Abstract: Systems and methods for optimizing key performance indicators (KPIs) using adversarial imitation deep learning include processing sensor data received from sensors to remove irrelevant data based on correlation to a final KPI and generating, using a policy generator network with a transformer-based architecture, an optimal sequence of actions based on the processed sensor data. A discriminator network is employed to differentiate between the generated action sequences and real-world high performance sequences employing. Final KPI results are estimated based on the generated action sequences using a performance prediction network. The generated action sequences are applied to the process to optimize the KPI in real-time.

Abstract: Systems and methods train a transformer-based policy network and Generative Adversarial Network (GAN) by initializing a transformer-based policy network to model action sequences by encoding temporal dependencies within sensor data. Multi-head self-attention mechanisms process sequential sensor inputs by being pre-trained on a labeled dataset having sensor data from known low-risk action sequences. A generator within the GAN is trained to produce generated action sequences, which mimic behavior of low-risk action sequences. A discriminator within the GAN is concurrently trained to differentiate between action sequences derived from the labeled dataset and synthetic action sequences produced by the generator. A feedback loop is employed to adjust parameters to produce sequences indistinguishable from real low-risk action sequences.

Abstract: Systems and methods are provided for classifying components include monitoring sensors to collect sensor data related to a state of a plurality of components; processing, by a computing system, the sensor data to generate an action sequence using a transformer-based policy network for each of the components. A risk score is generated for the action sequence using a Generative Adversarial Network (GAN), wherein the GAN includes a generator for generating action sequences and a discriminator to distinguish low-risk action sequences in accordance with a threshold. The low-risk action sequences are associated with components in the plurality of components based on the risk score. A status of the low-risk action sequences is communicated to the components.

Abstract: Systems and methods for an agent-based carbon emission reduction system. A carbon product of a supply chain system can be limited below a carbon product threshold by performing a corrective action to monitored entities based on a calculated carbon emission. The carbon emission can be calculated based on carbon-relevant data and a calculation route by utilizing an agent-based simulation model that simulates a learned relationship between a supply chain system and the carbon-relevant data. The calculation route can be determined based on the carbon-relevant data based on a relevance of a carbon product contribution of monitored entities to a goal of the monitored entities. Carbon-relevant data can be extracted from the monitored entities.

Abstract: Systems and methods for training a time-series-language (TSLa) model adapted for domain-specific tasks. An encoder-decoder neural network can be trained to tokenize time-series data to obtain a discrete-to-language embedding space. The TSLa model can learn a linear mapping function by concatenating token embeddings from the discrete-to-language embedding space with positional encoding to obtain mixed-modality token sequences. Token augmentation can transform the tokens from the mixed-modality token sequences with to obtain augmented tokens. The augmented tokens can train the TSLa model using a computed token likelihood to predict next tokens for the mixed-modality token sequences to obtain a trained TSLa model. A domain-specific dataset can fine-tune the trained TSLa model to adapt the trained TSLa model to perform a domain-specific task.

Abstract: Systems and methods for data fusion and analysis of vehicle sensor data, including receiving a multiple modality input data stream from a plurality of different types of vehicle sensors, determining latent features by extracting modality-specific features from the input data stream, and aligning a distribution of the latent features of different modalities by feature-level data fusion. Classification probabilities can be determined for the latent features using a fused modality scene classifier. A tree-organized neural network can be trained to determine path probabilities and issue driving pattern judgments, with the tree-organized neural network including a soft tree model and a hard decision leaf. One or more driving pattern judgments can be issued based on a probability of possible driving patterns derived from the modality-specific features.

Abstract: Systems and methods for log representation learning for automated system maintenance. An optimized parser can transform collected system logs into log templates. A tokenizer can tokenize the log templates partitioned into time windows to obtain log template tokens. The log template tokens can train a language model (LM) with deep learning to obtain a trained LM. The trained LM can detect anomalies from system logs to obtain detected anomalies. A corrective action can be performed on a monitored entity based on the detected anomalies.

Abstract: Systems and methods for a demonstration uncertainty-based artificial intelligence model for open information extraction. A large language model (LLM) can generate initial structured sentences using an initial prompt for a domain-specific instruction extracted from an unstructured text input. Structural similarities between the initial structured sentences and sentences from a training dataset can be determined to obtain structurally similar sentences. The LLM can identify relational triplets from combinations of tokens from generated sentences using and the structurally similar sentences. The relational triplets can be filtered based on a calculated demonstration uncertainty to obtain a filtered triplet list. A domain-specific task can be performed using the filtered triplet list to assist the decision-making process of a decision-making entity.

Abstract: Systems and methods for unsupervised multi-modal causal structure learning for root cause analysis. System logs of a cloud system can be transformed to time-series data using a log-tailored language model to obtain system log features of the cloud system. A metric causal graph and a log causal graph can be predicted from modality-specific representations and modality-invariant representations of extracted system metric features and system log features, respectively, using the deep neural network. The metric causal graph and log causal graph can be fused to obtain a fused causal graph. Root causes of system failure can be flagged for system maintenance based on ranked entities obtained from the fused causal graph to obtain flagged root causes. System maintenance can be performed autonomously based on the flagged root causes from identified system entities to optimize the cloud system with an updated configuration.

Abstract: Systems and methods for correlation-aware explainable online change point detection. Collected data metrics from the cloud system can be transformed to correlation matrices. Correlation shifts from the correlation matrices can be captured as differences of correlation between batches of collected data metrics through determined statistics of the batches of collected data metrics across timesteps. Change points in the cloud system can be detected based on the correlation shifts to obtain detected change points. System maintenance can be performed autonomously based on the detected change points from identified system entities to optimize the cloud system with an updated configuration.

Abstract: A method for early warning is provided. The method clusters normal historical data of normal cars into groups based on the car subsystem to which they belong. The method extracts (i) features based on group membership and (ii) feature correlations based on correlation graphs formed from the groups. The method trains an Auto-Encoder and Auto Decoder (AE&AD) model based on the features and the feature correlations to reconstruct the normal historical data with minimum reconstruction errors. The method reconstructs, using the trained AE&AD model, historical data of specific car fault types with reconstruction errors, normalizes the reconstruction errors, and selects features of the car faults with a top k large errors as fault signatures. The method reconstructs streaming data of monitored cars using the trained AE&AD model to determine streaming reconstruction errors, comparing the streaming reconstruction errors with the fault signatures to predict and provide alerts for impending known faults.

Abstract: A computer-implemented method for multi-model representation learning is provided. The method includes encoding, by a trained time series (TS) encoder, an input TS segment into a TS-shared latent representation and a TS-private latent representation. The method further includes generating, by a trained text generator, a natural language text that explains the input TS segment, responsive to the TS-shared latent representation, the TS-private latent representation, and a text-private latent representation.

Abstract: A computer-implemented method for identifying root cause failure and fault events is provided. The method includes detecting a trigger point, converting, via an encoder, previous system state data, new batch data in a next system state, and a causal graph to system state-invariant embeddings and system state-dependent embeddings, generating a learned causal graph, via a graph generation layer, by integrating state-invariant and state-dependent information, and predicting, by a prediction layer, future time-series data on the learned causal graph.

Abstract: Methods and systems for information extraction include configuring a language model with an information extraction instruction prompt and at least one labeled example prompt. Configuration of the language model is validated using at least one validation prompt. Errors made by the language model in response to the at least one validation prompt are corrected using a correction prompt. Information extraction is performed on an unlabeled sentence using the language model to identify a relation from the unlabeled sentence. An action is performed responsive to the identified relation.

Abstract: Systems and methods are provided for incident analysis in Cyber-Physical Systems (CPS) using a Temporal Graph-based Incident Analysis System (TGIAS) and/or Transition Based Categorical Anomaly Detection (TCAD). Dynamically gathered multimodal data from a distributed network of sensors across the CPS are preprocessed to identify abnormal sensor readings indicative of potential incidents, and a multi-layered incident timeline graph, representing abnormal sensor readings, relationships to specific CPS components, and temporal sequencing of events is constructed. Severity scores are calculated, and severity rankings are assigned to identified anomalies based on a composite index including impact on CPS operation, comparison with historical incident data, and predictive risk assessments.

Abstract: Systems and methods are provided for incident analysis in Cyber-Physical Systems (CPS) using a Temporal Graph-based Incident Analysis System (TGIAS) and/or Transition Based Categorical Anomaly Detection (TCAD). Dynamically gathered multimodal data from a distributed network of sensors across the CPS are preprocessed to identify abnormal sensor readings indicative of potential incidents, and a multi-layered incident timeline graph, representing abnormal sensor readings, relationships to specific CPS components, and temporal sequencing of events is constructed. Severity scores are calculated, and severity rankings are assigned to identified anomalies based on a composite index including impact on CPS operation, comparison with historical incident data, and predictive risk assessments.

Abstract: Methods and systems for anomaly detection include encoding a time series with a time series encoder and encoding an event sequence with an event sequence encoder. A latent code is generated from outputs of the time series encoder and the event sequence encoder. The time series is reconstructed from the latent code using a time series decoder. The event sequence is reconstructed from the latent code using an event sequence decoder. An anomaly score is determined based on a reconstruction loss of the reconstructed time series and a reconstruction loss of the reconstructed event sequence. An action is performed responsive to the anomaly score.

Abstract: Methods and systems for anomaly detection include encoding a multivariate time series and a multi-type event sequence using respective transformers and an aggregation network to generate a feature vector. Anomaly detection is performed using the feature vector to identify an anomaly within a system. A corrective action is performed responsive to the anomaly to correct or mitigate an effect of the anomaly. The detected anomaly can be used in a healthcare context to support decision making by medical professionals with respect to the treatment of a patient. The encoding may include machine learning models to implement the transformers and the aggregation network using deep learning.

Abstract: A computer-implemented method for ordinal prediction is provided. The method includes encoding time series data with a temporal encoder to obtain latent space representations. The method includes optimizing the temporal encoder using semi-supervised learning to distinguish different classes in the labeled space using labeled data, and augment the latent space representations using unlabeled training data, to obtain semi-supervised representations. The method further includes discarding a linear layer after the temporal encoder and fixing the temporal encoder. The method also includes training k-1 binary classifiers on top of the semi-supervised representations to obtain k-1 binary predictions. The method additionally includes identifying and correcting inconsistent ones of the k-1 binary predictions by matching the inconsistent ones to consistent ones of the k-1 binary predictions. The method further includes aggregating the k-1 binary predictions to obtain an ordinal prediction.

Abstract: A computer-implemented method for identifying attack origins is provided. The method includes detecting a trigger point from entity metrics data and key performance indicator (KPI) data, generating a learned causal graph by fusing a state-invariant causal graph with a state-dependent causal graph, backtracking from an attack detection point, via an incident backtrack and system recovery component, by using the learned causal graph to identify an attack origin when an intrusion or attack occurs, and displaying data relating to the attack origin on a visualization display for user analysis.