application based comparison of statistical versus deep learning approaches to reciprocating compressor valve condition monitoring

This paper presents a vibration-based condition monitoring approach for early assessment of valve wear in an industrial reciprocating compressor. Valve seat wear is a common fault mode that is caused by repeated impact and accelerated by chatter. Seeded faults consistent with valve seat wear are installed on the head-side discharge valves of a Dresser-Rand ESH-1 industrial reciprocating compressor. Due to the cyclostationary nature of these units a time-frequency analysis is employed where targeted crank angle positions can isolate externally mounted, non-invasive, vibration measurements. A region-of-interest (ROI) is then extracted from the time-frequency analysis and used to train a suitably sized convolutional neural network (CNN). The proposed deep learning method is then compared against a similarly trained discriminant classifier using the same ROIs where features are extracted using texture and shape image statistics. Both methods achieve > 90% success with the CNN classification strategy nearing a perfect result.

Degradation detection for internal gear pumps using pressure reduction time maps

In this paper, a novel approach for detecting degradation in internal gear pumps is proposed. In a data-driven approach, pressure reduction time maps (PRTMs) are identified as a useful indicator for degradation detection. A PRTM measures the time for reducing the internal pump pressure from certain levels to any lower level when the pump engine is stopped and the valves are closed. The PRTM can thus be interpreted as an internal leakage indicator of the pump. For simplified evaluation, PRTMs are compressed to a single scalar indicator by computing their volume (PRTMV). When the internal leakage increases due to wear, the pressure in the pump decreases faster (implying a decreased PRTMV). The proposed approach has been developed and tested with data of real internal gear pumps with different operating times. The PRTMV shows a close relation to the operating time of the pump. Moreover, we compare PRTMV with the commonly used and well known approach of observing pressure holding speed (PHS). Especially for medium degradation, PRTMV shows better sensitivity then PHS.

Data Management Framework & UAV Simulation Testbed for the Study of System-level Prognostics Technologies

Prognostics-enabled technologies have emerged over the last few years, primarily for Condition Based Maintenance (CBM+) applications, which are used for maintenance and operational scheduling. However, due to the challenges that arise from real-world systems and safety concerns, they have not been adopted for operational decision making based on system end of life estimates. It is typically cost-prohibitive or highly unsafe to run a system to complete failure and, therefore, engineers turn to simulation studies for analyzing system performance. Prognostics research has matured to a point where we can start putting pieces together to be deployed on real systems, but this reveals new problems. First, a lack of standardization exists within this body of research that hinders our ability to compose various technologies or study their joint interactions when used together. The second hindrance lies in data management and creates hurdles when trying to reproduce results for validation or use the data as input to machine learning algorithms. We propose an end-to-end object-oriented data management framework & simulation testbed that can be used for a wide variety of applications. In this paper, we describe the requirements, design, and implementation of the framework and provide a detailed case study involving a stochastic data collection experiment.

An End-to-End Learning Framework for Battery Capacity-Fade Trajectory Prediction Using Early Life Data

Predicting the capacity-fade trajectory of a lithium-ion (Li-ion) battery cell is a critical task given its broad utility throughout the battery product life cycle. Even more useful is estimating a battery cell’s capacity-fade trajectory when this cell has not exhibited any noticeable capacity degradation. Accurately predicting the entire capacity-fade trajectory using early life data enables more efficient cell design, operation, maintenance, and evaluation for second-life use. To accomplish this challenging task, we propose an end-to-end learning framework combining empirical capacity fade models and data-driven machine learning models, in which the two types of models are closely coupled. First, we evaluate the accuracy of a library of relevant empirical models which have been shown to model the observed capacity fade of Li-ion cells with reasonable accuracy. After selecting a model, we formulate an end-to-end learning problem that simultaneously fits the chosen empirical model to estimate the capacity fade curve and trains a machine learning model to estimate the best-fit parameters of the empirical model. By solving this end-to-end learning problem, rather than sequentially executing the separate tasks of fitting the capacity fade model and training the machine learning model, we achieve a more optimal solution which is shown to better balance these two objectives. Our proposed end-to-end learning framework is evaluated using a publicly available battery dataset consisting of 124 lithium-iron-phosphate/graphite cells charged with various fast-charging protocols. This dataset was split into training, primary test, and secondary test datasets. Our method performs on par with existing early prediction methods in terms of cycle life prediction, attaining root-mean-square errors of 84 cycles and 169 cycles for primary and secondary test datasets, respectively. In addition to the cycle life prediction, our method possesses a unique ability to predict the entire capacity-fade trajectory.

Remaining Useful Life Estimation from Event Data

Abstract—Prognostics aims to predict the degradation of equipment by estimating their remaining useful life (RUL) and/or the failure probability within a specific time horizon. The high demand of equipment prognostics in the industry have propelled researchers to develop robust and efficient prognostics techniques. Among data driven techniques for prognostics, machine learning and deep learning (DL) based techniques, particularly Recurrent Neural Networks (RNNs) have gained significant attention due to their ability of effectively representing the degradation progress by employing dynamic temporal behaviors. RNNs are well known for handling sequential data, especially continuous time series sequential data where the data follows certain pattern. Such data is usually obtained from sensors attached to the equipment. However, in many scenarios sensor data is not readily available and often very tedious to acquire. Conversely, event data is more common and can easily be obtained from the error logs saved by the equipment and transmitted to a backend for further processing. Nevertheless, performing prognostics using event data is substantially more difficult than that of the sensor data due to the unique nature of event data. Though event data is sequential, it differs from other seminal sequential data such as time series and natural language in the following manner, i) unlike time series data, events may appear at any time, i.e., the appearance of events lacks periodicity; ii) unlike natural languages, event data do not follow any specific linguistic rule. Additionally, there may be a significant variability in the event types appearing within the same sequence. Therefore, this paper proposes an RUL estimation framework to effectively handle the intricate and novel event data. The proposed framework takes discrete events generated by an equipment (e.g., type, time, etc.) as input, and generates for each new event an estimate of the remaining operating cycles in the life of a given component. To evaluate the efficacy of our proposed method, we conduct extensive experiments using benchmark datasets such as the CMAPSS data after converting the time-series data in these datasets to sequential event data. The event data conversion is carried out by careful exploration and application of appropriate transformation techniques to the time series. To the best of our knowledge this is the first time such event-based RUL estimation problem is introduced to the community. Furthermore, we propose several deep learning and machine learning based solution for the event-based RUL estimation problem. Our results suggest that the deep learning models, 1D-CNN, LSTM, and multi-head attention show similar RMSE, MAE and Score performance. Foreseeably, the XGBoost model achieve lower performance compared to the deep learning models since the XGBoost model fails to capture ordering information from the sequence of events.

The Future of PHM Could be Tiny under Cloud: Exploring Potential Application Patterns of TinyML in PHM Scenarios

Deep learning has shown impressive performance acrosshealth management and prognostics applications. Nowadays, an emerging trend of machine learning deployment on resource constraint hardware devices like micro-controllers(MCU) has aroused much attention. Given the distributed andresource constraint nature of many PHM applications, using tiny machine learning models close to data source sensors for on-device inferences would be beneficial to save both time andadditional hardware resources. Even though there has beenpast works that bring TinyML on MCUs for some PHM ap-plications, they are mainly targeting single data source usage without higher-level data incorporation with cloud computing.We study the impact of potential cooperation patterns betweenTinyML on edge and more powerful computation resources oncloud and how this would make an impact on the application patterns in data-driven prognostics. We introduce potential ap-plications where sensor readings are utilized for system health status prediction including status classification and remaining useful life regression. We find that MCUs and cloud com-puting can be adaptive to different kinds of machine learning models and combined in flexible ways for diverse requirement.Our work also shows limitations of current MCU-based deep learning in data-driven prognostics And we hope our work can

Data-Driven Diagnostics and Prognostics for Modelling the State of Health of Maritime Battery Systems – a Review

Battery systems are becoming an increasingly attractive alternative for powering ocean going ships, and the number of fully electric or hybrid ships relying on battery power for propulsion and maneuvering is growing. In order to ensure the safety of such electric ships, it is of paramount importance to monitor the available energy that can be stored in the batteries, and classification societies typically require that the state of health of the batteries can be verified by independent tests – annual capacity tests. However, this paper discusses data-driven diagnostics for state of health modelling for maritime battery systems based on operational sensor data collected from the batteries as an alternative approach. Thus, this paper presents a comprehensive review of different data-driven approaches to state of health modelling, and aims at giving an overview of current state of the art. Furthermore, the various methods for data-driven diagnostics are categorized in a few overall approaches with quite different properties and requirements with respect to data for training and from the operational phase. More than 300 papers have been reviewed, most of which are referred to in this paper. Moreover, some reflections and discussions on what types of approaches can be suitable for modelling and independent verification of state of health for maritime battery systems are presented.

Evaluating and Optimizing Analytic Signals

Component failures in complex systems are often expensive. The loss of operation time is compounded by the costs of emergency repairs, excess labor, and compensation to aggrieved customers. Prognostic health management presents a viable option when the failure onset is observable and the mitigation plan actionable. As data-driven approaches become more favorable, success has been measured in many ways, from the basic outcomes, i.e. costs justify the prognostic, to the more nuanced detection tests. Prognostic models, likewise, run the gamut from purely physics-based to statistically inferred. Preserving some physics has merit as that is the source of justification for removing a fully functioning component. However, the method for evaluating competing strategies and optimizing for performance has been inconsistent. One common approach relies on the binary classifier construct, which compares two prediction states (alert or no alert) with two actual states (failure or no failure). A model alert is a positive; true positives are followed by actual failures and false positives are not. False negatives are when failures occur without any alert, and true negatives complete the table, indicating no alert and no failure. Derivatives of the binary classifier include concepts like precision, i.e. the ratio of alerts which are true positives, and recall, the ratio of events which are preceded by an alert. Both precision and recall are useful in determining whether an alert can be trusted (precision) or how many failures it can catch (recall). Other analyses recognize the fact that the underlying sensor signal is continuous, so the alerts will change along with the threshold. For instance, a threshold that is more extreme will result in fewer alerts and therefore more precision at the cost of some recall. These types of tradeoff studies have produced the receiver operating characteristic (ROC) curve. A few ambiguities persist when we apply the binary classifier construct to continuous signals. First, there is no time axis. When does an alert transition from prescriptive to low-value or nuisance? Second, there is no consideration of the nascent information contained in the underlying continuous signal. Instead, it is reduced to alerts via a discriminate threshold. Fundamentally, prognostic health management is the detection of precursors. Failures which can be prognosticated are necessarily a result of wear-out modes. Whether the wear out is detectable and trackable is a system observability issue. Observability in signals is a concept rooted in signal processing and controls. A system is considered observable if the internal state of the system can be estimated using only the sensor information. In a prognostic application, sensor signals intended to detect wear will also contain some amount of noise. This case, noise is anything that is not the wear-out mode. It encompasses everything from random variations of the signal, to situations where the detection is intermittent or inconsistent. Hence, processing the raw sensor signal to maximize the wear-out precursors and minimize noise will provide an overall benefit to the detection before thresholds are applied. The proposed solution is a filter tuned to maximize detection of the wear-out mode. The evaluation of the filter is crucial, because that is also the evaluation of the entire prognostic. The problem statement transforms from a binary classifier to a discrete event detection using a continuous signal. Now, we can incorporate the time dimension and require a minimum lead time between a prognostic alert and the event. Filter evaluation is fundamentally performance evaluation for the prognostic detection. First, we aggregate the filtered values in a prescribed lead interval n samples before each event. Each lead trace is averaged so that there is one characteristic averaged behavior before an event. In this characteristic trace, we can consider the value at some critical actionable time, tac, before the event, after which there is insufficient time to act on the alert. The filtered signal value at this critical time should be anomalous, i.e. it should be far from its mean value. Further, the filtered value in the interval preceding tac should transition from near-average to anomalous. Both the signal value at tac­ as well as the filtered signal behavior up to that point present independent evaluation metrics. These frame the prognostic detection problem as it should be stated, as a continuous signal detecting a discrete event, rather than a binary classifier. A strong anomaly in the signal that precedes events on an aggregated basis is the alternate performance metric. If only a subset of events show an anomaly, that means the detection failure mode is unique to those events, and the performance can be evaluated accordingly. Thresholding is the final step, once the detection is optimized. The threshold need not be ambiguous at this step. The aggregated trace will indicate clearly which threshold will provide the most value.

Electric Power Steering Power Circuit Health Assessment and Mitigation Strategy

With recent developments of energy efficient design and control for electric motors, electrical subsystems and components have become integral parts of main actuators in vehicle systems (e.g., steering and propulsion systems). To ensure proper vehicle operations, it is important to make sure that electrical power is properly transmitted through the power circuit from vehicle power source to the electric motor. However, degradation in the power circuit health, which often manifests itself as increased resistance, may affect power transmission and degrade the system performance. For example, in Electric Power Steering (EPS) systems, if the EPS power circuit resistance is increased and the EPS is drawing power to assist the driver, voltage at the EPS module will drop significantly, causing the EPS to reset and, consequently, Loss of Assist (LOA) incidents. Due to compliance in the steering system and suspension design, drivers often feel that the steering system is fighting back when an LOA incident occurs. While previous work has partially addressed this issue by developing algorithms that estimate resistance increase in EPS power circuits, this paper further validates and refines the algorithms for vehicle on-board and off-board implementations using test drive data collected. Since on-board and off-board implementations impose different limits on signal sampling rates, a total of 250 and 465 minutes of data are respectively collected with various vehicle speeds and steering maneuvers. Moreover, a supervisory control solution, referred to as EPS Anti-Loss-of-Assist (ALOA), is proposed that gradually and proactively reduces EPS torque assist as resistance in the EPS power circuit increases so that the EPS voltage is kept above a resetting threshold. Stationary steering tests of the proposed solution as well as demonstrations on parking lot maneuvers at General Motors Milford Proving Grounds are conducted. The stationary steering tests and demonstrations show that, with the proposed supervisory control, negative effects of increased EPS power circuit resistance can be mitigated without noticeable changes in normal driving experience.

Machinery Faults Prediction Using Ensemble Tree Classifiers: Bagging or Boosting?

One of the main goals of predictive maintenance is to accurately classify the temporal trends as early as possible, detect faulty states and pinpoint the root cause of the fault. Undoubtedly, neither late nor early maintenance is desirable and incurs additional operating costs; however, early identification of faulty trends and scheduling on-time maintenance is crucial to smooth machinery operation. Various data-driven techniques have been used to identify faults; nevertheless, many of these techniques fail to perform when faced with missing values at run time or lack any explanation on the root cause of the fault. The present work offers a comprehensive study on different techniques used for fault type classification and compares their performance in identifying the mode of operation for the PHME21 dataset. We also evaluate the robustness of such classifiers against missing values. This study shows that tree-based techniques are best suited to perform root cause analysis for each faulty state and establish rules for faulty conditions.