Neural network vehicle braking system fault diagnosis and prediction system based on attention

Abstract

The invention discloses an attention-based neural network vehicle braking system fault diagnosis and prediction system and method, and the system comprises a data enhancement module which is used for obtaining time sequence data, and forming an enhanced time sequence through the control of multiple attention; and the fault diagnosis and forecast module is used for allowing the attention mechanism layer comprising at least one multi-head attention to perform joint processing at different positions of the enhanced time sequence so as to perform fault diagnosis and forecast on the braking system of the vehicle. The method has the advantages that the method is easy to implement, data easier to understand is provided for a user by enhancing the resolution and the signal-to-noise ratio of the input data, and especially when the input data is affected by errors such as noise or data missing, the performance and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of vehicle braking, and relates to a neural network vehicle braking system fault diagnosis and prediction system and method based on attention. BACKGROUND

[0002] Air brake systems are commonly used in heavy vehicles, such as buses and trucks. The air brake system is composed of an air compressor, a governor, an air tank, a brake pedal, a brake chamber, and various valves and pipes. The current maintenance method for transport vehicles, particularly their air brake systems, includes a combination of preventive maintenance and corrective maintenance. Preventive maintenance of the air brake system includes checking the function and components of the air brake system, while the initiation of corrective maintenance mainly relies on the driver's knowledge of the air brake system condition. Although the current approach enables the vehicle to operate reliably, there are some limitations. For example, resources are wasted due to repairs or replacements of still functioning components. This will increase costs, including materials, human resources, and service time of the vehicle. The repair schedule also lacks flexibility and cannot be adjusted to the equipment state. Moreover, these maintenance works cannot be performed until a failure occurs in the system. In addition, the need for corrective maintenance will raise concerns about reliability and unexpectedly increase costs. Therefore, predictive maintenance is adopted to overcome the shortcomings of existing preventive maintenance.

[0003] The above-mentioned predictive maintenance is achieved by integrating failure physics (PoF) analysis and artificial intelligence (AI). Failure physics analysis studies the regularity of product and system failure according to the failure cause and mechanism, thereby providing insights for implementing preventive maintenance. Artificial intelligence, particularly data-driven models (DDM), plays a crucial role in predictive maintenance, which relies on continuous data collection from on-board sensors and analyzes these data to diagnose usage patterns or wear symptoms. DDMs that employ machine learning techniques can determine the equipment condition by analyzing the data and automatically suggest appropriate maintenance. More specifically, DDMs can accomplish the tasks required for predictive maintenance, i.e., anomaly detection, fault diagnosis, and prognosis. However, such maintenance only considers the state of the machine and only suggests repairs or replacements when necessary.

[0004] The models in existing solutions are either machine learning models or deep neural networks, which use raw data or data pre-processed by typical procedures as input. The accuracy and robustness of machine learning models can be weak, while deep neural networks require huge computational costs and have poor interpretability of inference results. It is even more difficult to understand the inference results of the model if the input is raw data in the form of multivariate irregularly sampled time series with observation time misalignment, or data pre-processed by typical procedures.

[0005] The patent document US20220068051A1 of the United States discloses a system and method for predicting vehicle component failure and issuing a warning to the driver accordingly. Sensor data from sensors can include CAN signals, which will be further used to predict vehicle component failure through various machine learning techniques.

[0006] The patent document US11410048B2 of the United States discloses a deep learning-based anomaly event detection, which can include a system for device anomaly event detection. The system includes a computing device having a processor, an encoding module, and a decoding module. The processor is configured to receive sensor data. The encoding module generates reconstruction data from the sensor data, identifies at least one reconstruction error in the reconstruction data, and determines an anomaly score based on the at least one reconstruction error. The decoding module generates an action prediction from the sensor data and determines a likelihood value based on the action prediction. Then, the processor can calculate a scaled anomaly score based on the anomaly score and the likelihood value, so that the processor performs an operation based on the scaled anomaly score. A neural network composed of an autoencoder is used to calculate the anomaly score.

[0007] The patent document CN112418411A of China discloses an air brake system fault diagnosis method based on RBF neural network. The method uses genetic algorithm to optimize the training time of RBF neural network, and pre-trains a large number of samples on the system, so as to realize air brake system fault diagnosis. The fault diagnosis method includes the following steps: obtaining the data of the air brake system through the sensor, preprocessing the data, inputting the data into the trained RBF neural network, generating the output recognition result, and feeding back the diagnosis result to the trained RBF neural network.

[0008] However, when the above machine learning model or deep neural network cannot focus only on useful information in the time series, its accuracy and robustness can be weak. Using deep neural networks can also require high computational cost, and the inference result can not be optimal. Therefore, it is necessary to provide an improved system and method to predict vehicle component failure with high precision and high efficiency. In addition, deep neural network prediction based on multivariate irregularly sampled time series or noisy time series can not be easy to understand, which can be challenging for user analysis work. SUMMARY

[0009] The purpose of the present application is to overcome the above-mentioned shortcomings and provide a neural network vehicle brake fault diagnosis and prediction system based on attention, which can solve various problems existing in the prior art.

[0010] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is: a neural network vehicle brake system fault diagnosis and prediction system based on attention, comprising:

[0011] a data augmentation module based on a deep learning model, the module having at least one attention mechanism layer;

[0012] wherein the attention mechanism layer comprises a multi-head attention; and

[0013] a fault diagnosis module and a fault prediction module based on at least one deep learning model having at least one attention mechanism layer.

[0014] Further, the braking system comprises an air braking system.

[0015] Further, the vehicle comprises a heavy road vehicle.

[0016] Further, the fault diagnosis and prediction system further comprises a data collection module and a data pre-processing module.

[0017] Further, the attention mechanism layer is configured to select useful information from the time series data.

[0018] Further, the attention mechanism layer is trained to identify and prioritize the most useful information.

[0019] Further, the attention mechanism layer of the multi-head attention is configured to project the input of the time series data into multiple subspaces and perform scaled dot-product attention in each subspace.

[0020] Further, the attention mechanism layer comprises a masked multi-head attention configured to prevent unwanted dependencies from different positions of the time series.

[0021] Further, the output of the multi-head attention is formed by the outputs of the scaled dot-product attention from different subspaces.

[0022] Further, the multi-head attention is further configured to allow the deep learning model to jointly attend to information from different representation subspaces at different positions of the time series.

[0023] Another object of the present application is to provide an attention-based neural network vehicle braking fault diagnosis and prediction method, comprising:

[0024] data augmentation based on a deep learning model having at least one attention mechanism layer; and

[0025] fault diagnosis and prediction of the braking system based on a deep learning model having at least one attention mechanism layer. Further, the augmented data is based on previously collected and pre-processed data.

[0026] Further, the augmented data comprises:

[0027] selecting useful information from time series data;

[0028] projecting inputs of time series data into multiple subspaces and performing scaled dot-product attention in each subspace; and

[0029] forming outputs of multi-headed attention by concatenating outputs of scaled dot-product attention from different subspaces.

[0030] Further, the step of masking unwanted information is included before the forming outputs of multi-headed attention.

[0031] Further, the forming of the outputs of multi-headed attention further allows a deep learning model to jointly attend to information from different representation subspaces at different locations of time series.

[0032] Further, the data augmentation further includes training the attention mechanism layer to identify and prioritize the most useful information.

[0033] Further, after diagnosing and predicting brake system faults, the results are stored and displayed.

[0034] The present application has the following advantages compared with the prior art:

[0035] It is simple to implement, including: a data augmentation module for obtaining time series data, forming augmented time series through the control of multi-headed attention; a fault diagnosis and prediction module for allowing an attention mechanism layer containing at least one multi-headed attention to jointly process at different locations of the augmented time series to diagnose and predict faults of a vehicle's brake system. By enhancing the resolution and signal-to-noise ratio of the input data, it provides users with more easily understandable data, especially when the input data is affected by errors such as noise or data missing, improving performance and accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0036] The features of the present application will be more readily understood and appreciated when considered in reference to the following detailed description, taken in conjunction with the accompanying drawings, in which:

[0037] Figure 1 shows an attention-based neural network vehicle brake fault diagnosis and prediction system of the present application;

[0038] Figure 2 shows a schematic diagram of multi-headed attention of the present application;

[0039] Figure 3 is an implementation flowchart of an attention-based neural network vehicle brake fault diagnosis and prediction method of the present application. DETAILED DESCRIPTION

[0040] In accordance with the requirements of the present disclosure, specific embodiments of the present application are disclosed. It should be understood, however, that the disclosed embodiments are merely examples of the present application, which can be practiced in variously diverse forms. Accordingly, the specific structural and functional details disclosed herein are not to be interpreted as limiting, but merely as a basis for the claims. It should be understood that the drawings and detailed description thereto are not intended to limit the application to the particular form disclosed, but on the contrary, the application covers all modifications, equivalents and alternatives falling within the scope of the claims. As used throughout this application, the word "may" is used in the permissive sense (i.e., meaning having the potential to), rather than the mandatory sense (i.e., meaning must). Similarly, the words "include," "including," and "includes" mean including, but not limited to. Furthermore, the word "a" means "at least one" and the word "one" means "one or more." In using abbreviations or technical terms, these refer to the generally accepted meanings known in the art.

[0041] As Figure 1 shown, the present application discloses an attention-based neural network vehicle braking fault diagnosis and prediction system, comprising: a data augmentation module based on a deep learning model, the module having at least one attention mechanism layer; wherein the attention mechanism layer comprises multi-head attention 102; and a fault diagnosis module and a fault prediction module based on at least one deep learning model, the deep learning model having at least one attention mechanism layer. The attention mechanism layer can also include other ways of attention mechanism in the present application.

[0042] In a preferred embodiment of the present application, the braking system includes, but is not limited to, an air braking system.

[0043] In a preferred embodiment of the present application, the vehicle includes, but is not limited to, a heavy road vehicle.

[0044] In a preferred embodiment of the present application, the system further comprises a data collection module and a data preprocessing module.

[0045] In a preferred embodiment of the present application, the attention mechanism layer is configured to select useful information from time series data. The attention mechanism layer is trained to identify and prioritize the most useful information.

[0046] In a preferred embodiment of the present invention, the attention mechanism comprising the multi-head attention 102 is configured to project the input of the time series data into multiple subspaces and perform scaled dot-product attention 106 in each subspace. The output of the multi-head attention 102 is formed by concatenating the outputs of the scaled dot-product attention 106 from different subspaces. The multi-head attention 102 is further configured to allow the deep learning model to jointly focus on information from different representation subspaces at different locations of the time series. The attention mechanism layer comprises the masked multi-head attention 104, which is configured to prevent unwanted dependencies from different locations of the time series, for example, locations with erroneous data or locations of later time steps. The masked multi-head attention 104 is further configured to prevent invalid temporal relationships in the fault prediction module.

[0047] The output of the data augmentation module forms an augmented time series to improve the resolution and signal-to-noise ratio of the time series. For example, inputting a multivariate irregularly sampled time series to the data augmentation module, the augmented time series will output a multivariate regularly sampled time series with a higher sampling rate.

[0048] The present invention discloses a method of fault diagnosis and prediction of a vehicle braking system using an attention-based neural network, comprising the steps of: data augmentation based on a deep learning model comprising at least one attention mechanism layer; and fault diagnosis and prediction of the braking system based on a deep learning model comprising at least one attention mechanism layer. The method of the present invention comprises but is not limited to merging at least two deep learning models of data augmentation, fault diagnosis, or fault prediction into one model.

[0049] In a preferred embodiment of the present invention, the method comprises the steps of: collecting and pre-processing the data before augmenting the data.

[0050] In a preferred embodiment of the present invention, augmenting the data comprises the steps of: selecting useful information from the time series data; projecting the input of the time series data into multiple subspaces and performing scaled dot-product attention 106 in each subspace; and forming the output of the multi-head attention 102 by concatenating the outputs of the scaled dot-product attention 106 from different subspaces. The formation of the output of the multi-head attention 102 further allows the deep learning model to jointly focus on information from different representation subspaces at different locations of the time series.

[0051] In a preferred embodiment of the present invention, augmenting the data further comprises the step of masking unwanted information before forming the output of the multi-head attention 102.

[0052] In a preferred embodiment of the present invention, data augmentation further comprises the step of: training the attention mechanism layer to identify and prioritize the most useful information.

[0053] In a preferred embodiment of the present application, the method further comprises the steps of storing and displaying the results after diagnosing and prognosticating the brake system failure.

[0054] For example, the failure diagnosis and prognosis system and method in the present application is applicable to air brake systems of heavy road vehicles, and also applicable to other systems of the vehicle. The failure diagnosis and prognosis first collects and pre-processes the sensor data. In the next step of data augmentation, the filled time series and failure condition labels are preferably used. Figure 1 is a flowchart of the Transformer model in the present application. Figure 1 The Transformer model shown is a deep neural network with one or more attention mechanism layers used as the enhancement model in the present application. The multi-head attention 102 is further illustrated in the Transformer model. Figure 2 A flowchart of the multi-head attention 102 is shown. As Figure 2 shown, various inputs of the time series data are projected into multiple subspaces to perform scaled dot-product attention 106 in each subspace. Thus, the outputs from the scaled dot-product attention 106 from different subspaces form the output of the multi-head attention 102 by concatenation.

[0055] Figure 3 A flowchart of the Transformer model is shown, which performs failure diagnosis and prognosis using attention mechanism layers with multi-head attention 102. Part of the attention mechanism layers are masked. As Figure 3 shown, the Transformer model produces multiple outputs for simultaneous failure diagnosis and prognosis. The Transformer model can effectively focus on useful information in the time series, thereby improving the performance and accuracy of the modeling results. The enhanced time series, as well as the failure diagnosis and prognosis results of the air brake system, will then be displayed to the user or operator through a web-based interface, such as a dashboard, and simultaneously stored on a cloud server.

[0056] The above explanation of the present application is not limited to the foregoing embodiments and drawings, and it is obvious to those skilled in the art of the present application that various substitutions, modifications and changes can be made without departing from the scope of the present application.

Claims

1. An attention-based neural network vehicle braking system fault diagnosis and prediction system, characterized in that, Comprising: a data augmentation module based on a deep learning model having at least one attention mechanism layer; wherein the attention mechanism layer comprises multi-head attention; and a fault diagnosis module and a fault prediction module based on at least one deep learning model having at least one attention mechanism layer.

2. The attention-based neural network vehicle braking system fault diagnosis and prediction system according to claim 1, wherein, The brake system comprises an air brake system.

3. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 2, wherein, The vehicle comprises a heavy road vehicle.

4. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 1, wherein, The attention-based neural network vehicle brake system fault diagnosis and prediction system further comprises a data collection module and a data pre-processing module.

5. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 1, wherein, The attention mechanism layer is configured to select useful information from the time series data.

6. The attention-based neural network vehicle braking system fault diagnosis and prediction system according to claim 1 or 5, characterized in that, The attention mechanism layer is trained to identify and prioritize the most useful information.

7. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 1, wherein, The multi-head attention mechanism layer is configured to project the input of the time series data into multiple subspaces and perform scaled dot-product attention in each subspace.

8. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 5, wherein, The attention mechanism layer comprises a masked multi-head attention configured to prevent unwanted dependencies from different locations of the time series.

9. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 7, wherein, The output of the multi-head attention is formed by the outputs of the scaled dot-product attentions from different subspaces.

10. The attention-based neural network vehicle braking system fault diagnosis and prediction system of claim 7, wherein, The multi-head attention is further configured to allow the deep learning model to jointly attend to information from different representation subspaces at different locations of the time series.

11. An attention-based neural network vehicle braking fault diagnosis and prediction method, characterized in that, Comprising: data augmentation based on a deep learning model having at least one attention mechanism layer; and fault diagnosis and prediction of a brake system based on at least one deep learning model having at least one attention mechanism layer.

12. The attention-based neural network vehicle braking fault diagnosis and prediction method according to claim 11, characterized in that, The augmented data is preceded by collecting and pre-processing data.

13. The attention-based neural network vehicle braking fault diagnosis and prediction method according to claim 11, characterized in that, The augmented data comprises: selecting useful information from the time series data; projecting the input of the time series data into multiple subspaces and performing scaled dot-product attention in each subspace; and forming the output of the multi-head attention by concatenating the outputs of the scaled dot-product attentions from different subspaces.

14. The attention-based neural network vehicle braking fault diagnosis and prediction method according to claim 13, characterized in that, The step of masking unwanted information is included before the forming of the output of the multi-head attention.

15. The attention-based neural network vehicle braking fault diagnosis prediction method according to claim 13, characterized in that, The forming of the multi-head attention output further allows the deep learning model to jointly attend to information from different representation subspaces at different locations of the time series.

16. The attention-based neural network vehicle braking fault diagnosis and prediction method according to claim 13, characterized in that, The data augmentation further comprises training the attention mechanism layer to identify and prioritize the most useful information.

17. The attention-based neural network vehicle braking fault diagnosis prediction method according to claim 13, characterized in that, Upon diagnosing and predicting faults of the brake system, the results are stored and displayed.

Images