Civil aviation ups backup lead-acid battery online abnormality detection method and system

Abstract

The present application relates to a kind of civil aviation UPS backup lead-acid battery online abnormal detection method and system, it is related to civil aviation power supply and distribution operation maintenance technical field, by obtaining the timestamp of battery in float charging operation state, voltage and internal resistance;Calculate window data quality score;Variational mode decomposition is carried out to the stable internal resistance sequence of float charging, extract low-frequency mode trend slope, low-frequency-high-frequency energy ratio and residual mean square root as special feature for float charging degradation mechanism;First abnormal score relative to itself history and second abnormal score relative to same cluster normal sample of current window are calculated respectively;After the fusion of multiple scores, the alarm probability is output after intra-cluster calibration;When lower than threshold value, trigger data link abnormal handling procedure, work order is routed to acquisition maintenance group and distinguished from battery ontology fault;And operation and maintenance closed loop update is executed.The present application designs special detection features for the long-term float charging degradation mechanism of civil aviation lead-acid battery, improves early abnormal identification rate and reduces false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of civil aviation power supply and distribution operation and maintenance and battery status monitoring technology, and in particular to an online anomaly detection method and system for civil aviation UPS backup lead-acid batteries, which can be applied to online monitoring and anomaly early warning of backup lead-acid batteries in civil aviation data centers, terminal building critical power supply systems, and communication and navigation support systems. Background Technology

[0002] Civil aviation UPS backup lead-acid batteries undertake emergency power supply tasks during mains power outages, power distribution switching, or power fluctuations, and are a core infrastructure to ensure the continuous operation of critical loads in civil aviation. Battery packs are usually composed of multiple lead-acid batteries connected in series. Degradation, mismatch, poor wiring, or abnormal test links of a single battery can reduce the backup power capacity of the entire pack, thereby threatening power supply security.

[0003] Existing lead-acid battery monitoring and anomaly detection technologies can be mainly classified into the following categories: The first type is a combination of manual inspection and fixed threshold methods. This type of method judges anomalies based on the comparison of voltage, internal resistance, or capacity test values ​​with preset thresholds. It is simple to implement, but it is difficult to identify early slow degradation, local sudden changes, and operational mismatch anomalies, resulting in a high rate of missed detections and false alarms.

[0004] The second category is based on unsupervised anomaly detection methods. These methods take voltage sequences, internal resistance sequences, or statistical features as input and employ distance-based, density-based, or isolation-based algorithms to output anomaly scores. Among these, isolation-based algorithms such as iForest and iNNE achieve a good balance between detection performance and computational cost, and are currently the mainstream solutions. However, these methods often use general features, are not optimized for specific operating conditions, and generally lack supporting data governance and engineered alarm systems.

[0005] The third category is methods based on dynamic behavior analysis of time series. These methods identify trend drift, increased volatility, and sudden jumps in a sequence through techniques such as sliding windows, frequency domain decomposition, and subsequence anomaly detection, making them more suitable for detecting early anomalies compared to static thresholds. However, most existing solutions focus only on the algorithm itself, lacking comprehensive data governance, threshold calibration, and operational coordination mechanisms, resulting in insufficient engineering feasibility.

[0006] In existing patent literature, Chinese patent application CN113533986A discloses a method for measuring the internal resistance of a battery based on VMD-Hilbert. This method uses variational mode decomposition to denoise the acquired voltage and current signals to improve the measurement accuracy of the internal resistance parameter. However, this scheme only applies variational mode decomposition to the signal measurement stage and does not further convert the decomposed modal components into anomaly detection features, nor does it design a specific detection feature system for the unique degradation mechanism of lead-acid batteries under float charging conditions.

[0007] Chinese patent application CN120703623A discloses an intelligent monitoring system for predicting UPS battery life, including modules such as data preprocessing, time-domain and frequency-domain feature extraction, and life prediction model. However, it uses general time series features and does not customize features for the degradation mode of slowly increasing internal resistance under float charging steady state; at the same time, it does not establish an isolation mechanism between data link anomalies and battery failures, which can easily trigger invalid battery maintenance due to data acquisition failures; its operation and maintenance closed-loop update is only a conceptual description and lacks specific technical means that can be quantified and implemented.

[0008] Based on the summary and analysis of existing technologies, the current solution has the following three main shortcomings in the application of floating charging UPS in civil aviation: First, the existing methods lack detection features tailored to the unique degradation mechanisms of civil aviation float charging conditions. Civil aviation UPS backup lead-acid batteries are in a constant-voltage float charging state for the vast majority of the time, with voltage fluctuations typically not exceeding 0.05V per cell. No charge-discharge curves are available. Their main degradation modes are softening of the positive electrode active material, sulfation of the negative electrode, and a slow increase in internal resistance due to electrolyte dehydration. The degradation rate is measured in months, with extremely small fluctuations within a single window, making statistical differentiation difficult. Existing solutions directly transplant feature extraction methods from general scenarios without designing dedicated signal decomposition and feature extraction logic for the low-amplitude, long-cycle degradation characteristics of float charging, resulting in insufficient sensitivity for detecting early, slow degradation. Furthermore, the internal resistance signal in float charging exhibits extremely high steady-state characteristics; even minute changes in trends and fluctuations are indicative of anomalies. This steady-state characteristic is not effectively utilized by existing methods.

[0009] Second, the online detection architecture lacks a complete engineering chain from data governance to alarm output. At the data governance level, raw monitoring data is prone to sampling asynchrony, short-term missing data, spikes, duplicate reporting, and timestamp drift. Existing solutions lack explicit data quality quantification and filtering, which easily leads to false alarms. At the detection architecture level, comparing only its own historical data can easily miss group mismatches, and comparing only group samples can easily ignore the unique degradation trajectory of individuals. Existing solutions generally lack a dual-path mechanism that takes into account both longitudinal degradation and lateral mismatch. At the same time, batteries of different brands, specifications, and batches have multiple distribution differences, and existing solutions lack comprehensive cluster affiliation identification and cluster drift handling methods. At the alarm output level, civil aviation has extremely high requirements for power supply continuity, and alarm sensitivity must match the operation and maintenance response capability. Existing solutions lack a continuous confirmation strategy adapted to the handover cycle of civil aviation operation and maintenance shifts. Moreover, data link anomalies such as acquisition board failure and poor probe contact can easily be misjudged as battery degradation. Existing solutions do not strictly isolate the two types of faults in the handling process, which can easily lead to ineffective maintenance.

[0010] Third, the closed-loop update mechanism lacks technical feasibility. While existing solutions propose closed-loop concepts such as incremental update of the sample library, periodic retraining of the model, and adaptive threshold adjustment, they generally fail to clarify: how to differentiate between different types of operational feedback to prevent sample contamination; how to quantitatively adjust alarm thresholds based on false alarm and false negative rates; and how to detect distribution drift caused by overall battery aging and trigger re-clustering. Vague descriptions of the closed loop cannot form practical technical measures and are difficult to maintain detection accuracy in the long term. Summary of the Invention

[0011] The technical problem to be solved by this invention is to provide an online anomaly detection method and system for backup lead-acid batteries in civil aviation UPS. Without adding additional sensors, this method utilizes only three existing monitoring quantities—time stamp, voltage, and internal resistance—to solve three technical problems: insufficient sensitivity in detecting early slow degradation under float charging conditions, lack of a complete engineering link in the detection architecture, and lack of operable rules for the operation and maintenance closed loop. This method achieves high-precision, low-false-alarm, and long-term stable online anomaly detection for backup lead-acid batteries in civil aviation UPS.

[0012] The technical solution adopted by this invention to solve its technical problem is: an online anomaly detection method for backup lead-acid batteries in civil aviation UPS, comprising the following steps: Time-series data of timestamps, voltages, and internal resistances of backup lead-acid batteries in civil aviation UPS systems under long-term float charging operation were obtained, and an online sliding window was constructed. Time alignment, missing data repair, glitch suppression, and deduplication were performed on the data within the window, and the data quality score of the window was calculated. ; Historical statistical features of each battery are extracted, and the operating cluster to which the battery belongs is identified through density clustering. A normal sample library of the same cluster is constructed for each operating cluster. To address the degradation mechanism of lead-acid batteries under long-term float charging, variational mode decomposition is performed on the internal resistance sequence within the window to obtain multiple mode components. Based on these mode components, the low-frequency mode trend slope, the low-frequency-to-high-frequency energy ratio, and the root mean square of the residual are extracted as core detection features. Perform historical behavior detection and calculate the first anomaly score of the current window relative to the battery's own historical normal windows; perform cluster verification detection and calculate the second anomaly score of the current window relative to the normal sample library in the same cluster; combine the first and second anomaly scores with the voltage consistency score and the data quality penalty score to obtain the total anomaly score; The alarm probability is obtained by performing intra-cluster calibration on the total anomaly score. Based on a continuous confirmation window strategy matching the handover cycle of civil aviation maintenance shifts, a four-level differentiated alarm is output; when the data quality score... When the data link is below the preset threshold, the data link abnormality verification process is triggered to strictly distinguish between data link abnormalities and battery failures in work order routing, handling process and sample inclusion rules. Furthermore, based on the operation and maintenance feedback results, samples are included in the classification according to the verification type. The alarm threshold is adaptively updated according to the cumulative false alarm rate and false negative rate. Periodic re-clustering is triggered through the cluster center drift sensing mechanism to continuously update the sample library and model parameters.

[0013] Furthermore, the present invention When the data link is below a preset threshold, the distinction between the data link anomaly handling process and the battery fault alarm is as follows: the data link anomaly work order is pushed to the data acquisition and maintenance group, while the battery fault work order is pushed to the battery maintenance group; the data link anomaly handling action involves checking the acquisition board and the upload link. Battery alarms will not be triggered before recovery; battery fault handling actions include arranging capacity testing and replacement evaluation; data from periods of data link anomaly will not be included in the anomaly sample database, pending further investigation. Reassess after recovery.

[0014] Furthermore, this invention employs the HDBSCAN density clustering algorithm to cluster each battery by extracting statistical feature vectors from its most recent 30-day historical sequence.

[0015] Furthermore, the variational mode decomposition of this invention decomposes the internal resistance sequence into K modal components; the low-frequency mode trend slope The low-frequency to high-frequency energy ratio is obtained by least-squares fitting of the trend reconstruction components and is used to characterize the extremely slow internal resistance increase trend in the floating charge state on a monthly basis. Used to characterize the enhanced high-frequency fluctuations after the float charge steady state is disrupted; the residual mean square These features are used to characterize the short-term fluctuation amplitude of internal resistance under float charging conditions. All three features have a very small normal value range under steady-state float charging. The low-frequency trend component remains stable during stable float charging, the low-frequency energy dominates under normal float charging conditions, and the residual amplitude does not exceed 1% of the mean internal resistance during normal float charging. However, in charging and discharging scenarios, the internal resistance itself exhibits obvious cyclic fluctuations, and the above features do not have the same distinguishing ability.

[0016] Furthermore, the first abnormal score of this invention The minimum subsequence distance is calculated by performing MatrixProfile on the current window feature representation and the historical normal window set, then normalized, and the online change point detection confidence score is superimposed. Enhancement: the second abnormal score The iNNE anomaly score, iForest anomaly score, and structural consistency score are calculated and weighted together by the current window feature vector within its respective running cluster.

[0017] Furthermore, in step S4 of this invention, the total anomaly score for:

[0018] in: As a secondary primary criterion, it highlights deviations from one's own historical behavior; The primary criterion is to highlight the group differences from similar normal batteries; This is an auxiliary item for voltage consistency. This refers to data quality constraints.

[0019] Furthermore, the alarm probability of this invention Through intra-cluster calibration function Total abnormal score The alarm levels are obtained through mapping; alarm levels include: Level 1 Reminder: However, if the continuous confirmation condition is not met, the candidate event will not be pushed. Level II Warning: Continuous Each window satisfies Real-time alerts; Level 3 alarm: Or exceeding multiple times within a given time period Real-time push notifications; Data link anomaly pending verification: and leading, and Triggered when not rising synchronously.

[0020] Furthermore, the sample inclusion method of this invention according to the verification type includes: those confirmed to be normal by capacity testing are included in the normal sample library of the same cluster; those confirmed to be faulty by work orders are marked as confirmed abnormalities for sensitivity calibration; and those with unknown causes are marked as pending verification and isolated. The alarm thresholds are adaptively updated based on the cumulative false alarm rate and false negative rate, including:

[0021] in, To confirm the false alarm rate, To preset the target value, and This is the step size parameter, and the adjustment range in a single instance shall not exceed 0.05; The cluster center drift quantification index is: when the distance between the current cluster center vector and the historical cluster center vector exceeds 10% of the initial internal resistance mean, re-clustering is triggered.

[0022] Meanwhile, the present invention also provides an online anomaly detection system for backup lead-acid batteries in civil aviation UPS systems, comprising: The data access module is used to receive timestamp, voltage, and internal resistance data uploaded by the UPS monitoring system. The data governance module is used to perform time alignment, missing data repair, spike suppression, deduplication, and generate data quality scores. ; The historical sample management module is used to store historical sequences, clustering results, and a normal sample library within the same cluster. The clustering module is used to cluster historical statistical features and identify the operating cluster to which the battery belongs; The online feature extraction module is used to extract the slope of low-frequency modal trends by sliding window. Low-frequency to high-frequency energy ratio and root mean square of residuals Dynamic characteristics of internal resistance under float charging conditions; The historical behavior detection module is used to calculate the first abnormal score of the current window relative to its own historical normal windows. ; The same-cluster verification module is used to calculate the second anomaly score of the current window relative to the normal samples in the same cluster. ; The alarm calibration module is used to fuse various abnormal scores and perform intra-cluster calibration. When the data level falls below a preset threshold, a data link anomaly handling procedure is triggered. The exception interpretation module is used to output the exception type and handling suggestions; The operation and maintenance linkage module is used to push alarm results to the monitoring platform, maintenance work order system and manual review interface, and route them to the data collection maintenance group or battery maintenance group according to the work order type. The model update module is used to include samples in a hierarchical manner according to the validation type, adaptively update the threshold based on the cumulative false positive rate and false negative rate, and perform re-clustering when the cluster center drift trigger condition is met.

[0023] Furthermore, the data governance module described in this invention... Below the preset threshold The data link anomaly handling process is activated: the work order is routed to the data acquisition and maintenance group, in... Restore to The above did not trigger a battery malfunction alarm previously. Reassess the window data after recovery.

[0024] Furthermore, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method.

[0025] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0026] The beneficial effects of this invention are: 1. Existing VMD applications are all aimed at lithium battery voltage signals in charging and discharging scenarios, while this invention applies VMD to the steady-state internal resistance signal of lead-acid battery float charging, and specifically designs a low-frequency mode trend slope. Low-frequency to high-frequency energy ratio and root mean square of residuals Three specific characteristics. Because the internal resistance signal has extremely high stability in the float charge state, It can amplify the slow, gradual increase in internal resistance that has a very small amplitude within a single window but accumulates to a considerable value over a long period. It can detect the enhancement of high-frequency fluctuations after the float charge steady state is disrupted. It can capture short-term fluctuations that are almost non-existent during normal float charging; however, these three features themselves exhibit significant fluctuations in internal resistance during charging and discharging scenarios, and therefore lack the same distinguishing ability. Thus, compared to existing general feature extraction methods, this invention has higher detection sensitivity for early, slow degradation.

[0027] 2. Existing data quality assessment methods only use quality scoring as a data preprocessing tool, while this invention uses data quality scoring... As an active participant in alarm decision-making: when When the data link is below a preset threshold, an abnormal data link is actively triggered for verification. This is strictly distinguished from battery-related faults in terms of work order routing, handling procedures, and sample inclusion rules. This prevents invalid battery repair work orders triggered by data collection link failures from being triggered at the source, significantly reducing false alarm rates and the ineffective use of maintenance resources.

[0028] 3. While existing iForest and iNNE have been proven to be suitable for anomaly detection of lead-acid batteries in civil aviation, they are both single-path detection methods. In contrast, this invention adopts a dual-path approach that combines historical behavior detection with cluster verification. The historical behavior path captures the unique degradation trajectory of the battery, while the cluster verification path captures the mismatch with similar batteries. The two scores are fused together, taking into account both longitudinal degradation and lateral mismatch, thus overcoming the coverage blind spots of the single-path method.

[0029] 4. This invention implements a general closed-loop update through three clear operating methods: three-level inclusion according to the verification type to prevent sample contamination, a quantitative threshold update formula based on the cumulative false alarm rate minus the false negative rate, and re-clustering triggered by cluster center drift quantification, so that the system can maintain detection accuracy continuously during long-term operation. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the overall method executed by the system of the present invention; Figure 2 This is a flowchart of the online anomaly detection process of the present invention; Figure 3This is a data flow diagram illustrating the linkage between data governance, central review, and operation and maintenance in this invention. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0032] An online anomaly detection system for backup lead-acid batteries in civil aviation UPS includes a data access module, a data governance module, a historical sample management module, an operation clustering module, an online feature extraction module, a historical behavior detection module, a cluster verification module, an alarm calibration module, an anomaly interpretation module, an operation and maintenance linkage module, and a model update module.

[0033] Data access module: Receives timestamp, voltage, and internal resistance data uploaded by the UPS monitoring system, and performs equipment identification mapping and basic integrity checks.

[0034] Data governance module: performs timestamp alignment, missing data repair, spike suppression, deduplication, and generates data quality scores. ;when When the data link falls below a preset threshold, a data link anomaly handling process is triggered.

[0035] Historical sample management module: Stores the most recent 30 days of historical sequences, clustering results, historical templates, and a normal sample library within the same cluster.

[0036] The clustering module performs HDBSCAN clustering on historical statistical features to identify the current battery's operating cluster; it also supports periodic re-clustering with cluster drift awareness.

[0037] Online feature extraction module: Extracts dynamic internal resistance features (corresponding to float charging degradation mechanism) with VMD trend slope, low-frequency to high-frequency energy ratio, and root mean square residual as the core by sliding window, as well as voltage-assisted features and voltage-internal resistance coupling features.

[0038] Historical Behavior Detection Module: Calculates the first anomaly score of the current window relative to its own historical normal windows using MatrixProfile. .

[0039] Cluster verification module: Calculates the second anomaly score of the current window relative to normal samples in the same cluster through iNNE, iForest, and consistency check. .

[0040] Alarm calibration module: It integrates various abnormal scores and completes intra-cluster calibration, outputs alarms according to the four-level differentiated handling logic, and strictly distinguishes between data link abnormality to be verified and battery body fault alarms in terms of routing and handling.

[0041] The anomaly interpretation module provides the anomaly type, contribution characteristics, and handling suggestions. Anomaly types include slow degradation, local mutation, population mismatch, link anomaly, and comprehensive anomaly.

[0042] Operation and maintenance linkage module: pushes alarm results to the monitoring platform, maintenance work order system and manual review interface; routes work order to the data collection and maintenance group or battery maintenance group according to work order type.

[0043] Model update module: Includes samples in a hierarchical manner according to the validation type, adaptively updates the threshold using the cumulative false positive rate minus false negative rate formula, and performs re-clustering when the trigger condition is met.

[0044] The overall execution process of the entire system is as follows: Figure 1 As shown.

[0045] Online anomaly detection process as follows Figure 2 As shown, it includes the following steps: 1. Input Data and Basic Definitions For any lead-acid backup battery of a civil aviation UPS under test, the raw observation collected at time t is recorded as:

[0046] in, For timestamps, For voltage, This is internal resistance.

[0047] For an online window of length L, it is denoted as:

[0048] Where L is 120, corresponding to a 120-minute behavioral segment; in other embodiments, L can be in the range of 60 to 180.

[0049] 2. Data Governance and Quality Scoring

[0050] Data processing is performed before the raw sequence enters the detection model. (In the window...) Inside, statistics are compiled separately: Missing Repair Ratio , Burr point ratio , Time alignment deviation ratio , Duplicate data ratio , Define a window data quality penalty item:

[0051] in, , , , The weights are non-negative and satisfy:

[0052] Further define the window data quality score:

[0053] The value range is [0,1], and the closer it is to 1, the higher the data quality.

[0054] In this embodiment, the weights are preferably set as follows: Quality threshold The value is set to 0.75, but in other embodiments it can be adjusted within the range of 0.70 to 0.80.

[0055] when When this happens, the system marks the window as a low-confidence window, activates the data link anomaly verification logic, and... Restore to Prior to the above, other abnormal scores will not be upgraded to battery fault alarms.

[0056] 3. Clustering and Cluster Affiliation Identification

[0057] To address the multi-distribution problem caused by batteries of different brands, specifications, and batches, a statistical feature vector is extracted from the historical sequence of the most recent 30 days for each battery. :

[0058] in: This represents the average internal resistance. The standard deviation of internal resistance; The slope of the internal resistance trend; This is the average voltage value; The standard deviation of voltage; The correlation coefficient between voltage and internal resistance; This represents the average absolute increment of the internal resistance.

[0059] It can be obtained through least squares fitting:

[0060] in, The sampling time sequence number; This represents the total number of sampling points in the historical sequence. For the first The internal resistance value at each sampling time; This represents the average of the sampling time sequence numbers; This is the average value of the internal resistance.

[0061] The clustering module uses HDBSCAN to analyze feature vectors. Density clustering is performed to obtain the cluster to which each battery belongs. Subsequently, a normal sample library and threshold distribution are constructed within each cluster.

[0062] 4. Dynamic feature extraction of internal resistance for float charging

[0063] This scheme uses internal resistance as the primary detection channel and is designed with specific features to address the float charge degradation mechanism. It measures the internal resistance sequence within the window. Perform variational mode decomposition (VMD) to solve the following constrained variational problem:

[0064] Satisfy constraints:

[0065] in, For the first One modal component, This corresponds to the center frequency. Preferably, Take 3 to 5.

[0066] The essence of solving the above variational problem is to rigorously reconstruct the original internal resistance sequence from the sum of all modes. Under the constraints, let each modal component The bandwidth (i.e., the squared gradient of its analytic signal after frequency shift demodulation) The sum of the norms is minimized, thus forcing each mode to be compact and non-overlapping in the frequency domain, and adaptively estimating their respective center frequencies. Thus, the internal resistance signal is effectively separated into components of different frequency bands: the mode with the lowest center frequency carries a slow low-frequency trend (corresponding to monthly degradation processes such as positive electrode softening, negative electrode sulfation, and water loss); the mid-frequency band mode carries a mid-frequency fluctuation (corresponding to fluctuations caused by the temporary disruption of the float charge steady state); and the highest frequency band mode carries a high-frequency noise (corresponding to measurement noise and transient disturbances). This frequency division result provides a physically interpretable basis for subsequent feature extraction: the low-frequency trend term, after trend reconstruction, is used for calculation. This amplifies the gradual increase in internal resistance; mid-to-high frequency components are used for calculation. and To characterize the enhanced fluctuations and short-term abrupt changes after the steady state is disrupted, the three together form a logical closed loop from signal decomposition to feature representation and then to anomaly detection.

[0067] After obtaining each modal component, the following core detection features are extracted: (1) Modal energy ratio

[0068] (2) Low-frequency to high-frequency energy ratio Rlh

[0069] in, This is a set of low-frequency modal indexes. A set of high-frequency modal indexes; To prevent extremely small positive numbers with a denominator of zero.

[0070] This feature is used to detect the phenomenon of increased proportion of high-frequency energy after the float charge steady state is disrupted: under normal float charge conditions, low-frequency energy dominates. Stable; when the battery shows signs of internal short circuit, increased plate bubbles, or abnormal contact, the high-frequency components increase. decline.

[0071] (3) Low-frequency mode trend slope

[0072] Reconstruct the low-frequency mode components into a trend sequence The trend slope is obtained through least squares fitting:

[0073] in, The sampling time sequence number within the window; For the first The measured internal resistance value at each sampling time; The first one is obtained by reconstructing the low-frequency trend mode. Trend values ​​at each sampling time; This represents the average of the sampling time sequence numbers; This is the mean of the trend reconstruction values.

[0074] This feature is used to amplify the extremely slow upward trend of internal resistance under float charging conditions: the low-frequency trend component remains almost unchanged under normal float charging steady state, and a continuous increase is statistically significant.

[0075] (4) Root mean square of residuals

[0076]

[0077] This feature is used to capture sudden increases in internal resistance residuals after removing trend components, quantifying short-term fluctuations during float charging. During normal float charging, the residual amplitude typically does not exceed 1% of the mean internal resistance; any increase strongly indicates a localized abrupt or critical anomaly.

[0078] (5) Local fluctuation intensity of internal resistance

[0079]

[0080] The above three core features ( It closely matches the degradation mechanism of floating charging in civil aviation; in charging and discharging cycle scenarios such as electric vehicles and energy storage power stations, the internal resistance fluctuates significantly with the charging and discharging process itself. The above characteristics do not have equivalent anomaly differentiation capabilities, reflecting the differentiated design for floating charging scenarios.

[0081] 5. Voltage-assisted characteristics and coupling characteristics

[0082] For the voltage sequence within the window, the following auxiliary features are extracted to provide consistency constraints: (1) Average voltage

[0083] (2) Voltage fluctuation intensity

[0084] (3) Voltage range

[0085] (4) Voltage-internal resistance coupling residual

[0086] in, and For the coupled regression parameters of the corresponding running cluster, and These represent the average changes in voltage and internal resistance within the window, respectively.

[0087] Define voltage consistency score:

[0088] in, For the Sigmoid function, The scaling parameters for the coupling residuals of normal samples within the running cluster.

[0089] 6. Historical Behavior Detection Path

[0090] The historical behavior detection path is used to compare the current window with the battery's own historical normal windows to capture the unique degradation trajectory of an individual.

[0091] Perform MatrixProfile calculation on the feature vector of the current window and the set of historical normal windows to obtain the minimum subsequence distance:

[0092] in, The feature representation of the current window, This is a set of historical normal window indexes.

[0093] Further define the basic historical behavioral anomaly score:

[0094] in, This indicates a normalized mapping based on the historical normal distribution. The larger the value, the more the current window deviates from the battery's normal behavior.

[0095] As an enhancement, online change point detection can be overlaid:

[0096] The enhanced historical score is defined as:

[0097] Preferably, Take a value between 0.7 and 0.9.

[0098] 7. Cluster-based verification detection path

[0099] The same-cluster verification path is used to compare the current window with the normal battery group in the same cluster to capture group mismatch anomalies.

[0100] For the current window feature vector In the cluster to which it belongs Calculate separately: 1. iNNE outlier score: ; 2. iForest Abnormal Score: ; 3. Structural consistency score based on IsolationKernel or nearest-neighbor stability: .

[0101] Define the clustered kernel score as:

[0102] Among them, iNNE has a higher weight because isolation algorithms usually have both high efficiency and good performance on this type of data; iForest is used to supplement the identification of isolated anomalies; and the consistency term is used to enhance the ability to identify cluster boundary samples and rare mismatch samples.

[0103] 8. Calculation of total abnormal score

[0104] This invention does not directly concatenate the original voltage sequence and internal resistance sequence, but instead employs a fractional-stage fusion process. The low-quality penalty score is defined as:

[0105] Then the total abnormal score Defined as:

[0106] in: The primary criterion is to highlight the group differences from similar normal batteries; As a secondary primary criterion, it highlights deviations from one's own historical behavior; This is an auxiliary item for voltage consistency. This refers to data quality constraints.

[0107] The above weights are preferred values. In other implementations, they can be updated using an adaptive mechanism based on operation and maintenance feedback and historical results.

[0108] 9. Intra-cluster calibration and hierarchical alarm

[0109] Because the normal distribution of different operating clusters varies, intra-cluster calibration is used instead of a globally fixed threshold. An intra-cluster calibration function is defined. :

[0110] in, It can be constructed in at least one of the following ways: Piecewise function based on historical normal sample quantiles; Monotonic calibration function based on order-preserving regression; Intra-cluster based on Mondrianconformalprediction Value mapping function.

[0111] Based on calibrated alarm probability It adopts a four-level differentiated alarm rule, combined with a continuous confirmation strategy matched with civil aviation operation and maintenance shifts: (1) Level 1 reminder: meets the requirements However, the continuous confirmation condition was not met. The system only records abnormal candidate events on the monitoring platform and does not push them to the on-duty maintenance personnel; if subsequent continuous confirmations are not met... If none of the windows exceed T1, the candidate event will be automatically closed and no work order will be generated. This level is used to filter out occasional noise interference.

[0112] (2) Level II warning: continuous Each window satisfies An alert will be sent to the on-duty maintenance personnel. It is recommended to arrange a follow-up inspection and capacity verification within 1-3 days. There is no need to shut down the system immediately.

[0113] In this embodiment, the continuous confirmation window length is... The value of 4 corresponds to an 8-hour detection period, which matches the 8-hour shift handover cycle of civil aviation, ensuring that a clear alarm record is formed within a complete shift handover cycle, making it easier for the incoming staff to take over and handle the situation. It can be adjusted within the range of 3 to 5, corresponding to a duration of 6 to 10 hours.

[0114] (3) Level 3 alarm: meets the requirements Or repeated multiple times within a given time period exceeding The alert is immediately sent to the responsible maintenance personnel, who are required to complete a preliminary check within 30 minutes and arrange emergency response without affecting the support mission; if the problem is determined to be a battery malfunction, the emergency replacement procedure is initiated. This level corresponds to the immediate response requirements for civil aviation UPS support.

[0115] (4) Data link anomaly pending verification: meets the requirements ,and It dominates the total score, while and The alarm level is not synchronized and is significantly increased. This type of alarm is not escalated to a battery failure; instead, a data link anomaly verification work order is pushed and routed to the data acquisition and maintenance team.

[0116] In this embodiment, Take 0.85 (can be adjusted within the range of 0.80~0.90). Take 0.97 (can be adjusted within the range of 0.95~0.99).

[0117] The procedures for handling data link anomalies and battery malfunctions are strictly distinguished: Data link anomaly handling: The work order is routed to the data acquisition and maintenance team. The handling procedures include checking the data acquisition board, verifying the probe contact status, and troubleshooting the upload link. Restore to Prior to the above, all test results for this battery were marked as pending observation, without triggering any alarms on the battery itself or including it in the abnormal sample database; Reassess the battery status after recovery.

[0118] Battery malfunction handling: only in and and The system is activated when at least one of the two warning thresholds is reached; the work order is routed to the battery maintenance group, and the handling content includes capacity testing, wiring verification, and replacement assessment; the handling results are fed back to the model update module.

[0119] 10. Explanation and Handling Suggestions for Abnormalities

[0120] The system interprets anomaly types based on feature contribution and score distribution, and provides corresponding handling suggestions: like and Continuously changing, and The slow increase is identified as a slow degradation anomaly, and it is recommended to arrange capacity retesting and continuous monitoring. like and Simultaneous sudden increase indicates a localized abrupt anomaly. It is recommended to prioritize checking the wiring status, testing the link, and checking the battery status. like Significantly higher than The result is determined to be a group mismatch anomaly, and it is recommended to replace it based on the comparison results with batteries from the same batch. like Dominant and If the data remains below the threshold, it is determined to be an abnormality in the data acquisition link, triggering the data link abnormality handling process. like , , At the same time, it is at a relatively high level and is judged as a high-confidence comprehensive anomaly.

[0121] 11. System Deployment Method

[0122] This embodiment adopts a deployment method that combines edge-light processing with center-heavy computing: The UPS field gateway or data acquisition server is responsible for data access, time alignment, and basic quality checks. The central monitoring service is responsible for maintaining 30-day historical sequences, sample databases, and operational cluster parameters; The online detection service is triggered once per minute to perform feature extraction, dual-path detection, and alarm calibration on newly added windows. The nighttime offline task is responsible for updating the running clusters, rebuilding the normal sample library, and optimizing the threshold parameters; The results of manual review, maintenance work orders, and capacity tests are periodically fed back to the model update module.

[0123] The database uses a time-series database to store the raw observation data, a feature table to store window-level features, and a model configuration table to store cluster parameters, thresholds, and fusion weights.

[0124] 12. Closed-loop operation and maintenance updates

[0125] This invention achieves a closed-loop operation and maintenance mechanism through three quantifiable and executable mechanisms to continuously maintain detection accuracy: (1) Samples were included in a hierarchical manner according to the type of validation. Operations and maintenance feedback results are categorized into three levels for processing: Capacity testing confirmed normal operation: The corresponding historical window was included in the normal sample library of the same cluster and participated in the cluster boundary update. When a maintenance work order confirms a battery malfunction and completes replacement, the corresponding alarm trigger window is marked as confirmed abnormal and used for sensitivity calibration; it is not included in the normal sample library. If the review conclusion is "cause unknown" or "delayed handling", the corresponding window will be marked with a "pending review" label. It will not be included in the normal database or participate in threshold updates, and will be isolated pending further confirmation.

[0126] This tiered strategy can prevent erroneous verification results from contaminating the normal sample distribution and ensure the long-term purity of the sample library.

[0127] (2) Adaptive threshold update based on cumulative false positive rate minus false negative rate

[0128] The system statistics show the proportion of false positives in maintenance feedback. Confirmed underreporting rate The alarm threshold is adaptively adjusted according to the following formula:

[0129] in, and The preset target false alarm rate and target false alarm rate can be adjusted according to the intensity of operation and maintenance support; , The step size parameter is set to 0.03 in this embodiment, but can be taken in the range of 0.01 to 0.05. The single threshold adjustment range shall not exceed 0.05, and shall be executed on a weekly basis to prevent the parameter from fluctuating too quickly.

[0130] (3) Periodic re-clustering based on cluster drift perception

[0131] To address the issues of overall internal resistance drift and cluster center shift caused by long-term float charging aging of batteries, two types of triggering conditions are set: The number of newly confirmed normal data in the sample library has exceeded the corresponding data volume of more than 30 batteries; The distance between the current cluster center vector and the initial cluster center vector exceeds 10% of the initial average internal resistance.

[0132] When any of the conditions are met, re-clustering is performed in the offline nighttime task, updating cluster labels and intra-cluster threshold parameters without affecting the online detection process. Within 48 hours of the re-clustering completion, the system will label the alarm results with the "just updated model" identifier for maintenance personnel to reference.

[0133] The following example illustrates the specific implementation process.

[0134] Example 1: Online Testing Process for UPS in Civil Aviation Data Centers

[0135] In a UPS backup battery system of a civil aviation data center, timestamp, voltage, and internal resistance data are collected for each lead-acid battery at 1-minute intervals. The system maintains historical data for the most recent 30 days, performs HDBSCAN clustering based on historical statistical characteristics, and establishes a normal sample library and calibration parameters for each operating cluster.

[0136] During online operation, the system constructs a sliding window with a length of 120 minutes. For newly added windows, it first performs missing data repair, glitch suppression, and time alignment, and then calculates the data quality score. Then, VMD decomposition was performed on the internal resistance signal to extract... , , and Features are extracted, including voltage fluctuations and voltage-internal resistance coupling features. These are then calculated using historical behavior detection paths. Calculated from the same cluster kernel path , fusion The final alarm level is then output after in-cluster calibration.

[0137] Example 2: Slow Degradation Recognition Scenario

[0138] When a lead-acid battery exhibits a slow increase in internal resistance and small voltage fluctuations over several consecutive days, but with a gradual imbalance in the coupling relationship, the fixed threshold method fails to provide timely alarms. In this solution, and Continuous abnormal changes, The score rose slowly across multiple adjacent windows; a cluster review revealed a significant deviation from normal batteries in the same batch, indicating a calibration error. If the thresholds for Level 1 and Level 2 are exceeded consecutively, a slow-degrading abnormal alarm will eventually be output.

[0139] Tests showed that when the internal resistance increased by 20% compared to the normal average, and The abnormality usually started 7 to 14 days ago, and the system can complete continuous confirmation and output a level 2 warning within this period, which is at least 7 days earlier than the fixed threshold method.

[0140] Example 3: Local Mutation Identification Scenario

[0141] When a lead-acid battery experiences a sudden jump in internal resistance due to a loose connection, abnormal test link, or a sudden change in internal state, MatrixProfile distance... Confidence level at change points The iForest score increases in sync. After fusion and calibration, the system will output a level 3 alarm within no more than 3 windows (6 hours), prompting users to prioritize checking the wiring status and battery health.

[0142] Example 4: Data Link Anomaly Differentiation Scenarios

[0143] When a faulty acquisition board, poor probe contact, or packet loss in the upload link causes a large number of missing, spiked, and misaligned parts in the original sequence, Significantly decreased to the following, It dominates the total score, while and The system failed to synchronize its data link increase. The system triggered the data link anomaly handling procedure, and the work order was routed to the data acquisition and maintenance team. Before recovery, the battery itself will not trigger an alarm, nor will it be included in the abnormal sample.

[0144] When the data acquisition link is repaired, Restore to After the above, the system re-evaluates the window data. , If the battery level rises simultaneously, a separate battery-related judgment process will be triggered to avoid missing any real battery anomalies.

[0145] Example 5: Detection Performance Verification

[0146] Forty-eight lead-acid batteries from a civil aviation data center were selected and continuously monitored for six months, accumulating approximately 52,000 available windows. The test set covered 30 batteries with normal float charging degradation, 8 batteries with abnormal slow increase in internal resistance due to artificial injection, 4 batteries with abnormal sudden change in internal resistance caused by loose wiring, and 6 batteries with simulated acquisition link failure.

[0147] Compared with pure fixed threshold methods, single iForest models, and pure historical behavior detection methods, the test results of the complete solution of this invention are as follows: the recall rate for slow degradation anomalies with an internal resistance increase of ≥20% is 87.5%, the F1 score is 0.83, and the average early warning lead time is 9 days; the false alarm rate for level 2 and above alarms on normal batteries is 4.2%; for local mutation anomalies with an internal resistance change of ≥15%, the average response time for level 3 alarms is 4.2 hours; and the correct isolation rate for data link anomalies is 96.7%, all of which meet the expected performance indicators.

[0148] The above description is only a specific embodiment of the present invention. Various examples and illustrations do not constitute a limitation on the substantive content of the present invention. Those skilled in the art can make modifications or variations to the above-described specific embodiments after reading the specification without departing from the substance and scope of the invention.

Claims

1. A method for online anomaly detection of backup lead-acid batteries in civil aviation UPS, characterized in that: Includes the following steps, Obtain the timestamp of the backup lead-acid battery of the civil aviation UPS under long-term float charging operation status. ,Voltage Internal resistance For time-series data, construct an online sliding window of length L; The data within the window is processed for time alignment, missing data repair, spike suppression, and deduplication, and a quality score for the window data is calculated. ; Historical statistical features of each battery are extracted, and the operating cluster to which the battery belongs is identified through density clustering. A normal sample library of the same cluster is constructed for each operating cluster. To address the degradation mechanism of lead-acid batteries under long-term float charging, variational mode decomposition is performed on the internal resistance sequence within the window to obtain multiple mode components. Based on these mode components, the low-frequency mode trend slope, the low-frequency-to-high-frequency energy ratio, and the root mean square of the residual are extracted as core detection features. Perform historical behavior detection and calculate the first anomaly score of the current window relative to the battery's own historical normal windows; perform cluster verification detection and calculate the second anomaly score of the current window relative to the normal sample library in the same cluster; combine the first and second anomaly scores with the voltage consistency score and the data quality penalty score to obtain the total anomaly score; The alarm probability is obtained by performing intra-cluster calibration on the total anomaly score. Based on a continuous confirmation window strategy matching the handover cycle of civil aviation maintenance shifts, a four-level differentiated alarm is output; when the data quality score... When the data link is below the preset threshold, the data link abnormality verification process is triggered to strictly distinguish between data link abnormalities and battery failures in work order routing, handling process and sample inclusion rules. Furthermore, based on the operation and maintenance feedback results, samples are included in the classification according to the verification type. The alarm threshold is adaptively updated according to the cumulative false alarm rate and false negative rate. Periodic re-clustering is triggered through the cluster center drift sensing mechanism to continuously update the sample library and model parameters.

2. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 1, characterized in that: The window data quality score satisfy: in: This represents the percentage of missing parts repaired within the window. For the proportion of burrs, This represents the time alignment deviation ratio. The proportion of repeated data. , , , Non-negative weights and ; when When the data link is below a preset threshold, the distinction between the data link anomaly handling process and the battery fault alarm is as follows: the data link anomaly work order is pushed to the data acquisition and maintenance group, while the battery fault work order is pushed to the battery maintenance group; the data link anomaly handling action involves checking the acquisition board and the upload link. Battery alarms will not be triggered before recovery; battery fault handling actions include arranging capacity testing and replacement evaluation; data from periods of data link anomaly will not be included in the anomaly sample database, pending further investigation. Reassess after recovery.

3. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 1, characterized in that: The density clustering employs the HDBSCAN density clustering algorithm, extracting statistical feature vectors for each battery from its most recent 30-day historical sequence. Perform clustering, where: The average internal resistance, The standard deviation of internal resistance, The slope of the internal resistance trend. The average voltage. For voltage standard deviation, This is the correlation coefficient between voltage and internal resistance. This represents the average absolute increment of the internal resistance.

4. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 1, characterized in that: The variational mode decomposition will transform the internal resistance sequence Decomposed into Modal components ,satisfy , Take a value of 3 to 5; the slope of the low-frequency mode trend. The trend reconstruction components are obtained by least-squares fitting. ;in, The sampling time sequence number within the window; For the first The measured internal resistance value at each sampling time; The first one is obtained by reconstructing the low-frequency trend mode. Trend values ​​at each sampling time; This represents the average of the sampling time sequence numbers; The mean of the trend reconstruction values; the low-frequency-high-frequency energy ratio ,in: The modal energy percentage, It is a set of low-frequency modes. It is a set of high-frequency modes. To prevent extremely small positive numbers with a denominator of zero; the residual mean square... .

5. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 1, characterized in that: First abnormal score The minimum subsequence distance is calculated by performing MatrixProfile on the current window feature representation and the historical normal window set, then normalized, and the online change point detection confidence score is superimposed. Enhancement: in: Take a value of 0.7 to 0.9; Based on the score for abnormal historical behavior; The second abnormal score The iNNE anomaly score is calculated for the current window feature vector within its respective running cluster. iForest Abnormal Score and structural consistency score Then, a weighted fusion was performed to obtain: 。 6. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 5, characterized in that: The total abnormal score for: in: As a secondary primary criterion, it highlights deviations from one's own historical behavior; The primary criterion is to highlight the group differences from similar normal batteries; This is an auxiliary item for voltage consistency. These are data quality constraints.

7. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 1, characterized in that: The alarm probability Through intra-cluster calibration function Total abnormal score The alarm levels are obtained through mapping; alarm levels include: Level 1 Reminder: However, if the continuous confirmation condition is not met, the candidate event will not be pushed. Level 2 warning: N consecutive windows meet the requirements Timely push notifications, N is 3 to 5; Level 3 alarm: Or push notifications immediately if the time exceeds T1 multiple times within a given time period; Data link anomaly pending verification: and leading, and Triggered when not rising synchronously; in, Take a value between 0.80 and 0.

90. Take a value between 0.95 and 0.

99. Take a value between 0.70 and 0.

80.

8. The online anomaly detection method for backup lead-acid batteries in civil aviation UPS as described in claim 1, characterized in that: The operation and maintenance feedback results are included in the sample according to the verification type: those confirmed to be normal by capacity testing are included in the same cluster of normal sample library; those confirmed to have battery failure by work order are marked as confirmed abnormal for sensitivity calibration; and those with unknown causes are marked as pending verification and isolated. The alarm thresholds are adaptively updated based on the cumulative false alarm rate and false negative rate, including: in, To confirm the false alarm rate, To preset the target value, and This is the step size parameter, and the adjustment range in a single instance shall not exceed 0.

05. The cluster center drift quantification index is: when the distance between the current cluster center vector and the historical cluster center vector exceeds 10% of the initial internal resistance mean, re-clustering is triggered.

9. An online anomaly detection system for backup lead-acid batteries in civil aviation UPS systems, characterized in that: include: The data access module is used to receive timestamp, voltage, and internal resistance data uploaded by the UPS monitoring system. The data governance module is used to perform time alignment, missing data repair, spike suppression, deduplication, and generate data quality scores. ; The historical sample management module is used to store historical sequences, clustering results, and a normal sample library within the same cluster. The clustering module is used to cluster historical statistical features and identify the operating cluster to which the battery belongs; The online feature extraction module is used to extract the slope of low-frequency modal trends by sliding window. Low-frequency to high-frequency energy ratio and root mean square of residuals Dynamic characteristics of internal resistance under specific floating charging conditions; The historical behavior detection module is used to calculate the first abnormal score of the current window relative to its own historical normal windows. ; The same-cluster verification module is used to calculate the second anomaly score of the current window relative to the normal samples in the same cluster. ; The alarm calibration module is used to fuse various abnormal scores and perform intra-cluster calibration. When the data level falls below a preset threshold, a data link anomaly handling procedure is triggered. The exception interpretation module is used to output the exception type and handling suggestions; The operation and maintenance linkage module is used to push alarm results to the monitoring platform, maintenance work order system and manual review interface, and route them to the data collection maintenance group or battery maintenance group according to the work order type. The model update module is used to include samples in a hierarchical manner according to the validation type, adaptively update the threshold based on the cumulative false positive rate and false negative rate, and perform re-clustering when the cluster center drift trigger condition is met.

10. The online anomaly detection system for backup lead-acid batteries in civil aviation UPS as described in claim 9, characterized in that: The data governance module is in Below the preset threshold The data link anomaly handling process is activated: the work order is routed to the data acquisition and maintenance group, in... Restore to The above did not trigger a battery fault alarm previously. Reassess the window data after recovery.

11. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that: The device contains a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images