Elevator maintenance quality evaluation method based on SVR-NSGA-II algorithm

Abstract

The invention discloses an elevator maintenance quality evaluation method based on an SVR-NSGA-I algorithm. The method comprises a data acquisition stage, a data preprocessing stage, an SVR model training stage, an NSGA-I optimization stage and an evaluation system construction stage. Compared with the prior art, the method has the advantages that PSO optimized SVR and adaptive NSGA-I are deeply fused, the multi-target optimization problem of elevator maintenance is solved, an isolated forest-KNN combined preprocessing method is provided, high-dimensional, nonlinear and noisy maintenance data are effectively processed, a time decay factor and fuzzy clustering are introduced, and the maintenance efficiency is improved. The real-time dynamic adjustment of the evaluation standard is realized, and the model transparency and practicability are improved in combination with SHAP value analysis and visual decision support.

Description

Technical Field

[0001] The present invention relates to the technical field of elevator maintenance and quality assessment, and in particular to an elevator maintenance quality evaluation method based on an SVR-NSGA-II algorithm. Background Art

[0002] Currently, elevator maintenance is implemented under the "Regulations on Safety Supervision of Special Equipment," which stipulates that "elevators should be cleaned, lubricated, adjusted, and inspected at least every 15 days." This regulation aims to ensure that elevators receive minimal maintenance, but it is difficult to adapt to elevators of varying quality levels or in varying operating environments. Elevators are wearable equipment, and based on the degradation patterns of their performance, they can be divided into three stages: premature failure, occasional failure, and failure-based wear and tear. When an elevator enters the wear and tear stage, the failure rate gradually increases and performance declines. Preventive maintenance is crucial during this stage. Over time, continuing to perform preventive maintenance every 15 days increases the risk of elevator accidents, meaning a major incident could occur before preventive maintenance is performed. Furthermore, when an elevator first enters the wear and tear stage, wear and tear is not yet severe. Experience suggests performing preventive maintenance at intervals greater than 15 days to reduce maintenance costs.

[0003] Existing methods mostly rely on a single indicator, which cannot fully reflect the comprehensive effect of maintenance and is prone to evaluation bias. Traditional methods have difficulty processing high-dimensional and nonlinear maintenance data and cannot tap the potential value in the data. The operating status of elevators changes over time, and existing methods are difficult to adjust the evaluation criteria in real time, resulting in delayed evaluation results. Cost, reliability, efficiency and other goals conflict with each other during the maintenance process. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above technical defects and provide an elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm.

[0005] In order to solve the above problems, the technical solution of the present invention is: an elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm, including a data acquisition stage, a data preprocessing stage, an SVR model training stage, an NSGA-II optimization stage and an evaluation system construction stage;

[0006] The acquisition phase includes deploying IoT sensors, establishing a maintenance management system, and integrating historical fault data;

[0007] The data preprocessing stage includes using the isolation forest algorithm to eliminate abnormal data points, using the KNN algorithm to fill missing data and constructing 28 composite features;

[0008] The SVR model training phase includes dividing the training set into a training set, a validation set, a test set, and optimizing the SVR parameters using PSO;

[0009] The NSGA-II optimization phase includes setting the population size, number of iterations, crossover rate, and mutation rate, and introducing an adaptive mechanism to generate a Pareto front containing 86 solutions after multiple iterations.

[0010] The evaluation system construction stage includes using AHP to determine reliability weight, cost weight, efficiency weight, introducing entropy weight method to adjust weight and determining membership functions of five evaluation levels based on fuzzy C-means clustering.

[0011] Furthermore, the adaptive threshold introduced in the isolation forest optimization is to dynamically adjust the outlier detection threshold according to the data distribution characteristics to avoid excessive elimination of normal data. The parallel computing architecture uses the MapReduce framework to implement the distributed isolation forest algorithm to improve the efficiency of big data processing. Time series anomaly detection targets time series data such as elevator vibration and speed, combining sliding windows and autoregressive models to identify trend anomalies.

[0012] Furthermore, the KNN algorithm used to fill missing data is based on similarity-weighted KNN, which weights the distance calculation according to the importance of the features to improve the accuracy of filling missing values. The time series KNN considers the time series characteristics of the data and gives priority to samples with close time proximity for filling. The domain knowledge constraints are combined with the elevator maintenance domain knowledge to verify the rationality of the filling results.

[0013] Furthermore, the PSO optimization of SVR parameters includes a multi-output SVR architecture, a PSO-grid search hybrid optimization, and model validation and evaluation. The shared feature layer in the multi-output SVR architecture is designed to construct a shared input layer and feature extraction layer. Different output tasks share underlying features to improve the model's generalization ability. A task correlation matrix is ​​introduced through task correlation modeling to dynamically adjust the weights of different output tasks. At the same time, a residual connection mechanism introduces residual connections in the deep SVR structure to alleviate the gradient vanishing problem. The PSO-grid search hybrid optimization of the adaptive inertia weight introduces a nonlinear decreasing inertia weight into the PSO algorithm to balance global exploration and local development. The elite particles guide the retention of historically optimal particles and guide the population to search for more optimal areas. The grid search refinement performs a high-density grid search near the PSO search results to obtain the optimal parameter combination. Nested cross-validation is used in the model validation and evaluation, using 5×2 cross-validation. The outer layer divides the training / test set and the inner layer optimizes the parameters to avoid overfitting. Noise data and adversarial samples are introduced through robustness evaluation to test the stability of the model under abnormal conditions. Domain adaptability testing is performed to verify the model's generalization ability on different types of elevators.

[0014] Furthermore, the population size is set based on the crossover rate adjustment of population diversity. When the population diversity decreases, the crossover rate is dynamically increased to maintain the exploration ability. The mutation rate is adjusted based on the Pareto front density, and the mutation rate is increased in the sparse area of ​​the solution space to improve the distribution uniformity of the solution. An adaptive elite retention strategy is adopted to dynamically adjust the proportion of elite individuals according to the convergence speed.

[0015] Furthermore, the entropy weight method is introduced to adjust the weights, and the weights of each indicator are dynamically adjusted according to changes in data distribution. Feedback weight learning is used to introduce user feedback, and weight distribution is optimized through reinforcement learning. Domain knowledge is integrated and expert experience is combined to make rational corrections to the weights determined by the algorithm.

[0016] Furthermore, the membership function dynamically adjusts the fuzzy clustering center and membership function parameters according to the distribution of historical data, divides the evaluation results into dynamic grades, dynamically adjusts the evaluation grade boundaries, and simultaneously performs multi-dimensional fuzzy evaluation to construct a multi-dimensional fuzzy evaluation space, taking into account multiple dimensions such as reliability, cost, and efficiency.

[0017] Furthermore, the Pareto frontier exploration is an interactive Pareto frontier exploration, which supports users to perform interactive operations such as dragging, zooming, etc. on the Pareto frontier, view the indicator changes of different schemes in real time, and display the impact of each parameter change on the evaluation results through sensitivity analysis to assist decision-making. At the same time, scenario simulation supports maintenance strategy simulation in different scenarios and predicts possible effects.

[0018] The elevator maintenance quality evaluation system architecture based on the SVR-NSGA-II algorithm is characterized by: a distributed computing architecture, a real-time monitoring and early warning mechanism, and system security and privacy protection;

[0019] The data acquisition layer in the distributed computing architecture uses edge computing devices for data preprocessing and feature extraction to reduce cloud computing pressure. The model training layer implements distributed SVR and NSGA-II algorithms based on Spark MLlib, supporting large-scale parallel processing of elevator data. The application service layer adopts a microservice architecture to decouple functions such as data processing, model prediction, and evaluation analysis.

[0020] The real-time monitoring and early warning mechanism includes dynamic threshold adjustment, which dynamically adjusts the fault warning threshold based on the elevator's historical operating data and current status. It also includes a multi-level early warning mechanism, which sets different levels of early warning signals corresponding to different response strategies. The early warning is interpretable, provides early warning cause analysis and recommended measures, and enhances system credibility.

[0021] The system security and privacy protection includes data encryption transmission: the TLS1.3 protocol is used to encrypt the sensor data transmission process, differential privacy technology is introduced in the data preprocessing stage to protect user privacy, and blockchain evidence storage uses blockchain technology to store maintenance records in an unalterable manner to improve data credibility.

[0022] Compared with the existing technology, the advantages of the present invention are: deep integration of PSO-optimized SVR and adaptive NSGA-II to solve the multi-objective optimization problem of elevator maintenance, proposal of isolation forest-KNN combined preprocessing method to effectively process high-dimensional, nonlinear and noisy maintenance data, introduction of time decay factor and fuzzy clustering to achieve real-time dynamic adjustment of evaluation criteria, and combination of SHAP value analysis and visual decision support to improve model transparency and practicality. DETAILED DESCRIPTION

[0023] The specific embodiments of the present invention are further described below in conjunction with specific embodiments, wherein the same components are represented by the same specific embodiment numbers.

[0024] It should be noted that the words "front", "rear", "left", "right", "up" and "down" used in the following description refer to directions in a specific embodiment, and the words "inside" and "outside" refer to directions toward or away from the geometric center of a specific component, respectively.

[0025] In order to make the contents of the present invention more clearly understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the specific implementation methods in the embodiments of the present invention.

[0026] The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm is characterized by comprising a data collection stage, a data preprocessing stage, an SVR model training stage, an NSGA-II optimization stage, and an evaluation system construction stage;

[0027] The collection phase includes deploying IoT sensors, establishing a maintenance management system, and integrating historical fault data;

[0028] The data preprocessing stage includes using the isolation forest algorithm to remove abnormal data points, using the KNN algorithm to fill missing data, and constructing 28 composite features;

[0029] The adaptive threshold introduced in isolation forest optimization dynamically adjusts the outlier detection threshold based on data distribution characteristics to avoid excessive elimination of normal data. The parallel computing architecture uses the MapReduce framework to implement the distributed isolation forest algorithm, improving big data processing efficiency. Time series anomaly detection targets time series data such as elevator vibration and speed, combining sliding windows and autoregressive models to identify trend anomalies.

[0030] The KNN algorithm is used to fill missing data based on similarity-weighted KNN. Distance calculations are weighted according to feature importance to improve the accuracy of missing value filling. Time series KNN considers the time series characteristics of the data and prioritizes samples with close time proximity for filling. Domain knowledge constraints are combined with elevator maintenance domain knowledge to verify the rationality of the filling results.

[0031] The SVR model training phase includes dividing the training set into training set, validation set, and test set and using PSO to optimize SVR parameters;

[0032] PSO optimization of SVR parameters includes a multi-output SVR architecture, PSO-grid search hybrid optimization, and model validation and evaluation. The shared feature layer in the multi-output SVR architecture is designed to construct a shared input layer and feature extraction layer. Different output tasks share underlying features to improve model generalization. Task correlation modeling introduces a task correlation matrix, dynamically adjusting the weights of different output tasks. At the same time, the residual connection mechanism introduces residual connections in the deep SVR structure to alleviate the gradient vanishing problem. PSO-grid search hybrid optimization of adaptive inertia weights introduces nonlinear decreasing inertia weights into the PSO algorithm to balance global exploration and local development. Elite particles guide the retention of historically optimal particles and guide the population to search for more optimal areas. Grid search refinement performs high-density grid search near the PSO search results to obtain the optimal parameter combination. Nested cross-validation is used in model validation and evaluation, using 5×2 cross-validation. The outer layer divides the training / test set and the inner layer optimizes parameters to avoid overfitting. Noise data and adversarial samples are introduced through robustness evaluation to test the stability of the model under abnormal conditions. Domain adaptability testing is also carried out to verify the model generalization ability on different types of elevators.

[0033] The NSGA-II optimization phase includes setting the population size, number of iterations, crossover rate, and mutation rate, and introducing an adaptive mechanism. After multiple iterations, a Pareto front containing 86 solutions is generated.

[0034] The population size is set based on the crossover rate adjustment of population diversity. When the population diversity decreases, the crossover rate is dynamically increased to maintain the exploration ability. The mutation rate is adjusted based on the Pareto front density. The mutation rate is increased in sparse areas of the solution space to improve the distribution uniformity of the solution. The adaptive elite retention strategy dynamically adjusts the proportion of elite individuals according to the convergence speed.

[0035] The evaluation system construction stage includes using AHP to determine the reliability weight, cost weight, efficiency weight, introducing the entropy weight method to adjust the weight, and determining the membership function of the five evaluation levels based on fuzzy C-means clustering.

[0036] The membership function dynamically adjusts the fuzzy clustering center and membership function parameters according to the distribution of historical data, divides the evaluation results according to the dynamic level, dynamically adjusts the evaluation level boundaries, and simultaneously performs multi-dimensional fuzzy evaluation to construct a multi-dimensional fuzzy evaluation space, taking into account multiple dimensions such as reliability, cost, and efficiency;

[0037] The entropy weight method is introduced to adjust the weights. The weights of each indicator are dynamically adjusted according to the changes in data distribution. Feedback weight learning is introduced, user feedback is introduced, and weight distribution is optimized through reinforcement learning. Domain knowledge is integrated and expert experience is combined to make reasonable corrections to the weights determined by the algorithm.

[0038] Pareto frontier exploration is an interactive Pareto frontier exploration that supports interactive operations such as dragging, zooming, and other interactive operations on the Pareto frontier. Users can view the indicator changes of different schemes in real time, and display the impact of each parameter change on the evaluation results through sensitivity analysis to assist in decision-making. At the same time, scenario simulation supports maintenance strategy simulation in different scenarios and predicts possible effects.

[0039] The elevator maintenance quality evaluation system architecture based on the SVR-NSGA-II algorithm is characterized by: a distributed computing architecture, a real-time monitoring and early warning mechanism, and system security and privacy protection;

[0040] The data collection layer in the distributed computing architecture uses edge computing devices for data preprocessing and feature extraction to reduce cloud computing pressure. The model training layer implements distributed SVR and NSGA-II algorithms based on Spark MLlib, supporting large-scale parallel processing of elevator data. The application service layer adopts a microservice architecture to decouple functions such as data processing, model prediction, and evaluation analysis.

[0041] The real-time monitoring and early warning mechanism includes dynamic threshold adjustment. According to the historical operation data and current status of the elevator, the fault warning threshold is dynamically adjusted. The multi-level early warning mechanism sets different levels of early warning signals corresponding to different response strategies. The early warning is explainable and provides early warning cause analysis and recommended measures to enhance system credibility.

[0042] System security and privacy protection include data encryption transmission: the TLS1.3 protocol is used to encrypt the sensor data transmission process, differential privacy technology is introduced in the data preprocessing stage to protect user privacy, and blockchain evidence storage uses blockchain technology to store maintenance records in an unalterable manner to improve data credibility.

[0043] The present invention and its embodiments are described above. Such description is not restrictive. The specific embodiments shown in the specific embodiments are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by the above, and does not deviate from the purpose of the invention, without creatively designing a structure and embodiment similar to the technical solution, they shall fall within the scope of protection of the present invention.

Claims

1. An elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm is characterized by: It includes data collection stage, data preprocessing stage, SVR model training stage, NSGA-II optimization stage and evaluation system construction stage; The acquisition phase includes deploying IoT sensors, establishing a maintenance management system, and integrating historical fault data; The data preprocessing stage includes using the isolation forest algorithm to eliminate abnormal data points, using the KNN algorithm to fill missing data and constructing 28 composite features; The SVR model training phase includes dividing the training set into a training set, a validation set, a test set, and optimizing the SVR parameters using PSO; The NSGA-II optimization phase includes setting the population size, number of iterations, crossover rate, and mutation rate, and introducing an adaptive mechanism to generate a Pareto front containing 86 solutions after multiple iterations. The evaluation system construction stage includes using AHP to determine reliability weight, cost weight, efficiency weight, introducing entropy weight method to adjust weight and determining membership functions of five evaluation levels based on fuzzy C-means clustering.

2. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1 is characterized in that: The adaptive threshold introduced in the isolation forest optimization is to dynamically adjust the outlier detection threshold according to the data distribution characteristics to avoid excessive elimination of normal data. The parallel computing architecture uses the MapReduce framework to implement the distributed isolation forest algorithm to improve the efficiency of big data processing. Time series anomaly detection targets time series data such as elevator vibration and speed, combining sliding windows and autoregressive models to identify trend anomalies.

3. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1 is characterized in that: The KNN algorithm used to fill missing data is based on similarity-weighted KNN, which weights distance calculation according to feature importance to improve the accuracy of missing value filling. The time series KNN considers the time series characteristics of the data and gives priority to samples with close time proximity for filling. The domain knowledge constraint is combined with the elevator maintenance domain knowledge to verify the rationality of the filling results.

4. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1, characterized in that: The PSO-grid search hybrid optimization method for optimizing SVR parameters includes a multi-output SVR architecture, a PSO-grid search hybrid optimization method, and model verification and evaluation. The shared feature layer in the multi-output SVR architecture is designed to construct a shared input layer and a feature extraction layer. Different output tasks share underlying features to improve the generalization ability of the model. The task correlation matrix is ​​introduced through task correlation modeling to dynamically adjust the weights of different output tasks. At the same time, the residual connection mechanism introduces residual connections in the deep SVR structure to alleviate the gradient vanishing problem. The PSO-grid search hybrid optimization method for adaptive inertia weight introduces a nonlinear decreasing inertia weight into the PSO algorithm to balance global exploration and local development. The elite particles guide the retention of historical optimal particles and guide the population to search for better areas. The grid search refinement performs a high-density grid search near the PSO search results to obtain the optimal parameter combination. Nested cross-validation is used in the model verification and evaluation, using 5×2 cross-validation. The outer layer divides the training / test set and the inner layer optimizes the parameters to avoid overfitting. Noise data and adversarial samples are introduced through robustness evaluation to test the stability of the model under abnormal conditions. Domain adaptability testing is performed to verify the generalization ability of the model on different types of elevators.

5. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1, characterized in that: The population size is set based on the crossover rate adjustment of population diversity. When the population diversity decreases, the crossover rate is dynamically increased to maintain the exploration ability. The mutation rate is adjusted based on the Pareto front density, and the mutation rate is increased in the sparse area of ​​the solution space to improve the distribution uniformity of the solution. The adaptive elite retention strategy dynamically adjusts the proportion of elite individuals according to the convergence speed.

6. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1, characterized in that: The entropy weight method is introduced to adjust the weights, and the weights of each indicator are dynamically adjusted according to changes in data distribution. Feedback weight learning introduces user feedback, and weight distribution is optimized through reinforcement learning. Domain knowledge is integrated and expert experience is combined to make reasonable corrections to the weights determined by the algorithm.

7. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1, characterized in that: The membership function dynamically adjusts the fuzzy clustering center and membership function parameters according to the historical data distribution, divides the evaluation results distribution according to the dynamic level, dynamically adjusts the evaluation level boundary, and simultaneously performs multi-dimensional fuzzy evaluation to construct a multi-dimensional fuzzy evaluation space, while considering multiple dimensions such as reliability, cost, and efficiency.

8. The elevator maintenance quality evaluation method based on the SVR-NSGA-II algorithm according to claim 1, characterized in that: The Pareto frontier exploration is an interactive Pareto frontier exploration that supports users to perform interactive operations such as dragging, zooming, etc. on the Pareto frontier, view the indicator changes of different schemes in real time, and display the impact of each parameter change on the evaluation results through sensitivity analysis to assist decision-making. At the same time, scenario simulation supports maintenance strategy simulation in different scenarios and predicts possible effects.

9. The elevator maintenance quality evaluation system architecture based on the SVR-NSGA-II algorithm is characterized by: Including distributed computing architecture, real-time monitoring and early warning mechanism, and system security and privacy protection; The data acquisition layer in the distributed computing architecture uses edge computing devices for data preprocessing and feature extraction to reduce cloud computing pressure. The model training layer implements distributed SVR and NSGA-II algorithms based on Spark MLlib to support large-scale parallel processing of elevator data. The application service layer adopts a microservice architecture to decouple functions such as data processing, model prediction, and evaluation analysis. The real-time monitoring and early warning mechanism includes dynamic threshold adjustment, which dynamically adjusts the fault warning threshold based on the elevator's historical operating data and current status. It also includes a multi-level early warning mechanism, which sets different levels of early warning signals corresponding to different response strategies. The early warning is interpretable, provides early warning cause analysis and recommended measures, and enhances system credibility. The system security and privacy protection includes data encryption transmission: the TLS1.3 protocol is used to encrypt the sensor data transmission process, differential privacy technology is introduced in the data preprocessing stage to protect user privacy, and blockchain evidence storage uses blockchain technology to store maintenance records in an unalterable manner to improve data credibility.