Advanced Pattern Recognition Innovations

---

TNP has a suite of advanced pattern recognition innovations that preprocess and detect anomalies in any/all types of time series signals relating to the health of dynamically executing components, subsystems, and integrated hardware-software systems. Sensor time-series signals from manufacturing assets provide quantitative metrics associated with physical variables, performance variables, energy efficiency, and various quality-of-service (QOS) metrics. These signals are processed in real time using an advanced pattern-recognition technique, called AI-MSET, for predictive and prescriptive maintenance in all types of manufacturing facilities. TNP’s suite of AI-MSET innovations achieves extremely high prognostic sensitivity for detection of the incipience or onset of subtle anomalies in all types of manufacturing assets and processes, but with ultra-low false-alarm and missed-alarm probabilities, to help manufacturing customers increase production-asset reliability margins and system availability goals while reducing costly downtime from spurious false alarms in the customer's critical assets, and achieving significantly reduced scrap rates.

TNP's suite of AI-MSET innovations brings significant advantages over conventional machine monitoring and machine learning (ML) approaches for real-time surveillance of business-critical manufacturing assets. These advantages include:

• The ability to proactively catch very subtle incipient disturbances, even when the disturbance signature is a tiny fraction of the inherent variance in the monitored metrics
• Ultra-low False-alarm and Missed-alarm probabilities (FAPs and MAPs)
• Separately specifiable FAPs and MAPs 
• Real-time signal validation and sensor operability validation. Note: Most false and missed alarms in prognostic health management (PHM) surveillance of business-critical and even safety-critical systems are due to sensor degradation events. AI-MSET has the proven capability to continuously disambiguate between anomalies in the sensors, vs anomalies in the assets the sensors are supposed to protect.
• Low compute cost for large-scale prognostic monitoring applications, i.e., lots of sensors and/or high sampling rates. (Three orders of magnitude lower compute cost versus competitive Neural Networks for “Big Data” surveillance of time-series signals).
• Remaining Useful Life (RUL) estimation with quantitative confidence factors 
• Highly accurate “inferential variable” capability, i.e., one doesn't have to shut down a $1M asset because a $2 internal sensor failed.  AI-MSET can swap in a highly accurate inferential variable, so the sensor fix/replacement can be postponed to a scheduled maintenance window.

By extending the prognostic surveillance envelope to include customer's production assets, programmable logic controllers (PLCs), power supplies, motor-operated valves, and interconnecting networks, TNP’s AI-MSET solutions can help the customer achieve higher availability with lower Operation and Maintenance (O&M) costs for manufacturing operations.

This brief white paper gives an overview of the key elements embodied in AI-MSET-related portfolio of innovations. 

Many industrial processes have embedded diagnostic systems and online statistical process control techniques that perform real-time analysis of process variables. Most of these systems employ simple tests (e.g., threshold, mean value + three-sigma, SPC control-chart thresholds, etc.) that are sensitive only to gross changes in the process mean, or to high step changes or spikes that exceed some threshold-limit test, to determine whether or not a failure has occurred or a process is drifting out of control. These conventional methods suffer from either high false-alarm rates (if thresholds are set too close) or high missed alarm rates (if the thresholds are set too wide).

For high-throughput industrial facilities, false alarms are very costly in terms of plant or physical-asset down time. Missed alarms can be even more costly when incipient problems are not identified and expensive assets fail catastrophically.

AI-MSET provides a superior surveillance tool because it is sensitive not only to disturbances in signal mean, but also to very subtle changes in the statistical moments of the monitored signals and the patterns of correlation between/among multiple types of signals. AI-MSET employs a statistical pattern recognition technique called the Sequential Probability Ratio Test (SPRT), which provides the basis for detecting very subtle statistical anomalies in noisy process signals at the earliest mathematically possible time, thereby providing actionable warning-alert information on the type and the exact time of onset of the disturbance. Instead of simple threshold limits that trigger faults when a signal increases beyond some threshold value, the SPRT technique is based on user-specified false-alarm and missed-alarm probabilities, allowing the end user to control the likelihood of missed detection or false alarm. For sudden, gross failures of sensors or system components the SPRT annunciates the disturbance as fast as a conventional threshold limit check. However, for slow degradation that evolves over a long time period, such as gradual decalibration bias or drift in a sensor, the SPRT raises a warning of the incipience or onset of the disturbance long before it would be apparent to any conventional threshold-based rules.

TNP’s AI-MSET comes from the class of mathematics called nonlinear, nonparametric (NLNP) regression that was originally developed by Argonne National Laboratory (ANL) in the 1990’s for high-sensitivity proactive fault monitoring applications in commercial nuclear power applications. The original ANL AI technique was called MSET (for the Multivariate State Estimation Technique). MSET was spun off to a variety of mission-critical and safety-critical industries. TNP was the first company to develop MSET-based prognostic tools for enterprise computing health-monitoring applications. TNP now has a portfolio of MSET and SPRT enabled innovations that perform Intelligent Data Preprocessing, are tuned automatically, and provide high-sensitivity prognostics with ultra-low false-alarm probabilities. This end-to-end prognostic framework, which solves sensor, signal, and asset prognostic challenges that no other ML techniques can solve (documented herein), is called AI-MSET.

The AI-MSET dataflow structure is similar to that of most time-series regression ML approaches, consisting of a training phase and a monitoring phase (Figure 1 below). The training procedure is used to characterize the monitored assets and processes using historical, fault-free operating data covering the envelope of possible operating regimes for the system variables under surveillance. This training procedure evaluates the available training data and automatically selects a subset of the data observations (using a similarity operator) that are determined to best characterize the monitored asset's normal operation. It creates a stored model of the equipment that is used in the monitoring procedure to estimate the expected values of the signals under surveillance. In the monitoring step, new observations for all the asset signals are first acquired. These observations are then used in conjunction with the previously trained AI-MSET model to estimate the expected values of the signals. AI-MSET  estimates are extremely accurate, with error rates that are typically less than 1 percent of the standard deviation of the input signal. [Important note for Manufacturing Prognostics: Conventional threshold-alarm monitoring cannot catch the onset of anomalies when the degradation is smaller than the noise band. TNP’s AI-MSET proactively catches the onset of subtle anomalies even when the anomalies are “below the noise floor”.]

The end-to-end processing steps taken during the AI-MSET surveillance phase are shown in Figure 1.

Figure 1: AI-MSET surveillance-phase block diagram

The difference between a signal's real-time AI-MSET estimate and its directly sensed value is termed a residual. The residuals for each monitored signal are used as an anomaly indicator for sensor and equipment faults. Instead of using simple thresholds to detect fault indications, AI-MSET's fault detection procedure employs a SPRT (sequential probability ratio test) to determine whether the residual error value is uncharacteristic of the learned process model and thereby indicative of a sensor or equipment fault. The SPRT algorithm is a significant improvement over conventional threshold detection processes in that it provides more definitive information about signal validity with a quantitative confidence factor through the use of statistical hypothesis testing. This approach allows the user to specify false alarm and missed alarm probabilities, allowing end-customer control over the likelihood of false alarms or missed detection. 

In addition to detecting signal, process, and system degradation, AI-MSET estimates also are used to continuously validate sensor operability. Detecting sensor inoperability is important for two reasons: (1) if the sensor is protecting the system from extreme operating conditions (e.g. high temperatures, overvoltage events, abnormal vibrational level), then a malfunctioning sensor effectively incapacitates the intended protection mechanism, and (2) if the sensor is part of a feedback-control loop, then the system may potentially enter an unstable state because of faulty input provided by the malfunctioning sensor.

Redundant sensors can ensure sensor operability at the expense of duplicated sensor hardware, but for industrial and computer applications where redundant sensors would be prohibitively costly, AI-MSET can be used to provide analytical redundancy for the sensors. The AI-MSET estimate for a signal at time t leverages the redundant information content read at t from multiple, separate (but correlated) sensors to provide a highly accurate “inferential sensor”.

Although the primary objective of applying AI-MSET is to achieve proactive fault monitoring, the fact that AI-MSET provides continuous sensor operability validation is a side benefit that is extremely valuable for asset monitoring. A typical industrial manufacturing plant or compute farm can contain thousands of physical sensors. It is an unfortunate fact that these physical sensors often have a shorter mean-time-between-failure (MTBF) than the expensive assets they are supposed to protect.

 TNP's AI-MSET prognostic solutions have the unique capability to disambiguate between sensor disturbances versus disturbances in the assets being monitored. It is no longer necessary to shut down a $1M asset and discover that the cause was a $2 sensor drifting out of calibration.

One common failure mode for sensors is known as "stuck-at" failures (meaning the transducer retains its last mean value but is no longer responding to changes in the sensed variable). When this occurs in an industrial asset, then the asset is vulnerable to serious degradation if there is a thermal, mechanical, or electrical event or excursion. Even more likely, however, is that as the sensor eventually drifts out of calibration, which may cause an asset to be unnecessarily be shut down from a "false alarm" event. TNP's prognostic system surveillance continuously monitors the telemetry signals associated with physical sensors and actuates a warning flag when it detects the incipience or onset of sensor degradation or sensor decalibration events. AI-MSET2 has unique benefits for equipment surveillance because it unambiguously distinguishes between sensor degradation and component/process degradation.

It is important to note that conventional threshold-limit-based industrial asset monitoring techniques will never catch stuck-at faults in sensors. AI-MSET not only detects all types of sensor degradation events (including stuck-at faults), but can then swap in a highly accurate inferential sensor. The inferential sensor can be used indefinitely, or until a scheduled maintenance window, when the identified degraded sensor(s) can be replaced (or recalibrated) with no loss of business-critical uptime for the assets. 

Figure 2: AI-MSET sensor failure detection

Although MSET and SPRT are very powerful algorithms, like all ML techniques they require accurate training data to be useful. There can be many problems with time series data such as missing values, uneven sampling rates, quantization errors, and so forth. TNP has a suite of data pre-processing tools that fix these problems with the data, thus allowing MSET to provide robust anomaly detection.

(1) Correcting for unevenly sampled signals - Analytical Resampling Process (ARP)

Because the various sensor data streams may originate from differing sampling rates, this important pre-processing step uses interpolation-based upsampling and downsampling methods to generate uniform sampling intervals for all telemetry time series. Moreover, it is very common for the various asset clocks, control network clocks, and environmental variable-monitoring clocks to be out of sync. Clock mismatch issues will cause almost all machine learning prognostic algorithms to fail. TNP's patented ARP prevents this issue with real-time phase synchronization

Figure 3: ARP--Correcting for unevenly sampled signals

(2) Dequantization of Quantized Sensor Signals

A second challenge with using telemetry signatures in computational machine learning algorithms is quantization, which can severely impact the resolution of the telemetry signals and hence accuracy of the computed results. Quantization occurs from the low resolution A/D chips typically used in industrial equipment transducers. TNP's data pre-processing solution has built-in TNP-patented techniques to "un-quantize" signals in real time, in effect producing high-accuracy output signatures from low-resolution input signals. 

Figure 4: Correcting for signal quantization

(3) Missing Value Imputation (MVI)

AI_MSET’s valuable ability to distinguish between a failed component and a failed sensor was described earlier. The same technology can be used to accurately calculate missing values, which can be useful if those values are being used in other applications.  

TNP’s suite of advanced pattern-recognition innovations for time-series anomaly detection is unmatched in the industry. The MSET2 ecosystem is comprised of a synergistic integration of the AI-MSET and SPRT techniques with a set of data pre-processing innovations. Together, they produce a system with unique surveillance capabilities that surpass any conventional approaches, including neural networks, in sensitivity, reliability, robustness to unreliable and possibly degrading sensors, simplicity of training, adaptability, and computational efficiency.