Elevator acceleration fault detection method and device based on improved meta learning

Abstract

The invention discloses an elevator acceleration fault detection method and device based on improved meta learning, and the method comprises the steps: obtaining an original acceleration response data sequence of a target elevator, carrying out the preprocessing, randomly selecting a support set and a query set after the preprocessing, constructing and training an acceleration fault detection model, the accelerated fault detection model is used for performing feature extraction on the support set and the query set, then performing prototype calculation on the extracted features to obtain loss, and introducing an intra-class distance regularization item in the prototype calculation; acquiring a difficult sample library; the accelerated fault detection model is tested through the test set to obtain the accuracy rate of the model, and the support set and the query set are adjusted according to the accuracy rate; retraining the accelerated fault detection model according to the difficult sample library and the support set and the query set which are redivided and adjusted; and performing fault detection by using the retrained accelerated fault detection model. The technical problems of few-sample generalization and class imbalance in the prior art are solved.

Description

Technical Field

[0001] This application relates to the technical field of elevator acceleration fault detection, and particularly to an elevator acceleration fault detection method and device based on improved meta-learning. Background Art

[0002] In the implementation process of intelligent fault diagnosis methods based on machine learning or deep learning, there are three main problems. First, it is often difficult to obtain sufficient fault data in the actual elevator scenario, and the training of deep learning fault diagnosis models usually requires a large amount of labeled data, which restricts the practical application of such methods. Second, existing elevator fault diagnosis methods for small samples of elevator operation data under complex conditions require manual experience intervention in the configuration of the support set and the query set. For example, it is necessary to understand relevant meta-learning methods and obtain results through experiments such as comparing 1-shot, 5-shot, and 10-shot, resulting in complex experiments and limited model generalization ability. Finally, the performance of the prototype network depends to a large extent on the quantity and quality of samples in the support set. On an imbalanced dataset, the prototype network may be greatly affected. Summary of the Invention

[0003] The purpose of the embodiments of this application is to provide an elevator acceleration fault detection method and device based on improved meta-learning to solve the technical problems of few-shot generalization and class imbalance in related technologies.

[0004] According to the first aspect of the embodiments of this application, an elevator acceleration fault detection method based on improved meta-learning is provided, including: Obtain the original acceleration response data sequence of the target elevator; Preprocess the original acceleration response data sequence to obtain a two-dimensional image containing time information; Randomly select a support set and a query set from the two-dimensional image; Construct and train an acceleration fault detection model according to the support set and the query set. The acceleration fault detection model is used to extract features from the support set and the query set, and then perform prototype calculation on the extracted features to obtain a loss, where an intra-class distance regularization term is introduced in the prototype calculation; Use the trained acceleration fault detection model for classification prediction. When a sample is misclassified, add it to the difficult library of the corresponding category to obtain a difficult sample library, and the query set is composed of the difficult sample library and ordinary samples; Test the trained acceleration fault detection model through a test set to obtain the accuracy of the model, and adjust the support set and the query set according to the accuracy; According to the difficult sample library, re-divide the adjusted support set and query set, and retrain the accelerated fault detection model using the re-divided support set and query set to predict the operating state of the elevator; Input the preprocessed acceleration response data sequence into the retrained accelerated fault detection model to obtain a fault detection result.

[0005] According to the second aspect of the embodiments of the present application, there is provided an elevator acceleration fault detection device based on improved meta-learning, including: An acquisition module for acquiring the original acceleration response data sequence of a target elevator; A preprocessing module for preprocessing the original acceleration response data sequence to obtain a two-dimensional image containing time information; A data set selection module for randomly selecting a support set and a query set from the two-dimensional image; A model training module for constructing and training an accelerated fault detection model according to the support set and the query set. The accelerated fault detection model is used to extract features from the support set and the query set, and then perform prototype calculation on the extracted features to obtain a loss, wherein an intra-class distance regularization term is introduced in the prototype calculation; A sample acquisition module for using the trained accelerated fault detection model to perform classification prediction. When a sample is misclassified, it is added to the difficult library of the corresponding class to obtain a difficult sample library, and the query set is composed of the difficult sample library and ordinary samples; A data set adjustment module for testing the trained accelerated fault detection model through a test set to obtain the accuracy of the model, and adjusting the support set and the query set according to the accuracy; A model re-training module for re-dividing the adjusted support set and query set according to the difficult sample library, and retraining the accelerated fault detection model using the re-divided support set and query set to predict the operating state of the elevator; A fault detection module for inputting the preprocessed acceleration response data sequence into the retrained accelerated fault detection model to obtain a fault detection result.

[0006] According to the third aspect of the embodiments of the present application, there is provided an electronic device, including: One or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in the first aspect.

[0007] According to a third aspect of the embodiments of the present application, there is provided a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the method described in the first aspect are implemented.

[0008] The technical solutions provided by the embodiments of the present application may include the following beneficial effects: As can be seen from the above embodiments, the present application innovatively integrates three core technologies: a dynamic meta-training mechanism, adaptive sample scheduling, and a regularized prototype network, systematically solving key problems such as scarce fault samples, unbalanced class distribution, and insufficient generalization ability of diagnostic models in industrial scenarios. The trained model can accurately identify high-risk faults such as elevator emergency stops and severe vibrations under extreme conditions that only require a small number of samples / classes, and output highly reliable fault classification results, meeting the stringent requirements of industrial intelligent operation and maintenance systems for fault diagnosis under complex working conditions. This solution not only breaks through the performance bottleneck of traditional deep learning models in small-sample scenarios but also forms a systematic technological breakthrough in aspects such as dynamic task construction, difficult sample mining, and unsupervised feature alignment, significantly improving the practicality and engineering capabilities of elevator fault diagnosis in data-scarce environments.

[0009] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

[0011] Figure 1 is a flowchart of a method for elevator acceleration fault detection based on improved meta-learning shown according to an exemplary embodiment.

[0012] Figure 2 is an overall framework diagram of a method for elevator acceleration fault detection based on improved meta-learning shown according to an exemplary embodiment.

[0013] Figure 3 is a distribution diagram of the length of the original acceleration response data sequence shown according to an exemplary embodiment.

[0014] Figure 4 is a diagram of the number of categories of the acceleration fault data set after being unified to a length of 400 shown according to an exemplary embodiment.

[0015] Figure 5 is a network model structure diagram of MetaRes-DMT-AS shown according to an exemplary embodiment.

[0016] Figure 6It is a diagram comparing the iterative processes of each model shown according to an exemplary embodiment.

[0017] Figure 7 It is a diagram comparing the running time, accuracy, and number of model parameters of each model shown according to an exemplary embodiment.

[0018] Figure 8 It is a confusion matrix diagram of each model shown according to an exemplary embodiment.

[0019] Figure 9 It is a block diagram of an elevator acceleration fault detection device based on improved meta-learning shown according to an exemplary embodiment. Detailed implementation mode

[0020] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0021] The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "the", and "said" used in the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc. may be used in the present application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0023] As Figure 1 and Figure 2 As shown, the embodiment of the present invention provides a method for detecting elevator acceleration faults based on improved meta-learning, which may include the following steps: S1: Obtain the original acceleration response data sequence of the target elevator; Specifically, the experimental environment was set up in elevators in daily use areas such as residential communities and commercial buildings. A gyroscope sensor installed above the elevator captured the three-axis acceleration of the elevator during operation, capturing a total of 16,794 operation processes. The complete operation process was automatically captured during the operation of the elevator, i.e., a complete operation process was considered as one process from the elevator door closing, the elevator running, to the elevator door opening. Five categories of raw acceleration response data sequences were screened using platform detection data (elevator reset operation, severe car vibration, emergency stop failure, normal operation, and overspeed). After obtaining the raw acceleration response data sequence, it was also cleaned, as errors occasionally occur during the elevator data collection process, resulting in data anomalies (0 and excessive acceleration). This was done to ensure data quality.

[0024] In one embodiment, the cleaning is to remove the running process containing incomplete and missing data in the data. In addition, the time length corresponding to each running state is different, such as Figure 3 As shown, most of the overspeed intervals are between 400 and 600, most of the reset operation intervals are above 800, most of the emergency stop intervals are between 100 and 400, most of the violent car oscillation intervals are between 100 and 400, and most of the normal operation intervals are between 100 and 400. Finally, the most corresponding intervals in each different operating state are taken and processed with a uniform length of 400. The insufficient intervals are interpolated with cubic spline, and the excessive intervals are extracted.

[0025] After processing, it was found that Figure 4 As shown in the figure, the data of emergency stop and violent car oscillation are far less than 1,500 for each of the other three types, with only 157 and 116 respectively.

[0026] S2: Preprocessing the original acceleration response data sequence to obtain a two-dimensional image containing time information; this step includes the following sub-steps: S21: performing normalization processing on the original acceleration response data sequence; Specifically, after cleaning the data, Z-score normalization is performed on the one-dimensional raw acceleration response data sequence. This involves subtracting the mean and dividing by the variance of each data point in the data sequence, so that the processed data approximately conforms to a standard normal distribution (0, 1). Before training a neural network, Z-score normalization is often performed on the input features to ensure a consistent distribution for each feature, making the optimization algorithm more stable during gradient descent and less likely to fall into local optima.

[0027] S22: converting the normalized acceleration response data sequence into polar coordinate representation, where the acceleration value corresponds to the cosine value of the angle and the timestamp corresponds to the radius; Specifically, according to the Gram matrix, calculate the angle between vectors at different times and convert it into a value between 0 and 1, thereby generating a new angle matrix. Convert the one-dimensional time series into a two-dimensional representation containing spatio-temporal information, that is, convert the normalized values into polar coordinate angles, retaining the complete information of the original signal. This step can amplify the fault characteristics: abnormal impacts (such as emergency stops) cause sudden changes in amplitude, which are reflected in the sharp corners that will appear in the polar coordinate trajectory due to the drastic changes.

[0028] S23: Reconstruct the acceleration response data sequence represented in polar coordinates into a two-dimensional image using the Gram matrix.

[0029] Specifically, use the angle matrix as the pixel values of the image to generate a two-dimensional image. Each pixel value in the image corresponds to the cosine value of the angle between different times in the time series data, thereby retaining the time information and dynamic characteristics of the data. After converting the one-dimensional data into a two-dimensional image, it is convenient to use image analysis methods to extract features, such as using a convolutional neural network (CNN) for feature extraction, so as to identify patterns in the time series data.

[0030] S3: Randomly select a support set and a query set from the two-dimensional image; Specifically, divide the two-dimensional image dataset into a training set and a test set, and set the ratio to 8:2. Initialize the support set and the query set size to 1 from the divided training set. That is, select one for training and one for testing for each class of data.

[0031] S4: Construct and train an accelerated fault detection model according to the support set and the query set. The accelerated fault detection model is used to extract features from the support set and the query set, and then perform prototype calculation on the extracted features to obtain a loss, where an intra-class distance regularization term is introduced in the prototype calculation; Specifically, refer to Figure 5 , select MetaRes-DMT-AS as the backbone of the PNML model. After feature extraction by ResNet-18, perform prototype calculation on the features input to the prototype network.

[0032] The expression of the loss L is as follows: ; where is the cross-entropy loss (Cross-EntropyLoss), is the regularization coefficient and is set to 0.1, N is the number of classes, K is the number of support samples per class, the feature of the k-th sample of class c, is the prototype of class c.

[0033] First, for each class c, calculate its class center , then input the QuerySet sample features x and class prototypes, calculate the probability that the sample x belongs to class c, then calculate the cross-entropy loss of all QuerySet samples and the squared average Euclidean distance from the support set samples to the class center, and finally weighted-combine the total loss. This loss function reduces the distance between the query samples and the true class prototypes and increases the distance from other class prototypes, ultimately accelerating the convergence of meta-training.

[0034] S5: Use the trained accelerated fault detection model for classification prediction. When a sample is misclassified, add it to the difficult sample library of the corresponding class to obtain a difficult sample library; Specifically, use the trained accelerated fault detection model for classification prediction. When a sample is predicted incorrectly, add the sample to the difficult sample library of its true class. The query set Q is composed of the difficult sample library H and ordinary samples, and is used to improve the model's recognition ability for difficult cases (such as emergency stops and severe vibrations). For example, if an emergency stop sample is predicted as overspeed, mark this emergency stop sample as a difficult sample and add it to the difficult sample library.

[0035] S6: Test the accelerated fault detection model through a test set to obtain the accuracy of the model, and adjust the support set and query set according to the accuracy; Specifically, if the accuracy rises or falls, the support set and query set will be dynamically adjusted. The specific formula is as follows: ; where is the change in accuracy, p is the dynamic adjustment parameter, P = {S, Q}, S is the size of the support set, Q is the size of the query set, t is the current training round, is p at the t-th round, is p at the (t + 1)-th round, is p the upper bound constraint of, is p the lower bound constraint of.

[0036] When the accuracy drops, expand the support set and query set to alleviate overfitting; when the accuracy improves, reduce the support set and query set to improve the generalization ability. When the accuracy has not improved for consecutive t rounds and reaches the trigger condition, parameter adjustment is performed.

[0037] S7: According to the difficult sample library, re-partition the adjusted support set and query set, and use the re-partitioned support set and query set to retrain the accelerated fault detection model; this step includes the following sub-steps: S71: Allocate the difficult sample library to the adjusted support set and query set, and retrain the accelerated fault detection model; Specifically, the support set and query set adjusted in the previous round are used for the next round of training, and parameter adjustment and performance prediction are carried out simultaneously to prepare for the next round of optimization.

[0038] First, the emergency stop samples that were predicted incorrectly in the previous round are added to the difficult sample library. At the start of the current training, the emergency stop samples in the difficult sample library are preferentially added to the query set for testing after this round of training. Then, the support set and query set are randomly selected for training, and finally, prediction is carried out after the training is completed. If the prediction is correct, the difficult sample library is removed. If the prediction fails, the next round of training will be carried out until the prediction is correct.

[0039] S72: Update the difficult sample library after training and make predictions to obtain the accuracy rate; Specifically, the accuracy rate obtained each time is compared with the accuracy rate of the previous round to detect whether the threshold is reached.

[0040] First, after the difficult sample library is obtained in the previous round and the current training is completed, the difficult sample library is preferentially predicted. If the prediction is correct, the difficult library is removed. If the prediction fails, the next round of training will be carried out until the prediction is correct.

[0041] S73: Adjust the support set and query set according to the accuracy rate until the iteration is completed.

[0042] Specifically, the support set, query set, and difficult sample library are cyclically trained and adjusted until the iteration is completed to obtain the optimal result. First, set Epochs to 100. Each Epochs trains 100 tasks, and each task undergoes different training. The performance is checked every T rounds of training (e.g., T = 10). Calculate whether t mod T = 0 reaches the detection point, and then compare the accuracy rate to record the consecutive rounds without improvement. If it reaches 5 times, the threshold is reached. If the accuracy rate drops, the support set or query set will be increased. If it rises, the samples in the support set or query set will be reduced.

[0043] S8: Input the preprocessed acceleration response data sequence into the retrained acceleration fault detection model to obtain the fault detection result.

[0044] In the experiment of collecting elevator acceleration, a total of six models were compared, and 100 iterations were used for all of them. The accuracy rates of the comparative experiments are shown in Table 1.

[0045] Table 1: As Figure 6As shown, there are significant differences in the convergence characteristics of each model during training: MetaRes-DMT-AS proposed in this application reaches the optimal recognition performance at the initial stage (initial accuracy of 97.85%), and stably converges to the optimal accuracy of 98.22% during the iterative process; the WDCNN model shows sub-optimal initial performance (96.12%), but there is an obvious learning stagnation phenomenon, and the final accuracy is 96.65%; ViT, WDCNN-GRU and WDCNN-DLSTM have similar convergence trajectories, with an average initial accuracy of 90.3±0.8%, and are respectively improved to 96.44%, 96.65% and 96.97% after iterative optimization; it is worth noting that the CCT model shows a unique overtaking characteristic, with an initial accuracy of only 82.14%, but finally reaches 97.28%.

[0046] From the analysis of the dimensions of computational efficiency and model complexity ( Figure 7 ), MetaRes-DMT-AS still maintains excellent time efficiency (34 minutes) at a relatively large parameter scale (11.18M), verifying the computational optimization effect of its architecture design. The WDCNN model shows the best resource efficiency with its streamlined number of parameters (0.66M) and the shortest iteration time (29 minutes), but there is a 1.57% performance gap between its accuracy and advanced models. The derivative models of WDCNN (WDCNN-DLSTM: 0.86M / 57min, WDCNN-GRU: 0.64M / 57min) and ViT (2.76M / 34min) form an intermediate performance cluster, and the accuracy differences among the three are less than 0.53%. The CCT model shows a significant efficiency bottleneck, with its iteration time (203 minutes) being the highest in the experimental group, but the accuracy advantage is only 97.28%.

[0047] To explore the robustness of the model in few-shot scenarios, this study conducts a fine-grained analysis of two typical few-shot anomalies (emergency stop, severe car vibration) through a confusion matrix ( Figure 8 ). While maintaining a recognition rate of 98.2±0.6% for regular samples, MetaRes-DMT-AS reaches recognition rates of 94% and 96% for few-shot categories respectively, significantly outperforming the comparison models. Although CCT, WDCNN-GRU and ViT perform well in multi-sample categories (>95%), their recognition rates for few-shot samples drop to the range of 75-84%, indicating their sensitivity to data distribution. WDCNN-DLSTM shows significant imbalance characteristics, with its emergency stop recognition rate reaching 91%, but the recognition rate for severe car vibration dropping sharply to 67%, revealing the limitations of the time series feature fusion mechanism. The WDCNN model performs poorly in both few-shot tasks (81% and 71%), verifying the deficiencies of the shallow architecture in complex pattern learning.

[0048] As can be seen from the above embodiments, the present application constructs a meta-learning framework MetaRes-DMT-AS for elevator few-shot fault diagnosis, innovatively integrating three core technologies: a dynamic meta-training mechanism, an adaptive sample scheduling module, and a regularized prototype network, systematically solving key problems such as scarce fault samples, unbalanced class distribution, and insufficient generalization ability of diagnostic models in industrial scenarios. The trained model can accurately identify high-risk faults such as elevator emergency stops and severe vibrations under extreme conditions with only a small number of samples / classes, and output highly reliable fault classification results, meeting the stringent requirements of industrial intelligent operation and maintenance systems for fault diagnosis under complex working conditions. This solution not only breaks through the performance bottleneck of traditional deep learning models in small-sample scenarios but also forms systematic technical breakthroughs in aspects such as dynamic task construction, difficult sample mining, and unsupervised feature alignment, significantly improving the practicality and engineering ability of elevator fault diagnosis in data-scarce environments.

[0049] Corresponding to the foregoing embodiment of the elevator acceleration fault detection method based on improved meta-learning, the present application also provides an embodiment of an elevator acceleration fault detection device based on improved meta-learning.

[0050] Figure 9 FIG. is a block diagram of an elevator acceleration fault detection device based on an exemplary embodiment. Referring to Figure 9 , the device includes: An acquisition module 1 for acquiring the original acceleration response data sequence of the target elevator; A preprocessing module 2 for preprocessing the original acceleration response data sequence to obtain a two-dimensional image containing time information; A data set selection module 3 for randomly selecting a support set and a query set from the two-dimensional image; A model training module 4 for constructing and training an acceleration fault detection model according to the support set and the query set. The acceleration fault detection model is used to extract features from the support set and the query set, and then perform prototype calculation on the extracted features to obtain a loss, where an intra-class distance regularization term is introduced in the prototype calculation; A sample acquisition module 5 for using the trained acceleration fault detection model for classification prediction. When a sample is misclassified, it is added to the difficult library of the corresponding class to obtain a difficult sample library, and the query set is composed of the difficult sample library and ordinary samples; A data set adjustment module 6 for testing the trained acceleration fault detection model through a test set to obtain the accuracy of the model, and adjusting the support set and the query set according to the accuracy; The model retraining module 7 is configured to re-partition the adjusted support set and query set according to the difficult sample library, and use the re-partitioned support set and query set to retrain the accelerated fault detection model to predict the operating state of the elevator; The fault detection module 8 is configured to input the preprocessed acceleration response data sequence into the retrained accelerated fault detection model to obtain a fault detection result.

[0051] Regarding the device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated herein.

[0052] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this application. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0053] Correspondingly, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the elevator acceleration fault detection method based on improved meta-learning as described above.

[0054] Correspondingly, this application also provides a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the elevator acceleration fault detection method based on improved meta-learning as described above is implemented.

[0055] Those skilled in the art will readily think of other implementation schemes of this application after considering the specification and practicing the content disclosed herein. This application is intended to cover any variations, uses, or adaptive changes of this application, which follow the general principles of this application and include the common general knowledge or conventional technical means in the technical field not disclosed in this application. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of this application are pointed out by the claims.

[0056] It should be understood that this application is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is only limited by the appended claims.

Claims

1. An elevator acceleration fault detection method based on improved meta - learning, characterized in that, Including: Obtain the original acceleration response data sequence of the target elevator; Preprocess the original acceleration response data sequence to obtain a two-dimensional image containing time information; Randomly select a support set and a query set from the two-dimensional image; According to the support set and the query set, construct and train an acceleration fault detection model, which is used to extract features from the support set and the query set, and then perform prototype calculation on the extracted features to obtain a loss, where an intra-class distance regularization term is introduced in the prototype calculation; Use the trained acceleration fault detection model for classification prediction. When a sample is misclassified, add it to the difficult sample library corresponding to the category to obtain a difficult sample library, and the query set is composed of the difficult sample library and normal samples; Test the trained acceleration fault detection model through a test set to obtain the accuracy of the model, and adjust the support set and the query set according to the accuracy; According to the difficult sample library, re-divide the adjusted support set and query set, and use the re-divided support set and query set to retrain the acceleration fault detection model to predict the operating state of the elevator; Input the preprocessed acceleration response data sequence into the retrained acceleration fault detection model to obtain a fault detection result.

2. The method according to claim 1, wherein Preprocess the original acceleration response data sequence to obtain a two-dimensional image containing time information, including: Perform normalization processing on the original acceleration response data sequence; Convert the normalized acceleration response data sequence into a polar coordinate representation, where the acceleration value corresponds to the cosine value of the angle and the timestamp corresponds to the radius; Use the Gram matrix to reconstruct the acceleration response data sequence in polar coordinate representation into a two-dimensional image.

3. The method according to claim 1, characterized in that, The expression of the loss L is as follows: ； where is the cross-entropy loss, is the regularization coefficient, N is the number of classes, K is the number of support samples per class, the feature of the k-th sample of class c, is the prototype of class c.

4. The method according to claim 1, characterized in that According to the accuracy, adjust the support set and the query set, including: If the accuracy rises or falls, the support set and the query set will be dynamically adjusted, and the specific formula is as follows: ； where is the change in accuracy rate, p is the dynamically adjusted parameter, P = {S, Q}, S is the size of the support set, Q is the size of the query set, and t is the current training round, is p the value at the t-th round, is p the value at the (t + 1)-th round, is p the upper bound constraint of, is p the lower bound constraint of; When the accuracy drops, expand the support set and the query set to alleviate overfitting; when the accuracy improves, reduce the support set and the query set to improve the generalization ability.

5. The method according to claim 1, wherein According to the difficult sample library, re-divide the adjusted support set and query set, and use the re-divided support set and query set to retrain the acceleration fault detection model, including: Allocate the difficult sample library to the adjusted support set and query set, and perform retraining on the acceleration fault detection model; Update the difficult sample library after training and perform prediction to obtain the accuracy; According to the accuracy, readjust the support set and the query set until the iteration is completed.

6. The method according to any one of claims 1 to 5, characterized in that The operating state includes elevator reset operation, severe car vibration, emergency stop fault, normal operation, and overspeed.

7. An elevator acceleration fault detection device based on improved meta - learning, characterized in that, Including: An acquisition module for acquiring the original acceleration response data sequence of the target elevator; A preprocessing module for preprocessing the original acceleration response data sequence to obtain a two-dimensional image containing time information; A data set selection module for randomly selecting a support set and a query set from the two-dimensional image; A model training module, configured to construct and train an accelerated fault detection model according to the support set and the query set, where the accelerated fault detection model is used to extract features from the support set and the query set, and then perform prototype calculation on the extracted features to obtain a loss, and a within-class distance regularization term is introduced in the prototype calculation; A sample acquisition module, configured to use the trained accelerated fault detection model for classification prediction, and add the sample to the corresponding difficult sample library when the sample is misclassified, to obtain a difficult sample library, and the query set is composed of the difficult sample library and normal samples; A data set adjustment module, configured to test the trained accelerated fault detection model through a test set to obtain the accuracy of the model, and adjust the support set and the query set according to the accuracy; A model retraining module, configured to re-partition the adjusted support set and query set according to the difficult sample library, and use the re-partitioned support set and query set to retrain the accelerated fault detection model to predict the operating state of the elevator; A fault detection module, configured to input the preprocessed acceleration response data sequence into the retrained accelerated fault detection model to obtain a fault detection result.

8. An electronic device, characterized in that, Comprising: One or more processors; A memory, configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-6.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the instruction is executed by the processor, the steps of the method according to any one of claims 1-6 are implemented.

Images