Elevator rotating member failure detection method, electronic device, and program product

Abstract

The application provides an elevator rotating part fault detection method, an electronic device and a program product. The method comprises: obtaining to-be-detected vibration data of a to-be-detected elevator rotating part; performing frequency domain feature extraction on the preprocessed to-be-detected vibration data to obtain to-be-detected features; mapping the to-be-detected features from a target domain to mapping features of a source domain based on a pre-established cross-domain mapping relationship through a self-encoder model; inputting the mapping features into a trained classification model to obtain a detection result of the to-be-detected elevator rotating part, the classification model being trained through a first vibration data set of a source structure, and the first vibration data set comprising a fault data set and a non-fault data set of the source structure. In this way, the fault detection of the elevator rotating part can be realized without using the fault data of the elevator rotating part for model training, and the problems of false positives and false negatives caused by the lack of fault data of the elevator rotating part and the failure of the model to learn fault features are avoided.

Description

Technical Field

[0001] This invention relates to the field of elevator testing technology, and more specifically, to a method, electronic equipment, and program product for detecting faults in rotating elevator components. Background Technology

[0002] As a vertical transportation facility in cities, the health of elevators' rotating components, such as the traction machine, door operator system, and guide wheels, directly affects operational safety. Under long-term dynamic loads, these rotating components are prone to faults such as bearing wear, gear damage, and rotor imbalance. Failure to detect these faults in a timely manner can lead to shutdowns or safety accidents. Fault detection of elevator rotating components can be achieved by collecting vibration signals and using appropriate deep learning models. However, due to the limited amount of fault data for elevator rotating components, it is often impossible to collect sufficient data, leading to issues such as false alarms and missed alarms in deep learning models. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, electronic device and program product for detecting faults in elevator rotating parts, which can improve the problem of false alarms and missed alarms in the detection of elevator rotating parts.

[0004] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0005] In a first aspect, embodiments of this application provide a method for detecting faults in rotating components of an elevator, the method comprising:

[0006] Acquire the vibration data of the rotating components of the elevator under test;

[0007] The preprocessed vibration data to be measured is subjected to frequency domain feature extraction to obtain the measured features;

[0008] Using an autoencoder model, based on a pre-established cross-domain mapping relationship, the features to be measured are mapped from the target domain to the features to be measured in the source domain, which are used as mapping features. The target domain represents the elevator rotating component, and the source domain represents the source structure. The source structure is an experimental test platform with similar features to the elevator rotating component. The cross-domain mapping relationship is obtained based on the fault-free dataset and the second vibration dataset in the first vibration dataset of the source structure. The second vibration dataset is the vibration data of the elevator rotating component in a fault-free state.

[0009] The mapping features are input into the trained classification model to obtain the detection results of the rotating component of the elevator under test. The detection results include results indicating that there is no fault or that there is a fault. The classification model is trained using a first vibration dataset of the source structure. The first vibration dataset includes a fault dataset and a fault-free dataset of the source structure.

[0010] In conjunction with the first aspect, in some optional embodiments, before acquiring the vibration data of the rotating component of the elevator under test, the method further includes:

[0011] Obtain a first vibration dataset of the source structure and a second vibration dataset of the elevator rotating component in a fault-free state. The first vibration dataset includes a fault-free dataset representing the vibration of the source structure in a fault-free state and a fault dataset representing the vibration in a fault state.

[0012] Frequency domain features are extracted from the preprocessed first vibration dataset and second vibration dataset to obtain a first feature set and a second feature set, respectively. The first feature set includes the fault feature set corresponding to the fault dataset and the fault-free feature set corresponding to the fault-free dataset.

[0013] Using the second feature set as input data and the fault-free feature set as label data, a cross-domain mapping relationship is established through an autoencoder model to transform the second feature set into a feature set whose distribution is consistent with that of the fault-free feature set.

[0014] The classification model is trained using the fault feature set and the fault-free feature set from the first feature set until the classification model converges, resulting in a well-trained classification model.

[0015] In conjunction with the first aspect, in some optional embodiments, between the step of obtaining the first vibration dataset of the source structure and the second vibration dataset of the elevator rotating component in a fault-free state, and the step of performing frequency domain feature extraction on the preprocessed first and second vibration datasets, the method further includes:

[0016] The first vibration dataset and the second vibration dataset are preprocessed to obtain preprocessed first vibration dataset and second vibration dataset, wherein the preprocessing includes:

[0017] The first vibration dataset and the second vibration dataset are filtered and denoised to obtain the denoised first vibration dataset and the second vibration dataset.

[0018] The first and second vibration datasets after noise reduction are divided according to the multiple load levels of the elevator to obtain a sample subset corresponding to each load level.

[0019] Adopt a data augmentation strategy to expand the sample subset, and the data augmentation strategy includes performing at least one of time stretching, Gaussian noise addition, and time domain translation on part of the data in the sample subset.

[0020] Combined with the first aspect, in some optional embodiments, the frequency domain feature extraction of the preprocessed first vibration data set and second vibration data set respectively obtains a first feature set and a second feature set, including:

[0021] Perform a fast Fourier transform on the vibration data of any vibration signal in the preprocessed first vibration data set and second vibration data set to obtain the amplitude of the frequency domain signal, expressed as:

[0022]

[0023] In the formula, is the amplitude of the frequency domain signal; f represents frequency; n is the time series index, and 0 ≤ n < N, where N is the number of signal sampling points; is the vibration data of any vibration signal at the nth sampling point; j is the imaginary unit; [[ID=1&]] is the natural constant;

[0024] If the vibration signal is a signal in a fault state, according to the preset frequency band characterizing fault sensitivity, extract all the amplitude values of the frequency domain signals within the preset frequency band from all the amplitude values of the frequency domain signals of the vibration signal to form fault features; [[ID=&4]]

[0025] If the vibration signal is a signal in a fault-free state, take the amplitude value of the frequency domain signal corresponding to the vibration signal as the fault-free feature;

[0026] Take all the fault features and all the fault-free features obtained based on the first vibration data set as the first feature set; take all the fault-free features obtained based on the second vibration data set as the second feature set.

[0027] Combined with the first aspect, in some optional embodiments, the autoencoder model includes an encoder and a decoder;

[0028] The establishment of the cross-domain mapping relationship by using the second feature set as input data and the fault-free feature set as label data through the autoencoder model includes:

[0029] Input the features corresponding to each vibration signal in the second feature set into the encoder to compress the features of the target domain into the latent space, obtain the features in the latent space, and reconstruct, through the decoder, features consistent with the feature distribution of the source domain based on the features in the latent space as the reconstructed features;

[0030] Based on the reconstructed features and corresponding label data, the parameters of the autoencoder model are optimized using a first preset loss function. The autoencoder model is then iteratively trained using the second feature set and the fault-free feature set until it converges. This enables the autoencoder model to convert target domain features into features aligned with the source domain feature distribution, serving as the established cross-domain mapping relationship. The first preset loss function is:

[0031]

[0032] In the formula, X represents the reconstruction error; Y represents the reconstruction feature; M represents the label data corresponding to the reconstruction feature, which is the fault-free feature of the source domain; M and Q represent the number of samples in the fault-free feature set and the number of samples in the second feature set, respectively. Refers to the reconstructed feature corresponding to the i-th feature in the second feature set; It refers to the j-th feature in the set of fault-free features.

[0033] In conjunction with the first aspect, in some optional implementations, the step of training a classification model using the fault feature set and the fault-free feature set from the first feature set until the classification model converges to obtain a trained classification model includes:

[0034] A41, using stratified sampling, the first feature set is divided into a training set, a validation set, and a test set according to a preset ratio;

[0035] A42, using the training set, iteratively train the classification model to obtain a preliminary trained classification model;

[0036] A43, the grid search method is used to traverse the adjustable parameters in the classification model to obtain multiple parameter sets, wherein the adjustable parameters include kernel parameters and regularization parameters. Each parameter set and the initially trained classification model form a candidate classification model. The candidate classification model is then optimized using the validation set, and the candidate classification model with the highest fault identification accuracy is selected as the optimized classification model.

[0037] A44. Using the test set, the optimized classification model is tested to obtain test results indicating whether the classification model's performance is acceptable.

[0038] A45. If the test result is unqualified, repeat steps A43 and A44 until the test result is qualified and a trained classification model is obtained.

[0039] A46. When the test results are satisfactory, a trained classification model is obtained.

[0040] In conjunction with the first aspect, in some optional implementations, the iterative training process includes:

[0041] A batch of samples is selected from the training set, the batch of samples including evenly distributed fault-free features and fault features;

[0042] The batch samples are used as the input feature vector for the current round and input into the initial classification model;

[0043] The radial basis function (RBF) kernel is invoked through the classification model to calculate the similarity between each feature in the batch of samples and the features of the baseline samples. The formula for the RBF kernel is as follows:

[0044]

[0045] In the formula, for and The similarity value; The currently input feature in the batch of samples; γ represents the pre-stored features of the i-th benchmark sample; γ is the kernel parameter. It is an exponential function;

[0046] By using a classification model, based on the similarity of the batch samples, the features in the batch samples are mapped to a high-dimensional space to obtain high-dimensional features;

[0047] The optimal decision boundary in the high-dimensional space is determined by minimizing the objective function, which is:

[0048]

[0049] The constraints of the objective function are:

[0050]

[0051] In the formula, For high-dimensional space weights, The high-dimensional features are the features in the batch of samples after mapping. This represents the offset of the decision boundary. As slack variables, The total number of samples in the training set. For regularization parameters;

[0052] Based on the optimal decision boundary, the decision function is obtained, and the decision function is:

[0053]

[0054] Where f(x) represents the value of the decision function. At that time, it was determined to be in a fault-free state. When this occurs, it is determined to be a fault state; U is the sign function; U is the total number of support vectors involved in constructing the decision boundary. The Lagrange coefficients obtained during training when the decision boundary is optimal;

[0055] The loss value of the classification model is backpropagated using the gradient descent method to update the relevant parameters of the classification model until a preset stopping condition is met, thus obtaining a pre-trained classification model. The preset stopping condition includes a single round loss value being lower than a first preset threshold, or a continuous R round loss value decreasing by a factor lower than a second preset threshold, where R is a preset integer greater than or equal to 5. The loss value is obtained based on the value output by the decision function.

[0056] In conjunction with the first aspect, in some optional embodiments, the elevator rotating component includes at least one of the elevator's traction machine, door operator system, guide wheels, and counterweight wheels, and the method further includes:

[0057] When the detection result indicates a fault, a fault prompt is sent to the management terminal.

[0058] Secondly, embodiments of this application also provide an electronic device, which includes a processor and a memory coupled to each other. The memory stores a computer program, and when the computer program is executed by the processor, the electronic device performs the above-described method.

[0059] Thirdly, embodiments of this application also provide a program product, including a computer program, which, when executed by a processor, implements the above-described method.

[0060] The invention employing the above technical solution has the following advantages:

[0061] In the technical solution provided in this application, an autoencoder model is used to convert the features to be tested corresponding to the rotating components of the elevator under test into the features to be tested in the source domain using cross-domain mapping relationships. Then, a classification model trained based on the first vibration dataset (including fault and non-fault data) corresponding to the source domain is used to detect the features to be tested in the source domain, thereby realizing the fault detection of the corresponding rotating components of the elevator. In this way, it is not necessary to use the fault data of the rotating components of the elevator for model training. Instead, the classification model trained on the complete vibration dataset of the source domain is used to detect the features in the source domain (the features to be tested of the rotating components of the elevator after the target domain-source domain mapping). This avoids the problem of false alarms and missed alarms caused by the lack of fault data of the rotating components of the elevator, which prevents the model from learning the fault features. Attached Figure Description

[0062] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0063] Figure 1 This is a schematic diagram of an elevator rotating component fault detection method provided in an embodiment of this application.

[0064] Figure 2 This is a schematic diagram of the structure of the AE model provided in the embodiments of this application. Detailed Implementation

[0065] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0066] The following is a definition of some terms used in this application:

[0067] Target domain: refers to the rotating parts of the elevator in operation, that is, the rotating parts of the elevator in actual operation and the domain to which they belong.

[0068] Source domain: refers to the source structure, that is, the experimental test platform and its associated field that possesses similar dynamic characteristics to the rotating components of an elevator. The experimental test platform can be used to simulate the normal (fault-free) operation of an elevator under different loads, as well as its operation under fault conditions.

[0069] Cross-domain mapping relationship: A correspondence obtained through autoencoder training that can transform the features of the target domain into a distribution consistent with the features of the source domain.

[0070] Autoencoder (AE) model: A neural network model that includes an encoder and a decoder, used to achieve cross-domain feature distribution alignment.

[0071] Encoder: The first half of the AE model network, whose function is to compress the high-dimensional input features of the target domain into a low-dimensional latent space, thereby achieving feature dimensionality reduction and key information extraction.

[0072] Decoder: The second half of the AE model network, whose function is to reconstruct high-dimensional features that are consistent with the feature distribution of the source domain from low-dimensional latent space features, thereby achieving cross-domain feature alignment.

[0073] Latent space: The low-dimensional feature space of the encoder output. It is an intermediate feature representation connecting the input and output, which can retain core feature information and remove redundancy.

[0074] Classification models can be one-class support vector machines (OC-SVM) supporting fault and non-fault identification, or multi-class support vector machines (Multi-Class SVM) or convolutional neural networks (CNN) supporting multi-class fault identification. The specific type of classification model is not limited here, as long as it meets the corresponding fault identification requirements. It should be understood that OC-SVM supports binary classification, i.e., detecting faults and non-faults. Multi-class support vector machines and convolutional neural networks can support multi-class fault identification; that is, in addition to identifying faults and non-faults, they can further identify fault types, such as traction machine bearing damage or guide wheel failure.

[0075] The first vibration dataset is a collection of the fault-free dataset and the fault dataset of the source structure under fault-free conditions and fault conditions, respectively. The fault condition can be called the damaged condition, and the fault-free condition can be called the undamaged condition.

[0076] Second vibration dataset: Vibration dataset collected from actual elevator rotating parts under fault-free conditions.

[0077] First feature set: The feature set extracted based on the first vibration dataset of the source structure, including the fault feature set (sensitive frequency band amplitude features of fault state signals) and the fault-free feature set (complete frequency domain amplitude features of fault-free state signals).

[0078] The second feature set is a set of fault-free features extracted from the second vibration dataset of the elevator rotating components. It should be noted that the features in both the first and second feature sets are feature vectors, used collectively to characterize the frequency domain vibration patterns of the corresponding components.

[0079] Load Class: The operating load category based on the elevator's rated load capacity. As an example, the load class can include no load, half load, and full load, covering the main operating scenarios of the elevator.

[0080] Data augmentation strategies: These are techniques that generate new samples by reasonably transforming the original samples, in order to solve the problem of insufficient sample size and improve the generalization ability of the model.

[0081] Sample subset: A local data set obtained by dividing the data according to the load level. Each subset corresponds to vibration data under a certain load scenario.

[0082] Filtering and denoising: This involves using signal processing algorithms to remove irrelevant noise from the original vibration data in order to preserve fault-sensitive signals.

[0083] Please refer to Figure 1 This application provides a method for detecting faults in rotating elevator components, which can be applied to electronic devices and whose steps can be executed or implemented by the electronic devices. The electronic devices can be, but are not limited to, personal computers, servers, etc. The method for detecting faults in rotating elevator components may include the following steps:

[0084] Step S10: Obtain the vibration data of the rotating component of the elevator under test;

[0085] Step S20: Extract frequency domain features from the preprocessed vibration data to obtain the features to be measured.

[0086] Step S30: Using an autoencoder model, based on a pre-established cross-domain mapping relationship, the feature to be tested is mapped from the target domain to the feature to be tested in the source domain, which is used as the mapping feature. The target domain represents the elevator rotating component, and the source domain represents the source structure. The source structure is an experimental test platform with similar features to the elevator rotating component. The cross-domain mapping relationship is obtained based on the fault-free dataset and the second vibration dataset in the first vibration dataset of the source structure. The second vibration dataset is the vibration data of the elevator rotating component in a fault-free state.

[0087] Step S40: Input the mapped features into the trained classification model to obtain the detection results of the rotating component of the elevator under test. The detection results include results indicating whether there is no fault or a fault exists. The classification model is trained using the first vibration dataset of the source structure. The first vibration dataset includes the fault dataset and the fault-free dataset of the source structure.

[0088] The following is a detailed explanation of each step in the method for detecting faults in the rotating parts of an elevator:

[0089] The rotating components of an elevator may include, but are not limited to, at least one of the following: traction machine (elevator power source, including rotating components such as motors and reducers), door operator system (rotating mechanism that controls the opening and closing of elevator doors), guide wheel (rotating component that guides the movement of steel wire rope), and counterweight wheel (rotating component that balances the weight of the car).

[0090] Prior to step S10, the method may further include steps such as establishing cross-domain mapping relationships and model training. For example, before acquiring the vibration data of the rotating component of the elevator under test in step S10, the method may further include:

[0091] A10, obtain the first vibration dataset of the source structure and the second vibration dataset of the elevator rotating component in a fault-free state. The first vibration dataset includes a fault-free dataset representing the vibration of the source structure in a fault-free state and a fault dataset representing the vibration in a fault state.

[0092] A20, frequency domain features are extracted from the preprocessed first vibration dataset and second vibration dataset to obtain a first feature set and a second feature set, respectively. The first feature set includes the fault feature set corresponding to the fault dataset and the fault-free feature set corresponding to the fault-free dataset.

[0093] A30, using the second feature set as input data and the fault-free feature set as label data, a cross-domain mapping relationship is established through an autoencoder model to transform the second feature set into a feature set with a distribution consistent with the fault-free feature set;

[0094] A40, using the fault feature set and the fault-free feature set in the first feature set, train the classification model until the classification model converges, and obtain the trained classification model.

[0095] Specifically, the source structure is an experimental test platform with similar characteristics to the rotating parts of an elevator. Vibration data of the fault-free state and simulated fault states (such as bearing damage and gear wear) are collected by sensors to form the first vibration dataset. The fault-free vibration data of the elevator rotating parts are collected by sensors deployed on actual elevators to form the second vibration dataset.

[0096] The vibration data collected at each sampling point includes frequency, amplitude, and phase information. Furthermore, the raw data can be standardized in format (e.g., by unifying sampling frequency and data length) and invalid data (such as abrupt changes in data collected by the sensor) can be removed. The vibration data is collected using a conventional method.

[0097] Between step A10, which obtains the first vibration dataset of the source structure and the second vibration dataset of the elevator rotating component under fault-free conditions, and step A20, which extracts frequency domain features from the preprocessed first and second vibration datasets, the method further includes:

[0098] The first vibration dataset and the second vibration dataset are preprocessed to obtain preprocessed first vibration dataset and second vibration dataset, wherein the preprocessing includes:

[0099] The first vibration dataset and the second vibration dataset are filtered and denoised to obtain the denoised first vibration dataset and the second vibration dataset.

[0100] The first and second vibration datasets after noise reduction are divided according to the multiple load levels of the elevator to obtain sample subsets corresponding to each load level.

[0101] A data augmentation strategy is adopted to expand the sample subset. The data augmentation strategy includes at least one operation among time stretching, Gaussian noise addition, and time-domain shifting on some data in the sample subset.

[0102] Specifically, the filtering and noise reduction methods can be wavelet filtering or adaptive filtering algorithms, setting the frequency band corresponding to the normal operation of the elevator and the fault-sensitive frequency band (such as 10-1000Hz) as the effective frequency band, filtering out irrelevant signals such as environmental noise and electromagnetic interference outside the frequency band, and retaining the vibration signals related to the operating status of the elevator's rotating parts.

[0103] Example of elevator load level classification: Unloaded: 0% of rated load; Half-loaded: (0, 50%) of rated load; Full-loaded: (50%, 100%) of rated load. In other embodiments, load levels can be further divided into finer granularities. For the denoised first and second vibration data, based on the load information during elevator operation, the vibration data are classified according to the corresponding load levels, forming a sample subset for each load level, ensuring that each subset contains a sufficient number of samples. Then, the sample subsets for each load level are passed to the data augmentation step.

[0104] For sample subsets with small or insufficient sample sizes, at least one data augmentation operation is employed: For the vibration data corresponding to the vibration signals in the sample subset, time stretching (stretching or compressing the signal length by 10%-20% while maintaining frequency characteristics), Gaussian noise addition (adding Gaussian noise with a mean of 0 and a variance of 0.01-0.05 to the signal), and time-domain shifting (shifting the signal on the time axis by 5%-10% to avoid boundary effects). The vibration data obtained after data augmentation is added to the original sample subset, forming an expanded sample subset, thus obtaining the preprocessed first and second vibration datasets. The expanded sample subset will then be passed to the frequency domain feature extraction step.

[0105] In this embodiment, filtering and denoising helps improve the signal-to-noise ratio of vibration data, prevents irrelevant noise from masking fault features, and provides high-quality data for subsequent feature extraction; dividing samples according to load level enables the classification model to adapt to the feature differences under different elevator loads, improving the comprehensiveness of detection; data augmentation expands the sample diversity, solves the overfitting problem of the classification model caused by insufficient sample size, and enhances the robustness of the model.

[0106] Step A20 involves extracting frequency domain features from the preprocessed first and second vibration datasets to obtain a first feature set and a second feature set, which may include:

[0107] Perform a Fast Fourier Transform (FFT) on the vibration data of any vibration signal in the pre - processed first vibration data set and second vibration data set to obtain the amplitude of the frequency - domain signal, which is expressed as:

[0108] (1)

[0109] In the formula, is the amplitude of the frequency - domain signal; f represents frequency; n is the time - series index, and 0 ≤ n < N, where N is the number of signal sampling points; is the vibration data of any vibration signal at the n - th sampling point; j is the imaginary unit; is the natural constant;

[0110] If the vibration signal is a signal in a fault state, according to the preset frequency band characterizing fault sensitivity, extract all the amplitude values of the frequency - domain signals within the preset frequency band from all the amplitude values of the frequency - domain signals of the vibration signal to form fault features;

[0111] If the vibration signal is a signal in a fault - free state, take the amplitude value of the frequency - domain signal corresponding to the vibration signal as the fault - free feature;

[0112] Take all the fault features and all the fault - free features obtained based on the first vibration data set as the first feature set; take all the fault - free features obtained based on the second vibration data set as the second feature set.

[0113] Specifically, for the vibration data corresponding to a single vibration signal in the pre - processed first and second vibration data sets, perform FFT transformation according to the above formula (1) to calculate the amplitude of the frequency - domain signal corresponding to each frequency. Then, for the amplitude of the frequency - domain signal of the vibration signal in the fault state, screen out all the amplitude data within the preset frequency band sensitive to faults from the amplitude values of the frequency - domain signals in the full frequency band, arrange them in the order of frequency to form fault features, and汇入 the fault feature vectors into the first feature set. The preset frequency band sensitive to faults can be determined according to the characteristic frequencies corresponding to common faults of elevator rotating components (such as bearing damage, gear wear, rotor imbalance), that is, it can be flexibly set according to the actual situation. As an example, for the inner ring of the traction machine bearing, it corresponds to 60 - 300 Hz, and for gear wear, it corresponds to 500 - 1000 Hz.

[0114] For the amplitude of the frequency - domain signal of the vibration signal in the fault - free state, retain the amplitude values of the frequency - domain signals in the full frequency band, arrange them in the order of frequency to form fault - free features. Among them, the fault - free feature vectors corresponding to the source structure (the first vibration data set) are汇入 the first feature set; the fault - free features corresponding to the elevator rotating components (the second vibration data set) are used as the second feature set. The first feature set is used for training the classification model, and the second feature set is used for training the AE model.

[0115] In this embodiment, the time-domain signal is converted into a frequency-domain signal by FFT transformation, highlighting the characteristic frequency corresponding to the fault and solving the problem of unclear fault characteristics in the time-domain signal; the fault characteristics are focused on sensitive frequency bands, reducing interference from irrelevant frequency band data, which helps to improve the identification of fault characteristics.

[0116] In this embodiment, the autoencoder model includes an encoder and a decoder. As an example, the encoder uses a 3-layer fully connected neural network (input layer → hidden layer 1 → hidden layer 2 → latent space layer), with ReLU activation function for each layer. The decoder uses a 3-layer fully connected neural network (latent space layer → hidden layer 3 → hidden layer 4 → output layer), with Sigmoid activation function for the output layer to ensure that the output features are consistent with the dimension of the fault-free features in the source domain.

[0117] Step A30, using the second feature set as input data and the fault-free feature set as label data, establishes a cross-domain mapping relationship through an autoencoder model, which may include:

[0118] The features corresponding to each vibration signal in the second feature set are input into the encoder to compress the features of the target domain into the latent space to obtain the features of the latent space. The decoder then reconstructs features consistent with the feature distribution of the source domain based on the features of the latent space, which are used as the reconstructed features.

[0119] Based on the reconstructed features and corresponding label data, the parameters of the autoencoder model are optimized using a first preset loss function. The autoencoder model is then iteratively trained using the second feature set and the fault-free feature set until it converges. This enables the autoencoder model to convert target domain features into features aligned with the source domain feature distribution, serving as the established cross-domain mapping relationship. The first preset loss function (mean squared error loss function) is:

[0120] (2)

[0121] In the formula, X represents the reconstruction error; Y represents the reconstruction feature; M represents the label data corresponding to the reconstruction feature, which is the fault-free feature of the source domain; M and Q represent the number of samples in the fault-free feature set and the number of samples in the second feature set, respectively. Refers to the reconstructed feature corresponding to the i-th feature in the second feature set; It refers to the j-th feature in the set of fault-free features.

[0122] Specifically, each feature vector in the second feature set (elevator fault-free feature vector) is input into the encoder one by one. Through hierarchical dimensionality reduction (e.g., input dimension 500 → hidden layer 1 dimension 128 → hidden layer 2 dimension 64 → latent space dimension 32), low-dimensional latent space features are obtained. After the latent space features are input into the decoder, the decoder performs hierarchical dimensionality increase on the latent space features (latent space dimension 32 → hidden layer 3 dimension 64 → hidden layer 4 dimension 128 → output dimension 500), reconstructing features consistent with the distribution of fault-free features in the source domain. Next, the reconstructed features and label data (the corresponding fault-free features in the first feature set) are input into the first preset loss function to calculate the reconstruction error. The Adam optimizer (learning rate 0.001) is used to backpropagate the error and update the weight parameters of the encoder and decoder.

[0123] After each training round, the reconstruction accuracy is verified using a validation set (extracted from the second feature set and the source domain fault-free feature set). When the reconstruction error is lower than the corresponding threshold (e.g., 0.005) for 10 consecutive rounds, the AE model converges. At this time, the AE model has the function of cross-domain feature mapping and can be called the "target-source domain feature mapping model", and the cross-domain mapping relationship is established.

[0124] The first preset loss function is used to measure the difference between the reconstructed features and the fault-free features of the source domain, guiding the optimization of model parameters. Model convergence means that the reconstruction error of the AE model stabilizes at a low level and no longer decreases significantly with training epochs. At this point, the cross-domain mapping effect of the model reaches its optimal level.

[0125] In this embodiment, the AE model with encoder-decoder architecture can achieve feature distribution alignment between the target domain and the source domain, solving the problem of model transfer difficulties caused by cross-domain feature differences; minimizing the loss function ensures the high similarity between the reconstructed features and the source domain features, providing a foundation for accurate recognition by the subsequent classification model; the trained AE model has stable cross-domain mapping capabilities, which can adapt to the feature differences of different brands and models of elevators, which is conducive to improving the generalization of this method.

[0126] In this embodiment, the core function of the AE model is to achieve Domain Adaptation (DA). By using an encoder and decoder architecture, the feature distribution gap between the target domain (elevator rotating parts) and the source domain (experimental test platform / source structure) is narrowed, providing a cross-domain adaptation basis for zero-shot diagnosis. It follows the core logic of "encoder dimensionality reduction - decoder reconstruction": the encoder maps the high-dimensional features of the input to a low-dimensional latent space to achieve feature compression and key information extraction; the decoder reconstructs features from the latent space that are consistent with the feature distribution of the source structure without damage, thus completing cross-domain alignment.

[0127] Step A40: Using the fault feature set and the fault-free feature set from the first feature set, train the classification model until the classification model converges, obtaining the trained classification model, including:

[0128] A41, using stratified sampling, the first feature set is divided into a training set, a validation set, and a test set according to a preset ratio;

[0129] A42, using the training set, iteratively train the classification model to obtain a preliminary trained classification model;

[0130] A43, the grid search method is used to traverse the adjustable parameters in the classification model to obtain multiple parameter sets, wherein the adjustable parameters include kernel parameters and regularization parameters. Each parameter set and the initially trained classification model form a candidate classification model. The candidate classification model is then optimized using the validation set, and the candidate classification model with the highest fault identification accuracy is selected as the optimized classification model.

[0131] A44. Using the test set, the optimized classification model is tested to obtain test results indicating whether the classification model's performance is acceptable.

[0132] A45. If the test result is unqualified, repeat steps A43 and A44 until the test result is qualified and a trained classification model is obtained.

[0133] A46. When the test results are satisfactory, a trained classification model is obtained.

[0134] Specifically, for the first feature set (including fault feature set and fault-free feature set), fault features and fault-free features are extracted stratified according to a preset ratio (e.g., 70% for training set, 20% for validation set, and 10% for test set) to ensure that the ratio of fault and fault-free features in each dataset is consistent with the original first feature set, avoiding data distribution imbalance. The training set is used to pass in the classification model training steps, the validation set is used to pass in the parameter tuning steps, and the test set is used to pass in the classification model testing steps.

[0135] After the training set features are input into the initial classification model, the classification model can call the Radial Basis Function (RBF) to calculate feature similarity and map it to a high-dimensional space; the fault-free feature boundary is determined by minimizing the objective function; the loss value is backpropagated using the gradient descent method to update the model parameters, and the model is iterated for 50-100 rounds to obtain the initially trained classification model.

[0136] In step A42, the iterative training process includes:

[0137] A batch of samples is selected from the training set, the batch of samples including evenly distributed fault-free features and fault features;

[0138] The batch samples are used as the input feature vector for the current round and input into the initial classification model;

[0139] The radial basis function (RBF) kernel is invoked through the classification model to calculate the similarity between each feature in the batch of samples and the features of the baseline samples. The formula for the RBF kernel is as follows:

[0140] (3)

[0141] In the formula, for and The similarity value; The currently input feature in the batch of samples; γ represents the pre-stored features of the i-th benchmark sample; γ is the kernel parameter. It is an exponential function;

[0142] By using a classification model, based on the similarity of the batch samples, the features in the batch samples are mapped to a high-dimensional space to obtain high-dimensional features;

[0143] The optimal decision boundary in the high-dimensional space is determined by minimizing the objective function, which is:

[0144] (4)

[0145] The constraints of the objective function are:

[0146] (5)

[0147] In the formula, For high-dimensional space weights, The high-dimensional features are the features in the batch of samples after mapping. This is the offset of the decision boundary; These are slack variables (allowing a small number of samples to deviate from the boundary). The total number of samples in the training set. This is the regularization parameter (the initial value can be 0.5);

[0148] Based on the optimal decision boundary, the decision function is obtained, and the decision function is:

[0149] (6)

[0150] Where f(x) represents the value of the decision function. At that time, it was determined to be in a fault-free state. When the condition is met, it is determined to be a fault state; sign(·) is the sign function, which determines no damage when the output is +1 and fault when the output is -1; U is the total number of support vectors participating in the construction of the decision boundary. Support vectors are sample features (from the training set of the first feature set) that play a key role in determining the optimal decision boundary during the training process of the classification model. The Lagrange coefficients obtained during training when the decision boundary is optimal can be used to solve the Lagrange dual problem to obtain the optimal Lagrange coefficients.

[0151] The loss value of the classification model is backpropagated using the gradient descent method to update the relevant parameters of the classification model until a preset stopping condition is met, thus obtaining a pre-trained classification model. The preset stopping condition includes a single round loss value being lower than a first preset threshold, or a continuous R round loss value decreasing by a factor lower than a second preset threshold, where R is a preset integer greater than or equal to 5. The loss value is obtained based on the value output by the decision function.

[0152] Specifically, in each training round, a batch of samples is selected (e.g., batch size 32, 64), ensuring that the ratio of fault-free features to fault features within the batch is 1:1 (if the training set ratio is unbalanced, oversampling fault features or undersampling fault-free features is used to achieve balance). Each feature vector in the batch samples is used as input. Features of the baseline samples pre-stored in the model The similarity is calculated using the radial basis function kernel formula, where γ is the kernel parameter (initial value 0.1). Then, through the nonlinear mapping of the radial basis function kernel, the low-dimensional input features are transformed into feature vectors in a high-dimensional space. This makes the previously linearly indistinguishable fault and fault-free features distinguishable in a high-dimensional space. Next, based on the aforementioned minimization objective function, the optimal decision boundary is determined, and a decision function is constructed based on this boundary. The output of the decision function, the true labels (fault / fault-free) of the training set samples, can be used to calculate the loss value of the classification model. The loss value of the classification model is backpropagated using gradient descent (learning rate 0.001) to update the model's weight parameters and kernel parameters. With regularization parameters After each training round, check the stopping condition: if the loss value in a single round is lower than the first preset threshold (e.g., 0.01), or if the loss value decreases less than the second preset threshold (e.g., 0.001) for R consecutive rounds (R≥5, preset R=10), then stop training and obtain the preliminary training classification model.

[0153] In this embodiment, the nonlinear mapping to the basis kernel function can solve the problem of indistinguishable fault and fault-free features in low-dimensional space, which is beneficial to improving feature discrimination. Minimizing the objective function ensures the optimality of the decision boundary, which can distinguish fault and fault-free features to the greatest extent. The parameter update and stopping condition setting of the gradient descent method ensure that the classification model converges quickly and has good generalization ability, reducing the risk of overfitting.

[0154] In step A43, the adjustable parameters may include: kernel parameter γ: 0.001-10; regularization parameter ν: 0.01-0.9. All parameter combinations are iterated (e.g., γ takes values ​​of 0.001, 0.01, 0.1, 1, 10; ν takes values ​​of 0.01, 0.1, 0.3, 0.5, 0.7, 0.9), and each parameter combination is applied to the initially trained classification model. The fault identification accuracy is tested using a validation set, and parameter tuning is performed in a conventional manner. The model corresponding to the parameter combination with the highest accuracy is selected as the optimized classification model.

[0155] During parameter tuning, a comprehensive loss function can be constructed, focusing on cross-domain mapping error and fault diagnosis error.

[0156] (7)

[0157] in, For the comprehensive error, This refers to the cross-domain mapping error of the AE model; These are weighting coefficients used to balance the accuracy of cross-domain mapping with the accuracy of fault diagnosis. This represents the classification error of the classification model.

[0158] After each batch of testing is completed, the overall loss value is calculated. If the loss value does not reach the corresponding threshold, adjust the key parameters of the AE model and the classification model; repeat the iterative training until the comprehensive loss value converges to ensure the diagnostic accuracy and stability of the model in the elevator scenario.

[0159] In step A44, the test set features can be input into the optimized classification model to calculate the fault identification accuracy, false alarm rate, and false negative rate. As an example, if the accuracy is ≥95%, the false alarm rate is ≤3%, and the false negative rate is ≤2%, the test result is acceptable; otherwise, it is unacceptable.

[0160] If the test results do not meet the passing standard, expand the parameter search range (e.g., expand γ to 0.0001-100), and repeat the parameter tuning and testing process until the test results are qualified.

[0161] If the test results are satisfactory, the classification model obtained from the test will be used as the trained classification model.

[0162] In this embodiment, stratified sampling avoids model overfitting caused by imbalanced dataset distribution, ensuring the fairness of model training; grid search can efficiently find the optimal parameter combination, significantly improving the fault identification accuracy of the classification model; and validation on the test set ensures that the trained model has stable performance, reducing the risk of false positives and false negatives in practical applications, and providing a guarantee for the reliability of fault detection.

[0163] In step S10, the original vibration signal can be acquired by the sensor at a preset sampling frequency (e.g., 10kHz) and stored in a time series data format as the original vibration data.

[0164] Next, the original vibration data can be filtered and denoised (such as wavelet filtering) to remove irrelevant noise such as car vibration and electromagnetic interference; then, it can be classified into corresponding datasets according to the elevator load level (empty, half-loaded, full-loaded) to obtain the preprocessed vibration data to be measured.

[0165] In step S20, the method of performing fast Fourier transform on the preprocessed vibration data can be referred to above for the implementation process of performing FFT transform on the vibration data in the first vibration dataset using formula (1). In this way, the frequency domain signal amplitude of the vibration data to be tested can be obtained as the feature to be tested.

[0166] In step S30, after the feature to be tested is input into the trained autoencoder model, the feature to be tested is compressed into the latent space by the encoder and then reconstructed by the decoder to obtain the reconstructed feature to be tested, which is the mapped feature. The mapped feature is consistent with the distribution of the source domain feature.

[0167] In step S40, after inputting the mapping features into the classification model, the classification model can perform fault classification and detection, and obtain the detection results. For example, if Determined to be fault-free. It has been determined to be a malfunction.

[0168] Based on the above design, there is no need to collect actual elevator fault data (i.e., zero-sample faults of elevator rotating parts). Only through the non-damaging vibration data of normal operation, cross-domain feature mapping with the source structure can be achieved with the help of the AE model, that is, domain adaptation, which solves the pain point of scarce elevator fault data and difficulty in collection.

[0169] Furthermore, the domain-adaptive driven feature alignment mechanism can accurately adapt to the structural differences of elevators of different brands, load capacities, and rated speeds, eliminating the need for repeated model training and reducing deployment and time costs. Combining FFT frequency domain feature extraction with elevator load hierarchical processing strategies, it can effectively filter out multi-source noise such as car vibration and electromagnetic interference, reducing false alarms and missed alarms. It only requires utilizing existing or a small number of additional sensors deployed for vibration data acquisition in the elevator, making it simple to operate, low in implementation cost, and easy to scale up and apply. Therefore, this method possesses reliable anti-interference performance and economic practicality.

[0170] In this embodiment, the classification model captures abnormal features of source domain faults through training and learning, solving the technical problems of scarce actual elevator fault data and difficulty in cross-model generalization. This method is applicable to fault monitoring of rotating components such as elevator traction machines, door operators, and guide wheels.

[0171] In this embodiment, the method may further include:

[0172] When the detection result indicates a fault, a fault prompt is sent to the management terminal.

[0173] Fault notifications can include the elevator number, name of the faulty component, inspection time, and suspected fault type (e.g., traction machine bearing damage). The notification is sent to the elevator management terminal (e.g., monitoring center computer, management personnel's mobile app) via a wireless communication module (e.g., 4G, WiFi). Alarms are then displayed on the management terminal via pop-up windows, SMS, or voice messages.

[0174] The fault indication function enables real-time alarms for faults, facilitating timely response and troubleshooting by management personnel, preventing the fault from escalating and causing outages or safety accidents, and ensuring the safe operation of elevators.

[0175] This application also provides an electronic device that may include a processor and a memory. The memory stores a computer program, which, when executed by the processor, enables the electronic device to perform the corresponding steps in the aforementioned elevator rotating component fault detection method.

[0176] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0177] The memory can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc. In this embodiment, the memory can be used to store an autoencoder model and a classification model. Of course, the memory can also be used to store a program, which the processor executes after receiving an execution instruction.

[0178] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0179] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the elevator rotating component fault detection method described above. The computer program product may exist in a computer-readable storage medium in forms including, but not limited to, source files, executable files, and installation package files.

[0180] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0181] In the embodiments provided in this application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0182] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for detecting faults in rotating components of an elevator, characterized in that, The method includes: Acquire the vibration data of the rotating components of the elevator under test; The preprocessed vibration data to be measured is subjected to frequency domain feature extraction to obtain the measured features; Using an autoencoder model, based on a pre-established cross-domain mapping relationship, the features to be measured are mapped from the target domain to the features to be measured in the source domain, which are used as mapping features. The target domain represents the elevator rotating component, and the source domain represents the source structure. The source structure is an experimental test platform with similar features to the elevator rotating component. The cross-domain mapping relationship is obtained based on the fault-free dataset and the second vibration dataset in the first vibration dataset of the source structure. The second vibration dataset is the vibration data of the elevator rotating component in a fault-free state. The mapping features are input into the trained classification model to obtain the detection results of the rotating component of the elevator under test. The detection results include results indicating that there is no fault or that there is a fault. The classification model is trained using the first vibration dataset of the source structure. The first vibration dataset includes the fault dataset and the fault-free dataset of the source structure. The method further includes, prior to acquiring the vibration data of the rotating component of the elevator under test: Obtain a first vibration dataset of the source structure and a second vibration dataset of the elevator rotating component in a fault-free state. The first vibration dataset includes a fault-free dataset representing the vibration of the source structure in a fault-free state and a fault dataset representing the vibration in a fault state. Frequency domain features are extracted from the preprocessed first vibration dataset and second vibration dataset to obtain a first feature set and a second feature set, respectively. The first feature set includes the fault feature set corresponding to the fault dataset and the fault-free feature set corresponding to the fault-free dataset. Using the second feature set as input data and the fault-free feature set as label data, a cross-domain mapping relationship is established through an autoencoder model to transform the second feature set into a feature set whose distribution is consistent with that of the fault-free feature set. The classification model is trained using the fault feature set and the fault-free feature set from the first feature set until the classification model converges, resulting in a well-trained classification model.

2. The method according to claim 1, characterized in that, Between the step of obtaining the first vibration dataset of the source structure and the second vibration dataset of the elevator rotating component under fault-free conditions, and the step of extracting frequency domain features from the preprocessed first and second vibration datasets, the method further includes: The first vibration dataset and the second vibration dataset are preprocessed to obtain preprocessed first vibration dataset and second vibration dataset, wherein the preprocessing includes: The first vibration dataset and the second vibration dataset are filtered and denoised to obtain the denoised first vibration dataset and the second vibration dataset. The first and second vibration datasets after noise reduction are divided according to the multiple load levels of the elevator to obtain a sample subset corresponding to each load level. A data augmentation strategy is adopted to expand the sample subset. The data augmentation strategy includes at least one operation among time stretching, Gaussian noise addition, and time-domain shifting on some data in the sample subset.

3. The method according to claim 1, characterized in that, The step of extracting frequency domain features from the preprocessed first and second vibration datasets to obtain a first feature set and a second feature set, respectively, includes: Perform a Fast Fourier Transform on the vibration data of any vibration signal from the preprocessed first and second vibration datasets to obtain the frequency domain signal amplitude, represented as: ； In the formula, is the amplitude of the frequency-domain signal; f represents the frequency; n is the time-series index, and 0 ≤ n < N, where N is the number of signal sampling points; is the vibration data of any vibration signal at the n-th sampling point; j is the imaginary unit; is the natural constant; If the vibration signal is a signal under fault conditions, based on the preset frequency band that characterizes fault sensitivity, all frequency domain signal amplitudes within the preset frequency band are extracted from all frequency domain signal amplitudes of the vibration signal to form fault characteristics; If the vibration signal is a signal under fault-free conditions, the amplitude of the frequency domain signal corresponding to the vibration signal is taken as the fault-free characteristic. All fault features and all fault-free features obtained based on the first vibration dataset are used as the first feature set; all fault-free features obtained based on the second vibration dataset are used as the second feature set.

4. The method according to claim 1, characterized in that, The autoencoder model includes an encoder and a decoder; The step of using the second feature set as input data and the fault-free feature set as label data to establish a cross-domain mapping relationship through an autoencoder model includes: The features corresponding to each vibration signal in the second feature set are input into the encoder to compress the features of the target domain into the latent space to obtain the features of the latent space. The decoder then reconstructs features consistent with the feature distribution of the source domain based on the features of the latent space, which are used as the reconstructed features. Based on the reconstructed features and corresponding label data, the parameters of the autoencoder model are optimized using a first preset loss function. The autoencoder model is then iteratively trained using the second feature set and the fault-free feature set until it converges. This enables the autoencoder model to convert target domain features into features aligned with the source domain feature distribution, serving as the established cross-domain mapping relationship. The first preset loss function is: ； In the formula, X represents the reconstruction error; Y represents the reconstruction feature; M represents the label data corresponding to the reconstruction feature, which is the fault-free feature of the source domain; M and Q represent the number of samples in the fault-free feature set and the number of samples in the second feature set, respectively. Refers to the reconstructed feature corresponding to the i-th feature in the second feature set; It refers to the j-th feature in the set of fault-free features.

5. The method according to claim 1, characterized in that, The step of training a classification model using the fault feature set and the fault-free feature set from the first feature set until the classification model converges, resulting in a well-trained classification model, includes: A41, using stratified sampling, the first feature set is divided into a training set, a validation set, and a test set according to a preset ratio; A42, using the training set, iteratively train the classification model to obtain a preliminary trained classification model; A43, the grid search method is used to traverse the adjustable parameters in the classification model to obtain multiple parameter sets, wherein the adjustable parameters include kernel parameters and regularization parameters. Each parameter set and the initially trained classification model form a candidate classification model. The candidate classification model is then optimized using the validation set, and the candidate classification model with the highest fault identification accuracy is selected as the optimized classification model. A44. Using the test set, the optimized classification model is tested to obtain test results indicating whether the classification model's performance is acceptable. A45. If the test result is unqualified, repeat steps A43 and A44 until the test result is qualified and a trained classification model is obtained. A46. When the test results are satisfactory, a trained classification model is obtained.

6. The method according to claim 5, characterized in that, The iterative training process includes: A batch of samples is selected from the training set, the batch of samples including evenly distributed fault-free features and fault features; The batch samples are used as the input feature vector for the current round and input into the initial classification model; The radial basis function (RBF) kernel is invoked through the classification model to calculate the similarity between each feature in the batch of samples and the features of the baseline samples. The formula for the RBF kernel is as follows: ； In the formula, for and The similarity value; The currently input feature in the batch of samples; γ represents the pre-stored features of the i-th benchmark sample; γ is the kernel parameter. It is an exponential function; By using a classification model, based on the similarity of the batch samples, the features in the batch samples are mapped to a high-dimensional space to obtain high-dimensional features; The optimal decision boundary in the high-dimensional space is determined by minimizing the objective function, which is: ； The constraints of the objective function are: ； In the formula, For high-dimensional space weights, The high-dimensional features are the features in the batch of samples after mapping. This is the offset of the decision boundary; As slack variables, The total number of samples in the training set. For regularization parameters; Based on the optimal decision boundary, the decision function is obtained, and the decision function is: ； Where f(x) represents the value of the decision function. At that time, it was determined to be in a fault-free state. When this occurs, it is determined to be a fault state; U is the sign function; U is the total number of support vectors involved in constructing the decision boundary. The Lagrange coefficients obtained during training when the decision boundary is optimal; The loss value of the classification model is backpropagated using the gradient descent method to update the relevant parameters of the classification model until a preset stopping condition is met, thus obtaining a pre-trained classification model. The preset stopping condition includes a single round loss value being lower than a first preset threshold, or a continuous R round loss value decreasing by a factor lower than a second preset threshold, where R is a preset integer greater than or equal to 5. The loss value is obtained based on the value output by the decision function.

7. The method according to claim 1, characterized in that, The elevator rotating component includes at least one of the following: traction machine, door operator system, guide wheels, and counterweight wheels. The method further includes: When the detection result indicates a fault, a fault prompt is sent to the management terminal.

8. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1 to 7.

9. A program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images