Evolving Systems (2018) 9:119–133 DOI 10.1007/s12530-017-9206-8 ORIGINAL PAPER Evolving ANN‑based sensors for a context‑aware cyber physical system of an offshore gas turbine 1 1 2 Farzan Majdani · Andrei Petrovski · Daniel Doolan Received: 27 December 2016 / Accepted: 17 October 2017 / Published online: 27 October 2017 © The Author(s) 2017. This article is an open access publication Abstract An adaptive multi-tiered framework, that can be applied across a wide range of problem domains requiring utilised for designing a context-aware cyber physical system processing, analysis and interpretation of data obtained from to carry out smart data acquisition and processing, while heterogeneous resources. minimising the amount of necessary human intervention is proposed and applied. The proposed framework is applied Keywords Smart condition monitoring · Context within the domain of offshore asset integrity assurance. awareness · Cyber physical system · Asset integrity · The suggested approach segregates processing of the input Artificial neural network stream into three distinct phases of Processing, Prediction and Anomaly detection. The Processing phase minimises the data volume and processing cost by analysing only inputs 1 Introduction from easily obtainable sources using context identification techniques for finding anomalies in the acquired data. Dur - There exists a growing demand for smart condition monitor- ing the Prediction phase, future values of each of the gas tur- ing in engineering applications often achieved through evo- bine’s sensors are estimated using a linear regression model. lution of the sensors used. This is especially true when some The final step of the process— Anomaly Detection—clas- constraints are present that cannot be satisfied by human sifies the significant discrepancies between the observed intervention with regard to decision making speed in life and predicted values to identify potential anomalies in the threatening situations (e.g. automatic collision systems, operation of the cyber physical system under monitoring exploring hazardous environments, processing large vol- and control. The evolving component of the framework is umes of data). Because computer-assisted instrumentation based on an Artificial Neural Network with error backpropa- is capable of processing large amounts of heterogeneous data gation. Adaptability is achieved through the combined use much faster and is not subject to the same level of fatigue as of machine learning and computational intelligence tech- humans, the use of machine-based condition monitoring in niques. The proposed framework has the generality to be many practical situations is preferable. Cyber physical systems (CPSes) integrate information processing, computation, sensing and networking, which * Farzan Majdani allows physical entities to operate various processes in firstname.lastname@example.org dynamic environments (Lee 2008). Many of these intelli- Andrei Petrovski gent CPSes carry out smart data acquisition and processing email@example.com that minimise the amount of necessary human intervention. Daniel Doolan In particular, a considerable research interest lies in the area Daniel.Doolan@bcu.ac.uk of managing huge volumes of alerts that may or may not 1 correspond to incidents taken place within CPSes (Pierazzi School of Computing Science and Digital Media, Robert Gordon University, Aberdeen, UK et al. 2016). 2 The integration of multiple data sources into a uni- School of Computing and Digital Technology, Birmingham City University, Birmingham, UK fied system leads to data heterogeneity, often resulting Vol.:(0123456789) 1 3 120 Evolving Systems (2018) 9:119–133 into difficulty, or even infeasibility, of human process- 2 Cyber physical systems ing, especially in real-time environments. For example, in real-time automated process control, information about Rapid advances in miniaturisation, speed, power and mobil- a possible failure is more useful before the failure takes ity have led to the pervasive use of networking and informa- place so that prevention and damage control can be car- tion technologies across all economic sectors. These tech- ried out in order to either completely avoid the failure, or nologies are increasingly combined with elements of the at least alleviate its consequences. physical worlds (e.g. machines, devices) to create smart or Computational Intelligence (CI) techniques have been intelligent systems that offer increased effectiveness, produc- successfully applied to problems involving the automation tivity, safety and speed (Lee 2008). Cyber physical systems of anomaly detection in the process of condition monitor- (CPS) are a new type of system that integrates computa- ing (Khan et al. 2014). These techniques however require tion with physical processes. They are similar to embedded training data to provide reliable and reasonably accurate systems but focus more on controlling the physical entities specification of the context in which a CPS operates. The rather than processes embedded computers monitor and con- context enables the system to highlight potential anoma- trol, usually with feedback loops, where physical processes lies in the data so that intelligent and autonomous control affect computations and vice versa. Components of cyber of the underlying process can be carried out. physical system (e.g. controllers, sensors and actuators) Anomalies are defined as incidences or occurrences, transmit the information to cyber space through sensing a under a given circumstances or a set of assumptions, that real world environment; also they reflect policy of cyber are different from the expected outcome (for instance space back to the real world (Park et al. 2012). when generator rotor speed of the gas turbine goes below Rather than dealing with standalone devices, cyber 3000 rpm). By their nature, these incidences are rare and physical systems are designed as a network of interacting often not known in advance. This makes it difficult for the elements with physical inputs and outputs, similar to the Computational Intelligence techniques to form an appro- concepts found in robotics and sensor networks. The main priate training dataset. Moreover, dynamic problem envi- challenge in developing a CPS is to create an interactive ronments can further aggravate the lack of training data interface between the physical and cyber worlds; the role of by occurrence of intermittent anomalies. this interface is to acquire the context information from the Computational Intelligence techniques that are used physical world and to implement context-aware computing to tackle dynamic problems should therefore be able to in the cyber world (Lun and Cheng 2011). Figure 1 illus- adapt to situational/contextual changes. A multi-tiered trates a conceptual framework for building context-aware framework for CPSes with heterogeneous input sources cyber physical systems (Rattadilok et al. 2013), adapted is proposed in the paper that can deal with unseen anoma- from a widely used modern sensor system reference model lies in a real-time dynamic problem environment. The standardised by the CENSIS Innovation Centre for Sensor goal is to develop a framework that is as generic, adaptive and Imaging Systems (www.sensorsystems.org.uk). The and autonomous as possible. In order to achieve this goal component parts and function of this reference model need both machine learning and computational intelligence to be delineated by function and interface in order to effec- techniques are applied within the framework, together tively develop effective instrumentation system in particular with the online learning capability that allows for adap- and cyber physical system in general. tive problem solving. Each layer of the framework is dedicated to a certain con- The application of the CI techniques to provide evolv- text processing task, ranging from low-level context acquisi- ing functionality of the intelligent sensors deployed tion up to high level context application using either existing within cyber physical systems is the first novel contribu- or acquired knowledge. In particular, the context acquisition tion of the presented work. The second contribution is layer corresponds to the exploration of the available sensory the implementation of the generic framework to make data, including their visual representation, identification of the CPSes context-aware by processing a large amount of the appropriate sampling periods, and data transformation heterogeneous data. Finally, the application of these novel (for example, differencing) for further analysis. The context approaches to developing evolving sensory systems for processing layer deals with pre-processing of measured sig- optimising the operation of an offshore gas turbine consti- nals (e.g. identification of outliers, signal validation, etc.) tutes another original contribution of the paper that dem- and with detection of their salient features (e.g. the presence onstrates practical benefits of the suggested methodology. of outliers). The main function of the second layer is to make necessary preparations for building data-driven models with good generalisation capabilities. Of particular interest to the authors are the models based on computational intel- ligence techniques artificial neural networks, support vector 1 3 Evolving Systems (2018) 9:119–133 121 Fig. 1 Framework for designing context-aware CPS machines, etc., built and tuned with the help of genetic algo- rithms, particle swarm optimization and artificial immune systems. The remaining layers of the proposed framework operate at a higher abstraction level. The third context selec- tion layer is responsible for building, evaluating, and correct- ing (if necessary) the data-driven models based on empirical data supplied by the lower layers. The final context acquisi- tion layer purports to examine the outputs of the models built at the previous layer in order to obtain or refine knowledge about the principles or rules that govern the dynamics of the processes under investigation (Petrovski et al. 2013). Of a particular interest in the context of the present work is the data acquisition and processing layers that in context- aware CPSes are often implemented on the basis of intel- ligent and evolving sensors. Figure 2 illustrates a possible structure of evolving CPS sensors, wherein the adaptation or evolution of the sensors is done through building a data- driven process model (typically implemented in the con- text selection layer of the framework) and its tuning using machine learning techniques (Rattadilok et al. 2013). Thus, referring back to Fig. 1, the context processing and selection layers of the CPS framework are merged together to form Fig. 2 Structure of an evolving sensor evolving sensors within the CPS under investigation. Cyber physical systems may consist of many intercon- nected parts that must instantaneously exchange, parse and An optimal balance needs to be attained between data act upon heterogeneous data in a coordinated way. This cre- availability and its quality in order to effectively control the ates two major challenges when designing cyber physical underlying physical processes. Figure 3 illustrates a system- systems: the amount of data available from various data atic approach to handling the challenges related to context sources that should be processed at any given time and the processing, which has been successfully applied by the choice of process controls in response to the information authors to various real world applications (Petrovski et al. obtained. 2013; Rattadilok et al. 2013). 1 3 122 Evolving Systems (2018) 9:119–133 Fig. 3 Systematic approach to context processing As can be seen from Fig. 3, the suggested approach seg- intelligence techniques and expert systems have been regates processing of the input stream into three distinct successfully applied to tackling many anomaly detection phases. The Processing phase minimises the volume of data problems, where anomalies are known a priori (El-Abbasy and the data processing cost by analysing only input streams 2014). More interesting, however is to detect previously from easy to obtain data sources using context selection unseen anomalies, especially for real-time control of the techniques for finding anomalies in the acquired data. If any cyber physical systems, which is the focus of the approach anomalies are detected at this stage, Alert 1 gets activated. suggested in the current paper. This phase of the process is used to analyse real-time data Statistical analysis and clustering are examples of tech- and is a safe guard process on scenarios where the frame- niques that are commonly used when the characteristics works prediction fails to highlight an occurrence of unex- of anomalies are unknown (Chandola et al. 2009). Fig- pected changes in the environment. ure 4 illustrates a more detailed process for the systematic In the Prediction phase, future values of each sensor in approach where machine learning and computational intelli- the CPS under investigation (gas turbine’s sensor in our gence techniques are combined to tackle the unknown anom- case) are estimated, using a linear regression model. Moreo- alies and learn from the experience when similar anoma- ver, a new parameter is added which gets populated with the lies occur again. In Fig. 4, a circle labelled “b” represents a “predicted status” value for each data instance, indicating belief function of the output from both the statistical analysis with Alert 2 whether any of the future predicted value of the and computational intelligence nodes, such that sensors goes beyond the set threshold. n m The final step of the process—Anomaly Detection— f (t)= w (X)+ w (X) (1) i i j j classifies the meaning and implications of overall predicted i=1 j=1 future values so that anomalies being present in the under- lying operation process of the cyber physical system are shown. If any anomalies are detected at this stage, Alert 3 The weights (w and w ) of this belief function are adap- i j is triggered. tively adjusted depending on how much knowledge related Such an approach allows for the acquisition of data to the problem context has been obtained. The contribution and/or activation of the necessary physical entities on an of the CI nodes increases with collection of more normal ad-hoc basis, depending on the outcome at each phase. and abnormal data points that can be used for training. This Moreover, the accuracy attained at the specified phases allows the system to run autonomously if required, and any can be enhanced by incorporating additional data sen- potential anomalies are flagged for closer inspection at the sors or additional environmental factors. Computational anomaly classification phase. 1 3 Evolving Systems (2018) 9:119–133 123 Fig. 4 Context processing in a CPS With the use of parallelisation and/or distributed systems, System (CMS) to prevent any possible system failure, with- multiple machine learning, CI techniques and various belief out resorting to the ground truth values rarely available for functions can be evaluated simultaneously with their param- real CPSes, in particular used in the oil and gas industry. eters being adaptively chosen. Anomaly identification using On CMS there are varieties of formulae and thresholds to a combination of such techniques, as described in Fig. 4, has measure and assure safety conditions and efficiency of the been successfully applied to a traffic surveillance applica- turbine. These sensors’ data, although not very important tion (Rattadilok et al. 2013), a smart home environment and as part of the CMS, nevertheless is used for controlling dif- automotive process control (Petrovski et al. 2013), and in ferent divisions of the Turbine and is passed straight into some other applications (Duhaney et al. 2011). HMI. In addition to this, FMMS is also connected to HMI, enabling the SCADA software on the HMI to read all sensor values from Turbine as well as being able to write some val- 3 Experimental results ues into some of actuators. On the HMI there is also another software called OPC Server, which is capable of writing data It is a common practice that most of the sensory data on a into OPC client that then writes data into the historian. The platform are stored in a historian system (e.g. the PI system), proposed Cyber Physical System then reads data from his- which act as a repository of sensory information gathered torian, as shown in the Fig. 5 illustrating the entire process. from one or multiple installation. For this study we used his- torical sensor data of a gas turbine from an offshore instal- 3.2 Data cleaning process and challenges lation in the North Sea. This data in real-time is transmitted offshore via satellite Internet. The integration of smart sen- One of the challenges in exporting data from a historian sors with networked computing clearly indicated the appro- system, such as PI, is the necessity of interpolating values priateness of considering the gas turbine under investigation that are calculated by the PI system during the export pro- as an example of a cyber-physical system, since it utilises the cess and are not real data. Another challenge is that some computing–networking combination. sensors can have an assigned text when the value goes below (or beyond) the admissible range and that text get written 3.1 Data monitoring flow to the PI system instead of an actual number. For example, for some of the sensors during the reboot process the word, Data from most of the Turbine’s sensors goes straight to Configuration get stored in PI instead of a value;another the connected High Frequency Machine Monitoring System example of this is I/O timeout, which gets written into PI (HFMMS). This is due to the high volume of data generated when a connection to a sensor is temporarily lost. Unfortu- every fraction of a seconds, which makes it almost impossi- nately in such scenarios, where the expected value is a num- ble for any other system to handle such data volumes. These ber rather than text, the entire instance needs to be removed sensor values are then passed into Conditional Monitoring since it is problematic for many machine learning algorithms 1 3 124 Evolving Systems (2018) 9:119–133 system is packed with 800 + sensors and to be able to run any study on that we needed to reduce number of attributes to around only 20 sensors. For this study, initially 432 sensors have been identified, which assumed to have direct impact on the performance of the turbine engine. However these factors are mainly selected based on those sensors, which are used as part of calculation in CMS to monitor real-time malfunctions. To identify the significance of each of these contributing sensors, we used a factorial Design subcategory of Design of Experiment using the Minitab statistical software. A factorial design aims at carrying out experiments to identify the effect of several factors simultaneously. To conduct this experiment, instead of varying one element at the time, all the factors change concurrently. The most common approaches for conducting the studies is to run either a Fractional or a Full Factorial Design. One of the known approach to Full Factorial Design Fig. 5 Data monitoring flow is 2-level Full Factorial, when the experimenter assigns only two value of maximum and minimum to each factor. There- fore the number of runs necessary for a 2-Level Full Facto- to combine textual and numeric data in the same input pro- rial Design is 2 k where k is the number of factors. cessing approach. Since Minitab only allows total of 15 factors for each The issues, which highlighted above and any other similar experiment, similar sensors have been grouped into a total issues, make the use of approaches such as time-series very of 29 groups and a separate set of experiments have been ran hard since after deleting the records with missing values on each sensor group. Having a scenario where 15 elements the expected time frame changes. As a result, to be able match the expected pattern is very rare, therefore percentage to work with the data we need to use bigger window size, thresholds have been used as part of the filtering process. which might not be ideal and fail to capture vital informa- Depending on the expected minimum or maximum value for tion. Another challenging issue with the output dataset from each sensor, as part of the full factorial scenarios, thresholds the PI system is selecting the right sensors for the study. The have been added in accordance with the following formulae: reason for that is because not only the oil and gas instal- f (x)=n − ((n − n )× ) (2) max max min lations are packed with many sensors, but in the majority of cases having redundant sensors is very common—hav- f (x)=n + ((n − n )× ) (3) min max min ing many sensors can make selection of the right data input If an instance of the dataset satisfies the scenario, the record channels very difficult. However as well as being a chal- with the performance rate of the engine gets stored into a file lenge, these redundant sensors might work to an advantage for further analysis. This means for each scenario multiple where by comparing the value of the main and redundant instances satisfy the requirement. After going through all sensors, it becomes possible to validate sensor inputs before the elements, a new cut down version of the dataset gets writing them into the data historian. formed. Then once more application goes through all the Notwithstanding, care should be taken while doing input scenarios, one by one, and if the scenario expect more mini- validation, because sometimes the value Doubtful gets writ- mum values than maximum values the least sensor value of ten to the PI system, indicating a potentially broken sensor. all the instances get selected and vice versa for the maxi- In addition to this, it might also happen from time to time mum value. Moreover, if the expected minimum and maxi- that a sensor due to various reasons temporarily goes offline, mum are equal, then the average performance value of the which in those scenarios results in an Out of Service mes- instances get selected. This process lead to a single perfor- sage written to PI. Therefore removing instances from the mance value for each scenario, which will then gets feed dataset due to various reasons as illustrated above, makes it into Minitab. Generated p-values using Minitab then help to difficult to have a solid dataset. By trial and error, from the identify statistically significant factors. Since each p-value is initial interval of every 1 s we have increased the dataset a probability, it ranges from 0 to 1, and measures the prob- into the interval of 5 s to populate the sensor dataset used ability of obtaining the observed values due to randomness for further investigations. only, therefore the lower the p-value of a parameter is the Another vital challenge in the area of data cleaning is the more significant this parameter appears. If the p-value of a attribute selection. In something like gas turbine the whole 1 3 Evolving Systems (2018) 9:119–133 125 factor is less than 0.05, this means that a factor is statisti- Table 1 Gas turbine sensors cally significant. Sensor description Unit Count This approach has been used on three-month worth of Power turbine rotor speed rpm 2 data from a PI historian, which led to a selection of total 25 Gas generator rotor speed rpm 2 sensors from different parts of the gas turbine out of initial Power turbine exhaust temperature F 6 432 sensors (see Fig. 6). Within this period system expe- None drive end direction mm/s 1 rienced eight failures, which are indicated by blue arrows Drive end vibration × direct um P-P 1 in Fig. 7. The sample data for the 3-month period includes Turbine inlet pressure psia 1 around 217,000 instances. Sensors used from the turbine are Compressor inlet total pressure psia 1 listed in Table 1. Ambient temperature F 1 In addition to all the sensors we also had a turbine status, Axial compressor inlet temperature F 2 which has each of the instances of the dataset labelled as Mineral oil tank temperature F 1 either False, True or I/O timed out. False indicates the tur- Synthetic oil tank temperature F 1 bine failure state, True indicates the engine is running, and OB bearing temperature C 1 I/O Timed out indicates when the engine is getting restarted IB bearing temperature C 1 or communication between the PI historian and offshore is IB thrust bearing temperature C 1 temporarily lost. The importance of having the I/O Timeout OB thrust bearing temperature C 1 state is to prevent the system from sending an alarm when Generator active power Mwatt 1 the system is actually in a state of reboot. Grid voltage V 1 3.3 Processing The Processing phase of the proposed context-aware CPS and Psyllos 2012). Moreover, additional algorithms used to detect anomalies on offshore turbines includes k-Nearest implements a computational intelligence technique [an artifi- cial neural network (ANN)] to classify the input stream. The Neighbour (kNN), Support Vector Machine (SVM), Logistic Regression and C4.5 decision tree (Duhaney et al. 2011). ANN was chosen as many studies have shown that it is the most effective classification model to predict the condition of Based on these studies,seven algorithms have been com- pared to identify the best performing one. These algorithms offshore oil and gas pipelines on varieties of factors, includ- ing corrosion (El-Abbasy 2014; Schlechtingen and Santos are listed in Table 3. Moreover Table 2 lists the most signifi- cant hyperpatameters used for each algorithm. 2011). Also these studies highlighted the effective use of Bayesian and decision tree approaches in condition-based 3.3.1 Evolving process maintenance of offshore wind turbines (Nielsen and Srensen 2011). Random Forest Tree is another algorithm, which is In the processing phase when the input stream is analysed widely used in the field of predictive maintenance in the oil and gas industry to forecast a remote environment condi- and classified, it gets appended to the training dataset.The whole framework is wrapped by a Linux bash file and gets tion, where visual inspection is not sufficient (Topouzelis Fig. 6 Gas turbine process design 1 3 126 Evolving Systems (2018) 9:119–133 Fig. 7 Turbine’s fail scenarios Table 2 Comparison of algorithm performance Table 3 Comparison of algorithm performance Algorithm Hyperparameters Algorithm Accuracy (%) Error (%) Multi-layer perceptron (MLP) Iteration: 5000 Multi-layer perceptron (MLP) neu- 100 0 neural network Hidden layers: 4 ral network Neurons per layer: 24 C4.5 decision tree (%) 94.74 5.26 C4.5 decision tree (%) Confidence factor: 0.25 Decision tree random forest 94.73 5.27 Number of folds: 3 k-nearest neighbour (%) 94.07 5.93 Minimum number of objects: 2 Support vector machine (SVM) 87.21 12.79 Number of leaves: 30 Size of the tree: 59 Logistic regression (%) 46.5 53.5 Decision tree random forest Minimum number of records per Nave Bayes (%) 40.45 59.55 node: 10 Number of threads: 4 Quality measure: Gini Index Number of leaves: 29 primary and secondary. The two machines run side by side. Size of the tree: 58 The secondary virtual machine runs the cycle with 2.5 k-nearest neighbour (%) Number of neighbours to consider min lag which provide enough time for the primary virtual (k): 3 machine restart the cycle with the updated training dataset Support vector machine (SVM) Overlapping penalty: 1.7 before itself restart the cycle. the process create a continues Kernel:polynomial monitoring system which every 5 min evolves and retrain the Power: 1.3 Bias: 0.7 model with the updated dataset without downtime. Gamma: 0.3 As it is illustrated in Table 3, Multi-Layer Perceptron Logistic regression (%) – (MLP) Neural Network generates the best result amongst Nave Bayes (%) Default probability: 0.004 other algorithms. All the results listed are the best results Maximum number of unique for each of the algorithms considered, obtained by adjust- nominal values per attribute: 20 ing their hyper-parameters to achieve the best performance using the Auto-Weka package for comparing CI techniques. executed using a timer every 5 min. To prevent downtime Therefore to implement the Processing phase of the sug- while the framework cycle gets restarted with the updated gested framework, a Multilayer Perceptron is used, which is training dataset there are two parallel virtual machines called a feedforward Artificial Neural Network (ANN). Funahashi 1 3 Evolving Systems (2018) 9:119–133 127 (1989), Hornik et al. (1989) and Qin et al. (2016) have all 3.4 Prediction shown that only one hidden layer can effectively generate highly accurate results and to improve the processing time. The second stage of the proposed model is the Prediction Therefore initially an ANN Multilayer Perceptron with Phase. The purpose of this phase is to predict the future Backpropagation of error with one hidden layer has been values for the next 24 h of all 25 sensors. During this phase used. However, in addition to that the chosen algorithm has three-month historical data has been used to train a linear been been trained with 1, 2, 3 and 4 hidden layers and ten- regression model for each sensors. In addition to that, the fold cross validation. The experiments had been carried out thresholds for each of the sensors, provided from currently up until four hidden layers, which eventually generated an installed Conditional Monitoring System have been used to excellent result. Table 4 lists the results obtained from the set threshold alarms. After training the models the developed experiments with 1–4 hidden layers. anomaly detection framework was put into practice for each Although by using only one hidden layer we have man- sensors times series with the lag period of 24 h for each aged to classify 92.77% of the instances correctly, by sensor to predict the next 24-h datasets. Therefore, if any increasing the number of hidden layers to 4, all test instances of the predicted values for each of the sensors fall below or could be correctly classified. beyond the allowed threshold interval, then Alarm 2 gets Figure 8 illustrated an artificial neural network design. activated. Figure 9 illustrates the predicted results for all the The input layer corresponds to the 25 input sensors of the 25 sensors chosen. gas turbine. The middle layers are used to form the rela- tions between the neurons, their number being determined 3.5 Anomaly detection at runtime. The output neurons are the three classifications which indicates the status of the turbine. Since the combination of all the sensors together reflects the status of the turbine, after predicting future sensor values, all the predicted values get merged into a single test dataset. Table 4 ANN multilayer perceptron optimisation A Multi-Layer Perceptron (MLP) Neural Network model, which has been selected as the best performing algorithm as Layers count One Two Three Four part of the Processing phase, was used again for labelling the Correctly classified (%) 92.77 92.77 94.95 100 status of the turbine for each of the instances. After predict- Incorrectly classified (%) 7.23 7.23 5.05 0 ing the status of the turbine for all instances of the dataset, Kappa statistic 0.60 0.60 0.74 1 the developed framework iterates through all labels and, if Mean absolute error 0.09 0.09 0.062 0 any of the instances are labeled as failed, Alarm 3 gets trig- Root mean squared error 0.21 0.21 0.17 0 gered. The system then picks the timestamp of the predicted Relative absolute error (%) 57.32 57.79 39.10 0.34 time and deducts it from the current time to provide the Root relative squared error (%) 74.89 74.97 62.71 0.77 estimated hours left until the system failure. In the final step Coverage of cases (0.95 level) (%) 100 100 100 100 of the Anomaly Detection phase, the total remaining hours Mean rel. region size (0.95 level) 4.65 64.65 55.25 33.33 gets included into an automatically generated email and is Fig. 8 ANN multilayer percep- tron proposed model 1 3 128 Evolving Systems (2018) 9:119–133 Fig. 9 Predicted sensor values sent out to the preset list of email addresses, as well as play- framework using Knime (www.knime.org/). Knime is ing an audio alarm on the PC. an open source data analytics, reporting and integration platform. Although there are other alternatives, such as 3.6 Overall automated process Weka’s KnowledgeFlow and Microsoft Azure’s Machine Learning, Knime was chosen since it has the capability of Initially Weka (www.cs.waikato.ac.nz/ml/weka/) was used importing most of Weka’s features through the addition of to run each of the phases separately. However, in the final a plugin. Also being able to run java snippets and write the stage of the process we have actually formed the proposed developed model into disk to free up space on memory, it 1 3 Evolving Systems (2018) 9:119–133 129 Fig. 9 (continued) was considered to be a preferred option in comparison to Network Multilayer Perceptron (MLP) using backpropaga- Azure’s Machine Learning. tion of error (Pal and Mitra 1992). The dataset was divided into two sets of training and After training the model, it was tested against the devel- test data as illustrated in Fig. 10. Two-month worth of data oped ANN MLP to classify the status of the engine. This was used for training, which included eight cases of tur- implementation covered the Processing phase of the pro- bine failure, with the remainder set aside for testing. The posed Cyber Physical System. This was followed by intro- training dataset has been used to form an Artificial Neural ducing times series lag and a linear regression model to 1 3 130 Evolving Systems (2018) 9:119–133 Fig. 10 Overall automated framework of the process predict the next 24 h on the test dataset. By looking at the of four months into four separate datasets. Then the test for eight failure situations, corresponding thresholds were iden- each month was run individually by removing 5-day worth tified for each of the input sensors based on the pre-labeled of data from each dataset. This led to developing a model dataset generated by the CMS. Therefore, if during the pre- used to predict each of eliminated days on an hourly basis. diction stage any of the sensor’s value go below or above To achieve this, the operational performance of the turbine the set threshold, the second Alarm goes off. However, this for the next 1, 3, 6, 9, 12, 14, 16, 18 and 24 h on each days alarm is an amber alarm, which does not necessarily imply has been predicted. Then the average performance across that the turbine will fail. With 24 h of predicted data for the these 5 days was estimated, using twenty experiment that sensor data gathered in the final stage of the process, all the have been averaged, as illustrated in Fig. 11. predicted data is put together as a test dataset and is tested The average performance shows that within the first 12 against the model developed in the Processing Phase. If the h the proposed framework could predict the status of the status of the engine gets classified as False, then the third turbine with nearly 99% accuracy, which is a very high per- and last alarm gets activated. formance. Even for the 15 h period, prediction was around 84.28%, where other studies (Topouzelis and Psyllos 2012) 3.7 Evaluation support that predictive accuracy over 84% by having 25 features (sensors) or above is considered to be of high per- To test the accuracy and performance of the proposed model, formance. Also, Naseri and Barabady (2016) showed the we have divided the available dataset which covers a total waiting downtime associated with each item of corrective Fig. 11 Hourly performance evaluation 1 3 Evolving Systems (2018) 9:119–133 131 maintenance for gas turbine is considered to be about 7210 of Exhaust temperature is increasing, so is the expected Gen- h. Therefore having performance of even 73% after 16 h can erator Active Power. very ee ff ctively reduce the downtime by 20%. However, after To visualise the accuracy of the framework in the Anom- 18 h the prediction performance shows a sudden decline and, aly Detection phase (Fig. 14), the correlation between Rotor when it gets to prediction of the next 24 h, the result is really Speed, Exhaust Temperature, Generator active power and poor by being around 58%. Table 5 lists the average value of the framework’s prediction are shown. As expected, in the the results for each prediction. scenario illustrated in the Figure, although all three input Figure 12 shows the correlation of the 4 factors of Rotor values showing an increase in expected correlation, but speed, Exhaust temperature, generated active power and the all the performances are below the expected rate and, as a prediction in processing phase of the framework. As the result, the Turbine state is identified as failure, which is the speed of rotor increases, this results in a rise of exhaust tem- expected result. perature, which leads to higher generated power. The figure illustrates the prediction is clearly matching the scenario where the rotor speed and exhaust temperature is low, system 4 Conclusion is generating low power or it is in the fail state. As illustrated in Fig. 13, the prediction phase of the An implementation of a context-aware cyber physical system framework where future values of the sensors are predicted using evolving inferential sensors for condition monitoring and correlation of the values matching what is expected to predict the status of a gas turbine on an offshore instal- where when Rotor speed is increasing, the predicted values lation has been successfully developed. In this research, a three-phase approach has been proposed: In the process- ing phase, historical data of 25 sensors was collected from Table 5 Comparison of real-time status vs. predicted status different areas of turbine to train an evolving component Hours Accuracy (%) Error (%) (ANN-based) used as the basis of the prediction model. In the second phase, future value of each physical sensor were 1 100 0 predicted for a certain period of time using linear regression. 3 100 0 The final phase makes use of the model developed in phase 6 100 0 one to label the predicted data in order to detect anomalies 9 100 0 prior to their occurrence. 12 98.716 1.284 The developed evolving sensor proved to be capable of 14 84.287 15.713 highly accurate predictions of gas turbine status up to 15 h in 16 73.539 26.461 advance with the accuracy of about 84.28%. The clear chal- 18 65.221 34.779 lenge in these sort of problem is dealing with imbalanced 24 58.545 41.455 data and taking advantage of a time-series algorithm, such as Fig. 12 Processing output 1 3 132 Evolving Systems (2018) 9:119–133 Fig. 13 Prediction output Fig. 14 Anomaly detection output time series prediction with Feed-Forward Neural Networks, Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://crea- to help improving the length of a predication period. Fur- tivecommons.org/licenses/by/4.0/), which permits unrestricted use, ther research will focus on normalising the imbalanced data distribution, and reproduction in any medium, provided you give appro- using approaches, such as ensemble methods that include priate credit to the original author(s) and the source, provide a link to bagging and boosting, as well as extending the prediction the Creative Commons license, and indicate if changes were made. time frame by assuring high accuracy in anomaly identifica- tion through exploring various combinations of computa- References tional intelligence techniques with conventional classifica- tion approaches. Chandola V, Banerjee A, Kumar V (2009) Anomaly detection: a sur- vey. ACM Comput Surv (CSUR) 41(3):15 1 3 Evolving Systems (2018) 9:119–133 133 Duhaney J, Khoshgoftaar TM, Sloan JC, Alhalabi B, Beau-Jean PP Pal SK, Mitra S (1992) Multilayer perceptron, fuzzy sets, and clas- (2011) A dynamometer for an ocean turbine prototype: reli- sification. Neural Netw IEEE Trans 3(5):683–697 ability through automated monitoring. In: 2011 IEEE 13th inter- Park KJ, Zheng R, Liu X (2012) Cyber-physical systems: milestones national symposium on high-assurance systems engineering and research challenges. Comput Commun 36(1):1–7 (HASE). IEEE Computer Society, Boca Raton, FL, pp 244–251 Petrovski S, Bouchet F, Petrovski A (2013) Data-driven modelling El-Abbasy MS et al (2014) Artificial neural network models for of electromagnetic interferences in motor vehicles using intel- predicting condition of offshore oil and gas pipelines. Autom ligent system approaches. In: Proceedings of IEEE symposium Constr 45:50–65 on innovations in intelligent systems and applications (INISTA), Funahashi K (1989) On the approximate realization of continuous FL, pp 1–7 mappings by neural networks. Neural Netw 2(3):183–192 Pierazzi F, Casolari S, Colajanni M, Marchetti M (2016) Exploratory Hornik K, Stinchcombe M, White H (1989) Multilayer feedfor- security analytics for anomaly detection. Elsevier Comput Secur ward networks are universal approximators. Neural Netw 56:28–49 2(5):359–366 Qin M, Song Y, Akagi F (2016) Application of artificial neural network Khan Z, Shawkat Ali ABM, Riaz Z (eds) (2014) Computational for the prediction of stock market returns: the case of the Japanese intelligence for decision support in cyber-physical systems. stock market. Chaos Solitons Fractals 85:1–7 Springer, Berlin Rattadilok P, Petrovski A, Petrovski S (2013) Anomaly monitoring Lee EA (2008) Cyber physical systems: design challenges. Tech- framework based on intelligent data analysis. Intelligent data engi- nical report no. UCB/EECS-2008-8. University of California, neering and automated learning (IDEAL 2013), vol 8206. Lecture Berkeley notes in computer science, pp 134–141 Lun Y, Cheng L (2011) The research on the model of the context-aware Schlechtingen M, Santos IF (2011) Comparative analysis of neural for reliable sensing and explanation in cyber-physical system. Pro- network and regression based condition monitoring approaches ced Eng 15:1753–1757 for wind turbine fault detection. Mech Syst Signal Process Naseri Masoud, Barabady Javad (2016) An expert-based approach to 25(5):1849–1875 production performance analysis of oil and gas facilities consider- Topouzelis K, Psyllos A (2012) Oil spill feature selection and clas- ing time-independent Arctic operating conditions. Int J Syst Assur sification using decision tree forest on SAR image data. ISPRS J Eng Manag 7(1):99–113 Photogramm Remote Sens 68:135–143 Nielsen JJ, Srensen JD (2011) On risk-based operation and mainte- www.knime.org/. Accessed 25 Nov 2016 nance of offshore wind turbine components. Reliab Eng Syst Saf www.sensorsystems.org.uk. Accessed 17 Oct 2016 96(1):218–229 www.cs.waikato.ac.nz/ml/weka/. Accessed 15 Dec 2016 1 3
Evolving Systems – Springer Journals
Published: Oct 27, 2017
It’s your single place to instantly
discover and read the research
that matters to you.
Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.
All for just $49/month
Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly
Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.
Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.
Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.
All the latest content is available, no embargo periods.
“Hi guys, I cannot tell you how much I love this resource. Incredible. I really believe you've hit the nail on the head with this site in regards to solving the research-purchase issue.”Daniel C.
“Whoa! It’s like Spotify but for academic articles.”@Phil_Robichaud
“I must say, @deepdyve is a fabulous solution to the independent researcher's problem of #access to #information.”@deepthiw
“My last article couldn't be possible without the platform @deepdyve that makes journal papers cheaper.”@JoseServera