Elevator fault prediction method and system based on big data technology

Abstract

The invention discloses an elevator fault prediction method and system based on the big data technology, and the method comprises the steps: collecting and preprocessing elevator multi-source data, and generating a standardized data set; performing kernel function mapping and kernel principal component analysis dimensionality reduction to obtain a dimensionality-reduced feature sequence; the sequence is sent into a gating Transform in a segmented mode, and a time sequence modeling result is output; identifying a fault type based on a time sequence modeling result and predicting a future fault probability; the risk threshold is compared, the fault risk is judged, and early warning information is generated; and sending a judgment result and early warning information to an operation and maintenance end to assist in maintenance decision making. According to the method, efficient prediction and intelligent early warning of elevator faults are achieved by fusing kernel principal component analysis and gating Transform, and the operation and maintenance response efficiency and the equipment operation safety are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent elevator operation and maintenance technology, and in particular to an elevator fault prediction method and system based on big data technology. Background Art

[0002] Against the backdrop of accelerating urbanization, elevators, as an indispensable means of vertical transportation in high-rise buildings, are increasingly attracting significant attention for their operational safety and reliability. Traditional elevator maintenance relies primarily on a combination of regular inspections and passive repair reporting. This approach not only suffers from delayed response times and weak fault prediction capabilities, but also fails to fully utilize the growing amount of elevator operational data, making it difficult to promptly identify potential equipment risks. This results in a high elevator failure rate, potentially impacting passenger safety and building operations.

[0003] In recent years, with the development of the Internet of Things and information technology, a large number of sensors and smart terminals have been applied to elevator systems. These sensors can collect multi-dimensional and multi-type elevator operation data in real time, providing a data foundation for intelligent operation and maintenance and accurate prediction. However, existing elevator fault prediction technologies are mostly based on a single data source, simple feature analysis, or static modeling, failing to achieve efficient integration and deep mining of massive multi-source data. Furthermore, their processing capabilities for high-dimensional and complex data are limited, making it difficult to capture subtle changes and temporal dependencies during elevator operation. Traditional models generally lack the ability to intelligently respond to abnormal situations, have relatively simple early warning mechanisms, and cannot dynamically adapt to the variability of the elevator operating environment. As a result, both prediction accuracy and practical application value need to be improved.

[0004] Therefore, how to provide an elevator fault prediction method and system based on big data technology is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of the present invention is to propose a method and system for predicting elevator failures based on big data technology. This method leverages multi-source big data acquisition, kernel principal component analysis dimensionality reduction, gated Transformer time series modeling, and neural network classification and probabilistic prediction techniques. It describes in detail the entire process from efficient processing of elevator operation data, deep feature extraction, to multi-level intelligent risk warning, enabling dynamic and accurate prediction of elevator failures. Compared with existing technologies, this method offers advantages such as strong data fusion capabilities, effective feature extraction, accurate time series dependency modeling, and timely risk response, contributing to improved intelligence and safety management of elevator operations and maintenance.

[0006] An elevator fault prediction method based on big data technology according to an embodiment of the present invention includes:

[0007] Collect multi-source data of elevators, pre-process the multi-source data, and obtain a standardized data set;

[0008] The kernel function is used to perform feature mapping on the standardized data set, and nonlinear dimensionality reduction is performed through kernel principal component analysis to extract the principal component features and obtain the feature sequence after dimensionality reduction;

[0009] The feature sequence after dimensionality reduction is segmented according to the set time window. Each segment of the feature sequence is input into the gated Transformer. The gating mechanism is used to filter key temporal features. The multi-head self-attention mechanism is combined to model the long-term dependencies between features and output the temporal modeling results.

[0010] Based on the output of time series modeling results, fault type identification and probability prediction are carried out to obtain the elevator operation status score and the probability of future faults;

[0011] Compare the probability of future failures with the preset risk threshold to determine whether the elevator currently has a potential failure risk. If the risk reaches a high level, a corresponding warning message will be generated;

[0012] The judgment results and warning information are sent to the operation and maintenance personnel, and maintenance recommendations are formulated based on historical maintenance records.

[0013] Optionally, the multi-source data specifically includes real-time operating parameters collected by elevator sensors, environmental monitoring data, equipment maintenance records and basic elevator attribute data.

[0014] Optionally, the preprocessing of multi-source data specifically includes time synchronization, outlier removal, missing value filling and normalization.

[0015] Optionally, the step of performing feature mapping on the standardized data set using a kernel function, performing nonlinear dimensionality reduction through kernel principal component analysis, extracting principal component features, and obtaining a feature sequence after dimensionality reduction includes:

[0016] According to different types of features in the obtained standardized data set, a variety of kernel functions are selected, including Gaussian radial basis kernel, polynomial kernel, and Laplace kernel, and initial weight coefficients are set for each type of kernel function;

[0017] Based on the dynamic measurement results of the correlation between each type of feature and fault prediction, an adaptive weight is assigned to each type of kernel function, and an adaptive adjustment coefficient for sample pairs is introduced to construct a fusion kernel function;

[0018] Using the fusion kernel function, feature mapping is performed on each pair of samples in the standardized data set. The fusion kernel function values ​​between all sample pairs are calculated in sequence. All fusion kernel function values ​​are filled into a square matrix according to the order of sample arrangement in the data set to form a weighted fusion kernel matrix. The mean of each row and column of the weighted fusion kernel matrix is ​​calculated respectively, and the overall mean is corrected. Each item of the weighted fusion kernel matrix is ​​centered with zero mean to obtain a centralized fusion kernel matrix.

[0019] In the feature extraction process, the local relationship between samples is combined to construct a k-nearest neighbor weight matrix for each sample, and the local preservation constraint is introduced into the principal component optimization objective. The principal component optimization objective is to maximize the global variance of the principal component while minimizing the distance difference between the low-dimensional representations of the k-nearest neighbors.

[0020] According to the set principal component optimization goal, the centralized fusion kernel matrix is ​​decomposed into eigenvalues ​​to obtain the principal component direction with both global structure and local preservation capabilities;

[0021] Select the first m principal component directions whose cumulative contribution rate is greater than the set threshold. For each sample in the standardized data set, the feature vector obtained by fusion kernel function calculation is projected on each principal component direction in turn, and the projection values ​​on all principal component directions are recorded. The projection values ​​of all principal component directions are combined into the reduced dimension feature vector of the corresponding sample. The reduced dimension feature vectors of all samples are collected to form a reduced dimension feature sequence.

[0022] In actual application, the distribution drift of newly collected data and historical data in the fusion kernel space is monitored in real time. When the drift metric exceeds the preset threshold, the type, weight and local preservation parameters of the fusion kernel function are automatically adjusted to reconstruct the fusion kernel matrix and principal component direction.

[0023] Optionally, the feature sequence after dimensionality reduction is segmented according to a set time window, each segment of the feature sequence is input into a gated Transformer, the gating mechanism is used to screen key time series features, and the long-term dependency between features is modeled in combination with a multi-head self-attention mechanism to output the time series modeling results, including:

[0024] According to the pre-set time window, the feature sequence after dimensionality reduction is divided into multiple non-overlapping feature segmentation sequences, each of which contains the dimensionality reduction feature data of consecutive time steps;

[0025] For each feature segmentation sequence, all input features are divided into several feature groups according to the source or statistical characteristics of the input features, and a set of independent gating parameters is set for each feature group;

[0026] When the feature segmentation sequence is input into the gating mechanism, samples containing known fault state labels in the standardized data set are selected. The prediction results generated by each feature group are compared with the corresponding true fault state labels. The proportion of samples with consistent prediction results and true labels to the total number of labeled samples is used as the prediction accuracy of the feature group. Combined with the distribution characteristics and prediction accuracy of the feature group, the sharing relationship of the gating parameters between the feature groups is dynamically adjusted. Some gating parameters are shared between feature groups, and some gating parameters are dedicated to a single feature group.

[0027] When calculating each gating weight, an activation function is dynamically selected from the three activation functions of Sigmoid, ReLU, and Swish based on the statistical distribution of the current input features and the input order or iteration rounds when adjusting the gating weight parameters using a standardized dataset. The selected activation function is used to perform a nonlinear transformation on the input features to obtain the gating weight at the current moment.

[0028] For each time step and each feature in the feature segmentation sequence, the variance calculation method is used to evaluate the uncertainty of the input data. According to the uncertainty result, the gating weight is assigned, and the uncertainty lower limit threshold and the uncertainty upper limit threshold are set as the basis for adjusting the gating weight. When the uncertainty evaluation value is less than the uncertainty lower limit threshold, a gating weight close to 1 is assigned. When the uncertainty evaluation value is greater than the uncertainty upper limit threshold, a gating weight close to 0 is assigned. When the uncertainty evaluation value is between the upper and lower limit thresholds, the gating weight is calculated between 0 and 1 according to the linear interpolation method, and the gating weight distribution is dynamically adjusted.

[0029] Using a gating mechanism that has undergone dynamic gating parameter sharing, adaptive activation function switching, and input uncertainty adjustment, all features of each feature segmentation sequence are gated to obtain a dynamically filtered feature sequence.

[0030] The dynamically filtered feature sequence is input into the multi-head self-attention mechanism to analyze and capture the long-term dependencies between different feature groups and between different time steps, and obtain the feature sequence fused by the multi-head self-attention;

[0031] The multi-head self-attention fusion feature sequence is sequentially input into the multi-layer gated Transformer, the temporal dependency features across time periods are recursively extracted, and a complete temporal modeling feature sequence is output as the final temporal modeling result.

[0032] Optionally, the method of performing fault type identification and probability prediction based on the output time series modeling results to obtain an elevator operation status score and a probability of future fault occurrence includes:

[0033] The output time series modeling feature sequence is used as input data and divided into a feature set for classification tasks and a feature set for regression tasks;

[0034] The feature set used for the classification task is input into a fully connected neural network structure consisting of multiple linear transformation layers and nonlinear activation layers. The fault type prediction probability distribution at each time step is calculated through forward propagation. The fault type corresponding to the maximum fault type prediction probability is selected from the fault type prediction probability distribution as the fault type prediction label for the current time step.

[0035] The feature set used for the regression task is input into another fully connected neural network structure, which continuously performs multi-layer linear mapping and activation operations, outputs the numerical results corresponding to each time step, and normalizes the numerical results to the interval [0,1] as the probability of future fault occurrence;

[0036] Based on the fault type prediction label and future fault probability corresponding to each time step, combined with the preset risk factor weights and scoring calculation standards, the elevator's operating status score for the current time step is calculated to quantify the risk level of the elevator's fault in the current state;

[0037] The fault type prediction label, future fault probability and operating status score corresponding to each time step are used as the final prediction output results.

[0038] Optionally, the probability of future failure is compared with a preset risk threshold to determine whether the elevator currently has a potential failure risk. If the risk reaches a high level, corresponding warning information is generated, including:

[0039] Set up multi-level risk assessment criteria to assign multiple risk levels to the probability of future failures. Each risk level corresponds to a specific probability threshold, and the probability thresholds are arranged from low to high.

[0040] Obtain the future failure probability corresponding to each time step, and compare each future failure probability value with the probability threshold of each risk level set;

[0041] For each time step, determine which threshold interval the future failure probability of the time step falls into, and determine the risk level to which the time step belongs accordingly;

[0042] Set the judgment criteria for high risk levels. When the risk level of a time step reaches or exceeds the judgment criteria, the time step is judged as a high-risk state for potential failure. An early warning prompt containing the time step sequence number, risk level, and predicted probability information is generated and submitted to the alarm management platform.

[0043] All time steps judged as high risk are summarized and output as a fault warning time series.

[0044] An elevator fault prediction system based on big data technology according to an embodiment of the present invention includes the following modules:

[0045] Data acquisition module, used to collect multi-source data of elevator operation;

[0046] Data preprocessing module, used to preprocess multi-source data and generate standardized data sets;

[0047] The kernel principal component analysis module is used to perform kernel function feature mapping and principal component analysis on the standardized data set to obtain the feature sequence after dimensionality reduction;

[0048] The time series modeling module is used to segment the feature sequence after dimensionality reduction, input it into the gated Transformer, filter key time series features, model long-term dependencies between features, and output the time series modeling results;

[0049] The fault prediction module is used to identify fault types and predict probabilities based on the time series modeling results, obtaining the elevator operation status score and the probability of future faults;

[0050] The risk assessment module compares the probability of future failures with the risk threshold to determine potential failure risks and generates warning information when the risk reaches a high level.

[0051] The information push module is used to send risk assessment and warning information to operation and maintenance personnel, and output maintenance suggestions based on historical maintenance records.

[0052] The beneficial effects of the present invention are:

[0053] The present invention achieves efficient processing and deep feature extraction of elevator operation data by introducing multi-source data fusion, kernel principal component analysis and gated Transformer, overcoming the shortcomings of existing technologies in data utilization, feature modeling and risk response. Through systematic analysis and nonlinear dimensionality reduction of high-dimensional, heterogeneous and multi-time series data, it effectively captures subtle changes and potential abnormal signals in the elevator operation status. In the time series modeling and prediction stage, the gated Transformer structure has powerful time series dependency modeling and key feature screening capabilities, accurately identifying fault types and risk trends in complex operating environments. The classification and probability prediction method based on neural networks, combined with the principal component interpretation strength and dynamic threshold risk judgment mechanism, not only improves the accuracy and response speed of fault prediction, but also realizes real-time intelligent early warning of multi-level risks, and timely assists operation and maintenance personnel in formulating scientific maintenance strategies. The present invention improves the intelligence, automation and safety level of elevator operation and maintenance management, effectively reduces the operational risks and maintenance costs caused by faults, and has significant engineering application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0055] Figure 1 This is a flow chart of an elevator fault prediction method based on big data technology proposed by the present invention;

[0056] Figure 2 This is a structural diagram of an elevator fault prediction system based on big data technology proposed in the present invention. DETAILED DESCRIPTION

[0057] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0058] refer to Figure 1 , an elevator fault prediction method based on big data technology, including:

[0059] Collect multi-source data of elevators, pre-process the multi-source data, and obtain a standardized data set;

[0060] The kernel function is used to perform feature mapping on the standardized data set, and nonlinear dimensionality reduction is performed through kernel principal component analysis to extract the principal component features and obtain the feature sequence after dimensionality reduction;

[0061] The feature sequence after dimensionality reduction is segmented according to the set time window. Each segment of the feature sequence is input into the gated Transformer. The gating mechanism is used to filter key temporal features. The multi-head self-attention mechanism is combined to model the long-term dependencies between features and output the temporal modeling results.

[0062] Based on the output of time series modeling results, fault type identification and probability prediction are carried out to obtain the elevator operation status score and the probability of future faults;

[0063] Compare the probability of future failures with the preset risk threshold to determine whether the elevator currently has a potential failure risk. If the risk reaches a high level, a corresponding warning message will be generated;

[0064] The judgment results and warning information are sent to the operation and maintenance personnel, and maintenance recommendations are formulated based on historical maintenance records.

[0065] In this embodiment, the multi-source data specifically includes real-time operating parameters collected by elevator sensors, environmental monitoring data, equipment maintenance records, and basic elevator attribute data.

[0066] In this embodiment, the preprocessing of multi-source data specifically includes time synchronization, outlier removal, missing value filling and normalization.

[0067] In this embodiment, the standardized data set is subjected to feature mapping using a kernel function, nonlinear dimensionality reduction is performed through kernel principal component analysis, principal component features are extracted, and a feature sequence after dimensionality reduction is obtained, including:

[0068] According to different types of features in the obtained standardized data set, a variety of kernel functions are selected, including Gaussian radial basis kernel, polynomial kernel, and Laplace kernel, and initial weight coefficients are set for each type of kernel function;

[0069] According to the dynamic measurement results of the correlation between each type of feature and fault prediction, an adaptive weight is assigned to each type of kernel function, and an adaptive adjustment coefficient for sample pairs is introduced to construct the fusion kernel function K fused (x i ,x j ):

[0070]

[0071] Among them, K q (x i ,x j ) is the qth type kernel function, r q is the weight coefficient determined by the correlation between the feature and the target variable, γ q (x i ,x j ) is the qth type kernel function for sample pair x i ,x j The adaptive adjustment coefficient, p is the total number of kernel functions;

[0072] The fusion kernel function describes a similarity measurement method for the fusion of multiple kernel functions, aiming to achieve accurate feature mapping of complex data, dynamically allocate kernel function weights according to the correlation between each type of feature and the fault prediction target, and introduce adaptive adjustment coefficients for different sample pairs to achieve an organic combination of local and global features. By weightedly fusing the output results of multiple kernel functions, the advantages of various kernel functions under different data distributions are fully utilized, the expressive power of the feature space and the generalization performance of the model are improved, and it can adapt to the complexity of the data structure and enhance the robustness and accuracy of principal component analysis and model training.

[0073] Using the fusion kernel function, feature mapping is performed on each pair of samples in the standardized data set. The fusion kernel function values ​​between all sample pairs are calculated in sequence. All fusion kernel function values ​​are filled into a square matrix according to the order of sample arrangement in the data set to form a weighted fusion kernel matrix. The mean of each row and column of the weighted fusion kernel matrix is ​​calculated respectively, and the overall mean is corrected. Each item of the weighted fusion kernel matrix is ​​centered with zero mean to obtain a centralized fusion kernel matrix.

[0074] In the feature extraction process, the local relationship between samples is combined to construct a k-nearest neighbor weight matrix for each sample, and the local preservation constraint is introduced into the principal component optimization objective. The principal component optimization objective is to maximize the global variance of the principal component while minimizing the distance difference between the low-dimensional representations of the k-nearest neighbors.

[0075] According to the set principal component optimization goal, the centralized fusion kernel matrix is ​​decomposed into eigenvalues ​​to obtain the principal component direction with both global structure and local preservation capabilities;

[0076] Select the first m principal component directions whose cumulative contribution rate is greater than the set threshold. For each sample in the standardized data set, the feature vector obtained by fusion kernel function calculation is projected on each principal component direction in turn, and the projection values ​​on all principal component directions are recorded. The projection values ​​of all principal component directions are combined into the reduced dimension feature vector of the corresponding sample. The reduced dimension feature vectors of all samples are collected to form a reduced dimension feature sequence.

[0077] In actual application, the distribution drift of newly collected data and historical data in the fusion kernel space is monitored in real time. When the drift metric exceeds the preset threshold, the type, weight and local preservation parameters of the fusion kernel function are automatically adjusted to reconstruct the fusion kernel matrix and principal component direction.

[0078] In this embodiment, the feature sequence after dimensionality reduction is segmented according to the set time window, and each segment of the feature sequence is input into the gated Transformer. The gating mechanism is used to screen key temporal features, and the long-term dependency between features is modeled in combination with the multi-head self-attention mechanism to output the temporal modeling results, including:

[0079] According to the pre-set time window, the feature sequence after dimensionality reduction is divided into multiple non-overlapping feature segmentation sequences, each of which contains the dimensionality reduction feature data of consecutive time steps;

[0080] For each feature segmentation sequence, all input features are divided into several feature groups according to the source or statistical characteristics of the input features, and a set of independent gating parameters is set for each feature group;

[0081] When the feature segmentation sequence is input into the gating mechanism, samples containing known fault state labels in the standardized data set are selected. The prediction results generated by each feature group are compared with the corresponding true fault state labels. The proportion of samples with consistent prediction results and true labels to the total number of labeled samples is used as the prediction accuracy of the feature group. Combined with the distribution characteristics and prediction accuracy of the feature group, the sharing relationship of the gating parameters between the feature groups is dynamically adjusted. Some gating parameters are shared between feature groups, and some gating parameters are dedicated to a single feature group.

[0082] When calculating each gating weight, an activation function is dynamically selected from the three activation functions of Sigmoid, ReLU, and Swish based on the statistical distribution of the current input features and the input order or iteration rounds when adjusting the gating weight parameters using a standardized dataset. The selected activation function is used to perform a nonlinear transformation on the input features to obtain the gating weight at the current moment.

[0083] For each time step and each feature in the feature segmentation sequence, the variance calculation method is used to evaluate the uncertainty of the input data. According to the uncertainty result, the gating weight is assigned, and the uncertainty lower limit threshold and the uncertainty upper limit threshold are set as the basis for adjusting the gating weight. When the uncertainty evaluation value is less than the uncertainty lower limit threshold, a gating weight close to 1 is assigned. When the uncertainty evaluation value is greater than the uncertainty upper limit threshold, a gating weight close to 0 is assigned. When the uncertainty evaluation value is between the upper and lower limit thresholds, the gating weight is calculated between 0 and 1 according to the linear interpolation method, and the gating weight distribution is dynamically adjusted.

[0084] Using a gating mechanism that has undergone dynamic gating parameter sharing, adaptive activation function switching, and input uncertainty adjustment, all features of each feature segmentation sequence are gated to obtain a dynamically filtered feature sequence.

[0085] The dynamically filtered feature sequence is input into the multi-head self-attention mechanism to analyze and capture the long-term dependencies between different feature groups and between different time steps, and obtain the feature sequence fused by the multi-head self-attention;

[0086] The multi-head self-attention fusion feature sequence is sequentially input into the multi-layer gated Transformer, the temporal dependency features across time periods are recursively extracted, and a complete temporal modeling feature sequence is output as the final temporal modeling result.

[0087] In this embodiment, the fault type identification and probability prediction are carried out based on the output time series modeling results to obtain the elevator operation status score and the probability of future faults, including:

[0088] The output time series modeling feature sequence is used as input data and divided into a feature set for classification tasks and a feature set for regression tasks;

[0089] The feature set used for the classification task is input into a fully connected neural network structure consisting of multiple linear transformation layers and nonlinear activation layers. The fault type prediction probability distribution at each time step is calculated through forward propagation. The fault type corresponding to the maximum fault type prediction probability is selected from the fault type prediction probability distribution as the fault type prediction label for the current time step.

[0090] The feature set used for the regression task is input into another fully connected neural network structure, which continuously performs multi-layer linear mapping and activation operations, outputs the numerical results corresponding to each time step, and normalizes the numerical results to the interval [0,1] as the probability of future fault occurrence;

[0091] Based on the fault type prediction label and future fault probability corresponding to each time step, combined with the preset risk factor weights and score calculation standards, the elevator's operating status score for the current time step is calculated to quantify the risk level of the elevator failure in the current state:

[0092]

[0093] Among them, R t is the operating status score at time step t, C is the total number of fault types, w i is the preset risk factor weight of fault type i, P t,i is the predicted probability of fault type i at time step t, α is the coefficient for regulating the fault trend weight, p t represents the maximum predicted probability of fault type, β t is the time-dependent regulatory factor, γ t,i is the explanatory strength of the principal component features for fault type i at time step t, where γ t,i The generation includes:

[0094] According to the direction of each principal component obtained, the global variance proportion of each principal component in the feature mapping space is counted as the structural explanatory weight of the principal component;

[0095] According to the projection value of each sample in the direction of the principal component, the response degree of the principal component feature corresponding to each time step is extracted;

[0096] The structural explanation weight of the principal component is weightedly combined with the response degree of each principal component feature at time step t to obtain the explanation response value of the principal component at each time step;

[0097] The explained response value is cross-calculated with the predicted probability of the fault type i at time step t to generate γ t,i, as the structural explanation strength of the principal component for the fault type;

[0098] The elevator's operating status score at the current time step combines the predicted probability of various fault types at the current moment, the explanatory strength of the principal component characteristics for different fault types, the preset risk factor weights, and the dynamic adjustment factor of future fault trends. This achieves a multi-dimensional and accurate assessment of the current fault risk of the elevator. It not only integrates multi-source data and model prediction information, but also fully considers the differences between fault types and the impact of temporal evolution. The risk score results can truly reflect the current health level of the elevator and potential safety hazards in the future, providing a quantitative basis for intelligent early warning and scientific maintenance of elevators.

[0099] The fault type prediction label, future fault probability and operating status score corresponding to each time step are used as the final prediction output results.

[0100] In this embodiment, the probability of future failure is compared with a preset risk threshold to determine whether the elevator currently has a potential failure risk. If the risk reaches a high level, a corresponding warning message is generated, including:

[0101] Set up multi-level risk assessment criteria to assign multiple risk levels to the probability of future failures. Each risk level corresponds to a specific probability threshold, and the probability thresholds are arranged from low to high.

[0102] Obtain the future failure probability corresponding to each time step, and compare each future failure probability value with the probability threshold of each risk level set;

[0103] For each time step, determine which threshold interval the future failure probability of the time step falls into, and determine the risk level to which the time step belongs accordingly;

[0104] Set the judgment criteria for high risk levels. When the risk level of a time step reaches or exceeds the judgment criteria, the time step is judged as a high-risk state for potential failure. An early warning prompt containing the time step sequence number, risk level, and predicted probability information is generated and submitted to the alarm management platform.

[0105] All time steps judged as high risk are summarized and output as a fault warning time series.

[0106] refer to Figure 2 , an elevator fault prediction system based on big data technology, including the following modules:

[0107] Data acquisition module, used to collect multi-source data of elevator operation;

[0108] Data preprocessing module, used to preprocess multi-source data and generate standardized data sets;

[0109] The kernel principal component analysis module is used to perform kernel function feature mapping and principal component analysis on the standardized data set to obtain the feature sequence after dimensionality reduction;

[0110] The time series modeling module is used to segment the feature sequence after dimensionality reduction, input it into the gated Transformer, filter key time series features, model long-term dependencies between features, and output the time series modeling results;

[0111] The fault prediction module is used to identify fault types and predict probabilities based on the time series modeling results, obtaining the elevator operation status score and the probability of future faults;

[0112] The risk assessment module compares the probability of future failures with the risk threshold to determine potential failure risks and generates warning information when the risk reaches a high level.

[0113] The information push module is used to send risk assessment and warning information to operation and maintenance personnel, and output maintenance suggestions based on historical maintenance records.

[0114] Example 1:

[0115] To verify the feasibility of this invention, the system, based on big data technology, was fully deployed in a commercial international plaza. The plaza has 15 passenger elevators and four freight elevators, with a high daily traffic volume. Over the past two years, sudden elevator failures and reactive maintenance have become increasingly common. The property management department frequently faces outages and complaints due to door system failures, car vibrations, and motor anomalies. These issues impact business operations and result in high maintenance costs and management pressure.

[0116] In order to effectively improve the safe operation and intelligent management level of elevators, the project team installed multiple types of intelligent sensors in the main control room of each elevator, covering door machine current, motor temperature rise, car vibration, door lock switch times, load changes, etc. The data collected by all sensors are uploaded to the cloud platform in real time through the Internet of Things terminal. The cloud system uses the multi-source data preprocessing and kernel principal component analysis modules of the present invention to clean, reduce noise and standardize the nearly one million high-dimensional heterogeneous data uploaded daily, and then uses the principal component analysis method of the fusion kernel function to automatically screen out the core features that reflect the risk of faults. After that, the feature sequence after dimensionality reduction is segmented and input into the improved gated Transformer. The system can effectively model the long-term dependency and time-series evolution process between features. After the model is judged by the neural network, it can give the probability distribution of the current main fault type of each elevator, the risk probability of failure in the next seven days, and give risk scores and maintenance recommendations based on the principal component interpretation strength and historical maintenance data.

[0117] In May 2025, the door machine current in Elevator No. 2 in Building 1 experienced abnormal fluctuations for several consecutive days, but passengers were not aware of it. The system automatically identified a door system risk score exceeding 0.7 and predicted a 38% probability of door system failure within the next three days. It generated an early warning and notified maintenance personnel to conduct a special inspection. The door machine pulleys were found to be severely worn, and their timely replacement prevented an actual outage. That same month, Elevator No. 5 in Building 3 experienced an abnormal car vibration principal component score. The system predicted a motor failure probability exceeding 40%, and recommended guide rail lubrication maintenance, successfully avoiding a motor overheating shutdown. Similar early warning and proactive maintenance cases occurred numerous times during the six-month pilot period, effectively reducing elevator failures and downtime.

[0118] Table 1 Comparison of key operation and maintenance indicators before and after the launch of the elevator intelligent fault prediction system

[0119]

[0120] Table 1 shows that key operational and maintenance indicators showed significant improvements before and after implementing the intelligent elevator fault prediction system of the present invention. The total number of elevators remained constant at 19, ensuring data comparability. Before the system went live, the average number of sudden failures per elevator per month was 0.82. After the system went live, this number dropped to 0.38, a decrease of over 50%. During the same period, the number of sudden failures for projects without the system was essentially unchanged at 0.79. This demonstrates the outstanding effectiveness of the system in preventing and detecting faults in advance.

[0121] The average single response time was shortened from 29 minutes to 13 minutes, more than doubling the response speed. The control group remained at 28 minutes during the same period, indicating that intelligent early warning has greatly improved the initiative and timeliness of maintenance responses. The average total downtime has also dropped from 5.4 hours per month to 2.1 hours, which not only reduces the safety risks and user inconvenience caused by elevator downtime, but also effectively reduces business losses. The system's early warning detection rate has increased most significantly, from 19% to 86%, while the proportion of non-online system projects during the same period was only 21%. This shows that the solution of this invention can proactively identify the vast majority of potential faults, gaining valuable maintenance windows for maintenance and management departments.

[0122] The repeat repair rate also dropped from 10.2% to 5.8%, reducing resource waste caused by misdiagnosis of faults and incomplete repairs. User satisfaction increased from 85% to 96%, compared to only 87% for the control group during the same period, demonstrating that the system not only improves elevator safety and reliability but also enhances users' overall experience and management evaluation. This invention organically combines multi-source data fusion, accurate risk warning, and proactive maintenance strategies, possessing significant practical value in improving elevator operational safety, reducing maintenance costs, and optimizing user service.

[0123] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting elevator failures based on big data technology, characterized in that: include: Collect multi-source data of elevators, pre-process the multi-source data, and obtain a standardized data set; The kernel function is used to perform feature mapping on the standardized data set, and nonlinear dimensionality reduction is performed through kernel principal component analysis to extract the principal component features and obtain the feature sequence after dimensionality reduction; The feature sequence after dimensionality reduction is segmented according to the set time window. Each segment of the feature sequence is input into the gated Transformer. The gating mechanism is used to filter key temporal features. The multi-head self-attention mechanism is combined to model the long-term dependencies between features and output the temporal modeling results. Based on the output of time series modeling results, fault type identification and probability prediction are carried out to obtain the elevator operation status score and the probability of future faults; Compare the probability of future failures with the preset risk threshold to determine whether the elevator currently has a potential failure risk. If the risk reaches a high level, a corresponding warning message will be generated; The judgment results and warning information are sent to the operation and maintenance personnel, and maintenance recommendations are formulated based on historical maintenance records.

2. The elevator fault prediction method based on big data technology according to claim 1 is characterized in that: The multi-source data specifically includes real-time operating parameters collected by elevator sensors, environmental monitoring data, equipment maintenance records, and basic elevator attribute data.

3. The elevator fault prediction method based on big data technology according to claim 1 is characterized in that: The preprocessing of multi-source data specifically includes time synchronization, outlier removal, missing value filling and normalization.

4. The elevator fault prediction method based on big data technology according to claim 1, characterized in that: The method uses a kernel function to perform feature mapping on the standardized data set, performs nonlinear dimensionality reduction through kernel principal component analysis, extracts principal component features, and obtains a feature sequence after dimensionality reduction, including: According to different types of features in the obtained standardized data set, a variety of kernel functions are selected, including Gaussian radial basis kernel, polynomial kernel, and Laplace kernel, and initial weight coefficients are set for each type of kernel function; Based on the dynamic measurement results of the correlation between each type of feature and fault prediction, an adaptive weight is assigned to each type of kernel function, and an adaptive adjustment coefficient for sample pairs is introduced to construct a fusion kernel function; Using the fusion kernel function, feature mapping is performed on each pair of samples in the standardized data set. The fusion kernel function values ​​between all sample pairs are calculated in sequence. All fusion kernel function values ​​are filled into a square matrix according to the order of sample arrangement in the data set to form a weighted fusion kernel matrix. The mean of each row and column of the weighted fusion kernel matrix is ​​calculated respectively, and the overall mean is corrected. Each item of the weighted fusion kernel matrix is ​​centered with zero mean to obtain a centralized fusion kernel matrix. In the feature extraction process, the local relationship between samples is combined to construct a k-nearest neighbor weight matrix for each sample, and the local preservation constraint is introduced into the principal component optimization objective. The principal component optimization objective is to maximize the global variance of the principal component while minimizing the distance difference between the low-dimensional representations of the k-nearest neighbors. According to the set principal component optimization goal, the centralized fusion kernel matrix is ​​decomposed into eigenvalues ​​to obtain the principal component direction with both global structure and local preservation capabilities; Select the first m principal component directions whose cumulative contribution rate is greater than the set threshold. For each sample in the standardized data set, the feature vector obtained by fusion kernel function calculation is projected on each principal component direction in turn, and the projection values ​​on all principal component directions are recorded. The projection values ​​of all principal component directions are combined into the reduced dimension feature vector of the corresponding sample. The reduced dimension feature vectors of all samples are collected to form a reduced dimension feature sequence. In actual application, the distribution drift of newly collected data and historical data in the fusion kernel space is monitored in real time. When the drift metric exceeds the preset threshold, the type, weight and local preservation parameters of the fusion kernel function are automatically adjusted to reconstruct the fusion kernel matrix and principal component direction.

5. The elevator fault prediction method based on big data technology according to claim 1 is characterized in that: The feature sequence after dimensionality reduction is segmented according to the set time window, and each segment of the feature sequence is input into the gated Transformer. The gating mechanism is used to screen key time series features, and the long-term dependency between features is modeled in combination with the multi-head self-attention mechanism. The time series modeling results are output, including: According to the pre-set time window, the feature sequence after dimensionality reduction is divided into multiple non-overlapping feature segmentation sequences, each of which contains the dimensionality reduction feature data of consecutive time steps; For each feature segmentation sequence, all input features are divided into several feature groups according to the source or statistical characteristics of the input features, and a set of independent gating parameters is set for each feature group; When the feature segmentation sequence is input into the gating mechanism, samples containing known fault state labels in the standardized data set are selected. The prediction results generated by each feature group are compared with the corresponding true fault state labels. The proportion of samples with consistent prediction results and true labels to the total number of labeled samples is used as the prediction accuracy of the feature group. Combined with the distribution characteristics and prediction accuracy of the feature group, the sharing relationship of the gating parameters between the feature groups is dynamically adjusted. Some gating parameters are shared between feature groups, and some gating parameters are dedicated to a single feature group. When calculating each gating weight, an activation function is dynamically selected from the three activation functions of Sigmoid, ReLU, and Swish based on the statistical distribution of the current input features and the input order or iteration rounds when adjusting the gating weight parameters using a standardized dataset. The selected activation function is used to perform a nonlinear transformation on the input features to obtain the gating weight at the current moment. For each time step and each feature in the feature segmentation sequence, the variance calculation method is used to evaluate the uncertainty of the input data. According to the uncertainty result, the gating weight is assigned, and the uncertainty lower limit threshold and the uncertainty upper limit threshold are set as the basis for adjusting the gating weight. When the uncertainty evaluation value is less than the uncertainty lower limit threshold, a gating weight close to 1 is assigned. When the uncertainty evaluation value is greater than the uncertainty upper limit threshold, a gating weight close to 0 is assigned. When the uncertainty evaluation value is between the upper and lower limit thresholds, the gating weight is calculated between 0 and 1 according to the linear interpolation method, and the gating weight distribution is dynamically adjusted. Using a gating mechanism that has undergone dynamic gating parameter sharing, adaptive activation function switching, and input uncertainty adjustment, all features of each feature segmentation sequence are gated to obtain a dynamically filtered feature sequence. The dynamically filtered feature sequence is input into the multi-head self-attention mechanism to analyze and capture the long-term dependencies between different feature groups and between different time steps, and obtain the feature sequence fused by the multi-head self-attention; The multi-head self-attention fusion feature sequence is sequentially input into the multi-layer gated Transformer, the temporal dependency features across time periods are recursively extracted, and a complete temporal modeling feature sequence is output as the final temporal modeling result.

6. The elevator fault prediction method based on big data technology according to claim 1, characterized in that: According to the output time series modeling results, fault type identification and probability prediction are carried out to obtain the elevator operation status score and the probability of future faults, including: The output time series modeling feature sequence is used as input data and divided into a feature set for classification tasks and a feature set for regression tasks; The feature set used for the classification task is input into a fully connected neural network structure consisting of multiple linear transformation layers and nonlinear activation layers. The fault type prediction probability distribution at each time step is calculated through forward propagation. The fault type corresponding to the maximum fault type prediction probability is selected from the fault type prediction probability distribution as the fault type prediction label for the current time step. The feature set used for the regression task is input into another fully connected neural network structure, which continuously performs multi-layer linear mapping and activation operations, outputs the numerical results corresponding to each time step, and normalizes the numerical results to the interval [0,1] as the probability of future fault occurrence; Based on the fault type prediction label and future fault probability corresponding to each time step, combined with the preset risk factor weights and scoring calculation standards, the elevator's operating status score for the current time step is calculated to quantify the risk level of the elevator's fault in the current state; The fault type prediction label, future fault probability and operating status score corresponding to each time step are used as the final prediction output results.

7. The elevator fault prediction method based on big data technology according to claim 1, characterized in that: The probability of future failure is compared with the preset risk threshold to determine whether the elevator currently has a potential failure risk. If the risk reaches a high level, corresponding warning information is generated, including: Set up multi-level risk assessment criteria to assign multiple risk levels to the probability of future failures. Each risk level corresponds to a specific probability threshold, and the probability thresholds are arranged from low to high. Obtain the future failure probability corresponding to each time step, and compare each future failure probability value with the probability threshold of each risk level set; For each time step, determine which threshold interval the future failure probability of the time step falls into, and determine the risk level to which the time step belongs accordingly; Set the judgment criteria for high risk levels. When the risk level of a time step reaches or exceeds the judgment criteria, the time step is judged as a high-risk state for potential failure. An early warning prompt containing the time step sequence number, risk level, and predicted probability information is generated and submitted to the alarm management platform. All time steps judged as high risk are summarized and output as a fault warning time series.

8. An elevator fault prediction system based on big data technology, executing the elevator fault prediction method based on big data technology according to any one of claims 1 to 7, characterized in that: Includes the following modules: Data acquisition module, used to collect multi-source data of elevator operation; Data preprocessing module, used to preprocess multi-source data and generate standardized data sets; The kernel principal component analysis module is used to perform kernel function feature mapping and principal component analysis on the standardized data set to obtain the feature sequence after dimensionality reduction; The time series modeling module is used to segment the feature sequence after dimensionality reduction, input it into the gated Transformer, filter key time series features, model long-term dependencies between features, and output the time series modeling results; The fault prediction module is used to identify fault types and predict probabilities based on the time series modeling results, obtaining the elevator operation status score and the probability of future faults; The risk assessment module compares the probability of future failures with the risk threshold to determine potential failure risks and generates warning information when the risk reaches a high level. The information push module is used to send risk assessment and warning information to operation and maintenance personnel, and output maintenance suggestions based on historical maintenance records.

Images