Method for reducing data processing in the mechanical monitoring of an industrial process

Abstract

The invention relates to a method for reducing data processing in the machine monitoring of an industrial process, - acquiring training data comprising a plurality of data time series (S10); - training a machine learning algorithm to predict a feature in the data time series, comprising repeatedly performing the following steps: 1) determining a signal window in each data time series (S14), wherein each signal window is defined by a starting point in the original data time series and a signal length; 2) inputting the signal windows into the machine learning algorithm (S16); 3) determining a loss function (S18), wherein the loss function takes into account the starting point, the signal length, and the accuracy of the prediction of the feature (S22); 4) adjusting the starting points and signal length of the signal windows in the data time series to minimize the loss function;- providing the machine learning algorithm (S24), wherein the machine learning algorithm is configured to monitor the industrial process;

Description

The invention relates to monitoring an industrial process for a parameter which is detected sensorially when the industrial process is carried out. Further, the invention is in the field of automated anomaly, fault and fault detection in industrial processes.Anomaly detection is a method of data analysis, in which specific instances or patterns are detected that do not correspond to the norm or the statistically expected behavior. The aim here is to determine, by the analysis of data, whether a data point deviates from the normal or intended pattern or behavior. For this purpose, a reference must first exist, by which it is determined which behavior can be considered normal.For abnormality detection, a wide variety of methods are used, which have in common that they can detect outliers from large amounts of data. These typically use historical data of a machine or an entire installation.However, due to the stochastic character of many industrial processes and the possibly possible geometric complexity of the machined parts, it may be very difficult to detect an occurring process anomaly within the duration of the machining process. The sensor signals obtained can be quite large due to high sampling rates or long process durations, which leads to computational challenges with regard to data storage and the use of transmission methods. This applies in particular when the industrial process is to be scaled.Further, if the monitoring system analyzes all sensor data at each process, delays may occur in real-time monitoring. This is particularly problematic when fast responses to anomalies are required to prevent damage or quality issues.The sensor signals may have to be processed for real-time monitoring. This conditioning of the sensor signals for monitoring can be time-consuming and demanding. The data must frequently be digitized, cleaned or denoised, normalized, and placed in the proper format in order for the monitoring algorithm to operate effectively.It has therefore proved expedient to use only a small part, for example 0.5 seconds, of the sensor signal obtained in order to extract the corresponding features and to image the process anomaly which has occurred. The features extracted from these sections or signal windows may then be used in conjunction with machine learning techniques to solve regression or classification tasks.Approaches are already known from the prior art for identifying these signal windows with regard to position and length. A first possibility is to access a process expert's experience. A process expert analyzes the data time series based on his experience and available information such as engineering drawings, metrology protocols, numerical control code of the machine tool, etc., to identify an anomaly in the signal and define the signal window. The accuracy with which the anomaly is detected precisely and with a high information content depends largely on the experience, the available information and the time to solve this task. Moreover, an analysis of the industrial process by a process expert is always subject to subjective impressions of the process expert, which leads to a certain uncertainty in the process monitoring. Furthermore, a process expert is not always available.A test plan ("Design of Experiments", DoE) can likewise be used for the analysis of the process signal obtained. The success of DoE depends on the DoE setup used and the resources available.DoE planning and execution often requires considerable resources in terms of time, money, and labor. This may be a hurdle, especially for smaller companies or projects with limited resources. Furthermore, a deep understanding of the statistical methods and process being studied is usually required. This can be very complex for persons without corresponding training and experience.DoE therefore generally requires more time than other process monitoring methods. The planning and execution of experiments and the analysis of the data can take weeks or even months.The object of the present invention is to propose a simplified method for monitoring and a method for determining an optimum signal window which requires fewer resources.The object is achieved by the subject matters of the independent claims.Disclosure of the InventionAccording to a first aspect of the invention, this object is achieved by a method, in particular a computer-implemented method, for reducing data processing during machine monitoring of an industrial process,acquiring training data comprising a plurality of data time series;training a machine learning algorithm for predicting a feature in the data time series comprising repeatedly executing the following steps:1) determining a signal window in each data time series, each signal window defined via a starting point in the original data time series and a length;2) inputting the signal windows into the machine learning algorithm;3) determining a loss function, wherein the loss function takes into account the starting point, the length and the accuracy of the prediction of the feature;4) adjusting the start points and length of the signal windows in the data time series to minimize the loss function;providing the machine learning algorithm, wherein the machine learning algorithm is configured to monitor the industrial process, wherein the machine learning algorithm uses an optimized starting point and an optimized length as hyperparameters.An anomaly can be, in particular, a change in state in the industrial process or identify a fault or an irregularity in the industrial process. The change of state may also be intended or unintended.An industrial process may be any commercially applicable process for which at least one machine is used. A system within the scope of this invention denotes a combination of at least one machine, preferably a plurality of machines. Machines can perform all possible industrial processes. Among the industrial processes, there may be, but are not limited to, conveyance, cutting, drilling, milling, grinding, punching, casting, and injection molding, mounting, welding, painting, packaging, positioning, etc.The term "industrial system" can therefore include individual machines, such as robot arms, production lines over the full circumference, and the range between them.The data time series describes or characterizes a process size of the industrial process.A process variable represents an initial value or measured value that is obtained from the industrial process that is being carried out.The process variable quantitates the industrial process and therefore describes an actual value. For example, if the process is repositioning a workpiece using a robot arm, the process variable may be determined based on a position of the workpiece.Furthermore, a process variable can also be determined on a workpiece produced or processed using the industrial process. For example, a sensor can measure the workpiece, it then being possible to detect anomalies on the basis of the measured property of the workpiece. An example may illustrate this.The industrial process may include, for example, drilling a wellbore into a workpiece. The diameter and the depth of the hole to be drilled and the material of the workpiece are predefined by the industrial process. The process variable could be the weight of the workpiece in this case. By determining the diameter and depth, the volume of the ejection material can be determined. Each measured weight of the workpieces may then represent a data point in the data time series.In particular in fully automated industrial systems, monitoring of the process variables may be useful.The acquisition of the training data can therefore comprise, in particular, an acquisition of the process variables measured directly in the industrial process.In each training step, all data time series in the training dataset are reduced to signal windows according to the parameters proposed by the acquisition function. The signal window is characterized by two parameters. The starting point (sp) and the signal length (sl) relative to the original data time series.The signal window may be defined as:Here, L is the length of the original data time series, and the condition x 1+ x 2 ≤ 1 causes the signal window to be within the data time series. In order to ensure that the condition that the partial signal is completely contained in the original signal is fulfilled, the search space must be mapped to a restricted space:To optimize the model quality while simultaneously minimizing the computing effort for the signal processing by minimizing the signal length, the following objective function can be defined, for example: where y acc is the accuracy of the model and a and β control the weighting between the optimization of the signal length and the optimization of the model accuracy.The training process is repeated until an optimum for the starting point and the signal length has been found. These are then referred to as the optimized starting point and optimized signal length.A system trained in this way can then be used for real-time monitoring of an industrial process. Since the signal window includes fewer data points than the original data time series, the data points to be examined are reduced. This makes it possible to save computation capacities and to reduce the storage requirement for the data points to be examined. This achieves the object of the invention.In one embodiment, a Bayesian optimization is used to determine the starting point and the length of the feature.Bayesian optimization is a powerful method for optimizing complex, expensive and nonlinear target functions. It is based on the Bayesian theorem, which updates the probability of an event based on existing information. Optimization is a probabilistic method that projects various configurations of parameters and, based on the results so far, draws conclusions for the next configuration. Bayesian optimization aims to minimize the number of evaluations of the objective function while at the same time finding a good solution.Bayesian optimization is an iterative method that goes through several steps until an abort criterion is fulfilled. A surrogate model, also known as a response surface model, is updated based on the available data. This is generally done using the Bayesian theory in order to estimate a probability distribution over possible target function values.An acquisition function is calculated to determine which configuration to evaluate next. Acquisition functions can be, for example, Expected Improvement (EI), Probability of Improvement (PI) and Upper Confidence Bound (UCB). These functions combine the current surrogate model and its uncertainty to identify the most promising configurations.After completion of the loop, the Bayesian optimization returns the best found configuration based on the collected data and the surrogate model. Bayesian optimization is particularly effective in optimizing expensive and nonlinear objective functions because it dependently decides which configurations should be tested to optimize the objective function with as little effort as possible. This makes them a useful technique in applications such as hyperparameter optimization of machine learning models.In one embodiment, the industrial process is a grinding process.The grinding of components can be differentiated into different variants. In grinding, geometrically undefined cutting edges are used to process components with partially complex geometry and high hardness (e.g. >55 HRc) with high precision. The duration of machining a workpiece with a grinding process is, as in most other machining processes, dependent on the machining parameters, the quality characteristics required and the amount of material to be removed. During this period, process anomalies may occur that may have a negative effect on the quality characteristics of the workpiece. These anomalies can affect the surface integrity of the workpiece and its finished properties.The surface integrity has a decisive significance for the performance and the service life of components. The characteristic values such as structure formation, hardness or residual stress are generally considered to be decisive parameters which can be set by the heat treatment and can be changed considerably by the final grinding step of the component surface depending on the material state after the heat treatment and the grinding conditions. The affected material depth is generally limited to a few tenths of a millimeter.The underlying processes are extremely complex and do not allow precise statements about the final state of the generated or influenced surface integrity. After grinding, therefore, final examinations are always required for assessing the surface integrity state. Deviations from the desired surface integrity state, such as abrasive firing or structural changes, can be summarized under the term process abnormality.Such anomalies can be detected at an early stage with the lowest possible resource load using a system which carries out an industrial process according to this embodiment.In one embodiment, the data time series were recorded using at least one sensor for recording a process variable, wherein the at least one sensor is designed for recording structure-borne sound, force, vibration, electrical power, current intensity and / or voltage as a process variable.Depending on the object or process phenomenon to be monitored, different sensors can preferably, but not exclusively, be used. These include, for example, sensors for detecting sound emission, force, vibration and the electrical power, current or voltage of drive components and other suitable sensors. The electrical power and / or current intensity is preferably measured at a drive and / or in the vicinity of a drive of a process tool, for example a component and / or tool spindle.In one embodiment, at least one of the process variables forms an electrical power and / or current intensity, wherein this process variable is detected with a sampling rate between 1 kHz and 1000 kHz, in particular between 1 kHz and 100 kHz. The sampling rate is in particular a frequency with which a process variable, in particular a signal, is sampled in a predefined time. This embodiment is based on the consideration of ensuring sufficient accuracy for determining the extent of the damage and at the same time being able to be executed in a data-saving manner.In one specific embodiment, it is provided that at least one of the sensors is designed to detect structure-borne noise as a process variable. The structure-borne sound is preferably recorded at a sampling rate between 1 kHz and 2 MHz. Optionally, the structure-borne sound is detected as the process variable with a sampling rate between 1 kHz and 10 MHz, in particular 1 kHz and 5 MHz, preferably 10 kHz and 2 MHz and in particular 1 MHz and 2 MHz.The sensor for detecting the structure-borne sound can be designed as a structure-borne sound sensor. The structure-borne sound sensor proposes a non-destructive testing method of structure-borne sound analysis, wherein at least one evaluation parameter is obtained from the structure-borne sound, which evaluation parameter is used for determining the damage, in particular the grinding burn. For example, the structure-borne sound sensor is designed as a vibration travel sensor, a vibration acceleration sensor and / or a vibration speed sensor. In particular, the structure-borne sound sensor is designed as a ring sensor, an SEA (sound emission sensor) and / or an AE (acoustic emission) fluid sensor.In one embodiment, the sensor for detecting the structure-borne sound is arranged at a distance of at most 15 cm, in particular at most 10 cm, in particular as close as technically possible, from the component and / or a contact zone of component and a tool. The structure-borne sound sensor is preferably mounted and / or attached at a short distance from the contact zone between the component and the tool. In particular, the structure-borne sound sensor is mounted and / or attached in or on the tool spindle, in or on a component spindle, on a tailstock and / or on a tailstock centring tip. Preferably, a number of contact zones between components, in particular bearings and / or screw connections, is small. This embodiment is based on the consideration of achieving measurement results which are as precise and good as possible on the basis of contactless measurement.In one embodiment, the training data comprises data time series from a plurality of sensors, wherein an optimized starting point and an optimized length are determined for each sensor and / or each sensor type.The more sensors used for data acquisition, the more possibilities for abnormality detection can be used. In principle, anomalies from more data can be better detected. On the other hand, too much data should not be acquired to not increase the resource load too much when processing the data points. Preferably, a compromise is found between the number of sensors and / or sensor types used on the one hand and the resources required for processing the data points. In advantageous embodiments, more sensors / more sensor types cause fewer data points per signal window to be needed to detect an anomaly.In one embodiment, the data time series are digitized, normalized, amplified, and / or filtered prior to input to the machine learning algorithm.For example, the charge amplifier is configured as a charge-voltage converter that converts charges into a proportional voltage. The amplified charges are adjusted, linearized and / or filtered, for example, using a filter. This embodiment is based on the consideration of reducing specific errors in the process variable determination, eliminating interference and / or suppressing specific frequency ranges. The filter can be designed in particular as a low-pass filter, a high-pass filter, a band-pass filter and / or band-stop filter. For example, the analog-to-digital converter is designed as an electronic device, in particular a component and / or part of a component for adaptation, in particular for converting analog input signals into a digital data stream. The process variables of the machining process detected by the measurement value sensors are preferably amplified in their charge by charge amplifiers. In particular, the process variables are filtered by the filter. The process variables are preferably digitized using the analog-to-digital converter, while complying with a WKS (Whittaker, Kotelnikow and Shannon) scanning theory.In one embodiment, the machine learning algorithm is a model of the following: linear models, decision trees, support vector machines, neural networks.A machine learning algorithm may take various forms, such as linear models, decision trees, support vector machines, neural networks, and many others. It is optimized by learning from the training data by recognizing patterns and rules to make the best possible predictions or classifications for new data.The effectiveness of a machine learning algorithm depends on various factors, including the quality and amount of training data, the choice of algorithm, the model configuration, and the evaluation of the model based on evaluation metrics. The model is continually improved and optimized to maximize accuracy and performance.The linear regression model assumes a linear relationship between a dependent variable and one or more independent variables and is used to make continuous value predictions.Support Vector Machines (SVM) is a model that is frequently used for classification or regression and recognizes patterns in the data. It searches for the optimal separation between different classes or tries to adapt a continuous function to the data.Decision trees are a model that creates decision rules in the form of a tree diagram. It divides the data based on features and allows predictions or classifications.A probabilistic model is Naive Bayes, which is based on the Bayes Theorem and is used for classification. It is assumed that features are independent of each other, and calculates the probability of a certain class based on the given features.Neural networks relate to models that are preferably used for processing images or other grid-based data. Preferably, a neural network is used as the machine learning algorithm for the invention.The neural network architecture includes multiple nodes, neurons, or nodes arranged in layers. The sub-networks may share one or more layers to accomplish their tasks with the same output values. Thereafter, each sub-network may have its own layers, which only serve to solve the specific task of the sub-network.In a further aspect, the invention relates to a computer program with program code for carrying out a method for reducing data processing in the machine monitoring of an industrial process as described above when the computer program is executed on a computer.In a further aspect, the invention relates to a computer-readable data carrier with program code of a computer program for carrying out a method for reducing data processing in the machine monitoring of an industrial process as above when the computer program is executed on a computer.Overall, a method for reducing data processing in the machine monitoring of an industrial process, a computer program for this, a computer-readable data carrier, a system for monitoring an industrial process, and an industrial system for executing an industrial process are thus specified.The described embodiments and developments can be combined with one another as desired.Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described above or below with respect to the exemplary embodiments, which combinations are not explicitly mentioned.Brief Description of the DrawingsThe accompanying drawings are intended to provide a further understanding of the embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain principles and concepts of the invention.Other embodiments and many of the advantages mentioned are evident with reference to the drawing. The elements of the drawing shown are not necessarily shown true to scale with respect to one another.It shows: FIG. 1 schematically shows the sequence of the proposed method according to one embodiment of the invention.In the drawing, identical reference numerals designate identical or functionally identical elements, components or components, unless indicated to the contrary.FIG. 1 schematically shows the sequence of the proposed method according to one embodiment. In a first step S 10, the training data is acquired. The acquisition of the training data is carried out by observing comparable industrial processes by means of at least one sensor. For example, a grinding process can be observed with a sensor for detecting structure-borne noise. Patterns can then be recognized from the acquired training data, which are characteristic of certain anomalies such as state changes or faulty processes.In step S 12, initial parameters for the signal windows are determined from the training data and adapted to these. For example, a Bayesian optimization can be used for this purpose. With the initial parameters, in step S 14, the data time series from the training data is reduced to signal windows. Reducing a data time series to a signal window reduces the number of data points per data time series for monitoring. As a result, the storage space and the necessary computing power during the monitoring can be reduced.In step S 16, the signal windows are examined for anomalies. The anomalies may be exhibited in the signal windows in different ways. For example, the signal may be amplitude modulated or frequency modulated once an anomaly occurs. Certain anomalies do not occur until after a certain time. Thereby, parts of a data time series in which an abnormality is not expected can be omitted from monitoring.In step S 18, the model determines a loss function, the result of which is a measure of the quality of the abnormality detection in relation to the computation time applied. The better the anomalies detected from the training data match those actually present in the training data, the lower the result of the loss function.If the loss function has not yet reached a target specification, new parameters for the model are determined in step S 20. For this purpose, in particular the starting points and the signal lengths are adapted. In step S 22, the model is adjusted according to the newly determined parameters. At this time, the starting point may be shifted forward and backward in the data time series, but the signal length may also be shortened or extended.The adapted parameters can then be further adapted to the training data in a further step S 12 by means of Bayesian optimization and the cycle begins from the beginning.If the loss function is sufficiently minimized in step S 18, the model may be applied in step S 24 to detect anomalies in the industrial process.

Claims

A method for reducing data processing in machine monitoring of an industrial process, - acquiring training data comprising a plurality of data time series (S10) - training a machine learning algorithm for predicting a feature in the data time series comprising repeatedly executing the following steps: 1) determining a signal window in each data time series (S14), each signal window being defined by a starting point in the original data time series and a signal length; 2) inputting the signal windows into the machine learning algorithm (S16); 3) determining a loss function (S18), wherein the loss function takes into account the starting point, the signal length and the accuracy of the prediction of the feature (S22); 4) adjusting the starting points and signal length of the signal windows in the data time series to minimize the loss function; - providing the machine learning algorithm (S 24), wherein the machine learning algorithm is configured to monitor the industrial process, wherein the machine learning algorithm uses an optimized starting point and an optimized signal length as hyperparameters.Method according to claim 1, wherein a Bayesian optimization is used for determining the starting point and the length of the feature.The method of any preceding claim, wherein the industrial process is a grinding process.Method according to one of the preceding claims, wherein the data time series have been recorded with at least one sensor for recording the one process variable, wherein the at least one sensor is designed for recording structure-borne sound, force, vibration, electrical power, current intensity and / or voltage as process variable.Method according to one of the preceding claims, wherein the training data comprise data time series from a plurality of sensors, wherein an optimized starting point and an optimized length are determined for each sensor and / or each sensor type.The method of any preceding claim, wherein the data time series are digitized, normalized, amplified, and / or filtered prior to input to the machine learning algorithm.The method of any preceding claim, wherein the machine learning algorithm is one of linear models, decision trees, support vector machines, neural networks.Computer program with program code for carrying out a method for reducing data processing in the machine monitoring of an industrial process according to one of the preceding claims, when the computer program is executed on a computer.Computer-readable data medium with program code of a computer program for carrying out a method for reducing data processing in the machine monitoring of an industrial process according to one of Claims 1 to 7 when the computer program is executed on a computer.A system for monitoring an industrial process, the system being configured to perform a method for reducing data processing in machine monitoring an industrial process according to any one of claims 1 to 7.An industrial system for executing an industrial process, the industrial system comprising a system according to claim 10.

Images