Machine Learning Solutions for Monitoring Door Vandalism

by Gabriela Roivainen, Matti Lin, Joel Cardenas, Kari Karen and Konsta Simonen

Abstract

Doors being misused or vandalized has been recognized as a major cause of elevator failures, especially in metro stations, airports or other market segments where the traffic is heavy.

The focus of this paper is to go through the findings in creating an economical and accurate solution for monitoring the doors’ health, identifying the vandalism failure type and notifying the maintenance team about the details of occurring failures.

The methodology consists of capturing the signals required for operating the doors during their opening and closing cycles. No additional sensors are required for data collection, which makes the solution attractive economically. Several Machine-Learning (ML) algorithms were developed and compared for analyzing the changes induced by passengers hitting the doors, the time when the hit occurred in relation with the running cycle of the doors, the position of the hit and the response time for identifying the fault and sending the notification.

Data collection and analysis were essential for the initial training of ML algorithms to identify the most accurate and reliable solutions. Due to the variability inherent in human-induced actions, which can lead to significant differences in data from case to case, it was imperative to utilize a comprehensive and validated dataset. This approach ensured the attainment of optimal results.

Finally, the most suitable algorithm was selected and deployed on an edge device located on the top of the car. The testing and validation of the solution was then performed to ensure the efficacy of the ML algorithms.

1. Introduction

Doors play a critical role in the operational reliability of elevator systems, particularly in environments characterized by high people traffic. Traditionally, door-related issues account for a significant portion of elevator service callouts across industry — approximately one-third of all reported incidents. Moreover, these door malfunctions are often associated with passenger entrapment, further complicating service and maintenance operations.

The importance of effective elevator monitoring and maintenance is leveraged by the expanding global market, driven by escalating urbanization and a significant rise in the elderly population. With an estimated annual growth rate of 6.4% from 2023 to 2030,^[1] the elevator industry faces increasing demands for reliability and longevity, with typical elevator lifespans ranging between 15 and 30 years across various market segments.^[2,3] Given this context, the ranking of elevator faults by frequency, impact, repair cost and safety implications reveals that vandalism and user- induced damage are among the top 10 most critical issues. These incidents, especially in high- traffic environments, are often indistinguishable from standard door malfunctions, complicating fault diagnosis and reporting.

The installation and maintenance of elevator doors is a meticulous process that often requires on-site assembly by technicians using delivered parts. The sensitivity of these processes can introduce variability in door performance, influencing their reliability and susceptibility to operational failures. The economic repercussions of poor-quality door installations are considerable, affecting both maintenance costs and customer satisfaction, as the primary goal remains ensuring elevator systems are operational with minimal downtime.

In response to these challenges, there is a growing emphasis on developing robust monitoring and diagnostic solutions aimed at identifying door malfunctions and instances of vandalism. These initiatives are crucial for enhancing system reliability and reducing service disruptions, particularly in infrastructure segments like airports, metros and train stations where the misuse or vandalism of doors is more prevalent due to higher traffic and availability requirements.

This paper explores a novel approach centered on leveraging ML algorithms to monitor the health of elevator doors, detect vandalism and promptly notify maintenance teams of potential faults. By analyzing signals captured during normal door operations without additional sensors, the proposed solution offers an economically viable and technically feasible method for enhancing elevator system reliability.

This introduction sets the stage for a detailed examination of the methodology, findings and results of deploying ML algorithms in door monitoring, aimed at achieving both economic efficiency and operational reliability in elevator systems.

2. Selecting the Solution

There are several proposed solutions for monitoring door vandalism: video monitoring; using accelerometers, pressure, sound or other sensors; monitoring elevator control systems, etc.

There are benefits and challenges for each of these solutions. Video monitoring of each landing door and car door can provide real time monitoring and identification of the user misusing the elevator;^[4,5] however, the complexity of the system makes the solution expensive. In addition, the identification of the user must be aligned with existing general data protection regulations.

The methodology for detecting elevator-related door faults or vandalism using external sensors such as accelerometers have been presented in many patents.^[6,7,8] The solution is more economical than video monitoring due to the cost of the sensors and the smaller volume of the data that must be stored and analyzed. The challenge of using accelerometers is the training of ML to differentiate normal behavior and misuse: The absolute value of the vibration level and its spectrum is highly dependent on the quality of sensors, their position in the car and the elevator configuration.^[9]

The faults reported by control systems, related to doors, have been proposed and studied as a solution for monitoring vandalism. The monitoring solution can identify with high accuracy when the door functionality is compromised and the elevator is out of order; however, there are challenges in identifying the root cause of door failure whenever there is a design fault, wear of equipment or vandalism.

The solution proposed in this article is combining the door expert’s knowledge with the data analyst’s expertise. Door controller operational data and synthetic data from simulation models are used for training ML algorithms.

Fault identification and notification are performed on the data collected from the door controller for the time with the highest risk of vandalism: people rushing in or out of the car with large heavy luggage, during the door opening or closing. The monitoring of landing doors when  the car is not located at the landing is considered a lower risk, and due to the additional cost induced by collecting data from each landing door, it is outside the scope.

The result is a simple edge device located on top of the car and connected to the door-operating system with a single cable [Figure 1].

2.1 Simulation Models

Developing algorithms for preventive and detection purposes in elevator systems is inherently challenging due to the complexity and variability of the components involved. To achieve optimal accuracy and minimize false positives in scalable algorithms, resource-intensive testing and data acquirement are essential. For labeling and validation purposes, a substantial number of sensors and monitoring systems in conjunction with existing elevator infrastructure are required for ensuring statistical reliability across diverse conditions. Additionally, elevator doors are highly configurable sub-systems with varying components, resulting in different data baselines for each configuration. 

In this study, simulation models representing elevator door systems were employed to analyze door behavior during nominal operation, misuse and failure scenarios. Findings from simulation analyses are utilized to develop specifications for algorithms designed to detect misuse and quantify its severity. These findings also determine the selection of specific outputs that should be monitored by ML algorithms for each malfunction case to ensure accurate and reliable detection. The plausibility of utilizing simulations to artificially recreate malfunctioning elevator door systems was investigated successfully in ^[9]. The study’s results demonstrated a sufficient correlation between simulated outcomes and data obtained from testing environments. The proposed methodology has also been applied in other industries, such as the automotive industry,^[10,11] for similar purposes.

Door vandalism or misuse is a challenge for all segments of the market, with a higher importance for the infrastructure segment: airports, metro stations and train stations. 

From simulation results, it was found that when elevator doors were struck with an external force, a fast change in door speed, position and door motor force intake are detectable [Figure 2].

The finding that vandalism induces variation in data collected from the door controller determines the next step required for validating the models and extending the computation for different door configurations, vandalism force, time and position.

2.2 Collecting Data

The data that was feasible for the vandalism and misuse detection defined from simulation models were collected from elevator door systems. The door motor and drive communicate with the door controller in a control loop. The controller receives motor encoder data and drive power outputs, which are logged by an edge device located in elevator infrastructure via a one-way secure data transfer protocol. From the logged data, the outputs feasible for detection algorithms could be calculated using conversion equations depending on the motor specifications.

Data available in the communication busses of elevators can be utilized as additional metadata to the prediction algorithms or to be added as condition parameters indicating door system functional failures caused by vandalism. From the metadata, the floor information from door events is the most important, as the data is essential to distinguish data from each floor with individual landing doors. The condition and configuration of each landing door may vary, which can make prediction algorithms based on historical data obsolete if floor information is not available. Additionally, this information is essential for preventive maintenance, guiding field technicians to the correct floor where the predicted fault can be found.

The feasibility study of external accelerometers installed on door panels for detecting vandalism showed that impact impulses manifest as sudden displacements in door components. For the sensors, precise placement is essential to accurately capture the impacts. Both the placement and orientation of the sensor are critical factors influencing both training data quality and the overall performance of a possible detection algorithm. The transfer path of the impact varies depending on door configurations and sensor placement, making it challenging to assess the severity of the impact. Additionally, the robustness of different door configurations affects the measured impulses, which closely correlate with the energy applied to the doors. The data collected from accelerometers have been tested only in the training phase of machine learning algorithms to label the events.

Door controller data collection from normal use of the doors and vandalism, when the doors have been impacted by a test trolley, have also been collected. This data has been used to label the events used for training ML algorithms.

3. Selecting ML Algorithm

The development of the solution for door impact detection was based on dividing the vandalism events into three classes: nominal, impact and impact leading to failure. In this phase, the problem was formulated as a multi-class classification task.

The data analysis revealed that the most significant correlations with impacts were found in the largest differences between actual and target door speeds, as well as in the cumulative accumulation of these differences. This finding was anticipated, as a malfunctioning door that cannot move will continue to accumulate differences between actual and target speeds for the duration of the logging cycle. This feature was identified as a potential indicator of door breakage. However, in some impact cases, the door may start reopening, meaning the cumulative error will not accumulate, or it becomes difficult to distinguish from cases where the door is naturally reopening. In addition, the maximum speed difference between actual and targeted value depends on if the impact occurred near the start or end of the door cycle or in the middle of the door cycle.

Time criticality is essential when developing solutions for vandal detection, as it is required to flag incidents as soon as possible. This is particularly relevant for impacts leading to failure, where the data cycle logging ends after several minutes. Therefore, in these cases, the inference must be performed during the data cycle.

Another criterion for selecting the final solution was to ensure it flags only instances of impacts and no other malfunctions.

The field of ML covers a wide range of different algorithms, many of which can be used for classification.^[12] The process of model selection typically involves several iterations of the design cycle, which includes steps such as dimensionality reduction and predictor design.^[12,13] In this case, in addition to accuracy, emphasis was placed on the algorithm’s interpretability. Logistic regression was chosen as the baseline model due to its linearity and high interpretability.^[14] The objective was to enhance its performance. The baseline model was designed to be a streaming solution, meaning that classification occurs concurrently with door movement. The initial accuracy rate achieved with the linear model was around 75%. Figure 3 visualizes the development of streaming solution type classifiers. The first cycle is classified as nominal, (green) then impact happens and the label transfers from impact (yellow) to impact leading to failure (red).

The relatively small size of the initial data set limits the use of more complex algorithms, as these could potentially lead to overfitting.^[13] The metrics for each algorithm were evaluated using the cross-validation method, a widely used technique in such scenarios. ^[12]

One of the methods tested was random forest classifier, which is a commonly used algorithm in ML.^[14] This method can effectively handle high-dimensional data and reduce the risk of overfitting by utilizing multiple trees.^[14] However, it also has disadvantages, including limited control over the model and increased complexity compared to a single decision tree, making them difficult to interpret when the number of trees increases. The initial accuracy rate achieved with the random forest was around 80%.

Neural networks were also considered as a potential solution. These methods are typically most advantageous when dealing with large data sets with complex relationships.^[14] Compared to previous methods, the complexity increases and interpretability decreases, especially when dealing with deep neural networks.^[12,14] In this case, the accuracy rate was found to be like that obtained with the random forest classifier.

The preliminary results confirmed the ability to detect impacts on doors. However, there were indications that performing the classification after the data cycle had been logged could enhance accuracy. Data analysis demonstrated that jammed doors could be identified during the door cycle using a linear model. This would then enable binary classification after the door cycle, separating nominal cases from impacts that do not lead to failure. Figure 4 illustrates a plane to differentiate landing-door failures from other types of cycles. The data-logging process was halted after the door became jammed, resulting in a higher cumulative speed error in practice.

The random forest classifier was selected for the classification of the subsequent data cycle due to its good performance in preliminary tests. Hyperparameters were optimized through an exhaustive search over a predefined parameter grid, resulting in an enhanced accuracy rate of up to 90% and a precision rate exceeding 90%. In these reported results, the classification was conducted correctly, meaning that all instances where an impact occurred, regardless of its magnitude, were labeled accordingly. It was found that car door impacts, for instance, are more challenging to detect than landing door impacts. The findings indicate that, by focusing on optimal impact events, there is potential to further enhance the precision rate. By prioritizing precision over recall, the false positive rate can be reduced even if it means some impacts may be missed.

To fulfill the fast-flagging requirements, two ML solutions have been proposed: a streaming solution, for impact leading to failure, which is doing the inference live, when the data is logged; and a delayed solution for impact not leading to failure when the data is inference after the whole door cycle has been completed.

The inputs for a delayed ML solution are maximum, minimum or derivatives of door speed, position and force. In addition, thresholds and scaling factors based on simulation models are used. The accuracy of this solution is higher.

The input for a streaming ML solution is limited, because cumulative errors are calculated only at the end of the cycle, which is too late for the fast-flagging requirement. The accuracy of this solution is not as high as the previous one and it can be improved by using metadata from the elevator communication buss.

4. Implementation

Vandalism detection necessitates the real-time analysis of data to ensure that acts of vandalism are identified as they occur. Consequently, the ML algorithm for vandalism detection is closely integrated with the data-logging software to minimize latency and enhance the promptness of detection.

Data logging and vandalism detection are conducted on-site through edge computing, implemented directly on the elevator car. The edge device is interfaced with the door operator via a proprietary protocol over a serial connection.

The software and hardware are designed for plug-and-play functionality, allowing a technician to simply position the device on top of the elevator car and connect it to the door operator and supply power. Once installed, the system will autonomously initiate data collection and analysis.

The functionalities of the solution operate entirely without the need for internet access. All data post processing and ML inferences are performed locally on the edge device, eliminating the need for remote computation. This architecture enables the solution to function effectively in offline environments, which is particularly advantageous in scenarios where cybersecurity is a primary concern. Additionally, critical notifications generated by the device can be transmitted through the local network directly to a local control room, ensuring continued operation even in the absence of internet connectivity.

In an online configuration, the analyzed data and inference results are transmitted from the edge device to a cloud-based solution. Notification rules and data storage are managed within the cloud environment,^[15] facilitating more efficient administration of notifications and enabling the data to be utilized for additional applications. Data transfers from the edge device to the cloud are conducted using the MQTT messaging protocol^[16] and Kafka,^[17] ensuring reliable and scalable communication.

The solution’s remote-access capabilities and ease of deployment enable continuous development, allowing containers and various components to be updated remotely without prolonged delays. This facilitates the incorporation of new features during their development, permitting real-time analysis in real-world environments to assess performance. Once validated, these features can be fully implemented and deployed for practical use.

5. Conclusions

Door vandalism or misuse is a challenge for all segments of the market, with a higher importance for the infrastructure segment: airports, metro stations and train stations. Identifying this type of elevator failure is relevant both for the company providing the maintenance and for the customer. The early identification of door-related issues allows for more effective maintenance planning, reducing the need for emergency repairs and enabling predictive maintenance strategies. This not only lowers overall maintenance costs but also minimizes disruptions to service.

A cost-efficient solution has been presented where ML algorithms deployed on an edge located on top of the car can successfully identify and notify the misuse and vandalism of the doors for the phase when the car is at the landing. The use of existing operating system signals for data collection eliminates the need for additional sensors. This makes the solution not only cost- effective, but also easier to implement without requiring significant changes to the elevator’s infrastructure.

While this solution has demonstrated its effectiveness in infrastructure segments with high foot traffic, the approach can be easily scaled and adapted to other market segments such as residential and commercial buildings, thereby offering a versatile tool for a wide range of elevator systems.

By reducing elevator downtime and minimizing incidents of passenger entrapment caused by door failures, this solution significantly improves the overall user experience and customer satisfaction, aligning with the core objective of maintaining maximum elevator availability.

References

[1] Elevator market report [grandviewresearch.com/industry-analysis/elevators-market-report], 2023.

[2] Elevator life expectancy [elevatorsource.com/elevator_life_expectancy.htm], 2019.

[3] Life expectancy of a residential lift [futureliftservices.co.uk/blog/what-is-the-life-expectancyof- a-residential-lift/], 2023.

[4] Elevator monitoring [2n.com], 2024.

[5] What is elevator monitoring [sirixmonitoring.com] 2023.

[6] EP 3424860 A1 “An elevator vandalism monitoring system,” European patent application, 2019.

[7] EP 3191395B1, Roberts, Randall Keith. “Vibration-based elevator tension member wear and life monitoring system.” European patent application, 2015.

[8] EP 4188862A1, Martin Zellhofer, “A computer-implemented method for training a machine learning model to detect installation errors in an elevator, in particular an elevator door, a computer-implemented method for classifying installation errors and a system thereof.” European patent application, 2022.

[9] Lin, M. “Physics-based simulations of elevator door systems.” Master thesis, Aalto University, 2021.

[10] Magargle, Ryan, et al., “A Simulation-Based Digital Twin for Model-Driven Health Monitoring and Predictive Maintenance of an Automotive Braking System.” Modelica. 2017.

[11] GE, Boston, MA., “How a ‘Digital Twin’ for physical assets can help achieve no unplanned downtime,” 2016.

[12] Li, H., Lin, L. & Zeng, H., “Machine Learning Methods.” 1st ed. 2024. Singapore: Springer Nature Singapore, 2024.

[13] Braga-Neto, U. “Fundamentals of Pattern Recognition and Machine Learning”. 2nd ed. Springer International Publishing, 2024.

[14] Vijayvargia, “A. Machine Learning with Python: An approach to Applied Machine Learning.” New Delhi: BPB Publications, 2018.

[15] Amazon Web Services [aws.amazon.com], 2024.

[16] Message Queuing Telemetry Transport (MQTT) [mqtt.org], 2024

[17] Apache Kafka [kafka.apache.org], 2024.