FIELD OF THE INVENTION

The present invention relates a fault state detection apparatus, a method of fault state detection, and a method of training a machine learning algorithm for a fault state detection apparatus.

BACKGROUND OF THE INVENTION

Condition monitoring data such as vibration measurements, infrared images, or current values, can be used to train artificial intelligence (Al) systems to automatically detect or even predict faulty states. However, typically, the amount of healthy data far outweighs the examples which correspond to faulty cases. This class imbalance has a negative effect on the performance of machine learning algorithms and other related approaches. In order to improve the performance, we need to circumvent the impact of class imbalance in the original (training) data. However, this is difficult to achieve.

There is a need to address this issue.

SUMMARY OF THE INVENTION

Therefore, it would be advantageous to have an improved technique to detect system faults based on measurement data from such systems.

The object of the present invention is solved with the subject matter of the independent claims, wherein further embodiments are incorporated in the dependent claims.

In a first aspect, there is provided a fault state detection apparatus, comprising: an input unit; and a processing unit.

The input unit is configured to receive condition monitoring data. The processing unit is configured to implement a trained machine learning algorithm to analyse the received condition monitoring data to determine if the received condition monitoring data is associated with a fault state. The trained machine learning algorithm was trained on the basis of a plurality of non-fault state condition monitoring data and associated ground truth information and on the basis of a plurality of fault state condition monitoring data and associated ground truth information. A subset of the plurality of fault state condition monitoring data was generated from one or more non-fault state condition monitoring data. Generation of fault state conditioning monitoring data in the subset of the plurality of fault state condition monitoring data comprises a transformation of non-fault state condition monitoring data to fault state condition monitoring data.

Thus, an artificial-intelligence-based approach is developed that takes examples of healthy condition monitoring data as input. This input is transformed into examples of faulty condition monitoring data. This reduces the class imbalance between healthy and faulty examples that is very common in condition monitoring data, because normally data associated with healthy operation far outweighs data associated with faulty operation, and now this imbalance is mitigated in a completely new way to provide for better machine learning fault detection determination. In other words, the balanced data provides for better training results than an imbalanced version.

Here, the ground truth information relates to whether the data is associated with a non fault state or associated with a fault state. The ground truth data associated with the fault state, can also detail the type of fault. Thus, the apparatus can in certain embodiments not only determine is acquired data is associated with healthy operation or with faulty operation, but if associated with faulty operation can also determine the type of fault that has lead to the faulty data.

In an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can then comprise an increase in a vibration amplitude peak. In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The increase in the vibration amplitude peak can comprise an increase in the amplitude in one of the different frequency bins.

In an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to fault state condition monitoring data can comprise a decrease in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The decrease in the vibration amplitude peak can comprise a decrease in the amplitude in one of the different frequency bins.

In an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to fault state condition monitoring data can comprise a shift in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The shift in the vibration amplitude peak can comprise an increase in the amplitude in a first one of the different frequency bins and an associated decrease in the amplitude in a second one of the different frequency bins.

In an example, the transformation of the non-fault state condition monitoring data to the fault state condition monitoring data comprises a preservation of a total vibrational energy.

In an example, the transformation of the non-fault state condition monitoring data to the fault state condition monitoring data comprises a change of a total vibrational energy. In an example, the received condition monitoring data comprises infrared image data and the plurality of non-fault state condition monitoring data comprises infrared image data and the plurality of fault state condition monitoring data comprises infrared image data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a hot spot to a non-fault state infrared image.

In an example, the hot spot is added at a random position within the non-fault state infrared image.

In an example, the hot spot is added at a position within the non-fault state infrared image associated with a conductive part of an imaged object.

In an example, the received condition monitoring data comprises visible image data and the plurality of non-fault state condition monitoring data comprises visible image data and the plurality of fault state condition monitoring data comprises visible image data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a scratch or dent to an object in a non-fault state visible image.

In a second aspect, there is provided a method of fault state detection, comprising: receiving by an input unit condition monitoring data; and analysing by a trained machine learning algorithm implemented by a processing unit the received condition monitoring data to determine if the received condition monitoring data is associated with a fault state.

The trained machine learning algorithm was trained on the basis of a plurality of non fault state condition monitoring data and associated ground truth information and on the basis of a plurality of fault state condition monitoring data and associated ground truth information. A subset of the plurality of fault state condition monitoring data was generated from one or more non-fault state condition monitoring data. Generation of fault state conditioning monitoring data in the subset of the plurality of fault state condition monitoring data comprises a transformation of non-fault state condition monitoring data to fault state condition monitoring data. In a third aspect, there is provided a method of training a machine learning algorithm for a fault state detection apparatus, the method comprising: providing a plurality of non-fault state condition monitoring data and associated ground truth information; providing a plurality of fault state condition monitoring data and associated ground truth information, the providing comprising generating a subset of the plurality of fault state condition monitoring data from one or more non fault state condition monitoring data, and wherein the generating of fault state conditioning monitoring data in the subset of the plurality of fault state condition monitoring data comprises transforming non-fault state condition monitoring data to fault state condition monitoring data; implementing a machine learning algorithm on a processing unit; and training the machine learning algorithm on the basis of the plurality of non-fault state condition monitoring data and the associated ground truth information and on the basis of the plurality of fault state condition monitoring data and the associated ground truth information.

The above aspects and examples will become apparent from and be elucidated with reference to the embodiments described hereinafter.

BRIEF DESCRIPTION OF THE DRAWINGS

Exemplary embodiments will be described in the following with reference to the following drawing:

Fig. 1 shows a schematic illustration of a hot spot added to a infrared image of a piece of equipment operating normally to transform the image to one associated with normal operation to one associated with faulty operation.

DETAILED DESCRIPTION OF EMBODIMENTS

Fig. 1 relates to data useable with a fault state detection apparatus, a method of fault state detection, and a method of training a machine learning algorithm for a fault state detection apparatus. According to an example of a fault state detection apparatus, the fault state detection apparatus comprises an input unit, and a process unit. The input unit is configured to receive condition monitoring data. The processing unit is configured to implement a trained machine learning algorithm to analyse the received condition monitoring data to determine if the received condition monitoring data is associated with a fault state. The trained machine learning algorithm was trained on the basis of a plurality of non-fault state condition monitoring data and associated ground truth information and on the basis of a plurality of fault state condition monitoring data and associated ground truth information. A subset of the plurality of fault state condition monitoring data was generated from one or more non-fault state condition monitoring data. Generation of fault state conditioning monitoring data in the subset of the plurality of fault state condition monitoring data comprises a transformation of non-fault state condition monitoring data to fault state condition monitoring data.

In an example, the apparatus is configured to output an indication that the received condition monitoring data is associated with a fault state.

In an example, the trained machine learning algorithm is a neural network.

In an example, the trained machine learning algorithm is a decision tree algorithm.

According to an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an increase in a vibration amplitude peak.

According to an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The increase in the vibration amplitude peak can comprise an increase in the amplitude in one of the different frequency bins.

According to an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to fault state condition monitoring data can comprise a decrease in a vibration amplitude peak.

According to an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The decrease in the vibration amplitude peak can comprise a decrease in the amplitude in one of the different frequency bins.

According to an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to fault state condition monitoring data can comprise a shift in a vibration amplitude peak.

According to an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The shift in the vibration amplitude peak can comprise an increase in the amplitude in a first one of the different frequency bins and an associated decrease in the amplitude in a second one of the different frequency bins.

According to an example, the transformation of the non-fault state condition monitoring data to the fault state condition monitoring data comprises a preservation of a total vibrational energy.

According to an example, the transformation of the non-fault state condition monitoring data to the fault state condition monitoring data comprises a change of a total vibrational energy.

According to an example, the received condition monitoring data comprises infrared image data and the plurality of non-fault state condition monitoring data comprises infrared image data and the plurality of fault state condition monitoring data comprises infrared image data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a hot spot to a non-fault state infrared image. According to an example, the hot spot is added at a random position within the non fault state infrared image.

According to an example, the hot spot is added at a position within the non-fault state infrared image associated with a conductive part of an imaged object.

According to an example, the received condition monitoring data comprises visible image data and the plurality of non-fault state condition monitoring data comprises visible image data and the plurality of fault state condition monitoring data comprises visible image data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a scratch or dent to an object in a non-fault state visible image.

In an example, the received condition monitoring data comprises current data and the plurality of non-fault state condition monitoring data comprises current data and the plurality of fault state condition monitoring data comprises current data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise a variation in a current value in non-fault state current data.

In an example, the received condition monitoring data comprises voltage data and the plurality of non-fault state condition monitoring data comprises voltage data and the plurality of fault state condition monitoring data comprises voltage data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise a variation in a voltage value in non-fault state voltage data.

According to a method of fault state detection, the method comprises: receiving by an input unit condition monitoring data; analysing by a trained machine learning algorithm implemented by a processing unit the received condition monitoring data to determine if the received condition monitoring data is associated with a fault state; wherein, the trained machine learning algorithm was trained on the basis of a plurality of non-fault state condition monitoring data and associated ground truth information and on the basis of a plurality of fault state condition monitoring data and associated ground truth information; wherein, a subset of the plurality of fault state condition monitoring data was generated from one or more non-fault state condition monitoring data, and wherein, generation of fault state conditioning monitoring data in the subset of the plurality of fault state condition monitoring data comprises a transformation of non-fault state condition monitoring data to fault state condition monitoring data.

In an example, the trained machine learning algorithm is a neural network.

In an example, the trained machine learning algorithm is a decision tree algorithm.

In an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an increase in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The increase in the vibration amplitude peak can comprise an increase in the amplitude in one of the different frequency bins.

In an example, the condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to fault state condition monitoring data can comprise a decrease in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The decrease in the vibration amplitude peak can comprise a decrease in the amplitude in one of the different frequency bins. In an example, the received condition monitoring data comprises vibration data and the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transformation of the non-fault state condition monitoring data to fault state condition monitoring data can comprise a shift in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The shift in the vibration amplitude peak can comprise an increase in the amplitude in a first one of the different frequency bins and an associated decrease in the amplitude in a second one of the different frequency bins.

In an example, the transformation of the non-fault state condition monitoring data to the fault state condition monitoring data comprises a preservation of a total vibrational energy.

In an example, the transformation of the non-fault state condition monitoring data to the fault state condition monitoring data comprises a change of a total vibrational energy.

In an example, the received condition monitoring data comprises infrared image data and the plurality of non-fault state condition monitoring data comprises infrared image data and the plurality of fault state condition monitoring data comprises infrared image data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a hot spot to a non-fault state infrared image.

In an example, the hot spot is added at a random position within the non-fault state infrared image.

In an example, the hot spot is added at a position within the non-fault state infrared image associated with a conductive part of an imaged object.

In an example, the received condition monitoring data comprises visible image data and the plurality of non-fault state condition monitoring data comprises visible image data and the plurality of fault state condition monitoring data comprises visible image data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a scratch or dent to a an object in a non-fault state visible image.

In an example, the received condition monitoring data comprises current data and the plurality of non-fault state condition monitoring data comprises current data and the plurality of fault state condition monitoring data comprises current data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise a variation in a current value in non-fault state current data.

In an example, the received condition monitoring data comprises voltage data and the plurality of non-fault state condition monitoring data comprises voltage data and the plurality of fault state condition monitoring data comprises voltage data. The transformation of the non-fault state condition monitoring data to the fault state condition monitoring data can comprise a variation in a voltage value in non-fault state voltage data.

According to a method of training a machine learning algorithm for a fault state detection apparatus, the method comprises: providing a plurality of non-fault state condition monitoring data and associated ground truth information; providing a plurality of fault state condition monitoring data and associated ground truth information, the providing comprising generating a subset of the plurality of fault state condition monitoring data from one or more non fault state condition monitoring data, and wherein the generating of fault state conditioning monitoring data in the subset of the plurality of fault state condition monitoring data comprises transforming non-fault state condition monitoring data to fault state condition monitoring data; implementing a machine learning algorithm on a processing unit; and training the machine learning algorithm on the basis of the plurality of non-fault state condition monitoring data and the associated ground truth information and on the basis of the plurality of fault state condition monitoring data and the associated ground truth information.

In an example, the machine learning algorithm is a neural network.

In an example, the machine learning algorithm is a decision tree algorithm.

In an example, the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transforming the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an increase in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The increase in the vibration amplitude peak can comprise an increase in the amplitude in one of the different frequency bins.

In an example, the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transforming the non-fault state condition monitoring data to fault state condition monitoring data can comprise a decrease in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The decrease in the vibration amplitude peak can comprise a decrease in the amplitude in one of the different frequency bins.

In an example, the plurality of non-fault state condition monitoring data comprises vibration data and the plurality of fault state condition monitoring data comprises vibration data. The transforming the non-fault state condition monitoring data to fault state condition monitoring data can comprise a shift in a vibration amplitude peak.

In an example, the non-fault state condition monitoring data are represented as a plurality of amplitudes in different frequency bins. The shift in the vibration amplitude peak can comprise an increase in the amplitude in a first one of the different frequency bins and an associated decrease in the amplitude in a second one of the different frequency bins.

In an example, the transforming the non-fault state condition monitoring data to the fault state condition monitoring data comprises a preservation of a total vibrational energy.

In an example, the transforming the non-fault state condition monitoring data to the fault state condition monitoring data comprises a change of a total vibrational energy.

In an example, the plurality of non-fault state condition monitoring data comprises infrared image data and the plurality of fault state condition monitoring data comprises infrared image data. The transforming the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a hot spot to a non-fault state infrared image.

In an example, the hot spot is added at a random position within the non-fault state infrared image.

In an example, the hot spot is added at a position within the non-fault state infrared image associated with a conductive part of an imaged object.

In an example, the plurality of non-fault state condition monitoring data comprises visible image data and the plurality of fault state condition monitoring data comprises visible image data. The transforming the non-fault state condition monitoring data to the fault state condition monitoring data can comprise an addition of a scratch or dent to a an object in a non-fault state visible image.

In an example, the plurality of non-fault state condition monitoring data comprises current data and the plurality of fault state condition monitoring data comprises current data. The transforming the non-fault state condition monitoring data to the fault state condition monitoring data can comprise a variation in a current value in non-fault state current data. In an example, the plurality of non-fault state condition monitoring data comprises voltage data and the plurality of fault state condition monitoring data comprises voltage data. The transforming the non-fault state condition monitoring data to the fault state condition monitoring data can comprise a variation in a voltage value in non-fault state voltage data.

Thus, in order to address training data imbalance, data that resembles the faulty case is generated and used to augment the training data. The generated faulty cases are then a “superset” of all possible failures. Additionally, there might even be some “failure patterns” which will never occur in real life in the data. However, the failure mechanism does not need to be fully understood as long as the basic principles (“Shotgun approach”) are understood, and if training data is used that will not actually be seen in real life, then this does not matter, because the trained machine learning algorithm will not assign this non-real life generated data to actual data being measured. What is important is that the trained data, associated with faults, has enough real-life faulty data. Thus, real faulty data can be used and faulty data generated from real data can be inspected by a human if necessary to validate that it is data that could be associated with a real fault.

This new concept can also be explained through through the following examples:

Vibration data consists of a list of amplitudes for different frequency bins. The healthy data can be changed by reducing and/or increasing specific bins. The total amount of energy can be preserved or not. It is known that failures manifest as increased peaks, sometimes accompanied by shifting peaks. There are strong physical arguments for changes in specific peaks corresponding to specific failure mechanisms. These rules can be considered but can be simplified for the purpose of data generation. Thus, the real failures will be a subset of all generated cases. However, since the “unrealistic” failures never occur, they do not negatively impact the performance of the system.

IR images can be used to monitor hot spots in electrical systems. Detecting hot spots and distinguishing them from normal heat up is easy for humans. For Al systems, it is a task of image recognition. One way to generate faulty data is to add visually convincing hot spots to images taken of the system under surveillance randomly. In Fig. 1, a healthy image is taken. An augmentation algorithm takes this image as input, selects a random spot (possibly with some guidance) and adds a hot spot through non- homogeneously adding values to the local temperature. While not all hot spots will correspond to realistic failure situations (for example a hot spot might show up on a non-conductive area), the algorithm will learn what it has to look for.

Also, current and voltage data can similarly be used to modify non-faulty data to augment healthy data to be used in training the machine learning algorithm that is then used for fault detection. While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing a claimed invention, from a study of the drawings, the disclosure, and the dependent claims.