METHOD FOR MONITORING A MACHINE, COMPUTER PROGRAM PRODUCT AND ARRANGEMENT

Description

[0001] The invention relates to a method for monitoring the operation of a machine, in particular an engine or electrically driven motor, comprising the following steps: a) Training phase: i. Providing training data comprising state variables of several operating points of the machine, ii. Detecting and summarizing operating points of the training data by means of clustering into operating point clusters, iii. Training a classifier that assigns operating points to the recognized operating point clusters, iv. Training at least one anomaly detection model for operational anomaly detection, b) Application phase: i. Recording operational data comprising state variables of operating points of the machine in an operating state, ii. Assigning the operating data to the operating point clusters using the classifier, iii. Detecting operational anomalies using the at least one anomaly detection model.

[0002] During the operation of machines, such as drive systems, errors and anomalies can occur. These problems can be caused by mechanical or electrical reasons and affect both the motor and the application. If these errors and changes are not detected early, they lead to increased downtime and additional and increased maintenance costs. Monitoring motor applications is therefore crucial for efficient, cost-effective, and flexible production. However, this is associated with technical challenges: 1) Many machines, for example motor applications (such as pumps and fans), are driven by inverters and therefore have different operating points. When monitoring the condition of such applications, the different operating points are typically not explicitly taken into account, or operating point-dependent monitoring is not carried out. This leads to faults being detected too late or not at all. It is also not possible to isolate the fault. In addition, vibration can vary greatly depending on the operating point. Monitoring the general vibration level across all operating points may not detect the different changes within an operating point and thus may not detect initial damage. 2) Furthermore, condition monitoring does not take into account the general conditions of the system, e.g. the application, environmental interference, installation, etc.These can affect the robustness and reliability of condition monitoring.

[0003] Motor applications (pumps, fans) are monitored by manually configuring standards-based limits. A distinction based on operating points is currently not available, and the underlying conditions are not taken into account.

[0004] A method of the type described above is already known from CN 111 985 546 A.

[0005] Various aspects of the method described above are known from the documents CN 111 275 198 A, WO 2021 / 175493 A1, US 2020 / 285997 A1, US 2021 / 133559 A1, US 2016 / 342903 A1 and DE 10 2019 110721 A1.

[0006] Based on the problems and disadvantages of the prior art, the invention aims to improve the monitoring of such machines, in particular engines, and thus to prevent machine damage as a result of unfavorable operating conditions.

[0007] To improve monitoring, the invention proposes a method of the type defined above, wherein the at least one anomaly detection model calculates an anomaly characteristic value (or anomaly score) whose value describes the severity of the anomaly, wherein an anomaly or an anomaly type is detected if the anomaly characteristic value exceeds a predetermined threshold value.

[0008] In the context of the invention, a state variable is a macroscopic physical quantity. State variables can be used to describe the state of a physical system. Examples include thermodynamic pressure, temperature, or quantities from other scientific disciplines, such as speed, force, torque, voltage, current, power, efficiency, or volume.

[0009] An advantageous variant of the invention provides that the at least one anomaly detection model is an unsupervised model.

[0010] The invention defines an unsupervised model (see also https: / / de.wikipedia.org / wiki / Un%C3%BCberwachtes_Lernen) as a model – i.e., a simplified representation of reality – that is designed to learn by machine learning without specific target values ​​for the learning process being predetermined. The unsupervised model recognizes patterns in the input data that deviate from structureless noise. A preferred option for implementing the unsupervised model is the use of artificial neural network algorithms. These are based on the similarity of input values ​​and adapt accordingly. This makes it possible to automatically create clustering (segmentation or grouping).

[0011] An advantageous variant of the invention provides that operating points comprise measured values ​​of at least one of the following types: vibration, temperature, torque, supplied power, output power, efficiency (e.g. electrical or thermal), energy consumption, electrical current, electrical voltage, frequency supply voltage (in the case of converter-controlled electric motors).

[0012] An advantageous variant of the invention provides that the training data comprise multiple operating points and / or historical data and / or reference measurements and / or fingerprints. This training data can originate from the same machine or a group of similar machines. It is also possible to use simulated data or data from similar machines. In this case, it may be expedient for the method to include further steps, in particular: Pre-processing of the training data and / or operating data by means of data cleansing and / or data scaling, in particular using additional data sources and / or data from a digital twin and / or machine nominal values.

[0013] According to an advantageous variant of the invention, it is useful if the anomaly or anomaly type is recognized as a faulty operating point, in particular if the anomaly detection model narrows down the operating point at which the anomaly is visible or occurs. It may also be expedient to assign one of different types of anomalies to an operating state based on the anomaly characteristic value and several threshold values, whereby each type of anomaly can have its own threshold value. For example, a low anomaly characteristic value can indicate a normal state, a medium one can indicate the need for a specific impending maintenance requirement, and a higher one can indicate the need to shut down the machine.

[0014] An advantageous variant of the invention provides that the method comprises further steps: Displaying the anomaly characteristic value and / or recommendations for operating the machine and / or controlling it based on the anomaly characteristic value, in particular automatically controlling the machine so that the current operating point of the machine changes.

[0015] In principle, an arrangement according to the invention can be connected to one or more possibly different human-machine interfaces. It is expedient if at least one display is provided under these human-machine interfaces so that anomaly characteristics and / or recommendations for machine operation based on anomaly characteristics can be displayed.

[0016] An advantageous variant of the invention provides that the classifier is designed as a supervised model (see also https: / / de.wikipedia.org / wiki / %C3%9Cberwachtes_Lernen), in particular working according to the nearest-centroid or random-forest method or as a neural network.

[0017] An advantageous variant of the invention provides that a separate anomaly detection model is assigned to each individual operating point cluster, in particular wherein each operating point cluster is assigned a separate anomaly detection model. This architecture enables particularly efficient anomaly detection because, according to the particular insight of the invention, the anomalies can generally occur in a specific operating point cluster. The classification into operating point clusters on the one hand and the modeling of anomaly detection specifically for these clusters makes the individual anomaly detection models faster and more efficient than with anomaly detection without prior cluster formation. The separate assignment of an anomaly detection model to each operating point cluster affects both the training phase and the application phase.This means that the individual anomaly detection models are preferably trained in a specific operating point cluster in the training phase and then preferably applied in the application phase for anomaly detection in a specific operating point cluster.

[0018] Furthermore, the invention relates to a computer program product for carrying out a method according to the description of the invention or exemplary embodiments or according to at least one or a combination of the claims using at least one computer. The computer program product can be stored on a data carrier and can be transported, sold, or otherwise used by means of this data carrier and / or by downloading from the data carrier.

[0019] Furthermore, the invention relates to an arrangement, in particular comprising at least one computer for carrying out a method according to at least one or a combination of the claims by means of at least one computer, wherein the arrangement comprises the at least one computer prepared to carry out the method. For this purpose, the computer is preferably prepared with a computer program product according to the invention to carry out the method. The computer can also be a computer network, wherein the method can be applied by means of several computers that carry out the method in parallel or jointly. Alternatively or additionally, the computer can be designed as an edge device, preferably arranged in the spatial environment or in the preferably immediate environment of the machine or attached to the machine.The arrangement may preferably also comprise the machine, in particular a motor, for example an electric motor, in particular comprising a converter.

[0020] According to one possible exemplary embodiment of the invention, the monitoring of a machine, in particular an engine, and its application data is carried out on an operating point-specific basis. The different operating points are automatically detected and grouped using clustering algorithms. An unsupervised model for anomaly detection is then trained for each operating point. This allows various variables (e.g., vibration, temperature, etc.) to be monitored within each detected operating point.

[0021] In addition, a supervised classifier is created based on the clustering results to classify the operating points during operation. The basis for this creation is preferably historical data in the form of fingerprints / reference measurements.

[0022] At the beginning of the training phase of the possible exemplary embodiment of the invention, a fingerprint should be created. The fingerprint is a set of measurements from sensors (e.g. Siemens SIMOTICS CONNECT 400) or other data sources (e.g. Motor Digital Twin) within a defined time range and describes the reference or normal behavior of the motor application. The fingerprint contains several operating points. This provided fingerprint is preprocessed, i.e. data cleansing and data scaling are performed in order to obtain a usable standard. Other data sources can also be used for preprocessing, e.g. rated motor values ​​obtained from a digital twin of the motor.

[0023] In this exemplary embodiment of the invention, the preprocessed data is then used as input for the clustering algorithm based on machine learning (here, particularly preferred hierarchical clustering). The clustering of the operating points is preferably carried out primarily based on speed and torque measurement curves in the fingerprint, but other values ​​can also be used. The identified clusters are then used to train a verified classifier (e.g., of the nearest centroid, random forest, or neural network type) to classify new data into a cluster.

[0024] For each cluster, an anomaly detection model 1 is implemented based on one or more KPIs (e.g., vibration, temperature, etc.).

[0025] The models are then saved to be applied to new data later.

[0026] An application phase of an exemplary embodiment of the invention provides for the created models to be applied to new measurements to detect anomalies. First, a new measurement is assigned to a specific operating point. Then, the corresponding anomaly detection model is triggered, which calculates an anomaly metric or anomaly score that describes the severity of the anomaly. Based on the anomaly metric, specific recommendations / activities can be triggered.

[0027] These models preferably also detect during operation when a new measurement does not belong to a known operating point, i.e., the application is running at a new operating point. In this case, the method can provide for the affected model(s) to either be automatically retrained or for a recommendation to perform a training phase to be given via a human-machine interface.

[0028] Advantages of the invention are particularly: Early detection of faults and anomalies to avoid unplanned downtimes and better planning of maintenance activities, efficiency in maintenance and troubleshooting by isolating the anomaly by operating point, extension of production by switching to a safe operating point and avoiding critical operating points, differentiation between operational anomalies (for example, blockages in the pump) and anomalies caused by increased vibrations (imbalance, misalignment, bearing damage, etc.).

[0029] The invention is described in more detail below using a specific embodiment for clarity. It shows: Figure 1 shows a schematic flow diagram of the method according to the invention.

[0030] Figure 1shows a schematic flow diagram of the method PRC according to the invention for monitoring the operation of a machine MCH, in this case a motor ENG designed as an electrically driven motor EEG. A computer program product CPD according to the invention relates to the method PRC and is described in Figure 1 schematically indicated, wherein the computer program product CPD is stored on a data storage device and, when executed on at least one computer, causes the computer to execute the method. The method is accordingly computer-implemented. The computer and, if applicable, the machine (for example, comprising a motor) correspond to an arrangement ARR according to the invention.

[0031] First, a training phase TPH is carried out in several steps.

[0032] In step I, training data TDT comprising state variables PHC of several operating points OPP of the machine MCH is provided. The training data TDT comprises several operating points OPP and / or historical data and / or reference measurements and / or fingerprints. Subsequently, in step II, operating points OPP of the training data TDT are identified using clustering CLS and combined into operating point clusters OCL.

[0033] In step III, a classifier CLF is trained to assign the operating points OPP to the detected operating point clusters OCL. The CLF classifier operates according to the nearest-centroid or random-forest method.

[0034] In step IV, anomaly detection models ARM are trained for the purpose of operational anomaly detection.

[0035] Optionally, the invention provides that a separate anomaly detection model ARM is assigned to each of the individual operating point clusters OCL, in particular, wherein a separate anomaly detection model ARM is assigned to each operating point cluster OCL.

[0036] The training phase TPH is followed by an application phase APH. The application phase APH begins in step I. with the recording of operating data OPD comprising state variables PHC of operating points OPP of the machine MCH in a specific operating state. In step II. the operating data OPD is assigned to the operating point clusters OCL using the classifier CLF from the training phase TPH. The operating data OPD or operating points OPP comprise measured values ​​of at least one of the types vibration, temperature, torque, pressure, current, voltage, and power. Then in step III. the at least one trained, preferably operating point cluster-specific anomaly detection model ARM detects operating anomalies ANM. The anomaly detection model ARM is an unsupervised model.

[0037] In a step PRP between steps I and II of both the training phase TPH and the application phase APH (optionally only one of the two phases), the training data TDT and / or operating data OPD are preprocessed by means of data cleansing and / or data scaling, in particular using additional data sources and / or data from a digital twin and / or machine nominal values.

[0038] To classify a detected operational anomaly (ANM), the anomaly detection model (ARM) calculates an anomaly score (ASC). This value describes the severity of the anomaly. An operational anomaly (ANM) is detected when the anomaly score (ASC) exceeds a predetermined threshold (ATH).

[0039] In step IV, depending on the operating anomaly, ANM is carried out: Displaying the anomaly indicator ASC and / or at least one recommendation RCM based on the anomaly indicator ASC for operating the machine MCH and / or automatically controlling the machine MCH by means of a control command COD so that the current operating point OPP of the machine MCH changes.

Claims

1. Method (PRC) for monitoring the operation of a machine (MCH), in particular a motor (ENG) or electrically powered motor (EEG), comprising the following steps: a) training phase (TPH): i. providing training data (TDT) comprising state variables (PHC) of a number of operating points (OPP) of the machine (MCH), ii. recognising and combining operating points (OPP) of the training data (TDT) by means of clustering (CLS) to form operating point clusters (OCL), iii. training a classifier (CLF), which assigns operating points (OPP) to the recognised operating point clusters (OCL), iv. training at least one anomaly recognition model (ARM) for operating anomaly recognition, b) application phase (APH): i. recording operating data (OPD) comprising state variables (PHC) of operating points (OPP) of the machine (MCH) in an operating state, ii. assigning the operating data (OPD) to the operating point clusters (OCL) by means of the classifier (CLF), iii. recognising operating anomalies by means of the at least one anomaly recognition model (ARM), characterised in that an anomaly characteristic value (ASC) is calculated by the anomaly recognition model (ARM), the value of which describes the severity of the anomaly, wherein an anomaly and / or an anomaly type is recognised if the anomaly characteristic value (ASC) exceeds a predetermined threshold value (ATH).

2. Method (PRC) according to claim 1, wherein the anomaly recognition model (ARM) is an unsupervised model.

3. Method (PRC) according to claim 1 or 2, wherein operating points (OPP) and operating data (OPD) comprise measured values of at least one of the types: vibration, temperature, torque, pressure, current, voltage, power.

4. Method (PRC) according to at least one of the preceding claims, wherein the training data (TDT) comprises a number of operating points (OPP) and / or historical data and / or reference measurements.

5. Method (PRC) according to at least one of the preceding claims, comprising - preprocessing the training data (TDT) and or operating data (OPD) by means of data sanitisation and / or data scaling, in particular using further data sources and / or data from a digital twin and / or machine rated values.

6. Method (PRC) according to at least one of the preceding claims, comprising: - displaying the anomaly characteristic value (ASC) and / or recommendations (RCM) based on the anomaly characteristic value (ASC) for operating the machine (MCH) and / or - controlling, in particular automatically controlling the machine (MCH), so that the current operating point (OPP) of the machine (MCH) changes.

7. Method (PRC) according to at least one of the preceding claims, wherein the classifier (CLF) is embodied as a neural network or operates according to the nearest-centroid or random-forest methods.

8. Method (PRC) according to at least one of the preceding claims, wherein individual operating point clusters (OCL) are each assigned a separate anomaly recognition model (ARM), in particular wherein a separate anomaly recognition model (ARM) is assigned to each operating point cluster (OCL) in each case.

9. Computer program product for implementing a method according to one of the preceding claims by means of at least one computer (CMP) or computer network.

10. Arrangement, for implementing a method according to one of the preceding claims 1 to 8, by means of at least one computer (CMP), wherein the arrangement comprises the at least one computer (CMP) prepared for implementing the method.