Today most state-of-the-art manufacturing, mining, farming, and service machines (e.g., elevators) are actually quite “smart” in themselves. Many sophisticated sen- sors and computerized components are capable of delivering data concerning a machine’s status and performance. The problem is that little or no practical use is made of most of this data. We have the devices, but we do not have a continuous and seamless flow of information throughout entire processes. Sometimes this is because the available data is not rendered in a useable, or instantly understandable,
form. More often, no infrastructure exists for delivering the data over a network, or for managing and analyzing the data, even if the devices were networked.
Watchdog Agent®-based real-time remote machinery prognostics and health management (R2M-PHM) system has been recently developed by the IMS Center.
It focuses on developing innovative prognostics algorithms and tools, as well as remote and embedded predictive maintenance technologies to predict and prevent machine failures, as illustrated in Figure 3.2.
Figure 3.2. Key focus and elements of the Intelligent Maintenance Systems
The rest of the section is organized as follows. Section 3.1 deals with the platform of Watchdog Agent®-based real-time remote machinery prognostics and health management (R2M-PHM) system. Section 3.2 presents a generic and scalable prognostic methodology or toolbox, i.e., the Watchdog Agent® toolbox;
and Section 3.3 illustrates the effectiveness and potentials of this new development using several real industry case studies.
3.3.1Watchdog Agent®-based R2M-PHM Platform
A generic and scalable prognostics framework was presented by Su et al. (1999) to integrate with embedded diagnostics to provide “total health management”
capability. A reconfigurable and scalable Watchdog Agent®-based R2M-PHM platform is being developed by the IMS Center, which expands the well known open system architecture for condition-based maintenance (OSA-CBM) standard (Thurston and Lebold 2001) by including real-time remote machinery diagnosis and prognosis systems and embedded Watchdog Agent® technology. As illustrated in Figure 3.3, the Watchdog Agent® (hardware and software) is embedded onto machines to convert multi-sensory data to machine health information. The extracted information is managed and transferred through wireless internet or a satellite communication network, and service is automatically triggered.
Figure 3.3. Illustration of IMS real-time remote machinery diagnosis and prognosis system
3.3.1.1System Architecture
The system architecture of the Watchdog Agent®-based R2M-PHM platform is shown in Figure 3.4. In most products or systems, different sensors measure different aspects of the same physical phenomena. For example, sensor signals, such as vibrations, temperature, pressure, etc. are collected. A “digital doctor”
inspired by biological perceptual systems and machine psychology theory, the Watchdog Agent® consists of embedded computational prognostic algorithms and a software toolbox for predicting degradation of devices and systems. It is being built to be extensible and adaptable to most real-world machine situations. The health related information is saved to the database. The diagnostic and prognostic outputs of the Watchdog Agent®, which is mounted on all the machinery of interest, can then be fed into the decision support tools. Decision support tools help the operation personnel balance and optimize their resources, when one or more machines are likely to fail, by constantly looking ahead. For example, if a production line has three processes A, B and C, such that A has one machine, B has three machines, and C has one machine, what would we do if we could anticipate that one of the machines at station B is not behaving normally. Perhaps we would arrange a staging area for output from A, or perhaps we would ramp up production on the other two machines at station B. Whatever the case, we would be making our decision before experiencing the impending breakdown. These tools are critical to maintenance and process personnel, enabling them to stay ahead of the game, balancing limited resources with constant change in demand. Decision support tools also help minimize losses in productivity caused by downtime, and help production and logistics managers optimize their maintenance schedule to minimize downtime costs. The lean and necessary information for maintenance can then be determined and published to the internet through an embedded web server.
Embedded operating system Embedded software
Watchdog Agent®
toolbox
Web server Database
Embedded computer I/O cards
Sensor signals Vibration Temperature Pressure Current Voltage On/Off …
Remote computer
Client software Decision
support tools
Figure 3.4. System architecture of a reconfigurable Watchdog Agent®
The rapid development of web-enabled and cyber-infrastructure technologies is important in providing enablers for remote monitoring and prognostics. One of the major barriers is that most manufacturers adopt proprietary communication protocols which lead to difficulties in connecting diverse machines and products.
Currently, the IMS Center is developing a web-enabled remote monitoring Device- to-Business (D2B)™ platform for remote monitoring and prognostics of diversified products and systems. A system methodology and infotronics platform has been developed that enables the transformation of product condition data into more a useful health information format for remote and network-enabled prognostics applications. The MIMOSA (maintenance information management open system architecture) organization has adopted the IMS infotronic platform as one of its standard platforms and will use an IMS testbed to demonstrate MIMOSA standards in its future activities. As shown in Figure 3.5, the IMS infotronics platform includes the Watchdog Agent® toolbox (which contains adaptive algorithms for different situations and applications), decision support tools, data storage, and D2BTM (device-to-business) system level connectivity. The Watchdog Agent® tool- box includes signal processing, feature extraction, performance assessment, autonomous learning, prediction and prognostics functions. The lean and necessary information for maintenance from decision support tools can then be determined and sent out through D2BTM system level connectivity to remote workstations or computers.
Figure 3.5. Integrated infotronics platform
3.3.1.2Hardware Requirements
For a certain industry application, the selection of Watchdog Agent® hardware depends on characteristics of the input/output signals (for example, what type of input/output signal and how many channels needed), which tools or algorithms are selected (for example, different algorithms require different hardware computation and storage capacities), and the hardware’s working environment (for example, which decides the hardware’s storage type, temperature range, etc.). The hardware prototype currently used in the IMS Center is based on PC104 architecture, as shown in Figure 3.6a. PC104 architecture enables the hardware to be easily expanded to a multi-board system, which includes multiple CPUs and a large amount of input channels. It has a powerful VIA Eden 400MHz CPU and 128MB
of memory since all of the tools are embedded into the hardware. It has 16 high speed analog input channels to deal with highly dynamic signals. It also has various peripherals that can acquire non-analog sensor signals such as RS- 232/485/432, parallel and USB. The prototype uses a compact flash card for storage, so it can be placed on top of machine tools and is suitable for withstanding vibrations in a working environment. Once a certain set of tools/algorithms is determined for a certain industry application, commercially available hardware, such as Advantech and National Instruments (NI) as illustrated in Figure 3.6b and c, respectively, will be further evaluated for customized Watchdog Agent® applica- tions.
Figure 3.6a–c. Options of hardware prototypes for Watchdog Agent® application
3.3.1.3Software Development
The software system of the Watchdog Agent®-based IMS platform consists of two parts: the embedded side software and the remote side software, as shown in Figure 3.7. The embedded side software is the software running on the Watchdog Agent® hardware, which includes a communication module, a command analysis module, a task module, an algorithm module, a function module, and a DAQ module. The communication module is responsible for communicating with the remote side via TCP/IP protocol. The command analysis module is used to analyze different commands coming from the remote side. The task module includes multi- thread scheduling and management. The algorithm module contains specific watchdog agent tools. The function module has several auxiliary functions such as channel configuration, security configuration, and email list and so on. The DAQ module performs A/D conversion using either interrupt or software trigger to get data from different sensors. The remote side software is the software running on the remote computers. It is implemented by ActiveX control technology and can be used as a component of the Internet Explorer Browser. The remote side software is mainly composed of a communication module and a user interface module. The communication module is used for communicating with the embedded site via TCP/IP protocol. The user interface has a health information display, an ATC status display, and a discrete event display. It also possess an algorithm module, as well as error log database and data format interface.
Figure 3.7. Software structure of Watchdog Agent®
3.3.1.4Remote Monitoring Architecture and Human Machine Interface Standards A four-layer infrastructure for remote monitoring and human machine interface standards is illustrated in Figure 3.8. The data acquisition layer consists of multiple sensors which obtain raw data from the components of a machine or machines in different locations. The Network layer will use either traditional Ethernet connec- tions, or wireless connections for communication between the Watchdog Agent®s, or for sending short messages (SM) to an engineer’s mobile phone via GPRS ser- vices. The Application layer functions as a control server to save related information and control the behavior of the Watchdog Agent®s in the network. The Enterprise layer offers a user-friendly interface for maintenance-related engineers to access information either via an Internet browser or a mobile phone.
Figure 3.8. Illustration of Watchdog Agent®-based remote monitoring architecture
3.3.2Watchdog Agent® Toolbox for Multi-sensor Performance Assessment and Prognostics
The Watchdog Agent® toolbox, with autonomic computing capabilities, is able to convert critical performance degradation data into health features and quantitatively assess their confidence value to predict further trends so that proactive actions can be taken before potential failures occur. Figure 3.9 illustrates one of the developed enabling prognostics tools that can assess and predict the performance degradation of products, machines and complex systems.
Figure 3.9. MS innovation in advanced prognostics
The Watchdog Agent® toolbox enables one to assess and predict quantitatively performance degradation levels of key product components, and to determine the root causes of failure (Casoetto et al. 2003; Djurdjanovic et al. 2000; Lee 1995, 1996), thus making it possible to realize physically closed-loop product life cycle monitoring and management. The Watchdog Agent® consists of embedded computational prognostic algorithms and a software toolbox for predicting de- gradation of devices and systems. Degradation assessment is conducted after the critical properties of a process or machine are identified and measured by sensors. It is expected that the degradation process will alter the sensor readings that are being fed into the Watchdog Agent®, and thus enable it to assess and quantify the degradation by quantitatively describing the corresponding change in sensor signatures. In addition, a model of the process or piece of equipment that is being considered, or available application specific knowledge can be used to aid the degradation process description, provided that such a model and/or such knowledge exist. The prognostic function is realized through trending and statistical modeling of the observed process performance signatures and/or model parameters.
In order to facilitate the use of Watchdog Agent® in a wide variety of applications (with various requirements and limitations regarding the character of signals, available processing power, memory and storage capabilities, limited space, power consumption, the user’s preference etc.) the performance assessment module of the
Watchdog Agent® has been realized in the form of a modular, open architecture toolbox. The toolbox consists of different prognostics tools, including neural network-based, time-series based, wavelet-based and hybrid joint time-frequency methods, etc., for predicting the degradation or performance loss on devices, process, and systems. The open architecture of the toolbox allows one easily to add new solutions to the performance assessment modules as well as to easily interchange different tools, depending on the application needs. To enable rapid deployment, a quality function deployment (QFD) based selection method had been developed to provide a general suggestion to aid in tool selection; this is especially critical for those industry users who have little knowledge about these algorithms. The current tools employed in the signal processing and feature extraction, performance assess- ment, diagnostics and prognostics modules of Watchdog Agent® functionality are summarized in Figure 3.10.
Each of these modules is realized in several different ways to facilitate the use of the Watchdog Agent® in a wide variety of products and applications.
Figure 3.10. Watchdog Agent® prognostics toolbox
3.3.2.1Signal Processing and Feature Extraction Module
The signal processing module transforms multiple sensor signals into domains that are the most informative of a product’s performance. Time-series analysis (Pandit and Wu 1993) or frequency domain analysis (Marple 1987) can be used to process stationary signals (signals with time invariant frequency content), while wavelet (Burrus et al. 1998; Yen and Lin 2000), or joint time-frequency analysis (Cohen 1995; Djurdjanovic et al. 2002) could be used to describe non-stationary signals (signals with time-varying frequency content). Most real life signals, such as speech, music, machine tool vibration, acoustic emission etc. are non-stationary
signals, which place a strong emphasis on the need for development and utilization of non-stationary signal analysis techniques, such as wavelets, or joint time- frequency analysis. The feature extraction module extracts features most relevant to describing a product’s performance. Those features are extracted from the time domain into which the sensory processing module transforms sensory signals, using expert knowledge about the application, or automatic feature selection methods such as roots of the autoregressive time-series model, or time-frequency moments and singular value decomposition.
Currently the following signal processing and feature extraction tools are used in the Watchdog Agent® toolbox:
• The Fourier transformation method has been widely used in de-noising and feature extraction. Noise component in the signal can be distinguished after it is transformed, and feature components can be identified after the removal of noise. However, Fourier transformation is applicable to non- stationary signals only since frequency-band energies for applications are characterized by time-invariant frequency content.
• The autoregressive modeling method calculates frequency peak locations and intensities using autoregressive oscillation modes of sensor readings and bares significant information about the process (usually, mechanical systems are well described by the modes of oscillations).
• The wavelet/wavelet packet decomposition method enables the rapid calculation of non-stationary signal energy distribution at the expense of loosing some of the desirable mathematical properties.
• The time-frequency analysis method provides both temporal and spectral information with good resolution, and is applicable to highly non-stationary signals (e.g. impacts or transient behaviors). However, it is not applicable if a large amount of data has to be considered and calculation speed is a concern.
• The application specific features extraction method is applicable in cases when one can directly extract performance-relevant features out of the time-series of sensor readings.
3.3.2.2Performance Assessment Module
The performance assessment module evaluates the overlap between the most recently observed signatures and those observed during normal product operation.
This overlap is expressed through the so-called confidence value (CV), ranging between zero and one, with higher CVs signifying a high overlap, and hence performance closer to normal (Lee 1995, 1996). In case data associated with some failure mode exist, most recent performance signatures obtained through the signal processing and feature extraction module can be matched against signatures extracted from faulty behavior data as well. The areas of overlap between the most recent behavior and the nominal behavior, as well as the faulty behavior, are continuously transformed into CV over time for evaluating the deviation of the recent behavior from nominal to faulty.
Realization of the performance evaluation module depends on the character of the application and extracted performance signatures. If significant application
expert knowledge exists, simple but rapid performance assessment based on the feature-level fused multi-sensor information can be made using the relative number of activated cells in the neural network, or by using the logistic regression approach. For products with open-control architecture, the match between the current and nominal control inputs and the performance criteria can also be utilized to assess the product’s performance. For more sophisticated applications with intricate and complicated signals and performance signatures, statistical pattern recognition methods, or the feature map based approach can be employed.
The following performance assessment tools are currently being used in the Watchdog Agent® toolbox:
• The logistic regression method allows one to predict a discrete outcome, such as group membership, from a set of variables that may be continuous, discrete, dichotomous, or a mix of any of these. It can quantitatively represent the proximity of current operating conditions to the region of desirable or undesirable behavior. However, it is applicable when a good feature domain description of unacceptable behavior is available.
• The feature map method assesses the overlap between the normal and most recent process behavior, and is applicable in cases when the Gaussianness of extracted features cannot be guaranteed.
• The statistical pattern recognition method calculates overlap of feature distributions based on the assumption of Gaussian distribution of the features, and is applicable to a repeatable and stable process. However, it is not applicable to the highly dynamic systems in which feature distribution cannot be approximated as Gaussian
• The hidden Markov model method is applicable to highly dynamic phe- nomena when a sequence of process observations rather than a single observation is needed to describe adequately the behavior of process signatures.
• The particle filters performance assessment is able to describe quantitatively process performance, and is applicable in cases of complex systems that display multiple regimes of operation (both normal and faulty). In this case a hybrid description of the system is needed, incorporating both discrete and continuous states.
3.3.2.3Diagnostics Module
The diagnostics module tells not only the level of behavior degradation (the extent to which the newly arrived signatures belong to the set of signatures describing normal system behavior), but also how close the system behavior is to any of the previously observed faults (overlap between signatures describing the most recent system behavior with those characterizing each of the previously observed faults).
This matching allows the Watchdog Agent® to recognize and forecast a specific fault behavior, once a high match with the failure associated signatures is assessed for the current process signatures, or forecasted based on the current and past product’s performance. Figure 3.11 illustrates this signature matching process for performance evaluation.
Figure 3.11. Performance evaluation using Confidence Value (CV)
• The support vector machine method establishes a non-linear maximum margin classifier that infers the machine condition from a new set of measurements. It works by using a non-linear kernel to transform the input vector space (which is a set of measurements believed to be correlated with machine condition) to a much higher dimension feature space, and drawing a linear hyper-plane classifier there. It is especially applicable to the situa- tion when Gaussianity of the performance related features cannot be guaranteed and when a process may display multiple normal and faulty modes of behavior (multiple regimes of operation and/or multiple possible faults in the process). The main drawback to using this method is that the choice of a kernel in real applications is usually based on experience or trial-and-error test.
• The hidden Markov model method is especially applicable to a situation in which multiple signals exist and the system may have multiple failure modes. It is applicable to both stationary and non-stationary signals.
• The Bayesian belief network is a compact representation of cause-and-effect for a complex system, and is especially applicable to situations where there are multiple faults with multiple symptoms. The main drawback of this method is that no standard procedure exists to determine network structure and expert knowledge is needed to identify the node state.
• Condition diagnosis based on analytically calculated overlaps of Gaussians that describe the signatures corresponding to the current process behavior and the signatures corresponding to various modes of normal or faulty equipment behavior, is applicable to the cases in which performance related features approximately behave as Gaussians.
3.3.2.4Prediction and Prognostics Module
The prediction and prognostics module is aimed at extrapolating the behavior of process signatures over time and predicting their behavior in the future.
autoregressive moving average (ARMA) (Pandit and Wu 1993) modeling and match matrix (Liu et al. 2004) methods are used to forecast the performance behavior. Currently, autoregressive moving-average (ARMA) modeling and match matrix methods are used to forecast the performance behavior. Over time, as new