controller

Abstract

The invention provides a control device, wherein a controller is configured to automatically drive an actuator with an interpretation value. The control device has a controller configured to drive an actuator. The control device further comprises a setpoint controller configured to utilize a deviation between a machine-learned interpretation value and a machine-learned standard interpretation value. The setpoint controller forms a setpoint value for the controller.

Description

Technical Field

[0001] This invention relates to a controller that can be used, for example, to control a water treatment device or a papermaking machine. Background Technology

[0002] Today, machine learning algorithms are applied to controllers to make controller setup easier than ever before. Machine learning provides systems with the ability to learn automatically and improve from experience without explicit programming. Therefore, computer systems use practical machine learning (ML) algorithms and statistical models to perform one or more specific tasks without explicit instructions. Several ML algorithms exist. Only a few are introduced here: linear regression, logistic regression, K-means, feedforward neural networks, etc.

[0003] For example, machine learning algorithms are used to analyze multivariate measurements, such as in papermaking processes. Since it is difficult to explain how ML algorithms arrive at their predictions, explanatory values ​​are used to help users assess the contribution of each input parameter to the ML algorithm's prediction. Therefore, explanatory values ​​are used to explain how the ML algorithm arrives at a particular result and also to categorize how the process works. Explanatory values ​​are obtained using methods such as SHAP (Shapley additive explanations), LIME, or DeepLIFT.

[0004] Figure 1 An example of a known control device is shown, wherein process 1 is driven by actuator 2, and actuator 2 is controlled by controller 3. A measured value 4 is taken from this process and used as feedback data to the controller. Controller 3 compares the measured value with one or more setpoints 5 and generates one or more control commands for the actuator.

[0005] Measurement value 4 can also be used for other purposes, in which case it is convenient to preprocess it 6 before actually using the measurement data. Preprocessing may include, for example, data merging, aligning time formats, modifying metadata, data validation, etc. Figure 1 In the example, Machine Learning (ML) 7 is used to extract information and patterns from a large dataset. Machine learning algorithms are typically based on statistical models, which computers can use to perform specific tasks without requiring precise instructions, relying instead on pattern recognition. Recognized patterns are obtained by building a mathematical model based on a training dataset. Prediction (simulation) and pattern recognition can be performed by feeding new data into the mathematical model. The output of the ML algorithm can be used as input information to change the setpoint of controller 3 in the form of soft sensor values ​​(predictions of processed values).

[0006] Because it is difficult to see what happened during processing from the output of ML (prediction / simulation), the interpretation value 8 (e.g.) Figure 1 The SHAP values ​​in this embodiment are used to track how the ML predictions are linked back to the input variables. For each prediction, a rating number is calculated for each input variable, indicating how much that variable contributes to the final prediction. These rating numbers can be viewed as explanatory values, indicating the importance of the input value at a given point in time.

[0007] Explanatory values ​​are used to verify how well ML algorithms and models work.9 This is easier to do from explanatory values ​​than from ML predictions. Therefore, if an ML model is not working properly, it can be modified. For example, explanatory values ​​can also be used to update statistics.10

[0008] Although explanatory values ​​are currently used to help analyze systems (such as systems with multivariate measurements), there is no way to automatically utilize explanatory values ​​when controlling one or more actuators. Summary of the Invention

[0009] The object of this invention is to provide a control device in which a controller is configured to automatically drive an actuator using interpreted values. This object is achieved by means of the description in the independent claims. The dependent claims describe different embodiments of the invention.

[0010] The control device according to the invention has a controller configured to drive an actuator. The control device also includes a setpoint controller configured to utilize the deviation between a machine learning-derived explanatory value and a normal explanation value. The setpoint controller forms a setpoint value for the controller. The machine learning-derived explanatory value and the normal explanation value are, for example, SHAP values, values ​​from the LIME method, values ​​from the DeepLIFT method, or any other possible explanatory value. Attached Figure Description

[0011] The invention will now be described in more detail with reference to the accompanying drawings, wherein...

[0012] Figure 1 An example of a prior art control device is shown.

[0013] Figure 2 An example of a control device according to the present invention is shown.

[0014] Figure 3 An example of a setpoint controller according to the present invention is shown.

[0015] Figure 4 Another example of a setpoint controller according to the present invention is shown.

[0016] Figure 5 An example of an LE or fuzzy mapping is shown.

[0017] Figure 6 Other examples of LE or fuzzy mapping are shown, as well as

[0018] Figure 7 Another example of a setpoint controller according to the present invention is shown. Detailed Implementation

[0019] Figure 2 An example of a control device according to the present invention is shown. The control device has a controller 3 configured to drive an actuator 2. Furthermore, the control device includes a setpoint controller 11 configured to utilize the deviation 12 between an interpretation value 8 derived from machine learning 7 and a standard interpretation value 13 derived from machine learning. The setpoint controller 11 forms a setpoint value 11A for the controller. Figure 2 Other possible implementations are also shown, in which the deviation error is not only based on the interpreted value, but may also include measurement data and / or ML algorithm predictions. In these other embodiments, the standard value module 13 also has standard values ​​for the measurement data and / or standard values ​​for the ML values, and the deviation calculation module 12 calculates all deviations used in the implementation.

[0020] Therefore, the setpoint controller 11 can utilize the deviation of the interpreted values, but in addition to these values, the deviation of the measured data and / or the deviation of the ML value can also be used to form the setpoint value. All these deviations can be obtained from the normal operation period of the process 1. It can be seen that utilizing pre-processed measurement data is convenient because at least some measurement noise and other defects can be filtered out.

[0021] Interpretive values ​​for machine learning and standard interpretive values ​​for machine learning include SHAP values, values ​​from the LIME method, values ​​from the DeepLIFT method, or any other possible interpretive values.

[0022] The LIME method interprets a single model prediction based on local approximations of the model surrounding a given prediction. LIME refers to a simplified input x as an interpretable input. The mapping x = h x (x) transforms the interpretable input binary vector back to the original input space. Different types of h x Mapping is used for different input spaces.

[0023] DeepLIFT is a recursive predictive interpretation method. It provides a predictive interpretation for each input x. i Assign a value C ΔxiΔy This value represents the effect of setting the input to a reference value that is the opposite of its original value. This means that DeepLIFT maps x = hx (x) converts the binary value to the raw input, where 1 indicates that the input takes its raw value and 0 indicates that it takes a reference value. The reference value represents a typical uninformative background value of the feature.

[0024] SHAP (SHapley Additive exPlanation) explanatory values ​​assign each feature a change in the expected model prediction when that feature is conditionalized. These values ​​explain how to obtain an expected value E[f(z)] to be predicted from a base set if no features of the current output f(x) are known. The order in which features are added to the expectation is important. However, this is taken into account in the SHAP values.

[0025] Figure 2 Also shown (with) Figure 1 (Similarly) Process 1 is driven by actuator 2, which is controlled by controller 3. A measured value 4 is taken from this process, and the controller compares the measured value with one or more setpoints 11A and generates one or more control commands 3A for actuator 2. This process may include several processes, and therefore, as a whole, it can be a combination of processes operating together.

[0026] As mentioned above, measurement 4 can also be used for other purposes and can be preprocessed 6. Preprocessing may include, for example, data merging, aligning time formats, modifying metadata, data validation, etc. Figure 2 In the example, Machine Learning 7 is used to extract information and patterns from large datasets. Recognized patterns can be obtained by building a mathematical model based on the training dataset. Prediction (simulation) and pattern recognition can be performed by feeding new data to the mathematical model.

[0027] Explanatory values ​​(such as SHAP values) are used to track how the ML predictions are linked back to the input variables. For each prediction, a rating number is calculated for each input variable, indicating how much that variable contributes to the final prediction. These rating numbers are explanatory values, indicating the importance of the input values ​​at a given point in time.

[0028] It can be noted that the deviation / error between the standard explained value and the explained value from the current ML prediction / estimation is calculated. The standard explained value can be a repository value found from the normal operating period of the process. Therefore, the standard explained value has been derived from the normal operating period of the process controlled by the control unit. For example, the standard value can be derived as a simple value or median of these normal periods. Normal operation of the process occurs during the period when the process or combination of processes is operating normally. Therefore, for all data (preprocessing, ML prediction, and ML explained values), a standard (optimal) value can be given or estimated (from the stored values). Thus, there can be a standard historical value repository in which the process has been identified as operating optimally.

[0029] Therefore, during operations where individual or combined processing does not operate optimally, differences, deviations, or errors are detected from measurements. These are detected as divergences from the standard value. The difference 12A from the standard value 13 is used as an input to the setpoint controller 11. Although the deviation calculation module 12 is shown as a separate module, it may also be part of the setpoint controller 11. Generally speaking, deviation is related to error. The magnitude of the error indicates whether the setpoint needs to be changed or how much the setpoint should be changed.

[0030] Figure 3 An example of a setpoint controller 11 is shown, which uses deviation / error 12A ( Figure 2 ). Figure 3 The example shows two error values ​​14 and 15 for two variables, but more variables and error values ​​can be used if needed, as shown in the figure. Therefore, at least one error / deviation value is available in the setpoint controller of this invention.

[0031] The setpoint controller includes at least one P module 16, 16A, I module 17, 17A, or D module 18, 18A, or any combination of these modules. As described above, the deviation is data input into the module. The setpoint controller also includes one or more input mapping modules 19, 20, 21, 19A, 20A, 21A for each output 22, 23, 24, 22A, 23A, 24A of the module. Furthermore, the setpoint controller includes: a summing module 25 for summing one or more outputs 26, 27, 28, 26A, 27A, 28A of one or more input mapping modules 19, 20, 21, 19A, 20A, 21A; and an output scaling module 29 for scaling the output 30 of the summing module. Additionally, the setpoint controller includes: an output mapping module 31 to provide a normalized output 32; and a setpoint adjustment module 33 for using the normalized output 32 to change the setpoint value. It is worth mentioning that, depending on the implementation, the output scaling module can provide a positive or negative output, and the shape of the mapping curve of module 30 is determined to change, i.e., output 32. The output of the setpoint adjustment module 33 is the adjusted setpoint 34. The adjusted setpoint is used as the setpoint 11A of the controller 3. The adjusted setpoint also replaces the previous setpoint value 35. It can be seen that the setpoint adjustment module 33 includes a second summing module to sum the normalized output 32 and the existing setpoint value 35.

[0032] The P, I, and D modules 16, 16A, 17, 17A, 18A, and their combinations PI, PD, ID, and PID are known, but the deviation / error of the interpreted values ​​has not previously been used as input. P modules 16 and 16A have weighting coefficients that are multiplied by the input error value. The I modules include integrator units 117 and 117A, which integrate the input error value over a certain time period. The integrated input error value is multiplied by second weighting coefficients 170 and 170A. The D modules include differentiator units 118 and 118A, which form the derivative of the error value over a specific time period. This derivative is multiplied by third weighting coefficients 180 and 180A. It can be seen that all P, I, and D modules and their combinations have a weighting coefficient unit. These units can have the same weighting coefficients or different weighting coefficients. The weighting coefficients allow for weighting of the importance of the proportional (P), integral (I), and derivative (D) components of the error value, and also allow for tuning or fine-tuning of setpoint adjustment performance by increasing or decreasing the contribution from each individual input calculation.

[0033] It's not always necessary to have all the P, I, and D modules, but as mentioned above, they can be in the controller if they are truly used and needed. Figure 3 In this embodiment, the P, I, and D modules together provide PID calculations for error values ​​14 and 15, as well as other possible values. Note that in another embodiment, a single error value (such as error value 14) is sufficient to achieve good setpoint adjustment.

[0034] Figure 4 Another possible example is shown where the D module is not required, therefore the setpoint controller in this example has PI calculation. As mentioned above, the controller may only have those modules required for P, I, D, PI, PD, ID, or PID calculations in an embodiment of the setpoint controller. It is also worth mentioning that the setpoint controller may have different calculations for different error values. For example, Figure 3 The embodiment can be modified to another solution so that PID calculation is performed on error value 14 and P calculation is performed on another error value 15 (i.e., I module 17A and D module 18A have been removed).

[0035] As described above, the setpoint controller also includes input mapping modules 19, 20, 21, 19A, 20A, and 21A for each output 22, 23, 24, 22A, 23A, and 24A of the P, I, and D modules. See also Figure 3Input mapping transforms the result of each output of the P, I, or D module into a value between -2 and 2. This can be viewed as value normalization. Input mapping is formed by linguistic equations (LE) or fuzzy logic. Nonlinearity can be conveniently considered by using input mapping. Because the properties of the processing are considered in the input mapping, the tuning of the setpoint controller is also relatively smooth. The mapping module of the setpoint controller can utilize any mapping curve independently. For example, in... Figure 3 In this context, modules 19 and 19A can be formed by LE, or one module 19 can be formed by LE and the other module 19A can be formed by fuzzy logic.

[0036] Figure 4 Other possible implementations are also shown. In addition to the bias error 14 based on the interpreted values, ML bias 40 can also be used. The bias of the measurement data can also be used, but... Figure 4 It is not displayed in [the image / data]. For example, in [the image / data]... Figure 4 It can be noted that for ML deviation, there are also P, I and D modules 16B, 17B, and their combinations PI, PD, ID and PID, which are necessary in this implementation. Figure 4 The embodiment illustrates examples of P module 16B and I module 17B (with integrator unit 117B and second weighting coefficient 170B), which have outputs 22B and 23B to mapping modules 19B and 20B. These mapping modules have outputs 26B and 27B to summing module 25.

[0037] Figure 7 Another example of an embodiment of the present invention is shown. Figure 7 Implementation methods are in accordance with Figure 4 The implementation method utilizes a deviation of 14 in the interpreted values ​​(such as SHAP values) and a deviation of 40 in the ML values. Furthermore, Figure 7 The implementation also utilizes the deviation 70 of the measurement data. Furthermore, for the measurement data deviation, there may also be the required P, I, and D modules 16C, 17C, and combinations thereof (PI, PD, ID, and PID) as described in this implementation. Figure 7 The implementation shows examples of P module 16C and I module 17C (with integrator unit 117C and second weighting coefficient 170C), which have outputs 22C and 23C to mapping modules 19C and 20C. These mapping modules have outputs 26C and 27C to summing module 25.

[0038] Figure 5This shows an example of a mapping curve 50, formed by a linguistic equation or fuzzy logic. X is the input variable, which is transformed into the output variable Y. The maximum and minimum values ​​of X and Y are determined. A linear formula (such as y = ax + b) determines the values ​​of Y between X, where X falls between the maximum and minimum values. If X is greater than the maximum X value, Y is the maximum Y. If X is less than the minimum X value, Y is the minimum Y.

[0039] The mapping curve can also be a curve other than a linear curve. It can be another curve that better matches the features being processed. Figure 6 Two other possible examples of mapping curves are shown. The solid line describes the piecewise linear mapping curve 60, and the dashed line describes the S-curve mapping 61. Other curves are also possible. Therefore, refer to... Figure 3 The mapping module can use any mapping curve individually. For example, modules 19 and 19A can have the same mapping curve (such as a linear curve) or different curves (such as different linear curves, or piecewise linear curves and S-curves).

[0040] One or more outputs 26, 27, 28, 26A, 27A, 28A from one or more input mapping modules 19, 20, 21, 19A, 20A, 21A are summed in summing module 25. Therefore, all deviation / error values ​​are taken into account. The summed output 30 is then scaled by output scaling module 29, and the scaled sum is normalized by output mapping module 31 to provide normalized output 32. Setpoint adjustment module 33 uses the normalized output to change the setpoint value.

[0041] To provide the apparatus of the present invention, it is useful to know the process to be controlled. As stated, this process often has many variables being measured. Typically, not all measurements are needed to control a particular characteristic of the process; therefore, measurement data selected for specific control is chosen. (Reference) Figure 3 and Figure 4 Input and output scaling (i.e., weighting coefficients) and curve forms are selected for the input mapping module and output mapping module. P, I, D parameters and other possible modules are selected. All units / modules of the setpoint controller (P, I, D and their combinations, weighting coefficients, mapping module, summation module) and deviation calculation module 12, data preprocessing module 6, ML module 7, and interpretation value module 8 can be executed as software, an application-specific integrated circuit, or a combination of software and hardware. Module 13, which has standard values, is a memory, which can naturally include both software and hardware.

[0042] The method for controlling industrial processes according to the present invention utilizes the control device described herein. Therefore, the method can use this control device to administer one or more chemicals used in the process. Furthermore, the processes controlled by the method of the present invention (similar to industrial processes) can be pulp and paper processing, papermaking, board or weaving processing, industrial water or wastewater treatment, raw water treatment, water reuse treatment, municipal water or wastewater treatment, sludge treatment, mining processing, oil recovery processing, or any other industrial process.

[0043] As described above, the present invention provides an automatic method for providing setting inputs to the controller 3 of the control process 1. The process can be an industrial process, such as pulp processing, papermaking, board or paper towel manufacturing, industrial water or wastewater treatment, raw water treatment, water reuse, municipal water or wastewater treatment, sludge treatment, mining, oil recovery, or any other industrial process.

[0044] The process can be, for example, a water treatment process or a papermaking process. This process is typically multivariate, thus requiring numerous measurements. To understand how the ML algorithm arrives at predicted values, explanatory values ​​are formed to evaluate the input parameters. Standard explanatory values ​​indicating normal operation of the process are also provided, which can form deviation / error values ​​for the explanatory values, and these can be used to provide setpoint commands to controller 3. In practice, there can be several different actuators 2 and controllers 3 to drive the process. Therefore, the apparatus of the present invention can include more than one controller and setpoint controller, as well as a deviation calculation module. As illustrated in the examples above, embodiments of the present invention can utilize deviations in explanatory values, deviations in ML values, and / or deviations in measurement data.

[0045] The apparatus of the present invention can be located in the same location as the process being performed. However, it is also possible that it is partially located in another location, which makes remote control of the process possible. For example, measurement data 4 is transmitted via one or more communication networks to a further process according to the present invention, in which the measurement data is processed and a setpoint adjustment is sent to the controller 3.

[0046] As can be clearly seen from the above, the present invention is not limited to the embodiments described herein, but can be implemented using many other different embodiments within the scope of the independent claims.

Claims

1. A control device having a controller (3), said controller being configured to drive an actuator (2), characterized in that, The control device includes a setpoint controller (11) configured to utilize the deviation (12A) between a machine learning interpretation value (8) and a standard machine learning interpretation value, the deviation being input data to the setpoint controller. The setpoint controller (11) forms a setpoint value (11A) for the controller. The machine learning interpretation value and the standard machine learning interpretation value are SHAP values, values ​​from the LIME method, or values ​​from the DeepLIFT method. The setpoint controller (11) is also configured to utilize the deviation (12A) between the machine learning value and the standard machine learning value, and / or the setpoint controller (11) is configured to utilize the deviation (12A) between measurement data and standard measurement data, the standard interpretation value, the standard machine learning value, and the standard measurement data being values ​​derived from the normal operation period of the process controlled by the control device.

2. The control device according to claim 1, characterized in that, The setpoint controller includes at least one P module (16, 16A), I module (17, 17A), or D module (18, 18A), or any combination of these modules, the deviation (12A) being the input data of these modules. The setpoint controller also includes an input mapping module (19, 20, 21, 19A, 20A, 21A) for each output (22, 23, 24, 22A, 23A, 24A) of these modules, a summing module (25) for summing the outputs (26, 27, 28, 26A, 27A, 28A) of the input mapping module, an output scaling module (29) for scaling the output (30) of the summing module, an output mapping module (31) for providing a normalized output (32), and a setpoint adjustment module (33) for using the normalized output to change the setpoint value (11A).

3. The control device according to claim 2, characterized in that, The setpoint adjustment module (33) includes a second summation module for summing the normalized output (32) and the existing setpoint value (11A).

4. The control device according to claim 2, characterized in that, The mapping modules (19, 20, 21, 19A, 19B, 19C, 20A, 20B, 20C, 21A, 31) are formed by linguistic equations or fuzzy logic.

5. The control device according to claim 4, characterized in that, The mapping curves of the mapping modules (19, 20, 21, 19A, 19B, 19C, 20A, 20B, 20C, 21A, 31) are provided in the form of linear curves, piecewise linear curves, S-curves and / or other curve forms.

6. The control device according to any one of claims 1 to 5, characterized in that, The control device includes a deviation calculation module (12) to provide the deviation (12A).

7. The control device according to claim 6, characterized in that, The deviation calculation module (12) is part of the setpoint controller (11).

8. The control device according to claim 6, characterized in that, The deviation calculation module (12) is a module independent of the setpoint controller (11).

9. The control device according to claim 7 or 8, characterized in that, The control device includes more than one controller (3), a setpoint controller (11), and a deviation calculation module (12).

10. A method for controlling industrial processing, characterized in that, Use the control device according to any one of claims 1 to 9.

11. The method according to claim 10, characterized in that, The control device is used to add one or more chemicals used in industrial processes.

12. The method according to claim 10 or 11, characterized in that, The industrial processes mentioned are pulp processing, papermaking, board or paper towel manufacturing processes, industrial water or wastewater treatment processes, raw water treatment processes, water reuse processes, municipal water or wastewater treatment processes, sludge treatment processes, mining processes, or oil recovery processes.

Images