METHOD AND SYSTEM FOR TRAINING A GENERATIVE MODEL FOR MODELING OPERATIONAL PROFILES

Abstract

This document describes a method (400) and a system (101) for training a generative model to model operational profiles. The method (400) comprises training an initial model based on a plurality of input data samples and processing an input training data sample from a plurality of input training data samples to generate a data series profile corresponding to the input training data sample. Furthermore, the method comprises processing the data series profile using one or more pairs of models to generate profiles of one or more parameters. Finally, the method comprises determining the initial loss value by comparing the generated profiles of one or more parameters with the corresponding profiles of one or more ideal parameters.Finally, the procedure involves training the generative model by repeating the above steps for the remaining input training data samples until the determined first loss value is less than a first threshold loss value, in order to train the second, third, and fourth models.

Description

PREAMBLE TO THE DESCRIPTION:

[0001] The following description explains the invention and the manner in which it is implemented in more detail. TECHNICAL AREA

[0002] The present subject matter generally relates to generative models in generative AI (GEN AI) and in particular, but not exclusively, to a method and a system for training a generative model for modeling operational profiles. BACKGROUND OF THE REVELATION

[0003] In general, a generative model (e.g., a generative machine learning model) is a large-scale artificial intelligence (AI) model (e.g., a machine learning model or a deep learning model) that is trained on a large or broad dataset, making it applicable to various and diverse use cases. Time series analysis (using time series models) offers unique features such as scalability and efficiency in the modeling approach, multivariate time series modeling, handling of missing data, transfer learning and domain adaptation, uncertainty quantification, integration of external data, development time, flexibility and adaptability, and consistency. Modeling data series (e.g., time series data) is a well-established practice.

[0004] US patent US20220180249A1, for example, discloses techniques for modeling vehicle operating profiles. This patent uses Generative Adversarial Networks (GANs) to generate scenarios. Specifically, the patent uses a standard GAN (for the model architecture) that is a combination of linear and ReLU layers. For validation, the patent does not include a specific methodology to ensure realistic generation. However, realistic generation is possible through the use of KL-GAN losses. The patent utilizes high-level driving scenarios that the GANs can generate and employs a single latent space. The conventional or traditional modeling approach for time-based models is inefficient because it uses separate models for analyzing different data sets or for performing specific tasks.Traditional time series models are computationally intensive, especially as the number of data series increases. This impairs scalability and efficiency.

[0005] Traditional time series models typically struggle to capture complex dependencies between data series. Therefore, traditional time series models cannot handle multivariate time series models effectively. They require specific imputation techniques (e.g., error correction, estimation, and filling in missing values). Traditional models perform imputation manually, thus limiting their efficiency. Furthermore, traditional time series models are limited to the specific task or domain for which they were trained. Transfer learning and domain adaptation are therefore difficult with traditional time series models. Finally, traditional time series models require manual integration and feature engineering for additional data sources, resulting in a lack of external data integration capabilities.Traditional time series models lack built-in mechanisms for estimating uncertainty. They are also time-consuming, as multiple models must be developed and maintained for different series, and each model must be adapted separately to changes. Consequently, traditional time series models offer limited flexibility and adaptability. Finally, due to the variability in model construction and training, traditional time series models exhibit inconsistent performance. Therefore, there is a need to address these and other related challenges.

[0006] The information disclosed in this section concerning the background of the disclosure is provided solely for a better understanding of the general background of the invention and is not to be understood as an acknowledgment or an indication that this information constitutes prior art already known to a person skilled in the art. SUMMARY OF DISCLOSURE

[0007] In light of the above, it is necessary to train a generative model (e.g., a generative machine learning model) to perform efficient time series analysis. The generative model can then be used for data series analysis. Generative models for data series overcome these challenges by utilizing advanced architectures, pretraining, and modern deep learning techniques.

[0008] This document describes a method for training a generative model to model operational profiles. The method includes training a first model based on a plurality of input data samples, each of which comprises a plurality of data series, each data series containing at least one operational profile. Furthermore, the method includes processing an input training data sample from a plurality of input training data samples using the trained first model to generate a data series profile corresponding to the input training data sample, where each of the plurality of input training data samples comprises data series.Furthermore, the procedure comprises processing the data series profile using one or more pairs of models to generate profiles of one or more parameters associated with the input training data sample, each pair comprising a second model and a third model. The procedure then comprises determining a first loss value using a fourth model by comparing the generated profiles of the one or more parameters with corresponding profiles of one or more ideal parameters. Finally, the procedure comprises training the generative model by repeating steps (b)-(d) for the remaining input training data samples from the set of input training data samples until the determined first loss value is less than a first threshold loss value, in order to train the second, third, and fourth models.

[0009] This paper presents a system for training a generative model to model operational profiles. The system comprises a memory and a processor. The memory is communicatively coupled to the processor, which is configured to train a first model based on a plurality of input data samples, each of which comprises a plurality of data series, each data series containing at least one operational profile. Furthermore, the processor is configured to use the trained first model to process an input training data sample from a plurality of input training data samples to generate a data series profile corresponding to the input training data sample, where each of the plurality of input training data samples comprises data series.Furthermore, the processor is configured to process the data series profile using one or more pairs of models to generate profiles of one or more parameters associated with the input training data sample, with each pair comprising a second and a third model. Subsequently, the processor is configured to determine an initial loss value using a fourth model by comparing the generated profiles of the one or more parameters with the corresponding profiles of the one or more ideal parameters.Finally, the processor is configured to train the generative model by repeating steps (b)-(d) for the remaining input training data samples of the multitude of input training data samples until the determined first loss value is less than a first threshold loss value to train the second, third, and fourth models.

[0010] The foregoing summary serves only for illustration and is in no way intended to be limiting. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features will become clear by reference to the drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The accompanying drawings, which form part of this disclosure, illustrate exemplary embodiments and, together with the description, explain the disclosed principles. In the figures, the leftmost digit(s) of a reference number indicates the figure in which the reference number first appears. The same numbers are used throughout the figures to identify identical features and components. Some embodiments of systems and / or methods according to the embodiments of the present subject matter are now described only by way of example and with reference to the accompanying figures, in which: Fig. Figure 1a shows an exemplary environment 100a with a system 101 for training a generative model for modeling operational profiles according to some embodiments of the present disclosure. Fig. 1b shows a block diagram 100b of a in Fig.1b system101 according to some embodiments of the present disclosure. Fig. Figure 1c shows an exemplary network architecture of a generative model for modeling operational profiles according to some embodiments of the present disclosure. Fig. Figure 2 shows an application case for training a generative model for analyzing data series related to pallidity profiles in accordance with some embodiments of the present disclosure. Fig. Figure 3a shows an application case for training a generative model to analyze data series related to battery aging in accordance with some embodiments of the present disclosure. Fig.Figure 3b illustrates an application case for training a generative model for analyzing data series related to battery systems according to some embodiments of the present disclosure. Fig. Figure 3c shows a diagram of the training losses of a single latent space model and a model with multiple latent spaces according to some embodiments of the present disclosure. Fig. Figure 3d shows a diagram displaying the velocity data of a single latent space model and a multi-latent space model training loss in accordance with some embodiments of the present disclosure. Fig. Figure 4 shows a flowchart illustrating a method 400 for training a generative model for modeling operational profiles in accordance with some embodiments of the present disclosure.

[0012] The person skilled in the art should know that all block diagrams contained herein represent conceptual views of systems embodying the principles of the present subject matter. Likewise, it will be understood that all flowcharts, process diagrams, state transition diagrams, pseudocodes, and the like represent various processes, essentially represented in a computer-readable medium and executable by a computer or processor, regardless of whether such a computer or processor is explicitly depicted. DETAILED DESCRIPTION

[0013] In this document, the word "exemplary" is used to mean "serving as an example, instance, or illustration." Each embodiment or implementation of the subject matter described herein as "exemplary" is not necessarily to be understood as preferred or advantageous over other embodiments. Although the disclosure is open to various modifications and alternative forms, one specific embodiment has been shown by way of example in the drawings and is described in detail below. However, it should be understood that the disclosure is not intended to be limited to the specific forms disclosed, but rather, on the contrary, the disclosure is intended to cover all modifications, equivalents, and alternatives that fall within the scope of the disclosure.

[0014] The expressions "includes," "comprises," "includes," or other variations thereof are intended to cover non-exclusive inclusion, so that a setup, device, or process that includes a list of components or steps may include not only those components or steps but may also include other components or steps not expressly listed or belonging to such a setup, device, or process. In other words, one or more elements in a system or device introduced by "includes ... a" do not, without further limitations, exclude the existence of other or additional elements in the system or process.The following detailed description of embodiments of the disclosure refers to the accompanying drawings, which form part of this disclosure and illustrate specific embodiments in which the disclosure can be put into practice. These embodiments are described in sufficient detail to enable the person skilled in the art to implement the disclosure in practice, and it is understood that other embodiments may be used and that modifications may be made without departing from the scope of this disclosure. The following description is therefore not to be understood in a restrictive sense.

[0015] In this disclosure, a vehicle's "operating profile" refers to time-varying information relating to one or more parameters associated with the vehicle. The operating profile may include, but is not limited to, a speed profile, a battery aging profile, an engine temperature profile, an engine speed profile, a steering angle profile, and the like. For example, the speed profiles may include, but are not limited to, an acceleration profile, a gear profile, and a brake torque profile. Similarly, the battery aging profile may include, without limitation, a temperature profile, a speed profile, and a current profile. In this disclosure, "operating profile modeling" refers to training a model to predict, analyze, or generate vehicle operating profiles based on input data series.

[0016] Fig.Figure 1a shows an environment 100a comprising a system 101 for training a generative model configured to perform a variety of tasks in accordance with some embodiments of the present disclosure. As in Fig. As shown in Figure 1a, System 101 can comprise an interface 103, a memory 105, and a processor 107. System 101 can, for example, be any system configured to implement the generative model (in Fig.1a not shown) trained, which is further configured to analyze data series. In another example, System 101 can be a server that receives the input data, performs processing operations, and then provides the output 113. In another example, System 101 can be any tool that receives input data 111 and provides the output 113. In one example, System 101 can be connected via a wired or wireless communication network (in Fig.(1a not shown) communicate with a data platform 109 and receive input data (e.g., a multitude of input data samples 111) to train the generative model for analyzing data series or modeling operational profiles. The data platform 109 can be any data storage / database that stores the input data or a multitude of input data samples 111. The multitude of input data samples 111 can comprise a data series. In an example, the multitude of input data samples 111 can, without restriction, include data in floating-point form, data in tensor value form, data in vector form, random data, and the like.

[0017] During operation, the system 101 can, in a first step, receive the multitude of input data samples 111 from the data platform 109. Subsequently, the system 101 can generate a first model (in Fig.1a not shown) train on the basis of the plurality of input data samples 111. The plurality of input data samples comprises a plurality of data series (also referred to as "series data"). Each data series of the plurality of data series comprises at least one operating profile. The first model may comprise a transformer-based model (e.g., a transformer-based encoder-decoder model). The transformer-based model, or the transformer-based encoder-decoder model, comprises an encoder block and a decoder block (in Fig.(1a not shown). In particular, after receiving the plurality of input data samples 111, the encoder block can process one input data sample from a plurality of input data samples to generate an encoded data sample. In one example, the encoder block can comprise a linear rectified linear unit (ReLu) with adjustable latent size and input size. For example, the latent size can be L=10 and the input size i=256. The present disclosure is not limited to this. The present disclosure can include various latent sizes and input sizes. In the next step, the decoder block can process the encoded data pattern and generate the data series profile corresponding to the input data pattern 111. In one example, the decoder block can comprise a linear rectified linear unit (ReLu) with adjustable latent size and input size. For example, the latent size can be L=10 and the input size i=256.The present disclosure is not limited to this. Subsequently, System 101 can determine a loss value (referred to as the "second loss value") by comparing the generated data profile with a corresponding ideal data profile. The ideal data profile can be retrieved from the database. In other words, the ideal data profile can be a historical data profile stored in the database. After comparing the generated data profile with the corresponding ideal profile, System 101 can determine the second loss value. In one example, the second loss value can be a summation of Kullback-Leibler (KL) divergence losses, a structural similarity index (SSIM) loss, least square errors (L. 2The process involves, but is not limited to, losses due to intermittent data loss and Dynamic Time Warping (DTW) loss. The process of processing the input data sample 111, processing the encoded data sample, and generating the data profile, comparing the generated data profile with the ideal profile, and determining the loss value is repeated for the remaining data sample from the multitude of input data samples until the determined second loss value is less than a threshold loss value (referred to as the "second threshold loss value") to train the first model of the generative model. After training the first model, in the second step, the system 101 uses the trained first model and can process an input training data sample 111 from a multitude of input training data samples to generate a data series profile that corresponds to the input training data sample. Each of the multitude of input training data samples comprises data series.In the third step, System 101 can process the data series profile using one or more pairs of models to generate profiles of one or more parameters associated with the input training data sample. Each pair of models comprises a second model and a third model. The second model could, for example, be a coding model and the third model a generator model.

[0018] Specifically, system 101 uses one or more model pairs to process the data series profile, generating the profiles of one or more parameters, and the second model (in Fig.(1a not shown) mid processes the data series profile and generates the coded data profile. In one example, the second model (encoder) can use the algorithms ReflectionPad1d(3), conv1d (I, Dimension), and instanceNormalld ReLu to process the data profile. The present disclosure is not limited thereto. In particular, System 101 uses the second model (encoder). The encoder can process the data series profile and generate a coded data profile for the data series profile, and generate the profiles of one or more parameters by processing the coded data profile and a noise data vector using the third model. The third model can be a generator (in Fig.1a not shown). In one example, the encoder can use the algorithms ReflectionPadld(3), conv1d (I. dimension) and instanceNormal1d ReLu, reshape Up sample Down sample, ResidualBlock1d to process the coded data profile and the noise data vector. The present disclosure is not limited to this. After processing the data series profile and generating profiles of one or more parameters, the system 101 can, in the fourth step, generate a fourth model (in Fig.(1a not shown) to use to determine a loss value (referred to as the "first loss value") by comparing the generated profiles of one or more parameters with corresponding profiles of one or more ideal parameters. The one or more parameters can be application-specific. For example, if the application relates to the driving profile, the one or more parameters can, without limitation, include an accelerator pedal profile, a gear profile, a brake torque profile, and the like. In another example, if the application relates to battery aging, the one or more parameters can, without limitation, include a temperature profile, a voltage profile, a current profile, and the like. The present disclosure is not limited to this. The one or more parameters can relate to different applications. For example,on understanding vehicle dynamics and software systems for different profiles, on simulation and validation of driver assistance systems (ADAS), on fuel consumption and emission behavior for different profiles, on traffic simulation and traffic management, on battery management of electric vehicles, on simulation of vehicle safety systems such as airbags and electronic stability control, and the like.

[0019] In one example, the profiles of one or more ideal parameters could be historical information from the above-mentioned use cases, stored in the database (in Fig.(1a not shown). The profiles of one or more ideal parameters are retrieved from the database and compared with the generated profiles of one or more parameters, and then the first loss value is determined. The fourth model can be a discriminator that compares the generated profiles of one or more parameters with the profiles of one or more ideal parameters and determines the first loss value. The discriminator can use the summation of various losses to determine the first loss value, such as the Kullback-Leibler (KL) divergence loss, a structural similarity index (SSIM) loss, or the least squared error (L) loss. 2) and the loss of Dynamic Time Warping (DTW). In the present disclosure, a multi-latent space model is used, resulting in lower loss and better maturity for generating results. The training loss diagrams of the single latent space model 331 and the training loss diagrams of the multi-latent space model 333 are shown in Fig. Figure 3c shows the speed of the single latent space model and the multi-latent space model. Fig. 3D representation. In one example, the simple latent space model can result in a total loss of 1.057 and the multiple latent space model in a total loss of 0.052.

[0020] After determining the loss value using the fourth model, System 101 can, in the fifth step, train the generative model by repeating steps one through four until the determined initial loss value is less than a threshold loss value. This allows the second, third, and fourth models to be trained. Specifically, the determined initial loss value is fed back to the second, third, and fourth models, and their hyperparameters are adjusted. Once the loss is reduced and meets the threshold, a complete pipeline is implemented, including the first, second, third, and fourth models, for various data series analysis use cases. In one example, the number of data samples can be similar to the number of training data samples.In one example, the multitude of data samples can be used to train the first model, and the input training data sample from the multitude of input training data samples can be used to train the second, third, and fourth models. In another example, the multitude of data samples can be used to train the first, second, third, and fourth models.

[0021] In one embodiment, the system 101 can receive a real-time input data sample during real-time inference and process the received real-time input data sample using the first model to generate a data series profile corresponding to the real-time input data sample. In the next step, the system 101 can process the data series profile corresponding to the real-time input data sample and analyze the generated profiles of one or more parameters associated with the real-time input data sample. Subsequently, the data series can be analyzed for various use cases.

[0022] Fig. 1b shows a block diagram 100b of a in Fig.The system 101 shown in Figure 1a, according to some embodiments of the present disclosure. In one embodiment, the system 101 may comprise an interface 103, a memory 105, and a central processing unit (also referred to as "CPUs" or "the one or more processors") 107. In some embodiments, the memory 105 stores instructions that are coupled to the one or more processors 107. The memory 105 stores instructions that can be executed by the one or more processors 107. The one or more processors 107 may include at least one data processor for executing program components for executing user- or system-generated requests. The one or more processors 107 may execute one or more functions of the system 101 for training the generative model.Memory 105 can store instructions that can be executed by the one or more processors 107 and which, when executed, can cause the one or more processors 107 to train the generative model. Interface 103 can be coupled to the one or more processors 107, through which a variety of input data samples 111 can be received. For example, the one or more processors 107 can communicate with the data platform 109 to receive the variety of input data samples 111, as shown in [reference]. Fig. 1a shown.

[0023] In one embodiment, the one or more processors 107 may comprise one or more modules or hardware units, such as a training unit 115, a processing unit 117, and a determination unit 119, but not only these. In some embodiments, the one or more modules or units may be software modules that can be stored in the memory 105. The one or more modules or hardware units may be configured to perform the various operations of this disclosure in order to train the generative model so that it can perform a variety of tasks.

[0024] Fig. Figure 1c shows an exemplary network architecture of a generative model for analyzing data series according to some embodiments of the present disclosure.

[0025] The network architecture (especially the generative model) comprises a transformer-based model (201, 203) which includes an encoder block or encoder 201 and a decoder block or decoder 203, which are combined to form an encoder-decoder model, or the transformer-based model. The transformer-based model (201, 203) is configured to produce an output (referred to as the "first output") based on input data samples. In an example, the first output can be passed to the decoder. The decoder or decoder block 203 can process the encoded data and generate the data sequence profile. The output of the encoder and decoder model is passed to at least one pair of models connected to the output of the transformer-based model (201).203) are stacked and configured to generate a further output (referred to as the “second output”) based on the first output. In one example, the second output can be an encoded data series profile (213). Each pair of the one or more pairs of models comprises an encoder or encoding model 205, configured to receive the first output, and a generator model 207, stacked at an output of the encoding model 205. In . Fig. Figure 1c shows only one model pair (coding / generation model) for better understanding. However, there can be several model pairs for different use cases (as in Fig. 3a). First, a large number of input data samples 111 are taken from the data platform 109 (in Fig.(1c not shown). The multitude of input data patterns can be data series relating to driving profiles. For example, the multitude of input data samples 111 can be received from the e-drive data platform of a vehicle. The encoder or encoder block 201 can be trained based on the received multitude of input data samples 111. The multitude of input data samples 111 includes random values, vector values, or tensor values ​​of driving profiles. First, the encoder or encoder block 201 can process the multitude of input data samples 111 and generate a corresponding coded data sample. In the next step, the coded data pattern is forwarded to the decoder or decoder block 203. The decoder or decoder block 203 can process the coded data pattern and generate a data series profile 211 that corresponds to the input data pattern.The generated data series profile 211 is forwarded to the encoder or the code node 205.

[0026] The encoder or coding model 205 can process the data series profile and generates a coded data profile for the data series profile. The coded data profile, along with a noise data vector, is then passed to the generator model 207. The generator model 207 can generate the profiles of one or more parameters by processing the coded data profile and the noise data vector. The profiles of one or more parameters include, for example, an accelerator pedal profile, a gear profile, and a brake torque profile.In the next step, the profiles of the one or more parameters, along with the corresponding profiles of the one or more ideal parameters, are passed to a discriminator model 209, which is stacked at the second output of the one or more model pairs and configured to generate a loss value for training the transformer-based model (201, 203) and the one or more model pairs based on the second output. Specifically, the discriminator model 209 can compare the generated profiles of the one or more parameters with corresponding profiles of one or more ideal parameters (e.g., a historical data profile with temporal data) and determine the loss value 210. The loss value 210 is backpropagated, and the hyperparameters are fine-tuned to train the generative model. The encoder-decoder models 201 and 203 are trained separately.The encoder or coding model 205, the generator model 207, and the discriminator model 209 can be trained separately. The loss value is a sum of one or more of the following values: pair loss, Kullback-Leibler (KL) divergence loss, structural similarity index (SSIM) loss, least squared error (L.) loss. 2 ) and loss of Dynamic Time Warping (DTW), but not limited to that.

[0027] Fig. Figure 2 shows an application case for training a generative model for analyzing data series related to driving profiles in accordance with some embodiments of the present disclosure.

[0028] As in Fig.As shown in Figure 2, the input data sample 111 can be fed to the encoder or encoder block 201. The input data sample 111 can be a random tensor vector of certain dimensions (e.g., (256,1)) with a variety of values ​​that can be adjusted. Suppose the input data sample 111 includes the values ​​0.53, 0.27, 0.22, 0.22, 0.08, 0.67, 0.3, 0.25, 0.95, 0.21, 0.36, 0.44, 1.0, .... The data sample 111 can be received by the E-Drive data platform 109. Subsequently, the input data sample 111 is fed to the encoder or encoder block 201. The encoder or encoder block 201 can process the input data pattern 111 and generate a corresponding coded data pattern (e.g., 0.75, 0.29, 0.9, 0.37, 0.74, 0.89, 0.62, 0.64, 0.35, 0.26, 0.69, (3.64)). Let's assume the coded data sample has the dimensions (64, 1). In the next step, the coded data pattern is passed to the decoder or decoder block 203.The decoder 203 can process the encoded data pattern and generate a data series profile 211 (e.g. 5.78, 9.37, 13.28, 17.28, 21.25, 25.13, 28.88, 32.49, 35.95, 39.26, 42.42, 45.42), which is also referred to as a velocity profile and can be determined from the curve in . Fig.Figure 2 is shown. Let us assume that the velocity profile has dimensions (256, 1). The generated velocity profile, or data series profile 211, is passed to the encoder 205. The encoder 205 can process the data series profile 211 and generates a coded data profile 213 (for example, a coded velocity profile with size (64, 1)). Subsequently, the coded data profile 213, along with a noise data vector 215, is passed to the generator model 207. The generator model 207 can generate the profiles of one or more parameters 217 by processing the coded data profile 213 and the noise data vector 215. The profiles of one or more parameters 217 include, for example, an accelerator pedal profile, a gear profile, and a brake torque profile.In the next step, the profiles of the one or more parameters, together with the corresponding profiles of the one or more ideal parameters 219, are passed to a discriminator model 209. The discriminator model 209 can compare the generated profiles of the one or more parameters 217 with the corresponding profiles of one or more ideal parameters 219 (e.g., a historical data profile with temporal data) and determine the loss value.

[0029] The profiles of one or more parameters 217 can, for example, contain information about the accelerator pedal profile (APP), a gear profile (GP), and a brake torque profile (BTP) and have the dimensions (256, 3). The information for the accelerator pedal profile can include, for example, 6.28, 4.37, 3.96, 4.14, 5.62, ..., 0.50, 0.38, 0.50, 0.69, 1.31. The information for the gear profile can include, for example, 5.0, 5.0, 5.0, 5.0, 5.0, ..., 2.0, 1.0, 1.0, 2.0, 2.0. The information for the brake torque profile can include 19.46, 6.54. 3.59, 1.70, 1.98...805.87, 681.60, 614.69, 440.84, 323.72. The corresponding profiles of one or more ideal parameters 219 include, for example, historical / ideal information about the accelerator pedal profile (APP), a gear profile, and a brake torque profile (BTP). The historical information of the accelerator pedal profile may include, for example, 5.79, 5.07, 3.82, 2.42, 1.93...0.12, 0.08, 0.13, 0.23, 0.55.The historical information of the transmission profile can, for example, include 4.0, 4.0, 4.0, 4.0, 4.0, ..., 1.0, 1.0, 1.0, 1.0, 2.0. The historical information of the brake torque profile can, for example, include 64.12, 34.81, 27.91, 16.21, 19.17, ..., 1445.64, 1387.54, 1226.59, 829.73, 502.75. The profiles of the one or more parameters 217 and the corresponding profiles of the one or more ideal parameters 219 are fed to the discriminator model 209. The discriminator model 209 can compare the profiles of one or more parameters 217 and the corresponding profiles of one or more ideal parameters 219, and the loss value 210 is calculated. The loss value 210 can have certain dimensions (e.g., 256.1) and can include, for example, the values ​​0.74, 0.94, 0.31, 0.93, 0.39, 0.42, 0.62, 0.05, 0.33, 0.42, and 0.19.After calculating the loss value 210, it can be backpropagated, and the hyperparameters are fine-tuned to train the generative model. The encoder 201.205 can be used for feature extraction and dimensionality reduction. The encoder can transform raw input data into a set of high-level features, reducing dimensionality and capturing essential information. The encoder can compress the input data samples into a smaller or more manageable latent space representation, retaining the core information while discarding redundancies. The input for the encoder can be raw data (e.g., time series data), and the output can be a latent space representation (e.g., a compressed feature vector). The decoder, or decoder block 203, can be used for reconstruction and data generation.Decoder 203 can take the latent representation generated by the encoder and reconstruct it back into the original input space. In generative models, the decoder can generate new data samples from the latent representation. The input to the decoder can be a latent space representation, and the output can be reconstructed data (e.g., velocity). Generator model 207 can be used for data synthesis and can combine noise with latent features. In generative models such as GANs, the generator can produce new data patterns that resemble the training data. The generator can receive a random noise data vector and generate realistic patterns. The generator's input can be a noise data vector, and the output can be a data pattern. Discriminator model 209 can be used for differentiation and adversarial training.The discriminator model 209 can distinguish between real and generated (fake) data patterns and provides feedback to the generator during training, subsequently helping it to produce more realistic patterns by penalizing unrealistic data patterns. Table 1 below contains detailed information about the variety of input data samples, coded data samples, data series profiles, coded data profiles, profiles of one or more parameters, and historical information about one or more parameters. Table 1 A variety of input data samples Encoded data patterns Profile of the data series Profiles of one or more parameters Historical information on one or more parameters Loss 0.53 0.75 5.78 APP GP BTP APP GP BTP 0.74 0.27 0.29 9.37 6.28 5.0 19.46 5.795.0 4.0 64.12 0.94 0.22 0.9 13.28 4.37 5.0 6.54 7 4.0 34.81 0.31 0.22 0.37 17.28 3.96 5.0 3.59 3.82 4.0 27.91 0.93 0.08 0.74 21.25 4.14 5.0 1.70 2.42 4.0 16.21 0.39 0.67 0.89 25.13 5.62 5.0 1.98 1.93 4.0 19.17 0.42 0.3 0.62 28.88 ... ... ... ... 0.62 0.25 0.64 32.49 0.50 2.0 805.87 0.12 1.0 1445.6 0.05 0.95 0.35 35.95 0.38 1.0 681.60 0.08 1.0 1387.5 0.33 0.21 0.26 39.26 0.50 1.0 614.69 0.13 1.0 1226.5 0.42 0.36 0.69 42.42 0.69 2.0 440.84 0.23 1.0 829.73 0.19 0.44 0.64 45.42 1.31 2.0 323.72 0.55 2.0 502.75

[0030] Fig. Figure 3a shows an application case for training a generative model for analyzing data series related to battery aging according to some embodiments of the present disclosure.

[0031] The generative model comprises an encoder block or encoder 301 and a decoder block or decoder 303, which are combined to form an encoder-decoder model. The output of the encoder / decoder model is passed to a model pair comprising an encoder model 305 and a generator model 307. The output of generator model 307 is passed to another encoder model 309 and another generator model 311. Fig. Figure 3 shows only two model pairs (encoder-generator). However, there can be a number of model pairs for different use cases. First, a variety of input data samples 111 are taken from the data platform 109 (in Fig.(3a not shown) received. The multitude of input data samples 111 can be a data series relating to battery aging profiles. For example, the multitude of input data samples 111 can be received from a vehicle's e-drive data platform. The encoder 301 can be trained based on the received multitude of input data samples 111. The multitude of input data samples 111 includes random values, vector values, or tensor values ​​of the battery aging profiles. First, the encoder or encoder block 301 can process the multitude of input data samples 111 and generate a corresponding coded data sample. In the next step, the coded data pattern is passed to the decoder or decoder block 303. The decoder or decoder block 303 can process the coded data pattern and generate a data series profile 315 (e.g., a battery aging profile) that corresponds to the input data pattern.The generated battery aging profile or data series profile 315 is passed to the encoder or coding model 305. The encoder or coding model 305 can process the data series profile and generates a coded data profile (e.g., a coded battery aging profile) for the data series profile. Subsequently, the coded data profile 319, together with a noise data vector 317, is passed to the generator model 307. The generator model 307 can generate the profiles of one or more parameters 321 by processing the coded data profile 319 and the noise data vector 317. The profiles of one or more parameters include, for example, a temperature profile, a voltage profile, and a current profile.

[0032] In the next step, the generated profiles of the one or more parameters 321 can be forwarded to the encoder 309. The encoding node 309 can process the received data and generate coded profiles 325 of the one or more parameters, forwarding them to the generator model 311. The generator model 311 can process the generated profiles of the one or more parameters together with the noise data vector 323 and generate profiles of the one or more parameters 327 that differ from those of 321. Finally, the discriminator model 313 can compare the generated profiles of the one or more parameters 327 with the corresponding one or more ideal parameters 329 (e.g., a historical data profile with temporal data) and determine the loss value 210. The loss value 210 is backpropagated, and the hyperparameters are fine-tuned to train the generative model.The encoder-decoder models 301 and 303 are trained separately. The encoding or coding model 305, the generator model 307, the encoding model 309, the generator model 311, and the discriminator model 313 are trained separately.

[0033] The input data sample 111 can be a random tensor vector with specific dimensions (e.g., (256,1)) and can have a variety of values ​​that are adjustable. For example, the input data sample can include the values ​​0.18, 0.88, 0.03, 0.03, 0.98, 0.96, 0.26, 0.27, 0.43, 0.41, 0.23. The input data sample 111 can be received by the E-Drive data platform 109. The input data sample 111 is then forwarded to the encoder 301. The encoder or encoder block 301 can process the input data pattern 111 and generate a corresponding coded data pattern (e.g., 0.22, 0.28, 0.03, 1.0, 0.25, 0.1, 0.89, 0.33, 0.85). Let's assume the coded data sample has the dimensions (256, 1). In the next step, the coded data pattern is forwarded to the decoder or decoder block 303. The decoder or decoder block 303 can process the coded data sample and generate a data series profile 315 (e.g.,97, 97, 96, 95, 92, 92, 90, 90, 90, 89, 88, 88, 86, 83, 81) generate a data series, also known as a battery aging profile. Suppose the battery aging profile has dimensions (256, 1). The generated battery aging profile, or data series profile, 315 is passed to the encoder or coding model 305. The encoder or coding model 305 can process the data series profile 315 and generates a coded data profile 319 (e.g., a coded battery aging profile of size (64, 1)). The coded data profile 319, along with a noise data vector 317, is then passed to the generation node 307. The generator model 307 can generate the profiles of one or more parameters 321 by processing the encoded data profile 319 and the noise data vector 317. The profiles of the one or more parameters 321 are then passed to the encoder or encoding model 309.The encoder or coding model 309 can process the profiles of one or more parameters 321 and generate the coded data profile 325 together with a noise data vector 325. The generated coded data profile 325, along with a noise data vector 323, is passed to the generator model 311. This process is repeated for various combinations of encoder and generator, and finally, the profiles of the one or more parameters are generated. The profiles of the one or more parameters include, for example, a temperature profile (TP), a voltage profile (VP), and a current profile (CP). In the next step, the profiles of the one or more parameters 327, along with the corresponding profiles of the one or more ideal parameters 329, are passed to a discriminator model 313.The discriminator model 313 can compare the generated profiles of one or more parameters 327 with the corresponding profiles of one or more ideal parameters 329 (e.g., a historical data profile with temporal data) and determine the loss value. The profiles of one or more parameters 327 can, for example, contain information about a temperature profile (TP), a voltage profile (VP), and a current profile (CP). The temperature profile information can include, for example, 2.62, 76.72, 60.00, 46.35, 26.15, 41.57, 83.81, 52.0, 69.8, 74.02. The information for the voltage profile can include, for example, 7.73, 1.83, 0.09, 6.3, 5.74, 3.5, 6.93, 7.26, 8.55, 8.98. The information for the current profile can include 3.04, 3.59, 1.55, 1.49, 0.97, 1.38, 4.34, 0.03, 4.48, 1.28. The corresponding profiles of one or more ideal parameters 329 include, for example, historical information about the temperature profile.A voltage profile and a current profile. Historical information for the voltage profile might include, for example, 48.26, 23.54, 57.18, 25.4, 51.66, 17.17, 69.51, 4.49, 99.0, 0.24, 7.2. Historical information for the current profile might include, for example, 6.69, 1.63, 0.42, 9.87, 2.52, 9.21, 5.27, 9.84, 7.83, 4.33, 8.22. Historical information for the current profile might include, for example, 2.54, 0.98, 3.6, 2.31, 4.12, 0.4, 0.51, 0.47, 4.75, 4.32, 0.8. The profiles of one or more parameters 327 and the corresponding profiles of one or more ideal parameters 329 are fed to the discriminator model 313. The discriminator model 313 can compare the profiles of one or more parameters 327 and the corresponding profiles of one or more ideal parameters 329, and the loss value (e.g., with the magnitude (256, 1)) is calculated. The loss value can be, for example, 0.61, 0.08, 0.039, 0.61, 0.29, 0.21, 0.The values ​​are 43, 0.83, 0.83, 0.7. After calculating the loss, the loss value is traced back and the hyperparameters are fine-tuned to train the generative model. Table 2 below provides detailed information on the variety of input data samples, coded data samples, data series profiles, coded data profiles, profiles of one or more parameters, and historical information about one or more parameters. Table 2 A variety of input data samples Encoded data patterns Profile of the data series Profiles of one or more parameters Historical information on one or more parameters Loss 0.18 0.22 97 TP W CP TP VP CP 0.74 0.88 0.28 97 2.62 7.73 3.04 48.26 6.69 2.54 0.61 0.03 0.03 96 76.72 1.83 3.59 23.54 1.63 0.98 0.08 0.03 1.0 95 60.0 0.09 1.55 57.18 0.42 3.6 0.39 0.98 0.25 92 46.35 6.3 1.49 25.4 9.87 2.31 0.61 0.96 0.1 92 26.15 5.74 0.97 51.66 2.52 4.12 0.29 0.26 0.89 90 41.57 3.5 1.38 17.17 9.21 0.4 0.21 0,27 0.33 90 83.81 6.93 4.34 69.51 5.27 0.51 0.43 0,43 0.85 90 52.0 7.26 0.03 4.49 9.84 0.47 0.83 0.41 89 69.8 8.55 4.48 99.0 7.83 4.75 0.83 0.23 88 74.02 8.98 1.28 0.24 4.33 4.32 0.7 88 7.2 8.22 0.8 86 83

[0034] The other use case is battery systems for use in vehicles. During operation, there may be one or more scenarios, such as predictive maintenance and diagnostics, adaptable safety systems, improved navigation and route planning, and improvements to the battery management system (BMS). For one application, the generative model can predict potential failures or the maintenance needs of vehicle components (e.g., engine, transmission, and brakes). For the integration point, the generative model can use the vehicle's on-board diagnostics (OBD) system. For the training part, the model can be trained using historical maintenance data to predict potential failures. For warnings and recommendations, the generative model can provide alerts and maintenance recommendations to the driver or maintenance team via the vehicle's dashboard or a connected app.For adaptive safety systems, the model can enhance the vehicle's safety systems, such as airbags, electronic stability control (ESC), and advanced driver assistance systems (ADAS), in the application part. For the training part, the model can be trained to generate scenarios for various safety events (e.g., collisions, skidding, etc.). For warning and recommendation, the model can adjust the vehicle's safety systems based on these predictions to increase protection (e.g., seatbelt pretensioning, ESC parameter adjustments). In the case of improved navigation and route planning, the model can optimize navigation and route planning for autonomous or semi-autonomous vehicles in the application part. For the integration part, the model can utilize the vehicle's navigation system and the autonomous driving control unit.For the training component, the model can be trained to generate and simulate various navigation scenarios, such as traffic jams, road closures, and adverse weather conditions. For the warning and recommendation function, the model can provide the vehicle's control unit with optimized routes and navigation instructions. Finally, in the application component, when improving the battery management system (BMS), the model can optimize the performance and lifespan of batteries in electric vehicles. For the training component, the model can be trained to simulate battery degradation and performance under different conditions. For warnings and recommendations, the model can implement adaptable strategies for cell balancing, thermal management, and extending battery lifespan.

[0035] Fig.Figure 3b shows an application case for training a generative model for modeling the operating profile of battery systems in accordance with some embodiments of the present disclosure.

[0036] First, a noise 331 (which can be in the form of a multitude of data series) can be received by the data platform 109. Specifically, the data can be recorded using sensors or a database. However, the collected data cannot be limited to voltage values, current values, temperature values, and the like.

[0037] First, the encoder or encoder block 301 can process the noise 331 and generate a corresponding coded data pattern. In the next step, the coded data sample is forwarded to the decoder or decoder block 303. The decoder or decoder block 303 can process the coded data sample and generate a data series profile 333 (e.g., a battery system profile) that corresponds to the input data sample or the noise 331. The generated data series profile 333 can be forwarded to the coding model 305. The coding model 305 can process the data series profile and generate a coded data profile 335 (e.g., a coded battery system profile) for the data series profile. Subsequently, the coded data profile 335, together with a noise data vector 337, is forwarded to the generation node 307.The generator model 307 can generate the profiles of one or more parameters 341 by processing the encoded data profile 335 and the noise data vector 337. The profiles of one or more parameters include, for example, a temperature profile, a voltage profile, and a current profile.

[0038] In the next step, the generated profiles of the one or more parameters 341 can be forwarded to the coding model 309. The coding model 309 can process the received data and generate coded profiles 339 of the one or more parameters, forwarding them to the generator 311. The generator 311 can process the generated profiles of the one or more parameters together with the noise data vector 343 and generate profiles of the one or more parameters 345 that differ from those of 341. Finally, the discriminator model 313 can compare the generated profiles of the one or more parameters 347 with the corresponding one or more ideal parameters 345 (e.g., a historical data profile with temporal data) and determine the loss value. The loss value is backpropagated, and the hyperparameters are fine-tuned to train the generative model.The encoder-decoder models, or blocks 301 and 303, are trained separately.

[0039] Fig. Figure 4 shows a flowchart illustrating a method 400 for training a generative model for analyzing data series in accordance with some embodiments of the present disclosure.

[0040] As in Fig.As shown in Figure 4, Procedure 400 can comprise one or more steps. Procedure 400 can be described in the general context of computer-executable instructions. Computer-executable instructions can generally include routines, programs, objects, components, data structures, procedures, modules, and functions that perform specific functions or implement specific abstract data types. The order in which Procedure 400 is described is not a restriction, and any number of the described procedure blocks can be combined in any order to implement the procedure. Furthermore, individual blocks can be omitted from the procedure without affecting the scope of the subject matter described herein. Additionally, the procedure can be implemented in any suitable hardware, software, firmware, or a combination thereof.

[0041] In block 402, procedure 400(a) comprises training a first model based on a plurality of input data samples 111. Each of the plurality of data samples 111 can comprise a plurality of data series relating to an operating profile. The operations of block 402 can be performed by the processor 107 (in particular by the training unit 115) of Fig. 1b. In block 404, procedure 400(b) comprises processing an input training data sample from a plurality of input training data samples using the trained first model to generate a data series profile corresponding to the input training data sample, each of the plurality of input training data samples comprising data series. The operations of block 404 can be performed by processor 107 (in particular by processing unit 117) of Fig. 1b will be carried out.

[0042] In block 406, procedure 400(c) comprises processing the data series profile (211) using one or more pairs of models to generate profiles of one or more parameters associated with the input training data sample. Each pair of the one or more pairs comprises a second model and a third model. The operations of block 406 can be performed by processor 107 (specifically by processing unit 117) of Fig. 1b. In block 408, procedure 400(d) comprises determining a first loss value using a fourth model by comparing the generated profiles of one or more parameters with corresponding profiles of one or more ideal parameters. The operations of block 408 can be performed by the processor 107 (in particular by the determination unit 119) of Fig.1b. In block 410, procedure 400(e) comprises training the generative model by repeating steps (b)-(d) for the remaining input training data samples of the plurality of input training data samples until the determined first loss value is less than a first threshold loss value in order to train the second, third, and fourth models. The operations of block 410 can be performed by processor 107 (specifically by training unit 115) of Fig. 1b will be carried out. Advantages of the present disclosure

[0043] In this disclosure, the generated profiles are valid and closely resemble real-world driving profiles. Accordingly, this disclosure offers efficient plausibility. The stacked architecture of the encoder, decoder, generator, and discriminator can be adapted to any generative modeling scenario, such as driving, charging, battery aging, fault code generation, component failure generation, etc. Consequently, the generative model can quickly adapt to multiple use cases with minimal fine-tuning. Within the scope of this disclosure, the generative model can be deployed in a shorter timeframe, reducing the need for extensive retraining. This disclosure utilizes multiple latent spaces within a single generative model that captures various aspects of the data, resulting in more diverse outcomes.In this disclosure, each latent container can specialize in capturing features specific to its respective data type, resulting in improved feature representation and generation. The separate latent spaces can capture more nuanced and diverse aspects of the data, leading to more realistic and detailed results. The model can learn more effectively because each encoder-decoder pair focuses on a specific data type, thus reducing complexity and potential interference between different data types. This enhances the model's performance.

[0044] The terms “an embodiment”, “elaboration”, “elaborations”, “the embodiment”, “the embodiments”, “one or more embodiments”, “some embodiments”, and “an embodiment” mean “one or more (but not all) embodiments of the invention(s)” unless expressly stated otherwise. The terms “including”, “comprising”, “not” and variations thereof mean “including, but not limited to”, unless expressly stated otherwise. The listing of items does not imply that all or some items are mutually exclusive unless expressly stated otherwise. The terms “a”, “a”, and “the” mean “one or more” unless expressly stated otherwise.

[0045] The description of an embodiment with several interconnected components does not mean that all of these components are required. On the contrary, a multitude of optional components are described to illustrate the wide variety of possible embodiments of the invention. Where a single device or article is described here, it is clear that more than one device / article (regardless of whether they interact) can be used instead of a single device / article. Where more than one device or article is described here (regardless of whether they interact), it is clear that a single device or article can be used instead of the multiple devices or articles, or that a different number of devices or articles can be used instead of the specified number of devices or programs.The functionality and / or features of a device can alternatively be embodied by one or more other devices that are not explicitly described as having this functionality / features. Therefore, other embodiments of the invention need not include the device itself.

[0046] Finally, the language used in the description was chosen primarily for readability and guidance purposes, and not to delimit or describe the subject matter of the invention. It is therefore intended that the scope of the invention is not limited by this detailed description, but rather by all claims based on an application thereunder. Accordingly, the embodiments of the present invention are intended for illustration, but not to limit the scope of the invention set forth in the following claims. While various aspects and embodiments have been disclosed here, other aspects and embodiments will be obvious to the person skilled in the art.The various aspects and embodiments disclosed herein serve for illustration and are not to be understood as limiting, the true scope and spirit of the invention being specified by the following claims. REFERENCE MARK: 101 System 103 Interface 105 storage 107 processor 109 Data platform 111 A large number of input data samples Issue 113 115 training units 117 processing units 119 Unit of determination 201 / 203 or 3011303 encoder / decoder model 205 / 305 coding model 207 / 307 / 311 Generator Model 209 / 313 Discriminator model 211 / 311 Data series profile 213 / 313 / 319 Coded Data Profile 215 / 315 / 337 / 343 Noise data vector 217 / 317 / 323 Profiles of one or more parameters 219 / 325 Profiles of the ideal one or more parameters QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 20220180249A1

[0004]

Claims

Method (400) for training a generative model for modeling operational profiles, the method comprising: (a) training (402) a first model based on a plurality of input data samples (111), wherein each of the plurality of input data samples (111) comprises a plurality of data series relating to an operational profile; (b) processing (404), using the trained first model of an input training data sample from a plurality of input training data samples, to generate a data series profile (211) corresponding to the input training data sample (111), wherein each of the plurality of input training data samples comprises data series;(c) Processing (406) the data series profile (211) using one or more pairs of models to generate profiles of one or more parameters associated with the input training data sample, each pair of the one or more pairs comprising a second model and a third model; (d) Determining (408) using a fourth model a first loss value by comparing the generated profiles of the one or more parameters with corresponding profiles of one or more ideal parameters; and (e) Training (410) the generative model by repeating steps (b)-(d) for the remaining input training data samples of the plurality of input training data samples until the determined first loss value is less than a first threshold loss value, in order to train the second, third, and fourth models. Method (400) according to claim 1, wherein the first model comprises a transphonator-based encoder-decoder model comprising an encoder block and a decoder block, the second model comprises an encoder model, the third model comprises a generator model, and the fourth model comprises a discriminator model, and wherein the training of the first model comprises: (a.1) processing an input data pattern from the plurality of input data patterns using the encoder block to generate an encoded data pattern; (a.2) processing the encoded data pattern using the decoder block to generate a data profile corresponding to the input data pattern; (a.3) determining a second loss value by comparing the generated data profile with a corresponding ideal data profile; and (a.4) training the first model by repeating steps (a.1)-(a.3) for the remaining data samples of the multitude of input data samples, until the determined second loss value is less than a second threshold loss value. Method (400) according to claim 1, wherein the processing of the data series profile using one or more pairs of models comprises: processing the data series profile using the second model to generate an encoded data profile for the data series profile; and generating the profiles of the one or more parameters by processing the encoded data profile and a noise data vector using the third model. The method (400) according to claim 1 further comprises: receiving a real-time input data sample relating to a specific operating profile; processing the received real-time input data sample using the first model to generate a data series profile corresponding to the real-time input data sample;Processing the data file profile corresponding to the real-time input data sample to generate profiles of one or more parameters associated with the real-time input data sample; and analyzing the generated profiles of one or more parameters associated with the real-time input data sample, wherein: the special operating profile is a vehicle speed profile, wherein the real-time input data sample includes vehicle-related driving data and the profiles of the one or more parameters include an accelerator pedal profile, a gear profile, and a brake torque profile; or the special operating profile is a vehicle battery aging profile, wherein the real-time input data sample includes vehicle-related battery data and the one or more parameters include a temperature profile, a voltage profile, and a current profile. System (101) for training a generative model for modeling operational profiles, wherein the system (101) comprises: a processor (107); and a memory (105) communicatively coupled to the processor (107), wherein the processor (113) is configured to: (a) train a first model based on a plurality of input data samples, each of the plurality of input data samples comprising a plurality of data series relating to an operational profile; (b) using the trained first model, process an input training data sample from a plurality of input training data samples to generate a data series profile (211) corresponding to the input training data sample, each of the plurality of input training data samples comprising data series; (c) process the data series profile (211) using one or more pairs of models to generate profiles of one or more parameters.which are associated with the input training data sample, wherein each pair of the one or more pairs comprises a second model and a third model: (d) to determine a first loss value using a fourth model by comparing the generated profiles of the one or more parameters with corresponding profiles of one or more ideal parameters; and (e) to train the generative model by repeating steps (b)-(d) for the remaining input training data samples of the plurality of input training data samples until the determined first loss value is less than a first threshold loss value, in order to train the second, third, and fourth models. System (101) according to claim 5, wherein the first model comprises a transphonator-based encoder-decoder model comprising an encoder block and a decoder block, the second model comprises an encoding node, the third model comprises a generator model, and the fourth model comprises a discriminator model, and wherein the processor for training the first model is configured to: (a.1) process an input data pattern from the plurality of input data patterns using the encoder block to generate an encoded data pattern; (a.2) process the encoded data pattern using the decoder block to generate a data profile corresponding to the input data pattern; (a.3) determine a second loss value by comparing the generated data profile with a corresponding ideal data profile; and (a.4) train the first model by repeating steps (a.1)-(a.3) to train on the remaining data samples of the multitude of input data samples until the determined second loss value is less than a second threshold loss value. System (101) according to claim 5, wherein the processor is configured to process the data series profile using one or more pairs of models to generate profiles of one or more parameters: to process the data series profile using the second model to generate an encoded data profile for the data series profile: and to generate the profiles of one or more parameters by processing the encoded data profile and a noise data vector using the third model. System (101) according to claim 5, wherein the processor is configured to: receive a real-time input data sample relating to a specific operating profile; process the received real-time input data sample using the first model to generate a data series profile corresponding to the real-time input data sample; process the data series profile corresponding to the real-time input data sample to generate profiles of one or more parameters associated with the real-time input data sample;and analyzes the generated profiles of one or more parameters associated with the real-time input data sample, wherein: the particular operating profile is a vehicle speed profile, wherein the real-time input data sample includes vehicle-related driving data and the profiles of the one or more parameters include an accelerator pedal profile, a gear profile, and a brake torque profile; or the particular operating profile is an aircraft battery aging profile, wherein the real-time input data sample includes vehicle-related battery data and the one or more parameters include a temperature profile, a voltage profile, and a current profile. Network architecture of a generative model for modeling operational profiles, wherein the network architecture comprises: a transformer-based model (201, 203) comprising an encoder block (201) and a decoder block (203) joined together, wherein the transformer-based model (201, 203) is configured to produce an initial output (211) based on input data samples (111); one or more pairs of models connected to an output of the transformer-based model (201).203) are stacked and configured to generate a second output signal (213) based on the first output signal (211), wherein each pair of the one or more pairs of models comprises a coding model (205) configured to receive the first output signal and a generator model (207) stacked at one output of the coding model (205); and a discriminator model (209) stacked at the second output of the one or more pairs of models and configured to generate a loss value based on the second output for training the transphonator-based model and the one or more pairs of models for modeling the operating profiles. Network architecture according to claim 9, wherein the loss value is a sum of one or more of the following values: pair loss, Kullback-Leibler (KL) divergence loss, structure similarity index (SSIM) loss, least squared error (L2) loss, and dynamic time warping (DTW) loss.

Images