Abstract: Techniques are described for identifying root cause anomalies in time series. Information to be used for root cause analysis (RCA) is obtained from a graph neural network (GNN) and is used to construct a dependency graph having nodes corresponding to each time series and directed edges corresponding to dependencies between the time series. Nodes corresponding to time series that do not contain anomalies may be removed from this dependency graph, as well as edges connected to these nodes. This edge and node removal may result in the creation of one or more sub-graphs from the dependency graph. A root cause analysis algorithm may be run on these one or more sub-graphs to create a root cause graph for each sub-graph. These root cause graphs may then be used to identify root cause anomalies within the multiple time series, as well as sequences of anomalies within the multiple time series.

Abstract: Techniques for identifying anomalous multi-source data points and ranking the contributions of measurement sources of the multi-source data points are disclosed. A system obtains a data point including a plurality of measurements from a plurality of sources. The system determines that the data point is an anomalous data point based on a deviation of the data point from a plurality of additional data points. The system determines a contribution of two or more measurements, from the plurality of measurements, to the deviation of the data point from the plurality of additional data points. The system ranks the at least the two or more measurements, from the plurality of measurements, based on the respective contribution of each of the two or more measurements to the deviation of the anomalous data point from the plurality of prior data points.

Abstract: Embodiments determine an optimized maintenance schedule for a maintenance program that includes multiple levels, each level including at least one asset (i.e., asset type) and at least one of the levels including a plurality of assets. Embodiments receive historical failure data for each of the assets, the historical failure data generated at least in part by a sensor network. For each asset, embodiments generate a probability density function (“PDF”) using kernel density estimation (“KDE”). For each asset, based on a reliability rate threshold, embodiments determine a cumulative density function (“CDF”) using the PDF. For each asset, embodiments determine an optimized time to failure (“TTF”) using the CDF. Embodiments then create the schedule for each level that includes a minimum TTF for the assets at each level.

Abstract: Techniques for improving system performance based on data characteristics are disclosed. A system may receive updates to a first data set at a first frequency. The system selects a first storage configuration, from a plurality of storage configurations, for storing the first data set based on the first frequency, and stores the first data set in accordance with the first storage configuration. The system may further receive updates to a second data set at a second frequency. The system selects a second storage configuration, from the plurality of storage configurations, for storing the second data set based on the second frequency, and stores the second data set in accordance with the second storage configuration. The second storage configuration is different than the first storage configuration.

Abstract: Techniques are described for providing explanation information for time series-based predictions made using statistical models, such as linear statistical models, examples of which include various Exponential Smoothing models, Autoregressive Integrated Moving Average (ARIMA) models, and others. For a forecast predicted by a statistical model that has been trained upon and/or fit to a set of historical times series data points, an explanation is generated for the forecast, where the explanation for the forecast includes information indicative of the importance or impact or influence of individual time series data points in the set on the forecast. The explanation for the forecast may be output to a user along with the forecast. This enables the user to have some visibility into why the particular forecast was predicted by the statistical model.

Abstract: Techniques for improving system performance based on data characteristics are disclosed. A system may receive updates to a first data set at a first frequency. The system selects a first storage configuration, from a plurality of storage configurations, for storing the first data set based on the first frequency, and stores the first data set in accordance with the first storage configuration. The system may further receive updates to a second data set at a second frequency. The system selects a second storage configuration, from the plurality of storage configurations, for storing the second data set based on the second frequency, and stores the second data set in accordance with the second storage configuration. The second storage configuration is different than the first storage configuration.

Abstract: A model analyzer may receive a representative data set as input and select one of a plurality of analytic models to perform the analysis. Before deciding which model to use the model may be trained, and the trained model evaluated for accuracy. However, some models are known to behave poorly when the training data is distributed in a particular way. Thus, the cost of training a model and evaluating the trained model can be avoided by first analyzing the distribution of the representative data. Identifying the representative data distribution allows ruling out use of models for which the distribution of the representative data is unsuitable. Only models that may be compatible with the distribution of the representative data may be trained and evaluated for accuracy. The most accurate trained model whose accuracy meets an accuracy threshold may be selected to analyze subsequently received data related to the representative data.

Abstract: Embodiments determine anomalies in sensor data generated by a sensor by receiving an evaluation time window of clean sensor data generated by the sensor. Embodiments receive a threshold value for determining anomalies. When the clean sensor data has a cyclic pattern, embodiments divide the evaluation time window into a plurality of segments of equal length, wherein each equal length comprises the cyclic pattern. When the clean sensor data does not have the cyclic pattern, embodiments divide the evaluation time window into a pre-defined number of plurality of segments of equal length. Embodiments convert the evaluation time window and each of the plurality of segments into corresponding curves using Kernel Density Estimation (“KDE”). For each of the plurality of segments, embodiments determine a Kullback-Leibler (“KL”) divergence value between corresponding curves of the segment and the evaluation time window to generate a plurality of KL divergence values.

Abstract: Techniques for providing actionable recommendations for configuring system parameters are disclosed. A set of environmental constraints and a set of values for a set of parameters for a target device is applied to a machine learning model to predict a first performance value of the target device. Candidate values for the set of parameters are identified that are within a threshold range from the first set of values in a multi-dimensional space. For each particular candidate set of values of the candidate sets of values the machine learning model to predicts a performance value of the target device and identifies a subset of the candidate sets of values with corresponding performance values that meet a performance criteria. A subset of candidate sets of values that meets performance criteria is provided as a recommendation.

Abstract: Techniques for selecting universal hyper parameters for use in a set of machine learning models across multiple computing environments include detection of a triggering condition for tuning a set of universal hyper parameters. The set of universal hyper parameters dictate configuration of the set of machine learning models that are independently executing, respectively, in the multiple computing environments. Based on the detected triggering condition, a first subset of universal hyper parameters from the set of universal hyper parameters are altered to generate a second set of universal hyper parameters. The second set of universal hyper parameters are applied to the set of machine learning models across the multiple computing environments.

Abstract: Techniques for generating a composite score for data quality are disclosed. Univariate analysis is performed on a plurality of data points corresponding to each of a first feature, a second feature, and a third feature of a data set. The univariate analysis includes at least a first type of analysis generating a first score having a first range of possible values, and a second type of analysis generating a second score having a second range of possible values. A first quality score is computed for the data values for the first, second, and third features based on a normalized first score and a normalized second score. Machine learning is performed on the data points corresponding to one or both of the first feature and the second feature having a first quality score above a threshold value to model the third feature.

Abstract: Embodiments determine anomalies in sensor data generated by a plurality of sensors that correspond to a single asset. Embodiments receive a first time window of clean sensor input data generated by the sensors, the clean sensor data including anomaly free data comprised of clean data points. Embodiments divide the clean data points into training data points and evaluation data points, and divide the training data points into a pre-defined number of plurality of segments of equal length. Embodiments convert each of the plurality of segments into corresponding segment curves using Kernel Density Estimation (“KDE”) and determine a Jensen-Shannon (“JS”) divergence value for each of the plurality of segments using the segment curves to generate a plurality of JS divergence values. Embodiments then assign the maximum value of the plurality of JS divergence values as a threshold value and validate the threshold value using the evaluation data points.

Abstract: Techniques for identifying anomalous multi-source data points and ranking the contributions of measurement sources of the multi-source data points are disclosed. A system obtains a data point including a plurality of measurements from a plurality of sources. The system determines that the data point is an anomalous data point based on a deviation of the data point from a plurality of additional data points. The system determines a contribution of two or more measurements, from the plurality of measurements, to the deviation of the data point from the plurality of additional data points. The system ranks the at least the two or more measurements, from the plurality of measurements, based on the respective contribution of each of the two or more measurements to the deviation of the anomalous data point from the plurality of prior data points.

Abstract: Embodiments determine anomalies in sensor data generated by a plurality of sensors that correspond to a single asset. Embodiments receive a first time window of clean sensor input data generated by the sensors, the clean sensor data including anomaly free data comprised of clean data points. Embodiments divide the clean data points into training data points and evaluation data points, and divide the training data points into a pre-defined number of plurality of segments of equal length. Embodiments convert each of the plurality of segments into corresponding segment curves using Kernel Density Estimation (“KDE”) and determine a Jensen-Shannon (“JS”) divergence value for each of the plurality of segments using the segment curves to generate a plurality of JS divergence values. Embodiments then assign the maximum value of the plurality of JS divergence values as a threshold value and validate the threshold value using the evaluation data points.

Abstract: Embodiments determine anomalies in sensor data generated by a sensor. Embodiments receive a first time window of clean sensor data generated by the sensor, the clean sensor data including anomaly free data, and determine if the clean sensor data includes a cyclic pattern. When the clean sensor data has a cyclic pattern, embodiments divide the first time window into a plurality of segments of equal length, where each equal length includes the cyclic pattern. Embodiments convert the first time window and each of the plurality of segments into corresponding curves using Kernel Density Estimation (“KDE”). For each of the plurality of segments, embodiments determine a Kullback-Leibler (“KL”) divergence value between corresponding curves of the segment and the first time window to generate a plurality of KL divergence values.

Abstract: Embodiments determine an optimized maintenance schedule for a maintenance program that includes multiple levels, each level including at least one asset (i.e., asset type) and at least one of the levels including a plurality of assets. Embodiments receive historical failure data for each of the assets, the historical failure data generated at least in part by a sensor network. For each asset, embodiments generate a probability density function (“PDF”) using kernel density estimation (“KDE”). For each asset, based on a reliability rate threshold, embodiments determine a cumulative density function (“CDF”) using the PDF. For each asset, embodiments determine an optimized time to failure (“TTF”) using the CDF. Embodiments then create the schedule for each level that includes a minimum TTF for the assets at each level.

Abstract: Embodiments determine anomalies in sensor data generated by a sensor. Embodiments receive a first time window of clean sensor data generated by the sensor, the clean sensor data including anomaly free data, and determine if the clean sensor data includes a cyclic pattern. When the clean sensor data has a cyclic pattern, embodiments divide the first time window into a plurality of segments of equal length, where each equal length includes the cyclic pattern. Embodiments convert the first time window and each of the plurality of segments into corresponding curves using Kernel Density Estimation (“KDE”). For each of the plurality of segments, embodiments determine a Kullback-Leibler (“KL”) divergence value between corresponding curves of the segment and the first time window to generate a plurality of KL divergence values.

Abstract: Embodiments determine anomalies in sensor data generated by a sensor by receiving an evaluation time window of clean sensor data generated by the sensor. Embodiments receive a threshold value for determining anomalies. When the clean sensor data has a cyclic pattern, embodiments divide the evaluation time window into a plurality of segments of equal length, wherein each equal length comprises the cyclic pattern. When the clean sensor data does not have the cyclic pattern, embodiments divide the evaluation time window into a pre-defined number of plurality of segments of equal length. Embodiments convert the evaluation time window and each of the plurality of segments into corresponding curves using Kernel Density Estimation (“KDE”). For each of the plurality of segments, embodiments determine a Kullback-Leibler (“KL”) divergence value between corresponding curves of the segment and the evaluation time window to generate a plurality of KL divergence values.

Abstract: Techniques for improving system performance based on data characteristics are disclosed. A system may receive updates to a first data set at a first frequency. The system selects a first storage configuration, from a plurality of storage configurations, for storing the first data set based on the first frequency, and stores the first data set in accordance with the first storage configuration. The system may further receive updates to a second data set at a second frequency. The system selects a second storage configuration, from the plurality of storage configurations, for storing the second data set based on the second frequency, and stores the second data set in accordance with the second storage configuration. The second storage configuration is different than the first storage configuration.

Abstract: Embodiments implement a machine learning prediction model with dynamic data selection. A number of data predictions generated by a trained machine learning model can be accessed, where the data predictions include corresponding observed data. An accuracy for the machine learning model can be calculated based on the accessed number of data predictions and the corresponding observed data. The accessing and calculating can be iterated using a variable number of data predictions, where the variable number of data predictions is adjusted based on an action taken during a previous iteration, and, when the calculated accuracy fails to meet an accuracy criteria during a given iteration, a training for the machine learning model can be triggered.