of retraining tasks. The retraining task queue module is used to manage the status and queuing of retraining tasks, and its output is connected to the input of the model retraining execution module. The model retraining execution module is used to periodically poll the retraining task queue module and read labeled feature data from the database according to the task information, and perform retraining on the unsupervised anomaly detection model and the supervised binary classification model of the specified loop respectively; wherein the supervised branch loads samples from the database association table when the configuration is enabled.

Images