Method and apparatus for controlling a machine

Abstract

Method and apparatus for controlling a machine. A computer-implemented method for controlling a machine includes: sensing values of measurement variables of a plurality of measurement events during operation of the machine; determining values of target variables for each of the plurality of measurement events; determining an observation for each measurement event, the observation including the sensed value of the measurement variable and the determined value of the target variable; performing principal component analysis or kernel principal component analysis on the plurality of observations, thereby resulting in load values and principal components of the observations; determining one or more principal components for the target variable, for which the absolute value of the load value is greater than a first predetermined threshold; identifying one or more measurement variables for each determined principal component, for which the absolute value of the load value is higher than a second predetermined threshold; grouping the measurement variables identified for each determined principal component into corresponding groups; and performing joint control on operating parameters represented by the measurement variables grouped into the same group to modify the target variable.

Description

Technical Field

[0001] The present disclosure relates to a computer-implemented method for controlling a machine by controlling operating parameters of the machine. Background Art

[0002] Identifying the root causes of errors, anomalies, malfunctions, and / or faults in computer control systems such as production, manufacturing lines, robots, monitoring systems, or medical imaging systems enables the provision of the necessary information to correct and improve these systems. However, the search for and measurement of root causes is not a simple task and typically lends itself to different potential explanations.

[0003] Given an input data set D and a causally related target variable V, the main goal of root cause analysis is to find the "best" and "simplest" explanation for the behavior of V given the observations of D.

[0004] For example, in the case of a production line, the target variable V could be, for example, the probability of a fault or escape predicted by a deployed machine learning model. Understanding the reasons behind a fault is crucial for increasing the efficiency of the production line, for example, by learning and understanding which operating parameters of the production line affect the target variable V and modifying them accordingly to reduce the fault / escape rate of the production line.

[0005] If the target variable V or the variables in D are the output of a prediction model such as a machine learning model, root cause analysis can also help to shed light on some of its biases, allowing for a better understanding of the model development cycle and ensuring that the model is properly validated and verified before it is deployed in the field.

[0006] Finding causal relationships in a data set may be impossible due to data set limitations. Thus, in practice, finding causal relationships is often replaced by finding correlations in the data set. However, there is no problem in quantifying and representing the correlations in the data set with a clear cut solution. Common methods such as univariate and multivariate correlation methods have traditionally tried to explain the variable V in terms of a set of variables D from a linear approximation perspective. The results are typically summarized by a single metric / number, which may often be a misleading or insufficient source of information and thus unable to identify the underlying dynamics in the data set D. The same is true for other common correlation statistics such as the Pearson correlation coefficient, mutual information, etc.

[0007] Another method for discovering correlations in a dataset (and typically also for dimensionality reduction) is principal component analysis (PCA), which attempts to systematically decompose the variance in a given dataset by mapping the original data points onto so-called "principal component axes". The PCA decomposition finds how to construct (in decreasing order) the variance of the dataset from linear combinations of these axes.

[0008] There are non-linear extensions of the PCA method, such as the kernel PCA method, for finding non-linear relationships in a dataset. For example, in et al. ), the kernel PCA was described in the paper "Nonlinear component analysis as a kernel eigenvalue problem" published in Neural Computation in 1996.

[0009] In fact, a dataset D typically captures different groups of variables that describe separate but related dynamic processes. Thus, for many applications, relying on simple correlation measures or PCA to determine how a target variable V depends on a group (subset) of variables of the dataset D is insufficient.

[0010] For example, in the case of a production line (layout), data from multiple sensors can be sampled continuously along the production line. Finally, for example, for quality control purposes, some decision must be made as to whether the produced parts are good. Without an interpretable way of evaluating how the target variable V depends on a subset of the variables of the measurement array taken throughout the production process, debugging errors is difficult and expensive. Simple correlation statistics do not reveal the true underlying dynamics of such complex systems because the dependencies between individual variables are missed. Summary of the Invention

[0011] The computer-implemented method and the machine controller of the independent claims allow for solving the following situation: for a target variable V, finding the group (subset) of measurement variables of the dataset D that is most strongly correlated with the target variable V, i.e., allowing for finding an explanation of how the target variable depends on the group of variables. In cases where a control variable indicates an anomaly, the groups can help explain the root cause of the anomaly.

[0012] In particular, in the case of a production line (layout), the group of measurement variables returned by the method can be used to, for example, determine the cause of a defective or abnormal part or to explain a system malfunction, thereby reducing costs and increasing the efficiency of debugging errors, and helping to confirm and validate the production process.

[0013] As a result, this can, for example, address some of the drawbacks of simple correlation statistics. The groups determined by the method describe the dynamic processes in the dataset D and are more descriptive than pairwise variable statistics.

[0014] Specifically, the overall variance contained in the data set D is decomposed, i.e., it covers the solution space much more widely and allows capturing all effects that may affect the target variable V.

[0015] Further examples are described below:

[0016] A computer-implemented method for controlling a machine may include: sensing values of measurement variables of a plurality of measurement events during operation of the machine, wherein values of each measurement variable are sensed at each of the plurality of measurement events, and each measurement variable represents an operating parameter of the machine; determining values of a target variable for each of the plurality of measurement events; determining, for each measurement event, an observation that includes the sensed value of the measurement variable and the determined value of the target variable; performing principal component analysis or kernel principal component analysis on the plurality of observations, thereby resulting in load values and principal components of the observations; determining one or more principal components for the target variable, for which the absolute value of the load value is greater than a first predetermined threshold; identifying one or more measurement variables for each determined principal component, for which the absolute value of the load value is higher than a second predetermined threshold; grouping the measurement variables identified for each determined principal component into corresponding groups; and performing joint control on the operating parameters represented by the measurement variables grouped into the same group to modify the target variable. The computer-implemented method mentioned in this paragraph provides a first example.

[0017] Determining groups of measurement variables that have an impact on the target variable has the effect of capturing dependencies between individual measurement variables, thereby enabling the capture of more complex clusters of underlying dynamic behavior for the group of measurement variables that may have a potentially non-linear impact on the target variable.

[0018] The determined groups of measurement variables have the potential to indicate an explanation (the underlying reason) for the behavior of the target variable and can therefore be used jointly to control the operating parameters represented by the measurement variables grouped into the same group to modify the target variable accordingly.

[0019] The method may further include: filtering out each group that includes the same measurement variables as another group, if the eigenvalue of the principal component associated with the group is less than the eigenvalue of the principal component of the other group, and the group and the other group indicate an opposite correlation between the target variable and the measurement variables. The feature combination mentioned in this paragraph provides a second example in combination with the first example.

[0020] Filtering out groups that include the same measurement variables as a "larger" group (i.e., a group for which the eigenvalue of the principal component associated with the group is larger but has an opposite correlation with the target variable) enables filtering out contradictions between groups / findings.

[0021] Due to the nature of principal component analysis or kernel component analysis, some contradictions in the groups are expected because the variance of the dataset including all the measured variables is completely decomposed, i.e., the "smaller" group can explain the variance after reducing all the previously observed effects / variance, and thus can capture the effect that there is some variance in the dataset in the opposite direction of the "larger" group.

[0022] The method may further include setting a first threshold and / or a second threshold such that the total number of groups for the principal components includes at most a predetermined percentage of the measured variables. The features mentioned in this paragraph, in combination with any one of the first example to the second example, provide a third example.

[0023] Setting a first threshold and / or a second threshold such that the total number of groups for the principal components includes at most a predetermined percentage of the measured variables enables setting the maximum number of measured variables that are considered to have a large enough impact on the target variable. This enables reducing the search for the root cause of the behavior of the target variable to the measured variables (which are included in the groups) that have the maximum effect on the target variable.

[0024] A machine controller may be configured to execute the method of any one of the first example to the third example. The machine controller mentioned in this paragraph provides a fourth example.

[0025] The machine controller provides the advantages identified above by using the method of any one of the first example to the third example. Such a controller may be applied, for example, in the context of a production line layout.

[0026] A computer program may include instructions arranged to cause a computer system to execute a computer-implemented method of any one of the first example to the third example. The computer program mentioned in this paragraph provides a fifth example.

[0027] The computer-readable medium may include transient or non-transient data representing instructions arranged to cause a computer system to execute a computer-implemented method of any one of the first example to the third example. The computer-readable medium mentioned in this paragraph provides a sixth example. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In the drawings, throughout the different views, like reference characters generally refer to like parts. The drawings are not necessarily to scale; instead, emphasis is generally placed upon illustrating the principles of the invention. In the following description, various aspects are described with reference to the following drawings, in which:

[0029] Figure 1 An exemplary production line layout is shown.

[0030] Figure 2 Shows Anscombe's quartet.

[0031] Figure 3 Shows an exemplary principal component analysis.

[0032] Figure 4 Shows an exemplary method for determining a group of variables that potentially have a causal impact on a target variable.

[0033] Figure 5 Shows a flowchart illustrating an exemplary method for controlling a machine.

[0034] The following detailed description refers to the accompanying drawings, which illustrate by way of example specific details and aspects of the present disclosure in which the invention may be practiced. Without departing from the scope of the present invention, other aspects may be utilized and structural, logical, and electrical changes may be made. The various aspects of the present disclosure are not necessarily mutually exclusive, as some aspects of the present disclosure may be combined with one or more other aspects of the present disclosure to form new aspects. Detailed Description

[0035] Hereinafter, various examples will be described in more detail.

[0036] Figure 1 Shows an exemplary production line layout 100.

[0037] In Figure 1 the example, component 101 is positioned on production line 102.

[0038] Controller 103 includes a data processing component (e.g., a processor (e.g., a CPU (central processing unit))) 104 and a memory 105 for storing control software according to which controller 103 operates and data on which processor 104 operates.

[0039] The data stored in memory 105 may include, for example, sensor signals / data received from one or more sensors 106. One or more sensors 106 may be any type of sensor that outputs continuous data. One or more sensors may also be used jointly. For example, one or more sensors 106 may sense temperature, pressure, concentration, torque, humidity, etc.

[0040] Optionally, the memory 105 also receives image data from one or more image sensors 107 (e.g., cameras). In this case, the controller 103 may determine that the component 101 is defective (and / or has an anomaly) based on the image data received from the one or more image sensors 107. This determination may be achieved, for example, by using a machine learning model that classifies the received image data into "good" and "bad" images.

[0041] For a component 101 that is considered defective, the controller 103 may determine the (potential) root cause of the defect based on the input sensor data received from the one or more sensors 106. In particular, the controller 103 may determine one or more groups (subsets) of variables that are most likely responsible for the defect.

[0042] For example, the controller 103 may determine that the group of variables (heat, pressure, welding time) is the group most likely responsible for the component defect. This specific group of variables may be associated with a specific point in the production process, i.e., it may be possible to trace back to which step / level in the production process the problem is most likely to occur.

[0043] It should be noted that the values of the variables may be measured using sensors and / or via controlled operating parameters (e.g., set welding time, set temperature, etc.).

[0044] Then, the controller 103 may send a feedback signal 108 to the error handling module 109. The feedback signal 108 contains an explanation for the defect, i.e., information representing the cause of the defect. For example, the feedback signal 108 may contain one or more groups (sub) of variables that are most likely responsible for the defect, and if determinable, may contain the corresponding step / level of the production line at which the problem is most likely occurring.

[0045] Then, the error handling module 109 may, for example, use the feedback signal 108 to adapt the production process accordingly. For example, the error handling module 109 may modify the operating parameters of the production process (such as the applied pressure, heat, welding time, etc.) to reduce the risk of defects / failures. The error handling module 109 may further command the actuator 110 to adapt the production process and / or remove the defective component from the production line.

[0046] After receiving the feedback signal 108 and depending on the level of correlation between the identified one or more groups of variables responsible for the defect and the target (control) variable, the error handling module 109 may control the production process to operate in a safe mode or possibly shut it down. The target variable may be, for example, a continuous variable representing the probability that a component is defective and / or has an anomaly.

[0047] The error handling module 109 can implement other user-specified rules and / or countermeasures in response to the received feedback signal 108.

[0048] Optionally, the error handling module 109 is part of the controller 103.

[0049] And Figure 1 Systems similar to the system illustrated in can be used in other technical fields, such as systems for conveying information, such as monitoring systems or medical (imaging) systems, or computer-controlled machines, such as robots, household appliances, power tools, personal assistants, or access control systems.

[0050] In the following, the determination of the set of variables most likely responsible for component defects will be explained in more detail.

[0051] Figure 2 The Anscombe's quartet is shown.

[0052] In Figure 2 the Anscombe's quartet 200 is illustrated, which is a sample data set of four sample distributions, all of which have the same descriptive statistics, i.e., they all have the same mean, variance, correlation, etc.

[0053] However, as clearly shown in Figure 2 each of these four sets has unique properties, i.e., descriptive statistics are not sufficient to describe them. Commonly used correlation statistics (such as the Pearson correlation coefficient, odds ratio, mutual information, etc.) are generally not sufficient to capture the underlying dynamics of complex data sets / systems.

[0054] Therefore, for many applications, especially for root cause analysis, relying on traditional correlation methods and measures is insufficient.

[0055] Another method for finding correlations in a data set is principal component analysis PCA, which attempts to systematically decompose the variance of a given data set by mapping the original values of the variables (data points) to so-called principal components or "principal component axes".

[0056] Figure 3 An exemplary principal component analysis with two principal components (axes) is shown, where the first principal component (axis) starts with a positive slope from the y-axis, and the second principal component (axis) is orthogonal to the first principal component.

[0057] In Figure 3 the (linear) PCA for a simple two-dimensional data set with two axes is illustrated.

[0058] PCA decomposition finds out how to construct (in decreasing order) the variance of the dataset from the linear combination of these two axes. In this example, two new principal components are found.

[0059] More specifically, the PCA method finds most (the largest possible amount) of the variance contained in the dataset scattered along the first principal component, and then, among all other directions orthogonal to the axis of the first principal component, the PCA method finds the second principal component, which explains most (the largest possible amount) of the remaining variance contained in the dataset.

[0060] In other words, the PCA method is an orthogonal transformation that determines a set of linearly uncorrelated principal components for a sample set containing values of correlated variables.

[0061] It should be noted that the PCA method is also commonly used for dimensionality reduction of datasets.

[0062] Extensions of the PCA method typically use kernel / kernel methods and are called kernel principal component analysis, kernel PCA, or KPCA.

[0063] Performing such kernel PCA decomposition typically includes at least the following steps:

[0064] First, normalize the input dataset D, where each variable is normalized by subtracting the mean of the variable and dividing by the variance, i.e., the normalized variable has a mean equal to zero and a variance equal to one.

[0065] Second, calculate the so-called similarity kernel of the dataset D. In linear PCA, the similarity kernel is typically the covariance of D.

[0066] Third, calculate the previously output eigenvalues and eigenvectors, such as the eigenvalues and eigenvectors of the covariance matrix in linear PCA. The eigenvectors are typically ranked by decreasing eigenvalues and arranged in a matrix. The resulting matrix is typically called the loading matrix (or "coefficient matrix"), where the columns correspond to the principal components and the rows correspond to the variables of the dataset D.

[0067] The values of the loading matrix are called loadings (or loading values) and can be positive or negative. However, no specific meaning can be assigned to the sign of the loadings; only the relative signs within each column / principal component are important for determining the correlation or anticorrelation between variables.

[0068] Optionally, multiply the output of the first step by the output of the third step to retrieve the mapping of the original data in the PCA space.

[0069] For more details on kernel PCA, see, for example, the above-mentioned et al. ( the paper "Nonlinear component analysis as a kernel eigenvalue problem" published in Neural Computation in 1996.

[0070] However, PCA, kernel PCA, and other multivariate correlation analysis methods fail to capture how the groups of variables contained in dataset D are related to the target variable.

[0071] For example, in the case of a production line as Figure 1 illustrated, it is impossible to determine how the target variable V (representing the probability that a component is defective) depends on the different groups of variables / measurements taken throughout the production process, making error debugging difficult and expensive.

[0072] This is because the dependencies between individual variables are explicitly omitted in correlation analysis methods such as PCA.

[0073] Figure 4 FIG. shows an exemplary method 400 for determining the groups of variables that have a potential causal impact on a target variable, which method 400 is implemented, for example, by a controller of a machine (such as, for example, Figure 1 the controller 103 of the production line arrangement 100 illustrated in ).

[0074] In Figure 4 the example, the dependencies between the individual variables of dataset 401 (hereinafter labeled D) are explicitly considered.

[0075] This can provide a way to find the groups of measured variables of D that are potentially the root cause behind the behavior of the target (control) variable 402 (hereinafter labeled V).

[0076] The dashed line 403 represents the boundary between the process of determining the groups of measured variables of dataset D that have a potential causal impact on the target variable V and the external inputs / parameters.

[0077] In step 404, if this is not already the case, all variables of dataset D and the target variable V are resampled using signal processing resampling techniques so that they share the same sampling rate, i.e., all variables (including the target variable V) have the same number of samples.

[0078] In step 405, a new dataset D' is constructed, where:

[0079]

[0080] Therefore, the target variable V is added to the dataset D. The dataset D' can be constructed in such a way that the variables are given in the columns (of the matrix) and the corresponding samples are given in the rows (of the matrix).

[0081] In step 406, kernel PCA is applied to the dataset D' using the kernel 407. This results in the load (coefficient) values and the principal components of the dataset D'.

[0082] The kernel 407 can be received from an external source, for example, the kernel 407 can be input by the user.

[0083] In the case where the dataset D' is to be analyzed to find linear relationships, a linear kernel such as, for example, a linear covariance kernel can be used. To find non-linear relationships, a non-linear kernel such as, for example, a mutual information kernel can be used.

[0084] In step 408, the resulting load matrix / scree plot output is examined to find the (potential) causal sub-(group) of the dataset D that explains the behavior of the target variable V.

[0085] Without loss of generality, the principal components can be given, for example, as the columns of the load matrix, and the original variables of the dataset D' are given as the rows of the load matrix.

[0086] The columns of the resulting load matrix are searched to find the columns for which the absolute value of the load of the target variable V (given by one of the rows) is higher than a predetermined threshold 410. This results in a set S of columns, which is the set of columns where the target variable V is active.

[0087] The predetermined threshold 410 can be a user-specified threshold and will typically be a value between 0 and 1. For example, the predetermined threshold 410 can have a value of 0.01, 0.05, 0.1, 0.15, or 0.2. The predetermined threshold 410 can be chosen to be sufficiently discriminative, i.e., only a small percentage (e.g., 1%, 5%, 10%) of the total number of variables should be in each of the subsequently determined groups of measured variables. The correct value of the predetermined threshold 410 to achieve this typically depends on the dataset / distribution and the specific problem.

[0088] In step 409, for the determined set S of columns, all the measured variables, i.e., the variables of the dataset D for which the absolute value of the load is higher than the predetermined threshold 410, are identified / determined.

[0089] It should be noted that the predetermined threshold 410 used in step 409 can be different from the predetermined threshold used in step 408. For example, the predetermined threshold can be chosen in such a way that the group (formed in step 411) includes at most a predetermined percentage of the measured variables.

[0090] In step 411, for each column in set S, if at least one measurement variable has been identified for that column in step 409, the identified one or more measurement variables are grouped together to form variable groups.

[0091] These are variable groups that are significantly correlated with the target variable V (or inversely correlated with the target variable V if the load sign of the target variable V is opposite to the load signs of the identified measurement variables) when the load sign of the target variable V is the same as the load signs of the identified measurement variables.

[0092] Sometimes the groups can be further split into groups that are positively correlated with the target variable V and groups that are negatively correlated with the target variable V. However, it is not always possible and / or desirable to split the variables in this way because a single group may have variables that are both positively and negatively correlated with the target variable V.

[0093] In step 412, some of the groups provided by step 411 are filtered out. For this purpose, first the groups are sorted in decreasing order of the eigenvalues of the principal components (which are provided by kernel PCA in step 406). It should be noted that this operation can be performed once kernel PCA has been applied to the data set D (in other words, at any point after step 406).

[0094] Let D' 1, D' 2,..., D' n denote the ordered groups. If there is another group D' j ( j < k ) such that D' j contradicts the findings from D' k , then group D' k is filtered out / removed.

[0095] In other words, groups that include the same measurement variables are filtered out - if the eigenvalue of the principal component associated with that group is less than the eigenvalue of the principal component associated with another group and if the group and the other group indicate opposite correlations between the target variable and the measurement variables.

[0096] The need to filter out some groups is due to the nature of PCA decomposition, which fully decomposes the variance included in the data set. Thus, it is common for some groups to be contradictory to each other. For example, in Figure 3The figure illustrates the effect, where it is obvious that the plotted data has a positive correlation. This information is captured by the first principal component, but contradicts the second principal component. This is because the second principal component only explains the remaining variance included in the dataset, that is, only the variance after discounting all previously observed effects, and thus trivially captures the effect where there is some variance in the opposite direction in this data.

[0097] Then, the filtered groups can be output for further processing and are candidates for the root cause (e.g., the root cause in case of system failure / anomaly).

[0098] By Figure 4 The groups returned by the described method correspond to clusters of dynamic behavior that describe latent processes in dataset D that have a potentially non - linear effect on the behavior of target variable V. The group of measurement variables is more descriptive than pairwise variable statistics.

[0099] Furthermore, since the group of measurement variables is formulated in the form of some combination of pairwise orthogonal axes and due to the nature of PCA, this method decomposes the overall variance found in the dataset, that is, if a group does not explain the overall variance included in the dataset, other groups that further explain the variance will be found. This allows for systematically capturing effects that may significantly affect target variable V.

[0100] Using standard dataset visualization techniques (using PCA load plots), the output of this method, that is, the determined group of measurement variables, can be easily visualized and confirmed. For example, they can be visualized by mapping one variable to the X - axis of a scatter plot, another variable to the Y - axis, and another variable to marker size, color, transparency, etc.

[0101] Furthermore, the output of this method can be provided to a controller ( Figure 4 not shown in the figure), which can then use the provided information to control a system, such as the machine (arrangement) or production line as Figure 1 illustrated. The controller can, for example, use the provided information to derive countermeasures in case of system failure, or can use the information to automate a schedule or perform specific tasks.

[0102] In summary, according to various embodiments, as Figure 5 illustrated, a computer - implemented method 500 for controlling a machine is provided.

[0103] In step 501, values of measurement variables for a plurality of measurement events during machine operation are sensed, where values of each measurement variable are sensed at each of the plurality of measurement events, and each measurement variable represents an operating parameter of the machine.

[0104] It should be noted that sensing means recording, i.e., the values of the measured variables do not necessarily have to be sensed by a sensor. They can be controlled operating parameters, such as the set welding time. They can also be received from an external source.

[0105] In step 502, the value of the target variable for each of the plurality of measurement events is determined.

[0106] In step 503, for each measurement event, an observation is determined that includes the sensed value of the measured variable and the determined value of the target variable.

[0107] In step 504, principal component analysis or kernel principal component analysis is performed on the plurality of observations, resulting in load values and principal components of the observations.

[0108] In step 505, for the target variable, one or more principal components are determined for which the absolute value of the load value is greater than a first predetermined threshold.

[0109] In step 506, for each determined principal component, one or more measured variables are identified for which the absolute value of the load value is higher than a second predetermined threshold.

[0110] It should be noted that the second predetermined threshold can be the same as the first predetermined threshold.

[0111] In step 507, the measured variables identified for each determined principal component are grouped into corresponding groups (such that a plurality of groups are generated).

[0112] In step 508, joint control is performed on the operating parameters represented by the measured variables grouped into the same group to modify the target variable.

[0113] In other words, according to various embodiments, a machine is controlled by performing a plurality of measurements during operation of the machine, where each measured variable represents an operating parameter of the machine. Then, a target variable having the same number of samples (values) is determined and added to the data set of the measured variables. Then, PCA or KPCA is performed on the data set, resulting in load values and principal components. For each principal component for which the load value corresponding to the target variable is high (above a predetermined threshold), i.e., for which the target variable is active, all measured variables that also have a high (above the same or another predetermined threshold) load value are selected. The measured variables selected for each principal component form a group. Then, these groups of measured variables are used to control the operating parameters of the machine in order to modify the target variable.

[0114] For example, in the case of a production line where the target variable represents the probability that a component has an anomaly, the determined set of measurement variables has the strongest correlation with the target variable and is thus the most likely root cause of the anomaly. Based on this information, the operating parameters of the production line can be changed / adjusted to avoid future anomalies.

[0115] Figure 5 The method can be performed by one or more computers including one or more data processing units. The term "data processing unit" can be understood as any type of entity that allows data or signals to be processed. For example, data or signals can be processed according to at least one (i.e., one or more than one) specific function performed by the data processing unit. A data processing unit can include analog circuits, digital circuits, compound signal circuits, logic circuits, microprocessors, microcontrollers, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), field-programmable gate arrays (FPGAs) integrated circuits, or any combination thereof or formed therefrom. Any other means of implementing the corresponding functions described in more detail below can also be understood as a data processing unit or a logic circuit. It will be understood that one or more of the method steps described in detail herein can be performed (e.g., implemented) by a data processing unit, through one or more specific functions performed by the data processing unit.

[0116] Although specific embodiments have been illustrated and described herein, those of ordinary skill in the art will appreciate that various alternative and / or equivalent implementations may be substituted for the specific embodiments shown and described without departing from the scope of the present invention. This application is intended to cover any adaptations or variations of the specific embodiments discussed herein. It is thus intended that the present invention be limited only by the claims and their equivalents.

Claims

1. A computer-implemented method for controlling a machine, comprising: During operation of the machine, sensing values of measurement variables of a plurality of measurement events, wherein values of each measurement variable are sensed at each of the plurality of measurement events, and each measurement variable represents an operating parameter of the machine; Determining values of target variables for each of the plurality of measurement events; For each measurement event, determining an observation, the observation including the sensed value of the measurement variable and the determined value of the target variable; Performing principal component analysis or kernel principal component analysis on the plurality of observations, thereby resulting in load values and principal components of the observations; Determining one or more principal components for the target variable, for which the absolute value of the load value is greater than a first predetermined threshold; For each determined principal component, identifying one or more measurement variables, for which the absolute value of the load value is higher than a second predetermined threshold; Grouping the measurement variables identified for each determined principal component into corresponding groups; And Performing joint control on the operating parameters represented by the measurement variables grouped into the same group to modify the target variable.

2. The method according to claim 1, further comprising: If a group includes the same measurement variables as another group and the eigenvalue of the principal component associated with the group is less than the eigenvalue of the principal component of the other group, and the group and the other group indicate an opposite correlation between the target variable and the measurement variable, then filtering out the group.

3. The method according to claim 1 or 2, comprising: Setting the first threshold and / or the second threshold such that the total number of groups of principal components includes at most a predetermined percentage of the measurement variables.

4. A machine controller configured to execute the method according to any one of claims 1 to 3.

5. A computer program comprising instructions arranged to cause a computer system to execute the method according to any one of claims 1 to 3.

6. A computer-readable medium including transient or non-transient data representing instructions arranged to cause a computer system to execute the method according to any one of claims 1 to 3.

Images