A multi-sensor fusion-based power metering box state monitoring method

Abstract

The present application belongs to the technical field of power metering box state monitoring, and particularly relates to a power metering box state monitoring method based on multi-sensor fusion. The method comprises: acquiring displacement, temperature and humidity, water immersion and infrared sensing data, collecting calibration and time alignment, generating a unified time base and a sensor credibility table; then performing time series processing, feature construction and scene prior association to form a fusion feature set; then generating an evaluation result set through fusion evaluation and anomaly identification; then performing hierarchical early warning and strategy generation to output a evidence trigger list; finally, performing evidence collection and learning update to generate a baseline update package. The beneficial effect is to improve monitoring accuracy, reliability and self-adaptive ability, and to realize closed-loop management and intelligent optimization.

Description

Technical Field

[0001] This invention belongs to the field of power metering box status monitoring technology, and particularly relates to a power metering box status monitoring method based on multi-sensor fusion. Background Technology

[0002] As a critical terminal in the power distribution process, power metering boxes are often located outdoors or semi-outdoors, making them susceptible to both environmental and external factors. Existing status monitoring solutions often rely on single or limited sensor channels, resulting in scattered sampling clocks, varying data formats, and a lack of unified time reference across channels, making it difficult to align multi-source data on the time axis. Furthermore, sensor aging, installation differences, and on-site interference are common problems. Existing solutions lack quantitative registration and dynamic constraints on channel quality, and lack a sensor reliability table for cross-channel fusion, leading to missed detections and misjudgments in anomaly identification. Traditional processes typically focus on single-channel threshold comparisons or simplified statistical judgments, failing to provide structured representation at the level of aligned data and anomaly fragment index sets. They lack traceable connections between fragment start and end positions, involved channels, and time indices, making it difficult to support continuous invocation of subsequent tiered early warning and response links.

[0003] In terms of feature construction, existing methods mainly focus on features within the channel, with a loose introduction of scenario knowledge. They lack a mechanism to bind scenario priors such as regional climate, metering box type, and line load habits to channel features at the item level. This results in unclear applicability of features across different seasonal segments, structural regions, and energy consumption segments, and blurred boundaries of participating items in the fusion assessment. In the assessment and identification stages, existing technologies often fail to establish bidirectional references between the fusion feature set, the anomaly fragment index set, and the unified time base. The generation of comprehensive state quantities and the formation of anomaly candidates lack a searchable mapping relationship with upstream data objects. Anomaly pattern recognition lacks sufficient joint constraints on temporal continuity, channel participation, and scenario matching. Causal factors cannot be standardized and archived according to channel and scenario dimensions, and subsequent strategy generation and evidence triggering lack structured input.

[0004] In terms of operation and maintenance closed loop, existing solutions mostly stop at alarm reporting or simple recording, lacking a data and indexing system that connects evaluation result set, hierarchical early warning, handling strategy, evidence collection and reporting, remote handling receipt, and baseline update. They also lack a unified expression and feedback method for evidence fingerprint and standardized status report, making it difficult to use handling result receipt for learning and updating scenario priors and sensor credibility registration, and failing to form constraints and parameter accumulation for the next round of monitoring. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for monitoring the status of power metering boxes based on multi-sensor fusion, comprising:

[0006] Acquire multi-source sensing data, based on displacement sensing, temperature and humidity sensing, water immersion sensing and infrared sensing data, and perform access initialization, range confirmation and zero-point calibration through edge-side acquisition devices. Perform trigger threshold self-calibration and short-term stability check for acquisition and calibration. Extract time information from multi-source raw data streams, and perform sample interpolation, sample truncation and sample re-acquisition sequence consistency processing based on time series rearrangement for time alignment. Based on basic health quantitative indicators including channel activity, data continuity, short-term stability and lost record rate, perform normalization processing and anomaly labeling registration for health fingerprint modeling. Generate multi-source raw data streams, unified time base and sensor reliability table;

[0007] Based on multi-source raw data streams, a unified time base, and a sensor confidence table, time-series alignment, noise reduction, anomaly localization, and summary processing are performed to generate an aligned data and anomaly fragment index set.

[0008] Based on the aligned data and the abnormal fragment index set, and the time series window based on the unified time base, segment-level aggregation processing and weighted integration are performed to construct features. Based on regional climate information, meter box type information, line load habit information and historical baseline information, item-level mapping relationship is established to perform scenario prior association. Based on the unified time base, field naming is unified, value range is unified and default value is filled. A two-way reference with the abnormal fragment index set and the sensor confidence table is established to perform normalized index processing and generate a fused feature set.

[0009] Based on the fusion feature set, fusion evaluation, abnormal pattern recognition and structured processing are performed to generate an evaluation result set;

[0010] Based on the evaluation result set, hierarchical early warning, strategy generation and trigger configuration processing are carried out to generate an evidence collection trigger list;

[0011] Based on the evidence collection trigger list, evidence collection execution, remote processing, and learning update processing are carried out to generate a baseline update package.

[0012] Furthermore, multi-source sensing data includes:

[0013] The multi-source sensing data specifically includes sensing data acquired by displacement sensing units, temperature and humidity sensing units, water immersion sensing units, and infrared sensing units installed inside and outside the power metering box structure through edge-side acquisition devices. Among them, displacement sensing data is used to monitor the physical displacement and vibration state of the metering box structure, temperature and humidity sensing data is used to collect environmental temperature and humidity parameters inside the box, water immersion sensing data is used to detect liquid intrusion around or inside the box, and infrared sensing data is used to sense external heat sources or abnormal temperature rise inside the box.

[0014] Furthermore, health fingerprint modeling specifically includes:

[0015] At each moment, the edge side summarizes the sample readings, valid sample markers, and time indexes of the corresponding channel to form a channel state segment that can be analyzed moment by moment.

[0016] The edge side normalizes the channel status segments based on basic health quantification indicators to form a set of comparable health markers across channels;

[0017] For state segments that exhibit abnormal transitions or long periods of inactivity, the edge side registers anomaly markers within a unified time base and associates these anomaly markers with the channel state segments.

[0018] Furthermore, a sensor confidence table is generated, specifically including:

[0019] At the edge level, long-term stability is summarized and channel stability indicators are generated.

[0020] Access quality is archived at the data collection link level to generate access reliability identifiers;

[0021] The operating status of the sensing unit is updated at the device level, and an operating health indicator is generated.

[0022] The aforementioned channel stability identifier, access reliability identifier, and operational health identifier, combined with basic health quantitative indicators and anomaly markers, form a multi-dimensional health fingerprint entry, which is then indexed and stored in the database based on a unified time base.

[0023] By aggregating health fingerprint entries channel by channel and time by time, a sensor confidence table is generated at the edge.

[0024] Furthermore, the sensor confidence scale includes:

[0025] The sensor credibility table is a structured collection of credibility registrations, including channel identifier, time index, health fingerprint entries, and credibility level fields; the sensor credibility table is bound to a unified time base when it is generated.

[0026] Furthermore, the process of generating the sensor confidence table also includes:

[0027] Obtain a sensor confidence level table to prioritize the correction or necessary removal of channel samples with low confidence levels;

[0028] When locating abnormal regions and labeling segments, refer to the abnormal markers in the sensor confidence table;

[0029] When generating the aligned data and anomalous fragment index set, the confidence level in the sensor confidence table is referenced.

[0030] Furthermore, the sensor confidence scale also includes:

[0031] The sensor credibility table serves as a weighting basis, channel reliability constraint, threshold reference, participation ratio control, evidence collection channel priority determination basis, and evidence source reliability level. It is applied to the entire subsequent methodological chain from feature fusion to remote processing and learning updates.

[0032] Furthermore, the process of segment-level aggregation and weighted integration for feature construction based on a unified time base time series window includes:

[0033] The edge-side processing unit constructs a time-series window with a unified time base as a reference, performs segment-level aggregation processing on the channel samples within the window, and forms a basic quantitative description applicable to the unified feature space. After the basic quantitative description is formed, the edge-side processing unit performs weighted integration of the sample contributions of each channel within the window according to the confidence level registered in the sensing confidence table, so that the multi-channel samples at the same time and in the same segment are comparable and combinable.

[0034] Furthermore, the edge processing unit establishes an entry-level mapping relationship between the channel sample contribution of the fused feature vector and the scene prior, establishes a correspondence between seasonal segments and temperature and humidity features for regional climate information, establishes a correspondence between structural regions and displacement features for meter box type information, establishes a correspondence between energy consumption segments and infrared features for line load habit information, and records the segment reference pointer and time index in each mapping relationship.

[0035] Furthermore, the edge processing unit establishes a bidirectional reference between the normalized feature entries and the abnormal fragment index set: on the one hand, it records the corresponding fragment reference pointer in the feature entries for back lookup, and on the other hand, it registers the directly searchable feature entry index in the abnormal fragment entries for forward lookup; at the same time, it registers the confidence reading pointer between the feature entries and the sensing confidence table.

[0036] The beneficial effects of this invention are:

[0037] (1) This invention uses a unified time base to connect all stages of data collection, alignment, feature generation, evaluation, early warning, evidence collection, and write-back. All data entries flow under the same time index, avoiding cross-module temporal semantic fragmentation, thus constructing a complete monitoring closed loop. Specifically, it includes: multi-source perception collection and calibration to form an aligned data and abnormal fragment index set, then generating a fusion feature set under scenario prior constraints, and carrying out fusion evaluation and abnormal pattern recognition based on it. The output evaluation result set drives the generation of hierarchical early warning and disposal strategies, and finally generates a baseline update package through evidence collection reporting and feedback learning, forming an isomorphic link from data access, temporal governance, feature expression, evaluation and recognition to strategy execution and knowledge write-back.

[0038] (2) By mapping channel quality, access status, and operational health to searchable entries through a sensor credibility table, and integrating alignment and denoising, feature construction, fusion evaluation, and strategy generation, a self-constrained data entry point is formed, effectively suppressing the interference of low-quality input on the conclusion. Compared with the traditional single-channel threshold comparison or loose statistical judgment mode, the unified time reference eliminates cross-channel clock drift, the structured index of abnormal segments strengthens segment-level retrieval and backtracking, the item-level binding of fusion features and scenario priors converges the participation boundary, the fusion evaluation forms a comprehensive state expression for the whole segment under reliability constraints, the abnormal pattern recognition output includes causal elements with channel, time, and scenario dimensions, the draft of graded early warning and disposal strategy completes the trigger configuration accordingly, the evidence collection trigger list implements evidence collection and standardized reporting, and the disposal result receipt drives credibility registration and rolling correction of prior knowledge.

[0039] (3) By aligning data and anomaly fragment index sets to establish bidirectional references at the object, time, and channel levels, anomaly fragments have start and end positions, involved channels, and participation descriptions, providing fragment-level anchors for feature extraction, evaluation and identification, and evidence collection, significantly improving the accuracy and traceability of anomaly identification. The overall process reduces the burden of manual review, decreases the judgment bias caused by cross-channel data inconsistency, and improves the continuity and traceability of the link from alarm generation to policy execution, forming stable data assets and parameter assets for metering box scenarios. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a method for monitoring the status of a power metering box based on multi-sensor fusion, provided in an embodiment of this application. Detailed Implementation

[0041] Example 1: Refer to Figure 1 This is a flowchart illustrating a method for monitoring the status of a power metering box based on multi-sensor fusion, provided by an embodiment of the present invention. The process may include at least steps S100-S600:

[0042] S100: Acquire multi-source sensing data. Based on displacement sensing, temperature and humidity sensing, water immersion sensing, and infrared sensing data, perform access initialization, range confirmation, and zero-point calibration through edge-side acquisition devices. Perform trigger threshold self-calibration and short-term stability check for acquisition and calibration. Extract time information from the multi-source raw data stream. Based on time series rearrangement, perform sample interpolation, sample truncation, and sample re-acquisition sequence consistency processing for time alignment. Based on basic health quantitative indicators including channel activity, data continuity, short-term stability, and lost record rate, perform normalization processing and anomaly labeling registration for health fingerprint modeling. Generate multi-source raw data stream, unified time base, and sensor reliability table.

[0043] S200, based on multi-source raw data streams, a unified time base and a sensor confidence table, performs time-series alignment, noise reduction, anomaly localization and summary processing to generate an aligned data and anomaly fragment index set;

[0044] S300: Based on the aligned data and the abnormal fragment index set, and the time series window based on the unified time base, segment-level aggregation processing and weighted integration are performed to construct features. Based on regional climate information, meter box type information, line load habit information and historical baseline information, item-level mapping relationship is established to perform scenario prior association. Based on the unified time base, field naming is unified, value range is unified and default value is filled. A two-way reference with the abnormal fragment index set and the sensor confidence table is established to perform normalized index processing and generate a fused feature set.

[0045] S400: Based on the fusion feature set, perform fusion evaluation, abnormal pattern recognition and structured processing to generate an evaluation result set;

[0046] S500 performs hierarchical early warning, strategy generation, and trigger configuration processing based on the evaluation result set, and generates an evidence collection trigger list;

[0047] S600, based on the evidence collection trigger list, performs evidence collection execution, remote processing and learning update processing, and generates a baseline update package.

[0048] Step S100 includes at least steps S110-S130:

[0049] S110. Acquire multi-source sensing data, perform collection and calibration, and obtain multi-source raw data stream;

[0050] The multi-source sensing data specifically includes sensing data acquired in parallel by displacement sensing units, temperature and humidity sensing units, water immersion sensing units, and infrared sensing units installed inside and outside the power metering box structure through an edge-side acquisition device. Displacement sensing data is used to monitor the physical displacement and vibration state of the metering box structure; temperature and humidity sensing data is used to collect environmental temperature and humidity parameters inside the box; water immersion sensing data is used to detect liquid intrusion around or inside the box; and infrared sensing data is used to sense external heat sources or abnormal temperature rises inside the box. The aforementioned multi-source sensing data undergoes access initialization, range confirmation, and zero-point calibration via the edge-side acquisition device. Based on a preset calibration parameter library, the sampling interval, quantization resolution, and range upper limit are uniformly set. Simultaneously, trigger threshold self-calibration is performed on the transient characteristics of the displacement and infrared sensing units to ensure that short-term abrupt events are fully recorded. Short-term stability checks are performed on the gradually changing characteristics of the temperature, humidity, and water immersion sensing units to suppress transient disturbances. Finally, a multi-source raw data stream with channel identifiers and initial time stamps is formed, providing a basic data source for subsequent time alignment processing and health fingerprint modeling.

[0051] Specifically, the multi-source sensing data consists of displacement sensing, temperature and humidity sensing, water immersion sensing, and infrared sensing. Specifically, the displacement sensing unit, temperature and humidity sensing unit, water immersion sensing unit, and infrared sensing unit, installed inside and outside the metering box structure, are connected in parallel via an edge-side acquisition device to synchronously collect environmental and structural elements. The edge-side acquisition device initializes the electrical signals and data frames from each sensing unit, confirms the range, and performs zero-point calibration. It also sets the sampling interval, quantization resolution, and upper limit of the range for each channel consistently according to a pre-set calibration parameter library. Furthermore, to address the transient characteristics of the displacement sensing unit and infrared sensing unit, the edge-side acquisition device performs a trigger threshold self-calibration before sampling to ensure that short-term abrupt changes are completely recorded within the sampling window. To address the gradual changes of the temperature and humidity sensing unit and water immersion sensing unit, the edge-side acquisition device performs a short-term stability check after sampling to avoid abnormal readings introduced by transient disturbances. Understandably, the access initialization includes channel identifier binding, timestamp injection, and data format unification. The range confirmation and zero-point calibration include comparing the difference between the calibration source and the measured source and updating the channel offset within the allowable range. Through the above acquisition and calibration, a multi-source raw data stream containing displacement sensing data, temperature and humidity sensing data, water immersion sensing data, and infrared sensing data is obtained. The multi-source raw data stream is stored sequentially in an edge-side buffer and includes channel identifiers and initial timestamps for alignment processing and health fingerprint modeling in subsequent steps. To ensure the continuity of subsequent processing links, the multi-source raw data stream is registered as available input immediately after generation, allowing S120 to extract time information from the multi-source raw data stream and perform alignment processing. It also serves as the basic data source for S210 when obtaining the multi-source raw data stream, the unified time base, and the sensing reliability table. Through the above operation process, it is clear that the input of this step is the raw readings and calibration parameter library of the multi-channel sensing unit, and the output is a multi-source raw data stream with channel identifiers and initial time stamps. This output is directly referenced by S120 and continuously used by each sub-step of S200, so as to maintain the same name and structure as the subsequent steps.

[0052] S120. Extract time information from the multi-source raw data stream, perform alignment processing, and obtain a unified time base;

[0053] After the multi-source raw data stream is generated, the edge side extracts and verifies the time information in the multi-source raw data stream. Specifically, firstly, time information is extracted based on the initial time stamp attached to each channel to complete the investigation of channel clock drift, time jumps, and missing intervals, and the current offset and jitter range of each channel are registered in the time reference maintained by the edge side. Further, the multi-source raw data stream is time-series rearranged, splicing data segments from different sources in chronological order to resolve timing misalignments caused by inconsistent sampling intervals of different sensing units. To ensure that different channels participate in subsequent processing under the same reference, the edge side generates a unified time base based on the time reference. The unified time base is a time index set shared by all channels, including time sequences, corresponding channel mapping relationships, and traceable synchronization control markers. The formation process of the unified time base includes alignment processing based on the initial timestamp and integrity verification based on the channel mapping relationship. In the alignment processing, the edge side uses the time reference as a basis to perform consistency processing on the sample interpolation, sample truncation, and sample resampling order of each channel, thereby ensuring that at any given time, the corresponding channel sample can be found under the unified time base. Through the above alignment processing, a unified time base is obtained, and the unified time base is bound and stored with the multi-source original data stream at the edge side for retrieval in subsequent steps. The unified time base serves as a necessary input for S130 when performing health fingerprint modeling on the multi-source original data stream and the unified time base, and is used to unify the time reference of samples during the health fingerprint modeling process to avoid deviations caused by time drift of different channels in health fingerprint modeling. The unified time base serves as the basis for alignment control during time alignment and robust denoising in S210, guiding the generation of clean data after alignment. During segment annotation in S220, the unified time base defines the start and end positions of abnormal intervals. In S230, when generating aligned data and an abnormal segment index set, the unified time base unifies the index specification and the time definition of the aligned data. Understandably, the input to this step is the multi-source raw data stream and time reference, and the output is the unified time base. This output provides a unified time reference for S130, S210, S220, and S230, enabling subsequent steps to perform feature construction, fusion evaluation, and hierarchical early warning within the same time frame, ensuring consistency in terminology and objects.

[0054] S130. Perform health fingerprint modeling on the multi-source raw data stream and the unified time base to generate a sensor confidence table;

[0055] After binding the multi-source raw data streams to the unified time base, the edge side performs health fingerprint modeling on the multi-source raw data streams and the unified time base to generate a sensor reliability table for subsequent fusion evaluation and hierarchical early warning. Specifically, firstly, on the time series of the unified time base, basic health quantification indicators are calculated for each channel at each time step. The basic health quantification indicators include channel activity, data continuity, short-term stability, and record loss rate. At each time step, the edge side summarizes the sample readings, valid sample markers, and time indexes of the corresponding channel to form a channel state segment that can be analyzed time step by time. Further, the edge side normalizes the channel state segments based on the basic health quantification indicators to form a cross-channel comparable set of health markers. For state segments with abnormal jumps or long periods of blankness, the edge side registers anomaly markers within the unified time base and associates the anomaly markers with the channel state segments so that the same time reference can be used for segment-level tracing and verification in subsequent steps. Subsequently, the edge side summarizes long-term stability at the channel level to generate a channel stability identifier; archives access quality at the acquisition link level to generate an access reliability identifier; and updates the operating status of the sensing unit at the device level to generate an operational health identifier. The aforementioned channel stability identifier, access reliability identifier, and operational health identifier, combined with the basic health quantification indicators and anomaly markers, form multi-dimensional health fingerprint entries, which are indexed and stored in the database based on the unified time base. Through channel-by-channel and time-by-time aggregation of the health fingerprint entries, the edge side generates a sensing reliability table. The sensing reliability table is a structured reliability registration set, containing fields such as channel identifier, time index, health fingerprint entry, and reliability level, providing directly referable reliability information for any channel at any given time. To ensure the continuity of subsequent steps, the sensing reliability table is bound to the unified time base during generation to ensure consistency with the ternary input when S210 acquires the multi-source raw data stream, the unified time base, and the sensing reliability table. This facilitates the selection of alignment strategies and denoising intensity based on the reliability level during time-series alignment and robust denoising. Specifically, in step S210, the sensor confidence table is obtained during the alignment and denoising process to prioritize the correction or necessary removal of channel samples with low confidence, and the unified time base is used to ensure that multiple channels are aligned under the same time index. In step S220, when locating abnormal intervals and labeling segments, the abnormal markers in the sensor confidence table are referenced to define the start and end ranges and channel participation of segment-level anomalies. In step S230, when generating the alignment data and the abnormal segment index set, the confidence level in the sensor confidence table is referenced to attach confidence labels to the alignment data and record the channel confidence distribution in the abnormal segment index, ensuring that feature construction can be carried out accordingly when the alignment data and the abnormal segment index set are obtained in step S310.Furthermore, the sensor credibility table serves as the weighting basis when S410 obtains the fusion feature set, as the channel reliability constraint when S420 extracts key elements from the comprehensive state score and anomaly candidate list, as the threshold reference when S510 obtains the evaluation result set for graded early warning, as the participation ratio control when S520 extracts constraints from the early warning level for strategy generation, and is used to determine the priority of the evidence collection channel when S530 generates the evidence collection trigger list. Finally, it is used to determine the reliability level of the evidence source when S610 obtains the evidence collection trigger list for evidence collection execution and summary generation. In the remote processing and learning update process of S620 and S630, the validity evaluation results are written back to improve the credibility registration in subsequent cycles. Through the above process, it is clear that the input of this step is the multi-source raw data stream and the unified time base, and the output is the sensor confidence table. By binding with the unified time base and explicitly referencing it in subsequent steps, the consistency, traceability and callability of the multi-source raw data stream, the unified time base and the sensor confidence table in the method chain are ensured.

[0056] Step S200 includes at least steps S210-S230:

[0057] S210. Obtain the multi-source raw data stream, the unified time base, and the sensor confidence table, perform time-series alignment and robust denoising, and obtain aligned clean data.

[0058] The multi-source raw data stream consists of displacement sensing data, temperature and humidity sensing data, water immersion sensing data, and infrared sensing data. Specifically, the edge-side processing unit first reads the channel identifiers and timestamps of the multi-source raw data stream and matches them line by line with the time sequence of the unified time base to form a set of segments to be aligned. During the matching process, the edge-side processing unit performs consistency checks on the sampling interval, sample quantity, and time coverage of each channel, performs buffer registration for early-arriving samples, and performs delayed merging for late-arriving samples to ensure that there are channel samples available for calculation at any reference time. Further, the edge-side processing unit performs alignment processing based on the unified time base: interpolating channels with large sampling intervals, re-ordering channels with excessively high sampling frequencies and redundancy, and strictly truncating out-of-bounds samples to ensure that multi-channel samples at the same reference time have a consistent time index. Subsequently, the edge processing unit invokes the sensor reliability table to obtain the reliability level of each channel at each time step, and establishes the execution order and intensity of alignment and denoising accordingly. Channel samples with higher reliability levels are used as reference trajectories for time drift verification, while channel samples with lower reliability levels and continuous missing or abrupt changes are preferentially entered into correction or elimination channels to reduce noise diffusion across channels. To balance the transient and slowly varying characteristics of different sensor types, the edge processing unit employs a smoothing strategy prioritizing transient preservation for displacement sensing data and infrared sensing data, and a smoothing strategy prioritizing steady-state consistency for temperature and humidity sensing data and water immersion sensing data. These strategies are all constrained by the sensor reliability table levels and their application time range is limited by the unified time base. After the alignment and denoising linkage is completed, the edge processing unit merges and sorts the samples from each channel at the same time step according to the channel identifier, generating summary data that is time-consistent, channel-consistent, and noise-suppressed. The mapping relationship between this summary data and the unified time base is then registered in the edge storage medium. At this point, the output is the aligned clean data, which remains bound to the unified time base and the sensor confidence table. This output serves as the input for subsequent location of abnormal intervals and segment annotation from the aligned clean data, and also as a time-consistent data source for feature construction and fusion evaluation. Through the above implementation process, the inputs to this step are the multi-source raw data stream, the unified time base, and the sensor confidence table; the output is the aligned clean data. The output maintains complete consistency with subsequent steps in naming and structure, facilitating unified retrieval and referencing throughout the method chain.

[0059] S220. Locate abnormal intervals from the aligned clean data, label the segments, and obtain the abnormal segment index;

[0060] After obtaining the aligned clean data, the edge-side processing unit uses the unified time base as a time reference to perform continuity and stability checks on sample entries of each channel at the same reference time. Specifically, the edge-side processing unit first sequentially scans the aligned samples at adjacent times, identifies missing sample segments, abnormal rise segments, and abnormal fall segments, and records the corresponding start and end times; simultaneously, it checks the consistency of the change relationship of multiple channels at the same reference time. When cross-channel co-variation or mutual deviation occurs and the amplitude exceeds the normal boundary, candidate anomaly markers are generated under the unified time base. Further, the edge-side processing unit, in conjunction with the sensor confidence table, performs channel participation evaluation and confidence correction on the candidate anomaly markers: if the anomaly is supported by channels with higher confidence levels, it is preferentially retained; if the anomaly is triggered solely by a channel with a lower confidence level and lacks cross-channel support, its priority is reduced or it is marked as a segment to be reviewed. To ensure the retrievability and traceability of anomalous fragments, the edge processing unit segments candidate anomaly markers, confirmed through both channel and time layers, according to the time series of the unified time base. This results in anomalous fragment entries containing start and end positions, involved channels, channel participation, and a reference pointer. The reference pointer points to the corresponding sample position in the aligned clean data, facilitating fragment-level feature extraction and evidence backtracking in subsequent steps. After generating initial entries, the edge processing unit performs deduplication and merging processes on the anomalous fragment entries, merging adjacent fragments with similar features into continuous fragments and uniformly numbering entries that are repeatedly registered across channels, ensuring that anomalous fragments have clear boundaries and unique identifiers in both the time and channel dimensions. Finally, after location and labeling, the output is an anomalous fragment index. This output, in the form of an entry set, maintains consistency with the unified time base, establishing a bidirectional reference relationship between the entry and the aligned clean data at the entry level. This output serves two purposes: firstly, it acts as direct input for the next step of summarizing and organizing the aligned clean data and the anomalous fragment index; secondly, it is used to define key time periods and key channels in subsequent feature construction and scene prior binding, and to provide fragment-level contextual reference during fusion evaluation and hierarchical early warning. Through the above implementation process, the input of this step is the aligned clean data, the unified time base, and the indirect constraints of the sensor credibility table bound to it; the output is the anomalous fragment index, ensuring consistency in time reference and object identification in subsequent processing.

[0061] S230. Summarize and organize the aligned clean data and the abnormal fragment index to generate an aligned data and abnormal fragment index set.

[0062] After obtaining the aligned clean data and the anomalous fragment index, the edge processing unit performs structured aggregation and consistency verification on the two types of objects based on the unified time base. Specifically, the edge processing unit first performs structured rearrangement on the aligned clean data, sorts multi-channel samples at the same reference time according to channel identifiers, and establishes a time primary key and channel dependency relationship to form an aligned dataset with the time index as the primary key and channel samples as dependent entries. At the same time, the fragment entries in the anomalous fragment index are sorted by time according to their start and end positions, and a reference pointer to the aligned dataset is added to each entry to realize bidirectional retrieval and mapping between anomalous fragment entries and aligned data entries. Further, the edge processing unit performs consistency verification on the above two types of objects: at the time boundary level, it verifies whether the start and end times of the anomalous fragment entries completely fall within the time range of the aligned dataset; at the channel identifier level, it verifies whether the channels involved in the fragment entries are consistent with the channel set registered in the aligned dataset; at the entry integrity level, it verifies whether each anomalous fragment entry has the necessary fields such as start and end positions, involved channels, channel participation degree, and reference pointer. If inconsistencies are found between the two types of objects in terms of time boundaries or channel identifiers during the verification process, the unified time base is used as the final reference to reshape the boundaries of the abnormal fragment entries and verify the channel identifiers in the aligned dataset until the expressions of the two types of objects are completely consistent at the same time and on the same channel. Subsequently, the edge processing unit encapsulates the aligned clean data and the abnormal fragment index into a set object. The set object includes a time index area, a channel sample area, and a fragment entry area. The time index area corresponds one-to-one with the unified time base and is used for unified retrieval. The channel sample area stores the channel samples after time alignment and robust denoising for subsequent feature construction and fusion evaluation. The fragment entry area stores the abnormal fragment entries after location and annotation for subsequent graded early warning and evidence collection triggering. To maintain consistency with the constraints of the preceding steps, the edge processing unit registers the association method with the sensor credibility table in the metadata of the set object, so that subsequent steps can read the credibility level at any time and on any channel when needed and use it as a weighting or filtering basis. After the above encapsulation is completed, the output is an alignment data and anomaly fragment index set. This output is registered as a callable resource on the edge side and is directly called when subsequent feature construction and scene prior binding obtain the alignment data and anomaly fragment index set for feature construction. It is used as the basic input reference for time consistency and fragment traceability by fusion evaluation and anomaly pattern recognition, and is used as a fragment-level execution reference and policy implementation verification by hierarchical early warning and policy generation. At the same time, it serves as the retrieval entry point for evidence fragment location and summary generation during evidence collection and reporting and closed-loop write-back.In summary, the technical effects of S210 to S230 are to form aligned clean data, accurate abnormal fragment indexes, and a standardized set of aligned data and abnormal fragment indexes under a unified time reference and credibility constraints. This provides a consistent, traceable, and directly callable input system for subsequent feature construction, fusion evaluation, hierarchical early warning, and evidence collection and reporting. Furthermore, by binding with the unified time base and the sensor credibility table, it ensures the consistency and closed-loop connection of the method chain in terms of naming, structure, and reference.

[0063] Step S300 includes at least steps S310-S330:

[0064] S310. Obtain the alignment data and the abnormal fragment index set, perform feature construction, and obtain the fused feature vector;

[0065] The aligned data and the set of anomalous fragment indices are output from the preceding steps and establish a one-to-one correspondence with the unified time base. Specifically, the edge-side processing unit first reads the channel sample entries of the aligned data sequentially on the time series of the unified time base. The channel sample entries include displacement sensing data, temperature and humidity sensing data, water immersion sensing data, and infrared sensing data. By cross-referencing with the fragment entries in the set of anomalous fragment indices, it confirms whether the samples at each time point fall within the start and end position range of the anomalous fragment. Further, the edge-side processing unit constructs a time series window with the unified time base as a reference, performs segment-level aggregation processing on the channel samples within the window, and forms a basic quantitative description suitable for the unified feature space. After the basic quantitative description is formed, the edge-side processing unit performs weighted integration of the sample contributions of each channel within the window according to the confidence level registered in the sensing confidence table, so that the multi-channel samples at the same time and in the same segment have comparability and composability. Subsequently, the edge-side processing unit concatenates the weighted basic quantization description in chronological order, and simultaneously establishes fragment-level reference pointers with the abnormal fragment index set to mark which segments and fragment entries in the concatenated description have overlapping relationships. Through the above concatenation and annotation, the edge-side processing unit obtains a multi-channel time-series description sequence under the unified time base, and supplements the channel identifiers and time indices of the aligned data at both ends of the sequence, so that the sequence can be directly retrieved and called in subsequent steps. To ensure the expandability of subsequent binding processing, the edge-side processing unit performs a consistency check on the time-series description sequence, including time boundary consistency, channel set consistency, and abnormal fragment mapping consistency. If any inconsistency is found, the corresponding segment is re-aligned and the index is revised using the unified time base as the final reference. After completing the consistency check, the edge-side processing unit solidifies the time-series description sequence into a fusion feature vector. The fusion feature vector retains three types of information in its structure: time index, channel sample contribution, and fragment reference pointer. It is associated with the sensor confidence table through meta-information records so that the original registration does not need to be retrieved again when calling the confidence level in subsequent steps. At this point, the inputs for this step are the alignment data and the set of anomalous fragment indices, along with the unified time base and the sensing confidence table bound to them. The output is a fused feature vector, which maintains consistency with the preceding objects in terms of naming, structure, and indexing method. This output serves as the direct input for the next step to associate scene priors with the fused feature vector and perform binding processing.

[0066] S320. Associate scene priors with the fused feature vectors and perform binding processing to obtain feature prior binding results;

[0067] After obtaining the fused feature vector, the edge processing unit reads the scene priors within the same time range as the fused feature vector. These scene priors include regional climate information, meter box type information, line load behavior information, and historical baseline information. A time index with the same name is established with the unified time base to maintain time reference consistency. Specifically, the edge processing unit first aligns the time markers in the scene priors to the time index positions of the fused feature vector using the unified time base as a reference, confirming that the corresponding scene prior entry can be retrieved at any reference time. Subsequently, based on the confidence level registered in the sensor confidence table, the edge processing unit constrains the contribution of channel samples related to that time, giving higher confidence level channel samples a higher weight in scene prior association, while assigning a conservative participation ratio to lower confidence level channel samples in scene prior association, thus ensuring consistency in both time and reliability references during the binding process. Furthermore, the edge-side processing unit establishes an entry-level mapping relationship between the channel sample contribution of the fused feature vector and the scene prior. For regional climate information, it establishes a correspondence between seasonal segments and temperature and humidity characteristics; for meter box type information, it establishes a correspondence between structural regions and displacement characteristics; and for line load habit information, it establishes a correspondence between energy consumption segments and infrared characteristics. In each mapping relationship, it records a segment reference pointer and a time index for subsequent backtracking of abnormal segments. After the above mapping relationships are established, the edge-side processing unit performs consistency checks on the entries during the binding process. The checks include time index consistency, channel identifier consistency, and the completeness of scene prior entries. If any entries are missing or indexes are misaligned, the scene prior entries are filled in or offset using the unified time base as a reference until every moment in the fused feature vector has a referenceable scene prior entry. After completing the consistency verification, the edge processing unit encapsulates the mapping relationship between the fused feature vector and the scene prior into a feature prior binding result. This result retains the time index, channel sample contribution, scene prior entries, and fragment reference pointers. It also registers the association method with the sensor credibility table in the metadata, allowing the credibility level to be directly read as a filtering or weighting basis in subsequent steps. Thus, the inputs to this step are the fused feature vector and the scene prior, along with the unified time base and the sensor credibility table bound to them. The output is the feature prior binding result, which, in its structured expression, simultaneously contains three types of information: time reference, channel reference, and scene reference. This output serves as the direct input for the next step of standardizing and indexing the feature prior binding result, and also as the preliminary data source for subsequent fusion evaluation and abnormal pattern recognition using the fused feature set.

[0068] S330. The prior binding results of the features are normalized and indexed to generate a fused feature set;

[0069] After obtaining the feature prior binding results, the edge processing unit normalizes and indexes them to ensure that the reference relationships with the unified time base, the sensor confidence table, and the abnormal segment index set remain consistent at the object, time, and entry levels. Specifically, the edge processing unit first verifies the time index of the feature prior binding results using the unified time base as a reference to ensure that each feature entry can be uniquely located in the time index area. Subsequently, it normalizes the channels sample contributions and scene prior entries in the feature prior binding results, including unifying field naming, value range, and filling default values, so that similar entries maintain consistent expression during cross-time, cross-channel, and cross-segment retrieval processes. Furthermore, the edge-side processing unit establishes a bidirectional reference between the normalized feature entries and the abnormal fragment index set: on the one hand, it records the corresponding fragment reference pointer in the feature entries for backtracking; on the other hand, it registers the directly searchable feature entry index in the abnormal fragment entries for forward searching. Simultaneously, it registers a confidence level read pointer between the feature entries and the sensor confidence level table, enabling subsequent steps to read the confidence level as needed at any time and on any channel, and use it as a weighting or filtering basis. After the above normalization and bidirectional reference establishment, the edge-side processing unit constructs an index for the feature entries. The index construction includes three parts: time index, channel index, and fragment index, and records a one-to-one correspondence with the unified time base in the index metadata to ensure that subsequent searches are unambiguous at the time level. After index construction is completed, the edge-side processing unit encapsulates all normalized and indexed feature entries into a fusion feature set. This fusion feature set includes a time index area, a feature entry area, and a reference relationship area. The time index area is consistent with the unified time base, the feature entry area stores normalized features and scene information, and the reference relationship area stores bidirectional references to the abnormal segment index set and the sensor reliability table. Once generated, the fusion feature set is registered as a callable resource on the edge side, directly available for subsequent steps to obtain the fusion feature set for fusion evaluation, obtaining a comprehensive state score and an anomaly candidate list. It also serves as the sole feature entry point when extracting key elements from the comprehensive state score and anomaly candidate list and performing anomaly pattern recognition, maintaining consistency in data source and time reference across the method chain.In summary, the three consecutive steps enable the feature construction process, which takes the aligned data and the abnormal fragment index set as input, to obtain a fused feature vector; the binding process, which takes the fused feature vector and the scene prior as input, to obtain a feature prior binding result; and the normalization and indexing process, which takes the feature prior binding result as input, to generate a fused feature set. The above outputs are sequentially invoked by subsequent fusion evaluation and abnormal pattern recognition, hierarchical early warning and strategy generation, and evidence collection reporting and closed-loop write-back, maintaining consistent expression in object naming, time reference, and bidirectional referencing. The technical effect is to form a fused feature data system with a clear structure, complete index, and traceability, providing a stable data foundation for continuous retrieval and consistent calculation in subsequent steps.

[0070] Step S400 includes at least steps S410-S430:

[0071] S410. Obtain the fusion feature set, perform fusion evaluation, and obtain a comprehensive state score and an anomaly candidate list;

[0072] The fusion feature set is generated by the preceding steps and establishes a bidirectional reference with the unified time base, the abnormal segment index set, and the sensing confidence table. Specifically, the edge-side processing unit first reads the feature entries in the fusion feature set hourly on the time series of the unified time base. The feature entries include time indexes, channel sample contributions, scene prior entries, and segment reference pointers. Based on this, the edge-side processing unit establishes a channel feature view at the same reference time, and integrates it by weighting the channel sample contributions with the confidence levels in the sensing confidence table to form a multi-channel fusion metric for that reference time. To avoid evaluation bias caused by interruptions in time continuity, the edge-side processing unit constructs a sliding window between adjacent time points under the unified time base, aggregates the multi-channel fusion metrics within the continuous window, and limits the applicable scope of the participating entries according to the scene prior entries, so that seasonal segments, structural regions, and energy consumption segments are processed consistently under the same time reference. For the completed segment aggregation results, the edge-side processing unit continues to perform segment-level annotation based on the segment entries in the abnormal segment index set. Fusion metrics overlapping with the start and end positions of segments are uniformly archived into the segment candidate pool, and segment reference pointers and time indices are retained in the candidate pool for subsequent backtracking and comparison. Through the above-mentioned time-by-time integration, segment aggregation, and segment annotation, the edge-side processing unit standardizes the fusion metrics within the same time reference and calculates a comprehensive state score for the entire segment. Simultaneously, it summarizes the abnormal candidate list based on the archiving status in the segment candidate pool. The comprehensive state score and the abnormal candidate list correspond one-to-one with the unified time base during generation, and their reference relationships with the fusion feature set, the abnormal segment index set, and the sensor confidence table are registered in the metadata. Thus, the inputs to this step are the fusion feature set and its associated unified time base, the abnormal segment index set, and the sensor confidence table; the outputs are the comprehensive state score and the abnormal candidate list. This output serves as the direct input for the next step to extract key elements from the comprehensive state score and the abnormal candidate list for abnormal pattern recognition.

[0073] S420. Extract key elements from the comprehensive state score and the list of anomaly candidates, perform anomaly pattern recognition, and obtain anomaly confidence and cause elements.

[0074] After obtaining the comprehensive status score and the anomaly candidate list, the edge processing unit uses the unified time base as a time reference to perform segment-by-segment linked analysis of the two. Specifically, the edge processing unit first traverses the segment entries in the anomaly candidate list, looks up the corresponding feature entries in the fusion feature set according to the segment reference pointer, and simultaneously reads the comprehensive status score domain within the same time range as the segment entry to establish a segment-level key element set. The key element set includes at least the continuous temporal form of channel sample contributions, the matching status of scene prior entries, the distribution of confidence levels registered in the sensor confidence table within the time range, and the boundary overlap relationship with the anomaly segment index set. The edge processing unit performs two-level screening on the segment-level key element set: the first-level screening is based on the common threshold of channel sample contributions and confidence levels to eliminate candidate segments with insufficient participation or reliability; the second-level screening is based on the adaptability range of scene prior entries to transfer segments that do not conform to regional climate, meter box type, or line load habits to the review area. After completing the two-level screening, the edge-side processing unit performs pattern comparison processing on the retained segments: on the one hand, it performs morphological comparison of the channel sample contributions based on the time index recorded in the fusion feature set to identify key combinations related to structural anomalies, environmental anomalies, or intrusion signs; on the other hand, it limits the time segment and channel region of the segments based on the scene prior entries, ensuring that the identification process only takes place within the effective range. For segments that pass the pattern comparison, the edge-side processing unit generates anomaly confidence based on the robustness of the channel sample contributions, the support strength of the confidence level, and the matching completeness of the scene prior entries. It also extracts causal elements from the morphological comparison and range limitation, so that the cause of the segment's anomaly has referable entries in the three dimensions of channel, time, and scene. The output anomaly confidence and causal elements are consistent with the unified time base and establish bidirectional references with the anomaly candidate list and the fusion feature set at the entry level to ensure that subsequent structured processing, hierarchical early warning, and evidence collection triggering can accurately locate the corresponding segments and their upstream data sources. At this point, the inputs to this step are the comprehensive state score, the list of anomaly candidates and their associated fusion feature set, the unified time base, the set of anomaly fragment indices and the sensor confidence table. The outputs are the anomaly confidence score and causal elements. The naming and structure of this output are consistent with the objects used to generate the evaluation result set in the subsequent steps.

[0075] S430. The anomaly confidence level and causal elements are structured to generate an evaluation result set;

[0076] After obtaining the anomaly confidence level and the causal factors, the edge processing unit structures and organizes them to form a result object that can be directly referenced by subsequent hierarchical early warning and strategy generation. Specifically, the edge processing unit first arranges all anomaly confidence level entries according to the unified time base, and aggregates anomaly confidence level entries from different channels within the same segment into segment-level entries, while retaining the segment number and start-end position mapping relationship with the anomaly segment index set. Subsequently, the edge processing unit archives the causal factors according to the channel dimension and the scene dimension: registering key morphological descriptions involving channels and their channel sample contributions at the channel dimension; registering corresponding items and their adaptation ranges with the scene prior entries at the scene dimension; and registering the confidence level reading pointers with the sensor confidence table at the reliability dimension, so that any segment-level entry can be re-filtered or re-weighted based on reliability in subsequent steps. After completing dimensional archiving, the edge-side processing unit establishes cross-object indexes: a time index for rapid retrieval under the unified time base, a fragment index for precise location within the abnormal fragment index set, a feature index for backtracking original feature entries within the fused feature set, and a confidence index for backtracking registered values ​​in the sensor confidence table. To ensure consistency in subsequent steps, the edge-side processing unit records the source relationship between the comprehensive state score and the abnormal candidate list in the metadata of the evaluation result object, enabling any evaluation entry to be traced back to its upstream fusion evaluation process and candidate generation process. After completing the index and metadata registration, the edge-side processing unit encapsulates the integrated abnormal confidence entries and causal element entries into an evaluation result set. The evaluation result set includes a time index area, a fragment entry area, a causal element area, and a reference relationship area. The time index area strictly corresponds to the unified time base, the fragment entry area maps one-to-one with the entries in the abnormal fragment index set, the causal element area records referable entries related to channels, scenarios, and reliability, and the reference relationship area stores bidirectional references to the fused feature set, the comprehensive state score, the abnormal candidate list, and the sensor confidence table. After output, the evaluation result set is registered as a callable resource on the edge side, directly available for subsequent steps to obtain the evaluation result set for hierarchical early warning and strategy generation. It also provides a unified entry point for fragment location, upstream tracing, and reliability reading when evidence collection triggering, remote handling, and closed-loop write-back are required. In summary, this step takes the comprehensive state score and the anomaly candidate list as input. Under the constraints of the unified time base, the anomaly fragment index set, the fusion feature set, and the sensor confidence table, it first constructs anomaly confidence and causal elements, then completes structuring and indexing, and finally generates the evaluation result set. This provides a consistent, traceable, and directly referenceable data foundation for subsequent hierarchical early warning, strategy generation, and evidence collection reporting.

[0077] Step 500 includes at least steps S510-S530:

[0078] S510. Obtain the evaluation result set, perform graded early warning, and obtain the early warning level;

[0079] The evaluation result set is generated by the preceding steps and establishes a bidirectional reference with the unified time base, the abnormal segment index set, the fusion feature set, and the sensor confidence table. Specifically, the edge-side processing unit first reads the evaluation result set time-by-time according to the unified time base, merges the segment entries within the same reference time, and performs entry-level parsing on the abnormal confidence and causal elements registered in each segment entry. During the parsing process, the edge-side processing unit checks the abnormal segment index set through the reference relationship area in the evaluation result set to verify the start and end positions of the segment entries and the channels involved, ensuring that the object to be judged is consistent with the time boundary and channel identifier. Further, the edge-side processing unit performs reliability weighting on the contributions of different channels in the same segment entry based on the confidence level registered in the sensor confidence table, and retrieves the feature entry corresponding to the segment entry under the time index of the fusion feature set to confirm whether the scenario prior entry pointed to by the causal element is within the adaptation range. When the causal element involves regional climate, meter box type, or line load habits, the processing unit performs consistency verification on its effectiveness through the scenario reference relationship in the evaluation result set. After completing the fragment-level parsing, the edge-side processing unit performs segment integration within a continuous window of the unified time base. Adjacent fragment entries with related causal elements are merged into judgment segments in chronological order. An aggregation mapping is established between each judgment segment and the abnormal fragment index set, enabling subsequent strategy generation and evidence collection configuration to be executed segment by segment. After forming a judgment segment, the edge-side processing unit performs a tiered warning judgment for each segment according to a preset tiered judgment procedure. This judgment sequentially reads the abnormal confidence level, channel participation, and credibility level distribution of the fragment entries, and compares the reading results with the adaptation range of the scenario prior entries item by item. After the comparison is completed, the corresponding warning level is registered for the judgment segment under the unified time base, and reference pointers to the evaluation result set, the abnormal fragment index set, the fusion feature set, and the sensor credibility table are written into the segment metadata to ensure that subsequent steps can directly trace back to the upstream object. In summary, the input to this step is the evaluation result set and its associated objects, and the output is the warning level. The warning level is strictly aligned with the unified time base at the granularity of the segment, and is directly called as the sole basis for extracting constraints from the warning level and generating strategies.

[0080] S520. Extract constraints from the warning levels, generate a strategy, and obtain a draft handling strategy.

[0081] After obtaining the warning level, the edge processing unit uses the unified time base as a time reference to read the warning level segment by segment, and uses the associated reference pointer to look up the evaluation result set, the abnormal segment index set, and the fusion feature set to form a segment-level policy input set. Specifically, the processing unit first parses the segment entry range, involved channel set, and cause element entries corresponding to the warning level in the policy input set to confirm the target channel and target area within the segment; then, the processing unit reads the confidence level distribution of each channel within the segment through the sensor confidence level table, marks channels with higher confidence levels as priority handling channels, and marks channels with lower confidence levels and marked as suspicious participation in the cause elements as review channels, so as to form a reliability constraint for targeted handling. Furthermore, the processing unit extracts strategy constraints based on the warning level. These constraints include three types of information: execution timeliness constraints, execution scope constraints, and evidence integrity constraints. Execution timeliness constraints are determined by the start and end positions of the segment and the window setting of the unified time base, limiting the time interval for strategy effectiveness. Execution scope constraints are determined by the set of involved channels and the segment spatial description, limiting the channels and locations where the strategy operates. Evidence integrity constraints are determined by the causal element entries and segment reference pointers, limiting the source, span, and priority of evidence required for subsequent evidence collection. After extracting the constraints, the processing unit combines and sorts the strategy entries according to the warning level and strategy constraints: when the warning level is high, strategy entries with immediate execution attributes are prioritized, and priority channels are prioritized as the primary targets; when the warning level is low, strategy entries with review or monitoring attributes are prioritized, and review channels are prioritized as the primary targets. After combination and sorting, the processing unit registers an execution time indication for each strategy entry under the unified time base, and embeds a fragment reference to the abnormal fragment index set in the strategy entry to ensure that the strategy can accurately locate the segment and channel when triggered. Finally, the processing unit encapsulates the above strategy entries into a draft handling strategy, registering references to the warning level, the evaluation result set, the abnormal fragment index set, the fusion feature set, and the sensor reliability table in the draft meta-information, so that the draft handling strategy can be directly read and executed in the next step. In summary, the input of this step is the warning level and its associated objects, and the output is the draft handling strategy. This output retains the same naming and referencing methods as the upstream objects, ensuring that subsequent trigger configuration and forensic task issuance can be performed through a unified interface.

[0082] S530. Configure the execution trigger for the draft disposal strategy and generate an evidence collection trigger list;

[0083] After obtaining the draft handling strategy, the edge processing unit configures trigger conditions and binds execution parameters for the strategy entries in the draft to generate an evidence collection trigger list that can be directly invoked by subsequent evidence collection and handling processes. Specifically, the processing unit first verifies the execution time indication of each strategy entry based on the unified time base, merges overlapping or adjacent strategy entries that act on the same channel into execution segments in chronological order, and inherits the segment references from the abnormal segment index set in the execution segments so that the start and end positions of the abnormal segments can be directly located during the execution phase. Subsequently, the processing unit configures trigger conditions for each execution segment. The trigger conditions include three types of information: time trigger, event trigger, and confidence trigger. Time trigger is determined by the execution time indication and the unified time base and is used to start execution at a specified time or within a time window. Event trigger is determined by the strategy constraints and causal elements in the draft handling strategy and is used to start execution when a specified form appears in the channel or a specified state appears in the scenario. Confidence trigger is determined by the confidence level threshold read from the sensor confidence table and is used to allow execution when the confidence level meets specified boundary conditions. After the trigger conditions are configured, the processing unit binds the execution parameters, which include at least three types of information: the evidence collection object, the evidence collection method, and the evidence collection duration. The evidence collection object is determined by the set of channels associated with the execution segment; the evidence collection method is determined by the requirements for evidence integrity constraints in the draft handling strategy; and the evidence collection duration is determined by the execution time indication and the start and end positions of the segment. To ensure consistent referencing with upstream data, the processing unit records reference pointers to the evaluation result set, the warning level, and the draft handling strategy in the entries of the evidence collection trigger list, enabling the evidence to be linked back to the upstream judgment link after the evidence collection task is executed. During the entry generation stage, the processing unit also performs integrity and conflict checks on the evidence collection trigger list: In the integrity check, it checks whether each entry has at least one of time trigger, event trigger, or confidence trigger, and has an evidence collection object and an evidence collection method; in the conflict check, it checks whether there are mutually exclusive evidence collection methods among entries acting on the same channel within the same time window. If so, priority is decided and entries are merged based on the warning level and the strictness of the evidence integrity constraints. After completing the above verification, the processing unit will sort the verified items by time index according to the unified time base and encapsulate them into an evidence collection trigger list. The evidence collection trigger list records the mapping method with the abnormal fragment index set in the metadata, so that evidence collection can be carried out at the fragment level during subsequent execution. After execution, the downstream steps will register the evidence fingerprint and standardized status report. In summary, the input of this step is the draft disposal strategy and its associated objects, and the output is the evidence collection trigger list. This output is directly called by the subsequent steps of obtaining the evidence collection trigger list for evidence collection execution and summary generation, and serves as a unified entry point for evidence location and strategy implementation during remote disposal and closed-loop write-back processes.In summary, the three steps run continuously under the constraints of the unified time base and the sensor credibility table, sequentially completing the generation of the warning level, the formation of the draft handling strategy, and the output of the evidence collection trigger list. This forms a closed-loop and traceable execution link with the evaluation result set, the abnormal fragment index set, and the fusion feature set, providing a consistent, searchable, and verifiable trigger basis and execution configuration for subsequent evidence collection reporting and closed-loop write-back.

[0084] Step S600 includes at least steps S610-S630:

[0085] S610. Obtain the evidence collection trigger list, perform evidence collection execution and digest generation, and obtain evidence fingerprint and standardized status report;

[0086] In the power metering box status monitoring method based on multi-sensor fusion of the present invention, the evidence collection trigger list is generated by the preceding steps and establishes a reference relationship with the unified time base, the evaluation result set, the abnormal segment index set, the fusion feature set, and the draft handling strategy. Specifically, the edge processing unit first reads the evidence collection trigger list item by item according to the unified time base, and performs executability checks on the time trigger, event trigger, and confidence trigger in each item: in the time trigger check, the window definition of the unified time base is used as a reference to confirm that the evidence collection start time and duration are both within the execution interval registered in the list; in the event trigger check, the constraints recorded in the draft handling strategy are matched with the cause elements registered in the evaluation result set to confirm that the form or state pointed to by the trigger condition has been identified in the corresponding segment of the abnormal segment index set; in the confidence trigger check, the confidence level of the sensor confidence table at the corresponding time and corresponding channel is directly read to confirm that it meets the threshold boundary in the list item. After completing the executability check, the edge processing unit expands the evidence collection objects pointed to by the list entries. These objects include the relevant channels and the start and end positions of the corresponding segments. For the relevant channels, the edge processing unit uses the sensor reliability table as a constraint to determine the priority order of evidence collection methods and the allocation ratio of evidence collection time. For the start and end positions of segments, the edge processing unit uses the unified time base as a reference to align the evidence collection window with the segment boundaries, ensuring that the subsequently generated evidence data has a traceable time index. Subsequently, the edge processing unit calls the evidence collection device or function associated with the relevant channels to execute the evidence collection process, including image capture, short-time series acquisition, and related metadata archiving. During the evidence collection process, the edge processing unit performs integrity registration of the original evidence data, records the time index, channel identifier, and segment reference pointer, and establishes a bidirectional reference between the registration results and the abnormal segment index set for subsequent segment-level comparison of the evidence source. To ensure consistent verification of evidence in subsequent remote processing stages, the edge processing unit generates evidence fingerprints for the evidence data after evidence collection. Each evidence fingerprint corresponds one-to-one with an evidence item and includes a time index, channel identifier, fragment reference pointer, and evidence collection method summary. The fingerprints are also registered in the metadata for their reference relationships with the evaluation result set, the warning level, and the draft processing strategy. Next, the edge processing unit, based on the standardization requirements, uniformly expresses the aligned data, fragment descriptions, and evidence collection methods involved in this evidence collection, forming a standardized status report. This standardized status report uses the unified time base as the primary time key and the involved channels as subordinate entries, centrally summarizing the start and end positions of fragments, causal element references, and credibility level reading pointers, enabling the report entries to be retrieved item by item in subsequent remote processes.At this point, the input for this step is the evidence trigger list and its associated objects, and the output is the evidence fingerprint and the standardized status report. When these two are generated, they are bound to the unified time base, the abnormal fragment index set, and the sensor credibility table, serving as direct input for the next step to extract key records from the evidence fingerprint and the standardized status report, perform remote processing, and verify receipts.

[0087] S620. Extract key records from the evidence fingerprint and standardized status report, perform remote processing and verification receipt, and obtain processing result receipt;

[0088] After obtaining the evidence fingerprint and the standardized status report, the main station processing unit performs synchronous analysis and verification on both. Specifically, the main station processing unit first matches the time primary key in the standardized status report with the time index in the evidence fingerprint based on the unified time base, merging evidence entries at the same reference time into remote review units. During the merging process, the main station processing unit uses the segment number and start and end positions provided by the abnormal segment index set as a reference to perform boundary checks on the evidence entries, ensuring that the evidence entries cover the valid segments of the segments. Subsequently, the main station processing unit extracts key records from the standardized status report. The key records include at least the channel, segment description, causal element reference, and confidence level reading pointer. The main station processing unit then checks back the evaluation result set to verify whether the anomaly confidence level and causal element description corresponding to the key record are consistent with the edge-side judgment. At the same time, the main station processing unit uses the sensor confidence table as the reading entry point to confirm whether the confidence registration of the channel in the key record under the corresponding time index meets the reliability requirements of remote processing. After completing the extraction and consistency verification of key records, the main station processing unit generates a set of remote handling instructions according to the strategy category and execution order consistent with the draft handling strategy. This set of instructions is then aligned with the unified time base, clarifying the effective segment and action channel of each instruction. In cases requiring evidence review, the main station processing unit performs evidence integrity verification on the remote side based on the evidence fingerprint. The verification includes three elements: the time index of the evidence item, the channel identifier, and the fragment reference pointer. If any missing or inconsistent elements are found, supplementary evidence collection or a return request is triggered. After the remote handling instruction set is issued and executed, the main station processing unit collects execution feedback from the edge side or related execution units, forming a handling result record. This record is then aligned with the unified time base and linked back to the corresponding item in the standardized status report. Finally, the main station processing unit structures the handling results into a handling result receipt. This receipt includes a time index, involved channels, execution instruction description, execution status, and evidence reference pointers. The metadata records the reference methods to the evaluation result set, the warning level, the draft handling strategy, the evidence fingerprint, and the standardized status report, so that subsequent learning updates can complete item-level referencing and quality assessment simultaneously. Thus, the input to this step is the evidence fingerprint and the standardized status report, and the output is the handling result receipt. This output maintains consistency in naming and structure with subsequent learning update steps, serving as a direct basis for updating the baseline and adjusting the scenario priors.

[0089] S630. The processing result receipt is learned and updated to generate a baseline update package;

[0090] After receiving the processing result receipt, the edge side and the main station side collaboratively conduct learning updates under the same time reference. Specifically, the main station side processing unit first arranges the processing result receipts chronologically according to the unified time base, and checks the evidence fingerprint and the standardized status report back according to the evidence reference pointer in the processing result receipt, and confirms the completeness of the prior evidence and judgment source for each processing result record; during the confirmation process, the main station side processing unit checks the consistency of the channel and execution instruction descriptions involved in the record, and synchronously imports the abnormal confidence and causal elements corresponding to the evaluation result set to form the judgment-execution-evidence triplet required for learning updates. Subsequently, the main station processing unit uses the sensor confidence table as a reliability reference to aggregate triples according to time indexes. Triples with the same segment number or adjacent time windows and involving consistent channels are combined into learning segments. Within the learning segments, the record coverage of the channel layer and the registration completeness of the evidence layer are calculated. During this aggregation process, if a learning segment has incomplete evidence registration or insufficient reliability, the main station processing unit will trigger a supplementary lookup through bidirectional references with the abnormal segment index set until the learning segment has the minimum registration elements. After completing the aggregation of learning segments, the main station processing unit performs regularization on the learning segments: registering the start and end positions in the time dimension using the unified time base as the primary key; registering readable pointers involving channels and channel sample contributions in the channel dimension; registering corresponding entries and adaptation ranges with the scene prior in the scene dimension; and registering directly readable confidence levels in the reliability dimension. The above regularization establishes a reference relationship with the fused feature set for lookup of feature entries when needed. Based on the rule-based learning segments, the main station processing unit performs learning updates: on the one hand, based on the execution status of the handling result receipt record, it adjusts the adaptation range and fine-tunes the time segment of the scenario prior items within the corresponding time period; on the other hand, based on the evidence integrity and reliability registration, it registers and revises the credibility level of the corresponding channel in the sensor credibility table within the corresponding time period; at the same time, based on the matching degree of the judgment-execution-evidence triplet, it updates the parameters of the judgment threshold and item screening order used in subsequent graded early warning, ensuring that the parameter updates are consistent with the existing items in terms of naming and indexing methods. The aforementioned adjustments, revisions, and updates are uniformly packaged into a baseline update package. The baseline update package includes a time index area, a priori adjustment area, a confidence revision area, and a threshold update area. The metadata records reference pointers to the unified time base, the evaluation result set, the abnormal segment index set, the fusion feature set, and the sensor confidence table. This ensures that the updated content can be directly loaded in the next round when feature construction is performed from the aligned data and the abnormal segment index set, fusion evaluation is performed from the fusion feature set, and graded early warning is performed from the evaluation result set.The generated baseline update package is written back to the configuration storage that is visible to both the edge side and the main station side. When the edge side performs the multi-source perception and health fingerprint modeling in a new round, it reads the confidence revision and threshold update in the package. When the main station side performs the hierarchical early warning and policy generation in a new round, it reads the prior adjustment and policy parameters in the package, so that the entire link remains consistent in terms of time reference and object reference. In summary, the three consecutive steps, under the constraints of the unified time base, the abnormal fragment index set, the fusion feature set, the evaluation result set, the sensor credibility table, and the draft handling strategy, sequentially complete the closed-loop processing of evidence collection and summary generation, remote handling and verification receipt, and learning update and baseline encapsulation, forming an ordered output of evidence fingerprints, standardized status reports, handling result receipts, and baseline update packages. The technical effect of this step is to make the upstream judgment and execution processes evidence-based, structured, and rewritable under the same time reference, so that subsequent loops can carry out alignment, feature construction, fusion evaluation, and hierarchical early warning with consistent data objects, unified indexing methods, and continuous parameter updates, thereby forming a searchable, verifiable, and continuously updated closed loop within the method chain.

Claims

1. A multi-sensor fusion based power metering box condition monitoring method, characterized in that, Comprise: S100, acquire multi-source perception data, based on displacement perception, temperature and humidity perception, water immersion perception and infrared perception data, through edge side collection device initialization, range confirmation and zero point check, and execute trigger threshold self-checking and short time stability checking to collect and calibrate, extract time information from multi-source raw data stream, rearrange based on time sequence, sample interpolation, sample cutting and sample resampling consistency processing for time alignment, based on basic health quantization index including channel activity, data continuity, short time stability and loss record rate, normalization processing and abnormal mark registration for health fingerprint modeling, generate multi-source raw data stream, unified time base and sensor reliability table; Wherein, the health fingerprint modeling, specifically includes: The edge side summarizes the sample reading, sample effective mark and time index of the corresponding channel at each time, forming a channel state segment that can be analyzed by time; The edge side normalizes the channel state segment according to the basic health quantization index, forming a set of health marks comparable across channels; For the state segment with abnormal jump or long-time blank, the edge side registers the abnormal mark in the unified time base, and associates the abnormal mark with the channel state segment; Wherein, generating the sensor reliability table, specifically includes: The edge side summarizes the long-term stability at the channel level to generate a channel stability identifier; Archive the access quality at the collection link level to generate an access reliability identifier; Update the running state of the sensor unit at the device level to generate a running health identifier; The channel stability identifier, access reliability identifier and running health identifier are combined with the basic health quantization index and abnormal mark to form a multi-dimensional health fingerprint entry, and are indexed and stored in the database according to the unified time base; Through the channel-by-channel and time-by-time aggregation of the health fingerprint entry, the edge side generates the sensor reliability table; S200, based on the multi-source raw data stream, the unified time base and the sensor reliability table, perform time sequence alignment, denoising, abnormal positioning and summary processing to generate aligned data and abnormal segment index set; Specifically includes: Obtain the multi-source raw data stream, the unified time base and the sensor reliability table, perform time sequence alignment and robust denoising to obtain clean data after alignment; Including: the channel sample with higher reliability level as the reference track for time drift checking, the channel sample with lower reliability level and continuous loss or mutation is preferentially put into the correction or rejection channel; From the clean data after alignment, locate the abnormal interval, perform segment labeling, and obtain the abnormal segment index; Summarize and arrange the clean data after alignment and the abnormal segment index to generate aligned data and abnormal segment index set; S300, based on the alignment data and the abnormal segment index set, based on the unified time base timing window, segment-level aggregation processing and weighted integration are performed for feature construction, based on regional climate information, meter box type information, line load habit information and historical baseline information, entry-level mapping relationship is established for scene prior association, based on the unified time base, field naming is unified, value domain is unified and default value is filled, and bidirectional reference with the abnormal segment index set and the sensor credibility table is established for standardized index processing, and a fusion feature set is generated; Specifically, it includes: The edge side processing unit establishes an entry-level mapping relationship between the channel sample contribution of the fusion feature vector and the scene prior, establishes a corresponding relationship between the seasonal section and the temperature and humidity characteristics for the regional climate information, establishes a corresponding relationship between the structure area and the displacement characteristics for the meter box type information, establishes a corresponding relationship between the energy use section and the infrared characteristics for the line load habit information, and records the segment reference pointer and the time index in each mapping relationship; The edge side processing unit establishes bidirectional reference between the standardized feature entries and the abnormal segment index set: on the one hand, record the corresponding segment reference pointer in the feature entry for back checking, on the other hand, register the directly retrievable feature entry index in the abnormal segment entry for forward checking; At the same time, register the credibility reading pointer between the feature entry and the sensor credibility table; S400, based on the fusion feature set, perform fusion evaluation, abnormal pattern recognition and structured processing to generate an evaluation result set; Specifically, it includes: Obtain the fusion feature set, perform fusion evaluation, and obtain a comprehensive state score and an abnormal candidate list; Extract key elements from the comprehensive state score and the abnormal candidate list, perform abnormal pattern recognition, and obtain abnormal confidence and cause elements; Perform structured processing on the abnormal confidence and cause elements to generate an evaluation result set; S500, based on the evaluation result set, perform hierarchical early warning, strategy generation and trigger configuration processing to generate a forensic trigger list; Specifically, it includes: Obtain the evaluation result set, perform hierarchical early warning, and obtain a warning level; Extract constraint conditions from the warning level, perform strategy generation, and obtain a disposal strategy draft; Perform execution trigger configuration on the disposal strategy draft to generate a forensic trigger list; S600. Based on the evidence collection trigger list, perform evidence collection execution, remote handling, and learning update processing to generate a baseline update package; specifically, this includes: using the sensor credibility table as a reliability reference, aggregating the judgment-execution-evidence triples according to the time index, and merging triples with the same segment number or adjacent time windows and involving the same channel into learning segments; after regularizing the expression of the learning segments, adjusting the adaptation range and fine-tuning the time segment of the scenario prior items within the corresponding time period based on the execution status recorded in the handling result receipt; registering and revising the credibility level of the corresponding channel in the sensor credibility table within the corresponding time period based on the evidence integrity and reliability registration constraints; updating the judgment threshold and item selection order used in subsequent graded early warnings based on the matching degree of the judgment-execution-evidence triples; the above adjustments, revisions, and updates are uniformly packaged into a baseline update package containing a time index area, a priori adjustment area, a credibility revision area, and a threshold update area.

2. The method of claim 1, wherein, Multi-source sensing data includes: The multi-source sensing data specifically includes sensing data acquired by displacement sensing units, temperature and humidity sensing units, water immersion sensing units, and infrared sensing units installed inside and outside the power metering box structure through edge-side acquisition devices. Among them, displacement sensing data is used to monitor the physical displacement and vibration state of the metering box structure, temperature and humidity sensing data is used to collect environmental temperature and humidity parameters inside the box, water immersion sensing data is used to detect liquid intrusion around or inside the box, and infrared sensing data is used to sense external heat sources or abnormal temperature rise inside the box.

3. The method of claim 1, wherein, The sensor reliability table includes: The sensor credibility table is a structured collection of credibility registrations, including channel identifier, time index, health fingerprint entries, and credibility level fields; the sensor credibility table is bound to a unified time base when it is generated.

4. The method of claim 1, wherein, The process of generating the sensor confidence table also includes: Obtain a sensor confidence level table to prioritize the correction or removal of channel samples with low confidence levels; When locating abnormal regions and labeling segments, refer to the abnormal markers in the sensor confidence table; When generating the aligned data and anomalous fragment index set, the confidence level in the sensor confidence table is referenced.

5. The method of claim 1, wherein, The sensor reliability table also includes: The sensor credibility table serves as a weighting basis, channel reliability constraint, threshold reference, participation ratio control, evidence collection channel priority determination basis, and evidence source reliability level. It is applied to the entire subsequent methodological chain from feature fusion to remote processing and learning updates.

6. The method of claim 1, wherein, The process of segment-level aggregation and weighted integration for feature construction based on a unified time base time series window includes: The edge-side processing unit constructs a time-series window with a unified time base as a reference, performs segment-level aggregation processing on the channel samples within the window, and forms a basic quantitative description applicable to the unified feature space. After the basic quantitative description is formed, the edge-side processing unit performs weighted integration of the sample contributions of each channel within the window according to the confidence level registered in the sensing confidence table, so that the multi-channel samples at the same time and in the same segment are comparable and combinable.

Images