An elevator risk warning method based on Transformer-based channel clustering

Abstract

This invention provides an elevator risk early warning method based on Transformer-based channel clustering. It decomposes multivariate time-series data into trend and seasonal components using a time-series decomposition module. Channels with inherent similarity in the frequency domain are clustered, and a cross-attention module with routing mechanism is used to achieve channel interaction within and between clusters. This effectively captures the correlation between different channels, improving the accuracy of elevator risk prediction. Maintenance units can then develop more scientific maintenance plans, improve elevator operation and maintenance efficiency, and reduce maintenance costs.

Description

Technical Field

[0001] This invention belongs to the field of elevator fault early warning technology, and in particular relates to an elevator risk early warning method based on Transformer and channel clustering. Background Technology

[0002] Extensive engineering practice has shown that elevator operation data contains rich information about early warning signs of malfunctions: such as abnormal vibration spectra of traction machines, sudden changes in wire rope wear, and voltage fluctuations in control cabinets—all temporal characteristics—as well as discrete characteristics such as differences in manufacturing processes and the completeness of maintenance records. Deep learning models, with their multi-layer nonlinear mapping capabilities, can predict malfunctions by mining potential correlations in historical data without requiring a precise physical model. Therefore, researching and implementing a deep learning-based elevator fault early warning system has significant research importance and social value.

[0003] In practical applications such as elevator risk warning, time series typically exhibit multivariate characteristics, and their inherent correlations pose a key modeling challenge. Current multivariate time series forecasting methods mainly suffer from two limitations: channel independence strategies, while optimizing single-channel modeling, sever the potential dependencies across channels; while channel dependency strategies, although incorporating global correlations, introduce heterogeneous noise by obfuscating all channels. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, this invention provides an elevator risk warning method based on Transformer channel clustering.

[0005] To achieve the above objectives, this invention provides an elevator risk warning method based on Transformer-based channel clustering, the specific steps of which include:

[0006] S1: Acquire IoT elevator data and generate multivariate time-series features;

[0007] S2: In the feature extraction layer of the Transformer model, the multivariate time-series features are decomposed into trend and seasonal components through the time-series decomposition module;

[0008] S3: The decomposed temporal features are divided into blocks using the Patch block operation;

[0009] S4: Use Fourier transform to extract the frequency domain amplitude features of time series data, calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center, and convert it into a probabilistic relation matrix to generate a binary mask matrix.

[0010] S5: Learn the channel dependencies within clusters through a dual-route cross-attention mechanism and output the corresponding vectors;

[0011] S6: In the prediction layer, the clustered vectors are used to generate the corresponding fault types and obtain the corresponding fault probability distribution; the trained model is used to extract specific maintenance requirements.

[0012] Preferably, in step S2, when the time series decomposition module performs seasonal term decomposition, it transforms the time series from the time domain to the frequency domain, and the transformation formula is as follows:

[0013] ,

[0014] in, For a seasonal pattern, Φ and A are the phase and amplitude obtained from the Fourier transform DFT(X), respectively, and the set... The first K with the largest amplitude f Each frequency component.

[0015] Preferably, in step S2, when the time series decomposition module performs trend term decomposition, it extracts the residual signal X after removing the seasonal term based on the moving average method. rem =X—X sea The trend pattern in the formula is as follows:

[0016] ,

[0017] in, The trend pattern is extracted after padding and average pooling. Padding(·) is the padding operation, AvgPool(·) is the average pooling operation, and X is the time series. Local averages are calculated by sliding a fixed window along the time dimension. sea This is a seasonal pattern;

[0018] The extracted seasonal and trend patterns are overlaid with the original input time series to obtain the final input time series.

[0019] Preferably, the specific steps for generating the binary mask matrix in step S4 include:

[0020] Initialize a set of learnable clustering embeddings with a number smaller than the number of time-series data variables;

[0021] Perform a real-valued Fast Fourier Transform on the data from each channel, projecting it onto a finite frequency space with a fixed Fourier basis and a spatial dimension of d. m / 2;

[0022] The channel feature vector is obtained by calculating the norm, and the specific formula is as follows:

[0023] ,

[0024] ,

[0025] in, This represents the time series of the i-th channel in the p-th patch. Let rFFT(·) be the feature vector of the j-th cluster center in the p-th patch, where rFFT(·) represents the real fast Fourier transform, norm(·) represents the calculation of the amplitude value of the complex frequency domain vector, and chan is the feature vector through the Fourier transform.

[0026] The Euclidean distance metric is used to quantify the similarity between channels and cluster embeddings in the frequency domain. The calculation formula is as follows:

[0027] ,

[0028] in, Let be the value of the frequency domain vector of the i-th channel in the p-th patch at the n-th frequency component. Let d be the embedding value of the p-th patch in the j-th cluster embedding on the n-th frequency component. m / 2 represents the dimension of the frequency domain features.

[0029] Preferably, the specific steps for generating the binary mask matrix in step S4 further include:

[0030] Similarity distance in the frequency domain based on channel and cluster embedding The relationship matrix between channels and clustering is calculated and normalized to the [0,1] interval, as shown in the following formula:

[0031] , , ,

[0032] Where D, R, and P are the distance matrix, relation matrix, and probability matrix, respectively, and γ∈ (0, 1) is the discount factor. , and These represent the probability, distance, and relationship of channel i belonging to cluster j in the prediction task, respectively. Through the above process, the relationship between the channel and the cluster embedding is probabilistically transformed.

[0033] The probability matrix P is converted into a binary channel mask to obtain the clustering membership matrix M.

[0034] Preferably, step S5 includes the following specific steps:

[0035] Routing clustering phase: embedding the initial clusters As a router, it learns channel dependencies within clusters through a cross-attention mechanism, filtering out noise from outside the clusters. Its expression is:

[0036] ,

[0037] in, The updated clustering embeddings are given, where M is the clustering membership matrix. W represents element-wise multiplication, used for sparse masking of the attention matrix. Q W K and W V Here, is the learnable parameter matrix; exp(·) performs a non-linear transformation on the attention score, simulating the scaling effect of the Softmax function; Normalize(⋅) is a row-wise normalization operation, controlling the sum of the attention weights to be 1; P is the number of patches. This is the channel representation of the time series; T is the transpose of the matrix, d m is the dimension of the frequency domain features.

[0038] Preferably, in step S5, when performing the clustering operation, the steps further include:

[0039] Routing and distribution phase: embedding the updated clusters As key vector K and value vector V, they are represented by channels in the time series. As the query vector Q, information is exchanged between channels to construct fully connected relationships between different clusters; the formula for the routing distribution phase is shown below:

[0040] ,

[0041] in, , , These are learnable parameters. It is the output of a single patch block, which has different relationships between its channels; all patch blocks share the same cross-attention mechanism to obtain the output containing all patch blocks. ,Right now ;

[0042] X dim Add the features obtained from the temporal decomposition.

[0043] Preferably, the multivariate time-series features in step S1 include: elevator operating speed, elevator operating frequency, and temperature features.

[0044] Preferably, in step S6, the prediction layer converts the elevator aggregated features into elevator fault probabilities using an MLP structure and a Sigmoid function. The MLP structure has two fully connected layers, and its calculation steps include:

[0045] Employing a first-layer fully connected layer for timing cross-channel X out The calculation is as follows:

[0046] ,

[0047] Where W1 and b1 are the weight matrix and bias vector of the first layer, T1 is the activation function, and T1 is the output intermediate hidden layer feature.

[0048] Preferably, the step of calculating the failure probability in the prediction layer further includes:

[0049] The prediction score is generated through the second fully connected layer:

[0050] ,

[0051] Where W2 and b2 are the parameters of the second layer, The original score vectors for each fault category are normalized using the Softmax function to convert the predicted scores into a probability distribution.

[0052] This invention provides an elevator risk warning method based on Transformer and channel clustering, which has the following advantages compared with existing technologies:

[0053] To balance fine-grained modeling and correlation capture, this invention constructs a channel clustering module. First, channels with high structural similarity are dynamically aggregated in the frequency domain (using FFT amplitude features). Then, a dual-route cross-attention mechanism is designed for modeling. Through inter-cluster synergy enhancement and refined intra-cluster interaction, efficient modeling of complex multi-channel dependencies is achieved. This framework significantly improves the accuracy and robustness of risk prediction while preserving the local characteristics of channels. Attached Figure Description

[0054] Figure 1 The flowchart shows an elevator risk warning method based on Transformer channel clustering provided by the present invention.

[0055] Figure 2 This is a system diagram of an elevator risk warning method based on Transformer channel clustering, provided in an embodiment of the present invention. Detailed Implementation

[0056] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0057] like Figure 1 As shown, this invention provides an elevator risk early warning method based on Transformer-based channel clustering, which is applied to an elevator risk early warning big data model system. The specific steps include:

[0058] S1: Acquire IoT elevator data and generate multivariate time-series features;

[0059] S2: In the feature extraction layer of the Transformer model, the multivariate time-series features are decomposed into trend and seasonal components through the time-series decomposition module;

[0060] S3: The decomposed temporal features are divided into blocks using the Patch block operation;

[0061] S4: Use Fourier transform to extract the frequency domain amplitude features of time series data, calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center, and convert it into a probabilistic relation matrix to generate a binary mask matrix.

[0062] S5: Learn the channel dependencies within clusters through a dual-route cross-attention mechanism, shield noise outside clusters, and output the corresponding vectors;

[0063] S6: In the prediction layer, the clustered vectors are used to generate the corresponding fault types and obtain the corresponding fault probability distribution; the trained model is used to extract specific maintenance requirements.

[0064] Specifically, based on the multivariate time-series vector of IoT elevators—that is, the time-series vector formed by processing the original data of IoT elevators—this invention employs a time-series decomposition module to break down the multivariate time-series data into trend and non-trend terms (i.e., seasonal terms), and clusters them by channels with inherent similarity in the frequency domain. During the clustering stage, the frequency domain amplitude features extracted by Fourier transform are used to calculate the similarity distance between each channel of the multivariate time series and the learnable cluster centers, and this is transformed into a probabilistic relation matrix, which is then used to generate a binary mask matrix. Subsequently, a cross-attention module with a routing mechanism is used to realize channel interaction within and between clusters, thereby effectively capturing the correlation between different channels while shielding noise interference from outside the clusters, thus improving the accuracy of elevator risk prediction. Maintenance units can then formulate more scientific maintenance plans, improve elevator operation and maintenance efficiency, and reduce maintenance costs. During model training, the performance is improved through iterative optimization of the algorithm. When predicting maintenance needs, the prediction results are further refined through multi-layer convolutional blocks based on the fault type and probability distribution, making the entire early warning process logic clearer and execution more efficient.

[0065] In machine learning and clustering algorithms, "learnable cluster centers" typically refer to cluster centers that can adaptively adjust and optimize during model training. These centers are not fixed in advance but are learned from the training data in a data-driven manner to better fit the distribution characteristics of the data. Traditional clustering algorithms (such as K-means) often determine cluster centers through initialization or heuristic methods, while learnable cluster centers are dynamically optimized through model parameter updates (such as gradient descent) to better reflect the true distribution of the data.

[0066] In this embodiment, before training and prediction, the raw data of the IoT elevator is cleaned and preprocessed. Data cleaning is mainly used to handle abnormal information in the raw data, while data preprocessing includes handling missing values ​​and discretizing elevator-related text information.

[0067] Data cleaning primarily targets anomalies in the raw data, including garbled characters, mixed fields, numerical anomalies, and inconsistent formats. Differentiated processing strategies are needed based on the type of anomaly: for garbled data, manual filtering retains valid information or removes it directly; mixed fields are often caused by data export errors and can be corrected manually by checking and correcting field correspondences; numerical anomalies arise from missing values ​​or automatic imputation and are repaired using rule-based detection methods; inconsistent formats are handled through a combination of manual and machine processing to ensure data from different time points and standards can be integrated and used. Data preprocessing covers missing value handling and discretization of elevator-related text information. Fields with clear identifiers, such as dates and fault codes, can be quickly converted using automated rules, while text information lacking identifiers relies on keyword extraction and field recombination for structured processing, thereby constructing a standardized input dataset.

[0068] Based on multivariate time-series feature vectors, this invention performs time-series decomposition to capture their periodicity (i.e., seasonality) and trend. Periodicity refers to the recurring fluctuation patterns in a time series at fixed time intervals, and the fluctuation cycle is predictable. For example, elevator operation data may exhibit regular changes on a daily, weekly, or monthly basis. Trend refers to the overall increasing or decreasing trend of the time series over a longer time span, reflecting the long-term direction of data change.

[0069] The time series decomposition module includes seasonal decomposition and trend decomposition to extract cyclical patterns and long-term trends. When performing seasonal decomposition, the module uses Discrete Fourier Transform (DFT) to decompose the input time series X from the time domain to the frequency domain, representing it as a linear combination of Fourier bases. Specifically, this is achieved by selecting the largest amplitude K... f Using four Fourier bases to maintain sparsity in the frequency domain, the seasonal pattern X is reconstructed by inverse Fourier transform (IDFT).sea The specific process is as follows:

[0070] ,

[0071] Where Φ and A are the phase matrix and amplitude vector obtained by the Discrete Fourier Transform (DFT) (X), respectively, and the set {f1,...,f_K} is... f} represents the top K selected after sorting by amplitude in descending order. f Each frequency component.

[0072] When performing trend decomposition in the time series decomposition module, the moving average principle is used to highlight long-term trend characteristics by smoothing out periodic fluctuations. Specifically, trend patterns are extracted from the residual signal Xrem=X—Xsea after removing seasonal components. The specific process is as follows:

[0073] ,

[0074] Among them, X trend The trend pattern is extracted through padding and average pooling. Padding (·) represents the boundary padding operation, and AvgPool (·) is the average pooling operation that calculates the local mean by sliding a fixed window along the time dimension.

[0075] Finally, the extracted seasonal and trend patterns are superimposed on the original input time series X to obtain the final input time series.

[0076] After the above time-series decomposition process, the time-series data now contains seasonal and trend characteristics. To avoid the attention mechanism modeling only single-point time and to better capture the temporal relationships in the sequence, the time-series data is divided into blocks, specifically:

[0077] Suppose the input is a multivariate time series Where R is the vector space, T is the time series length, and C is the feature dimension. For a given patch size S, the time series is divided into P patches, where P = T / S, resulting in the patch sequence. This method addresses the problem that traditional attention mechanisms struggle to capture long-distance temporal dependencies when directly processing single-point temporal features. The block operation aggregates S consecutive time steps into local temporal units, enabling the attention mechanism to model in units of "temporal blocks." For example, by interacting with time steps within a patch, it learns the variation patterns of elevator vibration signals within short periods, thus extending single-point modeling to local temporal correlation analysis.

[0078] The time-series data after block processing is input into the channel clustering module to learn the dependencies between channels (cross-dimensions). The channel clustering module aims to capture the dependencies of multi-dimensional features in different clusters. This module first projects the time-series information into the frequency domain space through real-number Fourier transform, providing a frequency domain feature foundation for channel clustering analysis. Then, leveraging a dual mechanism of route aggregation and route distribution, it constructs an interactive correlation network between multiple variables (channels), achieving efficient aggregation and distribution of cross-channel features. Specifically, it can be divided into three stages: clustering (i.e., step S4), route clustering, and route distribution (i.e., step S5).

[0079] (1) Clustering stage

[0080] During the clustering phase, the model initializes a set of K learnable clustering embeddings (K less than the number of channels) and performs a patch operation on them. To enhance cross-channel modeling capabilities, a frequency domain modeling strategy is adopted to mine discriminative features. Specifically, for the multivariate time series processed by the dual time series module, for a single patch block Xp,:, real-valued Fast Fourier Transform is performed on the data of each channel, projecting it onto a finite frequency space (dimension d) with a fixed Fourier basis. m / 2). The channel eigenvector is obtained by calculating the norm (amplitude). Performing the same transformation operation on the cluster embedding yields... The specific calculation logic is as follows:

[0081] ,

[0082] ,

[0083] in, This represents the time series of the i-th channel in the p-th patch. Let rFFT(·) be the feature vector of the j-th cluster center in the p-th patch, where rFFT(·) represents the real fast Fourier transform, norm(·) represents the calculation of the amplitude value of the complex frequency domain vector, and chan is the feature vector through the Fourier transform.

[0084] To evaluate the similarity between channels and cluster embeddings in the frequency domain, Euclidean distance is used as a metric to quantify the similarity between channels and cluster embeddings in the frequency domain, i.e.:

[0085] ,

[0086] in, Let be the value of the frequency domain vector of the i-th channel in the p-th patch at the n-th frequency component. Let d be the embedding value of the p-th patch in the j-th cluster embedding on the n-th frequency component. m / 2 represents the dimension of the frequency domain features.

[0087] Similarity distance in the frequency domain based on channel and cluster embedding The relationship matrix between channels and clustering is calculated and normalized to the [0,1] interval, as shown in the following formula:

[0088] , , ,

[0089] Where D, R, and P are the distance matrix, relation matrix, and probability matrix, respectively, and γ∈ (0, 1) is the discount factor. , and These represent the probability, distance, and relationship of channel i belonging to cluster j in the prediction task, respectively. Through the above process, the relationship between the channel and the cluster embedding is probabilistically transformed.

[0090] Furthermore, to effectively filter out interference from irrelevant channels and enhance feature aggregation of similar channels, a Bernoulli resampling mechanism is introduced to convert the probability matrix P into a binary channel mask, generating a clustering membership matrix M. The elements... Follows Bernoulli distribution ≈Bernoulli ( ), through probability sampling, make high probability values ​​P_i,j( It is easier to map to 1, ensuring that channels with high correlation are assigned to the corresponding clusters with a higher probability, thus achieving adaptive filtering and cluster assignment of channel features.

[0091] (2) Routing clustering stage

[0092] During the routing clustering phase, the initial clusters generated earlier are embedded. As a router, d m The embedding dimension serves as the core component of routing, leveraging an improved cross-attention mechanism to strengthen intra-cluster channel dependencies while suppressing out-of-cluster noise interference. In specific implementation, cluster embedding... As a query vector, the time series representation after processing by the dual time series module will be used. Simultaneously serving as both key and value, the cross-attention mechanism aggregates feature information from similar channels, and its mathematical expression is as follows:

[0093] ,

[0094] Where M is the cluster membership matrix, which is an approximate binary mask matrix used to implement sparse attention on channels within a cluster. That is, by multiplying the attention matrix element by element, only channels within the cluster are allowed to participate in the attention calculation. W represents element-wise multiplication, used for sparse masking of the attention matrix. Q W K and W V The parameter matrix is ​​learnable. The attention score is nonlinearly transformed using the exp function to simulate the scaling effect of the Softmax function, and then normalized using Normalize() to make the weight sum equal to 1. T is the transpose of the matrix. The calculated updated clustering embedding is used both to iteratively optimize the clustering probability matrix and to provide a feature basis for cross-cluster information interaction in the subsequent routing distribution stage.

[0095] (3) Routing distribution phase.

[0096] During the routing phase, the updated clustering embedding is used. The key vector K and value vector V are represented by channels in a time series. As the query vector Q, cross-channel information sharing is achieved through a cross-attention mechanism, thereby constructing a fully connected interactive relationship between different clusters. The specific calculation logic in the routing distribution phase is as follows:

[0097] ,

[0098] in, , , These are learnable parameters. A single patch block aggregates the outputs of the relationships between different channels; all patch blocks share the same cross-attention mechanism to obtain the output containing all patch blocks. ,Right now .

[0099] Ultimately, X dim The features obtained from the temporal decomposition are added together to obtain the final temporal cross-channel representation Xout, which captures dependencies across time and dimensions.

[0100] Based on the feature similarity weights learned by the channel clustering module, this invention achieves fault probability prediction for IoT elevators through a prediction layer. The prediction layer employs a two-layer fully connected MLP structure for dimensionality reduction. Compared to the linear fitting capability of a single fully connected layer, this structure can effectively capture nonlinear mapping relationships in the data. The specific calculation process is as follows:

[0101] First, the timing cross-channel representation X out Calculated after the first fully connected layer:

[0102] ,

[0103] Where W1 and b1 are the weight matrix and bias vector of the first layer, T1 is the activation function (such as ReLU) and T1 is the output intermediate hidden layer feature.

[0104] The predicted score is then generated through a second fully connected layer:

[0105] ,

[0106] Where W2 and b2 are the parameters of the second layer. The original score vectors for each fault category are represented by '0'. To convert the predicted scores into a probability distribution, the Softmax function is used for normalization.

[0107] In its implementation, the output vector of the channel clustering module first enters the fully connected layer for dimensionality reduction. Through weighted summation and bias calculation, high-dimensional time-series features are mapped to a low-dimensional space. This process reduces feature dimensionality by learning latent patterns in the data, improving computational efficiency while avoiding information loss. For example, multi-channel features such as elevator traction machine vibration spectrum and door operator current can be compressed into a low-dimensional representation containing fault correlations after MLP processing, and then the probability distribution of each fault type is generated through the Softmax function.

[0108] The advantages of this prediction mechanism are: the two fully connected layers introduce nonlinear transformations through activation functions, which can capture complex correlations; the Softmax output has a clear fault probability distribution, which makes it easier for maintenance personnel to formulate strategies according to risk levels; and the interpretability is enhanced: key fault indicators can be traced through intermediate features, such as when a certain channel has a high weight, the corresponding sensor data is indicated as the main basis for early warning.

[0109] During model training, both hyperparameters and intrinsic model parameters need to be optimized simultaneously to improve prediction performance. Addressing the common sample imbalance problem in elevator fault warning scenarios, traditional cross-entropy loss functions tend to bias the model towards the majority class, reducing the accuracy of minority class fault identification. Therefore, this invention introduces a weighted cross-entropy loss function, Loss, which adaptively adjusts class weights to balance the training contributions of each class. The specific calculation formula is as follows:

[0110] ,

[0111] Where N is the total number of samples, M is the number of categories, and w c Let y be the weight coefficient of the c-th category. ic p represents the true probability of the sample. icLet w be the probability that the i-th sample belongs to class c. Let w be the weight and weight coefficient of class c. c The calculation formula is:

[0112] ,

[0113] Where, n c This represents the number of samples belonging to category c.

[0114] Tables 1-5 show the prediction results of the system of this invention under different fault types (trapped passengers, open door, door opening / closing, overspeed, and others). The study used 8000 KONE elevators as the test subjects, analyzing data from four periods: early May, late May, early June, and late June. The results show that the model exhibits good stability in predicting door opening / closing faults, with a hit rate of approximately 28%. The prediction accuracy for other types of faults also remained at around 20%, fully validating the model's effectiveness. Even with a small sample size, the model can still effectively predict serious faults such as trapped passengers, open door operation, and overspeed, highlighting its generalization ability and fault early warning value in small sample scenarios.

[0115]

[0116] Table 1. Prediction results for IoT elevators experiencing entrapment faults.

[0117]

[0118] Table 2 Prediction results when the IoT elevator experiences door opening failure

[0119]

[0120] Table 3. Prediction results for door opening / closing failures in IoT elevators

[0121]

[0122] Table 4. Prediction results when an IoT elevator experiences overspeed fault.

[0123]

[0124] Table 5. Predictive Results for Other Faults in IoT Elevators

[0125] like Figure 2As shown, the raw data contains three types of tables: elevator information table, fault record table, and weather temperature table. The elevator information table and fault record table are generated from IoT elevator data, while the weather temperature table comes from historical meteorological data of Hangzhou from the 2345 Weather King platform. As shown in Tables 6-9, the dataset used in this invention consists of three types of structured tables: elevator information table, fault record table, and weather temperature table. Table 9 lists common anomalies and their corresponding cleaning methods.

[0126]

[0127] Table 6 Elevator Information Sheet

[0128]

[0129] Table 7 Elevator Fault Record Form

[0130]

[0131] Table 8 Weather Temperature Table

[0132]

[0133] Table 9

[0134] Most of the raw elevator data is in text format, which is inconvenient for direct model processing. To improve training efficiency and model robustness, the text data needs to be discretized and encoded. After transformation, the discretized data becomes dense multivariate time-series feature vector data. Common discretization encoding methods include enumeration encoding and one-hot encoding. Enumeration encoding assigns a unique number to each unique value; one-hot encoding creates an independent field for each unique value, with a value of 0 or 1.

[0135] Normalization primarily addresses the issue of inconsistencies in the dimensions of different features. Its standardization formula is as follows:

[0136] ,

[0137] Here, x represents the input feature value. After normalization, all feature values ​​will be mapped to the [0,1] interval, thus avoiding model bias caused by differences in numerical scale. To avoid data inflation during the preprocessing stage, the normalization operation is performed in the prediction module, and only the maximum and minimum values ​​of each field are recorded in advance.

[0138] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. An elevator risk warning method based on Transformer-based channel clustering, characterized in that, The specific steps include: S1: Acquire IoT elevator data and generate multivariate time-series features; S2: In the feature extraction layer of the Transformer model, the multivariate time-series features are decomposed into trend and seasonal components through the time-series decomposition module; S3: The decomposed temporal features are divided into blocks using the Patch block operation; S4: Use Fourier transform to extract the frequency domain amplitude features of time series data, calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center, and convert it into a probabilistic relation matrix to generate a binary mask matrix. S5: Learn the channel dependencies within clusters through a dual-route cross-attention mechanism and output the corresponding vectors; Step S5 specifically includes the following steps: Routing clustering phase: The initial clusters are embedded as routers, and the channel dependencies within the clusters are learned through the cross-attention mechanism to shield noise outside the clusters; In the routing and distribution phase, the updated cluster embeddings are used as key vectors and value vectors, and the channel representation of the time series is used as the query vector to exchange information between channels in order to build a fully connected relationship between different clusters. All patch blocks share the same cross-attention mechanism to obtain the output of all patch blocks and the features after temporal decomposition, and then add them together to obtain the final temporal cross-channel representation; S6: In the prediction layer, the clustered vectors are used to generate the corresponding fault types and obtain the corresponding fault probability distribution; the trained model is used to extract specific maintenance requirements.

2. The elevator risk warning method based on Transformer-based channel clustering according to claim 1, characterized in that, In step S2, when the time series decomposition module performs seasonal term decomposition, it transforms the time series from the time domain to the frequency domain.

3. The elevator risk warning method based on Transformer-based channel clustering according to claim 1, characterized in that, In step S2, when the time series decomposition module performs trend term decomposition, it extracts the trend pattern in the residual signal after removing the seasonal term based on the moving average method. The extracted seasonal and trend patterns are overlaid with the original input time series to obtain the final input time series.

4. The elevator risk warning method based on Transformer-based channel clustering according to claim 1, characterized in that, The specific steps for generating the binary mask matrix in step S4 include: Initialize a set of learnable clustering embeddings with a number smaller than the number of time-series data variables; Perform a real-valued fast Fourier transform on the data of each channel and project it onto a finite frequency space with a fixed Fourier basis. Channel feature vectors are obtained through norm calculation; The similarity between channels and cluster embeddings in the frequency domain is quantified using Euclidean distance metric. Based on the similarity distance between channels and cluster embeddings in the frequency domain, the relationship matrix between channels and clusters is calculated and normalized to the [0,1] interval to probabilize the relationship between channels and cluster embeddings; The probability matrix is ​​converted into a binary channel mask to obtain the clustering membership matrix.

5. The elevator risk warning method based on Transformer-based channel clustering according to claim 1, characterized in that, The multivariate time-series features in step S1 include: elevator operating speed, elevator operating frequency, and temperature features.

6. The elevator risk warning method based on Transformer-based channel clustering according to claim 1, characterized in that, In step S6, the prediction layer converts the elevator aggregated features into elevator fault probabilities using an MLP structure and a Sigmoid function. The MLP structure has two fully connected layers, and its calculation steps include: The first fully connected layer is used to perform time-series cross-channel calculations; Prediction scores are generated through a second fully connected layer.

Images