METHOD FOR VALIDATION OF AN INVESTIGATED SENSOR AND CORRESPONDING MACHINE The invention relates to a method for validation of an investigated sensor within a sensor group and in particular to a method for validation of an investigated sensor of an industrial turbine. Complex technical systems such as industrial turbines can be outfitted with a plurality of sensors which can generate a high amount of data during testing and/or operation of the respective technical system. A technical system can comprise for instance a gas turbine having a large number of sensors working in a harsh environment. High-quality data is needed to provide a correct reasoning about the status of the gas turbine. However, with hundreds of sensors working in extreme conditions, the turbine can typically experience frequent sensor faults which may go on unnoticed and, thus, distort the true picture of the machine's status or machine

condition. Accordingly, an automatic detection of sensor faults is required for the identification of correct sensor information. Users or rule-based expert systems built on knowledge derived from human experts are able to validate some sensor signals by inspecting by-passing sensor data received from sensors of the respective technical system. However, these validations have deficiencies, because the underlying vague expert knowledge can usually provide only low-dimensional shallow relationships between the parameters of the technical system. The increasing complexity of a measurement system inside a modern complex technical system such as an industrial turbine with a rising number of sensors requires an automatic fault detection which can cope with a huge number of variables and high frequency dynamic sensor data. This object is achieved by a method comprising the features of claim 1. Accordingly, the invention provides according to a first aspect of the present invention, a method for validation of an investigated sensor within a sensor group comprising:

receiving sensor data from sensors of the sensor group measuring the same physical property,

calculating a deviation of the sensor data received from the investigated, sensor within the sensor group from sensor data reconstructed on the basis of sensor data received from all other sensors of the sensor group, and

signalling a sensor fault of the investigated sensor if the calculated deviation is out of a trusted range.

In a possible embodiment of the method according to the first aspect of the present invention, each sensor within the sensor group is investigated in consecutive order.

In a further possible embodiment of the method according to the first aspect of the present invention, the investigated sensor is validated using online and/or offline sensor data.

In a further possible embodiment of the method according to the first aspect of the present invention, the reconstructed sensor data is calculated on the basis of a multiplication matrix M as follows:

wherein M is a NxN multiplication matrix,

N is the number of sensors in the sensor group, i, j are indices and

X is the received sensor data.

In a further possible embodiment of the method according to the first aspect of the present invention, the used NxN multiplication matrix M is calculated as follows: wherein η ₍ m are indices indicating sensors of the

sensor group,

T is a time interval used for sensor data validation and

t is a time index.

In a further possible embodiment of the method according to the first aspect of the present invention, the analysis time interval T used for sensor data validation is adapted

dynamically depending on the expected group sensor data dynamics of the received sensor data. In a further possible embodiment of the method according to the first aspect of the present invention, received new available online sensor data received during a reception time interval is validated. In a still further possible embodiment of the method

according to the first aspect of the present invention, the multiplication matrix M is updated for each reception time interval as follows: M _M =e ^T * -M _k+ (X ^r X) _k wherein T _N is the reception time interval,

T _A is the analysis time interval and

X is the new available online sensor data received during the k-th reception time

interval .

In a still further possible embodiment of the method

according to the first aspect of the present invention, the validated new available online sensor data received during the reception time interval T _N is stored as historical sensor data for a storage time interval T _H .

In a still further possible embodiment of the method

according to the first aspect of the present invention, the reconstructed sensor data Y is calculated on the basis of a linear combination of the sensor data received from all other sensors of the sensor group such that wherein Y is the sensor data received from the

investigated sensor of the sensor group,

i is an index,

T is a predetermined time interval and

t is a time index.

The invention further provides a machine comprising the features of claim 11.

Accordingly, the invention provides according to a second aspect a machine having at least one sensor group consisting of sensors adapted to measure the same physical property, wherein the sensors of the sensor group are monitored by a sensor monitoring unit configured to execute a program comprising instructions performing a method for validation of an investigated sensor within the sensor group comprising receiving sensor data from sensors of the sensor group measuring the same physical property,

calculating a deviation of the sensor data received from the investigated sensor within the sensor group from sensor data reconstructed on the basis of sensor data received from all other sensors of the sensor group, and

signalling a sensor fault of the investigated sensor if the calculated deviation is out of a trusted range. In a possible embodiment of the machine according to the second aspect of the present invention, the sensor monitoring unit is configured to calculate a deviation of sensor data received from an investigated sensor within the sensor group from sensor data reconstructed on the basis of sensor data received from all other sensors of the same sensor group and configured to signal a possible sensor fault of the

investigated sensor if the calculated deviation is out of a trusted range .

In a possible embodiment of the machine according to the second aspect of the present invention, the machine comprises an industrial turbine. In a further possible embodiment of the machine according to the second aspect of the present invention, different

physical properties of the industrial turbine are measured by corresponding sensor groups. In a still further possible embodiment of the machine

according to the second aspect of the present invention, the machine is controlled in response to sensor faults signalled by the sensor monitoring unit. In the following, different aspects of the present invention are explained in more detail with reference to the enclosed figures

Fig 1 shows a flow chart for illustrating a possible

exemplary embodiment of a method for validation of an investigated sensor within a sensor group according to the first aspect of the present invention; Fig 2 shows a diagram for illustrating a possible

embodiment of a method for validation of an

investigated sensor within a sensor group; Fig . 3 shows a diagram for illustrating an alternative approach for signal reconstruction and validation;

Fig . 4 illustrates the processing of data of an online data stream of the first approach for signal reconstruction and validation illustrated in Fig.

2;

Fig. 5 shows a diagram for illustrating the processing of sensor data of the second approach for signal reconstruction and validation as illustrated in Fig. 3.

As can be seen in Fig. 1, the method for validation of an investigated sensor within a sensor group according to the first aspect of the present invention can comprise several steps .

In a first step SI, sensor data from sensors of the sensor group measuring the same physical property such as a

temperature is received for further data processing.

In a second step S2, a deviation of the sensor data received from the investigated sensor within the sensor group from sensor data reconstructed on the basis of sensor data received from all other sensors of the same sensor group is calculated.

In a further step S3, a signal fault of the investigated sensor is signalled if the calculated deviation is out of a trusted range,.

In a possible embodiment, each sensor within a sensor group can be investigated in a consecutive order and the

investigated sensor can be validated using online and/or offline sensor data. In a complex technical system such as turbine, there can be several redundant or well-correlated sensors of a sensor group measuring the same physical property or physical parameter such as a temperature or a pressure. The method according to the first aspect of the present invention can correctly detect sensor faults without confusing them with system dynamics of the investigated technical system. With the method according to the present invention as illustrated in Fig. 1, readings from the group of sensors are monitored dynamically.

In a possible embodiment, the sensor data of each sensor in a sensor group is periodically compared with the sensor data of the remaining sensors in the same sensor group. In a possible embodiment, a linear combination of the sensor data of the remaining sensors in the sensor group is constructed to fit data of the current investigated sensor in a best way. The residual of a fitting procedure is then calculated. If the residual is out of a trusted range, then a sensor error of the investigated sensor is reported.

Within a predetermined time interval, the following procedure can be performed. Each sensor in the group {X _{i=l N} } in

consecutive order is marked as a suspected sensor to be investigated. The group of sensors is divided in two parts or subgroups, i.e. the suspected sensor Y=X _n , wherein n is the index of the investigated suspected sensor, and all other sensors Z = {X _i≠n }. Then, the signal of the suspected sensor Y is reconstructed as a linear combination of the signals or data from the other sensors Z . This reconstructed signal can be labelled as Y and constructed in such a way that a mean square difference between Y and Y becomes minimal:

The reconstructed signal that fits the minimum of equation (1) can be constructed in two different ways. A first approach for suspected signal reconstruction and validation is illustrated in Fig. 2 showing a sensor group SG of a suspected investigated sensor Y and other sensors Z in the same sensor group SG. A singular value decomposition SVD of a matrix of other sensor data Z can provide three

matrices :

U matrix of normalized principal components where columns are orthogonal, size Tx(N-l))

5" matrix of eigenvalues (positive, diagonal,

size (N-l)x(N-l)) and

V rotation matrix to principal components "eigenvectors"

(orthogonal, size (N-l) x (N-l) ) . These matrices U, S, V are connected to matrix Z with equation (2) as follows:

Z ^SVD U,S,V;Z = U-S-V ^T (2) Then, the suspected sensor Y can be projected to orthonormal basis of principal components U as expressed in equation (3) . k = -i/, _,m , rn = l..(N-l) (3) Then, the vector of projections k (size Nxl) is used for reconstruction signal Y as expressed in equation (4) :

Y=U-k (4) Consequently according to equations (2) to (4) , the vector of the reconstructed signal Y is calculated that is closest with the suspected signal Y . With the reconstructed signal

Y and the suspected sensor signal Y an error estimation EE can be formed as illustrated in Fig. 2. If the error is within a trusted range, there is no fault detection whereas if the error is outside the trusted range a sensor fault is detected. Fig. 3 shows an alternative approach for signal

reconstruction and validation. In this embodiment, the sensor reconstruction procedure uses a multiplication matrix M. The reconstructed signal Y is obtained in this embodiment using a multiplication matrix M which is defined by equation (5) and has the matrix dimension NxN.

M = X ^T xX;M _n,m =∑X _i X, _,m (5)

(=1

The reconstructed signal Y for the suspected signal Y=X _n can be constructed as follows: wherein an inversion matrix {M _{i≠n j≠n} ) ^~l indicates a pseudo- inverted matrix. As illustrated in Fig. 3 using a forgetting factor F and the new sensor group data X a new

multiplication matrix M _N E _W is provided using the previous multiplication matrix M ₀ LD iteratively in a loop. The signal and by using the pseudo- inverted matrix Μ ^'λ which has been sorted for other sensors (M _i≠nJ≠n ) on the basis of the sensor data of the remaining sensors of the same sensor group as illustrated in Fig. 3. Sensor 1 to N of a sensor group SG can be

investigated consecutively in a loop as also illustrated in Fig. 3. The reconstructed sensor signal Y is compared with the sensor data Y of the suspected sensor as shown in Fig. 3. If the calculated error or deviation is out of a trusted range,, the sensor is classified as possibly faulty, whereas if the error is small, the sensor is classified to be

healthy. If the sensor data is stored and a method is applied in an offline mode according to the first approach as

illustrated in Fig. 2 or the second approach as illustrated in Fig. 3, then validation results using direct SVD (Fig. 2) or a multiplication matrix M (Fig. 3) are the same. However, if the validation methods illustrated in the embodiments of Figs. 2, 3 are applied for online sensor data, the results can differ. In an offline mode, a reconstruction method as illustrated in Fig. 2 can be applied for all available data sets. But sensor data relations can vary with time and a single vector of reconstruction coefficients k may not fit the whole available time interval. In this case, the data set can be divided in small sets being not longer than a time interval T _L of data linearity. However, the time interval T _L should not be too short. Sensor faults can be divided into fast and slow sensor faults. Fast sensor faults include extreme sensor noise, outliers and abrupt data changes. In contrast, slow sensor faults include sensor signal drift and other disturbances that can be seen only in an enough long observation time interval if applied. An analysis time interval T _A is a time interval for sensor data validation and detection of slowest sensor faults. These time intervals should fulfill T _A < T _L to prevent false alarms.

In case of online sensor data validation, new available sensor data can be validated for each new data time interval T _N . Usually, this new data time interval T _N lasts for seconds or minutes and is smaller than the analysis time interval T _A . Accordingly, a reconstruction method can be applied not only for new not validated data but even also for historical data: T _A = T _H + T _N , wherein T _H is the time interval historical already validated data is temporarily stored. Using the embodiment illustrated in Fig. 2, one can validate several times the same data set as also illustrated in Fig. 4. In Fig. 4, the approach of Fig. 2 is applied for online data processing of an online data stream ODS . The approach

illustrated in Fig. 4 uses direct SVD to validate V data sets comprising historical data H and new available data N in an analysis time interval T _A .

The scheme of Fig. 5 illustrates another alternative approach for online sensor data validation applied in the embodiment illustrated in Fig. 3. In this embodiment, a multiplication matrix M is used. The advantage of the presented embodiment is online validation of sensor data using a multiplication matrix M for data history encapsulation instead of buffering data itself. In a possible embodiment, the multiplication matrix M can be dynamically forgotten with time. This helps to encapsulate data dynamics in a certain analysis of time interval T _A . If new data in a reception time interval T _N becomes available, then a new updated multiplication matrix M _NEW can be calculated from an old multiplication matrix M _0L D with equation (7) : wherein T _N is the reception time interval,

T _A is the analysis time interval and

X is the new available online sensor data received during the reception time interval TN. In a possible embodiment, the analysis time interval T _A can be dynamically adjusted

according to the group data dynamics, i.e. it can be adjusted to be small for fast signal changes and to be big for steady state sensor readings.

With the method according to the present invention, it is possible to apply a principle component analysis for sensor validation in an unsupervised manner. The method can also be applied for any technical system that contains sensors working in a harsh environment . With the method according to the embodiment illustrated in Fig. 3, it is possible to encapsulate historical data using a multiplication matrix M for data validation. The method can be performed by a program executed by a processor of a sensor monitoring unit of a machine or complex technical system. In this embodiment, the sensor monitoring unit is configured to calculate a deviation of sensor data received from an investigated sensor within a sensor group SG from sensor data reconstructed on the basis of sensor data received from all other sensors of a same sensor group SG and further configured to signal a possible sensor fault of the investigated sensor if the calculated deviation is out of a trusted range. The sensor monitoring unit can be integrated in a machine having at least one sensor group SG consisting of sensors adapted to measure the same physical property, wherein the sensors of the sensor group SG are monitored by the sensor monitoring unit. The sensor monitoring unit can also be connected via a data interface to the monitored sensor groups. The machine can be an industrial turbine, for instance a gas turbine, wherein different physical properties of the industrial turbine are measured by corresponding sensor groups monitored by a sensor monitoring unit configured to execute a program comprising instructions performing the method according to the first aspect of the present invention. In a possible embodiment, the machine is controlled in response to sensor faults signalled by the sensor monitoring unit.