Power meter based monitoring of elevator usage

The present invention generally relates to a system and a method for monitoring at least one elevator.

The majority of the 12 million elevators operating around the world are not equipped with modern sensor technology. The latter is a prerequisite to enable elevator health-based ef ficient operations, utilization analysis and predictive maintenance. Retrofitting older elevators with modern sensor technology for utilization and health status monitoring is costly and time consuming, and therefore rarely done in prac tice. Operators of elevators are therefore facing tremendous operation costs (e.g. for fixed schedule maintenance) com pared to what would be feasible with modern monitoring tech nology .

The issue is usually solved by retrofitting or exchanging old systems by ones that are equipped with modern sensor technol ogy from the elevator OEM. Therefore, remote monitoring of utilization and health status is only used to greenfield in stallations of new elevators with dedicated sensor technology and to modernization projects where old equipment is replaced at the end of life or to retrofits. Old legacy equipment, which however forms most of the installed base, is not re motely monitored regarding the utilization and health status of elevators.

It is therefore an object of the present invention to provide a cost saving system and method for monitoring and/or assessing the status and/or utilization of an elevator. According to a first aspect, the invention provides a system comprising an elevator monitoring device for monitoring at least one elevator operatively coupled to at least one motor which is operatively connected to at least one power meter; wherein the elevator monitoring device is configured to receive, from the power meter, data regarding the at least one motor collected by the power meter,

wherein the elevator monitoring device is further configured to pre-process the data collected by the power meter (thus generating pre-processed data) , and

wherein the elevator monitoring device further comprises a machine learning model using an analytical algorithm for monitoring and/or assessing a status and/or a utilization of the elevator (or, in other words, for monitoring and/or assessing a state of operation of the elevator) based on the pre-processed data) .

The elevator monitoring device may be realised as any device, or any means, for computing, in particular for executing a software, an app, or an algorithm. For example, the computing device may comprise a central processing unit (CPU) and a memory operatively connected to the CPU. The computing device may also comprise an array of CPUs, an array of graphical processing units (GPUs), at least one application-specific integrated circuit (ASIC) , at least one field-programmable gate array, or any combination of the foregoing. The elevator monitoring device may comprise at least one module which in turn may comprise software and/or hardware. Some, or even all, modules of the elevator monitoring device may be imple mented by a cloud computing platform. The system may, apart from the elevator monitoring device, further comprise (but does not necessarily have to comprise) the power meter configured to collect data regarding the operation of the at least one motor, in particular data about the power consumption of the at least one motor. The term "at least one motor" may comprise at least one motor configured and arranged for moving the elevator up and/or down, at least one motor configured and arranged for opening doors of the elevator and/or at least one motor configured and arranged for closing the doors of the elevator.

The system may (but does not have to) further comprise the elevator and/or the at least one motor operatively coupled to the elevator.

Preferably, the analytical algorithm is based on any supervised learning approach, and more preferably comprises, or consists of, a classification algorithm like xgboost or random forest and/or a deep neural network (such as a convolutional neural network and/or a recurrent neural network) .

In an advantageous embodiment or refinement, the system comprises the elevator, and the elevator monitoring device is located close to the elevator on an edge device or remotely. When the elevator monitoring device is located remotely from the elevator, it may be located at least partially, or completely, in a cloud computing platform and/or partially, or completely, in an elevator monitoring center. The elevator monitoring center may comprise an elevator monitoring device for each of a plurality of elevators (or more precisely: motors of elevators), or it may comprise on elevator monitoring device for all of the plurality of elevators. Advantageously, the elevator monitoring device is distributed between a local device (e.g. a microcontroller, an application-specific integrated circuit, ASIC, a field- programmable gate array, FPGA, and/or the like), or a local module, located close to the elevator and a remote device (e.g. a microcontroller, an application-specific integrated circuit, ASIC, a field-programmable gate array, FPGA, and/or the like), or remote module. In this way, the computing resources of any remote device or remote module may be exploited .

In a preferred embodiment or refinement, the power meter is configured to collect the data by measuring at least one pow er quantity (such as current, voltage, frequency, power fac tor, active, reactive and apparent power in all employed phases) of the at least one motor, preferably with a temporal measurement resolution of at least 5 Hz, preferably in the range of from 5 Hz to 50 Hz. The inventors have found that such a temporal resolution produces the best suitable data for the applications as described herein.

In a further embodiment, the data collected, preferably measured, by the power meter are provided as a time series of sensor signals indicative of a usage of the elevator. The time series is advantageously segmented into (time) windows which may be overlapping or strictly separate. For the training data, the time series is divided into windows which are then labelled with labels, in particular with event descriptors such as door opening and closing, elevator riding up/down events, number of floors travelled per ride, elevator load per ride (for example number of passengers and/or total weight of the load), etc. These labels enable training the machine learning model to generate key performance indicators (KPIs)

According to another aspect, the invention thus also provides a method for training a machine learning model (in particular a machine learning model of an elevator monitoring device) , comprising: collecting data regarding power consumption of at least one elevator by at least one power meter, wherein the data are provided as a time series of sensor signals (events); dividing the time series of sensor signals (events) into windows; labelling the windows with event descriptors; and training a machine learning model in supervised training based on the labelled data (or, in other words: the labelled time series) .

Preferably, the machine learning model is a machine learning model trained with pre-processed data. Thus, the method for training a machine learning model may further comprise a step of pre-processing the data collected by the at least one power meter, wherein pre-processing may be performed before or after labelling, and wherein training of the machine learning model is performed using the labelled pre-processed (or: pre-processed labelled) data.

Pre-processing of the training data may include data cleansing such as filling up of missing values and/or resampling of the raw data to equidistant data. Advantageously, said data cleansing is also applied in the deployment phase. For the training of the machine learning model, the data of a limited number of elevators can be used as a training set. In a further embodiment, the system comprises an acceleration sensor configured for collecting rotational and/or vibration data, wherein said collected rotational and/or vibration data is preferably used for the calibration of the machine learning model.

According to a second aspect, the invention provides a method for monitoring at least one elevator operatively coupled to at least one motor operatively connected to at least one power meter, comprising:

- collecting data by the power meter;

- transferring the data to an elevator monitoring device and pre-processing, by the elevator monitoring device, the data;

- monitoring and/or assessing a status and/or a utilization of the elevator based on the pre-processed data using a machine learning model comprising an analytical algorithm.

In a preferred embodiment, the analytical algorithm is based on any supervised learning approach, and more preferably comprises, or consists of, a classification algorithm like xgboost or random forest and/or a deep neural network (e.g. comprising, or consisting of, a convolutional neural network and/or a recurrent neural network) .

In an advantageous embodiment, the data collected by the power meter is provided as a time series of sensor signals (events) indicative of usage of the elevator, wherein the time series used for training the trained machine learning model is optionally segmented into time windows that are labelled with event descriptors such as door opening and closing, elevator riding up/down events, number of floors travelled per ride, elevator load such as number of passengers per ride, etc. By using such training data in the training phase of the machine learning model, the machine learning model is enabled to detect and recognize these sensor signals (events) in time series of data during the deployment phase.

Preferably, the machine learning model is trained with pre- processed data, wherein pre-processing of the data especially includes data cleansing such as filling up (or: substituting) missing values and/or resampling the raw data to equidistant data. For the training of the machine learning model, training data from a limited number of known elevators (or elevator/motor arrangements) can be provided, and later in the deployment phase also data from previously unknown elevators (or elevator/motor arrangements) can be processed.

In a further embodiment, the system comprises an acceleration sensor configured for collecting rotational and/or vibration data, wherein the rotational and/or vibration data are preferably used for the calibration and/or parametrization of the machine learning model.

According to a third aspect, the invention provides a computer program product comprising executable program code configured to, when executed, perform the method according to the second aspect.

According to a fourth aspect, the invention provides a non- transitory, computer-readable data storage medium comprising executable program code configured to, when executed, perform the method according to the second aspect.

According to a fifth aspect, the invention provides a data stream comprising, or configured to generate, executable program code configured to, when executed, perform the method according to the second aspect.

In particular, the program code of the third, fourth and/or fifth aspect may comprise the analytical algorithm, in particular a trained machine learning model.

According to a sixth aspect, the invention provides an elevator monitoring device configured for use in a system according to an embodiment of the first aspect of the invention and/or configured for use in a method according to an embodiment of the second aspect of the invention. The above-described method for training a machine learning model may be used for training the machine learning model of the elevator monitoring device.

Additional features, aspects and advantages of the invention or of its embodiments will become apparent on reading the detailed description in conjunction with the following figures :

Fig. 1 provides a schematic overview of a system for the monitoring of an elevator according to an embodiment of the first aspect of the invention as well as of an elevator monitoring device according to an embodiment of the sixth aspect of the invention;

Fig. 2 provides a schematic overview of a time series for data collection; Fig. 3 provides a schematic flow diagram illustrating an embodiment of a method according to the second aspect of the present invention;

Fig. 4 schematically illustrates a computer program product according to an embodiment of the third aspect of the invention as well as a data storage medium according to an embodiment of the fourth aspect of the present invention.

In the following description, for purposes of explanation and not limitation, specific details are set forth in order to provide a thorough understanding of the present invention. It will be apparent to one skilled in the art that the present invention may be practiced in other implementations that depart from these specific details.

Fig. 1 shows a general overview of a system 100 for data analytics of the status and/or utilization of at least one elevator 200. The system 100 comprises an elevator monitoring device 300 which will be described in the following. The elevator monitoring device 300 itself may be provided according to an embodiment of the sixth aspect of the present invention .

The elevator 200 in this embodiment is operatively coupled to at least one motor 240, for example one motor for moving the elevator up and/or down, one motor for opening doors and one motor for closing doors and/or the like. At least one feeder 220 is arranged for the power supply of the at least one motor 240. Each of the at least one motor 240 may be provided with its own feeder 220, or all of the motors 240 may be provided with one and the same feeder 220. Additionally, at least one power meter 260 is provided and operatively connected to the at least one motor 240 (and/or the at least one feeder 220) for collecting data from the at least one motor 240 (and/or directly from the at least one feeder 220), in particular for measuring at least one electrical power quantity of the at least one 240, for example a current, voltage, frequency, power factor, active, reactive and apparent power in all employed phases. One power meter 260 may be provided for all of the at least one feeder 220, or one meter 260 for each of the at least one feeder 220 may be provided.

The at least one power meter 260 can advantageously be installed retroactively at the at least one feeder 220 of a built-in elevator 200. The power meter (or "smart meter") 260 is placed at the feeder 220 of a motor 240 for the collection of data 400, which helps in early detection of various failure modes of the elevator 200. However, additional local devices and sensors can be added to the elevator 200. For example, acceleration sensors can be placed inside of the elevator 200 to capture rotational and/or vibration data which help in early detection of various failure modes encountered in mechanical equipment.

In the following, for the sake of simplicity, explanations of details and optional refinements will be given with respect to a single power meter 260 connected between a single motor 240 and a single feeder 220 for that motor. However, it will be readily understood that the same concepts apply also to any of the above-described variants with multiple motors 240, multiple power meters 260 and/or multiple feeder 220. The power meter 260 measures power quantities such as cur rent, voltage, frequency, power factor, active, reactive and/or apparent power in all employed phases and provides them as data 400. Advantageously, the measurement resolution is larger than 5 Hz, preferably in the range of 5Hz to 50Hz, so that a high temporal resolution is given.

The data 400 which are collected by the power meter 260 can refer to (and be indicative of) events such as door opening and closing events, elevator riding up/down events, time that the elevator 200 is idle, number of floors travelled per ride, elevator load (such as number of passengers per ride and/or total weight of load) and/or a distance that the ele vator cable has covered which is typically measured in kilo meters .

The data 400 are transferred from the power meter 260 to the elevator monitoring device (which may also be designated as an "analytical unit") 300. The elevator monitoring device 300 in some embodiments or refinements constantly analyses the data stream 400 ("real time mode") and identifies a set of predefined events that are known to provide useful infor mation about the usage of the elevator 200. Such an analysis can be conducted in real time but also subsequently such as in batch mode, for example once an hour. The elevator moni toring device 300 learns to identify predefined events by us ing artificial intelligence (AI) . A supervised machine learn ing model 500 is trained on a training set of data 400 of an notated elevator rides.

While the smart meter 240 is typically deployed directly at the elevator motor feeder 240 and/or the elevator motor 220, the elevator monitoring device 300 can either be located close to the elevator 200 on an edge device or remotely such as in a cloud computing platform, for example a cloud-based service like Siemens Mindsphere (registered trademark) .

A third option are hybrid solutions in which the elevator monitoring device 300 is distributed (or: divided) between a local module close to the elevator 200 (e.g. used for imple menting a data pre-processing unit performing a data pre processing and/or some preliminary feature calculation) and a remote module on which the machine learning model 500 is run ning on precalculated features for the calculation of rele vant key performance indicators (KPIs) .

This hybrid approach can significantly reduce the amount of data 400 that has to be transferred over a wide area connec tion - such as the Internet - between the location of the at least one elevator 200 and the elevator monitoring device 300. In the first two cases all raw power measurements (i.e. raw data 400) are transferred from the power meter 240 to the elevator monitoring device 300 (e.g. in the cloud) while in the hybrid approach preferably only the pre-processed data, for example precalculated features, are transferred to the elevator monitoring device 300.

As shown in Fig. 2, the data 400 collected by the power meter is provided as a time series 600. For the training dataset for training the machine learning model, the time series 600 is segmented (e.g. by an expert) into windows 650 that are labelled with a corresponding event descriptor such as door opening and closing, elevator riding up/down events, number of floors travelled per ride, elevator load such as number of passengers per ride, etc. For the training of the machine learning model 500 it is typically sufficient to acquire and annotate data from a few days or one week of elevator usage (depending on the defined usage event) .

The data 400 labelling process can be carried out by a human domain expert that is manually segmenting the data 400 of a time series 600 into the respective event windows 650. The labelling process can be supported by a low-level automated method using a set of threshold-based rules followed by a do main-expert validation of the segmentation proposal.

In a next step, which is advantageously performed in the training phase and optionally in the deployment (or infer ence) phase, the labelled data 400 are pre-processed which may include data cleansing steps on time series data like filling up of missing values and resampling of the raw time series data to equidistant data. In the deployment phase, where the power meter 260 will usually provide non-windowed raw data, the pre-processing step may optionally comprise windowing i.e. splitting up the time series 600 received from the power meter 260 in a sequence of overlapping or non overlapping chunks of data, with the size of the window 650 and/or the delta between consecutive windows as configurable parameters. However, the system 100 advantageously does not require domain specific pre-processing steps and is transfer able to other technical types of smart meters 260 or eleva tors 200 without, or with small, adaptations.

The machine learning model 500 is trained on the pre- processed data 400. Each window 650 of a time series 600 is processed by an algorithm (A) that estimates the most proba ble event for a given window 650. The parameters of the ma chine learning model 500 can be optimized based on a detec- tion error between the prediction or classification and the label until a sufficient performance index is reached.

For the algorithm it is possible to use classification algo rithms like xgboost or random forest. Such algorithms typi cally require the windowed data to be transformed into a set of informative and non-redundant features to facilitate the subsequent learning. Other algorithms like (deep) neural net works, especially convolutional neural networks, can be trained directly on the raw window data. A deep neural net work denotes a class of artificial neural networks that have a plurality of hidden layers between the input layer and the output layer. Such a model is able to process raw data with out prior time consuming feature engineering. A further al ternative option for the machine learning model 500 is an un supervised segmentation approach that divides the time series 600 into segments and assigns each segment to a particular sensor signal (event) .

After the machine learning model 500 has been trained with the annotated data set 400 and the classification performance is sufficient (which may be determined using a test data set) , and optionally after a calibration and setup phase, the machine learning model 500 can be deployed (optionally to gether with the power meter 260 if it is not already present) even at a formerly unknown elevator 200 from which no data 400 was used during the training.

During the operation phase of the machine learning model 500 relevant elevator operation key performance indicators (KPIs) are calculated, of which a few are discussed exemplarily in the following: the number of door openings/closings and total ride length are calculated from the sequence of elevator op- eration events. The sequence of elevator operations is clas sified into segments. The KPI "number of door open ings/closings" can be directly calculated from the respective segments. The calculation of each ride length is done based on the length of the "elevator riding up" segments and "ele vator riding down" segments and a physical model of the ele vator acceleration and breaking using the characteristic ac celeration and breaking parameters from the elevator specifi cation/setting.

Alternatively, the length of the rides can be determined by applying a trained ride length model on features calculated from the "ride up" or "ride down" segments. Failures can be identified using the domain knowledge like the time required for door closing or door opening and the allowed sequence of events, for example an elevator 200 can only ride after the door has closed, or a door can only be opened again after having been opened before when it has been closed in between.

The present invention allows for low-cost monitoring and/or assessing of the utilization and/or health status of an ele vator 200 by attaching a power meter 260 to a motor 240 (in particular to a feeder 220 of a motor 240) of an installed elevator 200 and by analyzing the power meter data 400 by us ing machine learning models 500 in a elevator monitoring de vice 300. Additional sensors and costly retrofits of the ele vator 200 are not required. Therefore, there is no new certi fication of the elevator system by approving authorities re quired. Due to the usage of artificial intelligence, manual adjustments to the system 100 when deployed to a new site are reduced to a minimum. The elevator monitoring device 300 may further comprise an analysis module configured to compare KPIs (or predictions) of the machine learning model 500, in particular of the analytical algorithm, to expectations and/or subject them to predefined rules about "suspicious" or "faulty" behavior.

For example, the machine learning model 500 may predict the elevator 200 to be moving up after a predicted "doors open" event that has, however, not yet been followed by a predicted "doors close" event. One of the predefined rules may dictate that the elevator must not move without each "doors open" event having been followed by a "doors close" event, or, in simple terms, that the elevator must not move with open doors .

The results of the analysis module may be output in a monitoring signal. The monitoring signal may be used, for example, to control or feed an elevator KPI dashboard which may be displayed by an optional display device of the system 100. Such an elevator KPI dashboard may receive, and illustrate, a plurality of monitoring signals from a plurality of elevator monitoring devices 300 monitoring a plurality of elevators 200 at a plurality of sites.

The cost advantage of the proposed solution is realized by attaching (if it did not previously exist) a power meter 260 to the feeder 220 of an elevator 200 without changes to the elevator core system and calculating relevant elevator moni toring KPIs like the number of door openings/closings and the total ride length by applying trained machine learning models 500 to the data 400 measured by the power meter 260. According to some advantageous embodiments of the present in vention, a machine learning model 500 is trained and applied to time series data 400 from a power meter 260 attached to an elevator 200 in order to generate a sequence of elevator op eration events from which relevant elevator monitoring KPIs like the number of door openings/closings and the total ride length can be calculated. The machine learning model 500 typ ically involves a classification or segmentation approach.

In Fig. 3, processing steps of a method according to an embodiment of the second aspect of the present invention are illustrated schematically. The method may be performed advantageously using the system 100 as described in the foregoing, or any other system according to any embodiment of the first aspect of the invention. Method according to embodiments of the second aspect of the invention may be adapted, modified and refined according to any options, modifications and refinements as described with respect to any embodiment of the first aspect of the present invention and vice versa.

In a step S100, data 400 are collected by the power meter 260, in particular by measuring at least one power quantity (such as current, voltage, frequency, power factor, active, reactive and apparent power in all employed phases) related to at least one motor 240 of an elevator 200, in particular by measuring at a feeder 220 of said at least one motor 240.

In a step S200, the data 400 are transferred to an elevator monitoring device 300 and pre-processed therein. As has been discussed in the foregoing, the pre-processing in step S200 may comprise data cleansing, windowing and/or feature precalculation . In a step S300, a machine learning model 500 comprising an algorithm is used (preferably by the elevator monitoring device 300) for the data 400 for monitoring and/or assessing the status and utilization of the elevator 200. To that end, the predictions of the machine learning model 500, in particular of the analytical algorithm, may be compared to expectations and/or be subject to predefined rules about "suspicious" or "faulty" behavior.

The predictions of the machine learning model 500 and/or the results of the application of the predefined rules to said predictions may be output in a monitoring signal. The monitoring signal may be used, for example, to control or feed an elevator KPI dashboard which may be display by a display device.

Fig. 4 schematically illustrates a computer program product 700 comprising executable program code 750 configured to, when executed, perform a method according to an embodiment of the second aspect of the present invention. Fig. 4 may also be used to illustrate a non-transitory, computer-readable data storage medium comprising executable program code configured to, when executed, perform a method according to an embodiment of the second aspect of the present invention.

The system 100 makes it possible to monitor and/or assess the utilization and/or health status of an elevator 200 by evalu ating the time series 600 of electrical power quantities of the elevator 200 in all electrical phases employed. The sys tem 100 makes use of artificial intelligence (AI) to automat ically identify patterns in the time series data 400, thereby reducing the manual effort needed for installation and maintenance. As the system 100 only requires input data 400 that can be drawn directly from a power meter 260, it can be used without the need of additional and potentially expensive sensors. During training, the machine learning model 500 may be trained with data 400 from a limited number of elevators 200 of e.g. a type A, and this trained machine learning model 500 can then be deployed for further elevators 200 of a dif ferent types such as type B in the deployment phase.

Reference Signs

100 monitoring system

200 elevator

220 feeder

24 0 motor

2 60 power meter

300 elevator monitoring device

400 data

500 machine learning module

600 time series

650 window

700 data storage medium

750 program code