Method and system for monitoring a process

Abstract

A method of monitoring a process (1) is disclosed. The invention makes use of an explanation value. An ML model of the process is created. The ML model makes use of measurements from the process as input to the ML model and forms a model output. The inputs are classified into groups. An explanation value is calculated for each input, the explanation value indicating the contribution of the input to the model output, and a sum of the explanation values of each group is calculated. The calculated sums are monitored, each sum indicating a state of the group in question.

Description

Technical Field

[0001] This invention relates to process monitoring. Such processes include, for example, water treatment plants and papermaking machines. Background Technology

[0002] Today, machine learning algorithms are used in conjunction with systems that model, analyze, and estimate the behavior of processes such as paper machines or water treatment. These processes are often multivariate, requiring numerous measurements, and can be very difficult to monitor or understand. This is especially true when measurements are taken online daily, generating and processing massive amounts of data.

[0003] Machine learning provides systems with the ability to learn automatically and improve through experience without explicit programming. Therefore, machine learning (ML) utilizes algorithms and statistical models, allowing computer systems to perform specific tasks or multiple tasks using these algorithms and statistical models without explicit instructions. Several ML algorithms exist. Only a few are mentioned here: linear regression, logistic regression, K-means, feedforward neural networks, etc.

[0004] The results of ML algorithms are often difficult to interpret, especially those from complex processes. Therefore, interpretable values ​​are used to help users interpret the results of ML. Interpretive values ​​are obtained using methods such as SHAP (Shapley Additive Interpretation), LIME, or DeepLIFT.

[0005] It is still necessary to identify the source or origin of process interference and monitor it. Summary of the Invention

[0006] The object of this invention is to provide a method for monitoring a process. Another object of this invention is to provide a method for identifying, and optionally evaluating, process disturbance sources (or multiple sources) or quality defect sources in a process. A further object of this invention is a method for controlling the process using information about the disturbance sources / origins. This invention utilizes interpretable values. Process monitoring can be automated. The objectives are achieved in the manner described in the independent claims. The dependent claims illustrate different embodiments of the invention.

[0007] An inventive method for monitoring a process and / or identifying sources of process disturbance or quality defects includes a preliminary step for creating an ML model of the process. The ML model utilizes measurements from the process as input and forms a model output. The method further includes a second preliminary step for classifying the inputs into groups. This can be accomplished using a predetermined database containing information about how the inputs belong to a group. The method has computational steps for calculating an explanatory value for each input, indicating the input's contribution to the model output, and for calculating a sum of indicator values ​​for each group, correlated with the explanatory values. Further, the method includes a monitoring step for monitoring the calculated sums, each sum indicating the status of the group in question. Monitoring can be performed in a group-specific manner. Grouping can include a main group and subgroups within the main group. The output of the method can be used for manual or automatic control of the process or a subprocess of the process. Attached Figure Description

[0008] The invention will now be described in more detail with reference to the accompanying drawings, in which:

[0009] Figure 1 A simple example of ML modeling is shown for illustration.

[0010] Figure 2 A flowchart example of the method according to the present invention is shown.

[0011] Figure 3 An example of a system according to the present invention is shown.

[0012] Figure 4 An example of calculating the sum of the main groups for input according to the present invention is shown.

[0013] Figure 5 An example of calculating the sum of the subgroups of the main group according to the present invention is shown.

[0014] Figure 6 An example is shown of the sum of SHAP values ​​calculated for chemical groups with warning and alert thresholds (lower and higher level lines, respectively).

[0015] Figure 7 An example of grouping inputs into main groups and subgroups is shown. Detailed Implementation

[0016] Figure 1 A simple example of ML modeling is shown for illustration. Real-world models are likely to be much more complex, containing a large number of variables. Figure 1In this diagram, the X-axis represents one variable, such as the power supplied to the rollers of the paper machine. The Y-axis represents another variable, such as the rotational speed of the rollers. This simple example models the relationship between these variables. As the supplied power increases, the rotational speed also increases. Point 1 is the measured rotation point, and the corresponding X-axis represents the measured supplied power. Another example could be an X-axis showing the amount of hydrophobic contaminants in the paper machine's wire wate, and a Y-axis showing the paper machine's washing interval.

[0017] As we can see, the measurement can be modeled as an increasing straight line 2. The value of this line is the average of the actual measurements. In the equation, the line can be y = Ax + b, and the line and its equation can be a simple ML model of the relationship between two variables. As mentioned, the actual process is usually much more complex, so ML models are also more complex in practice. Figure 1 In the example, it's easy to see the extent to which electricity affects the rotational speed, so the explanatory values ​​aren't very useful. However, with more variables, such as ten, it becomes difficult to understand the contribution of each variable to the model output.

[0018] Although the process is referred to as a single process in this document, it can encompass many different processes that collectively perform the overall process. Therefore, a process in this description can actually include several processes and subprocesses. For example, a paper machine process can have one or more subprocesses, such as water circulation, wet end, and break line. A wastewater treatment process can include, for example, primary, secondary, and tertiary treatment, or some of these. Similarly, the ML model of the entire process can include multiple models, which can also be called sub-models. Therefore, the entire ML model can have several outputs (i.e., the outputs of the sub-models). Each output simulates / predicts a certain property of the process.

[0019] The process can be an industrial process. This process can be, for example, pulp treatment, papermaking, board making or sub-organization manufacturing processes, paper machines, pulp mills, tissue paper machines, paperboard machines, water treatment processes, wastewater treatment processes, raw water treatment processes, water reuse processes, industrial water treatment processes, municipal water or wastewater treatment processes, sludge treatment processes, mining processes, oil recovery processes, or any other industrial process.

[0020] Subprocesses and model outputs can be selected appropriately based on the process being discussed.

[0021] Now, in this invention, it has been found that machine learning explanatory values ​​(e.g., SHAP values) can be used to monitor the process. In this document, the example refers to SHAP values, but the example could also utilize values ​​from the LIME method, the deepLIFT method, or any other possible explanatory values. SHAP (Shapley Additive Interpretation) explanatory values ​​attribute each feature to the change in the expected model prediction given that feature. These values ​​explain how the predicted expectation e[f(z)] is obtained from the base values ​​if we do not know any features of the current output f(x). The order in which features are added to the expected facts is important. However, this is taken into account in SHAP values.

[0022] The LIME method interprets individual model predictions based on local approximations of the model around a given prediction. LIME refers to a simplified input x as an interpretable input. The mapping x = hx(x) transforms the binary vector of the interpretable input into the original input space. Different types of hx mappings are used for different input spaces.

[0023] DeepLIFT is a recursive predictive interpretation method. It assigns a value CΔxiΔy to each input xi, representing the effect of setting the input as a reference value instead of its original value. This means that the DeepLIFT mapping x = hx(x) converts the binary value back to the original input, where 1 indicates the input uses its original value and 0 indicates the reference value is used. The reference value represents the typical non-informative background value of the feature.

[0024] For each input that indicates its contribution to the model output, a SHAP value (or other machine learning interpretation value) is calculated. For example, when considering the headbox of a paper machine, pump speed and pulp pH can be input into a sub-model, and their SHAP values ​​can be calculated. Another part of the paper machine can be modeled by another sub-model using other inputs, such as the wire speed of the machine. Now, when the inputs for the model used for the entire process and / or the inputs for the models used for sub-processes are categorized into groups, and using the SHAP values ​​calculated for the model output, and summing the SHAP values ​​for each group, the contribution of the grouped inputs to the model output can be monitored.

[0025] Grouping the inputs into categories is a preliminary step. It utilizes a pre-defined database containing information about how the inputs belong to which groups. This knowledge of how to group the inputs is possessed by experts in the relevant process. In many cases, expert knowledge may include experience and process knowledge, such as chemical knowledge, knowledge of chemical phenomena, etc. Expert knowledge can be in the form of forms, charts, tables, or other formats used to provide the pre-defined database. For example, the database can be in the form of forms, charts, or tables. It can be seen that the database contains expert knowledge. Such a pre-defined database can be established in any known manner. Data analysis can be performed on historical data during the database's creation.

[0026] Another example could be that the output of a waste paper line sub-model is the amount of harmful contaminants. The inputs to this sub-model could be, for example, the volume or mass flow rate of uncoated and coated waste paper, anionic charge, or colloid number. Another exemplary sub-model output could be, for example, the amount of harmful compounds from water circulation and pulp treatment, or the mass of raw water. The output of a process-wide model could be, for example, the operating capacity of a paper machine (e.g., waste paper) or a certain quality of the final product, such as defects (e.g., spots, holes), strength, sizing performance, edge wicking, formation, spore count, microbial count, certain water quality parameters in the water treatment process, or any suitable target parameters for the process in question. Defects could be, for example, defects in the quality of paper or paperboard, such as spots or holes, or defects in strength, sizing performance, edge wicking, formation, spore count, or microbial count. For example, a strength defect or other defect means that the value describing the defect does not meet the target value. For example, a strength defect might mean that the strength value is too low or uneven / varies over time.

[0027] Figure 2 An example of the method of the present invention is shown. A method for monitoring a process and / or identifying sources of process disturbances or quality defects includes a preparatory step 21 for creating an ML model of the process. Historical data (e.g., 1-2 years of data) can be used for model training. The preparatory step can be performed once, at regular intervals (regularly, e.g., annually), or irregularly, as needed. As mentioned above, the process can have several processes (sub-processes). The ML model utilizes measurements from the process as input to the ML model and forms the model output. (Refer to...) Figure 3 Therefore, the process of expressing the output of the ML model or multiple ML models is as described above. This is as known above. The method also includes a second preparatory step 22 for classifying the inputs into groups using a predetermined database. Grouping may include a main group and optional subgroups within the main group. These groups are predefined, for example, by the user or the owner of the process. Grouping may reflect predetermined quality indicators and / or process status. Grouping may include one or more inputs, and inputs may be included in one or more groups.

[0028] Measurements from the process can be, for example, chemically related measurements (e.g., pH, redox, conductivity, charge, cation requirement, wood extract concentration, turbidity, phosphorus, phosphate, chemical oxygen demand, total solids, suspended solids), which can be real-time, continuous, or laboratory measurements; chemical dosage; flow rate of chemicals to be added to the process; flow rate and concentration of water and pulp streams; production information (e.g., grade, basis weight, tonnage of product produced, machine speed, breakage, breakage type); tower and tank fill levels; process delays; and product quality measurements (e.g., defects, strength, Cobb value). Relevant / suitable measurements depend on the process in question and may include any relevant / suitable measurements from that process.

[0029] The model has a calculation step 23 for calculating explanatory values ​​(explanatory values) for machine learning, such as the SHAP value for each input indicating its contribution to the model output, and the calculation step calculates the sum of the explanatory values ​​for each group. The calculation step can typically be performed at a regular time frequency, such as every minute, every 10 minutes, or every 60 minutes. It should be noted that the explanatory values ​​are calculated for all inputs of a known ML model. Furthermore, the invention calculates a specific sum of explanatory values ​​for the predetermined groups. Therefore, the explanatory value is related to the explanatory value (e.g., the SHAP value). The explanatory value is an explanatory value, such as the SHAP value, or an explanatory bias value, such as the SHAP bias value. The explanatory bias value (e.g., the SHAP bias value) is the difference between an explanatory value (e.g., the SHAP value) and a standard explanatory value (e.g., the SHAP value). The standard value can be obtained from a well-run cycle of the process. Therefore, it can be said that the standard explanatory value (e.g., the standard SHAP value) is the explanatory value (e.g., the SHAP value) of a stable process (without disturbances in terms of operability or final product quality). Furthermore, the method includes a monitoring step 24 for monitoring the calculated sums, each sum indicating the status of the group. The input can be a real-time value, and the monitoring can also be real-time. The invention can also be used for other analytical purposes. For example, the input can be historical data used for analysis.

[0030] The monitoring process can define the state of each group of inputs. These groups are established to be associated with specific items for each group. In other words, the inputs for a group are related to a specific item. The sum of SHAP values, or the SHAP values ​​themselves (or other interpreted values), can be scaled to an appropriate level for ease of use. For example, when the sum is negative or relatively close to zero, the state of the group is good, meaning the model and process are functioning well. On the other hand, when the sum is positive, the state of the group is not good, meaning the process has one or more problems or is at risk of increased process problems / interferences. Therefore, the sum of interpreted values ​​calculated for a specific group is designed to determine one or more specific conditions of the process. Scaling of the sum can be done in many ways, where zero or negative values ​​represent appropriate conditions. The interpretation / meaning of the calculated sum has become straightforward.

[0031] It should be noted that the interpretation / meaning of the sum depends on each individual group. For example, if the interpretation / meaning of the group represents, for example, paper strength, then a high strength value (high sum) is good, while a low strength value (low sum) is bad. The interpretation of the interpreted value is now different from the example above. When the sum is negative, the group is in a bad state (contributors reduce strength). When the sum is positive, the group is in a good state (contributors increase strength).

[0032] Therefore, the calculated sums are monitored in a specific grouping manner, with each sum indicating the grouping status under discussion.

[0033] The group status can be represented appropriately, such as by "traffic lights": green, yellow, and red. If a group is green, its contribution to the model output is low (i.e., the risk of process instability, disturbances, problems, or poor product quality is low or reduced). A green status indicates the group is in good condition. If a group is yellow, its contribution to the model output is moderate (i.e., the risk of process instability, disturbances, problems, or poor product quality is increased). A yellow status means the group is in acceptable condition or at a reasonable level. A yellow status can be called a warning status. If the status is red, the group's contribution to the model output is high (i.e., the risk of process instability, disturbances, problems, or poor product quality is high). A red status can be called an alarm status. The group status can also be, for example, numerical values ​​(e.g., 0, 1, 2, 3) or text values ​​(good, warning, alarm). The group status can be defined based on a predetermined threshold of the sum of SHAP values ​​(or other interpreted values). The predetermined threshold can depend on the model, the process, and the measurements input to the model. The predetermined threshold can be defined based on historical data processed.

[0034] The input categories can be easily divided into main groups and subgroups. As mentioned above, the classification utilizes a pre-defined database. Figure 4Examples of main groups 41 are shown, namely I, II, III, IV, V, and VI. The vertical axis shows the sum of the SHAP values ​​for each group. We assume the process to be monitored is the papermaking process, and the main groups are Production I, Chemical II, Tower III in Subprocess 1, Tower IV in Subprocess 2, Interruption V, and the other VI. Circles 42, rectangles 43, and crosses 44 show different scenarios, so... Figure 4 Three examples are actually shown. These examples are all based on real-world paper manufacturing data. Circles represent cases with over 500 quality defects in the roll. Rectangles represent cases with 50 to 500 quality defects in the roll, and crosses represent cases with fewer than 50 quality defects in the roll. As you can see, the crosses are relatively close to or equal to zero, indicating that the situation in each group is good. The defects are at an acceptable level. The situation within the rectangles is not so good, as you can see that the sum of the chemical groups is above zero. Therefore, monitoring indicates that something can be done in this group, namely, using the inputs and relating to the process that provides the inputs. These circles indicate that the situation with the chemical groups and production is now worse. Therefore, it can be noted that the hierarchy of sum values ​​can also indicate the severity of the defects. For example, when referencing... Figure 4 When the total value is below the threshold of 0.05, it indicates a good situation, meaning the number of defects is small and will not have a significant impact on the process. A total value between the threshold of 0.05 and 0.15 indicates a problem in the process; the process may be becoming unstable, meaning the risk of process disturbances may be increasing, but no immediate action is needed. Figure 4 In the diagram, a rectangular case indicates that the level is present in Group II, while a circular case indicates that the level is present in Group I. A sum exceeding the threshold of 0.15 indicates that immediate action is required, thus the defect at that level is the most severe. The circular case indicates that the level is present in Group II. The thresholds for the group status are predetermined and can depend on one or more of, for example, the model, the process, and the measurements used as input in the model.

[0035] The main grouping of a process or subprocess can be a chemical grouping (chemically related measurements of the process or subprocess), a production grouping (production-related measurements of the process or subprocess), product quality, tower, and / or any other suitable grouping. Examples of subgroupings within a chemical grouping include dissolved compounds, particles, hydrophobic particles, microbial activity, organic compounds, inorganic compounds, or chemical properties (e.g., pH, temperature, conductivity, any other measurement related to the chemical state of the process) in a specific water, pulp, or filtrate stream during a specific chemical treatment step in the process or subprocess. Examples of production subgroupings include grade, gram weight, amount of product produced, process interruption / stop, volume of treated water, and flow rate of treated water. For example, if the process is a paper or paperboard machine, the chemical subgroups could be: feed pulp quality, feed anionic waste, feed microbial activity, feed hydrophobic contamination, wet-end chemistry, wet-end microbial activity, wet-end anionic waste (amount or concentration of dissolved anionic compounds), wet-end hydrophobic contamination, retention, sizing, waste stream quality(s), hydrophobic contamination in the waste stream, microbial activity in the waste stream, anionic waste in the waste stream, raw water quality, and water recycling quality. The predefined subgroups can depend on the process and the model (the model's target values).

[0036] For example, the process can be a paper, paperboard, or tissue paper manufacturing process, with the main group including one or more of production, chemical, tower, and others.

[0037] Because main groups are relatively large, they may provide the correct direction for corrective action, but subgroups of main groups can provide even more accurate direction. While there are some advantages to using main and subgroups, it is also possible to use grouping without main and subgroups. In such embodiments, the input is simply classified into groups.

[0038] Figure 5 Examples of subgroups 51 are shown, namely A, B, C, D, E, F, G, H, I, J, K, and L. The vertical axis shows the sum of the SHAP values ​​for each subgroup. It is understood here that the process to be monitored is related to... Figure 4 The same papermaking process is described, with subgroups belonging to Chemical Main Group II. These subgroups can include retention, hydrophobic contamination, coagulation, anionic waste, pulp quality, sizing, raw water, wastewater, microbial activity, defoamers, batching temperature, and batching pH. (Circle 52 and...) Figure 4 The case of the circle in the rectangle 53 is related to the case of the circle in the rectangle 53. Figure 4 The case of the rectangle is related to the case of the cross 54. Figure 4The situation is related to the cross shape. It can be seen that the cross shape is in good condition. In the rectangular case, some pulp quality issues arise. Furthermore, in the circular case, there are problems with retention and pulp quality. The level of the sum value can also indicate the severity of defects in the subgroup.

[0039] Furthermore, the subgroup may include one or more of the following: feed pulp quality, feed pulp anionic waste, feed pulp hydrophobic contamination, feed pulp microbial activity, wet-end chemical processes, wet-end microbial activity, wet-end anionic waste (amount or concentration of dissolved anionic compounds), wet-end hydrophobic contamination, retention, sizing, quality of the waste stream, waste stream hydrophobic contamination, waste stream anionic waste, waste stream microbial activity, raw water quality, water recycling quality, tier, basis weight, amount of product produced, process interruption / stop.

[0040] Figure 3 An example of the system of the present invention for monitoring a process is shown. The system includes an ML module 7 for creating an ML model of the process. The ML model utilizes measurement results 4 from process 3 as input 6 and forms a model output 8. The measurement results 4 can be preprocessed 5 before being used as input in the ML model. Preprocessing may include, for example, data merging, data filtering, aligning time formats, modifying metadata, data validation, etc.

[0041] The system also includes a classification module 10 for classifying inputs into main groups and subgroups within those main groups. As previously mentioned, classifying inputs into groups is a preliminary step. It utilizes a predetermined database containing information about how the inputs belong to groups. The knowledge of how to group the inputs is expert knowledge of the process in question. Many case experiences and other process knowledge involve expert knowledge. Expert knowledge can be in the form of forms, diagrams, tables, or other formats used to provide the predetermined database. Thus, for example, the database can be in the form of forms, diagrams, or tables. It can be seen that the database contains expert knowledge. Therefore, the classification module utilizes the predetermined database containing information about how the inputs belong to groups.

[0042] The system also includes a calculation module 11A for calculating the SHAP value of each input, which indicates the contribution of the input to the model output 8, and for calculating the sum of the indication values ​​for each group. The indication values ​​are related to the SHAP values. The calculation module 11A may include different units for these calculations, namely the SHAP calculation unit 9 and the summation unit 11. Furthermore, the system includes a monitoring module 12 for monitoring the calculated sums, each sum indicating the state of the group in question. Note that input 6 is used when calculating the SHAP values; for clarity, a dashed line is shown from input 6 of the ML model to the SHAP module.

[0043] The classification module (10) for classifying inputs into groups can be configured to utilize a predetermined database containing information about how the inputs belong to groups, and the monitoring module (12) can be configured to monitor the calculated sum in a group-specific manner.

[0044] As described above, the indicated value is the SHAP value or the SHAP deviation value. The SHAP deviation value is the difference between the SHAP value and the standard SHAP value. The standard value can be obtained from a successful cycle of the process and can be stored in memory. Therefore, the system may also include a library module 13 to maintain a standard SHAP value that can be used by the SHAP calculation module 9. As mentioned above, other interpreted values ​​can be used instead of the SHAP value.

[0045] As described above, the grouping can be predetermined, and in the embodiments, when the sum is negative or relatively close to zero, the grouping status is good (acceptable) when the number of possible defects is small and the situation is within an acceptable range. When the sum is positive, the grouping status is bad. The sum may also indicate the severity of the situation. The larger the sum, the more severe the situation within the group. Refer to the above... Figure 4 Examples of different levels of the relevant summaries. Monitoring can be real-time when using real-time values ​​as input.

[0046] Figure 6 An example of chemical grouping status as a function of time (x-axis indicates hours) is shown, with predetermined thresholds 61 and 62 as a function of time. In this example, the grouping status is good when the sum of the SHAP values ​​60 (or other interpreted values) is equal to or less than the threshold of zero 62. Therefore, when the grouping status is good, the grouping does not increase the risk of process disturbances or defects in the final product quality of the process. When the sum is greater than zero and equal to or less than another threshold 61 with a value of 0.07, the chemical grouping status is at the warning level. At this level, the risk of process disturbances or defects in the final product quality of the process has increased, and corrective actions should be prepared. When the sum of the SHAP values ​​is greater than 0.07 (i.e., another threshold 61), the chemical grouping status is at the alarm level. When the threshold is at the alarm level, the risk of process disturbances or defects in the final product quality of the process is high. When the grouping status is at the alarm level, corrective actions are required immediately. The final product can be, for example, the final product of an industrial process. The final product can be, for example, paper, paperboard, or tissue paper from the corresponding process, or, for example, purified water from a water treatment plant or facility. Therefore, as can be seen from the example, the state of the group can be defined using at least one threshold for the sum.

[0047] Users can predefine a threshold for the sum of interpreted values, such as the SHAP values ​​of a group or subgroup. The threshold can be used to indicate the status of each group or subgroup and to indicate the need for intervention. The method can also include predetermined intervention recommendations for each group or subgroup. If the sum of the SHAP values ​​of a group or subgroup exceeds a predefined threshold for early warning or alarm purposes, the method provides intervention recommendations, such as increasing or decreasing the chemical dosage or the flow rate of pulp, filtrate, or water, decreasing or increasing the process temperature, decreasing or increasing the pH in the process stream, initiating or discontinuing chemical dosing. The sum of the SHAP values ​​of a group or subgroup can be used for manual or automatic process control, such as controlling chemical dosing, process flow, pH, temperature, etc.

[0048] The status of one or more groups and / or the sum of the interpreted values ​​of one or more groups can be used for automatic or manual process control, process troubleshooting, or process optimization. Controlling and / or optimizing the process may include one or more of the following: the dosage of chemicals, the point of chemical application, and the selection of the types of chemicals used in the process.

[0049] Therefore, the method and system according to the invention may include at least one next step for providing recommendations for corrective measures based on the status of one or more groups.

[0050] Therefore, the methods and systems according to the invention may include at least one next step for controlling, optimizing, or troubleshooting the process based on the status of one or more groups. The control and / or optimization steps according to the invention may include one or more of the following: controlling the dosage of chemicals, the point of chemical application, the interval between chemical applications, the selection of the type of chemicals used in the process, process conditions such as pH, temperature, process flow rate, and process flow delays, such as delays in the flow of pulp, waste paper, or water in towers, tanks, pulpers, basins, or other process equipment.

[0051] This method can also be used to predict the risk of process disturbances, operational and quality problems in the final product, such as the risk in the next few hours or days.

[0052] This invention enables the monitoring of complex processes that are cumbersome when arranged in a known manner. Monitoring can be performed at two levels without affecting existing ML models. Since input variables are categorized into groups or main groups and subgroups, the entities to be monitored can be changed by altering the classification, or even interleaved classification can be used. This means that the way inputs belong to groups can be changed based on expert knowledge. Process experts may notice changes that must be considered during the process, which can be accounted for by changing the grouping of inputs or creating new groupings. Changed or new groupings can be inserted into the database. Therefore, the predetermined database can be updated. Thus, the monitoring method and system of this invention are flexible. Monitoring can be automated.

[0053] Figure 7 The invention also illustrates how to place inputs (variables) into main groups and subgroups. We can consider an ML model using 20 inputs. For example, the interpreted value of the SHAP value is calculated in a known manner and used to interpret the output of the ML model. In addition, the invention features predetermined grouping... Figure 7 In the example, there is one main group and six subgroups. Grouping and variables are... Figure 7 While the information exists in tabular form, in actual solutions, the same information resides in a database in some data format. This information includes the aforementioned expert knowledge. Figure 7 In this table, inputs / variables 2-19 belong to the main group, and the calculated interpreted values ​​of these inputs are summed. The sum indicates a certain condition of the process, such as a general fault condition, or other conditions. Inputs 1 and 20 are not required for this indication.

[0054] Inputs / variables 4-1 belong to subgroup 1, and the calculated interpreted values ​​of these inputs are summed. Furthermore, the sum indicates a certain condition of the process, such as a more specific fault condition, or another more specific condition. Similarly, some inputs belong to other subgroups 2-6 that indicate certain conditions or faults.

[0055] It should be noted that the interpreted values ​​are calculated for all inputs of a known ML model. Furthermore, this invention calculates a specific sum of the interpreted values ​​for the predetermined groups. The advantages of this invention are obvious. Existing interpreted values ​​(such as SHAP values) can be used to indicate specific conditions or faults by utilizing predetermined groups of inputs for the ML model. Since the predetermined groups are based on expert knowledge, they can also be updated if needed. The formation of the predetermined groups is also flexible. It does not require changing one or more ML models, nor does it require creating new ML models.

[0056] Compared to using SHAP values, when using the difference in SHAP values, the monitoring system can better focus on the input variables that indicate problematic events in the process.

[0057] This invention can be implemented in software, using an application-specific integrated circuit (ASIC), or a combination of software and hardware. Module 13, which has standard values, is a memory, which can naturally include both software and hardware. The inputs are measurement results from the process. The ML model and interpreted values ​​are calculated in a known manner. Furthermore, classifying the inputs into predetermined groups and calculating the sum of the interpreted values ​​of the inputs for each group in a group-specific manner forms a multifunctional configuration for monitoring different states of the process, wherein each sum indicates the state of the group in question.

[0058] The device of the present invention can be located in the same place as the process being monitored. However, it is also possible to locate it in another place, which makes remote monitoring of the process possible. For example, measurement data 4 is transmitted via a communication network to the monitor of the present invention, which processes the measurement data and provides process status data as the monitor's output, which can be used for many purposes, such as making recommendations for process adjustments.

[0059] 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 method for monitoring and controlling a process and / or identifying sources of process disturbances or quality defects, the method comprising a preliminary step (21) for creating an ML model of the process, the ML model utilizing measurements from the process as inputs and forming a model output, characterized in that, The method further includes a second preparatory step (22) for grouping the input components, wherein the grouping reflects predetermined quality indicators and / or process conditions. Calculation step (23) is used to calculate an explanatory value for each input in the machine learning process, the explanatory value indicating the contribution of the input to the model output; and to calculate the sum of the indication values ​​for each group, the indication values ​​being related to the explanatory value in the machine learning process, and Monitoring step (24) is used to monitor the calculated sums, each sum indicating the state of the group in question, the state of the group indicating the degree of contribution of the group to the model output; and The process is controlled, optimized, or troubleshooted based on the status of one or more of the groups.

2. The method according to claim 1, characterized in that, The second preparatory step (22) for classifying the input into groups utilizes a predetermined database containing information about how the input belongs to the group, and the monitoring step (24) for monitoring the sum calculated in a group-specific manner.

3. The method according to claim 1 or 2, characterized in that, The machine learning interpretation value is a SHAP value, a value from the LIME method, or a value from the deepLIFT method.

4. The method according to claim 3, characterized in that, The grouping includes one or more main groups and one or more subgroups.

5. The method according to claim 1 or 2, characterized in that, The indication value is the explanatory value of the machine learning or the deviation of the explanatory value, wherein the deviation of the explanatory value is the difference between the explanatory value and the standard explanatory value.

6. The method according to claim 5, characterized in that, The grouping is predefined.

7. The method according to claim 6, characterized in that, The state of the group is defined using at least one threshold for the sum.

8. The method according to claim 7, characterized in that, The group is in good condition when the sum is negative or relatively close to zero, and in bad condition when the sum is positive.

9. The method according to claim 1 or 2, characterized in that, The process is pulp treatment, papermaking process, paperboard manufacturing process, tissue paper manufacturing process, paper machine, pulp mill, tissue paper machine, paperboard machine, water treatment process, wastewater treatment process, raw water treatment process, water reuse process, any industrial water treatment process, municipal water, municipal wastewater treatment process, sludge treatment process, mining process, or oil recovery process.

10. The method according to claim 4, characterized in that, The process is a paper, paperboard, or tissue paper manufacturing process, and the main group includes one or more of the following: production, chemistry, and tower.

11. The method according to claim 10, characterized in that, Subgroups include one or more of the following: feed pulp quality; feed pulp anionic waste; feed pulp hydrophobic contamination; feed pulp microbial activity; wet-end chemical processes; wet-end microbial activity; wet-end anionic waste; wet-end hydrophobic contamination; retention; sizing; waste stream quality; waste stream hydrophobic contamination; waste stream anionic waste; waste stream microbial activity; raw water quality; water recycling quality; grade; basis weight; quantity of product produced; process interruption / stop.

12. The method according to claim 1 or 2, characterized in that, The method includes at least one next step, which is used to provide recommendations for corrective measures based on the state of one or more of the groups.

13. The method according to claim 1, characterized in that, The control and / or optimization steps include one or more of the following: controlling the dosage of chemicals, the point of application of chemicals, the interval of application of chemicals, the selection of the type of chemicals to be used in the process, process conditions including pH and temperature, the flow rate of the process flow, and process flow delays, the process flow delays including pulp, waste paper, or water flow delays in process equipment, the process equipment including towers, tanks, pulpers, or basins.

14. The method according to claim 1 or 2, characterized in that, The inputs include real-time values.

15. A system for monitoring and controlling a process, the system comprising an ML module (7) for creating an ML model of the process, wherein, The ML model uses measurement results (4) from process (1) as input (6) to form model output (8), characterized in that the system further includes a classification module (10) for classifying the input into groups, wherein the grouping reflects predetermined quality indicators and / or process status. A calculation module (11A) is configured to calculate an interpretation value for machine learning for each input, each input indicating the contribution of the input to the model output (8), and to calculate the sum of the indication values ​​for each group, the indication values ​​relating to the interpretation value for machine learning, and A monitoring module (12) is used to monitor the calculated sums, each sum indicating the status of the group under discussion and the control, optimization or troubleshooting of the process based on the status of one or more of the groups, wherein the status of the group indicates the degree of contribution of the group to the model output.

16. The system according to claim 15, characterized in that, The classification module (10) for classifying the input into groups uses a predetermined database containing information about how the input belongs to the group to classify the input into the group, and the monitoring module (12) for monitoring the calculated sum is arranged in a group-specific manner.

17. The system according to claim 16, characterized in that, The machine learning interpretation value is a SHAP value, a value from the LIME method, or a value from the deepLIFT method.

18. The system according to claim 17, characterized in that, The grouping includes a main group and subgroups.

19. The system according to claim 17 or 18, characterized in that, The indicated value is the interpreted value or the deviation of the interpreted value, and the deviation of the interpreted value is the difference between the interpreted value and the standard interpreted value.

20. The system according to claim 19, characterized in that, The system includes a library module (13) for maintaining the standard interpretation values.

21. The system according to claim 19, characterized in that, The grouping is predefined.

22. The system according to claim 21, characterized in that, The group is in a good state when the sum is negative or relatively close to zero, and in a poor state when the sum is positive.

23. The system according to claim 21, characterized in that, The process to be monitored is the paper, paperboard, or tissue manufacturing process, and the main group includes one or more of the production, chemical processes, and towers.

24. The system according to claim 23, characterized in that, Subgroups include one or more of the following: feed pulp quality; anionic waste in feed pulp; hydrophobic contamination in feed pulp; wet-end chemical processes; microbial activity in wet-end; anionic waste in wet-end; hydrophobic contamination in wet-end; retention; sizing; quality of waste stream; hydrophobic contamination in waste stream; anionic waste in waste stream; raw water quality; water recycling quality; grade; basis weight; quantity of product produced; process interruption / stop.

25. The system according to claim 15, characterized in that, The input is a real-time value.

Images