A method for monitoring the state of elevator brakes based on multi-feature adaptive fusion

Abstract

This invention relates to the field of elevator brake condition monitoring, and provides a method for elevator brake condition monitoring based on multi-feature adaptive fusion. This method first filters the state parameters of the elevator brake during operation to remove abnormal information, and then fills in any existing missing values ​​or missing values ​​caused by the aforementioned reasons. Secondly, it uses an attention-based long-short-term anomaly feature adaptive fusion network model. The long-term trend feature network extracts long-term trend features during elevator brake operation, and the short-term anomaly feature network extracts short-term anomaly features. Then, an attention-based feature fusion network comprehensively extracts and fuses the long- and short-term features, thereby improving the accuracy and robustness of the monitoring results. Finally, it combines extreme value theory to realize elevator brake operation condition monitoring, which is of great significance for the application of deep learning models in the field of elevator brake condition monitoring.

Description

Technical Field

[0001] This invention relates to the field of elevator brake status monitoring, and in particular to an elevator brake status monitoring method based on multi-feature adaptive fusion. Background Technology

[0002] As a crucial component ensuring the safe stopping and precise control of elevators, the elevator brake bears a significant workload during actual operation, and its health status is paramount to elevator safety. However, due to various factors affecting elevator operation, such as usage time, load variations, and environmental conditions, the brake may experience performance degradation, malfunction, or even failure. Therefore, real-time monitoring of the brake is essential to ensuring elevator safety.

[0003] Currently, elevator brake monitoring methods are mainly divided into two types: monitoring based on traditional engineering techniques and monitoring based on data analysis. Traditional monitoring methods mostly focus on periodic inspections and manual patrols. This approach is not only inefficient but also prone to missing some sudden abnormal conditions. In addition, traditional physical model-based monitoring methods also have limitations. They rely heavily on empirical formulas and simple threshold judgments, making it difficult to cope with the complex and ever-changing elevator operating environment.

[0004] While some studies in recent years have begun to apply time-series data analysis techniques to monitor the condition of elevator brakes, existing technologies still have many limitations. Traditional time-series data analysis methods typically focus on a single time scale, either concentrating only on long-term trends to help identify overall changes in the elevator brake's operating status, or focusing on short-term abnormal fluctuations to capture immediate faults or anomalies. These methods often fail to comprehensively reflect the brake's health condition, and the information from different time scales is not effectively combined, resulting in low monitoring accuracy and an inability to detect potential faults in a timely manner.

[0005] Furthermore, although some methods attempt to fuse long-term and short-term features, existing fusion methods still have certain problems due to limitations in feature extraction and fusion techniques. Existing long-term and short-term feature fusion techniques often fail to achieve deep fusion across multiple dimensions, resulting in insufficient utilization of information at each time scale, thus affecting the accuracy and robustness of monitoring results.

[0006] Therefore, existing elevator brake condition monitoring technologies have significant shortcomings in dealing with complex operating conditions, improving monitoring accuracy, and enhancing system robustness. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an elevator brake status monitoring method based on multi-feature adaptive fusion. By filtering the state parameters of the elevator brake during operation and filling in missing values, and using an adaptive fusion network model of long-short-term anomaly features based on attention mechanism, combined with extreme value theory, the elevator brake operation status monitoring is realized.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] This invention discloses a method for monitoring the state of an elevator brake based on multi-feature adaptive fusion, comprising: Step 1: collecting various operating parameter data of the elevator brake during operation to establish a dataset, and deleting abnormal information from the dataset; Step 2: filling the dataset after deleting abnormal information in Step 1 with local mean method, and then standardizing the data; Step 3: selecting state parameters related to the operating state of the elevator brake based on Pearson and Spearman correlation coefficients, and then extracting a sample set through a sliding window. x The sample set is divided into normal historical data and the dataset to be tested. The normal historical data is used for further execution of step 4, and the dataset to be tested is used as a test set. Step 4: Construct an elevator brake status monitoring network model with adaptive fusion of long-term and short-term anomaly features, including setting the hyperparameters of the long-term trend feature module L, the short-term anomaly feature module S, and the adaptive feature fusion module F, and defining the loss function. L MSE Step 5: Randomly initialize network parameters, input the sample set x from Step 3 into the long-term trend feature module L and the short-term anomaly feature module S according to the batch size, and output long-term feature l and short-term feature s; Step 6: Input the long-term feature l and short-term feature s from Step 5 into the adaptive feature fusion module F, and output the... x Reconstructed samples of the same size ,calculate x and Mean square error reconstruction loss between L MSE Step 7: Based on the loss function L MSE Step 8: Backpropagation is used to calculate the gradient of each network layer, and the L, S, and F network parameters are updated using the optimizer; Step 9: Repeat steps 5 to 7. When the training reaches the preset number of rounds M, save the network parameters of the trained long-short-term anomaly feature adaptive fusion network model; Step 10: Input the dataset to be detected into the trained long-short-term anomaly feature adaptive fusion network model to obtain the reconstruction error of the dataset to be detected as the test set, and calculate the anomaly score; Step 11: Use extreme value theory to determine the anomaly threshold of the test set and output the elevator brake operation status monitoring results.

[0010] The advantages of this invention are as follows: The aforementioned method can filter and fill in missing values ​​for the state parameters during the operation of the elevator brake, and use an attention-based long-short-term anomaly feature adaptive fusion network model to extract long-term trend features during the operation of the elevator brake through a long-term trend feature network and short-term anomaly features through a short-term anomaly feature network. Then, an attention-based feature fusion network is used to comprehensively extract and fuse long- and short-term features, thereby improving the accuracy and robustness of the monitoring results. Finally, combined with extreme value theory, the elevator brake operation status monitoring is realized, achieving accurate and efficient elevator brake status monitoring. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a method for monitoring the state of an elevator brake based on multi-feature adaptive fusion in an embodiment of the present invention.

[0012] Figure 2 This is a visualization of the test results of Example 1 in this embodiment of the invention;

[0013] Figure 3 This is a visualization of the test results in Example 2 of this embodiment of the invention;

[0014] Figure 4 The diagram shows the effect of monitoring the elevator brake's operating status in Example 1 of the present invention, comparing various existing methods with those in the embodiments of the present invention.

[0015] Figure 5 The diagram shows the effect of monitoring the elevator brake's operating status in Example 2 of the present invention, comparing various existing methods with those in the embodiments of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1 This application discloses an elevator brake status monitoring method based on multi-feature adaptive fusion, comprising:

[0018] Step 1: Collect various operating parameter data of the elevator brake during operation to establish a dataset, and remove abnormal information from the data. For example, the abnormal information removal process includes deleting the dataset, removing variables with a large number of missing values ​​(null values) and variables with almost no change (variance less than 0.1) as part of the deleted dataset.

[0019] Step 2: For the dataset after Step 1, use the local mean method to fill in the data, so that the original missing data in the dataset and the missing data caused by the deletion of abnormal information in Step 1 are filled in, and then the data is standardized.

[0020] Specifically, step 2, which involves filling data using the local mean, includes: filling missing rows with data by using the data from the preceding and following rows. The missing row is filled by calculating the average of the five non-nullable rows before and after it, as shown in the following formula:

[0021]

[0022] In the formula, It is a missing data value. These are the five rows of data that are not null before the rows with missing data. These are the five rows of data that are not null after the missing rows. This indicates that the average value of the data is being calculated.

[0023] Step 3: Perform variable correlation analysis on the dataset processed in Step 2. Based on the Pearson and Spearman correlation coefficients, filter the state parameters related to the elevator brake's operating state, and then use a sliding window to extract a sample set of size [U×C]. x To create a data sample set, the sample set is divided into normal historical data and a dataset to be tested. The normal historical data is input into the elevator brake status monitoring network model in step 4, and the dataset to be tested is used as a test set.

[0024] Step 4: Construct an elevator brake state monitoring network model with adaptive fusion of long-term and short-term anomaly features, including setting the hyperparameters of the long-term trend feature module L, the short-term anomaly feature module S, and the adaptive feature fusion module F, and defining the loss function. L MSE .

[0025] Specifically, the hyperparameters of the long-short-term anomaly feature adaptive fusion network model in step 4 include at least the number of network layers, the number of hidden units, and the size of the convolutional kernel.

[0026] Step 5: Randomly initialize network parameters, input the sample set x from step 3 into the long-term trend feature module L and the short-term anomaly feature module S according to the batch size, and output long-term feature l and short-term feature s with size [U×T].

[0027] Specifically, in step 5, the long-term trend feature module L is characterized by using large convolutional kernels to increase the effective receptive field, and by decoupling information in the variable, time, and feature dimensions to comprehensively capture multiple dependencies in the entire long-term time series; the adaptive feature fusion module F is characterized by using channel and spatial attention mechanisms and an inverted residual multilayer perceptron to adaptively fusion and extract long- and short-term features, and its calculation formula is as follows:

[0028]

[0029] in, This represents element-wise multiplication. It is based on the combination generation of channel attention. It is a combination generation based on spatial attention, where LN represents the normalization layer and Concat represents concatenation, as shown in the formula. and Features are generated by connecting through IRMLP (Inverted Residual Multilayer Perceptron). CA represents channel attention mechanism, and SA represents spatial attention mechanism.

[0030] The formulas for the channel and spatial attention mechanisms are as follows:

[0031]

[0032] In the formula, For channel attention mechanism, For spatial attention mechanisms, MLP stands for Multilayer Perceptron, AvgPool is average pooling, and MaxPool is max pooling. It is the Sigmoid function. The kernel size is Convolution operation;

[0033] The formula for the inverted residual multilayer perceptron is as follows:

[0034]

[0035] In the formula, express Convolution dimensionality reduction operation, express Convolution dimensionality upscaling operation, express In depthwise convolution operations, GELU is the activation function.

[0036] Step 6: Input the long-term feature l and short-term feature s from Step 5 into the adaptive feature fusion module F, and output the result. x Reconstructed samples of the same size ,x and All dimensions are [U×C]. Calculate. x and Mean square error reconstruction loss between L MSE The model reconstruction error is obtained.

[0037] Step 7: Based on the loss function L MSE Backpropagation is used to calculate the gradients of each network layer, and the L, S, and F network parameters are updated using an optimizer.

[0038] Specifically, the purpose of backpropagating to update the parameters in step 7 is to obtain the minimum loss value. L MSE The formula for the mean square error loss is as follows:

[0039]

[0040] In the formula, It is a reconstructed sample. It is a 2-norm.

[0041] Step 8: Repeat steps 5 to 7. When the training reaches the preset number of rounds M, the trained long-short-term anomaly feature adaptive fusion network model is represented as follows: Save its network parameters;

[0042] Step 9: Input the dataset to be detected into the trained network model. The reconstruction error of the dataset to be detected is obtained as the test set, and the anomaly score is calculated.

[0043] Step 10: Use EVT extreme value theory to determine the abnormal threshold of the test set and output the elevator brake operation status monitoring results.

[0044] Specifically, the test set anomaly threshold determined by extreme value theory in step 10... The calculation formula is as follows:

[0045]

[0046] In the formula Let be the initial threshold for model error, q be the probability of anomalies in the entire sequence, and N be the number of observations. for Number of times, For parameters in the generalized Pareto distribution. This is the standard deviation estimated from the sample data.

[0047] It is understood that the embodiments of this application are based on a feature adaptive fusion architecture of parallel dual branches (long-term L + short-term S) and a fusion module F. The long-term trend module uses a large convolutional kernel (51×1) to capture macroscopic dependencies. Furthermore, it achieves deep fusion in multiple dimensions by fusing channel attention, spatial attention, and the multi-dimensional deep feature interaction mechanism of IRMLP, thus fully utilizing information at each time scale.

[0048] The method in this embodiment filters and fills in missing values ​​for the state parameters of the elevator brake during operation. It uses an attention-based long-short-term anomaly feature adaptive fusion network model to extract long-term trend features of the elevator brake during operation through a long-term trend feature network and short-term anomaly features through a short-term anomaly feature network. Then, it uses an attention-based feature fusion network to comprehensively extract and fuse long- and short-term features, thereby improving the accuracy and robustness of the monitoring results. Finally, it combines extreme value theory to realize the monitoring of the elevator brake's operating status, achieving accurate and efficient elevator brake status monitoring.

[0049] Next, we will further illustrate the above method with a specific example.

[0050] Step 1: Collect various operating parameter data of the elevator brake during operation to establish a dataset, and remove abnormal information from the data. For example, the abnormal information removal process includes deleting the dataset, removing variables with a large number of missing values ​​(null values) and almost no change (variance less than 0.1) as part of the deleted dataset. Specifically, this includes:

[0051] A dataset is built by collecting signals (such as temperature and speed) from various sensors installed on the elevator brake during its operation. The elevator brake is primarily used to control the elevator's movement, and abnormal conditions can lead to insufficient braking force, excessive clearance, uneven wear, partial wear, total wear, oil contamination, and surface carbonization. Then, for the collected elevator brake operation parameters, missing and unchanged data are removed.

[0052] Step 2: For the dataset after Step 1, use the local mean method to fill in the data, so that the original missing data in the dataset and the missing data caused by the deletion of abnormal information in Step 1 are filled in. Then, the data is standardized. For the specific process, please refer to the above embodiment, which will not be repeated here.

[0053] The Pearson and Spearman correlation coefficients were used to screen out the operating status parameters that are closely related to the elevator brake's operating status from the imputed data. The remaining operating status parameters after screening are shown in Table 1.

[0054] Table 1. Data remaining after deleting overall missing data, deleting unchanged data, and feature filtering.

[0055]

[0056] Step 3 involves performing a correlation analysis on the dataset processed in Step 2. State parameters related to the elevator brake's operating state are selected based on Pearson and Spearman correlation coefficients. Then, a sample set of size [U×C] is extracted using a sliding window. x To create a data sample set, the sample set is divided into normal historical data and a dataset to be tested. The normal historical data is input into the elevator brake status monitoring network model in step 4, and the dataset to be tested is used as a test set.

[0057] In this example, the preprocessed data is used to extract samples through a sliding window and divided into training and example test sets. The window length is set to 60 and the sliding step is 10, that is, the size is [60, 6]. The test set settings are shown in Table 2.

[0058] Table 2 Detailed Settings for Training and Case Test Sets

[0059]

[0060] Step 4: Construct an elevator brake state monitoring network model based on the PyTorch deep learning framework. The experimental equipment configuration is: PyTorch 1.11.0 + CUDA 11.3, Python 3.9.20, and training and computation are performed on an NVIDIA GeForce RTX 3090 graphics card. An elevator brake state monitoring network model with adaptive fusion of long- and short-term anomaly features is established. The specific parameters of the network structure of the method in this embodiment are shown in Table 3.

[0061] Table 3 Key parameter settings for network structure in the embodiments of the present invention

[0062]

[0063] The network model uses the AdamW optimizer with weight decay for parameter optimization, with an initial learning rate of 0.01 and a weight decay of 10. -5 The model was trained using a step size scheduler of 5, with a batch size of 128, for a total of 50 training rounds. This embodiment selects methods that extract features from a single feature perspective, such as TCN, SCINet, Transformer, and TranAD, as comparative methods for extracting features from both long-term and short-term perspectives simultaneously, such as MICN and DTAAD.

[0064] Then, steps 5-8 are performed to obtain the parameters of the adaptive fusion network model based on long- and short-term anomaly features. Finally, steps 9-10 are executed, where the dataset to be detected in the dataset is used as the test set and input into the trained model to obtain anomaly scores. The threshold is then determined using extreme value theory to output the state monitoring results.

[0065] See Figure 2 The image shows a t-SNE visualization of the test set data in Example 1, representing the normal phase (left) and the abnormal phase (right). The left image shows that during the normal period, the method of this embodiment can reconstruct the generated data well. However, during the abnormal phase, there is a significant difference between the actual data and the reconstructed data. Therefore, the method of this embodiment can effectively monitor abnormal conditions in the elevator brake's operating state. (See reference...) Figure 3 As can be seen, the same situation exists in Example 2, indicating that the monitoring effect of the method in this embodiment is not accidental, but rather universal and capable of monitoring various fault conditions. For the specific monitoring effects of each method on Example 1, please refer to... Figure 4 It can be seen that the method in this embodiment achieves good monitoring results, with significant advantages in terms of accuracy, recall, AUC value, and F1 score. (See also...) Figure 5 The specific monitoring effect scores of the methods in the existing technologies compared to Example 2 show that the method in this embodiment achieves better results in all indicators, especially in the F1 score, which reflects the comprehensive score.

[0066] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for monitoring the state of an elevator brake based on multi-feature adaptive fusion, characterized in that, include: Step 1: Collect various operating parameter data of the elevator brake during operation to establish a dataset, and delete abnormal information from the dataset; Step 2: For the dataset after removing outliers in Step 1, fill the data with the local mean method and then standardize the data; Step 3: Based on the Pearson and Spearman correlation coefficients, screen out the state parameters related to the elevator brake's operating state, and then use a sliding window to extract the sample set. x The sample set x is divided into normal historical data and the dataset to be tested. The normal historical data is used to further execute step 5, and the dataset to be tested is used as a test set. Step 4: Constructing an elevator brake state monitoring network model with adaptive fusion of long-term and short-term anomaly features, including: setting the hyperparameters of the long-term trend feature module L, the short-term anomaly feature module S, and the adaptive feature fusion module F, and defining the loss function. L MSE ; Step 5: Randomly initialize network parameters, input the sample set x from Step 3 into the long-term trend feature module L and the short-term anomaly feature module S according to the batch size, and output the long-term feature l and the short-term feature s; Step 6: Input the long-term feature l and short-term feature s from Step 5 into the adaptive feature fusion module F, and output the result. x Reconstructed samples of the same size Based on loss function L MSE calculate x and The mean squared error between them is used to reconstruct the model error; Step 7: Based on the loss function L MSE Backpropagation is used to calculate the gradients of each network layer, and the L, S, and F network parameters are updated using the optimizer. Step 8: Repeat steps 5 to 7. When the training reaches the preset number of rounds M, save the network parameters of the long-short-term anomaly feature adaptive fusion network model after training. Step 9: Input the dataset to be detected into the trained long-short-term anomaly feature adaptive fusion network model to obtain the reconstruction error and calculate the anomaly score; Step 10: Use extreme value theory to determine the abnormal threshold of the test set and output the elevator brake operation status monitoring results.

2. The elevator brake status monitoring method based on multi-feature adaptive fusion as described in claim 1, characterized in that, The method of filling data with local mean includes: To fill in missing data rows, the missing data row is filled by averaging the values ​​from the five non-null rows before and after the missing row. The calculation formula is shown below: In the formula, It is a missing data value. These are the five rows of data that are not null before the rows with missing data. These are the five rows of data that are not null after the missing rows. This indicates that the average value of the data is being calculated.

3. The elevator brake status monitoring method based on multi-feature adaptive fusion as described in claim 1, characterized in that, The hyperparameters of the adaptive fusion network model for long- and short-term anomalies include the number of network layers, the number of hidden units, and the size of the convolutional kernel.

4. The elevator brake status monitoring method based on multi-feature adaptive fusion as described in claim 1, characterized in that, The long-term trend feature module L uses a large convolutional kernel; The adaptive feature fusion module F employs channel and spatial attention mechanisms, along with an inverted residual multilayer perceptron, to adaptively fusion and extract long-short-term features. Its calculation formula is shown below: in, This represents element-wise multiplication. It is based on the combination generation of channel attention. It is based on spatial attention-based combinatorial generation, and finally... and Features are generated by connecting through IRMLP. LN represents the normalization layer, Concat represents concatenation, CA represents channel attention mechanism, SA represents spatial attention mechanism, and the subscript i represents the intermediate variable; The formulas for the channel and spatial attention mechanisms are as follows: In the formula, For channel attention mechanism, This is a spatial attention mechanism, in which, It is the Sigmoid function. The kernel size is The convolution operation is performed by MLP (Multilayer Perceptron), AvgPool (Average Pooling), and MaxPool (Max Pooling). The formula for the inverted residual multilayer perceptron (IRMLP) is as follows: In the formula, express Convolution dimensionality reduction operation, express Convolution dimensionality upscaling operation, express In depthwise convolution operations, GELU is the activation function.

5. The elevator brake status monitoring method based on multi-feature adaptive fusion as described in claim 1, characterized in that, The loss function L MSE : In the formula, It is a reconstructed sample. It is a 2-norm.

6. The elevator brake status monitoring method based on multi-feature adaptive fusion as described in claim 1, characterized in that, Determining the outlier threshold of the test set using extreme value theory The elevator brake operation status monitoring results are output, including: In the formula This is the initial threshold for the model error. Let N be the probability of an anomaly occurring in the entire sequence, and N be the number of observations. for Number of times, For parameters in the generalized Pareto distribution. This is the standard deviation estimated from the sample data.

Images