Self-evolution feedback type elevator steel wire rope fault prediction and life evaluation method

Abstract

The application belongs to the technical field of elevator safety detection, and provides a self-evolution feedback type elevator steel wire rope fault prediction and service life evaluation method, which utilizes multi-source heterogeneous data acquisition, deep learning feature fusion and feedback optimization mechanism with online self-evolution capability to accurately predict and evaluate the fault mode, health state and residual service life of the elevator traction steel wire rope. Through deep fusion of multi-source heterogeneous data and dynamic feature weighting of attention mechanism, early weak fault features of the elevator steel wire rope can be effectively captured, the identification accuracy of internal micro damage and progressive degradation can be significantly improved, and the risk of missed detection can be reduced; meanwhile, based on the self-evolution feedback closed loop of maintenance results, combined with model drift detection and hierarchical incremental updating strategy, the model can automatically adapt to individual differences of different elevators and changes in degradation rules caused by long-term aging, thereby solving the model drift problem of accuracy decline of static model over time from the root.

Description

Technical Field

[0001] This application belongs to the field of elevator safety inspection technology, specifically, it relates to a self-evolving feedback method for predicting elevator wire rope faults and assessing its lifespan. Background Technology

[0002] Elevators, as indispensable vertical transportation equipment in high-rise buildings, directly impact the safety of passengers' lives and property. The steel wire rope, as the core load-bearing component of the elevator traction system, plays a crucial role in bearing the car load and transmitting traction force; its health status is one of the core factors determining the safe operation of the elevator. Under the combined effects of long-term alternating loads, repeated bending fatigue, environmental corrosion, and mechanical wear, steel wire ropes are prone to progressive damage such as broken wires, wear, corrosion, and fatigue cracks. Failure to identify and assess the damage development trend in a timely manner can easily lead to safety accidents. Therefore, achieving accurate assessment of the health status of elevator steel wire ropes and predicting failure trends is a core prerequisite for improving elevator safety capabilities and promoting the upgrade of operation and maintenance models from scheduled maintenance to condition-based maintenance.

[0003] Currently, safety inspection of elevator wire ropes still mainly relies on regular manual inspections and single-mode non-destructive testing, which has several technical limitations: manual inspection is highly dependent on personnel experience, has low inspection efficiency, is highly subjective, and is difficult to identify internal minor damage and early fatigue degradation, and generally suffers from inspection lag; existing automated inspection solutions are mostly based on a single inspection principle, and real-time operation monitoring data is disconnected from historical maintenance records and environmental condition data, which cannot comprehensively characterize the multi-dimensional health status of the wire rope.

[0004] Existing predictive assessment model technologies, such as CN122126722A, "A method and device for detecting elevator cable defects," and CN122035676A, "A method and system for non-destructive testing of elevator wire ropes based on multimodal fusion," are mostly based on offline static training and fixed threshold judgment. They cannot adapt to individual differences such as the service conditions and usage intensity of different elevators, nor can they cope with the degradation law deviation caused by long-term aging of wire ropes, i.e., model drift. They lack a self-correction mechanism to feed back actual maintenance results to the model. The model accuracy gradually decreases over time, and the risk of false alarms and missed alarms is high, which cannot provide stable and reliable technical support for predictive maintenance. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a self-evolving feedback method for predicting elevator wire rope failures and assessing its lifespan.

[0006] In one approach, a self-evolving feedback-based method for predicting elevator wire rope failures and assessing its lifespan includes the following steps: S1. Deploy multi-source sensors in various parts of the elevator to collect the elevator's physical status, operating conditions, environmental parameters, and historical maintenance records in real time, forming a multi-source heterogeneous data stream; S2. Clean, align, extract features and label the multi-source heterogeneous data collected in S1; by eliminating noise interference, unifying the spatiotemporal benchmark and parsing unstructured text, the original multi-source heterogeneous data is transformed into a standardized tensor that can be recognized by the machine learning model, and a mapping relationship between sensor features and fault labels is established to output multimodal fusion data. S3. Extract feature indicators characterizing the wear state, fatigue damage, wire breakage risk and environmental aging of elevator wire ropes from the multimodal fusion data output by S2, and construct interpretable feature vectors by combining signal processing technology. S4. Construct and train a hybrid deep learning prediction model based on convolutional neural networks, long short-term memory networks and attention mechanisms, automatically extract deep features from multi-source data of steel wire ropes, capture their nonlinear degradation patterns, and output multi-dimensional steel wire rope health assessment index prediction data. S5. Based on the multi-dimensional wire rope health assessment index prediction data output by S4, combined with the wire rope safety specifications and real-time working conditions, a three-level dynamic early warning mechanism is constructed, and corresponding differentiated operation and maintenance decision instructions are generated. S6. By combining the early warning decision instructions generated by S5 with the actual execution results of maintenance personnel through confidence-weighted labeling, model drift detection, incremental learning retraining and federated learning strategies, the prediction model parameters are dynamically updated. S7. After completing the parameter update, the prediction model executes S4 and S5 again to update the prediction data and early warning mechanism, and waits for the next round of execution feedback to form a self-evolving feedback loop.

[0007] Furthermore, the elevator physical state data in step S1 includes vibration signals, tension and dynamic load, magnetic flux, and the three-phase current and voltage of the traction motor; the elevator operating condition data includes running direction and speed, floors and stops, number of starts and stops, acceleration and deceleration time, and load distribution; environmental parameters include real-time humidity and relative humidity; and historical maintenance file data includes maintenance records, replacement and scrapping records.

[0008] Furthermore, the logic for establishing the mapping relationship between sensor features and fault tags in step S2 is to take the time point of the fault maintenance record as the endpoint of the fault occurrence or confirmation, and trace back a feature window period; extract all sensor feature data collected in step S1 within this window period, and mark the data of this window period as the corresponding fault tags; the multimodal fusion data is formed by splicing batch size, time series length, time domain statistical features, frequency domain features, physical mechanism features and operating condition codes to form a fusion matrix.

[0009] Furthermore, in step S3, the wear characteristics of the elevator wire rope are characterized by tension fluctuation, vibration energy, and rope diameter change rate; fatigue damage characteristics are characterized by stress amplitude, stress cycle, and cumulative damage degree; wire breakage risk characteristics are identified by magnetic flux leakage and high-frequency shock waves; environmental aging characteristics include temperature and humidity aging factors and corrosion risk index.

[0010] Furthermore, in step S4, the multi-dimensional wire rope health assessment index prediction data includes health index, remaining life and failure probability; the comprehensive risk index of the wire rope is calculated based on the health index, remaining life and failure probability using a weighted fusion algorithm.

[0011] The output range of the health index is [0,1], where 1 represents brand new and 0 represents failure. The remaining life is determined by predicting the remaining time or number of operating cycles before the wire rope reaches the target failure threshold of the health index while maintaining the current working condition, and outputs a 95% confidence interval. The failure probability is the probability of the wire rope failing within the future target time window obtained by the model execution prediction.

[0012] Furthermore, in step S5, the three-level dynamic early warning mechanism includes the attention level, the maintenance recommendation level, and the emergency stop level.

[0013] The triggering conditions for a level-of-concern warning are a health index of [0.5, 0.8] and a remaining lifespan of more than 30 days, or a failure probability of less than 0.01 but showing an upward trend; the decision instruction is to strengthen monitoring, increase the data acquisition frequency, and initiate a special analysis of high-frequency vibration.

[0014] The recommended trigger conditions for maintenance-level early warnings are a health index of [0.2, 0.5], or a remaining lifespan of more than 7 days but less than 30 days, or a failure probability of [0.01, 0.05], or a confidence interval lower limit of less than 15 days; the decision instruction is preventive replacement or maintenance, and the system automatically generates a maintenance work order that includes the recommended replacement time window, estimated costs, and risk descriptions.

[0015] The emergency stop warning is triggered when the health index is less than 0.2, the remaining lifespan is less than 3 days, the failure probability is greater than 0.05, or a momentary wire breakage impact is detected. The decision command is emergency intervention, in which the system locks the elevator's operating permissions and automatically calls the rescue team.

[0016] Furthermore, in step S6, the actual execution results of the maintenance personnel are input as feedback data through the mobile terminal by entering maintenance work order data, on-site photos and sensor measurement data; the feedback data items include actual fault labels, timestamps and confidence level markers.

[0017] Using the timestamp of the maintenance time as the endpoint, a feature window period is traced back, and the original sensor data collected in step S1 and the feature vector calculated in step S3 within this window period are extracted as input to perform time backtracking; and a sample pair of the wire rope operating state before the failure and the actual failure state is constructed.

[0018] Furthermore, in step S6, the confidence weighted label is scored and weighted based on the confidence score. The confidence score determines the data integrity based on whether it contains complete vibration, tension, and magnetic flux data, determines the logical consistency based on whether the entered fault type matches the associated sensor characteristics, and determines the source authority based on whether it is entered by maintenance experts or verified by the system.

[0019] When the confidence score is greater than or equal to 0.8, it is considered a high-confidence sample, assigned a weight of 1.0, and directly added to the incremental training set as the main basis for the learning of the prediction model.

[0020] Samples with a confidence score greater than or equal to 0.4 and less than or equal to 0.8 are considered to have medium confidence, are assigned a weight of 0.6, and are added to the training set. However, their influence is reduced in the loss function calculation, and they are used as auxiliary learning data.

[0021] When the confidence score is less than 0.4, it is considered a low-confidence sample, assigned a weight of 0, temporarily excluded from training, marked as a difficult sample, and entered into the human expert review queue; after review and correction, if it is confirmed to be a misjudgment, it is converted into a high-confidence sample for retraining, and if it is confirmed to be an input error, it is discarded.

[0022] Furthermore, in step S6, the metrics for model drift detection include root mean square error, mean absolute percentage error, classification accuracy, and the sliding window trend of the error.

[0023] When the error is within the allowable range and there is no significant upward trend, the model performance is considered stable. Continuous monitoring is then carried out, but the model is not updated for the time being; only logs are recorded.

[0024] When the error shows a slow upward trend, it is determined that the model has slightly drifted, triggering incremental learning, freezing the low-level feature extraction parameters of the convolutional neural network and the long short-term memory network, and only fine-tuning the parameters of the attention mechanism layer and the output layer.

[0025] When the error suddenly increases or a fundamental change in the operating pattern is detected, it is determined that the model has drifted severely, triggering a full retraining to retrain the prediction model architecture using the accumulated historical full data.

[0026] The beneficial effects of this application are: This application effectively captures early, subtle fault characteristics of elevator wire ropes through deep fusion of multi-source heterogeneous data and dynamic feature weighting using an attention mechanism, significantly improving the accuracy of identifying minor internal damage and progressive degradation, and reducing the risk of missed detection. Simultaneously, based on a self-evolving feedback loop of maintenance results, combined with model drift detection and a hierarchical incremental update strategy, the model can automatically adapt to the individual differences of different elevators and the degradation patterns caused by long-term aging, fundamentally solving the model drift problem of decreased accuracy of static models over time. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a diagram showing the overall architecture and data flow of a self-evolutionary feedback elevator wire rope fault prediction and life assessment method in one embodiment of this application. Figure 2 This is a flowchart illustrating the multi-source data fusion and feature engineering process in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a hybrid deep learning prediction model based on CNN-LSTM-Attention in one embodiment of this application; Figure 4 This is a flowchart illustrating the online self-evolutionary feedback optimization mechanism in one embodiment of this application. Detailed Implementation

[0029] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application. Similarly, the following examples are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0031] In this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0032] To further illustrate the technical means and effects adopted by this application to achieve its technical objectives, the following detailed description of specific implementation methods based on this application is provided in conjunction with representative embodiments: In one embodiment, please refer to Figure 1 A self-evolving feedback-based method for predicting elevator wire rope failures and assessing its lifespan includes the following steps: Step S1: Multi-source heterogeneous data acquisition Multi-source sensors are deployed in various parts of the elevator to collect the elevator's physical state, operating conditions, environmental parameters, and historical maintenance records in real time. This constructs a comprehensive and high-precision elevator wire rope health status perception network, forming a multi-source heterogeneous data stream, which provides a complete data foundation for subsequent feature fusion and fault prediction.

[0033] Specifically, elevator physical condition data are direct characterization signals reflecting internal damage and external wear of the wire rope. They are acquired through high-frequency sampling sensors, and the data categories include vibration signals, tension and dynamic load, magnetic flux, and three-phase current and voltage of the traction motor.

[0034] The vibration signal acquisition includes low-frequency components (0.5Hz-100Hz) reflecting the overall stability and resonance during elevator start-up, acceleration, and deceleration; and high-frequency components (1kHz-20kHz) to capture transient impact signals caused by broken wires, strand fractures, or loose strands within the wire rope. Tension and dynamic load are the direct input sources for stress calculation and fatigue damage assessment in subsequent steps. The acquired data includes the initial tension when the elevator is stationary (static tension) and the real-time tension fluctuations caused by load changes, acceleration / deceleration, and uneven floor levels during elevator operation (dynamic tension fluctuations). Magnetic flux acquisition includes radial and axial magnetic leakage signals. When broken wires, corrosion, or wear reduce the cross-sectional area within the wire rope, magnetic field lines leak. Sensors capture changes in the intensity of the leaked magnetic field, recording the peak value, waveform distortion rate, and spatial distribution characteristics of the leakage magnetic signal. By analyzing the current harmonics and fluctuations during traction motor operation, abnormal running resistance and load conditions of the wire rope can be indirectly inferred.

[0035] In this step, elevator operating data includes direction and speed, floors and stops, number of starts and stops, acceleration and deceleration times, and load distribution. Data sources include the elevator mainboard, data transmission unit (DTU), and cloud center. The collected data corresponds to the elevator's upward / downward operating status, speed change curve, current floor, target floor, number of stops, total daily / monthly starts and stops, and real-time car load percentage. The acceleration and deceleration times reflect the smoothness of operation, and the load distribution can be obtained by inverse calculation using a weighing device or current.

[0036] In this step, environmental factors are key external variables affecting the aging and corrosion rate of wire rope materials. The main environmental parameters collected are real-time humidity and relative humidity, used to construct an environmental aging correction coefficient. High temperature and high humidity environments accelerate wire rope corrosion and grease failure. In addition, dust and corrosive gas data can be collected to assess chemical corrosion risk and assist in revising the fatigue life prediction model.

[0037] In this step, historical maintenance archive data, including unstructured or semi-structured data such as maintenance records, replacement and scrapping records, serves as a key source for constructing supervised learning labels. Maintenance record data primarily originates from unstructured documents such as scanned copies of paper work orders, electronic work order systems, and inspection reports from the maintenance unit. The recorded content includes qualitative descriptions of faults, quantitative data such as measured rope diameter, number of broken wires, wear depth, and lubrication dates, as well as corrective actions such as replacement, lubrication, adjustment, and monitoring, and timestamps for the specific date and time of each maintenance operation. Replacement and scrapping records include the wire rope replacement date, reason for scrapping, cumulative operating time, and cumulative number of operations.

[0038] In this step, data such as vibration, magnetic flux, and current are sampled at high frequencies above 50kHz using a triggered acquisition method. Full data is recorded when an abnormal impact is detected; under normal conditions, only statistical characteristics are recorded. Data such as tension, temperature, humidity, and operating conditions are sampled at low frequencies of 1Hz-10Hz and recorded continuously.

[0039] Step S2: Data Preprocessing and Multimodal Fusion Please see Figure 2 The multi-source heterogeneous data collected in step S1 is cleaned, aligned, feature extracted and labeled; by eliminating noise interference, unifying the spatiotemporal benchmark and parsing unstructured text, the original multi-source heterogeneous data is transformed into a standardized tensor that can be recognized by the machine learning model, and a mapping relationship between sensor features and fault labels is established to output multimodal fusion data.

[0040] In this step, the cleaning of multi-source heterogeneous data includes routine outlier removal, signal denoising, and missing value imputation. Timestamp alignment uses the system time of the elevator PLC controller or the time of the cloud server as a unified benchmark. High-frequency data such as vibration and magnetic flux are converted to a sampling rate consistent with low-frequency data such as operating conditions and environmental data through downsampling or statistical aggregation, which can be 1Hz or 10Hz, etc. Alternatively, low-frequency data can be upsampled and matched to the high-frequency time axis through hold-up or linear interpolation to construct a unified time series matrix. The continuous data stream is then divided into fixed-length time windows as input samples for the model. Furthermore, the multi-source heterogeneous data requires strict spatial correlation. By aligning the tension and vibration data of 4 to 8 steel wire ropes according to their rope numbers, a feature vector of multi-rope correlation is constructed. This allows for the identification of single-rope anomalies and multi-rope collaborative wear. The current operating state is also labeled with the corresponding sensor data segments to ensure that the model can distinguish between normal and abnormal vibrations under different operating conditions, achieving a mapping between elevator operating conditions and vibration states.

[0041] In this step, feature extraction and labeling of multi-source heterogeneous data involves transforming the unstructured historical maintenance data collected in step S1 into structured supervision labels required for machine learning, and establishing a correlation with its corresponding historical running data.

[0042] Specifically, unstructured texts such as maintenance work orders and inspection reports are input, and key entity information such as fault type, quantitative indicators, and handling results are extracted using a natural language processing model. Optical character recognition is performed on scanned copies of paper work orders to extract tabular data. In other words, by combining NLP technology and OCR recognition, a structured fault event record containing fault type, quantitative indicators, time, and handling results is output.

[0043] Furthermore, the logic for establishing the mapping relationship between sensor features and fault tags involves using the time point of the fault maintenance record as the endpoint of the fault occurrence or confirmation, and then tracing back a feature window period, the length of which depends on the speed of fault development. Then, all sensor feature data collected in step S1 within this window period are extracted, and the data in this window period is labeled with the corresponding fault tags. Simultaneously, qualitatively described fault tags also need to be quantified into numerical tags. For example, the qualitative description of wire corrosion is converted into a wear coefficient, the qualitative description of the number of broken wires is converted into a broken wire count, and the qualitative description of wire replacement is converted into failure flag 1, etc.

[0044] Specifically, before data fusion, key indicators such as stress and fatigue damage accumulation can be calculated based on physical mechanisms to enhance data interpretability. Stress is calculated in real-time using the tension and wire rope model parameters—effective cross-sectional area—collected in step S1. If historical tags include data indicating a reduction in rope diameter due to wear, the effective cross-sectional area is dynamically updated. Fatigue damage accumulation is calculated based on the calculated stress amplitude and cycle count, using Mine linear cumulative damage theory to calculate the current cumulative damage level, which is then added as a feature to the fusion matrix.

[0045] In this step, the fusion matrix of the multimodal fusion data has dimensions of [Batch_Size, Time_Steps, Features]. Here, Batch_Size represents the batch size, indicating the number of samples processed simultaneously in a single training or inference iteration; Time_Steps represents the length of the time series contained in each sample; Features include time-domain statistical features such as vibration RMS, peak value, kurtosis, mean tension, and stress amplitude; frequency-domain features such as FFT spectral energy distribution and wavelet packet energy entropy; physical mechanism features such as cumulative damage degree, environmental aging coefficient, and magnetic flux leakage peak; and operating condition codes for uplink, downlink, and stationary operating states. The fused feature matrix can also be Z-score normalized or Min-Max normalized to eliminate the influence of different physical dimensions and accelerate model convergence. The constructed fusion matrix is ​​used as the sample set and divided into training, validation, and test sets in a conventional ratio of 7:2:1.

[0046] Step S3: Construction of Wire Rope Fault Association Feature Engineering The multimodal fusion data output from step S2 is used to extract feature indicators that characterize the wear state, fatigue damage, wire breakage risk and environmental aging of elevator wire ropes. These indicators are then combined with signal processing techniques to construct interpretable feature vectors, providing high-quality input for subsequent deep learning models.

[0047] Specifically, the wear characteristics of elevator wire ropes are characterized by tension fluctuations, vibration energy, and the rate of change in rope diameter. As wire rope wear intensifies, the gap between strands increases, leading to uneven tension distribution and a significant increase in fluctuation amplitude during operation. Tension fluctuations are primarily reflected by calculating the variance of dynamic tension within a fixed time window. Surface wear and strand friction generate high-frequency noise at specific frequencies. Vibration energy is reflected by frequency domain decomposition of the vibration signal, calculating the ratio of the energy in the 1kHz-10kHz high-frequency band to the total energy across the entire frequency band; an increase in this ratio indicates intensified surface wear. The rate of change in rope diameter directly reflects the degree of loss in the wire rope's cross-sectional area and is a key indicator for assessing the decline in elevator load-bearing capacity. By combining a tension-elongation model, the equivalent stiffness change of the wire rope is inferred from the tension change, indirectly estimating the rate of reduction in rope diameter.

[0048] Specifically, fatigue damage characteristics are characterized based on stress amplitude, stress cycles, and cumulative damage. Based on the stress sequence calculated in step S2, the amplitude and mean stress of stress cycles can be extracted using the rainflow counting method. The amplitude of the stress cycle determines the fatigue crack propagation rate, while the mean stress affects the fatigue limit. Furthermore, by extracting all stress cycles ( Δσi , ni Based on the stress-life curve of the wire rope, the number of failure cycles at the corresponding stress amplitude can be found. Ni To calculate cumulative damage D , D The value changes from 0 to 1, visually quantifying the proportion of fatigue life consumed by the wire rope. Specifically, the number of effective stress cycles per unit time can be counted to reflect the working intensity of the wire rope; high-frequency start-stop cycles accelerate fatigue accumulation. Among these, the cumulative damage degree... .

[0049] Specifically, the risk characteristics of wire breakage are identified through magnetic flux leakage and high-frequency shock waves. Wire breakage in a steel wire rope leads to severe leakage of magnetic field lines, producing significant peak values ​​and waveform distortion. Simultaneously, the instantaneous breakage of a single wire generates a strong transient shock wave, and the surge in pulse count directly confirms the occurrence of wire breakage. The peak value and distortion rate of magnetic flux leakage can be obtained by extracting local maxima from the MFL magnetic flux leakage signal and calculating the correlation coefficient or mean square error between the current MFL magnetic flux leakage waveform and the standard new rope waveform. Simultaneously, envelope demodulation of the high-frequency vibration signal is performed, a threshold is set, and the number of shock pulses exceeding the threshold per unit time is counted to determine the high-frequency shock pulse count. Furthermore, the Shannon entropy of the wire breakage signal in its spatial distribution can be calculated by deploying a multi-sensor array. Low entropy indicates concentrated wire breakage at a single point, indicating high risk; high entropy may indicate uniformly distributed micro-wire breakage, with relatively lower risk.

[0050] Specifically, environmental aging characteristics include temperature and humidity aging factors and corrosion risk indices. Among these, the temperature and humidity aging factor is used to construct an aging acceleration coefficient based on the Arrhenius equation or empirical models. K env The calculated basic fatigue damage degree D Revised to The corrosion risk index is calculated by combining dust concentration, corrosive gas concentration, and humidity, and serves as a weighting coefficient for wear characteristics, increasing the weight of wear characteristics in corrosive environments.

[0051] Furthermore, Z-score standardization is applied to all features to eliminate the influence of dimensions, and PCA principal component analysis or LASSO regression is used to remove redundant features, retaining the principal components that contribute most to fault prediction. The final output is a multimodal fusion feature vector X. final It contains three sub-vectors: Xwear Wear feature sub-vector, X fatigue Fatigue feature vector, X break The feature vector of the broken wire, i.e., the vector structure, is as follows: Step S4: Construction and Training of Wire Rope Fault Prediction Model Please see Figure 3 We constructed and trained a hybrid deep learning prediction model based on convolutional neural network (CNN), long short-term memory network (LSTM), and attention mechanism (Attention). This model automatically extracts deep features from multi-source data of steel wire ropes, captures their nonlinear degradation patterns, and outputs multi-dimensional prediction data of steel wire rope health assessment indicators.

[0052] In this step, the model adopts a four-layer cascaded architecture of "feature extraction - temporal modeling - dynamic focusing - multi-task output", and sets up a CNN feature extraction module, an LSTM temporal modeling module, and an Attention mechanism module respectively. Using the multimodal fusion feature vector constructed in step S3 as input data, Gaussian noise or time warping is added to the input data during the training phase to enhance the model's robustness to noise and operating condition fluctuations.

[0053] In this step, the CNN feature extraction module extracts local frequency domain features and transient impact features from multi-source signals, such as instantaneous vibrations caused by wire breakage and local peaks in magnetic flux leakage. Utilizing the local receptive field of the CNN, it automatically identifies high-frequency shock waves in the vibration signal and leakage peaks in the magnetic flux signal, eliminating the need for manual design of frequency domain features. The LSTM time-series modeling module captures the long-term time dependencies of wire rope performance degradation, such as fatigue accumulation and gradual wear processes. LSTM can solve the gradient vanishing problem of traditional RNNs, effectively modeling the long-period nonlinear degradation trajectory of the wire rope from new rope to failure. The Attention mechanism module automatically identifies and weights the key time steps that contribute most to fault prediction, suppressing irrelevant noise time steps. The structure uses Query(Q), Key(K), Value(V) transformations to map the LSTM output into three vectors. The calculation formula for the attention score is: Based on the Attention mechanism, the model can automatically focus on the moment when the wire rope experiences a broken wire impact or a sudden stress change, while ignoring background noise during smooth operation.

[0054] In this step, the prediction model shares underlying features and outputs three key metrics: Health Index (HI), Remaining Lifetime (RUL), and Fault Probability (FPP), enabling multi-task learning and improving generalization ability. The multi-task output layer structure then includes a fully connected layer, a Health Index branch, a Remaining Lifetime branch, and a Fault Probability branch.

[0055] The fully connected layer uses 32 units, ReLU activation, and integrates context vectors. The output layer of the health index branch HI is a single unit, using the Sigmoid activation function, with an output range of [0,1], where 1 represents brand new and 0 represents failure. HI∈[0.8,1.0] is defined as a healthy state, HI∈[0.5,0.8] as a sub-healthy state, HI∈[0.2,0.5] as a warning state, and HI∈[0,0.2] as a dangerous state. Specifically, the threshold can be dynamically adjusted based on the wire rope model, service life, and environmental correction factor. The output layer of the remaining life branch RUL is a single unit, using the Linear activation function, with an output range of [0,+∞), in days or operating cycles. The remaining life is determined by predicting the remaining time or number of operating cycles before the wire rope reaches the target failure threshold of the health index while maintaining its current operating conditions, outputting a 95% confidence interval. The lower limit of the remaining life can also be used as a basis for conservative decision-making, ensuring early intervention even in extreme cases. The fault probability is the probability of the wire rope failing within the future target time window, obtained by the model's prediction. The output layer of the Fault Probability branch is a single unit with a Sigmoid activation function and an output range of [0,1].

[0056] In this step, the loss function for model training is designed using a weighted multi-task loss function to balance the optimization objectives of different tasks. The calculation formula is as follows: in, L HI For the health index, mean squared error loss is used to measure the deviation between the predicted HI and the actual HI. L RUL For the remaining lifetime, Huber Loss is adopted, which is insensitive to outliers and improves the robustness of RUL prediction; L Prob To determine the failure probability, a binary cross-entropy loss is used to optimize classification accuracy. λ 1 , λ 2 , λ 3 This is the task weighting coefficient, which is dynamically adjusted according to actual business needs.

[0057] The training process is as follows: The sample set constructed in step S2 is divided into a training set, a validation set, and a test set in a 7:2:1 ratio, and iterative training is performed; forward propagation is performed in the order of input features - CNN - LSTM - Attention - output; then the loss function and gradient are calculated, and the model parameters are updated. During the training process, the validation set loss is monitored. If it does not decrease for N consecutive rounds, training is stopped to prevent overfitting; the model weights with the lowest validation set loss are saved as the baseline model.

[0058] In particular, for newly installed elevators or data-scarce scenarios, a pre-trained general-purpose base model can be loaded, the underlying parameters of CNN and LSTM can be frozen, and only the Attention layer and output layer can be fine-tuned to quickly adapt using a small amount of local data.

[0059] In this step, the coefficient of determination R is used as the evaluation index for predicting the Health Index (HI). 2 The mean absolute error (MAE) and the root mean square error (RMSE) and early prediction penalty function (Score) were used to evaluate the remaining useful life (RUL) prediction. The AUC-ROC curve and the precision-recall curve were used to evaluate the failure probability. K-Fold cross-validation was used to ensure the model's generalization ability under different elevators and different operating conditions.

[0060] Step S5: Multi-level early warning and decision-making Based on the multi-dimensional wire rope health assessment index prediction data output in step S4, and combined with the wire rope safety specifications and real-time operating conditions, a three-level dynamic early warning mechanism is constructed, and corresponding differentiated operation and maintenance decision instructions are generated. While paying attention to the current health status of the wire rope, future risks are inferred based on the degradation trend, ensuring that sufficient intervention window is provided before the failure occurs.

[0061] In this step, the health index HI, remaining lifespan RUL, and failure probability P are based on step S4. fault The comprehensive risk index is calculated using a weighted fusion algorithm. CRI The calculation formula is: in, ω 1, ω 2, ω 3 represents the weighting coefficient. T max This is the maximum permissible maintenance cycle. CRI The higher the value, the greater the risk, according to CRI Based on the value and the development trend of various indicators, the system automatically triggers three levels of dynamic early warning: attention level, maintenance recommendation level, and emergency stop level, and pushes the warning content and decision instructions to the corresponding terminals.

[0062] Specifically, the trigger conditions for a Level 1 warning are a health index HI ∈ [0.5, 0.8] and a remaining lifespan greater than 30 days, or a failure probability less than 0.01 but showing an upward trend, or a significant increase in the environmental aging coefficient. The corresponding warning message is a notification that the wire rope is in a sub-healthy state, with accelerated wear / fatigue accumulation; the warning recommends adding a focused inspection of the corresponding items during the next routine maintenance, without interrupting elevator operation. The decision instruction is to strengthen monitoring, increase data collection frequency, and initiate a special analysis of high-frequency vibration.

[0063] Specifically, the recommended trigger conditions for maintenance-level early warnings are a health index HI ∈ [0.2, 0.5], remaining life RUL ∈ [7, 30] days, failure probability in the range of [0.01, 0.05], or a confidence interval lower limit of less than 15 days. The corresponding early warning content is a notification that the wire rope has a limited remaining life and a high risk of failure; the warning recommends immediate preventative maintenance, including replacing the wire rope, performing deep lubrication, and adjusting tension. The decision instruction is preventative replacement or repair, and the system automatically generates a maintenance work order containing the recommended replacement time window, estimated costs, and risk descriptions.

[0064] Specifically, the trigger conditions for an emergency elevator stop warning are: a health index (HI) < 0.2, remaining lifespan less than 3 days, failure probability greater than 0.05, or detection of instantaneous wire rope breakage impact characteristics. The corresponding warning message is a notification that the wire rope is about to fail or has suffered severe damage, indicating a risk of rope breakage; the warning recommends immediate elevator stoppage, passenger evacuation, and shaft closure. The decision command is emergency intervention: the system locks elevator operation permissions and automatically calls for rescue teams.

[0065] Furthermore, upon triggering any level of warning, the system automatically generates a structured decision support report. Utilizing the weight distribution of the Atteneion mechanism, it identifies the associated time window and characteristics leading to the warning. Based on different fault types such as wear, fatigue, and wire breakage, it recommends specific solutions: for wear, check lubrication or suggest replacement; for fatigue, check the traction sheave groove or suggest replacement; for wire breakage, check rope end fixing or replace immediately. The decision support report is output in the form of an electronic work order, containing items such as elevator number, floor location, warning level, trigger time, current predicted values ​​and confidence intervals for HI and RUL, recommended measures, estimated time, required spare parts, and historical degradation curves.

[0066] Step S6: Online self-evolutionary feedback optimization Please see Figure 4 By combining the early warning decision instructions generated by S5 with the actual execution results of maintenance personnel through confidence-weighted labeling, model drift detection, incremental learning retraining, and federated learning strategies, the prediction model parameters are dynamically updated.

[0067] Specifically, after the maintenance work is completed, the actual performance results of the maintenance personnel are input as feedback data via mobile devices, including maintenance work order data, on-site photos, and sensor measurement data. Feedback data items include actual fault labels, timestamps, and confidence level markers. The system automatically compares the predicted remaining life (RUL) with the actual replacement date and time to calculate the prediction error. Simultaneously, using the maintenance time timestamp as the endpoint, it traces back a feature window period, extracting the raw sensor data collected in step S1 and the feature vector calculated in step S3 within that window period as input for time backtracking. It also constructs sample pairs of the wire rope's operating state before the fault and the actual fault state. Abnormal samples with incorrect timestamps, severe data loss, or logical conflicts are cleaned and removed.

[0068] In this step, since the reliability of the feedback data needs to be verified, the system scores the confidence level of the data and assigns dynamic weights based on the data quality to prevent low-quality data from polluting the model.

[0069] Specifically, the confidence score ranges from [0,1]. Data integrity is determined by whether it contains complete vibration, tension, and magnetic flux data. Logical consistency is determined by whether the entered fault type matches the characteristics of the associated sensor. The authority of the source is determined by whether it was entered by maintenance experts or verified by the system.

[0070] When the confidence score is greater than or equal to 0.8, it is considered a high-confidence sample, assigned a weight of 1.0, and directly added to the incremental training set as the main basis for the learning of the prediction model.

[0071] Samples with a confidence score greater than or equal to 0.4 and less than or equal to 0.8 are considered to have medium confidence, are assigned a weight of 0.6, and are added to the training set. However, their influence is reduced in the loss function calculation, and they are used as auxiliary learning data.

[0072] When the confidence score is less than 0.4, it is considered a low-confidence sample, assigned a weight of 0, temporarily excluded from training, marked as a difficult sample, and entered into the human expert review queue; after review and correction, if it is confirmed to be a misjudgment, it is converted into a high-confidence sample for retraining, and if it is confirmed to be an input error, it is discarded.

[0073] In this step, the metrics for model drift detection include the root mean square error (RMSE) which measures the deviation of the remaining lifetime (RUL) prediction, the mean absolute percentage error which measures the relative error, the classification accuracy which measures the fault type identification rate, and the sliding window trend of the error.

[0074] When the error is within the allowable range and there is no significant upward trend, the model performance is considered stable. Continuous monitoring is then carried out, but the model is not updated for the time being; only logs are recorded.

[0075] When the error shows a slow upward trend, it is determined that the model has slightly drifted, triggering incremental learning, freezing the low-level feature extraction parameters of the convolutional neural network and the long short-term memory network, and only fine-tuning the parameters of the attention mechanism layer and the output layer.

[0076] When the error suddenly increases or a fundamental change in the operating pattern is detected, it is determined that the model has drifted severely, triggering a full retraining to retrain the prediction model architecture using the accumulated historical full data.

[0077] In this step, during federated learning, each elevator node or edge server trains the model locally using local feedback data. The original data remains within the domain; only the gradient updates of the model parameters are processed with differential privacy or homomorphic encryption before being uploaded to the cloud. The cloud-based federated aggregator receives the encrypted gradients from each node, calculates the weighted average, and updates the global base model. It then distributes the updated global model parameters back to each node. This process optimizes the model using online data while strictly protecting the data privacy of each elevator operator.

[0078] Specifically, for the individual differences of a single elevator, transfer learning can be used to generate a customized model. That is, a general base model is maintained in the cloud to understand the common degradation patterns of elevator products in the market, and a special feedback sample set is collected for the target elevator. The general base model is then fine-tuned to generate a personalized predictive sub-model. This model can accurately reflect the elevator's unique wear rate, environmental sensitivity, and operating habits.

[0079] Step S7, Closed-loop verification and iteration The updated prediction model is first validated on an independent test set. If the performance is satisfactory, it is deployed to the public; if the performance degrades, it is rolled back to the previous version. After the performance validation is completed and the updated model is confirmed for deployment, it is put back into steps S4 and S5 to continue updating the prediction data and the early warning mechanism, waiting for feedback from the next round of execution, thus forming a closed loop of self-evolutionary feedback.

[0080] In summary, this application constructs a closed-loop system for elevator wire rope health assessment, encompassing "perception-prediction-feedback-evolution." Through deep fusion of multi-source heterogeneous data and dynamic feature weighting using an attention mechanism, it effectively captures early, subtle fault characteristics, significantly improving the accuracy of identifying minor internal damage and progressive degradation, and reducing the risk of missed detections. Simultaneously, based on a self-evolutionary feedback loop of maintenance results, combined with model drift detection and a hierarchical incremental update strategy, the model can automatically adapt to individual differences in different elevators and the degradation patterns caused by long-term aging, fundamentally solving the model drift problem of decreased accuracy over time in static models. A multi-level dynamic early warning mechanism, combined with degradation trend projection, can identify fault risks in advance and provide sufficient intervention windows, effectively reducing ineffective maintenance and excessive component replacement. This ensures operational safety while reducing maintenance costs, providing reliable technical support for condition-based maintenance of elevator wire ropes.

[0081] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A self-evolving feedback-based method for predicting elevator wire rope failures and assessing its lifespan, characterized in that... The method includes the following steps: S1. Deploy multi-source sensors in various parts of the elevator to collect the elevator's physical status, operating conditions, environmental parameters, and historical maintenance records in real time, forming a multi-source heterogeneous data stream; S2. Clean, align, extract features and label the multi-source heterogeneous data collected in S1; by eliminating noise interference, unifying the spatiotemporal benchmark and parsing unstructured text, the original multi-source heterogeneous data is transformed into a standardized tensor that can be recognized by the machine learning model, and a mapping relationship between sensor features and fault labels is established to output multimodal fusion data. S3. Extract feature indicators characterizing the wear state, fatigue damage, wire breakage risk and environmental aging of elevator wire ropes from the multimodal fusion data output by S2, and construct interpretable feature vectors by combining signal processing technology. S4. Construct and train a hybrid deep learning prediction model based on convolutional neural networks, long short-term memory networks and attention mechanisms, automatically extract deep features from multi-source data of steel wire ropes, capture their nonlinear degradation patterns, and output multi-dimensional steel wire rope health assessment index prediction data. S5. Based on the multi-dimensional wire rope health assessment index prediction data output by S4, combined with the wire rope safety specifications and real-time working conditions, a three-level dynamic early warning mechanism is constructed, and corresponding differentiated operation and maintenance decision instructions are generated. S6. By combining the early warning decision instructions generated by S5 with the actual execution results of maintenance personnel through confidence-weighted labeling, model drift detection, incremental learning retraining and federated learning strategies, the prediction model parameters are dynamically updated. S7. After completing the parameter update, the prediction model executes S4 and S5 again to update the prediction data and early warning mechanism, and waits for the next round of execution feedback to form a self-evolving feedback loop.

2. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 1, characterized in that, The elevator physical status data in step S1 includes vibration signals, tension and dynamic load, magnetic flux, and three-phase current and voltage of the traction motor; the elevator operating condition data includes running direction and speed, floors and stops, number of starts and stops, acceleration and deceleration time, and load distribution; environmental parameters include real-time humidity and relative humidity; and historical maintenance file data includes maintenance records, replacement and scrapping records.

3. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 2, characterized in that, The logic for establishing the mapping relationship between sensor features and fault tags in step S2 is to take the time point of the fault maintenance record as the endpoint of the fault occurrence or confirmation, and trace back a feature window period. By extracting all sensor feature data collected in step S1 within the window period, and marking the data of the window period as the corresponding fault label; Multimodal fusion data is formed by splicing batch size, time series length, time domain statistical features, frequency domain features, physical mechanism features, and operating condition codes to form a fusion matrix.

4. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 1, characterized in that, In step S3, the wear characteristics of the elevator wire rope are characterized by tension fluctuation, vibration energy, and rope diameter change rate; fatigue damage characteristics are characterized by stress amplitude, stress cycle, and cumulative damage degree; wire breakage risk characteristics are identified by magnetic flux leakage and high-frequency shock waves; environmental aging characteristics include temperature and humidity aging factors and corrosion risk index.

5. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 4, characterized in that, In step S4, the multi-dimensional wire rope health assessment index prediction data includes health index, remaining life and failure probability; the comprehensive risk index of the wire rope is calculated based on the health index, remaining life and failure probability using a weighted fusion algorithm. The output range of the health index is [0,1], where 1 represents brand new and 0 represents failure. The remaining life is determined by predicting the remaining time or number of operating cycles before the wire rope reaches the target failure threshold of the health index while maintaining the current working condition, and outputs a 95% confidence interval. The failure probability is the probability of the wire rope failing within the future target time window obtained by the model execution prediction.

6. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 5, characterized in that, In step S5, the three-level dynamic early warning mechanism includes the attention level, the maintenance recommendation level, and the emergency stop level; The trigger conditions for a warning at the attention level are a health index of [0.5, 0.8] and a remaining lifespan of more than 30 days, or a failure probability of less than 0.01 but showing an upward trend; The decision-making directive was to strengthen monitoring, increase the frequency of data acquisition, and initiate a special analysis of high-frequency vibration. The recommended trigger conditions for maintenance-level early warnings are a health index of [0.2, 0.5], or a remaining lifespan of more than 7 days but less than 30 days, or a failure probability of [0.01, 0.05], or a confidence interval lower limit of less than 15 days; the decision instruction is preventive replacement or maintenance, and the system automatically generates a maintenance work order that includes the recommended replacement time window, estimated costs, and risk descriptions. The triggering conditions for emergency stop level warning are: health index less than 0.2, remaining life less than 3 days, failure probability greater than 0.05, or detection of instantaneous wire breakage impact characteristics. The decision command is for emergency intervention. The system locks the elevator's operating permissions and automatically calls for rescue teams.

7. The self-evolving feedback method for predicting elevator wire rope failures and assessing lifespan according to any one of claims 1 to 6, characterized in that, In step S6, the actual execution results of the maintenance personnel are input as feedback data through the mobile terminal, including maintenance work order data, on-site photos, and sensor measurement data. The feedback data items include actual fault labels, timestamps, and confidence level markers. Using the timestamp of the maintenance time as the endpoint, a feature window period is traced back, and the original sensor data collected in step S1 and the feature vector calculated in step S3 within this window period are extracted as input to perform time backtracking; and a sample pair of the wire rope operating state before the failure and the actual failure state is constructed.

8. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 7, characterized in that, In step S6, the confidence weighted label is scored and weighted based on the confidence score. The confidence score determines the data integrity based on whether it contains complete vibration, tension, and magnetic flux data, determines the logical consistency based on whether the entered fault type matches the associated sensor characteristics, and determines the source authority based on whether it is entered by maintenance experts or verified by the system. When the confidence score is greater than or equal to 0.8, it is considered a high-confidence sample, assigned a weight of 1.0, and directly added to the incremental training set as the main basis for the learning of the prediction model; When the confidence score is greater than or equal to 0.4 and less than or equal to 0.8, it is considered a medium confidence sample, assigned a weight of 0.6, added to the training set, but its influence is reduced in the loss function calculation, and it is used as auxiliary learning data; When the confidence score is less than 0.4, it is considered a low-confidence sample, assigned a weight of 0, temporarily excluded from training, marked as a difficult sample, and entered into the human expert review queue; after review and correction, if it is confirmed to be a misjudgment, it is converted into a high-confidence sample for retraining, and if it is confirmed to be an input error, it is discarded.

9. The self-evolving feedback elevator wire rope fault prediction and life assessment method according to claim 7, characterized in that, In step S6, the metrics for model drift detection include root mean square error, mean absolute percentage error, classification accuracy, and the sliding window trend of the error. When the error is within the allowable range and there is no significant upward trend, the model performance is considered stable. Continuous monitoring is carried out, but the model is not updated for the time being; only logs are recorded. When the error shows a slow upward trend, it is determined that the model has slightly drifted, triggering incremental learning, freezing the low-level feature extraction parameters of the convolutional neural network and the long short-term memory network, and only fine-tuning the parameters of the attention mechanism layer and the output layer. When the error suddenly increases or a fundamental change in the operating pattern is detected, it is determined that the model has drifted severely, triggering a full retraining to retrain the prediction model architecture using the accumulated historical full data.

Images