Villa elevator fault monitoring method
By constructing a learning model and attention mechanism fusion module, combining multi-domain feature vector sets and multiple base models, and dynamically adjusting weights, early and accurate identification and diagnosis of villa elevator faults are achieved, solving the problem of difficulty in early intervention in traditional methods and improving the safety of elevator operation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO LILONG ELECTROMECHANICAL
- Filing Date
- 2026-01-28
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional elevator fault monitoring methods are difficult to intervene in the early stages of a fault, leaving users in a dangerous operating state for a long time and failing to effectively ensure the operational safety of villa elevators.
A learning model combined with an attention mechanism fusion module is adopted. Vibration and sound signals of elevators are collected by sensors to construct a multi-domain feature vector set. Base models such as random forest, AdaBoost, SVM and CNN are used for fault diagnosis, and the weights are dynamically adjusted to improve the diagnostic accuracy.
It enables accurate identification of early-stage faults, reduces false alarms, improves the accuracy of fault diagnosis, ensures timely intervention by maintenance personnel, and reduces the risk to users.
Smart Images
Figure CN122020378A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of elevator fault prediction, and more specifically, to a method for monitoring faults in villa elevators. Background Technology
[0002] With the acceleration of urbanization and the improvement of residents' living standards, villa elevators, as an important supporting facility for high-end residences, have gradually become the first choice for many families. They not only enhance the comfort and convenience of living but also greatly improve residents' daily lives. However, the safety and reliability of villa elevators are increasingly attracting attention. Due to their low usage frequency, dispersed distribution, and long maintenance cycles, traditional fault monitoring methods are insufficient to effectively address these issues. Once an elevator malfunctions, it may lead to people being trapped, equipment damage, or even serious safety accidents, threatening the lives and property of residents. For example, malfunctions in the elevator's braking system, control system, or door operator system can all cause dangerous situations.
[0003] Patent CN114933221B, "An Elevator Fault Early Warning System," discloses an elevator fault early warning system, including a sensor module. The output of the sensor module is electrically connected to a communication module, and the output of the communication module is connected to a processor. The processor is bidirectionally connected to a comparison module, which is bidirectionally connected to a database. The processor is electrically connected to a server, and the processor communicates with a maintenance terminal via the server. An alarm is electrically connected to the output of the processor. This invention uses the sensor module to monitor in real time the elevator's load, elevator shaft air humidity, car door distance, car vibration, elevator operating sound, elevator shaft air temperature, car speed, and traction machine current and voltage. The collected data is then compared in real time with data in the database to accurately determine whether various parameters of the elevator are normal during operation.
[0004] However, directly comparing vibration signals with a database can only identify features that clearly match the fault characteristics. For distributed home elevators, users will be in a dangerous usage phase from the feedback to the elevator repair process. There is an urgent need for an efficient and accurate fault monitoring method to ensure the safe operation of villa elevators, so as to detect faults earlier and intervene earlier, thereby reducing the time users spend using "dangerous elevators". Summary of the Invention
[0005] This invention overcomes the shortcomings of existing methods that directly compare vibration data obtained from sensors with a database, making it difficult to intervene in the early stages of a fault. It provides a fault monitoring method for villa elevators that can detect faults in a timely manner, improve the accuracy of fault diagnosis, reduce false alarms, and help maintenance personnel seize the opportunity for maintenance.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0007] A method for monitoring elevator malfunctions in a villa, comprising the following steps: S1. Construct a learning model, which has an attention mechanism fusion module and several base models; S2. Generate a dataset by collecting data from sensors under different operating conditions of an elevator car. The data includes vibration signals, atmospheric pressure, and sound signals of the elevator. The operating conditions include normal conditions, loose guide shoes, elevator emergency stop, poor door sliding, and traction rope wear. S3. Data preprocessing, and dividing the dataset into training and test sets; S4. Input the training set into the pre-built learning model. Each base model receives the training set and trains independently, generating its own training results. The attention mechanism fusion module collects the training results of each base model and the elevator state labels corresponding to the samples in the training set as input, constructs a new feature matrix and trains it to generate the final classification result. S5. Evaluate the learning model by testing the learning model using a test set to ensure that the accuracy of the learning model in the elevator diagnosis task reaches the preset target. S6. Collect the signal of the target elevator, reconstruct it through denoising, build a multi-domain feature vector set, and input it into the learning model to obtain the final diagnostic result; S7. Upload the diagnostic results obtained in step S6 to the database and compare them, and notify the operation and maintenance personnel as needed.
[0008] The method for monitoring elevator malfunctions in this villa includes training a learning model and a process for denoising elevator data.
[0009] The learning model comprises several base models and an attention mechanism fusion module. The base models are trained independently on the training set data, each outputting predicted probabilities for five scenarios: normal state, guide shoe loosening, elevator emergency stop, car door malfunction, and traction rope wear, forming a probability vector. The outputs of each base model are concatenated and mapped to a high-dimensional feature space, no longer simply serving as the input feature matrix of the meta-model, but rather as context information for the attention module. This attention module, through a deep neural network structure, dynamically calculates the "confidence weights" of each base model for each input sample, thereby achieving adaptive weighted fusion of the decision results from different models.
[0010] The elevator data obtained from the original sampling contains a lot of noise. By denoising the elevator data and suppressing environmental and electrical noise, a signal with extremely high fidelity is obtained. The fault features under strong noise background can be clearly revealed, creating conditions for high-precision feature extraction. The clean input signal allows the model to focus more on the essential features of the fault without having to "laboriously" learn noise patterns, thus avoiding overfitting. This is the primary guarantee for the entire high-precision fault diagnosis system.
[0011] Preferably, the base model includes at least one decision tree-based model, at least one ensemble learning model, at least one support vector machine model, and at least one deep learning model. By selecting these four types of models as base models, maximum "diversity" and "complementary advantages" can be achieved: deep learning models provide a perspective based on deep representation learning, a capability lacking in traditional machine learning models; decision tree-based models have a natural ability to assess feature importance, are relatively robust to noise and overfitting, and can handle high-dimensional features without complex parameter tuning; support vector machine models excel at handling small sample sizes and nonlinear problems (elevator malfunction samples are often scarce), and have the advantage of finding an optimal hyperplane in the feature space that maximizes the margin between different classes; ensemble learning models effectively improve the ability to identify minority classes, imbalanced faults, or faults with unusual patterns, ensuring that samples ignored by other models receive sufficient attention. Their mathematical foundations and optimization objectives are completely different, meaning their errors are uncorrelated. By evaluating these four types of models, a comprehensive diagnostic perspective can be established.
[0012] Preferably, the base models are Random Forest, AdaBoost, SVM, and CNN. Random Forest is a decision tree-based model, AdaBoost is an ensemble learning model, SVM is a support vector machine model, and CNN is a deep learning model.
[0013] As a preferred method, each base model is trained using 5-fold cross-validation. Each base model outputs a 5-dimensional vector relating to the normal state, guide shoe loosening, elevator emergency stop, car door sliding problems, and traction rope wear. The output vectors of each base model are combined into 4... 5 characteristic matrix The training set is used as input to train the meta-model. The 5-fold difference validation method involves dividing the training set into five equal parts, training the model sequentially with four of these parts, and then predicting the remaining part. This process is repeated five times to ensure that each training sample receives an "out-of-bag" prediction (i.e., a prediction from a model that did not participate in training for that sample). The 5-dimensional vector corresponds to the probabilities of five scenarios: normal state, guide shoe loosening, elevator emergency stop, car door malfunction, and traction rope wear. The vectors generated by these four base models are recombine to form a feature matrix, resulting in a 20-dimensional feature matrix. An attention mechanism fusion module automatically adjusts the weights of different base models and different class probabilities to find a combination strategy that minimizes prediction error.
[0014] Preferably, the multi-domain feature vector set includes at least atmospheric pressure and time-domain, frequency-domain, and entropy features derived from the denoised and reconstructed signal. Based on the reconstructed signal, the time-domain, frequency-domain, and entropy features can be extracted, and then the atmospheric pressure can be obtained through a sensor.
[0015] Preferably, the attention mechanism fusion module includes a feature extraction layer, an attention layer, and a weight normalization layer. It uses the training results of each base model as input to iterate the parameters of the attention network to dynamically adjust the weights assigned to each base model. Unlike static meta-models such as XGBoost, the attention network of this invention acts as a "dynamic scheduler." It doesn't just learn fixed combination strategies from data; instead, it can perceive and identify in real time which base model is most accurate for a specific sample based on the feature distribution of the current input signal (e.g., the degree of nonlinearity of the signal, the significance of impact features, etc.). For example, when dealing with complex fault samples with strong nonlinear characteristics, the attention network automatically assigns higher weights to the CNN; while when dealing with small, sporadic faults, it automatically tilts towards the SVM. This mechanism makes the final diagnostic results more robust and accurate than any single model or static fusion strategy.
[0016] As a preferred option, denoising reconstruction includes: Perform VMD decomposition, initialize each mode function and its center frequency, and iteratively update the mode components, center frequency and Lagrange multipliers using the Alternating Direction Multiplier Method (ADMM) until the convergence condition is met. During the iteration process, each mode component is updated in the frequency domain, and its center frequency is recalculated according to the energy distribution. The Lagrange multipliers are used to ensure the accuracy of signal reconstruction. Perform spurious mode removal, calculate the correlation coefficient between each mode and the original signal, and retain the effective components with a correlation greater than 0.1.
[0017] By performing VMD, the original vibration / sound signal is adaptively decomposed into a set of narrowband intrinsic mode components (IMFs) with specific center frequencies. This step achieves preliminary separation of signal and noise in the frequency domain. It avoids the loss of effective modes (under-decomposition) or the generation of a large number of meaningless noise modes (over-decomposition) due to improper parameters, laying a high-quality foundation for subsequent processing.
[0018] In the multiple modal components obtained from VMD decomposition, not all components carry valid fault information. Some components may be entirely dominated by noise. The Pearson correlation coefficient between each modal component $u_k(t)$ and the original signal $f(t)$ is calculated. A threshold is set (>0.1 in this application), retaining components with correlation coefficients higher than this threshold and discarding those lower. This method removes spurious components while maximizing the retention of valid modal components containing fault information, significantly improving the signal-to-noise ratio.
[0019] As a preferred method, the parameters for VMD decomposition are obtained using the Eagle Algorithm. The population size of the Eagle Algorithm is set to 20-50 individuals, the maximum number of iterations is 20-100, and the values of K are defined as integers from 2 to 10, and the values of α are defined as continuous values from 1000 to 5000. The Eagle optimization algorithm optimizes the parameters by simulating four hunting strategies of eagles in nature: in the high-altitude exploration phase, a global search strategy is used to expand the search range; in the low-altitude exploration phase, the Levy flight mechanism is used to achieve a local fine search; in the ground attack phase, the current optimal solution and the population mean are combined for development; and in the precise capture phase, a dynamic weighting factor is used for final optimization, outputting the optimal combination of modal number K and penalty factor α. The AO algorithm simulates four hunting strategies of eagles (high-altitude exploration, low-altitude exploration, ground attack, and precise capture). It iteratively searches within a set parameter space (K=2-10, α=1000-5000), using the minimum envelope entropy or a similar index as the optimization objective, and automatically finds the optimal [K, α] parameter combination that best matches the essential characteristics of the signal, thus achieving global adaptive optimization.
[0020] As a preferred option, denoising reconstruction also includes: The elevator vibration signal and elevator sound signal are filtered in the frequency range of 70Hz to 75Hz using a Butterworth bandpass filter; the Butterworth bandpass filter is of order 3. A Savitzky-Golay filter was used for moving average processing with a window size of 3, and a first-order polynomial was used to fit the data to smooth the signal. While Butterworth filtering strictly confined the signal to the 70-75Hz frequency band where elevator fault characteristics are most pronounced, the transient response introduced by Butterworth filtering might produce a "ringing effect" or slight fluctuations at the signal edges. Therefore, the Savitzky-Golay filter was introduced to locally smooth the signal after Butterworth filtering, eliminating the slight oscillations that might be generated by Butterworth filtering, making the signal more suitable for calculating time-domain statistical characteristics. This ensured that the subsequent learning model could reliably identify weak early fault signs from noisy field data.
[0021] Preferably, data is collected via a sensor with a sampling frequency of 200Hz. The commonly used vibration analysis frequency band for elevators is between 0-200Hz; using a sensor of this specification achieves a good balance between performance and cost.
[0022] Compared with the prior art, the beneficial effects of the present invention are: (1) The Tianying optimization algorithm is introduced to perform global adaptive optimization of the key parameters of VMD, which solves the under-decomposition / over-decomposition problem caused by relying on experience or a single index to set parameters in the past, thereby obtaining modal components that are more in line with the essence of the signal. (2) After VMD denoising, Butterworth filter is added to suppress high-frequency residual noise, and spurious components are automatically eliminated by Pearson correlation coefficient. Compared with the scheme of using only EMD or single VMD, the signal-to-noise ratio is further improved.
[0023] (3) In the fault diagnosis stage, four types of significantly different models, namely CNN, RF, SVM and AdaBoost, are used as base models, and dynamic weighting is performed by the attention mechanism fusion module. Based on the attention mechanism fusion module, it can perceive and identify in real time which base model is most accurate in judging the current specific sample according to the feature distribution of the current input signal. Attached Figure Description
[0024] Figure 1 These are the training steps of the model of this invention; Figure 2 This describes the process of using the model of the present invention; Figure 3 This invention relates to the denoising and reconstruction of signals. Detailed Implementation
[0025] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0026] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0028] Example: A method for monitoring elevator faults in villas, which specifically includes training a model and using the model.
[0029] This embodiment provides a detailed description of the model training, specifically as follows: Figure 1 As shown, training the model includes the following steps: S1. Construct a learning model, which has an attention mechanism fusion module and several base models.
[0030] A learning model incorporating multiple models is constructed on the host computer. Specifically, the multiple models include a meta-model and several base models. The base models include at least one decision tree-based model, at least one ensemble learning model, at least one support vector machine model, and at least one deep learning model. By selecting these four types of models as base models, maximum "diversity" and "complementary advantages" can be achieved: deep learning models provide a perspective based on deep representation learning, a capability that traditional machine learning models lack; decision tree-based models have a natural ability to assess feature importance, are relatively robust to noise and overfitting, and can handle high-dimensional features without complex parameter tuning; support vector machine models excel at handling small sample sizes and nonlinear problems (elevator fault samples are often scarce), and have the advantage of finding an optimal hyperplane in the feature space that maximizes the margin between different classes; ensemble learning models effectively improve the ability to identify minority classes, imbalanced faults, or faults with peculiar patterns, ensuring that samples ignored by other models receive sufficient attention. Their mathematical foundations and optimization objectives are completely different, meaning that their errors are uncorrelated. By evaluating the above four types of models, a comprehensive diagnostic perspective can be established.
[0031] In some specific embodiments, the base models are random forest, AdaBoost, SVM, and CNN, respectively. Random forest is a decision tree-based model, AdaBoost is an ensemble learning model, SVM is a support vector machine model, and CNN is a deep learning model.
[0032] S2. Generate a dataset by collecting data from sensors under different operating conditions of an elevator car. The data includes vibration signals, atmospheric pressure, and sound signals of the elevator. The operating conditions include normal conditions, loose guide shoes, elevator emergency stop, poor door sliding, and traction rope wear. In a specific embodiment, the elevator is manually adjusted to five states: normal, guide shoe loose, elevator emergency stop, car door sliding sluggishly, and traction rope wear. The collected data includes sound signals from a sound sensor, vibration signals from a vibration sensor, and atmospheric pressure data from a sensor located on the car. The vibration and sound signals are collected by sensors with a sampling frequency of 200Hz. The commonly used vibration analysis frequency band for elevators is between 0-200Hz; using sensors of this specification achieves a good balance between performance and cost.
[0033] The input vibration and sound signals need to be processed. The denoised and reconstructed processed signals are used as the original input to the base model. A mixed feature set is formed by 26 features, including mean, standard deviation, root square amplitude, root mean square value, peak value, skewness, kurtosis, peak factor, margin factor, waveform factor, impulse index, amplitude mean, amplitude standard deviation, amplitude skewness, peak factor, frequency centroid, variance frequency, root mean square frequency, mean square frequency, spectral shape factor, spectral width factor, spectral skewness, spectral kurtosis, approximate entropy, sample entropy, and fuzzy entropy.
[0034] S3. Data preprocessing, and dividing the dataset into training set and test set; specifically, in this embodiment, the training set and test set are set in a ratio of 4:1.
[0035] S4. The training set is input into the pre-built learning model. Each base model receives the training set and trains independently, generating its own training results. The attention mechanism fusion module collects the training results of each base model and the elevator state labels corresponding to the samples in the training set as input, constructs a new feature matrix, trains it, and generates the final classification result. In some embodiments, each base model is trained using 5-fold cross-validation. The 5-fold cross-validation method involves dividing the training set into five equal parts, training the model sequentially with four of these parts, and then predicting the outcome of the remaining part. This process is repeated five times to ensure that each training sample receives an "out-of-bag" prediction (i.e., a prediction from a model that did not participate in the training of that sample). The 5-dimensional vector corresponds to the probabilities of five scenarios: normal state, guide shoe loosening, elevator emergency stop, car door malfunction, and traction rope wear. The vectors generated by these four base models are recombine to form a feature matrix, resulting in a 20-dimensional feature matrix. An attention mechanism fusion module automatically adjusts the weights of different base models and different class probabilities to find a combination strategy that minimizes prediction error.
[0036] Each base model outputs a 5-dimensional vector related to the normal state, guide shoe loosening, elevator emergency stop, car door sliding problems, and traction rope wear. The output vectors of each base model are combined to form a 4-dimensional vector. The feature matrix of step 5 is used as input to train the meta-model. The meta-model collects the training results of each base model as input, constructs a new feature matrix, and uses the MSE loss function to train the new feature matrix, generating the final classification result. Specifically, the classification result outputs the probability of being in one of the five states mentioned above.
[0037] Each base model outputs a probability distribution vector for five elevator states (normal, loose guide shoes, emergency stop, sluggish car door sliding, and worn traction ropes).
[0038]
[0039] in, This represents the probability that the i-th model predicts a sample belonging to the K-th type of fault.
[0040] The output vectors of all base models are concatenated to form a global decision feature vector. Its dimensions are 4 × 5 = 20.
[0041] by The input is fed into the attention mechanism fusion module.
[0042] The attention mechanism fusion module includes: The feature extraction layer uncovers the non-linear correlations between the outputs of different models; its expression is:
[0043] in, The weight matrix has the following dimensions. (That is, the hidden layer nodes are set to 64).
[0044] As a Transformer, the hidden layer set by the attention mechanism fusion module is used for feature mapping. For the bias phase, For the activation function, the ReLU (Rectified Linear Unit) function is preferably used in the embodiments to increase sparsity and nonlinear fitting ability.
[0045] Attention layering maps the hidden layer features H to raw scores for the four base models:
[0046] in, Its dimensions are 64x4. This represents the nonnormalized contribution score of the four base models.
[0047] A weight normalization layer is used to ensure that the sum of the weights is 1. A softmax transformation is applied to the score S to obtain the final dynamic attention weights α.
[0048] The i-th The physical meaning of is the degree of trust in the i-th base model in the diagnostic task of the x-th sample.
[0049] The final output is:
[0050] To achieve better prediction results, the attention mechanism fusion module is trained: During training, the parameters of each base model are frozen, and the parameters of the attention network are fine-tuned: Using the set of probability vectors obtained from the training set samples through four pre-trained base models as input, construct the objective function:
[0051] in, The actual status labels include normal status, loose guide shoes, elevator emergency stop, car door sliding problems, and traction rope wear.
[0052] Optimization process: The Adam optimizer is used to optimize the parameters in the attention network. Iterative updates are performed. By minimizing the loss function, the network automatically adjusts its parameters so that when the characteristics of a certain type of fault (such as "traction rope wear") are obvious, a larger weight value is given to the base model that is good at this type of fault (such as CNN) (e.g., α_CNN→0.7), while for another type of fault (such as "small sample sporadic faults"), the focus is automatically switched to SVM (e.g., α_SVM→0.6).
[0053] By training a specialized attention neural network, "sample-level" adaptive weight allocation was achieved.
[0054] When the elevator operating environment is noisy and exhibits strong nonlinear characteristics, the attention network automatically identifies this feature pattern and significantly increases the weights of the noise-resistant deep learning model (CNN) (α_CNN). When encountering fault categories with scarce data, the attention network automatically tilts the weights towards the support vector machine (SVM) which has the advantage of minimizing structured risk (α_SVM). This mechanism is similar to the "attending physician system" in expert consultation, dynamically assigning the most authoritative expert as the lead based on the characteristics of the condition, thereby significantly improving the system's diagnostic robustness and accuracy under complex and variable operating conditions.
[0055] The learning model comprises several base models and an attention mechanism fusion module. The base models are trained independently on the training set data, ultimately outputting a vector containing predicted probabilities for five scenarios: normal state, loose guide shoes, elevator emergency stop, malfunctioning car door, and worn traction rope. The outputs of each base model form a new feature matrix, encoding each model's "view" and "confidence level" regarding the sample from different perspectives. After obtaining the feature matrix, the attention mechanism fusion module assigns dynamic weights to each base model based on its characteristics, thereby enabling real-time perception and identification of which base model makes the most accurate judgment for a given sample.
[0056] S5. Evaluate the learning model by testing it with a test set to ensure that the model's accuracy in elevator diagnosis tasks meets the preset targets.
[0057] like Figure 2 As shown, the process of using the model includes: S6. Collect signals from the target elevator (including vibration and sound signals), denoise and reconstruct them into a multi-domain feature vector set, and input it into the learning model to obtain the final diagnostic result. The multi-domain feature vector set includes at least atmospheric pressure and time-domain, frequency-domain, and entropy features from the denoised and reconstructed signal. Based on the denoised and reconstructed signal, time-domain, frequency-domain, and entropy features can be extracted, and atmospheric pressure can be obtained through sensors.
[0058] In this embodiment, the following parameters are extracted: mean, standard deviation, root square amplitude, root mean square value, peak value, skewness, kurtosis, peak factor, margin factor, waveform factor, impulse index, mean amplitude, standard deviation amplitude, amplitude skewness, peak factor, frequency centroid, variance frequency, root mean square frequency, mean square frequency, spectral shape factor, spectral width factor, spectral skewness, spectral kurtosis, approximate entropy, sample entropy, and fuzzy entropy.
[0059] S7. Upload the diagnostic results obtained in step S6 to the database and compare them, and notify the maintenance personnel as needed. This step can also reduce the computing pressure on the cloud for maintenance, as it only requires comparing the diagnostic results with historical data in the database and setting relatively simple judgment logic. If the diagnostic result indicates that the elevator is in an abnormal state, an alarm will be issued to notify the owner of the villa elevator and the maintenance personnel, and provide relevant information.
[0060] like Figure 3 As shown, the denoising and reconstruction processes include: Perform VMD decomposition, initialize each mode function and its center frequency, and iteratively update the mode components, center frequency and Lagrange multipliers using the Alternating Direction Multiplier Method (ADMM) until the convergence condition is met. During the iteration process, each mode component is updated in the frequency domain, and its center frequency is recalculated according to the energy distribution. The Lagrange multipliers are used to ensure the accuracy of signal reconstruction. Specifically, the parameters for VMD decomposition are obtained using the Eagle Algorithm. The population size for the Eagle Algorithm is set to 20-50 individuals, the maximum number of iterations is 20-100, and the values of K are defined as integers from 2 to 10, and the values of α are defined as continuous values from 1000 to 5000. The Eagle optimization algorithm optimizes the parameters by simulating four hunting strategies of eagles in nature: in the high-altitude exploration phase, a global search strategy is used to expand the search range; in the low-altitude exploration phase, the Levy flight mechanism is used to achieve a local fine search; in the ground attack phase, the current optimal solution and the population mean are combined for development; and in the precise capture phase, a dynamic weighting factor is used for final optimization, outputting the optimal combination of modal number K and penalty factor α. The AO algorithm simulates four hunting strategies of eagles (high-altitude exploration, low-altitude exploration, ground attack, and precise capture). It iteratively searches within a set parameter space (K=2-10, α=1000-5000), using the minimum envelope entropy or a similar index as the optimization objective, and automatically finds the optimal [K, α] parameter combination that best matches the essential characteristics of the signal, thus achieving global adaptive optimization.
[0061] Perform spurious mode removal, calculate the correlation coefficient between each mode and the original signal, and retain the effective components with a correlation greater than 0.1.
[0062] By performing VMD, the original vibration / sound signal is adaptively decomposed into a set of narrowband intrinsic mode components (IMFs) with specific center frequencies. This step achieves preliminary separation of signal and noise in the frequency domain. It avoids the loss of effective modes (under-decomposition) or the generation of a large number of meaningless noise modes (over-decomposition) due to improper parameters, laying a high-quality foundation for subsequent processing.
[0063] In the multiple modal components obtained from VMD decomposition, not all components carry valid fault information. Some components may be entirely dominated by noise. The Pearson correlation coefficient between each modal component $u_k(t)$ and the original signal $f(t)$ is calculated. A threshold is set (>0.1 in this application), retaining components with correlation coefficients higher than this threshold and discarding those lower. This method removes spurious components while maximizing the retention of valid modal components containing fault information, significantly improving the signal-to-noise ratio.
[0064] The elevator vibration signal and elevator sound signal are filtered in the frequency range of 70Hz to 75Hz using a Butterworth bandpass filter; the Butterworth bandpass filter is of order 3. A Savitzky-Golay filter (SG filter) was used for moving average processing with a window size of 3, and a first-order polynomial was used to fit the data to smooth the signal. While Butterworth filtering strictly confined the signal to the 70-75Hz frequency band where elevator fault characteristics are most pronounced, the transient response introduced by Butterworth filtering might produce a "ringing effect" or slight fluctuations at the signal edges. Therefore, the Savitzky-Golay filter was introduced to locally smooth the signal after Butterworth filtering, eliminating the slight oscillations that might be generated by Butterworth filtering. This made the signal more suitable for calculating time-domain statistical characteristics, ensuring that the subsequent learning model could reliably identify weak early fault signs from noisy field data.
[0065] For each effective mode component obtained from VMD decomposition, the same 70-75Hz Butterworth bandpass filter is applied. Since elevator mechanical faults typically produce characteristic vibrations within the 70-75Hz range, the Butterworth bandpass filter retains only this critical frequency band, removing irrelevant noise and thus improving signal quality. The reconstructed signal is obtained by summing all the filtered components. The reconstructed signal is then smoothed in the time domain using an SG filter.
[0066] The elevator data obtained from the original sampling contains a lot of noise. By denoising the elevator data and suppressing environmental and electrical noise, a signal with extremely high fidelity is obtained. The fault features under strong noise background can be clearly revealed, creating conditions for high-precision feature extraction. The clean input signal allows the model to focus more on the essential features of the fault without having to "laboriously" learn noise patterns, thus avoiding overfitting. This is the primary guarantee for the entire high-precision fault diagnosis system.
[0067] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications may be made without departing from the technical solutions described in the claims.
Claims
1. A method for monitoring faults in a villa elevator, characterized in that, The method specifically includes the following steps: S1. Construct a learning model, which has an attention mechanism fusion module and several base models; S2. Generate a dataset by collecting data from sensors under different operating conditions of an elevator car. The data includes vibration signals, atmospheric pressure, and sound signals of the elevator. The operating conditions include normal conditions, loose guide shoes, elevator emergency stop, poor door sliding, and traction rope wear. S3. Data preprocessing, and dividing the dataset into training and test sets; S4. Input the training set into the pre-built learning model. Each base model receives the training set and trains independently, generating its own training results. The attention mechanism fusion module collects the training results of each base model and the elevator state labels corresponding to the samples in the training set as input, constructs a new feature matrix and trains it to generate the final classification result. S5. Evaluate the learning model by testing the learning model using a test set to ensure that the accuracy of the learning model in the elevator diagnosis task reaches the preset target. S6. Collect the signal of the target elevator, reconstruct it through denoising, build a multi-domain feature vector set, and input it into the learning model to obtain the final diagnostic result; S7. Upload the diagnostic results obtained in step S6 to the database and compare them, and notify the operation and maintenance personnel as needed.
2. The method for monitoring villa elevator faults according to claim 1, characterized in that, The base model includes at least one decision tree-based model, at least one ensemble learning model, at least one support vector machine model, and at least one deep learning model.
3. The method for monitoring villa elevator faults according to claim 2, characterized in that, The base models are Random Forest, AdaBoost, SVM, and CNN, respectively.
4. The method for monitoring villa elevator faults according to claim 3, characterized in that, Each base model is trained using 5-fold cross-validation. Each base model outputs a 5-dimensional vector representing the normal state, guide shoe loosening, elevator emergency stop, car door sliding problems, and traction rope wear. The output vectors of all base models are combined to form a 4-dimensional vector. The feature matrix of 5 is used as input to train the meta-model.
5. The method for monitoring villa elevator faults according to claim 1, characterized in that, The multi-domain feature vector set includes at least atmospheric pressure and time-domain, frequency-domain, and entropy features from the signal denoising reconstruction.
6. The method for monitoring villa elevator faults according to claim 1, characterized in that, The attention mechanism fusion module includes a feature extraction layer, an attention layer, and a weight normalization layer. It uses the training results of each base model as input to iterate the parameters in the attention network to dynamically adjust the assigned weights of each base model.
7. The method for monitoring villa elevator faults according to claim 1, characterized in that, Denoising and reconstruction include: Perform VMD decomposition, initialize each mode function and its center frequency, and iteratively update the mode components, center frequency and Lagrange multipliers using the Alternating Direction Multiplier Method (ADMM) until the convergence condition is met. During the iteration process, each mode component is updated in the frequency domain, and its center frequency is recalculated according to the energy distribution. The Lagrange multipliers are used to ensure the accuracy of signal reconstruction. Perform spurious mode removal, calculate the correlation coefficient between each mode and the original signal, and retain the effective components with a correlation greater than 0.
1.
8. A method for monitoring faults in a villa elevator according to claim 7, characterized in that, The parameters for VMD decomposition are obtained using the Eagle algorithm. The population size for the Eagle algorithm is set to 20-50 individuals, the maximum number of iterations is 20-100, and the values of K are defined as integers from 2 to 10, and the values of α are defined as continuous values from 1000 to 5000. The Eagle optimization algorithm optimizes the parameters by simulating four hunting strategies of eagles in nature: a global search strategy is used to expand the search range in the high-altitude exploration phase; a local fine search is achieved through the Levy flight mechanism in the low-altitude exploration phase; and a combination of the current optimal solution and the population mean is used in the ground attack phase. In the precise capture phase, dynamic weighting factors are used for final optimization, outputting the optimal combination of mode number K and penalty factor α.
9. A method for monitoring faults in a villa elevator according to claim 8, characterized in that, Denoising reconstruction also includes: The elevator vibration signal and elevator sound signal are filtered in the frequency range of 70Hz to 75Hz using a Butterworth bandpass filter; the Butterworth bandpass filter is of order 3. A Savitzky-Golay filter was used for moving average processing, with a window size of 3, and a first-order polynomial was used to fit the data to smooth the signal.
10. A method for monitoring faults in a villa elevator according to claim 1, characterized in that, Data is collected by a sensor with a sampling frequency of 200Hz.