System and method for analytic modeling based on onboard data

Abstract

The invention discloses a system and method for analytic modeling based on onboard data. The system comprises a data processing module, a feature engineering module and an automatic modeling module. The data processing module is used for receiving airborne data from the airborne member system and processing at least the airborne data to generate objectified data, and the objectified data comprises a preset data format; the feature engineering module is used for determining an optimal preset number of data features based on the objectified data by utilizing an automatic feature engineering method; and a modeling module to generate, using an automated model training framework, an output model based on the optimal predetermined number of data features, the output model deployed to analyze the onboard data.

Description

TECHNICAL FIELD

[0001] The present application relates to analysis modeling of onboard data. BACKGROUND

[0002] In the field of aviation related to flight safety, onboard data analysis is often applied to onboard equipment. By analyzing sensor data in onboard data in real time and identifying real-time state information of onboard equipment, more accurate and intelligent real-time monitoring warning information can be provided for flight safety. By analyzing historical data of onboard data, effective data means can be provided for onboard equipment in terms of residual life, fault diagnosis, and multi-dimensional evaluation of flight fleet, thereby providing effective operation and maintenance support for aircraft operators and reducing operating costs.

[0003] With the development of machine learning and deep learning technologies, the use of onboard data to assess the health of onboard systems and predictive maintenance of onboard systems is increasingly attracting attention in the industry. Currently, in the field of aviation, data anomalies or residual life are mainly predicted based on physical model characteristics or traditional machine learning methods. These applications require a large number of data experts to analyze features, perform feature engineering, select model training algorithms, and adjust parameters during the training process, which is highly dependent on the domain knowledge of data experts and model builders.

[0004] In addition, since onboard data comes from different onboard member devices, onboard data can be time series data, flight state real-time data, or cabin sound data, etc. Different onboard data requires different data experts for analysis, and data experts need to have domain knowledge of different onboard member devices. Data analysis based on traditional data-driven onboard data relies on data experts to analyze onboard data, select features, select modeling algorithms, etc. for each onboard member device.

[0005] Chinese patent application publication CN 114417501 A discloses a health management predictive modeling method for onboard deployment. The method processes onboard data and ground data and trains models to obtain a plurality of training models, evaluates and selects the training models to obtain a training model that meets task requirements, and deploys the training model that meets the task requirements in the onboard system. The method is a health management predictive modeling method for onboard deployment mainly based on onboard data, which uses traditional data processing, feature engineering, and data training modeling, but still needs to rely on experts' experience to select feature operators and training operators for model training.

[0006] Chinese patent application publication CN 114757307 A discloses an artificial intelligence automatic training method, system, device and storage device, which is designed for convolutional neural network, recurrent neural network, graph convolutional neural network and other network search frameworks. Basic operation sub is designed to adapt to power model structure. Feasible model search strategy is explored to realize automatic training of multi-task in power grid application scenario, reduce difficulty of artificial intelligence model development, solve problems of relying on personal experience and tedious time-consuming in feature selection, model evaluation and other aspects of traditional artificial intelligence algorithm, and improve development efficiency of artificial intelligence model. However, its model training still depends on uploading of sample data and manual labeling of sample data.

[0007] Chinese patent application publication CN 115640859 A discloses an enhanced learning method based on automatic machine learning, which includes a machine learning model automatic continuous integration framework and a four-stage incremental reinforcement learning MLops framework. AutoFE+AutoML automatic modeling and model integration based on incremental reinforcement learning are involved. However, this patent application does not mention specific automatic feature engineering and automatic model training methods.

[0008] There is a need in the art for more efficient and higher automated systems and methods for analyzing and modeling onboard data. SUMMARY

[0009] To provide more efficient and higher automated techniques for analyzing and modeling onboard data, the present invention is provided.

[0010] According to one aspect of the present invention, there is provided a system for analyzing and modeling based on onboard data, comprising: a data processing module configured to: receive onboard data from an onboard member system; and process at least the onboard data to generate objectified data, the objectified data comprising a predetermined data format; a feature engineering module configured to determine an optimal predetermined number of data features based on the objectified data using an automated feature engineering method; and a modeling module configured to produce an output model based on the optimal predetermined number of data features using an automated model training framework, the output model being deployed for analyzing onboard data.

[0011] The system as described above, wherein the on-board member system comprises on-board subsystems, and the data processing module comprises a data processing pipeline configured in a hierarchy, the data processing pipeline is configured to: detect the on-board data to determine information associated with the on-board subsystems; invoke a data format parser configured for the corresponding on-board subsystems to parse the on-board data associated with the on-board subsystems into on-board data metadata; parse the on-board data metadata into on-board data objects based on a predetermined data template format associated with the on-board subsystems, wherein the objectified data is generated based at least in part on the on-board data objects.

[0012] The system as described in any of the above, wherein the data processing pipeline is configured to detect a data type and a data format of the on-board data, or a data type, a data format, and a data label of the on-board data.

[0013] The system as described in any of the above, wherein the data processing pipeline is further configured to perform at least one of: remove invalid data in the on-board data objects with a data validation method; organize the on-board data objects in a hierarchy, the hierarchy is determined based at least in part on the on-board member system, the on-board subsystem, and the data type; normalize the on-board data objects.

[0014] The system as described in any of the above, wherein the feature engineering module is configured to generate candidate data features based on features of the on-board data corresponding to the objectified data with an automated feature engineering algorithm, wherein the features of the on-board data are determined based at least in part on the data type of the on-board data.

[0015] The system as described in any of the above, wherein the feature engineering module is configured to generate corresponding data features for time-series on-board data, discrete on-board data, or modalized on-board data.

[0016] The system as described in any of the above, wherein the feature engineering module is further configured to automatically select the optimal predetermined number of data features from the candidate data features with a feature selection algorithm.

[0017] The system as described in any of the above, wherein the feature selection algorithm comprises at least one of: a random forest algorithm, an evolutionary algorithm, a genetic algorithm.

[0018] The system of any of the above, wherein the automated model training framework comprises a predetermined number of model training algorithms, and the modeling module is configured to: train, using the predetermined number of model training algorithms, based on each of the optimal predetermined number of data features, respectively, to produce a corresponding trained model; score the produced trained models based on a pre-set target parameter; and determine the output model based on a result of the scoring.

[0019] The system of any of the above, wherein the modeling module is configured to validate the produced trained models using a validation data set.

[0020] The system of any of the above, wherein the target parameter is based on at least one of: precision, recall, f1 score, loss function, absolute error.

[0021] The system of any of the above, wherein the objectified data is stored to a data platform.

[0022] According to another aspect of the present disclosure, there is provided a method for analytical modeling based on onboard data, comprising: receiving onboard data from an onboard member system; processing at least the onboard data to generate objectified data, the objectified data comprising a predetermined data format; determining, using an automated feature engineering method, an optimal predetermined number of data features based on the objectified data; and producing, using an automated model training framework, an output model based on the optimal predetermined number of data features, the output model being deployed for analyzing onboard data.

[0023] The method of the above, wherein the onboard member system comprises an onboard subsystem, and processing at least the onboard data comprises: detecting the onboard data to determine information associated with the onboard subsystem; invoking a data format parser configured for the corresponding onboard subsystem to parse the onboard data associated with the onboard subsystem into onboard data metadata; parsing the onboard data metadata into an onboard data object based on a predetermined data template format associated with the onboard subsystem, wherein the objectified data is produced based at least in part on the onboard data object.

[0024] The method of any of the above, wherein detecting the onboard data comprises: detecting a data type and a data format of the onboard data, or detecting a data type, a data format, and a data label of the onboard data.

[0025] The method of any of the above, wherein processing at least the on-board data further comprises at least one of: removing invalid data in the on-board data objects using a data validation method; organizing the on-board data objects in a hierarchy, the hierarchy being determined based at least in part on on-board member systems, on-board subsystems, data types; normalizing the on-board data objects.

[0026] The method of any of the above, wherein determining the optimal predetermined number of data features comprises: generating candidate data features based on features of the on-board data corresponding to the objectified data using an automated feature engineering algorithm, wherein the features of the on-board data are determined based at least in part on data types of the on-board data.

[0027] The method of any of the above, wherein generating candidate data features comprises: generating corresponding data features for time-series on-board data, discrete on-board data, or modalized on-board data.

[0028] The method of any of the above, wherein determining the optimal predetermined number of data features further comprises: selecting the optimal predetermined number of data features from the candidate data features using a feature selection algorithm, wherein the feature selection algorithm comprises at least one of: a random forest algorithm, an evolutionary algorithm, a genetic algorithm.

[0029] The method of any of the above, wherein the automated model training framework comprises a predetermined number of model training algorithms, and generating the output model comprises: training each of the optimal predetermined number of data features based on the optimal predetermined number of data features respectively using the predetermined number of model training algorithms to generate corresponding training models; scoring the generated training models based on pre-set target parameters; and determining the output model based on results of the scoring.

[0030] The system and method for analytical modeling based on on-board data according to embodiments of the present application are more efficient and more automated. BRIEF DESCRIPTION OF DRAWINGS

[0031] Embodiments of the present application are described in conjunction with the appended drawings.

[0032] Figure 1 A schematic diagram showing a system for analytical modeling based on on-board data and its application environment according to some embodiments of the present application.

[0033] Figure 2 A block diagram showing a data processing module according to some embodiments of the present application.

[0034] Figure 3 A schematic diagram showing parsing of various types of on-board data into on-board data objects according to some embodiments of the present application.

[0035] Figure 4 A schematic diagram showing a hierarchically organized structured on-board data object according to some embodiments of the application.

[0036] Figure 5 A block diagram showing a feature engineering module according to some embodiments of the application.

[0037] Figure 6 A block diagram showing a modeling module according to some embodiments of the application.

[0038] Figure 7 A flowchart showing a method for analytical modeling based on on-board data according to some embodiments of the application.

[0039] Figure 8 A flowchart showing a process for processing on-board data according to some embodiments of the application.

[0040] Figure 9 A flowchart showing a process for determining an optimal predetermined number of data features according to some embodiments of the application.

[0041] Figure 10 A flowchart showing a process for generating an output model according to some embodiments of the application. DETAILED DESCRIPTION

[0042] According to an aspect of the application, there is provided a system for analytical modeling based on on-board data.

[0043] Figure 1 A schematic diagram showing a system 110 for analytical modeling based on on-board data and its application environment 100 according to some embodiments of the application.

[0044] In the application environment 100, the system 110 receives on-board data 120 and generates an output model for deployment 130. In some embodiments, the system 110 can be an integrated data platform located on the ground, for example, a server.

[0045] The system 110 can include a data processing module 111, a feature engineering module 113, a modeling module 115. Each of the data processing module 111, the feature engineering module 113, the modeling module 115 can be implemented by hardware, software, and / or firmware.

[0046] The data processing module 111 can be configured to receive the onboard data 120 from the onboard member system. As an example, the data processing module 111 can receive the onboard data 120 through a transmission mechanism that complies with the Universal Serial Bus (USB) protocol. As an example, the data processing module 111 can receive the onboard data 120 through a wired or wireless network.

[0047] The data processing module 111 can also be configured to process the onboard data to generate objectified data, which includes a predetermined data format. In some embodiments, the data processing module 111, in addition to being configured to process the onboard data, can also be configured to process other data, such as ground weather data, operations data provided by an airport operations platform, data provided by a component supplier for component analysis, etc. Further details of the data processing module 111 processing the onboard data 120 will be described below in connection with Figure 2 Further details.

[0048] The data processing module 111 according to embodiments of the present application automatically processes the onboard data 120 driven by the onboard data 120, while processing the onboard data 120 into objectified data to facilitate subsequent automatic feature engineering and modeling processes. This improves the efficiency and automation level of the entire analysis and modeling process of the onboard data 120, without the need for data experts relative to conventional methods.

[0049] The feature engineering module 113 can be configured to receive the objectified data generated by the data processing module 111, and determine an optimal predetermined number of data features based on the objectified data using an automated feature engineering method. As an example, an optimal K number of data features can be determined, where K is a self-defined constant. Further details of the feature engineering module 113 processing the onboard data 120 will be described below in connection with Figure 5 Further details.

[0050] The feature engineering module 113 according to embodiments of the present application automatically determines an optimal predetermined number of data features based on the objectified data from the data processing module 111, providing input for subsequent modeling.

[0051] The modeling module 115 can be configured to receive the optimal predetermined number of data features determined by the feature engineering module 113, and generate an output model based on the optimal predetermined number of data features using an automated model training framework. The generated output model can be deployed 130 for analyzing the onboard data.

[0052] The modeling module 115 according to embodiments of the present application utilizes an automated model training framework to produce an output model based on the optimal predetermined number of data features determined by the feature engineering module 113, without the need for a model builder to select a model training algorithm and tune parameters during the training process based on domain knowledge.

[0053] In general, the system 110 according to embodiments of the present application is directly driven by the on-board data 120 to automatically perform collection, processing, and feature engineering and modeling based on the processing results of the on-board data, improving the efficiency and automation of the analysis and modeling of the on-board data 120.

[0054] Figure 2 A block diagram of a data processing module 200 according to some embodiments of the present application is shown. The data processing module 200 can be used as the data processing module 111 of the system 110 in Figure 1 but the scope of the present application is not limited thereto.

[0055] The data processing module 200 can receive and process on-board data 220 to generate objectified data 230.

[0056] In some embodiments, the data processing module 200 can include a data processing pipeline 210 configured in a hierarchy. The data processing pipeline 210 can include a plurality of modules. In some embodiments, the plurality of modules can be executed serially. In some embodiments, the plurality of modules can be executed in parallel. In some embodiments, the plurality of modules can be executed partially serially and partially in parallel.

[0057] In some embodiments, the on-board data 220 is from an on-board member system, and the on-board member system can include on-board subsystems. In some embodiments, the data processing pipeline 210 can include a data source detection module 211. The data source detection module 211 is configured to detect the on-board data 220 to determine information associated with the on-board subsystems. Thereby, the on-board data 220 can be classified based on the on-board subsystems.

[0058] In some embodiments, the data source detection module 211 can be configured to detect data types and data formats. In some embodiments, the data source detection module 211 can be configured to detect data types, data formats, and data tags. As an example, data types can include discrete signal data, time series data, modal data, etc. As an example, discrete signal data can include signal data indicative of parameters or states of an onboard member system, time series data can include data indicative of maintenance records of an onboard member system, modal data can include audio, video data, etc. As an example, data formats can include A664 data packets, UDP data packets, text data, CSV data, cabin audio Wav, etc. Data tags can be derived from some tagging rules that require special handling of data when defining an onboard member system. As an example, data tags can include text data validity tags, audio validity timestamp tags, etc.

[0059] In some embodiments, the data processing pipeline 210 can include a format parsing module 212. In some embodiments, the format parsing module 212 can be configured with corresponding data format parsers for different member subsystems. The format parsing module 212 can be configured to invoke data format parsers configured for respective onboard subsystems to parse onboard data associated with the onboard subsystems into onboard data metadata. As an example, when the onboard data 220 includes text data, the format parsing module 212 can invoke a text parsing tool to read the text content and load the text content as binary byte metadata. As an example, when the onboard data 220 includes time series data, the format parsing module 212 can invoke a time series data parsing tool to read the time series data and load the time series data as binary byte metadata. As an example, when the onboard data 220 includes audio data, the format parsing module 212 can invoke an audio parsing tool to read the audio data and load the audio data as binary byte metadata.

[0060] By invoking format parsers configured for member subsystems based on the member subsystems through the format parsing module 212, automated processing of the onboard data 220 is facilitated. Meanwhile, the format parsing module 212 provides extensible custom configuration capabilities that can add, remove, and / or change corresponding data parsers based on member subsystems as needed.

[0061] In some embodiments, the data processing pipeline 210 can include an object resolution module 213. The object resolution module 213 can be configured to resolve the on-board data metadata into on-board data objects based on a predetermined data format associated with the on-board subsystems. The object resolution module 213 can customize the corresponding data object format for each on-board subsystem, and objectize the on-board data metadata from the format resolution module 212 into on-board data objects by resolving the mapping of the predetermined data format. The object resolution module 213 is mainly configured to convert the on-board data metadata associated with each member subsystem into on-board data objects in a predetermined format. As an example, the predetermined format can include data header information and data blocks (e.g., payloads). Details of the above process performed by the object resolution module 213 will be described below in conjunction with FIG. 3. Figure 3 The above process performed by the object resolution module 213 is further described.

[0062] In some embodiments, the objectized data 230 generated by the data processing module 200 can be generated directly based on the on-board data objects generated by the object resolution module 213. In some embodiments, the objectized data 230 generated by the data processing module 200 can be generated based on necessary post-processing of the on-board data objects.

[0063] In some embodiments, the data processing pipeline 210 can include a data cleansing module 214. The data cleansing module 214 can be configured to remove invalid data from the on-board data objects using data validation methods. As an example, the invalid data can be those data in the on-board data objects that do not conform to the data objects defined by the corresponding on-board subsystems. As an example, the data validation methods can include discrete quantity detection, data trend detection, data range detection, etc. By removing the invalid data through the data cleansing module 214, the quality of the data used for subsequent analysis modeling can be improved, and thus the quality of the resulting models can be improved. In some embodiments, the data cleansing module 214 can be omitted.

[0064] In some embodiments, the data processing pipeline 210 can include a data structuring module 215. The data structuring module 215 can receive the cleaned on-board data objects from the data cleansing module 214, or the on-board data objects directly from the object resolution module 213, and organize the received (cleaned or uncleaned) on-board data objects in a hierarchy. In some embodiments, the hierarchy can be determined based at least in part on the member systems, the on-board subsystems, the data types. In some embodiments, the data structuring module 215 can be omitted.

[0065] In some embodiments, the data structuring module 215 can organize the received (cleaned or uncleaned) on-board data objects into a tree structure. As an example, in the tree structure, the on-board member system can be a root node, child member systems can be generated as child nodes of the root node in a hierarchy of components, down to data types as further child nodes of the child nodes of the lowest level of child member systems to form a complete tree structure. An example of such a tree structure is described below in connection with Figure 4 FIG. 4.

[0066] In some embodiments, the data processing pipeline 210 can include a data normalization module 216. The data normalization module can receive data from the object parsing module 213, or the data cleaning module 214, or the data structuring module 215, and normalize the received data to generate objectized data 230. The data normalization module 216 is mainly used to map various on-board data to the range of [0, 1] to eliminate the dimensional influence between on-board data, so as to facilitate subsequent analysis modeling. In some embodiments, the data normalization module 216 can be omitted.

[0067] Figure 3 FIG. 1 shows a schematic diagram of parsing various types of on-board data into on-board data objects according to some embodiments of the present application. The schematic diagram can be used to describe the process performed by the object parsing module 213 in Figure 2 FIG. 2.

[0068] Figure 3 FIG. 3 shows discrete signal data 310, time series data 320, audio data 330, etc. These different types of data (e.g., by the object parsing module 213 in Figure 2 FIG. 4 shows a schematic diagram of structured on-board data objects 400 organized in a hierarchy according to some embodiments of the present application.

[0069] Figure 4 FIG. 4 shows a schematic diagram of structured on-board data objects 400 organized in a hierarchy according to some embodiments of the present application.

[0070] The structured on-board data objects 400 can include a tree structure, which can include a root node 410, first-level child nodes 420, and second-level child nodes 430. The root node 410 can include a member system A. The first-level child nodes 420 can include subsystems 1, 2, …, N belonging to the member system A in a hierarchy of components. The second-level child nodes 430 can include data types A, B, …, N belonging to the subsystem 1.Figure 4 Also shown in FIG. 4 are onboard data objects 440 associated with data type A in a tree structure, and onboard data objects 450 associated with data type N. One skilled in the art can also arrange various subsystems as sub-nodes in multiple levels of a tree structure as needed.

[0071] In some embodiments, the objectified data 230 can be stored in a data platform for selection by a subsequent feature engineering module (e.g., feature engineering module 113 in FIG. 1) for analysis modeling. As an example, the data platform can include an onboard data platform, a ground data platform, a cloud data platform, etc. In some embodiments, the objectified data 230 can also be directly provided to a feature engineering module (e.g., feature engineering module 113 in FIG. 1). In addition to providing data storage functionality, the data platform can also provide data query functionality, which can provide a unified data query interface for relevant modules to access the stored objectified data 230. Figure 1 Figure 1 In some embodiments, the objectified data 230 can be stored in a data platform for selection by a subsequent feature engineering module (e.g., feature engineering module 113 in FIG. 1) for analysis modeling. As an example, the data platform can include an onboard data platform, a ground data platform, a cloud data platform, etc. In some embodiments, the objectified data 230 can also be directly provided to a feature engineering module (e.g., feature engineering module 113 in FIG. 1). In addition to providing data storage functionality, the data platform can also provide data query functionality, which can provide a unified data query interface for relevant modules to access the stored objectified data 230.

[0072] Figure 5 A block diagram of a feature engineering module 500 according to some embodiments of the present application is shown. The feature engineering module 500 can be used as the feature engineering module 113 of the system 110 in FIG. 1, but the scope of the present application is not limited thereto. Figure 1

[0073] The feature engineering module 500 can receive objectified data 530, and determine an optimal predetermined number of data features 540 based on the objectified data 530 using an automated feature engineering method. As an example, an optimal K data features can be determined, where K is a custom constant. As described above in connection with FIG. 3, in some embodiments, the objectified data 530 can be stored in a data platform, and the feature engineering module 500 can receive the objectified data 530 from the data platform. In some embodiments, the feature engineering module 500 can also receive the generated objectified data 530 directly from a data processing module (e.g., data processing module 200 in FIG. 2). Figure 2 Figure 2

[0074] In some embodiments, the feature engineering module 500 can include a feature generation module 510. The feature generation module 510 can be used to generate candidate data features based on features of onboard data corresponding to the objectified data 530 using an automated feature engineering algorithm. The features of the onboard data can be determined based at least in part on data similarity of the onboard data. Feature engineering is a process of generating specific features for a certain type of data, with certain matching rules.

[0075] ​​​​In some embodiments, the feature generation module 510 can include time series data feature engineering 512: generating time series data features for time series on-board data. As an example, the time series data features can include at least one of statistical quantities associated with the time series, e.g., maximum, minimum, median, mean, variance, kurtosis, etc.

[0076] In some embodiments, the feature generation module 510 can include discrete data feature engineering 514: generating discrete data features for discrete on-board data. As an example, the discrete data features can include anomaly detection features based on the discrete data, e.g., one-hot encoding, label encoding, etc.

[0077] In some embodiments, the feature generation module 510 can include modalized data feature engineering 516: generating modalized data features for modalized on-board data. As an example, the modalized data features can include Mel-frequency Cepstral Coefficient (MFCC) features associated with audio data, Natural Language Processing (NPL) vector features associated with text data.

[0078] In some embodiments, the feature engineering module 500 can further include a feature selection module 520. The feature selection module 520 can be configured to select an optimal predetermined number of data features 540 from the candidate data features generated by the feature generation module 510 using a feature selection algorithm. The feature selection module 520 can select a suitable feature selection algorithm based on the data type of the on-board data, score the candidate data features, and determine the optimal predetermined number of data features 540 based on the score values.

[0079] In some embodiments, the feature selection module 520 can include a random forest algorithm 522. The random forest algorithm 522 computes the importance of each feature and ranks the score values in descending order to select a predetermined number of candidate data features with the lowest error rate as the optimal predetermined number of data features 540. In some embodiments, the feature selection module 520 can include an evolutionary algorithm 524. In some embodiments, the feature selection module 520 can include a genetic algorithm 526. In other embodiments, various other algorithms can also be used.

[0080] Figure 6 A block diagram of a modeling module 600 according to some embodiments of the present application is shown. The modeling module 600 can be used as the modeling module 115 of the system 110 in Figure 1 but the scope of the present application is not limited thereto.

[0081] In some embodiments, the modeling module 600 can include an automated model training framework 610. The modeling module 600 can be configured to utilize the automated model training framework 610 to generate an output model 660 based on the optimal predetermined number of data features 620. As an example, the automated model training framework 610 can include an AutoAI framework.

[0082] In some embodiments, the automated model training framework 610 can include a predetermined number of model training algorithms: model training algorithm 1 611, model training algorithm 2 612, model training algorithm 3 613, model training algorithm 4 614, …, model training algorithm N 615. The modeling module 600 can be configured to utilize the model training algorithm 1 611, model training algorithm 2 612, model training algorithm 3 613, model training algorithm 4 614, …, model training algorithm N 615 to generate respective training models based on each optimal data feature of the optimal predetermined number of data features 620, respectively. For example, a respective model 1, model 2, model 3, model 4, …, model N of the candidate models 650 can be generated for each optimal data feature of the optimal predetermined number of data features 620, where the number N of models of the candidate models 650 can correspond to the number N of model training algorithms of the automated model training framework 610. Accordingly, assuming there are K optimal data features 620, K sets of the above N models can be generated, i.e., a total of K*N models.

[0083] In some embodiments, the modeling module 600 can also be configured to receive a target parameter 630, and score the generated training models (e.g., the candidate models 650) based on the target parameter 630. In some embodiments, the modeling module 600 can determine the output model 660 based on the results of the scoring. As an example, Figure 6 In some embodiments, the accuracy of each model of the candidate models 650 is also labeled in the box of each model, and the ranking of the scores based on the accuracy of the models is labeled as “1”, “2”, “3” in the output model 660. The results show that the top three models 1, 3, N are selected as the output model 660.

[0084] In some embodiments, the target parameter can be based on accuracy, which can have a value range of 0 to 1. For example, the target parameter can be based on Average_precision, which can have a value range of 0 to 1.

[0085] In some embodiments, the target parameter can be based on Precision.

[0086] In some embodiments, the target parameter can be based on f1 evaluation metrics. For example, the target parameter can be based on multi-class f1 evaluation metrics.

[0087] In some embodiments, the target parameter can be based on a loss function (Neg_log_loss).

[0088] In some embodiments, the target parameter can be based on an absolute error.

[0089] In some embodiments, the modeling module 600 can also be configured to receive a validation dataset 640, and use the validation dataset 640 to validate the generated training model (e.g., candidate model 650). In some embodiments, the validation dataset 640 can include gold standard data, which is artificially labeled test data in artificial intelligence.

[0090] The system for analytical modeling based on onboard data according to embodiments of the present application is more efficient and more automated, which can significantly reduce the resource input of data experts and artificial intelligence algorithm engineers, greatly improve the efficiency of onboard data modeling work, and reduce the cost of onboard data intelligent analysis.

[0091] The system for analytical modeling based on onboard data according to some embodiments of the present application reduces the data processing work that needs to be done for onboard data analysis by converting onboard data into objectified data to reduce the data analysis work that needs to be done by data experts.

[0092] The system for analytical modeling based on onboard data according to some embodiments of the present application utilizes the automated feature engineering method (e.g., feature generation algorithm and feature selection algorithm) to automatically generate optimal data features from the characteristics of onboard data for subsequent analytical modeling, reducing the workload of data experts in researching the characteristics of onboard data.

[0093] The system for analytical modeling based on onboard data according to some embodiments of the present application utilizes the model training algorithm contained in the automated model training framework to automatically select the optimal model for output by evaluating the accuracy of the models generated by different model training algorithms, and deploy it to the target environment, solving the difficulty of selecting model algorithms and operator libraries in the model training process.

[0094] The system for analytical modeling based on onboard data according to some embodiments of the present application can be driven by massive onboard data, and build data analysis models to be more effectively applied to real-time state analysis of onboard equipment, predictive analysis of onboard equipment, data mining and visualization analysis of onboard equipment, etc., providing effective means for the intelligentization and digitization of aircraft.

[0095] According to an aspect of the present application, a method for analytical modeling based on onboard data is provided.

[0096] Figure 7A flow chart of a method 700 for analysis modeling based on onboard data according to some embodiments of the present application is shown. The method 700 can be performed by the system 110 in Figure 1 The scope of the present application is not limited in this regard, however.

[0097] The method 700 can include receiving onboard data from an onboard member system at block 710. As an example, the step at block 710 can be performed by the data processing module 111 of the system 110 in Figure 1 or the data processing module 200 in Figure 2 As an example, the onboard data can be received from the onboard member system through a transmission mechanism compliant with USB protocol. As an example, the onboard data can be received from the onboard member system through a wired or wireless network.

[0098] The method 700 can include automatically processing at least the onboard data to generate objectified data including a predetermined data format at block 730. As an example, the step at block 730 can be performed by the data processing module 111 of the system 110 in Figure 1 or the data processing module 200 in Figure 2 In some embodiments, in addition to processing the onboard data, other data can also be processed, such as ground weather data, operation data provided by an airport operation platform, data for component analysis provided by a component supplier, etc.

[0099] The method 700 can include determining an optimal predetermined number of data features based on the objectified data using an automated feature engineering method at block 750. As an example, the step at block 750 can be performed by the feature engineering module 113 of the system 110 in Figure 1 or the feature engineering module 500 in Figure 5 .

[0100] The method 700 can include generating an output model based on the optimal predetermined number of data features using an automated model training framework at block 770, the output model being deployed for analyzing the onboard data. As an example, the step at block 770 can be performed by the modeling module 115 of the system 110 in Figure 1 or the modeling module 600 in Figure 6 .

[0101] The method 700 according to embodiments of the present application is directly driven by the onboard data to automatically perform collection, processing of the onboard data, and feature engineering and modeling based on the processing result, improving the efficiency and automation level of analysis modeling of the onboard data.

[0102] Figure 8 A flow chart of a process 800 for processing onboard data according to some embodiments of the present application is shown. The process 800 can be performed by the system 110 inFigure 7 The specific implementation of the steps at block 730 of the method 700 in FIG. 7 is an example, and the scope of the application is not limited thereto. The process 800 can be performed by the data processing module 111 of the system 110 in FIG. 1, or the data processing module 200 (or its data processing pipeline 210) in FIG. 2, but the scope of the application is not limited thereto. Figure 1 The specific implementation of the steps at block 730 of the method 700 in FIG. 7 is an example, and the scope of the application is not limited thereto. The process 800 can be performed by the data processing module 111 of the system 110 in FIG. 1, or the data processing module 200 (or its data processing pipeline 210) in FIG. 2, but the scope of the application is not limited thereto. Figure 2 The specific implementation of the steps at block 730 of the method 700 in FIG. 7 is an example, and the scope of the application is not limited thereto. The process 800 can be performed by the data processing module 111 of the system 110 in FIG. 1, or the data processing module 200 (or its data processing pipeline 210) in FIG. 2, but the scope of the application is not limited thereto.

[0103] The process 800 can include, at block 810, detecting the on-board data to determine information associated with the on-board subsystem. Thereby, the on-board data can be classified based on the on-board subsystem. In some embodiments, the steps at block 810 can include detecting a data type and a data format of the on-board data, or detecting a data type, a data format, and a data tag of the on-board data.

[0104] As an example, the data type can include discrete signal data, time series data, modal data, etc. As an example, the discrete signal data can include signal data indicative of a parameter or a state of the on-board member system, the time series data can include data indicative of a maintenance record of the on-board member system, the modal data can include audio, video data, etc.

[0105] As an example, the data format can include A664 data packet, UDP data packet, text data, CSV data, cabin audio Wav, etc. The data tag can be derived from some tag rules that require special handling of the data when defining the on-board member system.

[0106] As an example, the data tag can include a text data validity tag, an audio validity timestamp tag, etc.

[0107] The process 800 can include, at block 820, automatically invoking a data format parser configured for the corresponding on-board subsystem to parse the on-board data associated with the on-board subsystem into on-board data metadata.

[0108] As an example, when the on-board data includes text data, a text parsing tool can be invoked to read the text content and load the text content as binary byte metadata. As an example, when the on-board data includes time series data, a time series data parsing tool can be invoked to read the time series data and load the time series data as binary byte metadata. As an example, when the on-board data includes audio data, an audio parsing tool can be invoked to read the audio data and load the audio data as binary byte metadata.

[0109] The process 800 can include, at block 830, parsing the on-board data metadata into an on-board data object based on a predetermined data format associated with the on-board subsystem. The objectized data generated by the process 800 is generated at least in part based on the on-board data object.

[0110] The corresponding data object format can be customized for each onboard subsystem, and the onboard data metadata can be objectized into onboard data objects by parsing and mapping the predetermined data format. The step at block 830 is mainly used to convert the onboard data metadata associated with each member subsystem into onboard data objects in a predetermined format. As an example, the predetermined format can include data header information and data blocks (e.g., payload), as shown in Figure 3 .

[0111] The process 800 can include, at block 840, removing invalid data in the onboard data objects using a data validation method. In some embodiments, the step at block 840 can be omitted.

[0112] The process 800 can include, at block 850, organizing the onboard data objects in a hierarchy, which is determined based at least in part on the onboard member systems, the onboard subsystems, and the data types.

[0113] The invalid data can be those data in the onboard data objects that do not conform to the data objects defined by the corresponding onboard subsystems. As an example, the data validation method can include discrete quantity detection, data trend detection, data range detection, etc.

[0114] By removing the invalid data through the step at block 850, the quality of the data used for subsequent analysis modeling can be improved, and thus the quality of the resulting model can be improved.

[0115] In some embodiments, the received onboard data objects can be organized into a tree structure. As an example, in the tree structure, the onboard member systems can be the root nodes, the subsystems can be generated as the child nodes of the root nodes in the hierarchy of components, and the data types can be further child nodes of the child nodes of the subsystems in the lowest hierarchy, to form a complete tree structure, as shown in Figure 4 . In some embodiments, the step at block 850 can be omitted.

[0116] The process 800 can include, at block 860, normalizing the onboard data objects.

[0117] The step at block 860 is mainly used to map various onboard data to the range of [0, 1] to eliminate the dimensional influence between the onboard data, so as to facilitate subsequent analysis modeling. In some embodiments, the step at block 860 can be omitted.

[0118] In some embodiments, one or more of the steps of the above-described blocks 810-860 of process 800 can be omitted. In some embodiments, the steps of the above-described blocks 810-860 of process 800 can be performed serially. In some embodiments, the steps of the above-described blocks 810-860 of process 800 can be performed in parallel. In some embodiments, the steps of the above-described blocks 810-860 of process 800 can be performed partially serially and partially in parallel.

[0119] Figure 9 A flowchart of a process 900 for determining an optimal predetermined number of data features according to some embodiments of the application is shown. Process 900 can be a specific implementation of the steps at block 750 of method 700 in Figure 7 the scope of the application is not limited in this regard. Process 900 can be performed by the feature engineering module 113 of system 110 in Figure 1 the scope of the application is not limited in this regard. Process 900 can be performed by the feature engineering module 113 of system 110 in Figure 5 the scope of the application is not limited in this regard. Process 900 can be performed by the feature engineering module 500 in

[0120] Process 900 can include, at block 910, generating candidate data features with an automated feature engineering algorithm based on features of on-board data corresponding to objectified data.

[0121] In some embodiments, the step at block 910 can include, at block 912, generating time series data features for time series on-board data. As an example, time series data features can include statistical quantities associated with time series, such as at least one of a maximum value, a minimum value, a median, a mean, a variance, a kurtosis, and the like.

[0122] In some embodiments, the step at block 910 can include, at block 914, generating discrete data features for discrete on-board data. As an example, discrete data features can include anomaly detection features based on discrete data, such as one-hot encoding, label encoding, and the like.

[0123] In some embodiments, the step at block 910 can include, at block 916, generating modalized data features for modalized on-board data. As an example, modalized data features can include MFCC features associated with audio data, NLP vector features associated with text data.

[0124] The process 900 can include selecting, at block 930, the optimal predetermined number of data features from the candidate data features using a feature selection algorithm. In some embodiments, the feature selection algorithm can include at least one of a random forest algorithm, an evolutionary algorithm, a genetic algorithm. In other embodiments, various other algorithms can also be used. In some embodiments, a suitable feature selection algorithm can be selected based on the data type of the onboard data, score the candidate data features, and determine the optimal predetermined number of data features based on the score values.

[0125] Figure 10 A flowchart of a process 1000 for generating an output model according to some embodiments of the present application is shown. The process 1000 can be a specific implementation of the steps at block 770 of the method 700 in Figure 7 Figure 1 Figure 6 The process 1000 can be performed by the modeling module 115 of the system 110 in

[0126] The process 1000 can include, at block 1010, training, using the predetermined number of model training algorithms, to generate respective training models based on each of the optimal predetermined number of data features respectively.

[0127] The process 1000 can include, at block 1030, scoring the generated training models based on a preset target parameter. In some embodiments, the target parameter can be based on at least one of precision, recall, f1 evaluation index, loss function, absolute error.

[0128] The process 1000 can include, at block 1050, determining the output model based on the results of the scoring.

[0129] The method for analytical modeling based on onboard data according to embodiments of the present application is more efficient and more automated, which can significantly reduce the resource input of data experts and artificial intelligence algorithm engineers, greatly improve the efficiency of onboard data modeling work, and reduce the cost of onboard data intelligent analysis.

[0130] The method for analytical modeling based on onboard data according to some embodiments of the present application reduces the data analysis work that needs to be done by data experts by converting onboard data into objectified data, reducing the data processing work that needs to be done for onboard data analysis.

[0131] ​​The method for analysis modeling based on onboard data according to some embodiments of the present application can utilize the characteristics of onboard data to generate optimal data features for subsequent analysis modeling by utilizing an automated feature engineering method (e.g., a feature generation algorithm and a feature selection algorithm), thereby reducing the workload of data experts in researching the characteristics of onboard data.

[0132] The method for analysis modeling based on onboard data according to some embodiments of the present application can automatically select an optimal model to output by evaluating the accuracy of models generated by different model training algorithms included in a model training framework, thereby solving the difficulty in selecting model algorithms and operator libraries in the model training process.

[0133] The method for analysis modeling based on onboard data according to some embodiments of the present application can be driven by massive onboard data, and can construct data analysis models to be more effectively applied to real-time state analysis of onboard equipment, predictive analysis of onboard equipment, data mining and visual analysis of onboard equipment, and the like, thereby providing an effective means for the intelligentization and digitization of aircraft.

[0134] Embodiments of the present application have been described in conjunction with the accompanying drawings. Each embodiment is illustrative and non-restrictive.

Claims

1. A system for analytical modeling based on onboard data, comprising: a data processing module configured to receive onboard data from an onboard member system and process at least the onboard data to generate objectified data, the objectified data comprising a predetermined data format; a feature engineering module configured to determine an optimal predetermined number of data features based on the objectified data using an automated feature engineering method; and a modeling module configured to produce an output model based on the optimal predetermined number of data features using an automated model training framework, the output model deployed for analyzing onboard data.

2. The system of claim 1, wherein, the onboard member system comprises onboard subsystems, and the data processing module comprises a data processing pipeline configured hierarchically, the data processing pipeline configured to: detect the onboard data to determine information associated with an onboard subsystem; invoke a data format parser configured for the corresponding onboard subsystem to parse the onboard data associated with the onboard subsystem into onboard data metadata; parse the onboard data metadata into an onboard data object based on a predetermined data object template format associated with the onboard subsystem, wherein the objectified data is generated based at least in part on the onboard data object.

3. The system of claim 2, wherein, the data processing pipeline is configured to: detect a data type and a data format of the onboard data, or detect a data type, a data format, and a data label of the onboard data.

4. The system of claim 2, wherein, the data processing pipeline is further configured to perform at least one of: remove invalid data in the onboard data object using a data validation method; organize the onboard data object hierarchically, the hierarchy determined based at least in part on the onboard member system, the onboard subsystem, the data type; normalize the onboard data object.

5. The system of claim 1, wherein, the feature engineering module is configured to generate candidate data features based on features of the onboard data corresponding to the objectified data using an automated feature engineering algorithm, wherein the features of the onboard data are determined based at least in part on the data type of the onboard data.

6. The system of claim 5, wherein, the feature engineering module is configured to generate corresponding data features for time series onboard data, discrete onboard data, or modalized onboard data.

7. The system of claim 5, wherein, the feature engineering module is further configured to select the optimal predetermined number of data features from the candidate data features using a feature selection algorithm.

8. The system of claim 7, wherein, the feature selection algorithm comprises at least one of: a random forest algorithm, an evolutionary algorithm, a genetic algorithm.

9. The system of claim 1, wherein, the automated model training framework comprises a predetermined number of model training algorithms, and the modeling module is configured to: train each of the optimal predetermined number of data features based on the optimal predetermined number of data features respectively using the predetermined number of model training algorithms to automatically produce a corresponding training model; score the produced training models based on a preset target parameter; and determine the output model based on a result of the scoring.

10. The system of claim 9, wherein, the modeling module is configured to validate the produced training models using a validation data set.

11. The system of claim 9, wherein, the target parameter is based on at least one of: precision, accuracy, f1 evaluation metric, loss function, absolute error.

12. The system of claim 1, wherein, The objectified data is stored to a data platform.

13. A method for analytical modeling based on onboard data, comprising: receiving onboard data from an onboard member system; automatically processing at least the onboard data to generate objectified data, the objectified data comprising a predetermined data format; determining an optimal predetermined number of data features based on the objectified data using an automated feature engineering method; and producing an output model based on the optimal predetermined number of data features using an automated model training framework, the output model being deployed for analyzing onboard data.

14. The method of claim 13, wherein, The onboard member system comprises onboard subsystems, and processing at least the onboard data comprises: detecting the onboard data to determine information associated with an onboard subsystem; automatically invoking a data format parser configured for the corresponding onboard subsystem to parse the onboard data associated with the onboard subsystem into onboard data metadata; parsing the onboard data metadata into an onboard data object based on a predetermined data format associated with the onboard subsystem, wherein the objectified data is produced based at least in part on the onboard data object.

15. The method of claim 14, wherein, Detecting the onboard data comprises: detecting a data type and a data format of the onboard data, or detecting a data type, a data format, and a data label of the onboard data.

16. The method of claim 14, wherein, Processing at least the onboard data further comprises at least one of: removing invalid data in the onboard data object using a data validation method; organizing the onboard data object in a hierarchy, the hierarchy being determined based at least in part on the onboard member system, the onboard subsystem, and the data type; normalizing the onboard data object.

17. The method of claim 13, wherein, Determining the optimal predetermined number of data features comprises: generating candidate data features based on features of the onboard data corresponding to the objectified data using an automated feature engineering algorithm, wherein the features of the onboard data are determined based at least in part on a data type of the onboard data.

18. The method of claim 17, wherein, Generating the candidate data features comprises generating corresponding data features for time series onboard data, discrete onboard data, or modalized onboard data.

19. The method of claim 17, wherein, Determining the optimal predetermined number of data features further comprises: automatically selecting the optimal predetermined number of data features from the candidate data features using a feature selection algorithm, wherein the feature selection algorithm comprises at least one of a random forest algorithm, an evolutionary algorithm, and a genetic algorithm.

20. The method of claim 13, wherein, The automated model training framework comprises a predetermined number of model training algorithms, and producing the output model comprises: training each of the optimal data features in the optimal predetermined number of data features based on the predetermined number of model training algorithms, respectively, to produce a corresponding training model; scoring the produced training models based on a pre-set target parameter; and determining the output model based on a result of the scoring.

Images