Tata Consultancy Services

In traditional ensemble modeling multiple individual models are trained on the same set of input variables without considering variations in spatial and temporal resolutions. A method and system for dynamic forest fire prediction and source prediction is proposed using dynamic weighting-based ensemble modeling and probabilistic cause determination. The adaptive ensemble model dynamically re-weights multiple base models, each optimized for different data scales, to provide accurate predictions based on the spatial and temporal granularity of incoming data. The method enables adjusting fire predictions probabilities due to shifts in the feature space caused by phenological and environmental changes. Key environmental variables are continuously monitored for recalibrating model predictions using shift detection algorithms, ensuring long-term accuracy without frequent retraining. Probabilistic determination is applied for likely causes of forest fires, which integrates data from various sources to provide a nuanced, data-driven understanding of the factors contributing to fire outbreaks.

The disclosure relates generally to method and system to generate decision-fused classification map using different modalities in remote sensing applications. In remote sensing applications, the problem of missing data from satellite images occurs due to various environmental factors and spatial resolution provides inconsistent class labels during segmentation. The method receives from each satellite among a plurality of satellites a plurality of remote sensing images capturing a land use and land cover (LULC) geographical area on earth. Each pretrained classifier obtains the plurality of remote sensing images to generate a segmentation map. Further, Kolmogorov Arnold networks (KAN) decision fusion combines two or more closest segmentation maps using an ontological knowledge tree to determine a minimal common parent class based on at least one satisfying criteria. Finally, a fused classification map is generated to obtain finer grained LULC classes for each remote sensing image using the ontological knowledge tree.

Embodiments herein provide a method and system for splitting an artificial neural network (ANN) between edge devices (edge) and a user equipment (UE) under supervision of an application cloud (AC) through a unified splitting AI framework (SAF). In AI operation splitting, a few layers of the artificial neural network (ANN) are computed at the endpoint itself. The intermediate result is transferred to the Edge/ Cloud for further computation. Besides distributing the computation load, this has another important advantage. If the first few layers of the ANN are computed at the endpoint itself, then that saves the resources required for transmission of the entire input signal because the intermediate result needs much less bandwidth than the actual input. Also, the ANN could start in the offloaded node only after receiving the entire input. So, splitting the ANN computation is beneficial in terms of the total computation latency as well.

This disclosure relates generally to a method and system for predicting cardiologist level clinical explanations for electrocardiogram signal classification using deep learning. Conventional methods for explainability of artificial intelligence models are influenced by data distribution and do not provide clinical level explanations in cardiac healthcare. The method disclosed provides clinical level explanations for rhythm diagnosis from ECG signals. The method validates that clinical concepts in ECG signals can be extracted by generating a set of synthetic ECG signals. The set of synthetic ECG signals are generated using trained generative adversarial networks and statistical parameterization techniques. Further the set of synthetic ECG signals are validated by classifying and comparing them with a set of real ECG signals. Then a clinical concept identifier model is trained for predicting the clinical concepts present in an ECG signal classified into a set of signal classes. These clinical concepts predicted are cardiologist level explanations.

The disclosure relates generally to methods and systems for generating design component knowledge graph and design component retrieval from the same. Conventional techniques that utilize customizable knowledge graphs for retrieval of the design components are very limited in the digital design domain. The methods and systems of the present disclosure generate the design component knowledge graph using filtered captions generated for the design component objects. The design component objects are of the design components and are selected using an instruction triplet generated based on a user requirement query of the digital design that the user is interested in. Then, a plurality of captions is generated and filtered using a semantic filtering to obtain the filtered captions. Then the design component knowledge graph is queried with the user input to retrieve the design components that best suit the design requirement of the user.

Usage of generative artificial intelligence (GenAI) models require appropriate prompts that describe a task to be completed. Conventionally complexities of large sizes and monolithic nature of input code are difficult for handling. The present disclosure resolves problems of conventional approaches by providing a system and method for generation of prompts for GenAI model using contextual information from legacy source code. The method of the present disclosure extracts relevant information from the legacy source code as context in natural language form which is provided as input to create the prompts required to enable successful usage of the GenAI model in multiple tasks of code analysis. In the present disclosure, information from legacy application code about business domain is extracted. Input legacy source code is parsed, and technical explanation is provided by handling the syntax and semantics of language. The legacy source code is split logically into blocks of manageable units.

The Range Migration Algorithm (RMA) is widely used for SAR image reconstruction from backscattered signals. Traditional RMA relies on the Fast Fourier Transform (FFT), which is unsuitable for irregular scanning trajectories. An interpolator based reconstruction method, referred as Local Interpolation Function-based Range Migration Algorithm (LIF-RMA), is provided for SAR imaging. The LIF-RMA efficiently addresses challenges associated with irregularly sampled trajectories. The LIF-RMA utilizes a complex-valued encoder-decoder network to transform irregularly sampled, complex-valued backscattered data into uniformly sampled complex raw data. The complex-valued encoder maps the distorted raw data into a feature space, while the complex-valued decoder interpolates and predicts the interpolated complex raw data at missing spatial coordinates. This data is then processed through a 2D FFT block before being fed to RMA for final SAR image reconstruction. In this process, as the LIF is applied prior to RMA, it aids in preserving phase integrity.

The embodiments of the present disclosure herein address unresolved problems of communication from the semantic aspects. Embodiments herein provide a method and system for a semantic communication with networked AI in sixth-generation technology for wireless communications (6G). Here, rather than encoding the input and transmitting the encoded symbols, the semantically relevant information is inferred from the input as per the context of the application and that information is encoded through conventional source encoding and transmitted. Conversion of the actual input to semantic information allows a huge reduction in the bandwidth requirement for the end-to-end channel. On the receiver side, these symbols are first decoded to extract semantic information. The semantic information is further used to predict the semantically relevant output to carry out the intended meaning at the receiver side.

Conventional traffic classification methods mainly depend on pre-defined criteria from known data, highlighting demand for advanced methods that can handle new types of traffic. The present disclosure receives an unlabeled data set comprising a plurality of data points from one or more Downlink Control Information messages and pre-processes the received unlabeled data set. A set of relevant features is selected from plurality of data points using a correlation matrix. One or more hyper-parameters of Gaussian Mixture Model (GMM) are evaluated for the unlabeled dataset with selected set of relevant features using Component-wise space Expectation Maximization technique (CEM) method. An optimal number of clusters is estimated using the evaluated one or more hyper-parameters of GMM. Each of the plurality of data points are labeled using the estimated optimal number of clusters to obtain a labelled data set. A classifier model is created to perform traffic classification of the obtained labelled dataset.

Conventional risk estimation techniques perform dynamic analysis of application or use models which require training data. Present disclosure provides method and system to estimate risk for a software application due to quantum threat by static analysis. A set of records pertaining to the application is received and parsed to obtain application, crypto and platform parameters. In addition, list of quantum vulnerable algorithms, number of Qubits required to break a cryptographic algorithm used by the application and a current Qubit number are also received. Then, value of Quantum Day is determined based on the current Qubit number and the number of Qubits required to break the cryptographic algorithm used by the application. Further, SOD (Severity, Occurrence, Detection) scores are calculated for each parameter, and they are multiplied to determine Risk Priority Number (RPN) for each parameter. Finally, RPNs of all parameters are summed up to estimate overall risk of the application.