Abstract: The disclosed embodiments provide a system that detects sensor anomalies in a univariate time-series signal. During a surveillance mode, the system receives the univariate time-series signal from a sensor in a monitored system. Next, the system performs a staggered-sampling operation on the univariate time-series signal to produce N sub-sampled time-series signals, wherein the staggered-sampling operation allocates consecutive samples from the univariate time-series signal to the N sub-sampled time-series signals in a round-robin ordering. The system then uses a trained inferential model to generate estimated values for the N sub-sampled time-series signals based on cross-correlations with other sub-sampled time-series signals. Next, the system performs an anomaly detection operation to detect incipient sensor anomalies in the univariate time-series signal based on differences between actual values and the estimated values for the N sub-sampled time-series signals.

Abstract: Systems, methods, and other embodiments associated with clustering of time series signals based on frequency domain analysis are described. In one embodiment, an example method includes accessing time series signals to be separated into clusters. The example method also includes determining similarity in the frequency domain among the time series signals. The example method further includes extracting a cluster of similar time series signals from the time series signals based on the similarity in the frequency domain. And, the example method includes training a machine learning model to detect anomalies based on the cluster.

Abstract: The disclosed system produces synthetic signals for testing machine-learning systems. During operation, the system generates a set of N composite sinusoidal signals, wherein each of the N composite sinusoidal signals is a combination of multiple constituent sinusoidal signals with different periodicities. Next, the system adds time-varying random noise values to each of the N composite sinusoidal signals, wherein a standard deviation of the time-varying random noise values varies over successive time periods. The system also multiplies each of the N composite sinusoidal signals by time-varying amplitude values, wherein the time-varying amplitude values vary over successive time periods. Finally, the system adds time-varying mean values to each of the N composite sinusoidal signals, wherein the time-varying mean values vary over successive time periods.

Abstract: A model-based approach to determining an optimal configuration for a data center may use an environmental chamber to characterize the performance of various data center configurations at different combinations of temperature and altitude. Telemetry data may be recorded from different configurations as they execute a stress workload at each temperature/altitude combination, and the telemetry data may be used to train a corresponding library of models. When a new data center is being configured, the temperature/altitude of the new data center may be used to select a pre-trained model from a similar temperature/altitude. Performance of the current configuration can be compared to the performance of the model, and if the model performs better, a new configuration based on the model may be used as an optimal configuration for the data center.

Abstract: Systems, methods, and other embodiments associated with quadratic acceleration boost of compute performance for ML prognostics are described. In one embodiment, a prognostic acceleration method includes separating time series signals into a plurality of alternative configurations of clusters based on correlations between the time series signals. Machine learning models are trained for individual clusters in the alternative configurations of clusters. One or more of the alternative configurations of clusters is determined to be viable for use in a production environment based on whether the trained machine learning models for the individual clusters satisfy an accuracy threshold and a completion time threshold. Then, one configuration is selected from the alternative configurations of clusters that were determined to be viable configurations. Production machine learning models are deployed into the production environment to detect anomalies in the time series signals based on the selected configuration.

Abstract: Systems, methods, and other embodiments associated with acoustic fingerprint identification of devices are described. In one embodiment, a method includes generating a target acoustic fingerprint from acoustic output of a target device. A similarity metric is generated that quantifies similarity of the target acoustic fingerprint to a reference acoustic fingerprint of a reference device. The similarity metric is compared to a threshold. In response to a first comparison result of the comparing of the similarity metric to the threshold, the target device is indicated to match the reference device. In response to a second comparison result of the comparing of the similarity metric to the threshold, it is indicated that the target device does not match the reference device.

Abstract: Systems, methods, and other embodiments associated with determining a quantity of exemplar vectors to select from available training vectors are described. In one embodiment, a method includes determining an available quantity of training vectors that are available in a set of time series signals. A boost function is automatically selected from a plurality of different boost functions based on the available quantity of the training vectors. A selection quantity of the exemplar vectors to select from the training vectors is generated by applying the selected boost function to the training vectors. A quantity of the exemplar vectors is selected from the training vectors based on the selection quantity. A machine learning model is trained to detect an anomaly in the time series signals based on the exemplar vectors that were selected.

Abstract: Systems, methods, and other embodiments associated with automatic clustering of signals including added ambient signals are described. In one embodiment, a method includes receiving time series signals (TSSs) associated with a plurality of machines (or components or other signal sources). The TSSs are unlabeled as to which of the machines the TSSs are associated with. The TSSs are automatically separated into a plurality of clusters corresponding to the plurality of the machines. A group of ambient TSSs is identified that overlaps more than one of the clusters. The group of the ambient TSSs is added into the one cluster of the clusters that corresponds to the one machine. A machine learning model is then trained to detect an anomaly based on the one cluster to generate a trained machine learning model that is specific to the one machine without using the TSSs not included in the one cluster.

Abstract: Systems, methods, and other embodiments associated with a merged-surface 3D fingerprint technique for improved prognostics for assets are described. In one embodiment, a method includes generating a set of time series signals from sensor readings of a reference device while the reference device is operated through multiple individual iterations of an exercise profile. The reference device operates in a known undegraded state. The method then separates the set of time series signals into segments that correspond to the individual iterations of the exercise profile. The method then aligns and merges the segments to generate a merged reference fingerprint. The method then trains a machine learning model to detect anomalous departures from the known undegraded state based on the merged reference fingerprint.

Abstract: Systems, methods, and other embodiments associated with inverse-density exemplar selection for improved multivariate anomaly detection are described. In one embodiment, a method includes determining magnitudes of vectors from a set of time series readings collected from a plurality of sensors. And, the example method includes selecting exemplar vectors from the set of time series readings to train a machine learning model to detect anomalies. The exemplar vectors are selected by repetitively (i) increasing a first density of extreme vectors that are within tails of a distribution of amplitudes for the time series readings based on the magnitudes of vectors, and (ii) decreasing a second density of non-extreme vectors that are within a head of the distribution based on the magnitudes of vectors. The repetition continues until the machine learning model generates residuals within a threshold in order to reduce false or missed detection of the extreme vectors as anomalous.

Abstract: Systems and methods are described that estimate a remaining useful life (RUL) of an electronic device. Time-series signals gathered from sensors in the electronic device are received. Statistical changes are detected in the set of time-series signals that are deemed as anomalous signal patterns. Anomaly alarms are generated, wherein an anomaly alarm is generated for each of the anomalous signal patterns. An irrelevance filter is applied to the set of anomaly alarms to produce filtered anomaly alarms, wherein the irrelevance filter removes anomaly alarms associated with anomalous signal patterns that are not correlated with previous failures of similar electronic devices. A notification may be generated indicating that the electronic device has a limited remaining useful life based on at least the anomalous signal patterns associated with the filtered anomaly alarms.

Abstract: The disclosed embodiments provide a system that performs seasonality-compensated prognostic-surveillance operations for an asset. During operation, the system obtains time-series sensor signals gathered from sensors in the asset during operation of the asset. Next, the system identifies seasonality modes in the time-series sensor signals. The system then determines frequencies and phase angles for the identified seasonality modes. Next, the system uses the determined frequencies and phase angles to filter out the seasonality modes from the time-series sensor signals to produce seasonality-compensated time-series sensor signals. The system then applies an inferential model to the seasonality-compensated time-series sensor signals to detect incipient anomalies that arise during operation of the asset. Finally, when an incipient anomaly is detected, the system generates a notification regarding the anomaly.

Abstract: We disclose a system that executes an inferential model in VRAM that is embedded in a set of graphics-processing units (GPUs). The system obtains execution parameters for the inferential model specifying: a number of signals, a number of training vectors, a number of observations and a desired data precision. It also obtains one or more formulae for computing memory usage for the inferential model based on the execution parameters. Next, the system uses the one or more formulae and the execution parameters to compute an estimated memory footprint for the inferential model. The system uses the estimated memory footprint to determine a required number of GPUs to execute the inferential model, and generates code for executing the inferential model in parallel while efficiently using available memory in the required number of GPUs. Finally, the system uses the generated code to execute the inferential model in the set of GPUs.

Abstract: Systems, methods, and other embodiments associated with detecting feedback control instability in computer thermal controls are described herein. In one embodiment, a method includes executing a workload on the computing system, wherein the workload varies between a minimum and a maximum at a workload frequency. The method includes recording thermal telemetry from the computing system during execution of the workload. The method includes converting the recorded thermal telemetry into a frequency domain. The method includes detecting whether thermal control of the computing system exhibits feedback control instability based on dissimilarity in the frequency domain between the transformed thermal telemetry and the workload frequency. And, the method includes generating an electronic alert that indicates whether the thermal control of the computing device exhibits the feedback control instability.

Abstract: Systems, methods, and other embodiments associated with auditing the results of a machine learning model are described. In one embodiment, a method accesses original time series data and machine learning estimates of the original time series data. The method generates reconstituted time series data from the machine learning estimates by reversing operations of a machine learning model trained for generating the machine learning estimates from the original time series data. The method detects tampering (or corruption) in the original time series data based on a difference between the original time series data and reconstituted time series data. And, the method generates an electronic verification report that indicates whether the tampering (or corruption) is detected in the original time series data.

Abstract: Systems, methods, and other embodiments associated with detecting unfairness in machine learning outcomes are described. In one embodiment, a method includes generating outcomes for transactions with a machine learning tool to be tested for bias. Then, actual values for a test subset of the outcomes that is associated with a test value for a demographic classification are compared with estimated values for the test subset of outcomes. The estimated values are generated by a machine learning model that is trained with a reference subset of the outcomes that are associated with a reference value for the demographic classification. The method then detects whether the machine learning tool is biased or unbiased based on dissimilarity between the actual values and the estimated values for the test subset of the outcomes. The method then generates an electronic alert that the ML tool is biased or unbiased.

Abstract: Systems, methods, and other embodiments associated with generating a stream of ML estimates from a stream of observations in real-time using a circular double buffer are described. In an example method, observations are received from the stream of observations. The observations are loaded in real time into a circular buffer. The circular buffer includes a first buffer and a second buffer that are configured together in a circular configuration. Estimates of what the observations are expected to be are generated by a machine learning model from the observations that are in the circular buffer. The generation of estimates alternates between generating the estimates from observations in the first buffer in parallel with loading the second buffer, and generating the estimates from observations in the second buffer in parallel with loading the first buffer. The estimates are written to the stream of estimates in real time upon generation.

Abstract: A double-blind comparison is performed between prognostic-surveillance systems, which are located on a local system and a remote system. During operation, the local system inserts random faults into a dataset to produce a locally seeded dataset, wherein the random faults are inserted into random signals at random times with variable fault signatures. Next, the local system exchanges the locally seeded dataset with a remote system, and in return receives a remotely seeded dataset, which was produced by the remote system by inserting different random faults into the same dataset. Next, the local system uses a local prognostic-surveillance system to analyze the remotely seeded dataset to produce locally detected faults. Finally, the local system determines a performance of the local prognostic-surveillance system by comparing the locally detected faults against actual faults in the remotely seeded fault information. The remote system similarly determines a performance of a remote prognostic-surveillance system.

Abstract: Systems and methods are described that estimates a remaining useful life (RUL) of an electronic device. Time-series signals gathered from sensors in the electronic device are received. Statistical changes are detected in the set of time-series signals that are deemed as anomalous signal patterns. Anomaly alarms are generated, wherein an anomaly alarm is generated for each of the anomalous signal patterns. An irrelevance filter is applied to the set of anomaly alarms to produce filtered anomaly alarms, wherein the irrelevance filter removes anomaly alarms associated with anomalous signal patterns that are not correlated with previous failures of similar electronic devices. A logistic-regression model is used to compute an RUL-based risk index for the electronic device based on the filtered anomaly alarms. When the risk index exceeds a risk-index threshold, a notification is generated indicating that the electronic device has a limited remaining useful life.

Abstract: Systems, methods, and other embodiments associated with frequency-domain resampling of time series are described. An example method includes generating a power spectrum for a first time series signal that is sampled inconsistently with a target sampling rate. Prominent frequencies are selected from the power spectrum. Sets of first phase factors that map the prominent frequencies to a frequency domain at first time points are generated. Coefficients are identified that relate the sets of first phase factors to values of the first time series signal at the first time points. Sets of second phase factors that map the prominent frequencies to a frequency domain at second time points are generated. A second time series signal that is resampled at the target sampling rate is generated by generating new values at the second time points from the coefficients and sets of second phase factors.