A bearing running state edge computing monitoring method and system

Abstract

The application relates to computer data processing and edge computing monitoring technology, in particular to a bearing operation state edge computing monitoring method and system. An edge computing node acquires and time-aligns bearing operation data and working condition data, divides working condition segments according to working condition change rates and operation continuity, retrieves health reference segments with adjacent working conditions and consistent evolution directions from a health sample library, reconstructs counterfactual health responses and calculates fault reservation residuals; a splittable and mergable multi-scale time window tree is constructed based on the residuals, abnormal evidence is extracted and cross-scale evidence consistency is established, the time window boundary and the health reference segment are corrected when the evidence conflicts, the bearing operation state is determined when the evidence is consistent, and the abnormal working condition segment is frozen and limited to be written into the health sample library. The application can separate working condition change responses and fault responses, reduce working condition switching misjudgment, retain fault features, and inhibit abnormal data pollution to the health sample library.

Description

Technical Field

[0001] This invention relates to computer data processing and edge computing monitoring technology, specifically to a method and system for edge computing monitoring of bearing operating status. Background Technology

[0002] Bearing operating status monitoring typically involves acquiring vibration, temperature, drive current, speed, and load data via a data acquisition link on the equipment side. Edge computing nodes then perform data alignment, segmentation, feature extraction, and status determination. Existing solutions often use fixed-duration windows to capture operating data, calculating time-domain amplitude, frequency-domain energy, impact factor, envelope spectrum peak value, or temperature rise trend within the window. The resulting features are then compared with preset thresholds, historical averages, or classification models trained offline. Some solutions establish multiple normal sample sets according to speed or load ranges, using the corresponding sample as a reference when the current operating condition falls within that range. Other solutions use moving averages, standardization, or periodic resampling to mitigate the impact of operating condition changes on monitoring features. For continuous monitoring results, the bearing operating status is typically determined by multiple consecutive windows exceeding a threshold, multiple features simultaneously exhibiting anomalies, or classification probabilities reaching set conditions. Data that did not trigger anomalies are added to the normal sample set to update the monitoring baseline or retrain the status determination model. The above scheme can complete routine monitoring under stable operating conditions, but health reference data, time window boundaries, anomaly detection process and sample update rules are usually independent of each other, and lack a data processing process for correlation correction based on monitoring results.

[0003] During periods of speed change, load change, start-stop transitions, and long-term wear, the normal operating response of bearings continuously changes with the operating condition trajectory, and may exhibit data characteristics similar to early faults at local time intervals, specific frequency bands, and different time scales. Existing solutions, using fixed operating condition intervals, global averages, or static classification boundaries as comparison benchmarks, struggle to generate healthy responses that match the initial state, direction of change, and evolution sequence of the current operating condition segment. The resulting discrepancy data simultaneously includes both normal operating condition changes and fault components. Fixed time windows and single-scale discrimination cannot correlate and verify the impact recurrence, periodic continuation, and frequency band migration in the discrepancy data. Normal operating condition transitions may be retained as anomalous, and early faults may be weakened during standardization or long-term window processing. Anomalous segments misjudged as normal are written into the healthy sample library and repeatedly used as reference data in subsequent baseline generation, causing the healthy reference data to gradually deviate from the bearing's true health state. Therefore, the main technical problem of existing bearing edge monitoring methods is that they lack a closed-loop data processing mechanism that can generate a healthy reference according to the current operating condition evolution process, perform cross-scale consistency verification on the remaining differences, and constrain the updating of healthy samples based on the verification results. This makes it difficult to stably distinguish between operating condition change response and fault response. Summary of the Invention

[0004] The application aims to provide a bearing operation state edge computing monitoring method and system, which can solve the problems in the background art.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the application is: A bearing operation state edge computing monitoring method, comprising: acquiring and time-aligning bearing operation data and corresponding working condition data by an edge computing node, dividing working condition segments according to the change rate and operation continuity of the working condition data; Retrieving a healthy reference segment from a healthy sample library, which is adjacent to the current working condition segment and consistent in evolution direction, and weighting and reconstructing the counterfactual healthy response according to the working condition position correspondence relationship between the two; Calculating the fault reservation residual between the bearing operation data and the counterfactual healthy response; Building a splittable and mergable multi-scale time window tree around the fault reservation residual, extracting abnormal evidence of each scale and establishing a cross-scale evidence consistency relationship; When the cross-scale evidence consistency relationship does not meet the consistency condition, the time window boundary and the healthy reference segment are corrected according to the evidence conflict position, and the fault reservation residual is regenerated; When the cross-scale evidence consistency relationship meets the consistency condition, the bearing operation state is determined according to the consistent abnormal evidence, and the current working condition segment is frozen and prohibited from being written into the healthy sample library when the bearing operation state is abnormal.

[0006] Preferably, dividing the working condition segments and retrieving the healthy reference segment comprise: reconstructing the working condition coordinates with the cumulative change amount of the working condition variables, and representing each working condition segment as a working condition trajectory index containing the starting state, the ending state, the change direction and the change sequence; Generating a bidirectional transition signature according to the forward adjacent segment and the backward adjacent segment of the current working condition segment; Screening candidate segments in the healthy sample library, which fall within the current working condition neighborhood and have consistent bidirectional transition signatures; Positionally corresponding the candidate segments and the current working condition segment along the working condition coordinates, and determining the candidate weight according to the working condition adjacency, the transition signature consistency and the position correspondence residual; Eliminating candidate segments with inconsistent forward and backward corresponding results, and determining the remaining candidate segments and the corresponding candidate weights as the healthy reference segment set.

[0007] Preferably, building the multi-scale time window tree and establishing the cross-scale evidence consistency relationship comprise: taking the time window covering the current working condition segment as the root time window, generating candidate split points according to the local energy aggregation position, the sign flip density and the periodic recurrence position of the fault reservation residual; A sub-time window is established for the residual sequence on both sides of the candidate split point, and the parent time window is restored when the dominant period, phase continuation relationship and frequency band migration direction of adjacent sub-time windows meet the merging conditions; Impulse sparsity evidence, period recurrence evidence, frequency band migration evidence and envelope phase evidence are formed within each time window; The time window and the corresponding abnormal evidence are taken as evidence nodes, and the evidence nodes are connected according to time overlap, integer period correspondence and parent-child scale inclusion relationship to form the cross-scale evidence consistency relationship.

[0008] Preferably, the time window boundary and the health reference segment are corrected, including: determining the conflicting abnormal evidence in the cross-scale evidence consistency relationship as a conflict evidence node, dividing the evidence conflict into boundary conflict, reference conflict and isolated conflict according to the position of the conflict evidence node in the multi-scale time window tree; For the boundary conflict, the split point is re-determined according to the local extreme value of residual continuity in the adjacent region of the conflict evidence node, and the corresponding parent-child time window is reconstructed; For the reference conflict, the local reconstruction components of each health reference segment at the conflict frequency band and the conflict time position are extracted, the median change direction of each local reconstruction component is determined as the reference direction, and the counterfactual health response is regenerated after removing the health reference segment whose local reconstruction component is opposite to the reference direction; For the isolated conflict, the corresponding working condition segment is marked as a to-be-confirmed segment and the health sample library remains unchanged; A frozen identifier and an associated evidence node set are generated for the working condition segment that meets the consistency condition.

[0009] Preferably, the counterfactual health response is generated according to the health reference segment set, including: mapping each health reference segment to a uniform position sequence along the working condition coordinate, and forming a health response set corresponding to the position according to the candidate weight; Each health reference segment is sequentially removed from the health response set, and the remaining health reference segments are used to reconstruct the removed health reference segment, and a reference reliability sequence is formed according to the reconstruction deviation of each position; A position interval with continuous and stable reliability is selected from the reference reliability sequence, and a local response gradient between the working condition coordinate increment and the health response increment is determined based on the position interval; The weighted response of the health response set at the starting position of the current working condition segment is taken as the recursive starting point, the health response is generated position by position according to the adjacent working condition coordinate increment and the corresponding local response gradient, and the generated health response is corrected according to the response continuity relationship at the end point of the adjacent working condition segment, to obtain the counterfactual health response and the counterfactual reliability corresponding to each position.

[0010] Preferably, the establishing of the cross-scale evidence consistent relationship further comprises: recording a time interval, a scale level, a frequency band level, a dominant period and a phase trajectory for each evidence node; For the evidence nodes with the parent-child scale containing relationship, a first connection weight is set according to a reserved proportion of the abnormal evidence in the parent time window and the child time window; For the evidence nodes with the integer period corresponding relationship, a second connection weight is set according to a period deviation and a phase trajectory continuity; For the evidence nodes located at adjacent frequency band levels, a third connection weight is set according to a consistent degree of the frequency band migration direction; According to the connection weights from high to low, the evidence marks are transmitted along the parent-child scale connection, the integer period connection and the frequency band adjacency connection respectively, the number of independent paths reaching the same evidence node and the adjacent period recurrence marks are recorded, the evidence nodes reached by only a single path and without the adjacent period recurrence marks are deleted, and the cross-scale evidence consistent relationship is updated.

[0011] Preferably, it further comprises: encapsulating the working condition segment with the frozen mark, the corresponding working condition segment index, the counterfactual health response, the fault reserved residual error, the multi-scale time window path, the evidence node set and the conflict correction record into a frozen evidence package, generating a version mark according to the arrangement order of the data in the package, and subsequently updating to establish a new frozen evidence package associated with the previous version; For the candidate working condition segment which is not frozen and is intended to be written into the health sample library, the health working condition segments before and after the candidate working condition segment are executed for counterfactual reconstruction under the reference configuration containing the candidate working condition segment and the reference configuration not containing the candidate working condition segment respectively, the evidence nodes and the time window boundaries corresponding to the two kinds of reference configurations are compared, and only the candidate working condition segment which does not introduce new conflict evidence nodes and maintains the original evidence node connection relationship is written into the health sample library; According to the time window path and the evidence node topology correspondence between the frozen evidence package and the existing fault prototype, the frozen evidence package is incorporated into the corresponding fault prototype or a new fault prototype is established.

[0012] Preferably, it further comprises: recording a health reference segment subset participating in reconstruction at each time position, forming a credibility gradient according to the time variation of the counterfactual credibility, finding the intersection of the mutation position of the credibility gradient and the local mutation position of the fault reserved residual error, and taking the obtained position as an initial candidate split point of the multi-scale time window tree; For the two abnormal evidences connected in the cross-scale evidence consistent relationship, the low counterfactual credibility of the corresponding positions of the two abnormal evidences and the overlapping proportion of the two health reference segment subsets are calculated, and when both are lower than the adjacent evidence connection value, the connection between the two abnormal evidences is deleted; When the evidence conflict focuses on a position where the counterfactual credibility is lower than the adjacent interval, the conflict position is fed back to the reference reliability sequence, and the healthy reference segment involving the conflict position is executed to move out reconstruction; When the abnormal evidence is cross-scale repeated at the position where the counterfactual credibility is continuously stable, a lock identifier is set for the corresponding time window path, and the boundary is re-determined only for the time window which is not set with the lock identifier.

[0013] Preferably, it further comprises: when the healthy reference segment is re-screened due to evidence conflict, a candidate version is established based on the current healthy sample library, the sample addition and deletion relationship between the candidate version and the original version, the counterfactual healthy response change relationship, and the reference relationship of each counterfactual healthy response to the healthy reference segment are recorded; The affected frozen evidence package is determined according to the sample addition and deletion relationship and the reference relationship, the working condition segment index and the multi-scale time window path are extracted from the frozen evidence package, and the directional reconstruction is performed on the corresponding working condition segment; After the directional reconstruction, when the abnormal evidence node in the frozen evidence package is not deleted, the determined bearing operating state is not changed, and the time window boundary of the adjacent working condition segment is continuous, the current healthy sample library version is replaced with the candidate version; Otherwise, the candidate version is revoked, and the healthy reference segment, the fault retention residual and the evidence node connection relationship are restored according to the counterfactual healthy response change relationship.

[0014] A bearing operating state edge computing monitoring system, comprising an edge computing node with a memory and a processor, the memory stores a monitoring program, and the processor executes the monitoring program to: acquire and time-align bearing operating data and working condition data, and divide working condition segments according to working condition change rate and operating continuity; Retrieving a healthy reference segment from a healthy sample library which is adjacent to the current working condition segment and consistent in evolution direction, and reconstructing a counterfactual healthy response according to the working condition position relationship between the two; Calculating a fault retention residual of the bearing operating data relative to the counterfactual healthy response; Based on the fault retention residual, a splittable and merged multi-scale time window tree is constructed, abnormal evidence of each scale is extracted, and cross-scale consistent relationship is established; When the relationship does not meet the consistent condition, the time window boundary and the healthy reference segment are corrected according to the evidence conflict position, and the fault retention residual is recalculated; When the relationship meets the consistent condition, the bearing operating state is determined according to the consistent abnormal evidence, and when the bearing operating state is abnormal, the current working condition segment is frozen and prohibited from being written into the healthy sample library.

[0015] Compared with the prior art, the beneficial effects of the present application are that: 1. According to the change rate and running continuity of working condition data, the working condition segment is divided, the working condition trajectory index is formed by starting state, ending state, change direction and change order, and the health reference segment with adjacent working condition and consistent evolution direction is screened from the health sample library. After the position correspondence, candidate weighting and removal reconstruction of each health reference segment, the counterfactual health response is generated, so that the comparison benchmark of bearing running data changes with the current working condition trajectory, and the uniform reference value in the fixed working condition interval is no longer used. After the formation of the fault reservation residual between the bearing running data and the counterfactual health response, the multiscale time window tree that can be split and combined is established around the fault reservation residual, and the cross-scale evidence consistent relationship is composed of impact sparse evidence, periodic recurrence evidence, frequency band migration evidence and envelope phase evidence. When evidence conflict occurs, the time window boundary is re-determined according to the conflict position, the health reference segment inconsistent with the reference direction is removed and the fault reservation residual is re-generated; when the evidence meets the consistent condition, the bearing running state is determined. Thus, the normal working condition change is followed in the health reference generation process, the health reference deviation is corrected in the cross-scale verification process, and the fault response is reserved through the evidence association in multiple scales and adjacent periods, reducing the mutual confusion of working condition change and early fault.

[0016] 2. For the working condition segment determined to be abnormal, the scheme generates a frozen identifier, saves the working condition segment index, counterfactual health response, fault reservation residual, multiscale time window path, evidence node set and conflict correction record, and prohibits writing the corresponding working condition segment into the health sample library. The candidate working condition segment to be written into the health sample library is executed for counterfactual reconstruction under the reference configuration containing and not containing the candidate working condition segment. Only when no new conflict evidence node is generated and the original evidence node connection relationship is maintained, the candidate working condition segment is written into the health sample library, so that the health sample update is constrained by the existing monitoring evidence. When the health reference segment is re-screened, the affected frozen evidence package is determined according to the sample addition and deletion relationship, the counterfactual health response change relationship and the reference segment reference relationship, the directional reconstruction is performed on the related working condition segment, and the health sample library version is replaced when the abnormal evidence node, bearing running state and adjacent time window boundary are consistent, otherwise the original version is restored. The processing forms a health sample update path that can be traced back, verified and revoked, prevents abnormal segments from continuously participating in the generation of subsequent counterfactual health responses, and maintains the corresponding relationship between the historical state determination result, monitoring evidence and health sample library version. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The overall flowchart of the bearing running state edge computing monitoring method is shown in the figure; Figure 2 The working condition division quantity and health division threshold change curve diagram is shown in the figure; Figure 3Flowchart for constructing a cross-scale evidence association process for counterfactual health responses; Figure 4 A weight distribution diagram of candidate reference segments; Figure 5 For reference, see the reliability and counterfactual credibility sequence diagram; Figure 6 Residual differential channel diagrams are preserved for operational response and faults; Figure 7 A cross-scale evidence connectivity weight graph; Figure 8 A map showing the number of candidate targets for border conflicts; Figure 9 Replace the decision graph with a candidate version from the sample library. Detailed Implementation

[0018] In one embodiment, please refer to Figure 1 An edge computing monitoring method for bearing operating status includes: acquiring and time-aligning bearing operating data and corresponding working condition data from edge computing nodes; dividing the working condition data into segments based on the rate of change and continuity of operation; retrieving healthy reference segments from a health sample library that are adjacent to the current working condition segment and have the same evolution direction; reconstructing a counterfactual health response based on the correspondence between the two working condition positions; calculating the fault retention residual between the bearing operating data and the counterfactual health response; constructing a splittable and mergeable multi-scale time window tree around the fault retention residual; extracting anomalous evidence at each scale and establishing cross-scale evidence consistency relationships; when the cross-scale evidence consistency relationship does not meet the consistency condition, correcting the time window boundary and the healthy reference segment based on the evidence conflict position and regenerating the fault retention residual; when the cross-scale evidence consistency relationship meets the consistency condition, determining the bearing operating status based on consistent anomalous evidence, and freezing the current working condition segment and prohibiting it from being written to the health sample library when the bearing operating status is anomalous.

[0019] In this embodiment, the bearing operation data is continuous time-series data formed by the existing acquisition link of the monitored equipment. It includes at least vibration data that can characterize the mechanical response of the bearing, and may also include temperature data or drive current data. The operating condition data includes at least speed data or load data that changes synchronously with the mechanical response. Edge computing nodes establish operation data records with a unified time identifier. A single record consists of time position, operation data vector, operating condition data vector, data integrity marker and source marker. During the processing, the original amplitude is not directly compared with a fixed threshold. Instead, the operating condition segment construction, health reference retrieval, counterfactual response reconstruction, residual generation, multi-scale evidence association and sample writing control are completed sequentially. This ensures that the health benchmark used for the current state determination corresponds to the current operating condition evolution process, and that subsequent updates to the health sample library are constrained by the current state evidence.

[0020] Specifically, time alignment is achieved using a common time axis and neighborhood matching. Edge computing nodes use the consecutive arrival times in the bearing operation data as target positions. Within the defined neighborhood of each target position, they search for operating condition data records. When multiple candidate operating condition records exist, the record with the smallest time difference and the one that maintains a continuous order with the previous matched record is selected. When no operating condition record exists in the neighborhood, piecewise linear interpolation is performed based on adjacent valid operating condition records, and interpolation marks are set for the interpolation positions. When the continuous missing range crosses the turning point of the operating condition, cross-turn interpolation is not performed, and the current segment is terminated at both ends of the missing range. This avoids forcibly classifying data that do not have the same evolutionary relationship into the same operating condition segment. After alignment is completed, only unit unification, outlier truncation, and time order verification are performed on each data channel, without covering the historical benchmark with the statistics of the current window.

[0021] The boundary of a working condition segment is determined by the degree of change in working condition, the degree of continuity of operation, and the degree of directional change. Edge computing nodes calculate the working condition boundary for adjacent time positions.

[0022] The operating condition boundary value is obtained by weighted summation of three indicators: The first item is the change in operating conditions per unit time: it is calculated by dividing the Euclidean distance between the operating condition data vectors of two adjacent time positions by the time interval between the two positions; The second item is the degree of operational discontinuity: it is calculated by subtracting the operational continuity of the current position from the value 1, where the continuity is calculated based on the data arrival order, interpolation markers, and equipment operation markers. The third item is the direction change amount: it represents the degree of change in the current operating condition direction relative to the previous direction of change.

[0023] Each of the three indicators is multiplied by its corresponding non-negative weighting coefficient, and the sum of the three weighting coefficients is 1. The final calculation result is the working condition boundary at the current time position.

[0024] When the operating condition boundary exceeds the boundary threshold determined by the distribution of the boundary of similar healthy operating condition segments, a segment boundary is established at that time position.

[0025] For example, if the three weighting coefficients are set to 0.5, 0.3, and 0.2 respectively, and the normalized rate of change at a certain position is 0.4, the continuity is 0.9, and the turning point is 0.6, the calculated boundary value is 0.35. If the health boundary threshold is 0.3, then this position is determined as the starting point of a new operating condition segment. This design ensures that no factor among sudden changes in operating conditions, data discontinuities, and directional turning points can be completely masked by other factors.

[0026] Please refer to Figure 2 , Figure 2The horizontal axis represents the unified position stage within the current working condition segment, unfolding sequentially from the starting point, the front segment, the front-middle segment, the middle segment, the rear-middle segment, the rear segment, and the end point; the vertical axis represents the working condition boundary quantity calculated by the edge calculation nodes based on the degree of working condition change, the degree of operational continuity, and the degree of directional change. Figure 2 The solid line represents the operating condition boundary curve, and the dashed line represents the health boundary threshold determined by the distribution of the boundary values ​​of similar healthy operating condition segments. Together, they are used to show whether the current operating data forms a new segment boundary at adjacent positions.

[0027] from Figure 2 As can be seen, the working condition boundary value at the starting point of the working condition segment shows a significant surge, and is significantly higher than the health boundary threshold. This indicates that this position is simultaneously affected by factors such as changes in working conditions, a decrease in continuity, or a change in direction, which is consistent with the processing logic of establishing the segment boundary at this position. After entering the first segment, the working condition boundary value drops rapidly and remains below the threshold for a long period, with only slight fluctuations in the middle and later segments. This indicates that the working condition evolution of the main part of this segment remains continuous, and no abrupt change position sufficient to split the segment has been formed.

[0028] The health sample library is organized according to operating condition segments rather than individual sampling points. Each health sample record saves the start and end times of the segment, the operating condition trajectory index, the running response sequence, the adjacency relationship before and after the segment, the written version, the reference status, and the historical reconstruction results. The operating condition trajectory index is obtained by encoding the cumulative change of the operating condition variables. The cumulative change is used to represent the position of the current segment in the operating condition space, the change direction is used to distinguish between acceleration, deceleration, loading, unloading, or stable operation, the change order is used to record the relationship between the sequential changes of multiple operating condition variables, and the adjacency relationship is used to describe the operating condition transition before and after the segment enters. This makes health segments with similar values ​​but different evolutionary sources distinguishable and avoids using data with the same instantaneous speed but in the acceleration and deceleration phases as the same reference.

[0029] In a preferred embodiment, please refer to Figure 3 The process of dividing operating condition segments and retrieving healthy reference segments includes: reconstructing operating condition coordinates based on the cumulative changes in operating condition variables; representing each operating condition segment as an index of operating condition trajectory containing the initial state, final state, direction of change, and order of change; generating bidirectional transfer signatures based on the forward and backward adjacent segments of the current operating condition segment; screening candidate segments from the healthy sample library whose operating condition trajectory index falls within the neighborhood of the current operating condition and whose bidirectional transfer signatures are consistent; performing positional correspondence between the candidate segments and the current operating condition segment along the operating condition coordinates, and determining the candidate weights based on the adjacentity of the operating conditions, the consistency of the transfer signatures, and the positional correspondence residuals; eliminating candidate segments whose forward and backward correspondence results are inconsistent, and determining the remaining candidate segments and their corresponding candidate weights as the set of healthy reference segments.

[0030] The bidirectional transfer signature is formed by sequentially encoding the working condition category before entering the current segment, the direction of entry, and the working condition category and direction of departure after leaving the current segment. When a segment is located at the beginning or end of a continuous running record, an empty adjacency marker occupies the corresponding position without using other segments to replace it. Candidate retrieval first limits the neighborhood according to the working condition trajectory index, and then compares the encoded positions in the bidirectional transfer signature. Only segments whose starting and ending states are within the neighborhood, whose main change directions are the same, and whose change order is not reversed are included in the candidate set. The position correspondence uses the cumulative working condition coordinates instead of the original time coordinates, mapping each response position in the candidate segment to the position in the current segment with the same or adjacent working condition coordinates, thereby allowing healthy segments with different running times but the same working condition evolution path to participate in reconstruction.

[0031] The weights of candidate segments are calculated using an exponential normalization method, as follows: For each candidate segment that passes the basic screening, calculate the weighted sum of its three difference indicators: the normalized adjacent distance between the current working condition trajectory and the candidate segment trajectory, the transfer difference of the bidirectional transfer signature, and the response continuous residual after the working condition coordinates are matched; each indicator is multiplied by the corresponding non-negative constraint coefficient and then summed.

[0032] Take the negative value of the above weighted sum and then calculate its natural index value.

[0033] The normalized candidate weight of a candidate segment is obtained by dividing the index value of a single candidate segment by the sum of the index values ​​of all candidate segments that passed the basic screening.

[0034] For example, the three difference indices of candidate fragment A are 0.1, 0, and 0.2, respectively, and the three difference indices of candidate fragment B are 0.2, 0.3, and 0.1, respectively. When all three constraint coefficients are 1, the index calculation result of candidate fragment A is greater than that of candidate fragment B. After normalization calculation, candidate fragment A gets a larger weight.

[0035] This design ensures that segments with similar operating conditions and consistent transfer paths account for a major proportion in the calculation of counterfactual health responses.

[0036] Please refer to Figure 4 , Figure 4 The horizontal axis represents the candidate reference segments after basic screening and bidirectional position correspondence verification, categorized by Chinese terms such as Reference A, Reference B, etc. The vertical axis represents the weight of the candidate reference segment in the counterfactual health response reconstruction of the current working condition segment. The height of the bar reflects the relative contribution of the candidate segment to the reconstruction result; the higher the weight, the better the reference segment matches the current segment in terms of adjacent working condition trajectories, consistent bidirectional transfer signatures, and position correspondence residuals.

[0037] from Figure 4As can be seen, the candidate weights of both Reference A and Reference B are at a high level and are very close, indicating that the current operating condition segment is not independently dominated by a single healthy reference segment, but rather formed by multiple healthy segments with consistent evolution directions and stable positional correspondences. This distribution reflects the result after candidate weight normalization, unifying the differences in operating condition distances, signature differences, and response continuity residuals between different candidate segments into a comparable contribution scale.

[0038] The forward correspondence result is the correspondence sequence obtained by mapping the position from the starting state of the current working condition segment to the ending state along the direction of change. The backward correspondence result is the correspondence sequence obtained by tracing back from the ending state to the starting state in the reverse order. The edge computing nodes compare the candidate response positions pointed to by the same working condition coordinates in the two correspondence sequences. When the position order intersects, adjacent positions are repeatedly mapped, or the direction of change of the corresponding residuals is opposite, the corresponding candidate segment is marked as bidirectionally inconsistent and removed from the candidate set. The remaining segments form a healthy reference segment set according to the candidate weights. When the number of remaining segments is insufficient to cover the current working condition coordinate range, the current segment is transferred to the pending confirmation state without calling segments with a greater working condition distance to make up for it, so as to maintain the traceability of the source of the counterfactual healthy response.

[0039] In a preferred embodiment, generating a counterfactual health response based on a set of health reference segments includes: mapping each health reference segment to a unified position sequence along the operating condition coordinates, and forming a set of health responses corresponding to each position according to candidate weights; sequentially removing each health reference segment from the health response set, reconstructing the removed health reference segment using the remaining health reference segments, and forming a reference reliability sequence based on the reconstruction deviation at each position; selecting a position interval with continuous and stable reliability from the reference reliability sequence, and determining the local response gradient between the operating condition coordinate increment and the health response increment based on the position interval; using the weighted response of the health response set at the starting position of the current operating condition segment as the recursive starting point, generating a health response position by position according to the adjacent operating condition coordinate increments and the corresponding local response gradients, and correcting the generated health response according to the response continuity relationship at the endpoints of adjacent operating condition segments, thereby obtaining the counterfactual health response and the counterfactual credibility corresponding to each position.

[0040] The reference reliability is obtained through the removal and reconstruction calculation.

[0041] The reference reliability is calculated using the shift-out reconstruction method, and the specific process is as follows: For each healthy reference segment, calculate the absolute difference between the response value mapped to the current position and the response value reconstructed from the remaining healthy reference segments after the segment is removed. This difference is the absolute reconstruction bias of the corresponding segment.

[0042] The median of the absolute reconstruction deviations of all candidate segments is taken as the representative value of the overall reconstruction deviation at that position.

[0043] Divide the value 1 by the result of "1 plus the median above", and the final value is the reference reliability of that position.

[0044] For example, the absolute reconstruction biases of three healthy reference segments at a certain location are 0.05, 0.08, and 0.4, respectively. Taking the median as 0.08, the calculated reference reliability is approximately 0.926. Under this calculation method, the outlier values ​​that deviate from the population do not directly dominate the reliability result.

[0045] When the reference reliability of adjacent locations remains continuous and changes in the same direction, the corresponding interval is determined as a reliable continuous and stable interval.

[0046] Please refer to Figure 5 , Figure 5 The horizontal axis represents the unified position stage after the current working condition segment is mapped, arranged sequentially from the start to the end; the vertical axis represents the reference reliability ratio, and the closer the value is to 1, the more stable the health reference source at that position is. Figure 5 The solid line represents the reference reliability sequence, and the dashed line represents the counterfactual reliability sequence. The two curves together reflect the impact of the removal and reconstruction bias, candidate coverage, bidirectional correspondence consistency, and endpoint continuous correction quality on the reliability of the current health benchmark.

[0047] from Figure 5 As can be seen, the reference reliability curve generally remains in a high range, but there are local dips in some stages, indicating that the reconstruction consistency between healthy reference segments at certain uniform locations has decreased. The counterfactual credibility curve is relatively smooth and maintains a similar direction of change as the reference reliability curve, indicating that although there are reference differences at local locations, candidate coverage and endpoint correction can still maintain a relatively stable counterfactual healthy response basis.

[0048] The local response gradient is calculated by dividing the health response increment of adjacent locations by the operating condition coordinate increment within the continuous and stable reliability interval. For cases where there are multiple reference segment gradients at the same location, the gradients are normalized and fused according to the product of the candidate weights and the reference reliability to obtain the final local response gradient.

[0049] The counterfactual health response is calculated using a recursive method of step-by-step accumulation: the weighted health response at the starting position of the current working condition segment is used as the initial reference value. Starting from the starting position, the calculation results of "local response gradient × corresponding working condition coordinate increment" between each adjacent position are accumulated sequentially to finally obtain the counterfactual health response at the target position.

[0050] For example, if the weighted health response at the starting position is 1.2, and the local response gradients at two consecutive positions are 0.4 and 0.3 respectively, and the corresponding working condition coordinate increments are 0.1 and 0.2 respectively, then the counterfactual health response at the third position is calculated to be 1.3.

[0051] The recursive result also needs to be continuously corrected at the endpoints with the health response at the end of the previous working condition segment, so that the generated response sequence maintains the local change relationship of the current working condition trajectory and does not produce unprovoked jumps at the segment connection.

[0052] Please refer to Figure 6 , Figure 6 The diagram uses a three-plane approach to display the axial, radial, and vertical responses. The horizontal axis of each sub-plot represents a unified positional stage, while the vertical axis represents the response amplitude in the corresponding direction. Within each sub-plot, solid lines represent the actual response after time alignment, dashed lines represent the healthy response obtained by recursion from the healthy reference segment and continuous endpoint correction, and dotted lines represent the fault residual retained after subtracting the healthy response from the actual response. The legend uniformly lists the three types of curves to avoid multiple-channel curves obscuring each other on the same coordinate system.

[0053] from Figure 6 As can be seen, the actual response in the axial response subplot exhibits strong periodic fluctuations, while the healthy response remains relatively flat. The retained residuals show continuous fluctuations in several stages, indicating that some response variations in this direction cannot be fully explained by the evolution under normal operating conditions. In the radial response subplot, the actual response shows more obvious stage-based changes, and the retained residuals change synchronously with the fluctuations in the actual response, suggesting that there may be local components in the radial direction related to the current operating trajectory but not fully explained by the healthy reference. In the vertical response subplot, the actual response amplitude is relatively small, but local peaks and valleys are dense. The residual curves can retain these fine-grained variations, providing input for subsequent extraction of impact sparsity evidence and periodic recurrence evidence.

[0054] Counterfactual credibility is jointly encoded by candidate coverage, reference reliability, bidirectional correspondence consistency, and endpoint correction. Candidate coverage represents the proportion of healthy reference segments with valid mapping values ​​at the current position. Bidirectional correspondence consistency represents the degree of overlap between the forward and backward mapping positions. Endpoint correction represents the adjustment magnitude made by the recursive response to satisfy the continuity relationship between adjacent segments. After the edge computing nodes complete the same-direction normalization of the four data items, they form a position-by-position credibility sequence. Positions with lower credibility still retain the counterfactual healthy response but are marked with low credibility. When establishing cross-scale evidence in the future, additional verification conditions are applied to the abnormal connections of low credibility positions, and the abnormal judgment is not directly replaced by low credibility.

[0055] The fault retention residual is obtained by subtracting the counterfactual health response at the corresponding position from the time-aligned bearing operating data. After the multi-channel operating data are used to form channel residuals, they are retained as independent sequences according to the channel time consistency and are not pre-merged into a single amplitude. For vibration data, the original direction or the corresponding relationship of the acquisition channel is retained. For temperature data and drive current data, the slowly changing components and local changing components are retained. The edge computing node synchronously records the health reference segment referenced by each residual position, the candidate weight, the counterfactual credibility, and the endpoint correction source, so that when evidence conflicts occur later, the reference data that generated the residual can be located, rather than just making sourceless modifications to the final state label.

[0056] In a preferred embodiment, constructing a multi-scale time window tree and establishing cross-scale evidence consistency relationships includes: using a time window covering the current operating condition segment as the root time window, generating candidate split points based on the local energy accumulation location, symbol flip density, and periodic recurrence location of the fault-preserving residuals; establishing sub-time windows for the residual sequences on both sides of the candidate split points, and restoring them to the parent time window when the dominant period, phase continuity relationship, and frequency band migration direction of adjacent sub-time windows meet the merging conditions; forming impact sparsity evidence, periodic recurrence evidence, frequency band migration evidence, and envelope phase evidence within each time window; using the time window and corresponding anomaly evidence as evidence nodes, connecting the evidence nodes according to time overlap, integer period correspondence, and parent-child scale inclusion relationship to form a cross-scale evidence consistency relationship.

[0057] The multi-scale time window tree is established by combining top-down splitting and bottom-up merging. The root time window covers the complete operating condition segment. The location of local energy accumulation is obtained by comparing the cumulative amount of local absolute value of residuals with the cumulative amount of residuals in adjacent regions. The sign flip density is obtained by statistically limiting the number of positive and negative changes of residuals in the neighborhood. The periodic recurrence location is determined by the local peak interval sequence of the residual envelope. Candidate split points must be located in the neighborhood pointed to by at least two types of location indicators. The sub-time windows formed after splitting inherit the operating condition coordinate range and counterfactual confidence index of the parent time window. If the dominant period difference of adjacent sub-time windows is within the same period tolerance range, the end phase and the start phase can be continuously spliced, and the frequency band energy center changes in the same direction, then the corresponding split is canceled and the parent time window is restored to avoid random spikes generating fragmented windows lacking continuity.

[0058] Shock sparsity evidence is represented by the occupancy relationship of a small number of high-amplitude residuals within the window in the total residual energy; periodic recurrence evidence is represented by the correspondence between adjacent shock intervals and the dominant period and its integer multiples; frequency band migration evidence is represented by the trajectory of the energy center between sub-time windows moving continuously to adjacent frequency bands over time; envelope phase evidence is represented by the phase dispersion of the envelope peaks of each period in a unified period coordinate system. Each type of evidence preserves the evidence source window, the corresponding residual position, the associated counterfactual credibility, and the traceable reference fragment. Evidence nodes are not named with a single feature value, but are jointly encoded by time window identifier, scale level, evidence category, and location range, so that the performance of the same anomaly at different scales can establish a clear parent-child correspondence.

[0059] In a preferred embodiment, the establishment of cross-scale evidence consistency further includes: recording the time interval, scale level, frequency band level, dominant period, and phase trajectory for each evidence node; for evidence nodes with parent-child scale inclusion relationships, setting a first connection weight based on the retention ratio of anomalous evidence in the parent and child time windows; for evidence nodes with integer period correspondence relationships, setting a second connection weight based on period deviation and phase trajectory continuity; for evidence nodes located in adjacent frequency band levels, setting a third connection weight based on the consistency of frequency band migration direction; transmitting evidence markers along parent-child scale connections, integer period connections, and frequency band adjacency connections according to the connection weights from high to low, recording the number of independent paths reaching the same evidence node and adjacent period recurrence markers, deleting evidence nodes that are only reached by a single path and have no adjacent period recurrence markers, and updating the cross-scale evidence consistency relationship.

[0060] The overall connection weight between evidence nodes is obtained by weighted summation of three connection metrics: The first item is the percentage of evidence retained that is abnormal on the father-son scale; The second term is the periodic connectivity, which is calculated from the integer periodic deviation and the phase trajectory continuity. The third term is the consistency quantity in the direction of frequency band migration.

[0061] Each of the three indicators is multiplied by its corresponding non-negative fusion coefficient, and the sum of the three fusion coefficients is 1. The final calculation result is the comprehensive connection weight between the two evidence nodes.

[0062] For example, when the parent-child scale retention ratio is 0.8, the periodic connection amount is 0.7, the frequency band consistency amount is 0.6, and the three fusion coefficients are set to 0.4, 0.35, and 0.25 respectively, the calculated comprehensive connection weight is 0.715.

[0063] The retention and deletion of evidence nodes follow these rules: An evidence node is deleted only if it is reached by a single path and has no adjacent periodic recurrence markers; evidence nodes that do not meet both of the above conditions are retained.

[0064] Please refer to Figure 7 , Figure 7 The horizontal axis represents the cross-scale evidence connection number, categorized by Chinese terms such as Connection A, Connection B, etc.; the vertical axis represents the evidence connection weight or component quantity. Figure 7 The solid line represents the overall connectivity weight, while the dashed, dotted, and light-colored lines represent the parent-child scale retention, periodic connectivity, and frequency band consistency, respectively. Each curve illustrates the degree of consistency among different evidence nodes at the time window level, integer period correspondence, and adjacent frequency band migration directions.

[0065] from Figure 7 As can be seen, the overall connectivity weight remains at a high level at most connectivity positions, indicating that these evidence nodes do not appear in isolation, but can form stable connections through parent-child scale inclusion relationships, periodic recurrence relationships, or frequency band adjacency relationships. The parent-child scale retention decreases at some connectivity positions, indicating that the retention of corresponding anomalies weakens after the window shrinks; the periodic connectivity and frequency band consistency show different patterns of change, indicating that evidence connectivity is not determined by a single indicator, but is supported by multiple types of relationships.

[0066] Independent paths refer to transmission paths originating from different root-level evidence or different evidence categories, with intermediate nodes not completely overlapping. When evidence is marked for transmission, it is processed from high to low according to the comprehensive connection weight. In parent-child scale connections, it is confirmed whether the anomaly is still retained after the window is narrowed. In integer period connections, it is confirmed whether the anomaly recurs across adjacent periods. In frequency band adjacency connections, it is confirmed whether the anomaly has continuous migration rather than instantaneous diffusion. After recording, a path source set is established for each evidence node. Single path nodes are only retained when there is a recurrence mark in adjacent periods. For multi-path nodes, it is also necessary to check whether the time positions pointed to by each path overlap. When the paths are completely separated, they are marked as conflict and not directly merged. This forms a cross-scale evidence consistency relationship that can be traced back to the window and residual positions.

[0067] In a preferred embodiment, correcting the time window boundaries and healthy reference segments includes: identifying conflicting anomalous evidence in cross-scale evidence consistency relationships as conflicting evidence nodes; classifying evidence conflicts into boundary conflicts, reference conflicts, and isolated conflicts based on the position of the conflicting evidence nodes in the multi-scale time window tree; for boundary conflicts, re-determining the split point in the adjacent region of the conflicting evidence node according to the local extrema of the residual continuity, and reconstructing the corresponding parent-child time windows; for reference conflicts, extracting the local reconstruction components of each healthy reference segment at the conflict frequency band and conflict time position, determining the median change direction of each local reconstruction component as the reference direction, removing healthy reference segments whose local reconstruction components are opposite to the reference direction, and regenerating the counterfactual healthy response; for isolated conflicts, marking the corresponding operating condition segment as a segment to be confirmed while keeping the healthy sample library unchanged; and generating a freeze identifier and associated evidence node set for operating condition segments that meet the consistency conditions.

[0068] Boundary conflict manifests as the presence of continuous abnormal evidence within the parent time window while neither of the two child time windows retains corresponding evidence, or as adjacent child time windows pointing to different halves of the same abnormality. In this case, the edge computing node constructs a set of candidate boundaries on both sides of the original split point and selects the optimal boundary from them to redivide the time window.

[0069] The optimal boundary is selected by minimizing the comprehensive objective value: for each candidate boundary, the weighted sum of three indicators is calculated as the comprehensive objective value. The three indicators are: The amount of crossover of anomalous evidence on both sides after segmentation by the candidate boundary; The discontinuity of the dominant cycles on both sides; The distance between the boundary location and the location of the counterfactual credibility mutation.

[0070] Each of the three indicators is multiplied by its corresponding non-negative constraint coefficient, and the sum of the three constraint coefficients is 1. Finally, the boundary with the smallest comprehensive objective value is selected from all candidate boundaries as the redefined time window boundary.

[0071] For example, when the combined target values ​​of the three candidate boundaries are 0.42, 0.25, and 0.37 respectively, the candidate boundary with a target value of 0.25 is selected to reconstruct the parent-child time window. This design allows anomalous evidence to fall into the same sub-time window as completely as possible, while matching the reconstructed boundary with the location of confidence changes.

[0072] Please refer to Figure 8 , Figure 8 The horizontal axis represents the candidate boundaries generated within the regions adjacent to the original split point, denoted as boundary A, boundary B, and boundary C, respectively; the vertical axis represents the target quantity of the candidate boundary. The lower the height of the bar, the more conducive the candidate boundary is to reducing the intersection of anomalous evidence on both sides, reducing the degree of discontinuity in the dominant cycle, and maintaining the correspondence between the boundary position and the position of counterfactual credibility abrupt change.

[0073] from Figure 8 As can be seen, boundary A has a higher target value, indicating that a large crossover of anomalous evidence may still occur after reconstructing the time window at that position; boundary B has the lowest target value, indicating that this candidate boundary can better place continuous anomalous evidence into the same sub-time window and maintain a relatively continuous periodic relationship on both sides; boundary C is at an intermediate level, which is better than boundary A, but its correspondence with the confidence mutation position or periodic continuity is still not as good as boundary B.

[0074] Reference conflict manifests as opposite residual change directions under different combinations of healthy reference segments within the same time window. Edge computing nodes locate the healthy reference segments referenced by the conflict frequency band and conflict time location from the residual generation record, extract the local reconstruction components of each reference segment after mapping by the working condition coordinates, calculate the local change direction according to the position order, and take the median direction as the reference direction. When a healthy reference segment is opposite to the reference direction in multiple consecutive corresponding positions, it is removed from the current reference set. The remaining candidate weights are renormalized and the counterfactual health response, fault retention residuals, and affected evidence nodes are recalculated. The removal operation only applies to the current reconstruction version. Whether the health sample library is modified synchronously depends on the verification results of subsequent versions.

[0075] Isolated conflicts manifest as anomalous evidence existing only at a single scale, a single period, or a low-confidence location, and cannot form a stable connection by reconstructing the time window or re-screening healthy reference segments. The corresponding operating condition segment is marked as a segment to be confirmed, and the original operating data, operating condition trajectory index, counterfactual generation record, and conflict evidence node are retained. After the arrival of subsequent adjacent operating condition segments, the periodic recurrence and frequency band migration relationship are re-examined. During the period of waiting for confirmation, the segment is neither written into the healthy sample library nor directly merged into the fault prototype library. When subsequent data still cannot form a consistent relationship, it is saved as an unresolved record and the current version of the healthy sample library remains unchanged.

[0076] The bearing operating status is determined based on the time range of consistent abnormal evidence, the number of independent paths, the recurrence of adjacent cycles, and the counterfactual credibility. When an evidence node can form at least two independent paths by connecting along parent and child scales and integer cycles, and the corresponding time position does not completely fall into the low credibility interval, the relevant operating condition segment is determined to be an abnormal state. When the evidence only forms a single-scale slow change trajectory under stable operating conditions without impact or periodic recurrence, the operating condition segment is kept in the unconfirmed state. When a cross-scale evidence node is deleted after redefining the boundary and regenerating the residual, the operating condition segment is determined to be a normal candidate state. However, the normal candidate state still needs to be verified by writing healthy samples before it can enter the healthy sample library.

[0077] In a preferred embodiment, the method further includes: recording a subset of healthy reference segments participating in reconstruction at each time point; forming a confidence gradient based on the time change of counterfactual confidence; intersecting the abrupt change position of the confidence gradient with the local abrupt change position of the fault-preserving residual; and using the obtained position as the initial candidate split point of the multi-scale time window tree; for two anomalous pieces of evidence connected in a cross-scale evidence consistency relationship, calculating the lower counterfactual confidence at their corresponding positions and the overlap ratio of the two healthy reference segment subsets; deleting the connection between the two anomalous pieces of evidence when both are lower than the adjacent evidence connection value; when evidence conflicts are concentrated at continuous positions where the counterfactual confidence is lower than the adjacent interval, feeding back the conflict position to the reference reliability sequence, and performing a removal reconstruction on the healthy reference segments involving the conflict position; when anomalous evidence recurs across scales at positions where the counterfactual confidence is continuous and stable, setting a locking flag for the corresponding time window path, and redefining the boundary only for time windows without a locking flag.

[0078] The credibility gradient is formed by the absolute value of the difference in counterfactual credibility between adjacent positions. Local mutation positions are determined by the condition that the difference in fault-preserved residuals between adjacent positions exceeds the range of positional variation in the same working condition segment. The two types of mutation positions are only used as initial candidate split points when they overlap in the time neighborhood. This avoids directly splitting the window due to credibility changes caused by reference segment switching or residual changes caused by a single random shock. When verifying evidence connections, the lower credibility within the corresponding intervals of the two pieces of evidence is taken as the common credibility basis. Then, the proportion of the intersection of the healthy reference segment subsets referenced by the two intervals to the union is calculated. When the overlap ratio of the common credibility basis and the reference source is lower than the current comprehensive connection weight, it indicates that the connection is mainly formed by the residual differences under different reference sources. The corresponding connection is deleted and a reference conflict check is triggered.

[0079] The locking identifier preserves the root window identifier, sub-window boundaries at each level, evidence node identifiers, and counterfactual credibility stable intervals in the time window path. When a local conflict occurs later, only the windows outside the unlocked path or the locked path are re-splitting and merging. If the screening of new healthy reference fragments changes the counterfactual response referenced by the locked path, the local locking overlapping with the changed position is released first while the locking of the remaining paths is retained. After the directional reconstruction is completed, the cross-scale recurrence relationship is re-examined. This avoids the repeated reconstruction of the entire multi-scale time window tree caused by a local reference correction and maintains the evidence structure that has been jointly confirmed by multiple paths and adjacent periods.

[0080] In a preferred embodiment, the method further includes: encapsulating the frozen condition segment, its corresponding condition segment index, counterfactual health response, fault retention residual, multi-scale time window path, evidence node set, and conflict correction record into a frozen evidence package; generating a version identifier based on the order of data within the package; and subsequently updating to establish a new frozen evidence package associated with the previous version; for unfrozen candidate condition segments intended to be written into the health sample library, performing counterfactual reconstruction on the health condition segments before and after the candidate condition segments under reference configurations that include and do not include candidate condition segments; comparing the evidence nodes and time window boundaries corresponding to the two reference configurations; and only writing candidate condition segments that do not introduce new conflict evidence nodes and maintain the original evidence node connection relationships into the health sample library; and merging the frozen evidence package into the corresponding fault prototype or establishing a new fault prototype according to the time window path and evidence node topology correspondence between the frozen evidence package and the existing fault prototype.

[0081] The frozen evidence package is stored using an immutable data structure. The package first records the index and time range of the operating condition segment, then records the counterfactual health response and its reference relationship, followed by the fault retention residual, multi-scale time window tree, evidence node connection relationship, conflict classification and correction process. The version identifier is formed by the sequential encoding of each data item, content summary and the previous version identifier. When any reference segment, time window boundary or evidence connection changes, a new version is created without overwriting the original version. After being frozen, abnormal operating condition segments are only allowed to have correction records added through version association, and are not allowed to re-enter the health sample writing process as normal candidate segments.

[0082] The candidate operating condition fragment writing verification adopts a dual-reference configuration. The inclusion configuration adds the candidate operating condition fragment to the temporary candidate library, while the non-inclusion configuration keeps the current healthy sample library unchanged. The edge computing node selects healthy operating condition fragments adjacent to its operating condition trajectory before and after the candidate fragment. Counterfactual health responses and multi-scale evidence relationships are generated under the two configurations respectively. When the inclusion configuration generates new conflicting evidence nodes, changes the connection direction of the original abnormal evidence, or causes the original time window boundary to cross the locked evidence position, the candidate fragment writing is rejected. When the original evidence node set and connection relationship remain consistent under the two configurations and the inclusion configuration does not generate new conflicts, the candidate fragment along with its verification record is written to the healthy sample library and a new sample library version is generated.

[0083] The fault prototype is composed of the common topological features of multiple frozen evidence packages. The common topological features include the hierarchical sequence of multi-scale time window paths, the combination of abnormal evidence categories, the integer period connection method, the frequency band migration direction, and the conflict correction type. The edge computing node first performs working condition coordinate alignment between the frozen evidence package and the existing fault prototype, and then compares the parent-child structure of the time window path and the connection order between the evidence nodes. When the two can establish corresponding connections on the main path and the difference nodes are only located in the leaf-level window, the frozen evidence package is merged into the corresponding fault prototype and its individual difference nodes are saved. When there is no corresponding prototype, a new fault prototype is established. The prototype establishment process does not change the original residuals and state records in the frozen evidence package.

[0084] In a preferred embodiment, the method further includes: when a health reference segment is re-screened due to evidence conflict, establishing a candidate version based on the current health sample library, recording the sample addition / deletion relationship, counterfactual health response change relationship, and the reference relationship of each counterfactual health response to the health reference segment between the candidate version and the original version; determining the affected frozen evidence package based on the sample addition / deletion relationship and the reference relationship, extracting the operating condition segment index and multi-scale time window path from the frozen evidence package, and performing targeted reconstruction on the corresponding operating condition segment; after targeted reconstruction, if the abnormal evidence nodes in the frozen evidence package are not deleted, the determined bearing operating state has not changed, and the time window boundaries of adjacent operating condition segments are continuous, replacing the current health sample library version with the candidate version; otherwise, revoking the candidate version, and restoring the health reference segment, fault retention residual, and evidence node connection relationship based on the counterfactual health response change relationship.

[0085] The replacement decision factor for candidate versions in the sample library is obtained by weighted summation of three indicators: The first item is the retention rate of the frozen abnormal evidence nodes after targeted reconstruction; The second item is the consistency of the bearing's operating state before and after the reconstruction; The third item is the continuous quantity at the boundary of the time window of adjacent operating condition segments.

[0086] Each of the three indicators is multiplied by its corresponding non-negative coefficient, and the sum of the three coefficients is 1. The final result is the replacement decision value for the candidate version.

[0087] Candidate versions must also meet a hard constraint: no frozen abnormal evidence nodes may be deleted.

[0088] For example, if the values ​​of the three indicators are 1, 1 and 0.8 respectively, and the three coefficients are set to 0.4, 0.4 and 0.2 respectively, the calculated replacement determination quantity is 0.96; if no frozen node is deleted at this time, the candidate version is allowed to enter the version replacement process.

[0089] If the bearing operating status consistency quantity is 0, or if a frozen node is deleted, the candidate version must be revoked even if the replacement judgment quantity reaches the set threshold.

[0090] Please refer to Figure 9 , Figure 9 The horizontal axis represents the components in the candidate version replacement judgment of the sample library, including the frozen node retention rate, state consistency, boundary continuity, and final judgment item; the vertical axis represents the version judgment quantity. The height of each bar is used to show whether the candidate version retains the original abnormal evidence nodes, maintains the bearing operating state, and whether the time window boundaries of adjacent working condition segments are continuous after directional reconstruction.

[0091] from Figure 9 As can be seen, the retention rate of frozen nodes and the consistency of state are both at a high level, indicating that the candidate version did not delete the frozen abnormal evidence nodes after being written, nor did it change the original bearing operating state. The boundary continuity is slightly lower than the first two items, indicating that although the candidate version can maintain the overall evidence structure, the connection between the boundaries of adjacent working condition segments still needs to be verified. The final judgment item remains at a high level, indicating that under the hard constraint of not deleting frozen nodes, the candidate version has the conditions to enter the replacement process.

[0092] Targeted reconstruction only processes counterfactual health responses that reference the added or deleted health reference fragments. Edge computing nodes locate the affected unified location sequence through reference relationships, then read the corresponding working condition trajectory index and locked time window path from the frozen evidence package, recalculate candidate weights, reference reliability, local response gradients, and fault retention residuals for the affected locations, retain the original version data for unaffected locations, recalculate the comprehensive connection weights for evidence connections that cross affected locations, and maintain the original records for evidence connections that are completely located in unaffected locations. When a candidate version is revoked, the reference fragment set and candidate weights are restored item by item according to the counterfactual health response change relationship, and the residuals, time window boundaries, and evidence node connection relationships associated with the original version identifier are restored.

[0093] In a specific operational example, when the bearing transitions from a low-speed stable section to a continuous acceleration section, the edge computing node establishes a working condition segment boundary at the end of the stable section based on the rate of change of the working condition and the direction of change. It generates a working condition trajectory index for the acceleration segment, which includes the initial speed state, the final speed state, the acceleration direction, and the load change sequence. It also retrieves healthy reference segments from the healthy sample library that are also accelerating and have consistent adjacent relationships. If a healthy segment covers the same speed range but comes from the deceleration process, its change direction is inconsistent with the bidirectional transfer signature and it is removed. The remaining segments are mapped along the cumulative working condition coordinates to generate a counterfactual healthy response that changes with the acceleration process. This allows the normal vibration amplitude changes caused by the increase in speed to enter the healthy response instead of being directly retained in the fault residual.

[0094] When there are local impacts with gradually shortening intervals in the acceleration segment, the fault retention residual forms impact sparse evidence in a sub-window with a small time scale, forms periodic recurrence evidence in a parent window covering multiple rotation cycles, and forms a continuous migration direction between adjacent frequency band levels. Evidence nodes form multiple independent paths through parent-child scale connections and integer period connections. Edge computing nodes compare the corresponding paths with the counterfactual confidence stable interval. When the path position does not fall into the low confidence interval, the abnormal state is determined and the operating condition segment is frozen. If the impact only occurs at the reference segment switching position and the two pieces of evidence connected across scales refer to different healthy reference segment subsets, the corresponding connection is deleted and reference conflict correction is initiated, without directly freezing the segment.

[0095] In another operational example, the bearing is under a long-term stable load. The counterfactual health response is mainly formed by weighting multiple health reference segments adjacent to the current stable operating condition. When the temperature data is slowly rising and the vibration residual does not form shock sparsity evidence or periodic recurrence evidence, the multi-scale time window tree only forms slowly changing leaf-level evidence nodes. The evidence nodes cannot be connected along integer periods to form independent paths. The current segment remains in a pending confirmation state. If the slow change continues in the same operating condition coordinate range after the arrival of subsequent adjacent segments and the corresponding residuals of other channels appear at the same time, the cross-scale evidence consistency relationship is re-established. If the subsequent data recovers and the original evidence node is deleted, the segment enters the normal candidate state and accepts dual-reference configuration write verification.

[0096] For cases where operational data is partially missing, edge computing nodes do not use a continuous filling method that crosses operating conditions. When the missing interval is located within a single operating condition segment and the operating conditions at both ends change in the same direction, only piecewise linear interpolation is performed on the operating condition data and interpolation markers are set. No fictitious response is generated at the location where bearing operating data is missing. Instead, null values ​​and coverage reduction markers are retained in the unified location sequence. Counterfactual health responses can still be generated from health reference segments, but the fault retention residuals at the corresponding locations are not included in the impact evidence calculation. Periodic connections that cross missing intervals must be established after at least one complete cycle is formed on each side of the missing interval. When the missing range crosses the operating condition boundary, the segments are directly split and processed separately.

[0097] In one system embodiment, a bearing operating status edge computing monitoring system includes an edge computing node with a memory and a processor. The memory stores a monitoring program, and the processor executes the monitoring program to: acquire and time-align bearing operating data and working condition data; divide the working condition into segments based on the rate of change of the working condition and the continuity of operation; retrieve healthy reference segments from a health sample library that are adjacent to the current working condition segment and have the same evolution direction, and reconstruct a counterfactual health response based on the positional relationship between the two working conditions; calculate the fault retention residual of the bearing operating data relative to the counterfactual health response; construct a splittable and mergeable multi-scale time window tree based on the fault retention residual, extract anomalous evidence at each scale, and establish cross-scale consistency relationships; when the relationship does not meet the consistency condition, correct the time window boundary and the healthy reference segment based on the position of evidence conflict and recalculate the fault retention residual; when the relationship meets the consistency condition, determine the bearing operating status based on consistent anomalous evidence, and freeze the current working condition segment and prohibit it from being written to the health sample library when the bearing operating status is anomalous.

[0098] In this system embodiment, the memory establishes an operating data area, a working condition segment area, a health sample library, a counterfactual response area, a residual area, a multi-scale time window area, an evidence relationship area, a frozen evidence package area, and a version relationship area. The processor reads the time stamps in the operating data area and the working condition segment area to complete the alignment, writes the segment division results and working condition trajectory index into the working condition segment area, writes the reference identifier and candidate weight of the candidate health reference segment into the counterfactual response area, writes the positional counterfactual response, counterfactual credibility, and reference relationship into the same record, writes the fault retention residual and the corresponding position index into the residual area, and then constructs a multi-scale time window tree and evidence node connection relationship based on the data in the residual area. All processed objects are associated through segment identifiers and version identifiers to avoid loss of correspondence between the status record and the referenced health sample.

[0099] The monitoring program includes time alignment instructions, condition segment construction instructions, reference retrieval instructions, counterfactual reconstruction instructions, residual processing instructions, multi-scale time window processing instructions, evidence relationship processing instructions, conflict correction instructions, state determination instructions, and version control instructions. The processor triggers the corresponding instructions according to the data arrival status of the condition segment. When the current segment is not yet closed, only the cumulative condition coordinates and temporary response cache are updated, and healthy sample writing is not performed. When the segment boundary is determined, the reference retrieval and counterfactual reconstruction instructions are called. When a conflict occurs in a cross-scale evidence relationship, the conflict correction instruction is called and the correction scope is limited to the position associated with the conflict node and the reference segment. When the state is determined to be abnormal, the version control instruction is called to generate a frozen evidence package and set a write prohibition flag.

[0100] Furthermore, after each state determination, the edge computing node saves the program version, health sample library version, operating condition segment index, counterfactual response record, time window tree structure, and evidence node relationship referenced in this processing. When a subsequent candidate version triggers targeted reconstruction, the processor reads the corresponding record based on the reference relationship and only recalculates the affected position. The recalculation result is saved alongside the original record through the new version identifier. When the candidate version meets the conditions of freezing abnormal evidence preservation, state preservation, and adjacent boundary continuity, the current version pointer is updated. When the conditions are not met, the original version pointer is restored and the temporary reference relationship of the candidate version is deleted. The original running data, frozen evidence package, and fault prototype are not overwritten due to the rollback operation.

Claims

1. A method for edge calculation monitoring of bearing operating status, characterized in that, include: The bearing operation data and corresponding working condition data are acquired and time-aligned by the edge computing nodes, and the working condition segments are divided according to the rate of change and the continuity of operation of the working condition data. Retrieve healthy reference segments from the health sample library that are adjacent to the current operating condition segment and have the same evolution direction, and reconstruct the counterfactual health response based on the correspondence between the two operating condition positions. Calculate the fault-preserved residual between the bearing operating data and the counterfactual health response; A splittable and mergeable multi-scale time window tree is constructed around the residuals of the aforementioned faults to extract anomaly evidence at each scale and establish cross-scale evidence consistency relationships. When the cross-scale evidence consistency relationship does not meet the consistency condition, the time window boundary and health reference segment are corrected according to the evidence conflict location and the fault retention residual is regenerated. When the cross-scale evidence consistency relationship meets the consistency condition, the bearing operating state is determined based on the consistent abnormal evidence, and when the bearing operating state is abnormal, the current operating condition segment is frozen and its writing to the health sample library is prohibited.

2. The bearing operating status edge calculation monitoring method according to claim 1, characterized in that, The process of dividing the working condition segments and retrieving health reference segments includes: reconstructing the working condition coordinates based on the cumulative changes of the working condition variables, and representing each working condition segment as a working condition trajectory index that includes the starting state, ending state, direction of change, and order of change. Generate a bidirectional transfer signature based on the forward and backward adjacent segments of the current operating condition segment; In the health sample library, candidate segments whose work condition trajectory index falls into the current work condition neighborhood and whose bidirectional transfer signatures are consistent are selected. The candidate segments are matched with the current working condition segments along the working condition coordinates, and the candidate weights are determined based on the working condition adjacency, the consistency of the transfer signature, and the position correspondence residual. Candidate segments whose forward and backward correspondence results are inconsistent are removed, and the remaining candidate segments and their corresponding candidate weights are determined as the set of healthy reference segments.

3. The bearing operating status edge calculation monitoring method according to claim 1, characterized in that, Constructing the multi-scale time window tree and establishing the cross-scale evidence consistency relationship includes: using the time window covering the current working condition segment as the root time window, and generating candidate split points based on the local energy accumulation location, symbol flip density, and periodic recurrence location of the fault-preserving residual; Sub-time windows are established for the residual sequences on both sides of the candidate split point, and the parent time window is restored when the dominant period, phase continuity relationship and frequency band migration direction of the adjacent sub-time windows meet the merging conditions. Within each time window, evidence of shock sparsity, periodic recurrence, frequency band shift, and envelope phase are formed. Using time windows and corresponding abnormal evidence as evidence nodes, the evidence nodes are connected according to time overlap, integer period correspondence and parent-child scale inclusion relationship to form the cross-scale evidence consistency relationship.

4. The bearing operating status edge calculation monitoring method according to claim 1, characterized in that, Correcting the time window boundaries and healthy reference segments includes: identifying conflicting anomalous evidence in the cross-scale evidence consistency relationship as conflicting evidence nodes, and classifying evidence conflicts into boundary conflicts, reference conflicts, and isolated conflicts based on the position of the conflicting evidence nodes in the multi-scale time window tree. For the boundary conflict, the split point is re-determined in the adjacent region of the conflict evidence node according to the local extremum of the residual continuity, and the corresponding parent-child time window is reconstructed. For the reference conflict, extract the local reconstruction components of each healthy reference segment at the conflict frequency band and conflict time position, determine the median change direction of each local reconstruction component as the reference direction, remove the healthy reference segment whose local reconstruction component is opposite to the reference direction, and regenerate the counterfactual healthy response. For the isolated conflict, the corresponding working condition segment is marked as a segment to be confirmed while the healthy sample library remains unchanged; For operating condition segments that meet the consistency conditions, a freeze marker and a set of associated evidence nodes are generated.

5. The bearing operating status edge calculation monitoring method according to claim 2, characterized in that, Generating the counterfactual health response based on the set of health reference segments includes: mapping each health reference segment to a unified position sequence along the working condition coordinates, and forming a set of health responses corresponding to the position according to the candidate weights; Each health reference segment is removed from the health response set in sequence, and the removed health reference segment is reconstructed using the remaining health reference segments. A reference reliability sequence is formed based on the reconstruction deviation at each position. Select a location interval with continuous and stable reliability from the reference reliability sequence, and determine the local response gradient between the operating condition coordinate increment and the health response increment based on the location interval; Using the weighted response of the health response set at the beginning position of the current working condition segment as the starting point of the recursion, a health response is generated position by position according to the coordinate increment of adjacent working conditions and the corresponding local response gradient. The generated health response is then corrected according to the response continuity relationship at the endpoints of adjacent working condition segments to obtain the counterfactual health response and the counterfactual credibility corresponding to each position.

6. The bearing operating status edge calculation monitoring method according to claim 3, characterized in that, The establishment of cross-scale evidence consistency also includes: recording the time interval, scale level, frequency band level, dominant period, and phase trajectory for each evidence node; For evidence nodes with the parent-child scale inclusion relationship, a first connection weight is set according to the retention ratio of the abnormal evidence in the parent time window and the child time window; For evidence nodes with integer period correspondence, a second connection weight is set based on period deviation and phase trajectory continuity. For evidence nodes located in adjacent frequency bands, a third connection weight is set based on the consistency of the frequency band migration direction; According to the connection weights from high to low, the evidence markers are transmitted along the parent-child scale connection, integer period connection and frequency band adjacency connection respectively. The number of independent paths to the same evidence node and the adjacent period recurrence markers are recorded. Evidence nodes that are only reached by a single path and have no adjacent period recurrence markers are deleted, and the cross-scale evidence consistency relationship is updated.

7. The bearing operating status edge calculation monitoring method according to claim 4, characterized in that, Also includes: The working condition segment with the frozen identifier, along with the corresponding working condition segment index, counterfactual health response, fault retention residual, multi-scale time window path, evidence node set, and conflict correction record, are encapsulated into a frozen evidence package. A version identifier is generated based on the arrangement order of the data within the package. Subsequent updates establish a new frozen evidence package associated with the previous version. For candidate operating condition segments that are not frozen and are intended to be written into the health sample library, counterfactual reconstruction is performed on the health operating condition segments before and after the candidate operating condition segments under reference configurations that include the candidate operating condition segments and do not include the candidate operating condition segments. The evidence nodes and time window boundaries corresponding to the two reference configurations are compared, and only candidate operating condition segments that do not introduce new conflicting evidence nodes and maintain the original evidence node connection relationship are written into the health sample library. According to the time window path and evidence node topology correspondence between the frozen evidence package and the existing fault prototype, the frozen evidence package is merged into the corresponding fault prototype or a new fault prototype is established.

8. The bearing operating status edge calculation monitoring method according to claim 5, characterized in that, Also includes: Record the healthy reference fragment subsets participating in the reconstruction at each time point, form a confidence gradient based on the time change of the counterfactual confidence, find the intersection of the abrupt change position of the confidence gradient with the local abrupt change position of the fault-preserving residual, and use the obtained position as the initial candidate split point of the multi-scale time window tree. For two connected anomalous pieces of evidence in the cross-scale evidence consistency relationship, calculate the low counterfactual credibility of their corresponding positions and the overlap ratio of the two healthy reference fragment subsets. When both are lower than the adjacent evidence connection value, delete the connection between the two anomalous pieces of evidence. When evidence conflicts are concentrated at consecutive positions where the counterfactual credibility is lower than that of adjacent intervals, the conflict positions are fed back to the reference reliability sequence, and the healthy reference segments involving the conflict positions are removed and reconstructed. When the anomalous evidence recurs across scales at locations where the counterfactual credibility is continuous and stable, a locking flag is set for the corresponding time window path, and the boundaries are redefined only for time windows where no locking flag is set.

9. The bearing operating status edge calculation monitoring method according to claim 1, characterized in that, Also includes: When the health reference segment is re-screened due to evidence conflict, a candidate version is established based on the current health sample library. The sample addition and deletion relationship, the counterfactual health response change relationship, and the reference relationship of each counterfactual health response to the health reference segment are recorded between the candidate version and the original version. Based on the sample addition / deletion relationship and the reference relationship, the affected frozen evidence package is determined, the working condition segment index and multi-scale time window path are extracted from the frozen evidence package, and targeted reconstruction is performed on the corresponding working condition segment; After targeted reconstruction, if the abnormal evidence nodes in the frozen evidence package are not deleted, the determined bearing operating status has not changed, and the time window boundaries of adjacent operating condition segments are continuous, the candidate version replaces the current healthy sample library version. Otherwise, the candidate version is revoked, and the health reference fragment, fault retention residual, and evidence node connection relationship are restored according to the counterfactual health response change relationship.

10. A bearing operating status edge computing monitoring system, characterized in that, It includes an edge computing node with a memory and a processor, wherein the memory stores a monitoring program and the processor executes the monitoring program to: acquire and time-align bearing operating data and working condition data, and divide the working condition into segments according to the rate of change of working conditions and the continuity of operation; Retrieve healthy reference segments from the health sample library that are adjacent to the current operating condition segment and have the same evolution direction, and reconstruct the counterfactual health response based on the weighted relationship between the two operating conditions; Calculate the fault-preserved residuals of the bearing operating data relative to the counterfactual health response; Based on the fault-preserving residuals, a splittable and mergeable multi-scale time window tree is constructed to extract anomaly evidence at each scale and establish cross-scale consistency relationships. When the relationship does not meet the consistency condition, the time window boundary and health reference segment are corrected according to the evidence conflict location and the fault retention residual is recalculated. When the relationship meets the consistency condition, the bearing operating status is determined based on the consistent abnormal evidence. When the bearing operating status is abnormal, the current operating condition segment is frozen and it is prohibited from being written into the health sample library.

Images