On-Demand Maintenance Method Using Fault Prediction to Reduce Elevator Entrapment

Abstract

With the rapid growth of elevator installations, conventional scheduled maintenance struggles to meet the dual demands of ensuring operational safety and cost control. This study proposes an innovative on-demand maintenance method that aligns with the Chinese policy directives on elevator maintenance reform. First, we conduct a historical fault cause analysis to identify the root causes of elevator entrapment incidents. Next, we establish an entrapment prediction model based on our historical data. Then, we design an elevator entrapment risk index report according to the prediction results. Finally, we formulate an on-demand maintenance plan that combines insights from the report with the conclusions of the cause analysis. Field implementation and comparative experiments demonstrate that the proposed on-demand maintenance method outperforms the scheduled one. The result shows significant reductions in accident and maintenance workload, justifying the practical value of this approach for the industry.

1. Introduction

According to the “Administration for Market Regulation Announcement on National Special Equipment Safety Status 2024,” the Chinese elevator inventory reached 11.5324 million units by the end of 2024. Elevators play a critical role in the eight major categories of special equipment [1]. However, the rapid expansion of the elevator industry has also led to serious safety challenges [2]. The maintenance methods and schedules of elevators directly impact the safety, stability, and reliability of operation [3,4,5]. The scheduled maintenance approach is inefficient, inconsistent, underqualified, and delayed and therefore significantly impairs elevator safety [6,7]. In 2024, there were 132 incidents involving special equipment nationwide, including 41 elevator accidents that accounted for 31.06% of all notable equipment accidents, resulting in 27 deaths [1]. Although the accident rate of 0.001% may seem statistically insignificant, the number of fatalities remains alarmingly high.

To mitigate the safety issues and stabilize elevator operation while reducing the accident rate, the Chinese government has initiated maintenance regime reforms, promoting the transition from scheduled to on-demand maintenance. Currently, implementing on-demand maintenance in China remains in the pilot exploration phase. Multiple regions have adopted methods such as “Self-reporting with credit-based supervision” systems and “IoT-enabled predictive maintenance” platforms, integrated with elevator liability insurance schemes, to facilitate with the implementation of on-demand maintenance. Although these measures have enhanced maintenance efficiency, certain operational limitations still persist [8,9].

With the iterative advancement of industrial technology, equipment maintenance methods have developed through three distinct phases: breakdown-based reactive maintenance, scheduled preventive maintenance, and predictive maintenance (PdM) [10]. The on-demand maintenance currently utilized for elevators in China essentially represents an application of predictive maintenance theory within the elevator domain [11,12]. PdM has emerged as the predominant operational maintenance methodology, providing significant advantages, including enhanced equipment reliability, improved operational status and performance, and reduced maintenance costs. This approach has been widely implemented across multiple industrial fields, including mechanical engineering, aerospace, nuclear power, maritime, and wind energy. The key to predictive maintenance lies in accurately forecasting the future state of equipment and promptly implementing maintenance measures under different conditions [13]. Maintenance technology demonstrates a clear trend of formulating decisions based on multivariable condition monitoring and fault prediction [14]. Against the backdrop of rapid advancements in information technology, machine learning and deep learning algorithms have been widely applied and continuously developed in the fields of fault diagnosis and fault prediction. For example, Mitici et al. [15] integrated probabilistic Remaining Useful Life (RUL) prediction into maintenance scheduling using Convolutional Neural Networks (CNNs), providing optimal timing for component replacement. Nguyen et al. [16] developed a novel dynamic PdM framework utilizing Long Short-Term Memory (LSTM) networks to predict defects. The authors thoroughly discussed the advantages of PdM over other maintenance strategies and presented a comprehensive framework spanning data-driven prediction to maintenance decision-making. Chen et al. [2] employed the Decision-Analytic Network Process to clarify the connections among various indicators, calculate their weights within the system, and derive a comprehensive evaluation score rating elevator health. The assessment model was applied to practical experiments to evaluate elevators and verify their effectiveness. Park et al. [17] proposed a risk-based inspection method, a systematic decision-making technique to identify components prone to fault and assess their potential consequences. This approach was introduced into elevator maintenance because it balances cost-effectiveness and safety well. Zhang et al. [18] analyzed historical malfunction data and identified the top four causes: controller faults, car door issues, landing door defects, and faults of other car interior equipment. The study demonstrated that risk-based controller maintenance could extend elevator lifespan. Feng et al. [19] established a defect risk factor index system for elevator drum brakes, analyzing hierarchical structures and intrinsic relationships among factors to identify critical maintenance points and reduce fault risks efficiently. Lin et al. [20] developed an elevator condition diagnosis model based on a simple vibration zone segmentation method. Liao et al. [21] validated the proposed fusion model during elevator installation and maintenance management. The model utilized a dataset of 110,698 safety inspection records from 25,729 construction sites (including elevator installation and maintenance sites) managed by an elevator enterprise to construct a hazard correlation network. This network enabled the identification of critical on-site hazards for proactive construction management. Niu et al. [22] introduced an intelligent defect diagnosis method for elevator traction induction motors based on a decision fusion system. Wu et al. [23] proposed a practical solution for defect diagnosis in smart building elevators, enabling maintenance personnel to detect malfunctions based on objective evidence rather than intuition. Koutsoupakis et al. [24] developed a novel damage identification method for elevator systems, which combines vibration data obtained from physical measurements and high-fidelity multibody dynamics models with deep learning algorithms. Pan et al. [25] presented a GNN-LSTM-BDANN deep learning model for variations in operating environments to ensure sound signal acquisition. The model leveraged historical acoustic data from other elevators to predict the remaining lifespan of target elevator door systems. To prevent mechanic accidents, Yao et al. [9] explored an IoT-based elevator fault monitoring system using the Relief-F algorithm to assess potential influencing factors.

The studies above compared different methods and improved predictive methods based on fault prediction research to obtain optimal results. However, applying predictive outcomes to subsequent maintenance operations remains unstudied. Some references focus only on specific elevator components or certain critical parts, which cannot effectively support the implementation of on-demand maintenance.

To mitigate the above issues, this paper proposes an elevator on-demand maintenance method based on the prediction of entrapment faults. First, using City H’s 2023 elevator fault records, the study identified the main causes of entrapment problems. Second, a hybrid model for elevator fault prediction was developed based on data from City H’s Elevator Emergency Response Center, enabling the forecasting of elevator entrapment events. Finally, an on-demand maintenance plan was implemented by integrating the predicted results with fault analysis findings. Comparative experiments with scheduled maintenance methods demonstrated that this approach effectively reduces the occurrence of entrapment incidents while decreasing maintenance workload.

The contributions of this paper are summarized as follows:

(1) Analysis of the 2023 historical fault data revealed that door system faults, human-related faults, and external factors are the three main causes of entrapment incidents.

(2) By applying the prediction model for entrapment events, a risk index checklist was designed, including predicted entrapment values, risk levels (red/green codes), and basic elevator information, which achieves multidimensional data visualization.

(3) According to the risk index checklist and historical root cause analysis, an on-demand maintenance plan was composed and examined in City H through field experiments. The results demonstrate that this solution achieves dual optimization compared with the scheduled method: entrapment incidents were reduced, and the maintenance workload decreased from 382 units to 54 units, representing an 85.86% reduction.

The subsequent sections of the paper are structured in the following manner: Section 2 provides a brief overview of the historical fault cause analysis; Section 3 introduces the data and models for entrapment fault prediction and the design of an entrapment risk index checklist; Section 4 presents the design and implementation of the on-demand maintenance plan; Section 5 compares and analyzes the results between the applied and the scheduled maintenance methods; and Section 6 presents the conclusion.

2. Fault Cause Analysis

By the end of 2023, the total number of operational elevators in City H had reached 211,432, meaning a 10.81% year-on-year increase, and this ranks as the city second among all provincial capitals in China in elevator number. Among these elevators, 46,000 had been working for over 10 years, of which 12,000 had exceeded 15 years of operation.

The historical elevator fault data was collected from the records of City H’s Elevator Emergency Response Center (96333). In 2023 alone, the city’s elevator emergency response platform received 30,865 calls related to elevator malfunctions and entrapment rescues and conducted 19,930 emergency interventions. The completed entrapment incident data were subsequently analyzed, and the overview is shown in Figure 1 and

Table 1. General overview of historical faults.

Fault Category	Cause of Fault	Number of Faults	Percentage of Faults
Human causes	…	4553	22.84%
External causes	…	2417	12.13%
Equipment causes	Door systems	5250	26.34%
	Traction System	653	3.28%
	Guiding and Weight Balancing System	557	2.79%
	Car System	681	3.42%
	Electrical Control System	2108	10.58
	Safety Protection System	1991	9.99%
	Electric Drive System	1720	8.63%

.

According to the statistical data of fault categories, the three leading causes of elevator stranding faults are door system faults (26.34%), human-caused faults (22.84%), and externally caused faults (12.13%).

Among all recorded faults, door system malfunctions accounted for 5250 incidents, representing the leading cause of elevator entrapments. These faults primarily resulted from equipment aging, improper operation, and inadequate maintenance. The specific fault manifestations are shown in

Table 2. Overview of door system faults.

Fault Category	Cause of Fault	Number of Faults	Percentage of Faults
Door systems	Car door lock fault/Landing door lock fault	1235	6.20%
	Door operator malfunction	870	4.37%
	Non-compliant clearance between the door knife and roller	421	2.11%
	Safety edges/light curtains protection device fault	166	0.83%
	Interlock fault between the active and passive door leaves	174	0.87%
	Door contact switch fault	544	2.73%
	Door hanger wheel damage	303	1.52%
	Door guiding system fault	155	0.78%
	Others	1382	6.93%

.

As shown in

Table 3. Overview of human-caused faults.

Fault Category	Cause of Fault	Number of Faults	Percentage of Faults
Human causes	The elevator stopped working because the door opening and closing was blocked by garbage.	2345	11.77%
	The door was deformed due to brutal handling by human, and the elevator could not operate.	466	2.34%
	Overloading	239	1.20%
	The elevator cannot operate because the door closing time is delayed for too long.	406	2.04%
	Others	1097	5.50%

, a total of 4553 human-caused faults were mainly triggered by uncivilized elevator behaviors, among which the more common phenomena included blocking the doors with domestic or renovative garbage, deforming the doors due to barbaric handling of heavy loads, delaying the door from closing, etc.

Externally induced entrapments totaled 2417 cases, primarily triggered by environmental factors including power outages, extreme summer heat, and typhoon-associated rainstorms. Specific fault manifestations included sudden pauses of elevator operation, short-circuiting of electrical components due to water ingression, the triggering of protection mechanisms by the elevator control system, etc., as shown in

Table 4. Overview of external causes of fault.

Fault Category	Cause of Fault	Number of Faults	Percentage of Faults
External causes	Power fault causing the elevator to stop running	494	2.48%
	Short circuit due to water ingress in electrical components	337	1.69%
	The temperature of the machine room is too high, and the control system turns on automatic protection.	123	0.62%
	Elevator restarts automatically after a fault	327	1.64%
	Others	1136	5.70%

. Therefore, when carrying out elevator fault prediction, it is necessary to consider the variations in the real operating environment parameters to obtain more accurate results.

In summary, door system faults stand as the most prevalent category among all elevator fault types. As a critical subsystem, door-related malfunctions may directly or indirectly lead to entrapment incidents, establishing them as the primary challenge requiring urgent resolution in contemporary elevator maintenance and safety management.

3. Prediction of Elevator Entrapment Incidents

3.1. Data Sources and Description

The data is obtained from the official database of the Emergency Response Center for Elevators in City H, covering the period from 1 January 2021 to 1 January 2024. The raw dataset contains diverse information extracted from various tables within the database, including basic elevator information specified in

Table 5. Elevator sample data.

No.	Data Item	Sample Data	No.	Data Item	Sample Data
1	Device encoding	301xxx	12	Installation unit	xx Enterprise
2	Address	xx City/xx Province	13	Date of installation	7 January 2020
3	Street	xx Street	14	Service life	Four-year
4	Address type	Public house	15	Maintenance unit	xx Enterprise
5	Address name	xx Hotel	16	Inspection unit	xx Enterprise
6	Elevator model	A TKJ1000/2.5—JX	17	Inspection time	3 July 2023
7	Elevator type	Traction drive	18	Repairer unit	xx Enterprise
8	Speed limit	2.5 m/s	19	In use	In use
9	Rated load	1000 kg	20	Manufacturer	xx Enterprise
10	Floor	10 Floors	21	Date of manufacture	23 December 2019
11	Units of use	xx Enterprise	22	……	……

, historical records of passenger entrapment incidents and rescue details specified in

Table 6. Elevator entrapment incident data.

No.	Data Item	Sample Data	No.	Data Item	Sample Data
1	Device encoding	301xxx	6	Rescue time	3 July 2023 5:30
2	The time of the fault	3 July 2023 5 pm	7	Rescue unit	xx Enterprise
3	The type of fault	Elevator entrapment	8	Cause of the fault	Mechanical jamming
4	Number of people rescued	Three persons	9	Fault	Door fault
5	Time to arrive	3 July 2023 5:15	10	……	……

, maintenance logs specified in

Table 7. Historical elevator maintenance records.

No.	Data Item	Sample Data	No.	Data Item	Sample Data
1	Device encoding	301xxx	4	Type of maintenance	Half-monthly /quarterly
2	Maintenance time	29 May 2024	5	Maintenance personnel	Li hua
3	Maintenance unit	xx Enterprise	6	……	……

, and other relevant factors specified in

Table 8. Ancillary factor datasets.

No.	Data Item	Sample Data	No.	Data Item	Sample Data
1	Date	1 May 2024	5	Minimum temperature	12 degrees Celsius
2	Season	Spring season	6	Maximum humidity	90%
3	Holiday	International Labor Day	7	Minimum humidity	30%
4	Maximum temperatures	21 degrees Celsius	8	……	……

.

In addition, this study collected the temperature and humidity from the elevator’s operating environment. Considering the differences in elevator design and structure that lead to variations in temperature and humidity conditions, this study selected three distinct elevator types for data collection: steel-structure retrofitted elevators (machine-room-less), glass panoramic elevators, and conventional elevators.

Temperature and humidity sampling points were established at two locations: the elevator machine room (or the top of the shaft for machine-room-less elevators) and the pit, as shown in Figure 2. Sampling point 1 was positioned at the outer side of the host machine’s “I”-shaped steel base (20 cm from the driving host for machine-room-less elevators), while sampling point 2 was placed on the pit wall 30 cm above the floor. The monitored environments included natural outdoor conditions in shaded, ventilated areas (Sample Group 1), steel-structure retrofitted elevators (Sample Group 2), glass panoramic elevators (Sample Group 3), and conventional elevators (Sample Group 4).

Data collection commenced on 1 January 2023, and continued uninterrupted to the present, with 24 h daily monitoring at a sampling frequency of once per hour. The collected environmental parameters are presented in Figure 3. The left panel (a) displays elevator machine room data. As most machine rooms are located at the top of buildings, their environmental parameters show strong correlation with outdoor conditions (e.g., direct sunlight or rain significantly affect temperature and humidity), resulting in more pronounced fluctuations in the graph. The right panel (b) shows elevator pit data. Since pits are typically situated at the lowest level of buildings, the temperature remains relatively stable as it’s less affected by external environmental factors, often maintaining a constant level in the graph. However, being underground, the pit humidity varies with elevator operation and changes in underground moisture levels.

To characterize the data properties comprehensively and establish appropriate preprocessing protocols, systematic analyses on the raw dataset containing elevator manufacturing specifications, performance parameters, geographical locations, operational records, fault histories, and maintenance logs are conducted. This included examining value ranges, distribution patterns, density concentrations, outlier identification, and missing data assessment.

The original dataset exhibits substantial volume and wide coverage. Directly incorporating all data into the modeling process would impose excessive computational burdens and adversely affect predictive accuracy and operational efficiency. Consequently, rigorous parameter screening and analytical evaluation become imperative.

Correlation and weight analyses were performed on the initial set of 150 input parameters. In the elevator risk assessment framework, the influence weight of each input parameter on the input layer is automatically determined through data-driven model training, generating a parameter weight distribution without manual intervention. These weighted parameters undergo transformations through complex neural network computations to derive the final entrapment probability values for individual elevators. It should be emphasized that a single parameter’s weight cannot be directly equated with its proportional impact on the output probability, owing to the sophisticated nonlinear transformations inherent in the neural network architecture. However, parameters with higher weight contributions generally demonstrate greater influence on the probability output. By selecting parameters exceeding the 0.5% weight threshold (while excluding those with lower weights that demonstrate negligible impact), the retained parameter subset accounts for 87.454% of the total predictive weight.

3.2. Predictive Modeling of Elevator Entrapment Incidents

A hybrid model is constructed based on Self-Attention and LSTM, and its network structure is shown in Figure 4. The model performance has been validated in our previous work [26].

Discrete data such as equipment code, geographic location, fault type and other discrete data are mapped to the feature space by solo thermal coding. Firstly, continuous data such as elevator age, maintenance cycle and other continuous data, are subjected to preliminary feature extraction, and constant and discrete data are spliced; then, the spliced feature information is passed into the LSTM network, and the (time-encoded) temporal abstract features are obtained through the first layer of the LSTM network. These temporal abstract features are subsequently passed to the Self-Attention Mechanism layer, and the current time point features are calculated using the different time point features between similarities and performing weighted summation to obtain the current time point features with integrated full-time information. Next, dropout is applied to the output of the Self Attention Mechanism to reduce the overfitting; it is then input to the second layer of the LSTM network and applied to produce the second layer of the LSTM again. These integrated temporal features are then passed into the two fully connected layers, and Sigmoid processing is performed in order to transform the features into predicted values of elevator entrapment probability.

Through grid search of the model parameters, several key parameters were determined: the sliding window size was set to 12; the step size was set to 1; the number of LSTM layers was 2; the number of hidden units was 128, and the dropout rate was set to 0.1. The model was then trained using the Adam optimization algorithm to adjust the learning rate for parameter updates. The initial learning rate was set to 0.0001, the batch size was 128, and the maximum number of training epochs was set to 1500. Early stopping was implemented, where training would terminate prematurely if the validation loss failed to decrease for 50 consecutive epochs to prevent overfitting. Cross-entropy loss was used for the loss function. The evaluation metrics ensured accuracy, recall, and precision. The loss curve of the model training is shown in Figure 5.

From the trajectory of the training loss curve, it can be observed that the loss value decreased rapidly during the initial training phase. This demonstrates that the model effectively learns from the data and that the parameter adjustments substantially, reducing predicted errors.

As the training progressed, the loss reduction rate gradually decreased, and the curve exhibited a trend toward stabilization, indicating that the model was progressively converging to a steady state. Although the loss curve presented fluctuations during certain phases, these oscillations may stem from inherent noise in the data or suboptimal configuration of optimization parameters. When training reached 1000 epochs, the early stopping mechanism was triggered, suggesting that the model is approaching its performance limit. The dynamic changes in the loss function curve profoundly reflected on the stable learning process and the gradual convergence characteristics during model training.

3.3. Prediction Results

To provide more intuitive prediction results of elevator entrapment failures and guide maintenance enterprises in preventive measures, an elevator entrapment risk index has been developed. It is a quantitative assessment metric designed to evaluate the probability of elevator entrapment incidents occurring within a 30-day period. The elevator entrapment risk index is derived from the entrapment probability values generated by the predictive model. The predicted entrapment probability value is the raw numerical output of the prediction results, while the prediction index is an alternative representation of this probability value. Whether the predictive index shows a red code or green code is determined by the magnitude of the predicted elevator entrapment probability. Empirical validation with City H’s data determined 0.010 as the optimal high-risk threshold, where an elevator entrapment risk index ≥ 0.010 classifies elevators as high-risk (red code) and an elevator entrapment risk index < 0.010 indicates operational normality (green code).

The system retrieves the results from the database to visualize the prediction and the precise localization of elevators at risk of entrapment. It extracts relevant elevator data, including equipment details, location, and maintenance provider information, to generate a monthly elevator entrapment risk index report. This report, presented in a tabular format, records each elevator’s predicted safety status and basic information, categorizing them based on their entrapment risk index, as shown in

Table 9. Elevator entrapment risk index report (Sample).

Elevator Code	Entrapment Probability Forecast	Entrapment Risk Index	District	Street	Address	Maintenance Enterprise
301xxx	0.998	Red Code	xx District	xx Street	xxx	xx Enterprise
…	…	…	…	…	…	…
301xxx	0.003	Green Code	xx District	xx Street	xxx	xx Enterprise

.

The table records multiple critical data elements about the elevator, specifically, the unique elevator number used for identification; model-generated entrapment risk probability values; the red and green code indices based on the threshold division; the district, county, and street where the elevator is located; detailed address information; and information about the enterprise responsible for the elevator’s maintenance. This approach facilitates the rapid identification of the elevator’s risk level, improves elevator safety monitoring effectiveness, and supports subsequent on-demand maintenance work.

4. On-Demand Maintenance Method

Safe and stable operations of elevators cannot be separated from every part of the process, such as prior prediction, on-demand maintenance, after-action emergency response, personnel skills training, emergency drills, and filing. Fault-prediction-based on-demand maintenance serves to predict the entrapment risk of the elevator, determine whether a fault will occur in the future, and take specific maintenance measures. The exact implementation process is shown in Figure 6.

4.1. Development of On-Demand Maintenance Content

According to the requirements of safety specifications, elevator maintenance is divided into four categories: semi-monthly, quarterly, semi-annual, and annual. The on-demand maintenance method proposed in this paper retains the quarterly, semi-annual, and yearly maintenance items while rationalizing the semi-monthly ones. This adjustment is based on the monthly entrapment risk index report generated by the entrapment prediction model, and different maintenance measures are taken for elevators marked with red and green codes. In the specific implementation, reference is made to the requirements and inspection rules for elevator maintenance in TSG-T5002-2017 “Lift Maintenance Regulation” and TSG-T7001-2023 “Regulation for Lift Supervisory Inspection and Periodical Inspection” [27,28]. Combined with the historical fault analysis results in Section 1, the door system fails frequently. The number of faults accounts for the highest proportion; therefore, when formulating preventive maintenance measures for the red code elevator, the door system is the key maintenance object, and the “※” mark in the table is the key maintenance and inspection items of the door system. The rest are the basic maintenance and inspection items, as shown in

Table 10. On-demand maintenance protocol for red code elevators (“※” mark is the key items).

Projects	Maintenance Contents
Electrical switching and wiring maintenance	Inspect and clean electrical switches to prevent switch faults due to dust and moisture.
	Test electrical wiring for insulation resistance with inspection tools to ensure the wiring is in good condition.
	Replace aged electrical wiring to avoid short circuits and false connections.
Door system maintenance	※Check the opening and closing flexibility of elevator doors and the reliability of door locks to ensure proper operation of the door system.
	※Lubricate the door guide rails to reduce friction and noise.
	※Clean up foreign objects in the door cracks to prevent door jamming and deformation.
	※Check whether there is severe wear and tear on the door roller on the door hanging plate, and replace it.
	※Check whether the door guide rail is deformed or loosely skewed, and correct it if it is.
	※Check whether the two rollers of the door lock and the door cutter are tightly affixed; if not, make adjustments.
	※Check the wear and tear of the door chain, door belt, and wire rope to avoid breakage.
	※Check whether the status of each relay of the door system is normal.
	※Check whether the light curtain or safety contact plate action is standard.
Maintenance of traction and guidance systems	Check the traction machine’s operating condition and the brakes’ reliability to ensure smooth operation of the traction system.
	Check and lubricate the guide sheave and counter rope sheave to reduce wear and noise.
	Adjust the verticality and spacing of the guide rails to ensure smooth running of the car.
Safety device inspection	Inspect and test safety devices, such as speed limiters, safety clamps, and buffers, to ensure their reliability.
	Overload protection devices are calibrated to prevent overloading of the elevator.

.

Elevators with a green code entrapment risk index have a relatively low likelihood of entrapment incidents. When elevators break down and become entrapped, IoT sensors detect the problem. Rescue teams can then respond quickly.

4.2. Implementation of On-Demand Maintenance

In selecting elevators for on-demand maintenance testing, the number of maintenance enterprises and maintenance project sites was reasonably controlled to reduce the field experiment’s implementation difficulty effectively. Relying on the database of City H’s Elevator Emergency Response Center, the selection focused on elevator maintenance enterprises. Through in-depth analysis and comprehensive comparison of various dimensions such as maintenance volume, maintenance quality, and the scope of business operations, elevators maintained by two enterprises, A and B, in City H were chosen as test samples.

Based on the latest data in the elevator database, the samples were combined with project sites as the main body. During the combing process, it was noted that some project sites had complex situations such as changes in maintenance units, division of areas within the project sites, but not uniformity in maintenance units, and deviations in the number of elevators. To ensure the accuracy and reliability of the study, these project sites with anomalies were eliminated. After rigorous screening and cleaning, the final sample of elevators included 456 elevators in seven project sites of enterprise A and 306 elevators in seven project sites of Enterprise B, for a total of 762 units.

Predictions of entrapment faults were made for a sample of elevators using existing entrapment fault prediction algorithms, and a report of entrapment risk indices was generated. In this section, the red code elevators were predicted to have a high incidence of entrapment faults and were labeled and categorized using data from the September 2024 entrapment risk index report. The red code elevators with a high predicted incidence of entrapment fault were divided into two groups for controlled maintenance tests, with the test group performing on-demand maintenance according to the maintenance schedule in Table 10. The other group is the control group, which maintains the existing maintenance status and performs only routine maintenance, and the specific information of the two groups of elevators is shown in

Table 11. Sample elevator of enterprise A after treatment.

Serial Number	Project Point Name	Maintenance Enterprise	Total Number of Elevators	Test Group Elevator	Test Group Red Code Elevator
1	Project point 1	Enterprise A	151	75	7
2	Project point 6	Enterprise A	40	20	0
3	Project point 7	Enterprise A	74	37	3
4	Project point 8	Enterprise A	72	36	13
5	Project point 9	Enterprise A	38	19	3
6	Project point 12	Enterprise A	50	25	4
7	Project point 14	Enterprise A	31	15	4

and

Table 12. Sample elevator of enterprise B after treatment.

Serial Number	Project Point Name	Maintenance Enterprise	Total Number of Elevators	Test Group Elevator	Test Group Red Code Elevator
1	Project point 2	Enterprise B	42	21	1
2	Project point 3	Enterprise B	60	30	1
3	Project point 4	Enterprise B	42	21	8
4	Project point 6	Enterprise B	40	20	4
5	Project point 8	Enterprise B	57	28	1
6	Project point 11	Enterprise B	34	17	1
7	Project point 15	Enterprise B	31	15	4

, respectively.

The basic information of the red code elevators in the test group and the work units are uploaded to the on-demand maintenance platform, and the maintenance personnel carry out on-demand maintenance on the uploaded red code elevators through the platform.

5. Results and Analysis

In the elevator emergency disposal center and maintenance enterprises, the entrapment fault on-demand maintenance effectiveness of the comparative experiment was implemented in collaboration. The experiment selected elevators of enterprises A and B as a sample, with 762 units. The data of entrapment faults in September 2024 provided by the Elevator Emergency Disposal Center of H city is being referenced. The data show that in September 2024, the number of elevators with entrapment faults among was 24, as shown in

Table 13. Empirical data on elevator entrapment occurrences.

Maintenance Enterprise	Project Point Name	The Number of Elevators Is Actually an Entrapment
A Enterprise	Project point 1	3
	Project point 7	4
	Project point 12	2
B Enterprise	Project point 2	5
	Project point 3	5
	Project point 6	2
	Project point 11	1
	Project point 15	2

.

A comparison of the results of implementing on-demand maintenance versus scheduled maintenance in enterprise A is shown in Figure 7.

In the test group of maintenance enterprise A, there are 34 red code elevators, of which only 1 experienced elevator entrapment after on-demand maintenance, and only 1 green code elevator experienced elevator entrapment. In contrast, in the control group, there are 28 red code elevators, of which 5 experienced elevator entrapment after scheduled maintenance, and 2 green code elevators experienced entrapment.

Figure 8 illustrates the results of implementing on-demand maintenance versus scheduled maintenance for the maintenance enterprise B.

In the test group of maintenance enterprises B, there were 20 red code elevators, 3 of which experienced elevator entrapment after on-demand maintenance, and 1 experienced elevator entrapment in the green code elevator; In contrast, in the control group, there were 13 red code elevators, 10 of which experienced elevator entrapment after scheduled maintenance, and 1 experienced elevator entrapment in the green code elevator.

After excluding the impact of uncontrollable external factors on elevator entrapment incidents and analyzing the actual entrapment data from both experimental groups, we conclude that the on-demand maintenance method demonstrates significant advantages in reducing elevator entrapment occurrences compared to scheduled methods. The entrapment case statistics from enterprise A and B provide compelling evidence for the effectiveness of on-demand maintenance measures.

In addition, as shown in

Table 14. Comparison of the number of maintenance units in different maintenance methods.

Category	Number of Maintenance Units for Enterprise A	Number of Maintenance Units for Enterprise B	Total
Scheduled maintenance	228 units	154 units	382 units
On-demand maintenance	34 units	20 units	54 units

, the on-demand maintenance mode reveals a significant difference in the number of units compared with the scheduled mode: enterprise A needs to deal with 228 units of equipment maintenance units when adopting the scheduled maintenance method. In comparison, the number of units is as low as 34 units after adopting the on-demand maintenance mode. Enterprise B units are reduced from 154 units to 20 units. In terms of the total volume, the total number of units under the conventional maintenance method was 382, while the total number for the on-demand method was only 54; this is an overall reduction of 85.86%. The significant difference between two methods fully demonstrates that on-demand maintenance effectively reduces the base number of maintenance units and thus significantly reduces the input of workforce and material resources and realizes the optimal allocation of maintenance resources.

6. Conclusions

This study proposes an on-demand maintenance method based on entrapment fault prediction to meet the elevator maintenance market’s current demands and reduce passenger entrapment incidents. Based on historical fault cause analysis and field validation results, the following conclusions are drawn:

(1) Statistical analysis of 19,930 historical fault records in City H confirmed that door system malfunction is the primary cause of elevator entrapment incidents, followed by human and external factors. This indicates that the door system is the key focus of on-demand maintenance strategies.

(2) We developed an entrapment prediction model based on Self-Attention and LSTM architectures with three years of elevator operation data from City H. Based on the model’s predictive outputs, we designed an entrapment risk index report containing predicted entrapment probability values, risk level classification (red/green codes), and essential elevator information.

(3) Implementing the on-demand maintenance plan derived from the risk index report and historical fault analysis, we conducted comparative field experiments including 2 maintenance enterprises, 14 project sites, and 762 elevators in City H. The experimental results demonstrate that implementing on-demand maintenance for high-risk “red code” elevators effectively reduces entrapment incidents. Additionally, the on-demand method significantly decreases the maintenance workload compared with conventional maintenance methods. Finally, we will conduct comparative experiments with existing methods proposed by other researchers in future work.

(4) Although this study focuses on on-demand maintenance for elevator entrapment prevention, the proposed maintenance framework demonstrates transfer potential to other equipment systems, subject to adaptations in feature selection and model architecture tailored to specific device characteristics. Compliance with local regulations and adaptation to regional maintenance system characteristics are required for application in other fields.