Abnormality detection, fault diagnosis and forecasting system and method for brake system

Abstract

The invention discloses an anomaly detection, fault diagnosis and forecasting system and method for a braking system, and the system comprises a data collection module which is used for collecting sensor data from vehicle braking system parameters, and storing the collected sensor data together with related parameters; the data preprocessing module is used for segmenting by collecting a data collection state of a time sequence marked with a missing value by a tag sequence; the data enhancement module is used for acquiring an enhancement error; and the result distribution module is used for identifying an anomaly and a fault according to the enhanced error, performing diagnosis and forecasting through a deep learning model when the anomaly and the fault are found, and storing and displaying an evaluation result. The method has the advantages that the operation condition of the braking system is reflected more clearly, better explanation is provided for a user, better data quality and higher resolution are provided, good performance is provided for fault diagnosis and forecasting, additional sensors are not needed, and anomaly detection, fault diagnosis and forecasting can be evaluated accurately and effectively.

Description

Technical Field

[0001] This invention belongs to the field of vehicle braking technology and relates to a system and method for abnormal detection, fault diagnosis and prediction of braking systems. Background Technology

[0002] Air braking systems are commonly used in heavy vehicles such as buses and trucks. The system consists of several key components, including an air compressor, governor, air tank, brake pedal, brake chamber, and various valves and pipes. Current maintenance methods for transport vehicles, including those with air braking systems, include preventative and corrective maintenance. Preventative maintenance of air braking systems involves inspecting the function and components of the system, while corrective maintenance relies heavily on the driver's understanding of the system's condition.

[0003] While current methods can ensure the reliability of vehicle operation, they also have some limitations. First, the need to repair or replace parts that are still functioning properly leads to a waste of resources. This increases costs, including materials, human resources, and vehicle service time. Second, fixed maintenance cycles lack flexibility and cannot be appropriately adjusted according to the condition of the equipment. Finally, proactive maintenance is not performed before a system failure occurs. If additional corrective maintenance is required, it raises concerns about reliability and unexpectedly increases costs.

[0004] To address these challenges, a new approach called predictive maintenance is being adopted. This maintenance method considers the condition of the machine and recommends repair or replacement only when necessary. The combination of Physical Failure (PoF) analysis and Artificial Intelligence (AI) is a suitable approach for implementing predictive maintenance. PoF analysis studies the patterns of product and system failures based on their causes and mechanisms. Therefore, PoF provides insights for implementing preventative maintenance. Artificial Intelligence (AI), particularly Data-Driven Models (DDM), plays a crucial role in predictive maintenance. It relies on continuously collecting and analyzing data from onboard sensors to diagnose usage patterns or wear symptoms. DDM, combined with machine learning techniques, can determine equipment condition by analyzing data and automatically suggest appropriate maintenance. More specifically, DDM can perform the tasks required for predictive maintenance: anomaly detection, fault diagnosis, and prediction.

[0005] Furthermore, recent road vehicles have integrated various sensors connected to onboard electronic control units (ECUs). The Controller Area Network (CAN) bus is a vehicle bus standard designed to reduce the number of signal cables in a vehicle, allowing multiple ECUs to connect to the same set of signal cables without a host computer. The ECU packages sensor data into messages and communicates them via the CAN bus. Introducing any additional sensors and communication interfaces is not ideal; therefore, sensor data should be collected indirectly via the CAN bus.

[0006] However, indirectly collecting sensor data via the CAN bus can affect the quality of the received time-series data. The frequency at which the ECU transmits sensor data via the CAN bus may not match the frequency at which it receives data from the sensors. The data transmission method of the CAN bus further limits the availability of the received time-series data, as the CAN bus can only transmit one message at a time. Sensor data transmitted by one ECU via a lower-priority CAN bus message is unlikely to be received by other ECUs. Generally, anomaly detection, fault diagnosis, and prediction require data from multiple sensors; therefore, the received time-series data will become a multivariate, irregularly sampled time series with misaligned observation times, which is challenging for DDM (Discrete Data Mechanism).

[0007] While existing systems may employ typical data preprocessing steps, each has its limitations. For example, discretization with a sufficiently long time step can convert a multivariate irregularly sampled time series into a multivariate regularly sampled time series, but this reduces the resolution over time. Alternatively, a shorter time step can be used to construct a multivariate regularly sampled time series, during which some values ​​are missing, requiring methods such as interpolation, mean estimation, or zero estimation to fill in the missing values. Filled values ​​may dominate the data representation but fail to provide information about the actual situation. All of these factors can impact the performance of predictive maintenance tasks. This is particularly acute in applications such as predictive maintenance of air braking systems, where high time resolution is required, but the frequency of received time series data is much lower.

[0008] Some examples of these are discussed in the preceding technologies below.

[0009] Chinese patent document CN113485302A discloses a method and system for fault diagnosis of vehicle operation process based on multivariate time series data. The method includes acquiring operating status information during vehicle operation; dividing each operating status information into time series segments to obtain multivariate time series data; extracting correlation characteristics from the multivariate time series data; extracting time dependency characteristics from the dependency characteristics of the multivariate time series data; and inputting the multivariate time series data and time dependency characteristics into a trained fault detection and diagnosis model to obtain fault detection and diagnosis results. This achieves fault detection and diagnosis of the vehicle operation process. However, the preprocessing methods disclosed in the prior art cannot handle multivariate irregularly sampled time series data, in which case the observations of each attribute are asynchronous or do not have equal frequencies.

[0010] Chinese patent CN112784965B discloses a method for anomaly detection in large-scale multi-factor time series data in a cloud environment. This method includes the following steps: establishing a multivariate time series anomaly detection model through offline training, and then using the offline-trained anomaly detection model to detect anomalies in online monitored data. This method improves the frontal feedback network of the native variational autoencoder, constructing the dependencies of the multivariate time series during the offline model training stage; it also improves the calculation method of the loss function, allowing the model to focus on data with normal patterns and ignore data with abnormal patterns during training. This results in a low model reconstruction probability and easier anomaly detection when performing anomaly detection online. However, existing anomaly detection methods are based on predefined thresholds and reconstruction errors. Especially in complex or dynamic environments, this method may not always provide robust and accurate results. Furthermore, existing methods learn the dependencies of multivariate time series through Long Short-Term Memory (LSTM) networks. When the proportion of missing values ​​in the multivariate time series is high, the dependencies may become biased.

[0011] Chinese patent document CN108958217A discloses a method for detecting anomalies in CAN bus messages, specifically a deep learning-based method for CAN bus message anomaly detection. This prior art method is mainly applied to automotive CAN network communication systems, including model analysis and anomaly detection. Model analysis involves analyzing the communication process points of the CAN bus network model and using this model as the basis for a deep learning network to perform anomaly detection based on message processing in the CAN bus. However, a large amount of data is required for effective training to utilize the deep learning network. Insufficient data can lead to overfitting and poor generalization.

[0012] Chinese patent document CN107703920A discloses a fault detection method for train braking systems based on multivariate time series. Existing fault detection methods involve collecting relevant sample data for train braking system fault detection and establishing a multivariate time series matrix of the sample data. A sliding time window is used to extract sample data of time series segments from the multivariate time series matrix, and feature matching detection is performed on abnormal patterns. The fault detection result of the train braking system in the time series segment is obtained based on the matching detection result. Existing methods, from a data analysis perspective, combine machine learning and multivariate time series mining algorithms to propose an algorithm for abnormal pattern matching based on a sliding time window. Through pattern matching, faults existing in the data can be monitored and intelligently diagnosed, thereby more accurately pointing out the root cause of the anomaly, and preferably locating the anomaly. However, in existing technologies, the operational data related to fault detection is acquired through sensors, and missing values ​​are filled into the operational data. Filling missing values ​​in the operational data may introduce bias into the dataset, as the filled values ​​may not accurately reflect the true values ​​that should have been observed when the data was complete. This may lead to biased estimations and inaccurate results in downstream analysis. Furthermore, the filled values ​​may fail to capture the true changes and patterns in the data, which could affect the reliability and validity of any conclusions drawn from the data. Summary of the Invention

[0013] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a system and method for abnormal detection, fault diagnosis and prediction of braking systems, which can solve various problems existing in the prior art.

[0014] To achieve the above objectives, the technical solution adopted by the present invention is: a system for anomaly detection, fault diagnosis, and prediction of a braking system, characterized in that it comprises:

[0015] The data collection module is used to collect sensor data from the vehicle braking system parameters and store the collected sensor data together with the relevant parameters.

[0016] The data preprocessing module is used to segment time series data by collecting data with missing values ​​from the labeled sequence.

[0017] The data augmentation module is used to acquire augmentation errors;

[0018] Evaluation module;

[0019] The evaluation module includes an anomaly detection module; a fault diagnosis module; and a fault prediction module; as well as...

[0020] The result allocation module is used to identify anomalies and faults based on the enhanced error, perform diagnosis and prediction using a deep learning model when anomalies and faults are detected, and store and display the evaluation results.

[0021] Furthermore, the braking system includes an air braking system.

[0022] Furthermore, the vehicles include heavy-duty highway vehicles.

[0023] Furthermore, the vehicle braking system parameters include: air compressor status, air tank pressure, brake chamber pressure, brake switch status, brake pedal position, vehicle speed, axle speed, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal position, engine speed, and engine torque.

[0024] Furthermore, the data preprocessing module is configured to convert multivariate irregular sampling time series into multivariate regular sampling time series with missing values.

[0025] Furthermore, the data augmentation module includes a deep learning model and an augmentation training model, the latter being configured to train the deep learning model to estimate missing values ​​in the segmented time series.

[0026] Furthermore, the data augmentation module is configured to generate an enhanced time series output.

[0027] Furthermore, the data augmentation module collects the label sequence in segments and fills the missing values ​​in the segmented time series with the values ​​in the augmented time series to obtain the filled time series.

[0028] Furthermore, the result allocation module displays the evaluation results through a web-based interface when an anomaly or fault is detected.

[0029] Furthermore, the evaluation results are stored in a cloud server.

[0030] Another object of the present invention is to provide a method for abnormal detection, fault diagnosis and prediction of a braking system, comprising:

[0031] Collect sensor data;

[0032] Preprocess sensor data;

[0033] Enhance the preprocessed data;

[0034] The results are evaluated through anomaly detection, diagnosis, and fault prediction; and

[0035] Store and display the evaluation results.

[0036] Furthermore, collecting sensor data includes connecting a data logger to the controller area network (CAN) bus to collect sensor data and storing CAN bus information with relevant parameters.

[0037] Furthermore, the preprocessed sensor data includes:

[0038] The sensor data is decoded into multiple univariate time series corresponding to the braking system parameters;

[0039] Remove inapplicable values ​​and perform normalization;

[0040] Align multiple univariate time series;

[0041] Remove redundant values;

[0042] Multiple univariate time series are concatenated into a multivariate conventional sampling time series;

[0043] Labeling the data collection status of multiple univariate time series; and

[0044] The multivariate regular sampling time series and collection label sequence with missing values ​​are segmented into segmented time series and segmented collection label sequences.

[0045] Furthermore, the method includes: training an augmentation model and an inference augmentation model.

[0046] Furthermore, the training enhancement model includes:

[0047] Segmented time series are randomly discarded to prepare discarded time series, which are used to artificially introduce missing values ​​into the discarded time series;

[0048] Prepare a discard collection tag sequence to reflect the data collection status of values ​​in the discard time series;

[0049] Training error was evaluated by comparing augmented time series and segmented time series; and

[0050] The training error is weighed based on the data collection state in the discarded collection label sequence to improve the missing value prediction of the augmented model.

[0051] Furthermore, the reasoning enhancement model includes:

[0052] The segmented time series and the segmented collected label sequences are fed into the trained augmented model;

[0053] Output enhanced time series;

[0054] A filled time series is obtained by replacing missing values ​​in a segmented time series with corresponding values ​​from the augmented time series; and

[0055] Based on the data collection status in the segmented collection label sequence, the augmentation error is obtained by comparing the collected values ​​with the augmented time series and the segmented time series.

[0056] Furthermore, the reasoning enhancement model includes:

[0057] The segmented time series and the segmented collected label sequences are fed into the trained augmented model;

[0058] Output enhanced time series;

[0059] A filled time series is obtained by replacing missing values ​​in a segmented time series with corresponding values ​​from the augmented time series; and

[0060] Based on the data collection status in the segmented collection label sequence, the augmentation error is obtained by comparing the collected values ​​with the augmented time series and the segmented time series.

[0061] Furthermore, the anomaly detection includes: identifying anomalies and faults when the enhancement error exceeds a predefined threshold; wherein the enhancement error is a vector value representing the error as a function of different parameters in the braking system.

[0062] Furthermore, the diagnosis and prediction of the deep learning model include:

[0063] The filled time series data are fed into the trained diagnostic model and the trained prediction model respectively to obtain the results;

[0064] When anomalies and faults are identified, the corresponding time series data will be included in the evaluation results.

[0065] Compared with the prior art, the present invention has the following advantages:

[0066] The implementation is simple and includes: a data collection module for collecting sensor data from vehicle braking system parameters and storing the collected sensor data along with relevant parameters; a data preprocessing module for segmenting the data collection status by collecting time series data with missing values ​​marked by label sequences; a data augmentation module for obtaining augmentation errors; and a result allocation module for identifying anomalies and faults based on the augmentation errors, performing diagnosis and prediction using a deep learning model when anomalies and faults are detected, and storing and displaying the evaluation results. It more clearly reflects the operating status of the braking system, provides better interpretation for users, offers better data quality and higher resolution, provides excellent performance for fault diagnosis and prediction, requires no additional sensors, and requires only minimal communication cables and interfaces, enabling accurate and effective evaluation of anomaly detection, fault diagnosis, and prediction. Attached Figure Description

[0067] The features of the invention will be more readily understood and recognized when reading the following detailed description in conjunction with the accompanying drawings of preferred embodiments of the invention, wherein:

[0068] Figure 1 The system structure of the present invention is shown;

[0069] Figure 2 a-2b shows preprocessed multivariate time series with missing values ​​filled by linear interpolation and with missing values ​​filled by augmentation models;

[0070] Figure 3 The workflow for preparing a model in this invention for enhancement, anomaly detection, fault diagnosis, and prediction is illustrated.

[0071] Enhanced error Figure 4 The workflow for preparing evaluation results in this invention is illustrated;

[0072] Figure 5 The training procedure for the enhanced model in this invention is shown;

[0073] Figure 6 The inference procedure of the enhanced model in this invention is shown;

[0074] Figure 7 The inference procedure of the incremental model in this invention, which can estimate the enhancement error of filling missing values, is shown. Detailed Implementation

[0075] Specific embodiments of the present invention are disclosed herein as requested. However, it should be understood that the disclosed embodiments are merely examples of the invention, which may be implemented in many different forms. Therefore, the specific structural and functional details disclosed herein should not be construed as limiting, but rather serve as the basis for the claims. It should be understood that the drawings and their detailed description are not intended to limit the invention to the specific forms disclosed herein; rather, the invention covers all modifications, equivalents, and alternatives falling within the scope defined by the claims. As used throughout this application, the word “may” indicates optional (i.e., possible) rather than mandatory (i.e., required). Similarly, the words “comprising” and “including” mean including but not limited to. Furthermore, unless otherwise mentioned, the word “a” means “at least one” and the word “a plurality” means one or more. When using abbreviations or technical terms, these refer to their generally accepted meanings known in the art.

[0076] like Figure 1 As shown, the present invention discloses an anomaly detection, fault diagnosis and prediction system for a braking system, comprising: a data collection module 101 with at least one sensor; a data preprocessing module 102; a data enhancement module 103; an evaluation module 104; wherein the evaluation module 104 includes an anomaly detection module 1041; a fault diagnosis module 1042; a fault prediction module 1043; and a result allocation module 105. Figure 1 The structure of the system is outlined.

[0077] According to a preferred embodiment of the invention, the braking system includes, but is not limited to, an air braking system, and the vehicle includes, but is not limited to, a heavy-duty road vehicle. The invention achieves maximum benefit in the application of air braking systems in heavy-duty road vehicles, but is also applicable to other systems in vehicles.

[0078] According to a preferred embodiment of the present invention, the data collection module 101 is configured to collect sensor data 210 from parameters, including but not limited to air compressor status, air tank pressure, brake chamber pressure, brake switch status, brake pedal position, vehicle speed, axle speed, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal position, engine speed, and engine torque. The invention further includes other parameters related to engine and vehicle dynamics. The sensor data 210 is packaged into information by the ECU in the vehicle and transmitted via the CAN bus. The data collection procedure requires a data logger connected to the CAN bus to collect the sensor data 210. The collected sensor data 210 is then stored along with relevant parameters. The data enhancement module 103 needs to extract large amounts of information. However, sensor data 210 for some parameters transmitted via the CAN bus may not be available at each relevant time step. Since the parameters of the air braking system are related to parameters of other systems during normal operation, the present invention utilizes relatively large amounts of data, including data from other systems, to enhance the parameters of the air braking system, thereby filling the sensor data 210 at each time step. The filled sensor data 210 can capture the operating status of the air braking system. This is crucial for human understanding and for the feature extraction process of the deep learning model in the fault diagnosis module 1042 and the fault prediction module 1043.

[0079] According to a preferred embodiment of the present invention, the data preprocessing module 102 is configured to convert a multivariate irregularly sampled time series into a multivariate regularly sampled time series with missing values. The data preprocessing module 102 is further configured to enhance the model 200 by collecting a label sequence to mark the data collection status of the time series with missing values, thereby filling in the missing values.

[0080] According to a preferred embodiment of the present invention, the data augmentation module 103 includes a deep learning model and an augmentation model training 1031, the latter being configured to train the deep learning model to estimate missing values ​​in the segmented time series 201. The data augmentation module 103 further includes an augmentation model inference 1032.

[0081] This invention enhances multivariate and univariate time series data used for fault diagnosis and prediction by extracting information from large datasets to improve data quality. Therefore, the data enhancement module 103 and the subsequent fault diagnosis and prediction modules 1042 and 1043 in this invention both employ deep neural networks.

[0082] According to a preferred embodiment of the present invention, the data augmentation module 103 is configured to generate the output of augmented time series 206. Based on the segmented collection label sequence 203, missing values ​​in the segmented time series 201 are filled with values ​​from the augmented time series 206, thereby obtaining the filled time series 208. To achieve better performance in subsequent tasks (such as fault diagnosis and prediction modules 1042, 1043), high-quality data must be provided, i.e., the time series of relevant parameters must have sufficiently high temporal resolution and be free of missing values. Therefore, augmentation model training 1031 is proposed to incentivize the augmentation model 200 to predict missing values. Augmentation model training 1031 follows a data preprocessing procedure. Augmentation model training 1031 trains the model to fill in missing values ​​in the time series data in the system (including, but not limited to, the system of the present invention).

[0083] Figure 5 An example of training an augmentation model is shown. In augmentation model training 1031, augmentation model 200 is trained in an unsupervised manner, where the output of augmentation model 200 is augmented time series 206. After the augmentation model training 1031 procedure, the trained augmentation model 200 is obtained.

[0084] The trained augmentation model 200 can also be used for Figure 6 The augmented model inference model 1032 shown is used to generate augmented time series 206, filled time series 208, and augmented error 209. (As shown...) Figure 7 As shown, the variance of the augmentation model inference procedure 1032 can be used to estimate the augmentation error 209 for filling missing values.

[0085] Figure 2 ab illustrates a multivariate time series preprocessing example where missing values ​​are filled in by linear interpolation and then filled in by augmentation model 200. Figure 2 The points shown in diagram 'a' represent the raw data points of two variables acquired from the CAN bus, and the lines represent the preprocessed data after linear interpolation, which is one of the typical preprocessing methods in existing inventions on the market. It can be seen that the resolution of both the raw data points and the preprocessed data is insufficient to reflect the fluctuations of the variables. And compared to... Figure 2 Compared to b, the enhancement module can effectively predict missing values, thus obtaining higher resolution for parameters related to the air braking system. Resolution is crucial for predictive maintenance. The aforementioned resolution is configured to capture the operation of the air braking system; therefore, the results are easier for users to understand than simple deep neural network outputs.

[0086] According to a preferred embodiment of the present invention, the result allocation module 105 is configured to display the evaluation results via a web-based interface when an anomaly or fault is detected. Furthermore, the evaluation results are stored in a cloud server, such as a dashboard. Additionally, the evaluation results for anomaly detection, fault diagnosis, and prediction are provided together with the filled time series 208 to facilitate user understanding of the evaluation results.

[0087] The present invention further teaches a method for detecting anomalies, diagnosing and predicting faults in a vehicle braking system, comprising the following steps: collecting sensor data 210; preprocessing the sensor data 210; enhancing the preprocessed data; evaluating the results by detecting anomalies, diagnosing and predicting faults; and storing and displaying the evaluation results. Figure 3 This invention illustrates the workflow of preparing models for enhancement, anomaly detection, fault diagnosis, and prediction. Figure 4 An example illustrates the workflow for preparing evaluation results in this invention.

[0088] According to a preferred embodiment of the present invention, the collection of sensor data 210 includes the following steps: connecting a data logger to a controller area network (CAN) bus to collect sensor data 210; and storing CAN bus information with relevant parameters.

[0089] According to a preferred embodiment of the present invention, the preprocessing of sensor data 210 includes the following steps: decoding sensor data 210 into multiple univariate time series corresponding to braking system parameters; removing inapplicable values ​​and performing normalization; aligning the multiple univariate time series; removing redundant values; concatenating the multiple univariate time series into a multivariate conventional sampling time series; marking the data collection status of the multiple univariate time series; and segmenting the multivariate conventional sampling time series and the collection mark sequence with missing values ​​into a segmented time series 201 and a segmented collection mark sequence 203.

[0090] The process of aligning multiple univariate time series is to create a multivariate irregularly sampled time series. Initially, all univariate time series are sampled asynchronously. Arranging multiple univariate time series according to predetermined time steps creates a multivariate regularly sampled time series.

[0091] In a multivariate periodically sampled time series, the value of a variable at a given time step corresponds to the associated univariate time series. If the univariate time series has a value within the corresponding duration of the time step, the data collection status for that value is collected. If no value is found, the data collection status for that value is missing, and the value can be filled with "not applicable" or a finite number. Furthermore, redundant values ​​(such as multiple values ​​within a predefined time step in a univariate time series) are removed. For example, the remaining values ​​can be determined by the first value, last value, average, maximum, or minimum value within that predefined time step. The time step should be less than 1 second to ensure sufficiently high system resolution.

[0092] The data collection status step for labeling multiple univariate time series includes creating a collection label sequence. For example, "1" and "0" in the collection label sequence correspond to collected and missing values ​​in a multivariate periodically sampled time series, respectively. Missing values ​​in multivariate periodically sampled time series can be replaced with some finite values ​​(such as "0") to facilitate further procedures.

[0093] According to a preferred embodiment of the present invention, the method includes the following steps: training the augmentation model 200; and inference augmentation model 200.

[0094] According to a preferred embodiment of the present invention, training the augmentation model 200 includes the following steps: randomly dropping (202) the segmented time series 201 to prepare a dropout time series 204, which is used to artificially introduce missing values ​​into the dropout time series 204; preparing a dropout collection label sequence 205 to reflect the data collection status of the values ​​in the dropout time series 204. The method further includes: inputting the dropout time series 204 and the dropout collection label sequence 205 into the augmentation model 200 to determine an augmentation time series 206; evaluating the training error 207 by comparing the augmentation time series 206 and the segmented time series 201; and weighing the training error 207 according to the data collection status in the dropout collection label sequence 205 to perform missing value prediction for the augmentation model 200. Figure 5 An example illustrates the training process of the augmented model.

[0095] The training error 207 can be the mean absolute error (MAE), mean squared error (MSE), or other error metrics. However, it is only evaluated based on the collected values ​​of the data collection status in the segmented collection label sequence 203.

[0096] Optionally, the training error 207 can be weighted according to parameters of the air braking system and other relevant systems. To improve the effectiveness of the augmentation model 200 in predicting missing values, the training error 207 can be weighted according to the data collection status in the discarded collection label sequence 205, thus giving higher weights to missing values ​​that were collected in the segmented time series 201 but artificially discarded in the discard time series 204. Furthermore, the augmentation model 200 can be optimized based on the training error 207 through a backpropagation process.

[0097] According to a preferred embodiment of the present invention, the inference of the augmentation model 200 includes the following steps: feeding the segmented time series 201 and the segmented collection label sequence 203 into the trained augmentation model 200; outputting the augmented time series 206; obtaining the filled time series 208 by replacing the missing values ​​in the segmented time series 201 with the corresponding values ​​in the augmented time series 206; and obtaining the augmentation error 209 by comparing the augmented time series 206 and the segmented time series 201 based on the data collection status in the segmented collection label sequence 203 and the collected values. Figure 6 An example is given of the enhanced model inference program 1032 in this invention.

[0098] According to a preferred embodiment of the present invention, the method further includes the following steps: randomly discarding 202 values ​​from the segmented time series 201 to prepare a discarded time series 204, thereby artificially introducing missing values ​​into the discarded time series 204; preparing a discarded collection label sequence 205 to reflect the data collection status of the values ​​in the discarded time series 204; inputting the discarded time series 204 and the discarded collection label sequence 205 into a trained augmentation model 200; outputting an augmented time series 206; replacing the missing values ​​in the segmented time series 201 with the corresponding values ​​in the augmented time series 206 to obtain a filled time series 208; and comparing the augmented time series 206 and the segmented time series 201 based on the collected values ​​according to the data collection status in the segmented collection label sequence 203 to obtain an augmentation error 209. Figure 7 A variant of the incremental model inference procedure 1032 is shown, which can estimate the augmentation error 209 for filling missing values.

[0099] According to a preferred embodiment of the present invention, anomaly detection includes the following steps: identifying an anomaly when the enhancement error 209 exceeds a predefined threshold; wherein, the enhancement error 209 is a vector value representing the error as associated with different parameters in the braking system.

[0100] Once any anomaly or fault is detected, the results will include the anomaly, fault, remaining useful life, and the corresponding fill time series 208. When no anomalies or faults are detected, the results will include a marker for the normal state.

[0101] According to a preferred embodiment of the present invention, the diagnosis and prediction of the deep learning model includes the following steps: feeding the filled time series 208 into the trained diagnostic model 211 and the trained prediction model 212 respectively to obtain results; wherein, when a fault is identified, the corresponding filled time series 208 is included in the evaluation results.

[0102] This invention offers several advantages and is simple to implement. It includes: a data collection module for collecting sensor data from vehicle braking system parameters and storing the collected sensor data along with relevant parameters; a data preprocessing module for segmenting the data collection status by collecting time series data with missing values ​​marked by tag sequences; a data augmentation module for obtaining augmentation errors; and a result allocation module for identifying anomalies and faults based on the augmentation errors, performing diagnosis and prediction using a deep learning model when anomalies and faults are detected, and storing and displaying the evaluation results. It more clearly reflects the operating status of the braking system, provides better interpretation for users, offers better data quality and higher resolution, provides excellent performance for fault diagnosis and prediction, requires no additional sensors, and requires only minimal communication cables and interfaces, enabling accurate and effective evaluation of anomaly detection, fault diagnosis, and prediction.

[0103] The above explanation of the present invention is not limited to the foregoing embodiments and drawings, and it will be apparent to those skilled in the art that various substitutions, modifications and alterations can be made without departing from the scope of the present invention.

Claims

1. A system for anomaly detection, fault diagnosis, and prediction in a braking system, characterized in that, include: The data collection module is used to collect sensor data from the vehicle braking system parameters and store the collected sensor data together with the relevant parameters. The data preprocessing module is used to segment time series data by collecting data with missing values ​​from the labeled sequence. The data augmentation module is used to acquire augmentation errors; Evaluation module; The evaluation module includes an anomaly detection module; a fault diagnosis module; and a fault prediction module; as well as... The result allocation module is used to identify anomalies and faults based on the enhanced error, perform diagnosis and prediction using a deep learning model when anomalies and faults are detected, and store and display the evaluation results.

2. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 1, characterized in that, The braking system includes an air braking system.

3. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 1, characterized in that, The vehicles mentioned include, but are not limited to, heavy-duty highway vehicles.

4. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 1, characterized in that, The vehicle braking system parameters include: air compressor status, air tank pressure, brake chamber pressure, brake switch status, brake pedal position, vehicle speed, axle speed, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal position, engine speed, and engine torque.

5. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 1, characterized in that, The data preprocessing module is configured to convert multivariate irregular sampling time series into multivariate regular sampling time series with missing values.

6. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 1, characterized in that, The data augmentation module includes a deep learning model and an augmentation training model, the latter being configured to train the deep learning model to estimate missing values ​​in the segmented time series.

7. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 6, characterized in that, The data augmentation module is configured to generate enhanced time series outputs.

8. The anomaly detection, fault diagnosis, and prediction system for a braking system according to claim 6 or 7, characterized in that, The data augmentation module collects segmented label sequences and fills the missing values ​​in the segmented time sequences with the values ​​in the augmented time sequences to obtain the filled time sequences.

9. The system for anomaly detection, fault diagnosis, and prediction of a braking system according to claim 1, characterized in that, When an anomaly or fault is detected, the result allocation module displays the evaluation results through a web-based interface.

10. The anomaly detection, fault diagnosis, and prediction system for a braking system according to claim 9, characterized in that, The evaluation results are stored on a cloud server.

11. A method for anomaly detection, fault diagnosis, and prediction of a braking system, characterized in that, include: Collect sensor data; Preprocess sensor data; Enhance the preprocessed data; The results are evaluated by detecting anomalies, diagnosing and predicting faults; as well as Store and display the evaluation results.

12. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 11, characterized in that, Collecting sensor data includes connecting a data logger to the controller area network (CAN) bus to collect sensor data and storing CAN bus information with relevant parameters.

13. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 11, characterized in that, Preprocessing of sensor data includes: The sensor data is decoded into multiple univariate time series corresponding to the braking system parameters; Remove inapplicable values ​​and perform normalization; Align multiple univariate time series; Remove redundant values; Multiple univariate time series are concatenated into a multivariate conventional sampling time series; Labeling the data collection status of multiple univariate time series; and The multivariate regular sampling time series and collection label sequence with missing values ​​are segmented into segmented time series and segmented collection label sequences.

14. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 11, characterized in that, The method includes: training an augmented model and inference an augmented model.

15. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 14, characterized in that, The training enhancement model includes: Segmented time series are randomly discarded to prepare discarded time series, which are used to artificially introduce missing values ​​into the discarded time series; Prepare a discard collection tag sequence to reflect the data collection status of values ​​in the discard time series; Training error was evaluated by comparing augmented time series and segmented time series; and The training error is weighed based on the data collection state in the discarded collection label sequence to improve the missing value prediction of the augmented model.

16. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 14, characterized in that, The reasoning enhancement model includes: The segmented time series and the segmented collected label sequences are fed into the trained augmented model; Output enhanced time series; A filled time series is obtained by replacing missing values ​​in a segmented time series with corresponding values ​​from the augmented time series; and Based on the data collection status in the segmented collection label sequence, the augmentation error is obtained by comparing the collected values ​​with the augmented time series and the segmented time series.

17. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 15 or 16, characterized in that, The reasoning enhancement model includes: The discard time series and the discard collection label sequence are fed into the trained augmented model; Output enhanced time series; A filled time series is obtained by replacing missing values ​​in a segmented time series with corresponding values ​​from the augmented time series; and Based on the data collection status in the segmented collection label sequence, the augmentation error is obtained by comparing the collected values ​​with the augmented time series and the segmented time series.

18. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 11, characterized in that, The anomaly detection includes: identifying anomalies and faults when the enhancement error exceeds a predefined threshold; wherein the enhancement error is a vector value representing the error as a function of different parameters in the braking system.

19. The method for abnormal detection, fault diagnosis, and prediction of a braking system according to claim 11, characterized in that, The diagnosis and prediction of the deep learning model include: The filled time series data are fed into the trained diagnostic model and the trained prediction model respectively to obtain the results; When anomalies and faults are identified, the corresponding time series data will be included in the evaluation results.

Images