The subject of the invention is a method and a computer program for assessing the condition of rotating machinery connected to an electric motor, where the motor and the rotating machinery comprise a drive train electrically powered from a three phase motor and said motor is connected with a computing device.

The work leading to this invention has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement n° PITN- GA-2010-264940.

Background

Today, there is an increasing realization in industry that asset monitoring is key to promoting efficiency of resources. There are mounting calls for more advanced asset management methods. Furthermore, customers are beginning to select individual components on the basis of whether their condition may be monitored. For example, gearboxes often form an important part of rotating machinery drive trains, finding applications in industries such as oil and gas (in the drive trains of compressors or pumps), mining (in driving conveyors) or marine (in propulsion drive trains), amongst others. Any downtime of gearboxes can result in large losses to a customer. Standard methods of monitoring the condition of rotating machinery include vibration analysis using casing-mounted accelerometers and oil debris analysis. In the case of oil debris analysis, the recent trend is to use an online/inline oil debris sensor. This sensor is connected to the lubrication oil loop of the gearbox in order to capture the wear particles that come from the gears. On the other hand there is an emerging trend in scientific literature related to the analysis of electrical signals _. (in particular stator currents) in order to evaluate the health of rotating machinery connected to an electric motor. Statement of problem

Due to spatial restrictions or environmental concerns, it is not always possible to instrument a shaft line component with oil debris or vibration sensors. Furthermore, sensors mounted to the components of a drive train may be considered as 'invasive' in that they are mounted directly on the rotating machinery of the system. It follows that in order to apply condition monitoring approaches based on these sensors the operation of the machine will have to be stopped in order to install the sensors. In such a situation, a non-invasive method of identifying the health of a component would be preferable. Whilst casing mounted sensors such as accelerometers may be considered as noninvasive, they are subject such as transmission path effects. Standard sensors such as accelerometers are typically casing-mounted. Usually, the relevant dynamics of the system occur on the rotating elements of the equipment, from which fault signatures propagate through the structure to the casing-mounted sensor. The propagation path, which will typically differ from installation to installation, can act to modulate and filter some of the dynamics useful for diagnostics. Hence, it can be difficult to properly set alarm threshold levels for a system using standard methods. Analysis of electrical signals potentially offers a solution which is less influenced by such uncertainties, as the electrical signals form a constituent part of the system. As previously noted, there is an emerging trend relating to analyzing electrical signals, in particular stator currents, to evaluate the health of rotating machinery connected to an electric motor. However, electric motors are complex systems, especially when considering the interactions of rotor and stator fluxes, and the links between flux linkage and slip, and as yet it is not fully understood exactly how torsional oscillations related to faults in a drive line component, are translated through to the stator current. This issue is further complicated in the case that the electrical machine is supplied by a drive since the control action of the inverter can act to distort the voltage supplied to the electric motor. Hence, whilst it has been proven that fault signatures of drive train components can appear in electrical signals, it can still prove difficult to set alarm thresholds for these new techniques. Neural networks can potentially avoid these problems; if a sufficiently large number of distinct training cases are provided, the trained network will contain an 'understanding' of the interactions between the electrical machine and other components in the drive train, and hence will be able to assess the health of the system. US patent description US20070282548 A1 describes a method in which electric signals (currents and voltages) from an electric motor are used to calculate the impedance, and then this impedance is processed with the use of Neural Networks to search for patterns that could indicate known faults of the drive mechanical system, in this approach features extracted from real and complex waveforms of the impedance are used to train the neural network. In the same way European Patent EP1398119B1 describes a method to detect rotor jamming in an electric driller, by applying pattern recognition of the electrical signals by means of Neural Networks. On the other hand Baqqar et. al. shows in "A General Regression Neural Network Model for Gearbox Fault Detection using Motor Operating Parameters", an application of the Neural Networks to build a data driven model through regression (instead of being used for classification) for the fault detection of gearboxes. These approaches use the Neural Networks in different ways (classification and regression), however, in addition to a proper Neural Network training another important aspect for the successful application of Neural Networks for classification purposes relies on the information that is provided to the Neural Network during the training process. The use of representative features in combination of a proper training of the neural network guarantees the reliability of the Neural Network classification capabilities. A solution with applications to classification is given in the present invention.

Summary of the invention

The present invention provides a method for assessing the condition of rotating machinery connected to an electric motor according to claims 1-4 and a computer program for assessing the condition of rotating machinery connected to an electric motor according to claims 5-8. The present invention provides also a drive train comprising an electrical motor connected with a rotating component wherein the motor is connected with a computing device, said computing device comprising means for performing the steps of claims 1-4.

The main advantage of the invention disclosed is that, by using neural networks to analyze the motor currents and voltages, the utilization of the technique requires only the acquisition of electrical signals, without the need of electromechanical sensors such as accelerometers, therefore providing a minimally invasive approach with high reliability.

Detailed description

A method according to the present invention is explained in an embodiment shown in the drawing , where fig.1 presents a drive train with unknown condition connected to a computing device having a processor with means for implementing the method according to the invention, fig.2 presents N different drive trains connected to the computing device, where the condition of each drive train is known and used for training purposes, fig.3 - is a diagram showing examples of the operations conducted in a signal processing unit of the computing device, fig.4 - is diagram showing schematically some examples of features obtained in a feature extraction unit of the computing device, fig.5 - presents an example of a neural network structure contained in the training and classification unit of the computing device, where the structure is suitable for assesing N conditions by analyzing K+M+R input signals, fig.6 - shows a flow chart of the example of assesing the condition of the rotating component connected to an electric motor according to the present invention.

In the example system, consisting of a drive train under investigation, is presented in fig.1 , which is comprised of an electrical motor MM, powered from three phase electrical source and a rotating machine RM connected to the motor. In the example embodiment the rotating machine is a gearbox. The electric motor MM is connected to a computing device CD having six units 3.1- 3.6 suitable for implementation of the method according to the invention. The computing device CD comprises:

a data acquisition unit 3.1 for acquiring electric currents (IA, IB, IC) and voltages (UA, UB, UC) signals from the electric motor MM,

a signal processing unit 3.2 for processing the acquired signals into different groups of signals, a feature extracting unit 3.3 for extraction of features from the processed signals into different groups of features, a training and classification unit 3.4 for training a neural network with data from N different drive trains for classifying the condition of the known N drive trains, said training and classification unit 3.4 is equipped with a classifier system 3.4.1 , a data storage unit 3.5 for storing data obtained from N different drive trains, presented in fig.2 and recorded during the training of the neural network, presented in fig.5, a user notification unit 3.6 for checking the health condition of the train drive under investigation and for displaying the result of the investigation.

The same computing device CD (CD') may be used for performing the training of the neural netwok presented, as presented in fig.2. This computing device CD (CD') is equipped with the units 3.1 - 3.6 for training the neural network with the data from the N different drive trains connected to it, where the condition of each drive train is known and indicated as Condition 1. Condition 2, ....Condition N.

In the example system a set of N drive trains are presented in fig.2 wherein each drive train consists of an electric motor MM' and a rotating machine RM' which is also a gearbox (N > 2). Each of the N drive trains is comprised of electric motors and rotating machines of the same type and model, the condition of the electric motors is considered to be healthy and each of the rotating machines has a known condition, where at least one rotating machine condition corresponds to a healthy and the rest of the rotating machines conditions correspond to faulty, which are indicated as Condition 1 , Condition 2, ....Condition N. Each of the motors MM' is connected to a computing device CD or CD' having six units 3.1-3.6 suitable for using a neural network in order to implement the method according to the invention. The computing device CD or CD' is equipped with the processor with all above mentioned units, standard memories RAM, ROM and other typical hardware that are not presented in the drawing. The computing device may be a typical computer or PC with the peripheral devices connected to it for displaying or printing the result of assessment. The result of the assessment will trigger another action for a safety maintaining the continuity of investigated drive train, that are not presented in the drawing.

The method according to the invention is realized in the following steps:

Step 0. Preliminaries.

Preparing a drive train (fig.1 ) for investigation whose unknown condition needs to be assessed. Said drive train has a motor MM and a rotating machine RM (for example a gearbox). The rotating component RC is working under unknown conditions. Instead of the gearbox, a compressor, a pump, a fan or any other type of rotating machinery may be connected to the electric motor.

Step 0'. Preliminaries.

Preparing N preliminary drive trains (N > 2) (fig.2), wherein each drive train is comprised of a motor MM' of the same type and model of the motor MM and a rotating machine, RM', of the same type and model of the rotating machine RM. All of the drive train conditions are different and known a priori. The known condition of the drive train having the motor MM' 1 is named as Condition 1 , the known condition of the drive train with the motor MM'2 is named as Condition 2 and so on up to the drive train with the MM'N motor and with Condition N. At least one of the N drive trains has healthy condition.

Step 1'. Data Acquisition.

Conduct data acquisition of electrical signals of currents (IAI , IA2.. . IAN; IBI , IB2. .. IBN _; ICI , IC2.. . ICN) and of voltages (UAI , UA2. .. UAN _; UBI , UB2 . . UBN _; UCI , UC2. .. UCN ) from the motors of the N different drive trains. This data acquisition is conducted in data acquisition campaigns in which a representative amount of data is acquired, wherein the amount of data is determined by the sampling frequency of the data acquisition unit 3.1 of a computing device CD or CD', and the duration in seconds of each data acquisition campaign. The data acquired in each data acquisition campaign represents one data set. The total amount of data sets is G (where G is a natural number). At least 10 data sets (G=10) are required to conduct the next steps concerning the classifier training phase of the neural network. After G acquisition campaigns a total amount G of data sets is available in the data acquisition unit 3.1 of the computing device CD or CD'. Step 2'. Signal Processing.

Once the electrical signals have been acquired, their time waveforms are processed in the signal processing unit 3.2 in the following three ways: a) by applying a Notch filter to the currents (IAI , IA2. .. IAN) and voltages (UAI , UA2. ..

UAN) signals in order to remove the fundamental frequency component and obtaining the filtered currents waveforms Wc and filtered voltages waveforms Wv,

b) by applying the Fourier transform to the filtered currents Wc and filtered voltages Wv waveforms in order to obtain the frequency spectra of the currents Sc and the frequency spectra of the voltages Sv,

c) by applying Time Synchronous Averaging method to the filtered currents signals (Wei , Wc2, ... ,WCN) in order to obtain a time waveform TSAc that represents the behavior of the filtered currents during one rotation of the electric motor shaft.

The result of the step 2' should be five total groups G1-G5 of data sets, and each group G1-G5 has data sets 1 , 2..N of the following signals:

G1- WCI , ... ,WCN: time waveforms of the filtered currents of the drive train with Condition 1 , Condition 2,....Condition N,

G2 - Wvi, ... ,WVN: time waveforms of the filtered voltages of the drive train with condition 1 , Condition 2,... , Condition N,

G3 - Sci SCN: frequency spectra of the filtered currents of the drive train with condition 1 , Condition 2,... , Condition N,

G4 - Svi, ... ,SVN: frequency spectra of the filtered voltages of the drive train with Condition 1 , Condition 2,... , Condition N,

G5 - TSAci, ... ,TSACN: Time Synchronous Averaged waveforms of the filtered currents of the drive train with Condition 1 , Condition 2,... , Condition N. Step 3'. Extraction of Representative Features

According to the different groups G1-G5 of signals, different features may be extracted from their data sets 1 , 2...N, this is performed in the feature extraction unit 3.3 of the computing device CD or CD'.

For groups G1 and G2 typical statistical features are calculated, preferably as: mean value, standard deviation, skewness, kurtosis, signal entropy, peak to peak value, maximum peak value, root mean square (RMS), crest factor and shape factor. The total number of features to be extracted from the time waveforms will be K. These indicators are preferable and not limited to the ones listed previously, other indicators may be also used.

For groups G3 and G4 extracted features are comprised of general features associated with the spectrum itself and the amplitude values of frequency components related with the rotating machine components (blades / vanes in pumps and compressors, gears in gearboxes). The features extracted from the frequency spectrum signal include, but are not limited to: frequency center, frequency at which the maximum peak is present, amplitude of the frequency peak corresponding to the first harmonic of the Gear Mesh Frequency (1xGMF), amplitude of the frequency peak corresponding to the second harmonic of the gear mesh frequency (2xGMF), amplitude of the frequency peak of the Blade Pass Frequency (1xBPF), amplitude of the frequency peak at the second harmonic of the Blade Pass Frequency (2xBPF). The total number of features to be extracted from the frequency spectra will be M.

For group G5 similar features to the ones extracted from the waveforms as in group G1 and G2 can be calculated. These features include, but are not limited to: kurtosis, skewness, maximum peak value, crest factor and signal entropy. The total number of features to be extracted from the Time Synchronous Averaged signal will be R.

In total, there are K+M+R features extracted from each data sets 1... N of the GIGS groups and these features are used to train the neural network in the classifier system 3.4.1 which is implemented in the training and classification unit 3.4 of the computing device CD or CD'.

Step 4'. Training of Classifier System.

The classifier system 3.4.1 is built using neural networks. The neural network is trained by inputting the features extracted previously in the step 3'. The training is conducted by using typical neural network training algorithms. These training algorithms include, but are not limited to: conjugate gradient methods, the Levenberg-Marquardt algorithm or gradient descent methods.

From the total G1-G5 data sets 1.2..N , 70% are used for training and 30% are used for testing.

After the Classifier system 3.4.1 is trained, with the 70% of the training data set, additional checks are carried out to test the performance of the classifier system 3.4.1. Table 1 illustrates the structure of a confusion matrix for a classifier system trained to classify between two different drive train conditions (Condition 1 and Condition 2).

Table 1

The Actual Condition, represents the real condition of the system under analysis and the Predicted Condition represents the condition predicted by the classifier system. The true positives represent the data sets whose actual condition is 1 and whose predicted condition is . The False positivies represent the data sets whose actual condition is 2 and whose predicted condition is 1. The False negatives represent the data sets whose actual condition is 1 and whose predicted condition is 2 and finally the True Negatives represent the data sets whose actual condition is 2 and whose predicted condition is 2.

The elements of the confusion matrix: True Positives, False Positives, False Negatives and True Negatives are used to evaluate the following factors: Precision, Recall and F score of the classifier. The value of the Precision, Recall and F score will determine if the Classifier system 3.4.1 requires more training or if it is already properly trained to be used with drive train with unknown condition.

The Precision (P) is defined as:

True Positives

p = —— :

True Positives + False Positives The Recall (R) is defined as: β True .·. P. o■s■ iti .v .e· .s

True Positives + False Negatives

The F Score is defined as:

After training the Classifier system 3.4.1 , the F score is calculated for the training data set and for the test data set. If both F scores are above 0.95, the Classifier system is considered to be properly trained, if not, then more data will be acquired and the classifier system will be re-trained until the F score of both data sets (training and testing) is above 0.95.

Step 1. Data Acquisition

Repeating the step 1 ' using an on-line data acquisition unit 3.1 to acquire current (IA, IB, IC) and voltage signals (UA, UB, UC) from the motor (MM) of the drive train with unknown condition. In the fig.6 this step is indicated in parentheses.

Step 2. Signal Processing.

Repeating the step 2' for the drive train with unknown condition by filtering the measured signals to remove the fundamental frequency components and processing the filtered signals to generate: time waveforms of the currents Wcx and voltages Wvx, frequency spectra of the currents Sex and voltages Svx, Time Synchronous Averaged TSAcx waveforms of the currents by using a Signal processing unit (3.2), for the drive train with unknown condition. In the fig.6 this step is indicated in parentheses.

Step 3. Extraction of Representative Features

Repeating the step 3' for the drive train under investigation by extracting representative features of time waveforms of the currents (G1-Wcx) and voltages (G2-Wvx), representative features of the frequency spectra of the currents (G3-Scx) and voltages (G4-Svx), and representative features of Time Synchronous Averaged (TSA) waveforms of the currents (G5-TSAcx), by using Feature extraction unit 3.4 for the drive train under investigation. In the fig. 6 this step is indicated in parentheses.

Step 5. Data Set Classification Process.

Once the Classifier system 3.4.1 has been trained it can be used to classify drive trains with unknown condition (fig. 1), given that this unknown condition is at least one of the N conditions used to train the Classifier system 3.4.1. To start with the classification process of a drive train of unknown condition it is required to acquire the electric signals from the motor MM, and next repeat the steps 1 ', 2' and 3' for receiving a real data of the drive train of unknown condition. These actions are indicated in the flowchart (fig. 6) as (1), (2), (3). Once the features are extracted they are used as inputs for the Classifier system 3.4.1 , which will provide an output. This output is a vector with Z = [ZI ... ZN], elements. From this vector Z, N-1 elements should be equal to 0 and one element should be 1 .

Step 6. Condition Identification.

The output vector Z of the Classifier system 3.4.1 is comprised of N-1 elements that are 0 and one element that is 1. The index i of said element equal to 1 determines the identified condition of the drive train analyzed. Classifier Output Z = [0 0 ... 1 ... 0]

Index 1 '

Index 2 '

Index / '

Index N

In the case that the N elements of the output vector of the classifier system are 0, the user is notified that the condition of the analyzed drive train component could not be assessed.

All steps presented above are implemented in a computer program product comprising non-transitory computer-readable program code which when executed on a computing device CD or CD' carries out the steps of assessing the condition of rotating machinery RC (RC) connected to an electric motor MM comprising a drive-train, and the computing device CD or CD' is provided with functional units (3.1 -3.6) suitable for using a known neural network algorithm.