### SysWatch Framework

SysWatch is a Windows program written in the Microsoft Visual Studio C# programming language running offline and online. By offline analysis we understand on the one hand the opportunity to visualize data but more importantly the training mode in which SysWatch reads a large, usually high-dimensional sensor time series database from a technical or industrial source, usually representing a physical process of some kind, eventually along with a database of historical events. The main objective in this mode is the establishment of correlations between the sensor readings and the events in question. In the online mode SysWatch uses computed and verified models for arriving data and calculated the learned outcomes.

#### Application Structure

The **Explorer** module is a professional visualization application of operational data. Furthermore it facilitates feeding technical data to the training step. A user-friendly GUI-Interface allows for a hierarchical structuring of all parameters. The **Models** module brings mathematical models into action. Models once generated are stored for analysis purposes. Besides the mathematical calculus a structured visualization and representation of KPIs is the main functionality of this module. The **Dashboard** visualizes the results of a Windows based service on the generated and active models. New data records are processed in the background and stored in MS SQL database.The chart below illustrates the data flow and the processing steps up to and including the visualization of results. The SysWatch framework has four well separated modules as executables. From the user’s perspective there is mainly a separation into the application **SysWatch**, which manages all relevant data handling, the visualization and the services as well as the application **Dashboard**, which is a customer specific visualization implementation of the relevant prognosis models.

#### Core Competence

The core competences of the framework lies in its algorithmic engine, including

- Multivariate abnormality technology
- Multivariate regressions
- Multivariate pattern recognition
- Empirical stochastic models
- High-dimensional response surface generation

Beside the development of mathematical models, FCE focuses on the technical practicability of their mathematical ideas. Following points accomplish a feasible implementation:

##### Flexibility in Data Source

Fault memory Analysis, Statistical Sensor Insights

##### Integration of System Know-How

Hierachical structuring, Structural Dependencies

##### Application of Predictive Algorithms

State of the art Algorithms, Embedding system-specific Features##### Automation

Automated Data Stream, Real-Time Monitoring, Customized Dashboard Conceivable#### Environment

SysWatch has its origin in the predictive maintenance sector and is primary a monitoring software. Nevertheless in the course of time new fields occurred like our response surface analysis being applied in the smart operation field, as well as multivariate time series forecasting for energy demand or weather prognosis.

In the year 2018, SysWatch lives in a highly competitive field, which has to merge many fields of expertise. Here we name some focuses that guarantee our contestability:

##### Supervised Learning Techniques

Development of highly performent and percise supervised learning techniques in the .NET environment.

##### Microsoft Azure Compatibility

SysWatch is a Microsoft Azure compatible software. Data storage interfaces enable the opportunity of developing cooperative software solutions. Recently the necessity to integrate cloud computing concepts has been increasingly noticed and forms part of our current activities.

##### University Cooperations

Several cooperations with universities in the field of applied mathematics increase the competence on the predictive analytics subject

##### PHM Society Membership

In order to maintain high quality of our results, FCE exposes itself on a regular basis to the peer review resulting from the publication process in magazins, journals and conferences.

### Predictive Maintenance

The SysWatch framework is a monitoring solution in the area of predictive maintenance. It is a data driven condition based monitoring system using different methodologies in order to provide reliable early warnings of deteriorating asset conditions. More and more sensors capture the state of modern systems periodically, whereby a complete coverage of stochastic dependencies is a must. Employing the routinely recorded sensor data, it generates an added value by detecting anomalies and predicting critical events in a system in order to avoid expensive, catastrophic machinery malfunctions and reduce the risk of unplanned down-time and costly repairs. The prevention of malfunctions

- Improves security
- Increases availability
- Reduces maintenance costs
- Protects the image of an organization

This prevention results from a delicate balancing procedure, trying to predict as many critical events as early as possible, while at the same time avoiding false alerts. SysWatch consistently handles this challenge as a solution to a complex multidimensional problem.

Like all pattern recognition work, these correlations will enable the program, to be applied to the sensor readings and detect or even predict the events in question.

As simple as this task sounds, as complicated it is. Some of the main challenges are the following:

- Does the entirety of sensor readings have enough predictive power to predict the events ?
- What are the optimal values of the parameters to minimize Type I and Type II errors simultaneously?
- Are there any significant influence variables missing

SysWatch is equipped with a powerful set of data analysis tools to decide the above questions. Data preprocessing steps such as correlation analysis, sensor grouping etc. will be activated before any evaluation is performed. We are convinced that SysWatch can enhance your existing monitoring capabilities and at the same time improve productivity metrics while lowering operating cost.

### SysWatch Smart Operation

Our Response Surface generation methodology is a very powerful data modeling and visualization tool which allows to find high-dimensional non-linear response surfaces in the form whereby u is a vector representing some ambient conditions of a process, x is a vector representing the input- or control variables and y is the vector of output- or response variables. Let the dimensions of x, u and y be n, k and m respectively. Then f maps the (n+k)-dimensional real space onto the m-dimensional real space.

For instance look at a combustion engine emitting u-, x- and y-values at a high sampling rate. Assume these data are collected in a large file or database. This file is used to find the unknown function f by means of parametric or non-parametric regression techniques.

Using a certain graphical “trick” called pseudo projection it becomes possible to represent each component of f as a function of x – ignoring u for the moment – as a 3D structure, although x may have more than two dimensions. The trick consists in selecting two arbitrary components of x and fixing the remaining n-2 variables by setting them equal to the arithmetic mean of their respective input values.

Figure below illustrates this method using an arbitrary set of input data:

The user has two additional degrees of freedom to manipulate this 3D representation:

⦁ First, he/she may select arbitrarily two different input variables

⦁ In principle the n-2 non selected variables can be set on any value between their empirical minimum and their empirical maximum

In the monitoring scenario our efficient algorithm solves a combinatorial optimization problem on the computed response surface and provides optimal operating modes for the customer.

### Power Demand Forecast

This SysWatch feature concentrates on time series forecasts of relevant parameters as for instance power consumption. We focus on a long term forecasting model dealing with main periodicities in addition to short term effects. The consolidation of Fast Fourier Transformations, which covers the basic oscillation of a time series, with machine learning algorithms approximating the error term, is at the heart of our forecasting model. The approximation part uses deifferent nonlinear regressions such as kernel and logistic regression, a combinatorial technique on sparse grids, neural networks and many more. Thereby we are able to forecast power consumption in certain locations of a given network and we show the results of those forecasts as functions of various inputs. The results presented are used for power demand planning of cities and are consequently prognostic in nature. In the context of Health Management, however, one usually works with anomaly detection and supervised learning methods. Nevertheless a time series forecast in neighboring applications, e.g. the power consumption of a traction system in railway vehicles, could substantially benefit from these prognosis functionalities. This also means that deviations of physical quantities measured on real-time conditions from their expected behavior most likely indicate a prevailing malfunction.