Elevator maintenance record cross-checking method and system based on cloud database

Abstract

This invention discloses a method and system for cross-verification of elevator maintenance records based on a cloud database. The method includes: constructing a maintenance project and expected parameter change model, storing the parameter change range before and after each maintenance project and the transient feature template of the process; obtaining an evidence chain to be verified, including evidence of the state before and after maintenance and the time series of the operation process; performing cross-verification of the operation trace dimension, extracting parameter snapshots for steady-state change verification, and simultaneously matching process data with transient feature templates to generate process verification results, comprehensively generating a trace dimension score; determining the anomaly type based on the score and outputting it. This invention integrates steady-state changes and transient process features for dual physical trace verification, solving the problem that existing technologies cannot identify forged process traces and improving the depth of verification of the authenticity of maintenance operations.

Description

Technical Field

[0001] This invention relates to the field of elevator maintenance and supervision technology, specifically to a method and system for cross-verification of elevator maintenance records based on a cloud database. Background Technology

[0002] With the continuous growth of the number of elevators in cities, the supervision of elevator maintenance quality has become a crucial aspect of ensuring public safety. In reality, the practice of maintenance personnel merely signing in without actually performing maintenance tasks—a practice known as "present but not operating"—remains rampant and is becoming increasingly sophisticated. Existing digital monitoring technologies suffer from the following shortcomings: The first type of solution verifies whether maintenance personnel arrive at the designated location and sign in at the specified time through IoT terminals, such as using RFID card swiping, facial recognition, or location information comparison to record their arrival status. This type of solution focuses on the authenticity of personnel identity and arrival location, but it cannot verify whether personnel actually performed each specific maintenance task after arriving on site.

[0003] The second type of solution collects and analyzes elevator operating status parameters through sensors to determine whether the elevator needs maintenance or fault diagnosis. While it can assess the health level of the equipment, it cannot reversely confirm whether the various operations in a completed maintenance task were actually performed.

[0004] In recent years, a verification approach based on parameter change comparison has emerged. This involves extracting snapshots of equipment operating parameters at two points in time, before and after maintenance, calculating the actual changes, and comparing them with the expected range of change to infer whether the operation occurred. However, this verification method, which relies solely on static snapshots, ignores the dynamic transient characteristics that equipment operating parameters inevitably exhibit during maintenance operations. Taking door lock gap adjustment as an example, actual operation requires repeated opening and closing of the door for testing and adjustment. This results in multiple jumps in the door operator's opening and closing duration parameter, gradually converging to the target range. If maintenance personnel simply create a momentary deviation within the expected range by manually pressing against the door operator once, this convergence process cannot be reproduced. Existing static comparison methods cannot capture this process "fingerprint," making it easy to bypass with simple methods.

[0005] In reality, genuine maintenance operations inevitably leave identifiable dynamic characteristics in sensor data, such as instantaneous fluctuations, gradual stabilization, or convergence through repeated tests. These characteristics are highly correlated with the operational actions and are difficult to forge using simple methods. However, current technologies lack the extraction and utilization of transient parameter characteristics during maintenance operations, and also fail to perform joint cross-validation of static changes and process characteristics. Therefore, there is an urgent need for a method that can integrate steady-state changes before and after the operation with dynamic transient characteristics of the operation process, using multi-dimensional traces to reverse-verify the authenticity of maintenance operations. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for cross-validation of elevator maintenance records based on a cloud database, comprising: A maintenance project and parameter expected change model is constructed in a cloud database. The maintenance project and parameter expected change model stores each maintenance project and its corresponding expected change range of parameters before and after operation, as well as transient feature templates of parameters during operation. Obtain the evidence chain to be verified corresponding to this maintenance task. The evidence chain to be verified includes at least the equipment status evidence collected before and after the maintenance and the time series data of equipment operating parameters collected during the maintenance operation. The cross-validation of the operation trace dimension of the evidence chain to be verified includes: extracting a first equipment status parameter snapshot before maintenance begins and a second equipment status parameter snapshot after maintenance ends from the evidence chain to be verified; determining the expected change range of the target parameter corresponding to the current maintenance project based on the maintenance project and the parameter expected change model; calculating the actual change amount of the first snapshot and the second snapshot on the target parameter and determining whether the actual change amount falls within the expected change range to generate a steady-state change verification result; and extracting the time series data of the target parameter within the maintenance operation time window from the evidence chain to be verified, performing pattern matching between the time series data and the transient feature template of the operation process parameter to generate a process verification result; and generating an operation trace dimension score based on the steady-state change verification result and the process verification result. Based on the operation trace dimension score, the anomaly type of this maintenance is determined and output.

[0007] Furthermore, the construction of the maintenance project and expected parameter change model in the cloud database includes: Collect multiple sets of maintenance records that have been verified and confirmed to have been actually performed in the past; For each maintenance record, the average target parameter value within the preset window before maintenance begins is extracted as the parameter value before execution, and the average target parameter value within the preset window after maintenance ends is extracted as the parameter value after execution. The actual change before and after is calculated. The actual change of multiple sets of the same maintenance project is statistically analyzed, and the range from the low quantile to the high quantile of the distribution is taken as the expected change range of the parameters before and after the operation. Furthermore, for each set of maintenance records, high-frequency sampling sequences of the target parameters collected at a preset frequency within the maintenance operation time window are extracted, and parameter sub-sequences reflecting the core operation stage are extracted; multiple parameter sub-sequences of similar maintenance projects are time-aligned and clustered to generate the transient feature template of the operation process parameters, which includes the typical change path and allowable fluctuation envelope of the target parameters.

[0008] Furthermore, the step of performing pattern matching between the time series data and the transient feature template of the operation process parameters includes: The cumulative distance between the actual collected time series data and the typical change path is calculated using a dynamic time warping algorithm, and the proportion of each sampling point in the time series data falling into the allowable fluctuation envelope is evaluated to obtain a matching score. If the matching score is lower than the preset matching threshold, the process verification result is determined to be abnormal.

[0009] Furthermore, if the evidence chain to be verified does not include the time series data of equipment operating parameters collected during the maintenance operation, then in the cross-verification of the operation trace dimension, the process verification result is set to a neutral value, and the operation trace dimension score is generated only based on the steady-state change verification result; wherein, if the steady-state change verification result is normal and the process verification result is a neutral value, then the operation trace dimension score is set to a preset normal score value.

[0010] Furthermore, the evidence chain to be verified also includes personnel identity verification evidence and operational process video evidence; after obtaining the evidence chain to be verified corresponding to this maintenance task, the method further includes: Perform cross-validation of the evidence chain to be verified in terms of space and content dimensions to generate a score in both space and content dimensions; The cross-validation of the spatial and content dimensions includes: extracting photo-type evidence and video-type evidence from the operational process image evidence; extracting local feature descriptors for the photo-type evidence and extracting local feature descriptors for each sampled frame of the video-type evidence; matching the descriptors of the photo-type evidence with the descriptors of each sampled frame and calculating the matching score for each frame; if the matching scores of all sampled frames do not exceed a preset spatial content score threshold, then the spatial content dimension score is set to a preset low score; otherwise, it is set to a preset high score.

[0011] Furthermore, the method also includes: Perform cross-validation of the evidence chain to be verified along the time dimension to obtain a time dimension score; Cross-validation of the evidence chain to be verified at the personnel dimension yields a personnel dimension score; The step of determining and outputting the anomaly type for this maintenance based on the operation trace dimension score specifically includes: weighting and fusing the time dimension score, the personnel dimension score, the spatial content dimension score, and the operation trace dimension score to obtain a comprehensive cross-validation value; when the comprehensive cross-validation value is lower than a preset qualified threshold, determining and outputting the anomaly type based on the lowest dominant dimension among the scores of each dimension.

[0012] Furthermore, the weights used in the weighted fusion are set differently according to the category of the current maintenance project: If the current maintenance project belongs to the mechanical adjustment category, then increase the weight of the operation trace dimension score and decrease the weight of the spatial content dimension score. If the current maintenance project belongs to the visual inspection category, then the weight of the operation trace dimension score is reduced, while the weight of the time dimension score and the spatial content dimension score is increased.

[0013] Furthermore, the target parameters are selected based on the characteristics of the elevator components corresponding to the maintenance project: for the door lock gap adjustment project, the door operator opening and closing time or the peak value of the door operator drive current is selected as the target parameter; for the brake gap adjustment project, the brake response time or the brake sliding distance is selected as the target parameter; for the guide rail lubrication project, the effective value of the car horizontal vibration acceleration is selected as the target parameter. The collection window for the first equipment status parameter snapshot before the start of maintenance ends before the maintenance check-in time, and the collection window for the second equipment status parameter snapshot after the end of maintenance starts after the maintenance check-out time. Before extracting the first and second snapshots, outliers in the window are removed using the median absolute deviation method, and the mean of the remaining data is used as the valid snapshot value. The maintenance operation time window is defined as the time from the maintenance check-in time to the maintenance check-out time, and the time series data is collected from the elevator controller or sensor at a preset frequency.

[0014] Furthermore, the method also includes: Collect confirmation information from subsequent manual reviews or law enforcement inspections regarding the determination of the anomaly type for this maintenance, and use it as a true label; The time dimension score, the personnel dimension score, the spatial content dimension score, the operation trace dimension score, the comprehensive cross-validation value, the dominant dimension, and the real label are associated and stored as training samples; When the number of accumulated training samples reaches a preset training threshold, an objective function is constructed with the weights of the scores of each dimension and the thresholds involved in the calculation of the scores of each dimension as variables. The objective function is an evaluation index of the anomaly detection accuracy. The relationship between the variables and the objective function is fitted using a Gaussian process as a surrogate model. The next set of variable combinations for evaluation is selected by the acquisition function. The iterative evaluation is completed and the surrogate model is updated until the convergence condition is met, and the optimal combination of weight and threshold parameters is obtained. The optimal combination of weights and threshold parameters is updated in the cloud database for subsequent cross-validation.

[0015] A cloud database-based elevator maintenance record cross-verification system includes: The model configuration module is used to execute the method as described in claim 2, and to build a maintenance project and parameter expected change model in the cloud database. The model stores each maintenance project and its corresponding expected change range of parameters before and after operation and transient feature templates of parameters during operation. The evidence acquisition module is used to acquire the evidence chain to be verified corresponding to this maintenance task. The evidence chain to be verified includes at least the equipment status evidence collected before and after maintenance, the time series data of equipment operating parameters collected during the maintenance operation, personnel identity verification evidence, and operation process video evidence. The operation trace verification module is used to perform the cross-verification of the above-mentioned operation trace dimensions, including extracting the first snapshot and the second snapshot for steady-state change verification, and extracting the time series data and the transient feature template of the operation process parameters for pattern matching to generate process verification results, and generating operation trace dimension scores based on the steady-state change verification results and the process verification results. The spatial content verification module is used to perform the cross-verification of the spatial and content dimensions mentioned above and generate spatial content dimension scores. The time dimension verification module is used to perform cross-verification of the evidence chain to be verified in the time dimension and generate a time dimension score. The personnel dimension verification module is used to perform cross-verification of the evidence chain to be verified in terms of personnel dimension and generate personnel dimension scores. The weighted fusion and anomaly determination module is used to execute the above method, weightedly fuse the time dimension score, the personnel dimension score, the spatial content dimension score and the operation trace dimension score to obtain a comprehensive cross-validation value, and determine the anomaly type based on the lowest dominant dimension and output it when the comprehensive cross-validation value is lower than the preset qualified threshold. The feedback learning module collects the results of subsequent manual review or law enforcement inspections as real labels to accumulate training samples. It also employs a Bayesian optimization method to adjust the weights of the weighted fusion and the threshold parameters involved in the score calculation for each dimension, thereby maximizing the accuracy of anomaly detection. The positive and progressive effects of this invention are as follows: By pre-establishing a model of maintenance projects and expected parameter changes, and extracting time-series data of the operation process for transient feature matching, it achieves dual physical trace verification of whether maintenance operations were actually performed, both statically and through the process. This fills the regulatory gap in existing technologies that only focus on personnel presence or static snapshots while neglecting process fingerprints. Simultaneously, it introduces cross-evidence content-level scene consistency matching, effectively identifying fabricated acts where photos and videos are mismatched. Differentiated weight allocation and verification modes are designed for different categories of maintenance projects, taking into account the verification sensitivity for mechanical adjustment and visual inspection projects. Overall, it constructs a closed-loop regulatory link from evidence chain collection and multi-dimensional cross-verification to anomaly classification and feedback optimization, significantly improving the accuracy and automation level of elevator maintenance quality supervision. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the elevator maintenance record cross-verification method based on a cloud database provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] Reference Figure 1 This embodiment provides a method for cross-validation of elevator maintenance records based on a cloud database, including: constructing a maintenance project and expected parameter change model in the cloud database, wherein the maintenance project and expected parameter change model stores the expected change range of parameters before and after each maintenance project and the transient feature template of parameters during the operation process; obtaining the evidence chain to be verified corresponding to the current maintenance task, wherein the evidence chain to be verified includes at least the equipment status evidence collected before and after maintenance and the time series data of equipment operating parameters collected during the maintenance operation; and cross-validating the execution operation trace dimension of the evidence chain to be verified, wherein the cross-validation of the execution operation trace dimension includes: extracting a first equipment status parameter snapshot before the start of maintenance and a second snapshot after the end of maintenance from the evidence chain to be verified. The system generates a steady-state change verification result by taking a snapshot of the equipment status parameters, determining the expected change range of the target parameters corresponding to the current maintenance project based on the maintenance project and the expected change model of the parameters, calculating the actual changes of the first and second snapshots on the target parameters, and determining whether the actual changes fall within the expected change range. It also extracts time-series data of the target parameters within the maintenance operation time window from the evidence chain to be verified, performs pattern matching between the time-series data and the transient feature template of the operation process parameters, and generates a process verification result. Based on the steady-state change verification result and the process verification result, it generates an operation trace dimension score. Based on the operation trace dimension score, it determines the anomaly type of this maintenance and outputs it. Furthermore, the process of constructing the maintenance project and parameter expectation change model in the cloud database is as follows: First, collect multiple sets of historically verified maintenance records. These records originate from maintenance tasks that have been confirmed as actual operations through on-site inspections or law enforcement checks. For each set of maintenance records, extract the average target parameter within a preset window before maintenance begins as the pre-execution parameter value, and extract the average target parameter within a preset window after maintenance ends as the post-execution parameter value. The preset window before maintenance begins is set from 30 minutes before the check-in time to the check-in time, and the preset window after maintenance ends is set from the check-out time to 30 minutes after the check-out time. Outliers in the original data within the window are first removed using the median absolute deviation method, and then the arithmetic mean of the remaining data is taken as the valid snapshot value. Calculate the actual change between the pre-execution snapshot value and the post-execution snapshot value. For multiple sets of actual changes for the same maintenance project, arrange them in ascending order, and take the interval covered by the 10th percentile to the 90th percentile as the expected parameter change range before and after the operation of that maintenance project. Simultaneously, for each set of real maintenance records, high-frequency sampling sequences of the target parameters are extracted within the maintenance operation time window at a preset frequency. The maintenance operation time window is defined as the period from check-in to check-out, with a preset frequency of once per second. Sliding window variance analysis is used to extract parameter subsequences reflecting the core operation phase. Then, a dynamic time warping algorithm is used to time-align multiple parameter subsequences of similar maintenance projects. Based on the aligned sequence set, K-means clustering is performed, with the cluster centers serving as typical change paths. The standard deviation of each sequence value at each time point is calculated, and the allowable fluctuation envelope is constructed by adding or subtracting twice the standard deviation from the cluster center value. This generates a transient feature template of the operation process parameters containing the typical change paths and the allowable fluctuation envelope. It is worth noting that the target parameters are selected based on the characteristics of the elevator components corresponding to the maintenance project: for door lock gap adjustment projects, the door operator opening / closing time or the peak value of the door operator drive current is selected as the target parameter; for brake gap adjustment projects, the brake response time or brake sliding distance is selected as the target parameter; for guide rail lubrication projects, the effective value of the car's horizontal vibration acceleration is selected as the target parameter. For visual inspection items, such as car lighting inspection, the target parameter is selected as the lighting circuit current, and the expected variation range is set to a small fluctuation range of ±0.5% relative to the reference value.

[0019] Furthermore, the process of matching the time series data with the transient feature template of the operation process parameters includes: using a dynamic time warping algorithm to calculate the cumulative distance between the actually collected time series data and the typical change path, and simultaneously evaluating the proportion of each sampling point in the time series data falling within the allowable fluctuation envelope, to obtain a matching score M = p_envelope × exp(-d_DTW / λ), where p_envelope is the proportion of points falling within the envelope, d_DTW is the cumulative distance, and λ is a normalization constant, which can be taken as the product of the sequence length and the average fluctuation amplitude. If the matching score is lower than a preset matching threshold (e.g., 0.7), the process verification result is determined to be abnormal; otherwise, it is normal. When the process verification result is normal, the process verification score is set to 100 points; when it is abnormal, it is set to 0 points. For visual inspection items, the transient feature template of the operation process is a stable straight line. During matching, it is required that each point in the actual sequence is within the small fluctuation envelope; otherwise, the process is determined to be abnormal.

[0020] Furthermore, if the evidence chain to be verified does not include time-series data of equipment operating parameters during the maintenance operation, then in the cross-validation of the operation trace dimension, the process verification result is set to a neutral value, that is, the process verification score is directly assigned 50 points, and the operation trace dimension score is generated only based on the fusion of the steady-state change verification result and the neutral value. One fusion rule is: if the steady-state change verification is normal, then the operation trace dimension score = 0.7 × steady-state verification score (100) + 0.3 × 50 = 85 points; if the steady-state change verification is abnormal, then the score is 0.7 × 0 + 0.3 × 50 = 15 points. In this way, a score with a certain degree of discrimination can still be given when data is missing.

[0021] Furthermore, the evidence chain to be verified also includes personnel identity verification evidence and operation process video evidence. After obtaining the evidence chain to be verified corresponding to this maintenance task, the method also performs cross-verification of the evidence chain in terms of spatial and content dimensions to generate a spatial content dimension score. The specific process is as follows: extract photo-type evidence and video-type evidence from the operation process video evidence; use the ORB algorithm to extract local feature descriptors for the photo-type evidence; sample frames for the video-type evidence at an interval of 2 frames per second, and extract ORB descriptors for each sample frame; match the descriptors of the photo-type evidence with the descriptors of each sample frame, use a distance ratio test (threshold 0.7) to filter out mismatches, and use the larger value between the number of matching points divided by the total number of photo feature points and the total number of frame feature points as the matching score for that frame; if the matching scores of all sample frames do not exceed the preset spatial content score threshold (e.g., 0.4), then the spatial content dimension score is set to 0 points; otherwise, it is set to 100 points. As further explained, this mechanism utilizes the inherent spatial and temporal constraints of different modal evidence to effectively identify fabricated historical photographs that have been misused.

[0022] Furthermore, the method also includes performing cross-validation on the evidence chain to be verified in terms of time dimension and personnel dimension, respectively obtaining time dimension scores and personnel dimension scores. The time dimension cross-validation process is as follows: extract the timestamps of all evidence in the evidence chain to be verified, calculate the time span between the earliest and latest timestamps as the total actual operation time, and compare it with the standard operation time range of the maintenance project in the database; at the same time, check whether the evidence pairs with preset time sequence dependencies meet the order requirements; if both meet the rules, the time dimension score is set to 100 points, if one does not meet the rules, it is set to 50 points, and if neither meets the rules, it is set to 0 points. The personnel dimension cross-validation process is as follows: extract the check-in live face image, the frontal face frame captured from the video, and the RFID pre-stored template image from the evidence chain to be verified, extract 128-dimensional depth feature vectors using the FaceNet model, calculate the cosine similarity between each pair, set the threshold to 0.8, and count the proportion of all evidence pairs that should be compared with a similarity greater than the threshold as the matching rate. The personnel dimension score = 100 × matching rate. Subsequently, the scores for the time dimension, personnel dimension, spatial content dimension, and operation trace dimension are weighted and fused to obtain a comprehensive cross-validation value. Specifically, the time dimension score is denoted as S_time, the personnel dimension score as S_person, the spatial content dimension score as S_scene, and the operation trace dimension score as S_result; the fusion weights for each dimension are denoted as w_time, w_person, w_scene, and w_result, respectively. Each weight is configured differently based on the category of the current maintenance project by querying the cloud database: if the current maintenance project belongs to the mechanical adjustment category (such as door lock gap adjustment, brake gap adjustment, guide rail lubrication, etc.), then w_time=0.20, w_person=0.20, w_scene=0.20, w_result=0.40; if the current maintenance project belongs to the visual inspection category (such as car lighting inspection, safety sign inspection, etc.), then w_time=0.30, w_person=0.25, w_scene=0.30, w_result=0.15. Calculate the comprehensive cross-validation value S_cross, with the formula S_cross=w_time×S_time+w_person×S_person+w_scene×S_scene+w_result×S_result; set the preset pass threshold S_threshold=75.When S_cross < S_threshold, determine the abnormal type according to the dominant dimension with the lowest score in each dimension and output it. For example, when the dominant dimension is the operation trace and the time and personnel are normal, output the abnormal type of "Compliant arrival but maintenance items not actually executed" and the freeze work order instruction; when the dominant dimension is the spatial content, output "There are spatial or scene conflicts in the evidence content"; when the dominant dimension is time and the operation trace is also low, output "Insufficient operation duration and no change in equipment status".

[0023] Furthermore, the method further includes a feedback learning optimization step. After each verification and output of the abnormal type, collect the confirmation information of the determination result of the current maintenance abnormal type by subsequent manual review or law enforcement inspection as the true label, and associate and store the scores of each dimension, the comprehensive cross-verification value, the dominant dimension, and the true label of the current maintenance as a training sample. When the number of accumulated training samples reaches a preset threshold (such as 200), construct an objective function with variables such as the weights of the scores of each dimension, the fusion coefficient of the steady state and the process in the operation trace dimension, and the spatial content matching score threshold. The objective function uses the macro-average F1 score between the abnormal determination result and the true label as the evaluation index. Use the Gaussian process as a surrogate model to fit the relationship between the variables and the objective function, and use the expected improvement function as the acquisition function. Randomly sample several groups of parameter combinations in the initialization stage and evaluate their objective function values to construct an initial surrogate model; then iteratively select the next group of optimal evaluated variable combinations through the acquisition function, actually evaluate the objective function values and update the surrogate model until the preset number of iterations or the optimal value converges. Select the parameter combination that maximizes the objective function in the historical evaluation as the optimal parameter and update it to the configuration table of the cloud database for subsequent cross-verification, so as to continuously improve the system accuracy over time.

[0024] As further illustrative, in one example, suppose a maintenance task is "adjusting the door lock gap." Maintenance personnel check in at 9:00 AM and check out at 9:20 AM. The system obtains a complete chain of evidence to be verified, including snapshots of the door operator's opening and closing time before and after check-in, the operation process duration sequence, on-site photos, and videos. In the cross-validation of the operation trace dimension, the first snapshot, after outlier removal, is 3200 milliseconds, and the second snapshot is 2900 milliseconds, with an actual change of 300 milliseconds, falling within the expected change range of 100-300 milliseconds. The steady-state change verification is normal, with a steady-state score of 100 points. The extracted operation process time series shows that the door opening and closing time underwent multiple jumps before converging to around 2.9 seconds. The matching score with the transient feature template is 0.85, exceeding the threshold of 0.7, indicating that the process verification is normal, with a process score of 100 points. The fused operation trace dimension score is 100 points. In the spatial content dimension verification, the highest matching score between the photo and video frames is 0.6, exceeding the threshold of 0.4, with a score of 100 points. In the time dimension verification, the actual operation time of 20 minutes is within the standard range and the evidence sequence is correct, scoring 100 points. In the personnel dimension verification, all face comparisons are consistent, scoring 100 points. The current project is a mechanical adjustment type, and the weighted fusion comprehensive cross-verification value is 100 points, which is greater than 75 points, so the maintenance is judged to be passed. In another embodiment, if the maintenance personnel did not actually operate, but only briefly pushed against the gantry to create a door opening and closing delay so that the change in the snapshot before and after is still 300 milliseconds, but the door opening and closing time sequence is basically stable without fluctuation during the operation, the matching degree score is only 0.3, the process verification is abnormal and scores 0 points, and the operation trace dimension score is reduced to 30 points according to the rules; if other dimensions are normal, the comprehensive score is 0.2×100+0.2×100+0.2×100+0.4×30=72 points, which is lower than 75 points. The dominant dimension is the operation trace dimension, and the output is an abnormal judgment of "on-site compliance but maintenance project not substantially executed".

[0025] It is worth noting that the specific scores, thresholds, and algorithm selections in the above embodiments are all illustrative examples, and those skilled in the art can make reasonable adjustments according to actual application scenarios.

[0026] Based on the same inventive concept, this embodiment also provides a cloud database-based elevator maintenance record cross-verification system, including a model configuration module, an evidence acquisition module, an operation trace verification module, a spatial content verification module, a time dimension verification module, a personnel dimension verification module, a weighted fusion and anomaly determination module, and a feedback learning module. The model configuration module is used to construct a maintenance project and expected parameter change model in the cloud database; the evidence acquisition module is used to acquire the evidence chain to be verified; the operation trace verification module is used to perform steady-state change verification and process verification and generate operation trace dimension scores; the spatial content verification module is used to perform photo and video matching verification; the time dimension verification module and the personnel dimension verification module calculate the time dimension score and the personnel dimension score, respectively; the weighted fusion and anomaly determination module is used to weightedly fuse the scores of each dimension to obtain a comprehensive cross-verification value and determine the anomaly type; the feedback learning module is used to collect real labels and perform Bayesian optimization to adjust parameters. The specific implementation methods of each module are consistent with the corresponding steps in the above method embodiments, and will not be repeated here.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-validating elevator maintenance records based on a cloud database, characterized in that, include: A maintenance project and parameter expected change model is constructed in a cloud database. The maintenance project and parameter expected change model stores each maintenance project and its corresponding expected change range of parameters before and after operation, as well as transient feature templates of parameters during operation. Obtain the evidence chain to be verified corresponding to this maintenance task. The evidence chain to be verified includes at least the equipment status evidence collected before and after the maintenance and the time series data of equipment operating parameters collected during the maintenance operation. The cross-validation of the operation trace dimension of the evidence chain to be verified includes: extracting a first equipment status parameter snapshot before maintenance begins and a second equipment status parameter snapshot after maintenance ends from the evidence chain to be verified; determining the expected change range of the target parameter corresponding to the current maintenance project based on the maintenance project and the parameter expected change model; calculating the actual change amount of the first snapshot and the second snapshot on the target parameter and determining whether the actual change amount falls within the expected change range to generate a steady-state change verification result; and extracting the time series data of the target parameter within the maintenance operation time window from the evidence chain to be verified, performing pattern matching between the time series data and the transient feature template of the operation process parameter to generate a process verification result; and generating an operation trace dimension score based on the steady-state change verification result and the process verification result. Based on the operation trace dimension score, the anomaly type of this maintenance is determined and output.

2. The method for cross-validation of elevator maintenance records based on cloud database according to claim 1, characterized in that, The construction of the maintenance project and expected parameter change model in the cloud database includes: Collect multiple sets of maintenance records that have been verified and confirmed to have been actually performed in the past; For each maintenance record, the average target parameter value within the preset window before maintenance begins is extracted as the parameter value before execution, and the average target parameter value within the preset window after maintenance ends is extracted as the parameter value after execution. The actual change before and after is calculated. The actual change of multiple sets of the same maintenance project is statistically analyzed, and the range from the low quantile to the high quantile of the distribution is taken as the expected change range of the parameters before and after the operation. Furthermore, for each set of maintenance records, high-frequency sampling sequences of the target parameters collected at a preset frequency within the maintenance operation time window are extracted, and parameter sub-sequences reflecting the core operation stage are extracted; multiple parameter sub-sequences of similar maintenance projects are time-aligned and clustered to generate the transient feature template of the operation process parameters, which includes the typical change path and allowable fluctuation envelope of the target parameters.

3. The method for cross-validation of elevator maintenance records based on cloud database according to claim 2, characterized in that, The step of performing pattern matching between the time series data and the transient feature template of the operation process parameters includes: The cumulative distance between the actual collected time series data and the typical change path is calculated using a dynamic time warping algorithm, and the proportion of each sampling point in the time series data falling into the allowable fluctuation envelope is evaluated to obtain a matching score. If the matching score is lower than the preset matching threshold, the process verification result is determined to be abnormal.

4. The method for cross-validation of elevator maintenance records based on cloud database according to claim 1, characterized in that, If the evidence chain to be verified does not contain time series data of equipment operating parameters collected during the maintenance operation, then in the cross-verification of the operation trace dimension, the process verification result is set to a neutral value, and the operation trace dimension score is generated only based on the steady-state change verification result; wherein, if the steady-state change verification result is normal and the process verification result is a neutral value, then the operation trace dimension score is set to a preset normal score value.

5. The method for cross-validation of elevator maintenance records based on cloud database according to claim 1, characterized in that, The evidence chain to be verified also includes personnel identity verification evidence and operation process video evidence; after obtaining the evidence chain to be verified corresponding to this maintenance task, the method further includes: Perform cross-validation of the evidence chain to be verified in terms of space and content dimensions to generate a score in both space and content dimensions; The cross-validation of the spatial and content dimensions includes: extracting photo-type evidence and video-type evidence from the operational process image evidence; extracting local feature descriptors for the photo-type evidence and extracting local feature descriptors for each sampled frame of the video-type evidence; matching the descriptors of the photo-type evidence with the descriptors of each sampled frame and calculating the matching score for each frame; if the matching scores of all sampled frames do not exceed a preset spatial content score threshold, then the spatial content dimension score is set to a preset low score; otherwise, it is set to a preset high score.

6. The method for cross-validation of elevator maintenance records based on cloud database according to claim 5, characterized in that, The method further includes: Perform cross-validation of the evidence chain to be verified along the time dimension to obtain a time dimension score; Cross-validation of the evidence chain to be verified at the personnel dimension yields a personnel dimension score; The step of determining and outputting the anomaly type for this maintenance based on the operation trace dimension score specifically includes: weighting and fusing the time dimension score, the personnel dimension score, the spatial content dimension score, and the operation trace dimension score to obtain a comprehensive cross-validation value; when the comprehensive cross-validation value is lower than a preset qualified threshold, determining and outputting the anomaly type based on the lowest dominant dimension among the scores of each dimension.

7. The method for cross-validation of elevator maintenance records based on cloud database according to claim 6, characterized in that, The weights used in the weighted fusion are set differently according to the category of the current maintenance project: If the current maintenance project belongs to the mechanical adjustment category, then increase the weight of the operation trace dimension score and decrease the weight of the spatial content dimension score. If the current maintenance project belongs to the visual inspection category, then the weight of the operation trace dimension score is reduced, while the weight of the time dimension score and the spatial content dimension score is increased.

8. The method for cross-validation of elevator maintenance records based on cloud database according to claim 1, characterized in that, The target parameters are selected based on the characteristics of the elevator components corresponding to the maintenance project: for the door lock gap adjustment project, the door operator opening and closing time or the peak value of the door operator drive current is selected as the target parameter; for the brake gap adjustment project, the brake response time or the brake sliding distance is selected as the target parameter; for the guide rail lubrication project, the effective value of the car horizontal vibration acceleration is selected as the target parameter. The collection window for the first equipment status parameter snapshot before the start of maintenance ends before the maintenance check-in time, and the collection window for the second equipment status parameter snapshot after the end of maintenance starts after the maintenance check-out time. Before extracting the first and second snapshots, outliers in the window are removed using the median absolute deviation method, and the mean of the remaining data is used as the valid snapshot value. The maintenance operation time window is defined as the time from the maintenance check-in time to the maintenance check-out time, and the time series data is collected from the elevator controller or sensor at a preset frequency.

9. The method for cross-validation of elevator maintenance records based on cloud database according to claim 6, characterized in that, The method further includes: Collect confirmation information from subsequent manual reviews or law enforcement inspections regarding the determination of the type of maintenance anomaly, and use this as a true label; The time dimension score, the personnel dimension score, the spatial content dimension score, the operation trace dimension score, the comprehensive cross-validation value, the dominant dimension, and the real label are associated and stored as training samples; When the number of accumulated training samples reaches a preset training threshold, an objective function is constructed with the weights of the scores of each dimension and the thresholds involved in the calculation of the scores of each dimension as variables. The objective function is an evaluation index of the anomaly detection accuracy. The relationship between the variables and the objective function is fitted using a Gaussian process as a surrogate model. The next set of variable combinations for evaluation is selected by the acquisition function. The iterative evaluation is completed and the surrogate model is updated until the convergence condition is met, and the optimal combination of weight and threshold parameters is obtained. The optimal combination of weights and threshold parameters is updated in the cloud database for subsequent cross-validation.

10. A cloud database-based elevator maintenance record cross-verification system, characterized in that, include: The model configuration module is used to execute the method as described in claim 2, and to build a maintenance project and parameter expected change model in the cloud database. The model stores each maintenance project and its corresponding expected change range of parameters before and after operation and transient feature templates of parameters during operation. The evidence acquisition module is used to acquire the evidence chain to be verified corresponding to this maintenance task. The evidence chain to be verified includes at least the equipment status evidence collected before and after maintenance, the time series data of equipment operating parameters collected during the maintenance operation, personnel identity verification evidence, and operation process video evidence. The operation trace verification module is used to perform cross-verification of the operation trace dimension as described in any one of claims 1, 3, 4, and 8, including extracting a first snapshot and a second snapshot for steady-state change verification, and extracting the time series data and performing pattern matching with the transient feature template of the operation process parameters to generate a process verification result, and generating an operation trace dimension score based on the steady-state change verification result and the process verification result. The spatial content verification module is used to perform the cross-verification of spatial and content dimensions as described in claim 5, and generate a spatial content dimension score; The time dimension verification module is used to perform cross-verification of the evidence chain to be verified in the time dimension and generate a time dimension score. The personnel dimension verification module is used to perform cross-verification of the evidence chain to be verified in terms of personnel dimension and generate personnel dimension scores. The weighted fusion and anomaly determination module is used to execute the method as described in claims 6-7, weightedly fuse the time dimension score, the personnel dimension score, the spatial content dimension score and the operation trace dimension score to obtain a comprehensive cross-validation value, and determine the anomaly type based on the lowest dominant dimension and output it when the comprehensive cross-validation value is lower than a preset qualified threshold. The feedback learning module is used to perform the method as described in claim 9, collect the results of subsequent manual review or law enforcement inspection as real labels, accumulate training samples, and use Bayesian optimization method to adjust the weights of the weighted fusion and the threshold parameters involved in the calculation of scores for each dimension, so as to maximize the accuracy of anomaly detection.

Images