19.3.1 Strategy for Anomaly Detection and Fault Isolation

There are many types of anomalies that can be defined in a complex manufacturing system. On the top level, product quality defects can be considered an anomaly caused by production system malfunctions, while at the system level, the occurrence of a critical parameter being out of bounds can be caused by component degradation or failure. In order to promptly detect and diagnose whenever a fault occurs, a causal relationship between failures and system status needs to be correlated.

The system behavior is assumed to be describable in a multivariable feature space, and different statuses will occupy different regimes in this hyper-dimension space. However, as the complexity of modern manufacturing systems increases dramatically, it is impractical to consider the system as a whole object and build a single model to describe its behavior. Thus, a divide-and-conquer strategy to first divide operation space into several sub-regimes must be adopted. The motivation is that it is possible to use linear modeling, rather than a nonlinear approach, within a small range space and reduce the global complexity to the local level. Unsupervised techniques such as a self-organizing network or k-mean cluster are candidates for use in achieving this goal. Furthermore, unsupervised methods require the least amount of understanding of the system and hence are good for complex problems. Within each separate region, supervised learning techniques are then applied for fault detection and diagnosis. Statistical methods such as the Bayesian Belief Network (BBN) or Markov Model are capable tools; distance or kernel-based techniques such as vector quantization or Support Vector Machines can also be applied. Finally, in order to have adaptability, reinforcement learning is used as a feedback mechanism to let the system have a second chance to reevaluate misdetections and update its failure profiles. With the aforementioned model built, it is possible to conduct:

• Fault detection and classification: When new inputs come, it will be projected to the closest operation region. Then, the diagnosis model from this regime will process the input and identify if the machine status is normal or faulty. If it's faulty, the model will further isolate the type of fault.

• Fault diagnosis: Based on the different fault detectors applied, the biggest variable contributor can be identified as the failure cause.

• Anomaly detection: When the input feature cannot be projected into any of the operation regions, similar anomaly conditions will be clustered when no known fault can be matched.

• Model adaptation: Reorganize the knowledge of faults when a new fault condition is confirmed from external inputs, and update the failure profile in the model.

19.3.2 Prognosis and Decision Making with the Self-Recovery Function

Prognosis can be done in several ways. For instance, a regression model can be used to fit the time series of input features, and the feature values predicted over time. Then, the same diagnosis procedure can be applied to evaluate how a system will behave in the future. A statistical approach may be used to check how the feature deviates from the normal regime and if it is moving toward any of the known fault zones. Thus, system performance can be quantified by the confidence value associated with each diagnosis scheme. Figure 19.5 illustrates a prognostics approach based on the feature map and statistical pattern recognition method.

Based on the prognosis results, a decision needs to be made regarding whether maintenance should be scheduled, or if system performance can be recovered by the control compensation approach. Maintenance decision making will be used in conjunction with production scheduling, downtime associated with failure, and the component's remaining useful life, while control compensation will follow heuristic philosophy to recover performance as soon as possible. It should be considered, however, that the component degradation cannot be compensated by utilizing a control strategy. Even if critical parameters can be adjusted to maintain the target value, component degradation will continue to reduce its reliability. Research will be done in this area to find the optimal way to balance system reliability and performance delivery.

19.3.3 An Industrial Agent Platform Based on an Embedded System

A major drawback of deploying an industrial agent platform in a stand-alone system is that the agent will not be capable of directly providing diagnostics and prognostics information to the process control system. It is also challenging to take advantage of the PLC controller processors to synchronize and trigger the external data acquisition. Hence, embedding the industrial agents into the commercial controllers provides the capability of raising flags automatically in the control system or changing the parameters in the control loops accordingly before an undesired event happens or the process gets out of control. Such an embedded agent has the potential to significantly reduce the downtime of the production line and/or avoid high costs imposed by repair and maintenance. On the other hand, the availability of adequate hardware resources, including computational power or memory, remains the major challenge. Figure 19.6 shows the architecture of the embedded PHM deployment platform. In such deployment platforms, KA (shown on the left side of in Figure 19.4) is responsible for loading the PHM models that are already developed by the system. EA (shown at the top of the Figure 19.6) runs the analytics tools on the data and provides the desired health information. SA (shown at the bottom of Figure 19.6) interfaces with the control system to provide feedback based on the obtained asset health information and avoids the progression of the potential faults in the system (Lee et al., 2013a).

19.3.4 Industrial Agent Platforms Based on Cloud Infrastructures

Many industries have recently been investing in PHM in order to improve the performance of their machineries, achieve better consistencies in the manufacturing process, and reduce machine or process downtime and the costs associated with them. In order to achieve such goals, new needs have arisen for more computational resources and more deployment options. Therefore, IT infrastructures have been developed to support the computational and network needs. Cloud computing is considered an important step in the advancement of IT technology and provides significant computational and storage resources that were not accessible before. Cloud computing is being used to provide a more secure and easy option for the storage of large-scale data files and improves the connectivity of services and data sets (Chun and Maniatis, 2009). Also, along with the rapid advancement of technologies around smart phones, tablets, and wireless networks there has been a boom in interactive applications and service delivery methods for use in platforms such as smart phones, tablets, and so on (Huang et al., 2011; Chun and Maniatis, 2009). Like many applications for cloud computing, cloud-based PHM has also received much attention and interest. Cloud-based PHM leverages many existing technologies such as distributed computing and grid computing and aims to integrate them in a form of a service-oriented architecture (SOA) (Zhang et al., 2010). An example of such a cloud-based PHM system is that for the machine tool builders provided in Kao et al. (2014). Three levels have been suggested for this structure: a manufacturing service channel, an in-factory resource management, and an intelligent service platform (Lee et al., 2013a).

A cloud-based monitoring platform requires a high level of flexibility because it targets multiple machines under different operating conditions. Moreover, such PHM workflow needs to be self-adaptive in order to work for different monitoring cases (Lee et al., 2013a).

In Lee et al. (2013a), a framework with three components is proposed for deploying cloud-based PHM systems. These components include (1) Machine Interface Agent, (2) Cloud Application Platform, and (3) Service User Interface. The framework of the deployment is shown in Figure 19.7. In this proposed framework, the Machine Interface Agent is used for communication among the machines in different locations and the cloud. The Machine Interface Agent collects a set of collected data from each monitoring target, and transfers them to the Cloud Application Platform. This Machine Interface Agent can be either embedded, or a stand-alone PC depending on the preferences and constraints of the deployment. Also in such a framework, PHM applications (APPs) can be developed to use the cloud platform as their computational resource. Such APPs can also be shared between different users so that each user can input their data and analyze them using a cloud-based PHM platform. The analytics part in this framework consists of different algorithms stored as separate modules. For different applications, these modules can be selected and put together to generate a customized PHM workflow depending on each application's needs and characteristics, with minimum configuration needed. Such APPs can also be expanded, improved, or customized for new PHM applications. In this platform, users are able to use their devices such as a laptop, smart phone, etc., to log in to the cloud-based platform through the service user interface and access their desired information through visualization tools (Lee et al., 2013a).

In manufacturing facilities, compressed air is a commodity and downtime from the air compressors can cause severe losses in productivity. A surge condition occurs when the compressor is operated at an insufficient flow rate for a given discharge pressure; and these large oscillations in the pressure and flow can create an unstable operating condition for the compressor. For a mild surge case, there is a higher level of vibration and noise for the compressor system; however, if the surge is more severe and the flow reverses, the pressure fluctuations can severely damage the compressor components and cause significant downtime. For this application, it is not only important to have early detection of a surge condition but also to integrate the detection method with the control system to avoid the surge from occurring. A common practice for avoiding a surge is to operate in a conservative manner and operate between 10% and 25% below the surge line. This prevents surge but is less efficient in terms of energy usage. For providing early surge detection and energy efficiency, the proposed method fuses multiple sensor signals and uses pattern classification algorithms to detect and avoid the surge from occurring.

For monitoring the compressor, various sensors were installed. The measured signals included the motor current, inlet guide vane (IGV) signal, inlet and outlet pressure and temperature signals for each stage, the air flow rate, and ambient temperature and humidity measurements. Calculated variables from the measured signals included compressor ratios for each of the three stages and also a pressure differential was calculated across the IGV. In total, there are 20 variables from the measured and calculated quantities. For developing the model, data was collected during a set of surge tests on an air compressor. The compressor was operated at a large flow rate, and then the IGV was slowly closed until a surge occurred. A set of 25 experiments were conducted and data samples right before the surge condition and at the surge condition were collected. It should be noted that for three of the experiments a surge condition was not reached. In total, there are 22 data files for the surge condition, and 25 data files with a nonsurge condition.

From an initial examination of two-dimensional variable plots between the pressure at the third stage and the air flow, it was noted that it was difficult to discriminate between the nonsurge and surge conditions. Thus, a principal component method was used to consider the 20 variables and identify key sensor signals that are most responsible for the surge condition. From the PCA result, the first two eigenvectors were retained and features that had the largest weight were then selected to be included into the health model. The two most significant variables from the PCA analysis were the stage 1 outlet temperature and the ambient air humidity temperature. Based on the facility engineer's experience, it was also decided to include the IGV opening signal into the health model, and this provided three variables to detect the surge condition.

An asymmetric support vector machine (ASVM) classification algorithm is used to automatically determine whether the compressor is in a nonsurge or surge condition. The ASVM algorithm was used, because the impact of a false alarm and a missed detection have different costs associated with them. If the operating point is predicted to be in a surge region and is actually not causing a surge, then changing the operating settings to move out of this region would only cause a slight reduction in energy efficiency. However, failure to detect the surge condition can have a significant impact, in that a surge can severely damage the compressor components and result in major downtime and losses in productivity. Using the ASVM algorithm, the separation boundary is moved closer to one of the two classes (e.g., the no surge class), and this provides a way to minimize the risk of a missed surge detection.

Using 30 samples for training and 15 samples for testing, the traditional support vector machine algorithm resulted in a 100% classification rate. This 100% classification rate was also achieved by the ASVM algorithm. A numerical study was further conducted on the ASVM classification result, in which a set of additional operating points were considered and the distance to the surge boundary and the nonsurge zone boundary were calculated. The additional test operating points highlighted that the ASVM method would not result in a missed detection, but would be slightly more susceptible to false alarms. Again, the risk associated with a false alarm is much lower than the risk and cost associated with a missed detection.

After developing the finalized set of processing methods and algorithms to provide an early detection and avoidance of the compressor surge, selecting the appropriate implementation strategy is needed before deploying the monitoring solution. Because the main objective was to detect and avoid the surge condition, there was not a requirement or need to transfer data and instead the customer's preference was to process all the data locally if possible. Access to the controller signals was imperative, because the health model included three sensor signals, and all three of these signals are from the controller. Also, in order to avoid the surge condition, the health model would have to be integrated into the control system. Unlike vibration-based monitoring applications, which require a high sampling rate, a much lower sampling rate would be sufficient for the temperature and IGV signals used to detect the compressor surge condition. Although the ASVM classification algorithm requires significant computation for determining the separation boundary, this could be done off-line. For performing the real-time monitoring, the already trained classification boundary could be used to provide early detection of the surge condition. Thus, a low complexity and computational requirement is needed for running the compressor surge health monitoring algorithm. The case of multiple operating regimes such as variable rotational speed or transient conditions did not apply to this compressor monitoring situation. In addition, data from both a baseline state and surge state were needed for training the classification algorithm.

For embedding this type of health monitoring algorithm into the compressor control system, KA would only be used once and would load the already trained classification boundary and model parameters for the ASVM algorithm. EA would perform the majority of the calculation work and continuously use the monitored signals and the trained classification algorithm to provide an early detection of the surge condition. Lastly, the SA would interface with the compressor control system to quickly avoid the surge condition once the problem was detected. Developing the appropriate health monitoring algorithm, the right deployment strategy, and a description of how all three agents in the system would interact to perform the calculations and use the computational and memory resources provided a complete solution for the manufacturing facility. In fact, the manufacturing facility worked with the original equipment manufacturer of the air compressors to eventually embed this algorithm and functionality into the compressor control system.

19.4.2 Horizontal Machining Center

As an illustration for the proposed cloud-based PHM framework, a tool condition monitoring (TCM) system is chosen as a case study. Machine tools are used in many manufacturing facilities and have a relatively complex mechanical and controller system. This engineering asset is very likely to experience unplanned breakdowns due to component failures. If the degradation process in the components happens gradually, then a PHM system can be designed and implemented in order to estimate the stage of degradation and the health condition of the component. Once the TCM application was triggered, the data stored in the cloud was parsed. After which, pre-processing techniques such as filtering and quantization were employed to increase the input signal's signal-to-noise ratio (SNR) and to aggregate samples by windowing. Next, a specific segment of the data was selected (duration when the tool is engaged on the machined part). Statistical summary metrics of the “segment-of-interest” were extracted. Finally, distance metrics were used to determine the current (or test) cut's condition relative to the first cut.

In this case study, the test-bed was Milltronics Horizontal Machining Center (HMC) with Fanuc CNC Controller 0i (Model C). The TCM APP required both the controller signals and sensory data. For controller signals, machine status, spindle speed, feed rate, and macro-variables were extracted and transformed based on an MTConnect protocol. In addition, two sensors were installed: a power sensor and an accelerometer. The power signal was extracted through a Universal Power Cell (UPC) from Load Controls Inc., while the vibration was measured using a PCB accelerometer which was powered through a 480D06 amplifier.

The CV value started out high (1 means a healthy status) and as the cutting tool was continuously used, the degradation manifests as an almost monotonic decrease in the health value. Eventually, the tool gets replaced when the CV reaches a value just below 0.5—the average power consumed when a part is machined. Intuitively, the power consumed increases with tool wear.

For this case study, Advantech 4711 was used as the Machine Interface Agent and was responsible for extracting both controller and sensory data, conducting signal preprocessing (data segmentation and MTConnect protocol translation), and transferring the dataset to the Cloud Application Platform. The TCM APP was implemented and deployed on the Cloud Application Platform, and can be used to analyze the incoming dataset. The TCM APP processed the data and provides a tool health estimate. After the related algorithms are integrated with the infrastructure, the analyzed results are continuously written to a file, which the Cloud Application Platform read in order to update the Service User Interface for visualization purposes. More than one machine can be connected and factory users may utilize a shared application with their own dataset on-demand.

In this case study, the SA would include the Machine Interface Agent, which could connect to the CNC system to trigger the data collection. On the other hand, the KA would be located on the Cloud Application Platform, where the history data and learning models are managed. Lastly, the EA would be implemented through the joint effort of the Machine Interface Agent and the Cloud Application Platform. The former would retrieve both controller and sensory data and use Ethernet to transfer the dataset, while the latter would invoke the TCM APP and link it with the KA. These components (or agents) would work together as a whole system to support factory users to take advantage of the developed TCM application.

19.6.1 Smart Factory Transformation From Component Level Point of View

The increasing feeding rate and high acceleration value have improved the machine tool efficiency and accuracy and also reduced the machining times. Each component of machine tools can directly affect the overall precision of the machine tool system. The critical components can be defined based on their criticality level, low failure frequency, and high impact. The condition-based monitoring can enable the manufacturing factories to observe the degradation of each critical component and calculate their health value. Therefore, the overall health condition of the machine will be evaluated based on the health condition of each component. This health indicator, called the confidence value, can provide and predict the information of the machine's health condition. Hence, the predictive analytical algorithms provide significant improvements of adding PHM to traditional maintenance schemes. Machine data is effectively transformed by PHM algorithms into valuable information that can be used by factory managers to optimize production planning, save maintenance costs, and minimize equipment downtime.

In a smart factory, all the physical entities in the factory are connected together physically and virtually. In the physical space, the sensors are utilized to measure and detect the occurrence of a fault or anomaly, and isolate the location of the fault and identify the failure types. Beside the sensor signals, the outputs from the machine controller, such as current, voltage, and power consumption, can also be used to reflect the machines' overall health degradation. These physical entities are connected in cyber space through the Internet. The sensor signals are collected through data acquisition hardware, and either stored locally or uploaded to a cloud server. After storing the data, pre-processing techniques are employed to check the data quality and increase the signals' SNR, such as data filtering, quantization, etc. Data segmentation is used to select the interesting segment from the input signals. Furthermore, the features (health indicators) are extracted, and a health value is calculated over time to evaluate the behavior of the machine components and detect incipient signs of failure even before the actual failure event. In addition to this data-driven method, the physical model and experimental knowledge are also applied to fully understand the performance of these components, such as the finite element analysis (FEA) models, modal analysis, etc. In cyber space, all of these models are integrated to convert the data to information, with which the customers can schedule the timely maintenance before the failure.

In order to model the components' behavior quickly and effectively, the accelerated degradation test in the lab is utilized in the machine reliability analysis, the different stress factors are screened and the important stress factors are picked according to the quantization matrix. After that, the characteristics tests are conducted to identify the initial condition. Thus, the data collected from multiple degradation processes under different stress levels can generate the degradation pattern for the component. Thus, the time conversion between these different degradation patterns will be validated, which can assist in identifying the degradation trend in the actual industrial application. After establishing the degradation model, this PHM model will be integrated into cyber space. The data collected from the machine components will be processed through this PHM model, and thus a virtual component will be modeled in cyber space, so it can simulate and predict the health pattern of the machine component in reality.

The component level of the machine system includes critical components such as motors, bearings, ball screws, gearboxes, etc. These critical components determine the performance of the machine system. Thus, solutions provided for the cyber-physical system of component levels will be:

1. Data acquisition, sensor installation in physical space

2. Establish the virtual model for the critical components using PHM solutions in cyber space

a. Component mechanism analysis and failure detection, for example:

• bearing inner race/outer race failure and ball wear, etc.

• ball screw's preload loss, internal surface wear, etc.

• gearbox's broken teeth, bent shaft, etc.

b. Prognostics and remaining useful life prediction for the components.

• Use the prediction model to estimate the life of each component, and evaluate the life of the whole system

c. Update the virtual model using the new data from the field operation data.

Once the condition of machine components has been analyzed and predicted, machine users and component suppliers can work together to improve uptime and reduce costs by employing an optimal management strategy. Of course, it won't happen without component transparency and remaining useful life predictions.

19.6.2 Smart Factory Transformation from Machine Level Point of View

Machine monitoring systems may put more focus on data acquisition and storage, but there is less emphasis on the analysis of the data. The data contains valuable health information that can be used to support factory decision-making procedures. Machinery prognostics and health management methodologies are able to make the whole process more systemized.

The traditional approach of reactive maintenance—essentially repairing a machine when it fails—may seem like the simplest option, but this approach is clearly inadequate in a modern factory. As throughput times have become increasingly fast due to improvements in plant automation, unexpected breakdowns have become prohibitively expensive and even catastrophic. While more recent preventive maintenance strategies (Qu et al., 2006; Sana, 2012; Schuman and Brent, 2005) may offer higher availability through time-based conditioning/repair/replacement activities that preclude unexpected downtime, this approach also has two major disadvantages. First, preventive maintenance is an expensive program to maintain, especially if its intervals are kept very tight. Second, although preventive maintenance activities ensure that components do not fail or exhibit significant behavioral changes, there is no insight learned about the equipment's actual degradation cycle that can be used to improve its design.

Therefore, beyond what has been discussed regarding component-level analytics, predictive analytics should be extended to a whole machine level. By trending degradation patterns, it can predict, with some level of confidence, when equipment is going to reach a failure condition. With the use of advanced predictive tools and algorithms, manufacturing asset behavior is modeled and tracked using a set of metrics known as “health value” or “confidence value.” Finally, prediction tools are utilized to infer when the machine is likely to fail. With such information, a much higher level of manufacturing transparency is achieved. Maintenance and production personnel can then collaboratively and proactively plan when to schedule repair/conditioning activities to avoid equipment failure so it does not interfere with planned production goals (Figure 19.8).

Besides, from the control perspective, there are two basic categories of machine errors: quasi-static errors and dynamic errors. Machine error compensation does not aim to reduce the absolute value of errors, but the effects of these errors on the machining accuracy and final dimensions of produced parts. The degree to which machining accuracy can be achieved by error compensation is highly dependent on the repeatability of the machine itself and the methods selected to demonstrate the interconnection between different errors. Currently, machine compensations, especially, quasi-state errors is achieved through offline error calibration and modeling, and online error correction. This method is highly dependent on the accuracy of the preset calibration table. Thus, it is labor-intensive and time-consuming. An online error measurement and compensation would alleviate the issues associated with the traditional method. An optical error measurement system installed on the machine would be able to measure the machine errors online and feed these errors back to the machine control module, which would then calculate the compensated values and make online error corrections. This closed-loop method would reduce the workload for the machine error calibration and significantly improve the machining accuracy.

19.6.3 Smart Factory Transformation from Factory Level Point of View

A factory or manufacturing facility can be described as a 5M system, which consists of Materials (properties and functions), Machines (precision and capabilities), Methods (efficiency and productivity), Measurements (sensing and improvement), and Modeling (prediction, optimization, and prevention). To realize a smart factory, the “Big Data” coming from the 5M systems should be further analyzed and integrated in a systematic way.

Big data has raised opportunities to develop “smarter” decision support tools in the factory. Especially, the increasing amount and accuracy of the real-time data from the Factory Information System (FIS) can be utilized to make more effective and dynamic online operational policies. For example, how to change the production sequences if there are some demand changes on the market? What is the bottleneck machine of the system in real time? Which maintenance tasks should be conducted, when, and by whom, so that the overall cost in the system can be minimized? What is the opportunity time window to stop one machine for maintenance while still satisfying the system throughout the requirement? Analyzing the real-time system condition and predicting its evolution can aid factory managers in answering these questions.

For example, by making the manufacturing capability transparent, plant and corporate managers have the right information to assess facility-wide OEE. Also, with the use of such advanced prediction tools, companies can plan more cost-effective, just-in-time maintenance to ensure equipment health over a longer period.

It was also noted that, to make optimal decisions, the information integrated from the component level is necessary, machine level and the system levels. For example, to make an optimal sequence of maintenance tasks, it is important to know: the current health state of the machines and how they will degrade over time; the availability and skill levels of the maintenance crews; the inventory level; and the production requirement of the system, etc. The more accurate this information is, the more effective these decision support tools will be.