Method, system and software product for detecting installation errors in an elevator
By deploying sensors in elevators and using machine learning models to detect installation errors in elevator doors, combined with a setpoint function and a Fourier transform model, the problem of poor elevator installation quality was solved, improving installation efficiency and customer satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WITTUR HLDG GMBH
- Filing Date
- 2022-06-17
- Publication Date
- 2026-07-24
AI Technical Summary
In the existing technology, poor elevator installation quality leads to a high recall rate. Existing monitoring systems cannot effectively detect and improve installation quality, which affects customer satisfaction and increases costs.
By deploying multiple sensors in the elevator and using machine learning models to detect installation errors, and combining set function models and Fourier transform models, the physical parameter features of the elevator doors are extracted and classified, achieving efficient detection and classification of installation errors.
It improved the quality of elevator door installation, reduced installation time and costs, enhanced the efficiency of maintenance operations, and reduced the chain reaction caused by poor installation.
Smart Images
Figure CN117412916B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a computer-implemented method for training a machine learning model to detect installation errors in elevators, particularly elevator doors, and a computer-implemented method and system for classifying installation errors.
[0002] One application of this invention is the installation control of elevator doors, particularly elevator landing doors and car doors.
[0003] This invention is also applied to predictive maintenance and remote monitoring of elevator doors.
[0004] Another application of the present invention relates to elevator drive systems, particularly motors, braking devices, and encoders.
[0005] Another application of the invention is in safety devices, such as overspeed regulators, safety equipment, regulator tension ropes, and razors.
[0006] The design also envisions the control of accelerometers, load weighing sensors, elevator controllers, and door safety switches.
[0007] Overall, the proposed invention is applicable to any component of an elevator. Background Technology
[0008] In the prior art, monitoring systems for elevator equipment based on data detected from sensors distributed throughout the equipment have been proposed.
[0009] According to its abstract, document US 10,196,236 B2 proposes a monitoring system for elevator equipment and a method for operating the monitoring system to generate elevator door usage data. The monitoring system includes sensors deployed in the elevator equipment and an evaluation unit, wherein at least one physical parameter of the environment of the sensors can be detected by the sensors, and the evaluation unit determines the operating status of the elevator doors by the changes of the physical parameter over time.
[0010] According to the abstract of document US 2020 / 0062542 A1, a method and system for determining the position of an elevator car are provided. This method and system are based on collecting vibration data associated with one or more components in the machine room of an elevator system by operating machine room sensors via a processor. The elevator system includes an elevator car and a hoistway, and the method analyzes the vibration data to determine the position of the elevator car within the hoistway.
[0011] The solutions described above aim for predictive maintenance, but are still susceptible to poor installation quality. Because the data captured by the sensors is still indirect, it is insufficient to improve installation quality, meaning they are not directly connected to the door operator.
[0012] There is a need to reduce the time and costs caused by incorrect installation. In fact, elevator companies and multinational corporations experience numerous recalls within six months of elevator release, primarily due to poor installation quality. The main component leading to these recalls is usually the door.
[0013] It is clear that installation quality is a key driver of customer satisfaction. Summary of the Invention
[0014] In this context, the technical objective of this invention is to propose a computer-implemented method for training a machine learning model to detect installation errors in elevators, particularly elevator doors, and a computer-implemented method and system for classifying installation errors, which overcomes the aforementioned disadvantages of the prior art.
[0015] Specifically, the purpose of this invention is to propose a computer-implemented method for training a machine learning model to detect installation errors in elevators, particularly elevator doors, and a computer-implemented method and system for classifying installation errors, thereby allowing for better detection of incorrect elevator door installations compared to existing solutions, and thus improving the quality of the installation process, particularly for doors.
[0016] Another object of the present invention is to provide a computer-implemented method and system for classifying installation errors in elevators, particularly elevator doors, which allows for the scheduling of maintenance operations and monitoring in a more efficient and easier manner, thereby reducing the time required to verify correct installation.
[0017] Another object of the present invention is to provide a computer-implemented method and system for classifying installation errors in elevators, especially elevator doors, which reduces the number of cascading effects caused by poor installation.
[0018] The technical task and specific objective are essentially achieved through a computer-implemented method for training a machine learning model to detect installation errors in elevators, particularly elevator doors. This machine learning model is a combination of a setpoint function model and a Fourier transform model. The method includes the following steps:
[0019] - Multiple sensors are installed at the elevator, each configured to detect physical parameters;
[0020] - By using sensors to detect the values of physical parameters, a dataset including at least one time series can be obtained;
[0021] - The first input layer is obtained by extracting features from the dataset;
[0022] - Obtain the second input layer by extracting features from the dataset;
[0023] - Feed the set function model using the first input layer;
[0024] - The Fourier transform model is fed in using the second input layer.
[0025] According to one aspect of the invention, the dataset includes a time series matrix.
[0026] According to one aspect of the invention, the dataset also includes one or more of static values or cyclic values.
[0027] According to one aspect of the invention, the dataset also includes audio samples.
[0028] According to one aspect of the invention, the extracted features of the audio sample include an audio spectrogram.
[0029] According to one aspect of the invention, the step of detecting the value of a physical parameter by means of a sensor is performed at different intervals depending on the type of sensor and the type of physical parameter involved.
[0030] The technical task and specific objective are essentially achieved through a computer-implemented method for detecting installation errors in elevators, particularly elevator doors, the method comprising the following steps:
[0031] - Multiple physical parameters are detected by using sensors placed at the elevator;
[0032] - Perform first and second feature extraction based on the detected physical parameters;
[0033] - The extracted first and second features are fed into a previously trained machine learning model to detect installation errors.
[0034] The technical task and specific objective are essentially achieved through a system for detecting installation errors in elevators, particularly elevator doors, the system comprising:
[0035] - Multiple sensors are placed at the elevator and configured to detect physical parameters;
[0036] - A first feature extraction unit and a second feature extraction unit, configured to extract features from the detected physical parameters;
[0037] - A machine learning model, consisting of a combination of a setpoint function model and a Fourier transform model, is configured to detect installation errors in response to receiving extracted features from a first feature extraction unit and a second feature extraction unit. The machine learning model has been previously trained to detect installation errors.
[0038] According to one aspect of the invention, the sensor is selected from the following: a position sensor, a speed sensor, and a microphone. Attached Figure Description
[0039] Additional features and advantages of the invention will become more apparent from the approximate, but not exclusive, description of preferred, but non-limiting, embodiments of computer-implemented methods for training machine learning models to detect installation errors in elevators, particularly elevator doors, and computer-implemented methods for classifying installation errors, as illustrated in the accompanying drawings, wherein:
[0040] - Figure 1 A system according to the invention for classifying installation errors in elevators, particularly elevator doors, is shown.
[0041] - Figure 2 This is a schematic diagram of the setting function used in a computer-implemented method according to the present invention for classifying installation errors in elevators, especially elevator doors;
[0042] - Figure 3 This is a schematic diagram of the FFT model used in a computer-implemented method according to the present invention for classifying installation errors in elevators, particularly elevator doors;
[0043] - Figure 4 A flowchart illustrating a computer-implemented method according to the present invention for training a machine learning model to detect installation errors in elevators, particularly elevator doors;
[0044] - Figure 5 A flowchart is shown of a computer-implemented method according to the present invention for classifying installation errors in elevators, particularly elevator doors. Detailed Implementation
[0045] Referring to the accompanying drawing, number 100 indicates a computer-implemented method for training a machine learning model to detect installation errors in elevators, particularly elevator doors.
[0046] This method can be applied to elevator landing doors or elevator car doors.
[0047] Various detectable installation errors exist in elevators.
[0048] According to a preferred embodiment, method 100 can identify two types of installation errors: binary faults and faults measured as a percentage.
[0049] Binary faults are indicated as "present" or "not present".
[0050] Two-factor faults include equipment shutdown failures and pulley contact with the belt.
[0051] Faults, measured as a percentage, are assessed within a range.
[0052] Among the faults measured as a percentage, the following are listed: reverse roller installation, horizontal misalignment of elevator floor / car door, vertical misalignment of elevator floor / car door, belt tension, zero position, etc.
[0053] like Figure 4 As shown, method 100 begins with the step of arranging multiple sensors 2 at the elevator (step 101).
[0054] Method 100 is actually based on measurements of physical parameters detected by sensors 2 located at different positions in the elevator 1 (step 102), for example, whether they are operationally valid on the door 10 of the elevator 1.
[0055] Sensor 2 can be of different types and numbers. For example, it can include a door speed sensor, a door position sensor, and a microphone.
[0056] Preferably, measurements are collected for a large number of gate loops, for example, 20,000.
[0057] The frequency of data collection can vary depending on the type of sensor.
[0058] Specifically, some sensors detect static values (i.e., values that do not change during the cycle). For example, static values are related to characteristics of a door, such as its width, material, motor type, etc.
[0059] Other sensors detect values that can change cyclically (meaning these values change periodically) (i.e., temperature, friction, vibration, etc.).
[0060] Other sensors detect audio samples. These may be related to the sound of a moving door, a clicking relay, etc.
[0061] Other sensors detect values that change at a high frequency within the gate period, thus generating a time series.
[0062] All values detected by the sensors form a database, which includes at least one time series.
[0063] For example, the dataset used to train method 10 includes:
[0064] - List of static values;
[0065] - A list of loop values;
[0066] -List of audio samples;
[0067] - Time series matrix.
[0068] Then, method 100 includes:
[0069] - The step of obtaining the first input layer for the machine learning model by extracting features from the dataset (step 103)
[0070] - The step of obtaining the second input layer for the machine learning model by extracting features from the dataset (step 104).
[0071] Features or labels are selected based on installation errors identified by the model.
[0072] Specifically, the features extracted to generate the first input layer are related to the time series of the dataset.
[0073] The features extracted to generate the second input layer are related to other fields of the dataset (static values, loop values, audio samples).
[0074] For audio samples, feature extraction is performed to transform the audio samples into a visual feature representation similar to an audio spectrogram.
[0075] The first approach to extracting features from audio samples is to use an autoencoder to learn latent feature vectors from the image by reconstructing the image itself (unsupervised learning).
[0076] The second approach uses a convolutional neural network to classify images using the collected labels, and fixes one of the last hidden layers as an additional feature vector for classification by both models. These features are then added to both models in a sequential step.
[0077] The extracted features are then fed into a machine learning model, following two different branches at the same level.
[0078] In fact, machine learning models consist of two parallel branches:
[0079] - First model, i.e., the defined function model
[0080] - The second model, namely the Fourier transform model
[0081] The first input layer is fed using a defined function model (step 105).
[0082] The second input layer is fed using the Fourier transform model (step 106).
[0083] Referring to the first model (i.e., the defined function model), the first input layer is obtained after preprocessing, including a normalization step.
[0084] The specification function for time series classification (SeFT) is a state-of-the-art method for time series classification and regression.
[0085] This method operates on the original (normalized) time series as a set and can process additional features that appear once in each time series.
[0086] In principle, time information is encoded via location coding, and a weighted average (possibly multiple) of the measurements is calculated. The weights are trained using an attention mechanism.
[0087] After these steps, a fixed-size vector is obtained, which describes the time series in a lower dimension.
[0088] This vector is concatenated with other static, cyclic, and possibly audio feature vectors to form the input to a classifier / regressor (such as a neural network).
[0089] It should also be remembered that the objective function is designed to handle both binary and bounded regression objectives. To achieve this, binary values are placed in {0, 1} and the regression values are normalized from [-100%, +100%] to [-1, 1].
[0090] SeFT training is performed by applying binary cross-entropy.
[0091] The training time is approximately one day, and the number of gate cycles used for training is approximately 20,000.
[0092] Figure 2 A schematic diagram is shown of the setting function used in the method 100 presented herein.
[0093] Referring to the second model (i.e., the Fourier transform model), a second input layer is obtained after preprocessing, which includes interpolating the values to obtain a uniformly sampled time series, converting the signal into a phase map, and normalizing the static features of the classifier.
[0094] The Fourier transform converts a time-series signal into its frequency component. To do this, a gate velocity-gate position phase diagram is generated and modified to form a sine curve for the Fourier transform.
[0095] For each sensor, the frequency with the largest coefficient and its corresponding coefficient are extracted as a two-element vector. All these vectors are concatenated with other cyclic features to form a feature column, which is then fed into the Fourier transform model.
[0096] Preferably, a so-called "XG starter" is used (which is an implementation of a gradient-boosted decision tree designed for speed and performance). A decoupled implementation uses multiple XG starter classifiers (where hinge loss is used as the loss function for binary classification) and squared loss is used as the loss function for regression.
[0097] Interpolate the irregularly sampled time series X to obtain p equidistant values.
[0098] Fourier transform training is performed by applying hinge loss to the classification target.
[0099] The training time is approximately 3 minutes, and the number of gate cycles used for training is approximately 20,000.
[0100] Figure 3 A schematic diagram of the FFT model used in the method 100 presented herein is shown.
[0101] Reference Figure 5 Number 200 indicates a method for detecting installation errors in elevators, particularly elevator doors. Method 200 includes the following steps: Figure 5 As shown in the flowchart:
[0102] - Multiple physical parameters are detected by means of sensors placed at the elevator (step 201).
[0103] - Perform first feature extraction and second feature extraction based on the detected physical parameters (step 202);
[0104] - Feed the first and second extracted features into the trained machine learning model to detect installation errors (step 203).
[0105] Reference Figure 1 Number 300 indicates a system for detecting installation errors in elevator 1, particularly elevator door 10, the system comprising:
[0106] - Multiple sensors 2, multiple sensors are arranged at elevator 1 and configured to detect physical parameters;
[0107] - A first feature extraction unit 3 and a second feature extraction unit 4, the first feature extraction unit and the second feature extraction unit being configured to extract features from the detected physical parameters;
[0108] - The machine learning model ML, which is a combination of the set function model SF and the Fourier transform model FT, is configured to detect installation errors in response to receiving features extracted from the first feature extraction unit 3 and the second feature extraction unit 4, wherein the machine learning model ML has been trained according to the above method.
[0109] According to one aspect of the invention, the detected physical parameters may be preprocessed before being fed to the first feature extraction unit 3 and the second feature extraction unit 4.
[0110] According to one aspect of the invention, the outputs of the first extraction unit 3 and the second extraction unit 4 can be processed before being fed into the machine learning model ML.
[0111] According to the present invention, the characteristics and advantages of the computer-implemented method for training a machine learning model to detect installation errors in elevators, especially elevator doors, and the computer-implemented method for classifying installation errors, as well as the system thereof, are clear, and the benefits are also clear.
[0112] Specifically, the proposed method allows for improved installation quality, particularly for elevator doors, thanks to the vast amount of data retrieved from sensors directly installed near the door and the chosen specific machine learning model.
[0113] This reduces installation time and costs and allows for the scheduling of maintenance operations and monitoring, as well as reducing the time required to obtain quality certification.
[0114] The proposed invention is also applicable to other components of elevators.
Claims
1. A computer-implemented method (100) for training a machine learning model to detect installation errors in an elevator, said machine learning model being a combination of a setpoint function model and a Fourier transform model, said method (100) comprising the following steps: - Multiple sensors (101) are arranged at the elevator, each sensor being configured to detect physical parameters; - The values of the physical parameters are detected by means of the sensor in order to obtain a dataset (102) including at least one time series. - The first input layer (103) is obtained by extracting features from the dataset; - The second input layer (104) is obtained by extracting features from the dataset; - The set function model (105) is fed in using the first input layer. - The Fourier transform model (106) is fed using the second input layer.
2. The method (100) according to claim 1, wherein the method (100) is used to train a machine learning model to detect installation errors in elevator doors.
3. The method (100) according to claim 1, wherein the dataset comprises a time series matrix.
4. The method (100) according to claim 1 or 3, wherein the dataset further includes one or more of static values or cyclic values.
5. The method (100) of claim 1, wherein the dataset further comprises audio samples.
6. The method (100) according to claim 5, wherein the extracted features of the audio sample include an audio spectrogram.
7. The method (100) according to claim 1, wherein the step (102) of detecting the value of the physical parameter by means of the sensor is performed at different cycles depending on the sensor type and the type of physical parameter involved.
8. A computer-implemented method (200) for detecting installation errors in an elevator, the method (200) comprising the following steps: - Multiple physical parameters are detected by means of sensors arranged at the elevator (201); - Perform first feature extraction and second feature extraction based on the detected physical parameters (202); - The extracted first feature and the extracted second feature are fed into the machine learning model trained according to any one of claims 1 to 7 to detect installation errors (203).
9. The method (200) according to claim 8, wherein the method (200) is used to detect installation errors in elevator doors.
10. A system (300) for detecting installation errors in an elevator (1), comprising: - Multiple sensors (2), the multiple sensors are arranged at the elevator (1) and configured to detect physical parameters; - A first feature extraction unit (3) and a second feature extraction unit (4), wherein the first feature extraction unit and the second feature extraction unit are configured to extract features from the detected physical parameters; - A machine learning model (ML), which is composed of a combination of a setpoint function model (SF) and a Fourier transform model (FT), configured to detect installation errors in response to receiving extracted features from the first feature extraction unit (3) and the second feature extraction unit (4), the machine learning model (ML) being trained according to any one of claims 1 to 7 to detect installation errors.
11. The system (300) of claim 10, wherein the system (300) is used to detect installation errors in elevator doors.
12. The system (300) according to claim 10, wherein the sensor (2) is selected from the following: a position sensor, a speed sensor, and a microphone.
13. A software product capable of being loaded into the memory of an electronic device, the software product comprising instructions that, when executed by the electronic device, determine the execution of the steps of the method according to claim 8.