Industrial Systems

Abstract

One aspect of this disclosure relates to determining actions to improve the operation of an industrial system. Reference data representing one or more states of the industrial system is obtained, and it is determined whether a deviation exists in the reference data. If a deviation exists, the impact of the deviation on the industrial system is evaluated, and based on the evaluated impact of the deviation, one or more mitigation control actions are selected from a plurality of predetermined control actions. A representation of the impact of performing one or more mitigation control actions is output to a user interface, and the user of the user interface can select an operation to initiate one or more mitigation control actions to modify one or more processes of the industrial system.

Description

Technical Field

[0001] This disclosure relates to the operation of supported industrial systems. In particular, but not limited to, this disclosure relates to providing operation support to operators of industrial systems. In particular, but not limited to, this disclosure relates to determining measures for improving the operation of an industrial system based on an evaluation of the impact of deviations detected in data related to the industrial system.

Background Art

[0002] Process operators in the food and beverage industry have conventionally equipped with conventional SCADA systems (supervisory control and data acquisition systems) to operate industrial systems such as processing plants. These systems have very detailed overview information of the plant, together with hundreds of sensor values, actuators, and alarms. Database-level local controllers and sequence processing systems (e.g., programmable logic control (PLC) unit systems) handle the basic operations of the process. Therefore, the role of the operator is limited to reacting to and corresponding to local events / alerts, or otherwise, the operator executes the assigned tasks depending on the PLC system.

Summary of the Invention

Problems to be Solved by the Invention

[0003] The challenge described above is that PLC systems are often programmed to a fixed set of parameters, regardless of process or upstream variations (e.g., raw material quality and / or quantity, seasonal changes in the surrounding environment, changes in utility supply to the process, etc.). While SCADA systems display these variations, the complexity of the information provided to operators exceeds human comprehension. Correlations and impact analyses are often overlooked, and operators are unable to take all the latest corrective actions in real time to optimize operations for productivity, quality, and production sustainability. Decision support systems are applied to industrial systems such as plants in the food and beverage industry, including breweries and liquid dairy processes, to help operators perform such real-time optimization operations.

[0004] This disclosure is intended to collect various types of data from various sections of the process, monitor changes in key performance indicators, predict future changes, assess the impact of deviations, and provide operators with a definitive set of information to propose corrective actions. [Means for solving the problem]

[0005] This disclosure relates to a method for predicting the future state of an industrial system and providing operational recommendations for modifying one or more processes in the industrial system.

[0006] One aspect of this disclosure relates to a method implemented in a computer for determining actions to improve the operation of an industrial system. The method includes obtaining reference data representing one or more states of the industrial system, the reference data being associated with data collected from one or more sensors of the industrial system. It is determined whether there is a deviation in the reference data that exceeds a predetermined threshold, the predetermined threshold being associated with a process of the industrial system. If a deviation exceeding the predetermined threshold exists, the impact of the deviation on one or more outputs of the industrial system is evaluated based on the difference between the performance of the industrial system with the deviation at a first time point and the predicted performance of the industrial system if there is no deviation at the first time point. The method further includes selecting one or more mitigation control actions from a plurality of predetermined control actions, each predetermined control action representing a predetermined operation to modify one or more processes of the industrial system, and the selection includes identifying a problem and the associated control action. The problem is identified from a plurality of predetermined problems as causing the deviation, based on the impact of the evaluated deviation. Based on the identified problem causing the deviation, one or more control actions are identified from a plurality of predetermined control actions, and the identified one or more control actions are selected to mitigate the impact of the deviation on one or more outputs of the industrial system. Subsequently, the impact of each mitigation control action on one or more outputs of the industrial system is calculated, and a representation of the impact of one or more mitigation control actions is output to the user interface. This representation allows the user of the user interface to select an operation to initiate one or more mitigation control actions to modify one or more processes of the industrial system.

[0007] Beneficial, complex information related to industrial systems is refined into information most relevant to production operations. Targeted communication of relevant integrated data improves operator decision-making, thereby increasing the efficiency of the industrial system. As a result, users of the user interface, such as operators of the industrial system, can quickly understand the current or future critical impact on one or more processes in the industrial system. Based on up-to-date information on the processes of the industrial system, operators can more accurately and confidently decide whether to implement recommended corrective actions to optimize the processes of the industrial system. Such a process allows small inefficiencies to be detected and corrected or overcome earlier than usual, and in some cases before larger inefficiencies occur.

[0008] In another aspect of this disclosure, this disclosure relates to a computer-implemented method for predicting the system performance of an industrial system. The method includes obtaining a first state value of the system from a trained machine learning model, which predicts the performance of the industrial system in a second period based on the performance of the industrial system in a first period. The performance of the industrial system in the first period is determined using data collected from a sensor set of the industrial system during the first period. The first state value of the system is associated with the predicted performance of the industrial system in the second period and indicates the future state of the industrial system. Real-time system information is also obtained, which indicates the current state of the industrial system and includes data collected by a sensor set of the industrial system. A second state value is determined using a predictive model configured to predict the performance of the industrial system in a second period based on the real-time system information and the first state value of the system. The method further includes calculating a third state value of the system using data collected from a sensor set of the industrial system during the second period, which is associated with the performance of the industrial system in the second period and represents the state of the industrial system in the second period. Subsequently, the trained machine learning model is adjusted to reduce the difference between the second state value determined by the predictive model and the third state value of the system.

[0009] Beneficially, this process combines the results of trained machine learning models and predictive models to generate initial predictions. These initial predictions are then refined to account for real-time data relevant to the predictions. Furthermore, the trained machine learning is kept up-to-date by incorporating a continuous learning loop, resulting in more accurate and reliable subsequent predictions. Additionally, incorporating predictive models allows for the incorporation of real-time fluctuations related to the relevant processes into the predictions, further aligning them with the state of the industrial system at the time of prediction. These predictions can be used by industrial system operators or optimization systems to determine actions to improve or maintain the efficiency of the industrial system before inefficiencies impact the system's processes.

[0010] In another aspect of this disclosure, the disclosure relates to a computer-implemented method for automated monitoring of an industrial system and for generating operational recommendations for one or more processes of the industrial system thereafter. The method includes obtaining a first state value from a trained machine learning model, which predicts the performance of the industrial system in a second period based on the performance of the industrial system in a first period, where the performance of the industrial system in the first period is determined using data collected from a sensor set of the industrial system during the first period. The first state value of the system is associated with the predicted performance of the industrial system in the second period and indicates the future state of the industrial system. Real-time system information is also obtained, which indicates the current state of the industrial system and includes data collected by the sensor set of the industrial system. Next, a second state value of the system is determined using a predictive model configured to predict the performance of the industrial system in a second period based on the first state value of the system and the real-time system information.

[0011] Using a second state value, it is determined whether there is a deviation greater than a predetermined threshold in the predicted performance of the industrial system during the second period. Here, the predetermined threshold is associated with a process in the industrial system. If a deviation greater than the predetermined threshold exists, this method includes evaluating the impact of the deviation on one or more outputs of the industrial system based on the difference between the predicted performance of the industrial system during the second period and the predicted performance of the industrial system when no deviation exists. One or more mitigation control actions are selected from a plurality of predetermined control actions, each predetermined control action representing a predetermined operation to modify one or more processes in the industrial system. The selection includes evaluating the difference between the real-time performance of the industrial system and the performance of the industrial system during the first period, and determining a problem from a plurality of predetermined problems that cause the deviation based on the evaluated difference. Then, one or more control actions are identified from the plurality of predetermined control actions based on the problem causing the deviation. The identified one or more control actions mitigate the difference between the predicted performance of the industrial system during the second period and the real-time performance of the industrial system.

[0012] One or more mitigation control actions are prioritized based on the calculation of the impact each action has on the industrial system. Furthermore, a display of the impact of one or more mitigation control actions on the industrial system is output to the user interface, allowing the user to select an action to initiate one or more mitigation control actions that modify one or more processes in the industrial system.

[0013] In this embodiment, the predicted state of the industrial system is used to determine operations to improve or maintain the efficiency of the industrial system. Deviations in the predicted data that may not be detectable by human operators can be used to identify current or future problems in the industrial system. Such problems may affect the performance of the industrial system. Impact calculations provide operators with relevant information so that they can decide whether to take recommended actions to overcome future problems. Such automated monitoring and operational support can result in a highly efficient industrial system optimized based on real-time fluctuations and the potential impact of changing trends.

[0014] Additional aspects and embodiments of the System are disclosed, and these aspects and embodiments should not be construed as limiting this disclosure.

[0015] Refer to the accompanying drawings for a more comprehensive understanding of this disclosure. In the drawings, similar elements are numbered similarly. These drawings are for illustrative purposes only and should not be construed as limiting this disclosure. [Brief explanation of the drawing]

[0016] [Figure 1] Figure 1 shows a flowchart illustrating how to determine actions to improve the operation of an industrial system. [Figure 2] Figure 2 shows a high-level system architecture diagram relating to this disclosure for determining actions to improve the operation of an industrial system. [Figure 3] Figure 3 shows an example of how the impact of a problem on the performance of an industrial system is displayed in the user interface of the industrial system. [Figure 4] Figure 4 shows a flowchart illustrating how to select one or more mitigation control actions based on their impact on the performance of the industrial system. [Figure 5]Figure 5 shows a flowchart illustrating a method for improving the selection of one or more relaxation control actions based on feedback data. [Figure 6] Figure 6 shows a flowchart illustrating a method for predicting the system performance of an industrial system. [Figure 7] Figure 7 shows a high-level system architecture diagram according to the present disclosure for predicting the system performance of an industrial system. [Figure 8] Figure 8 shows a flowchart illustrating a method for updating a trained machine learning model configured to predict the performance of an industrial system based on data collected from a series of sensors of the industrial system. [Figure 9] Figure 9 shows a flowchart illustrating a method for using quality labels when updating a trained machine learning model configured to predict the performance of an industrial system. [Figure 10A-10B] Figures 10A - 10B show a flowchart illustrating a method for automatic monitoring of an industrial system and generating operation recommendations for the industrial system. [Figure 11] Figure 11 shows a high-level system architecture diagram according to the present disclosure for automatic monitoring of an industrial system and generating operation recommendations for the industrial system. [Figure 12] Figure 12 shows an example of a computing system for operating any one or a combination of the methods disclosed herein. **DETAILED DESCRIPTION OF THE INVENTION**

[0017] In this disclosure, references to singular items should be understood to include plural items unless explicitly stated otherwise or clear from the context, and vice versa. Grammatical conjunctions are intended to represent any disjunctive and conjunctive combinations of clauses, sentences, words, etc., unless explicitly stated otherwise or clear from the context. Thus, the term "or" should generally be understood to mean "and / or", etc. The use of all examples or illustrative words provided herein (such as "for example", "such as", "including", etc.) is only intended to more clearly explain the embodiments and is not intended to limit the scope of the embodiments or the claims.

[0018] One of ordinary skill in the art will understand that the systems and methods of this disclosure are not limited to a single programming language or paradigm. In fact, the systems and methods of this disclosure are applicable to any suitable programming language or environment, including but not limited to Java, C, C++, any suitable assembly language, Python, C#, script language code (such as JavaScript, Ruby, PHP, etc.).

[0019] Some embodiments described herein may relate to computer storage products comprising a temporary or non-temporary computer-readable medium (also called a temporary or non-temporary processor-readable medium) having instructions or computer code for performing various computer implementation processes. A computer-readable medium (or processor-readable medium) is non-temporary in the sense that it does not contain the temporary propagating signal itself (e.g., propagating electromagnetic waves that carry information on a transmission medium such as space or a cable). The medium and computer code (also called code) are designed and constructed for a specific purpose. Examples of non-temporary computer-readable media include, but are not limited to, magnetic storage media such as hard disks, floppy disks, and magnetic tapes; optical storage media such as compact disks / digital video discs (CDs / DVDs), compact disk read-only memory (CD-ROMs), and holographic devices; magneto-optical storage media such as optical disks; carrier signal processing modules; and hardware devices specifically configured to store and execute program code, such as application-specific integrated circuits (ASICs), programmable logic devices (PLDs), read-only memory (ROMs), and random access memory (RAM) devices. Other embodiments described herein relate to temporary computer program products, which include, for example, instructions and / or computer code described herein.

[0020] Some embodiments and / or methods described herein can be implemented by software (running on hardware), hardware, or a combination thereof. Hardware modules include, for example, general-purpose processors, field-programmable gate arrays (FPGAs), and / or application-specific integrated circuits (ASICs). Software modules (running on hardware) can be expressed in a variety of software languages ​​(e.g., computer code), including C, C++, Java, Ruby, Visual Basic, Python, and other object-oriented, procedural, or other programming languages ​​and development tools. Examples of computer code include, but are not limited to, microcode or microinstructions, machine code instructions generated by a compiler, code used to generate web services, and files containing high-level instructions executed by a computer using an interpreter. For example, embodiments can be implemented using imperative programming languages ​​(e.g., C, Fortran, etc.), functional programming languages ​​(e.g., Haskell, Erlang, etc.), logic programming languages ​​(e.g., Prolog), object-oriented programming languages ​​(e.g., Java, C++, etc.), or other suitable programming languages ​​and / or development tools. Other examples of computer code include, but are not limited to, control signals, encryption codes, and compression codes.

[0021] Figure 1-5 relates to operational decision support, in which operational recommendations are automatically provided to the user interface, allowing them to select an action to modify one or more processes in the industrial system. Figure 6-9 relates to predicting the future performance of the industrial system. Figure 10A-11 relates to applying the predicted future performance of the industrial system in Figure 6-9 to enable the operational decision support in Figure 1-5. The computing system 1200 in Figure 12 is an exemplary system for implementing any of the methods described herein, but those skilled in the art will understand that alternative systems may also be used.

[0022] The industrial system in Figure 1-11 is a processing plant configured to perform at least one process, such as a chemical or physical process. The processing plant comprises at least one component, such as machinery, necessary to perform the process. The industrial system further comprises one or more sensors, which collect data related to the processes and / or components of the industrial system. The processes of the industrial system yield one or more inputs to the industrial system, which are used to produce one or more outputs of the industrial system. The industrial system is preferably monitored and / or operated at least partially by a skilled person, such as an operator. The operator is responsible for performing operations that result in process changes to improve efficiency, output quality, output volume, plant sustainability, or other performance indicators.

[0023] Monitoring of industrial systems by automated systems or operators involves determining or acquiring plant status values. Status values ​​are associated with the efficiency, output quality, output volume, plant sustainability, or other performance indicators of the industrial system. Status values ​​indicate the state of the industrial system at a specific point in time or over a specified period and are associated with the performance of the industrial system. For example, one status value might be directly obtained from data collected from one or more sensors in the industrial system, such as a status value associated with the quality of output from the industrial system over a predetermined period (e.g., protein density in the output). Another status value might be calculated from data collected from one or more sensors in the industrial system, such as the progress of the industrial system's process at a given point in time (e.g., 30% complete). One or more status values ​​are used as indicators of whether the industrial system's processes and / or components are functioning optimally, such as whether the process is efficient, whether the components are in good condition, whether the output yield is high, or whether other performance indicators are optimized.

[0024] One or more sensors that make up an industrial system typically collect vast amounts of data during system operation. Traditional analysis and monitoring of such complex data collection is time-consuming, and the operator's role is limited to reacting to local events detected or analyzed from the collected data based on process and / or static boundary conditions associated with the component, such as safety operating constraints assigned to the component by the component manufacturer.

[0025] The decision support methods and / or systems detailed in Figures 1-5 below assist operators in optimizing the performance of industrial systems by providing information and recommended actions related to performance optimization. This decision support is based on evaluating information, including one or more state values ​​and / or real-time data, obtained from one or more sensors in the industrial system, which are typically presented directly to the operator or obtained by a PLC system. Determining whether the information is relevant to the operator is a crucial step in decision support, and this can be achieved by detecting deviations in the information. This provides an objective measure for determining the relevance of the information.

[0026] A deviation is the difference between information associated with the state of an industrial system and its expected value, which can be based on historical or predicted trends based on multiple data points, or on a single data point associated with a specific condition and / or value (i.e., a deviation corresponds to the difference between one or more values ​​in reference data and one or more expected values). In other words, deviations are detected by comparing reference data with the expected value of the reference data (e.g., an expected value based on historical or future state data). Since the information associated with an industrial system has natural variations due to fluctuations such as ambient conditions and input quality, detecting the presence or absence of a deviation involves determining whether the deviation is greater than a given threshold. A given threshold can be dynamic, such as based on the standard deviation calculated from historical or predicted future data, or static, such as based on the specifications of components within the industrial system.

[0027] In an example of multiple data points, information associated with an industrial system includes state values ​​related to the progress of processes within the industrial system. The progress is calculated using data collected in real time from one or more sensors, providing an indicator of the industrial system's state at the current point in time.

[0028] For example, if the calculated state value at the current point in time is 82%, this value is compared to the past state value at a previous point in time. In this example, if the calculated state value follows the trend present in past state values, the calculated state value would be 90%, resulting in an 8% deviation. A predetermined threshold for this state value is based on the historical fluctuations of progress across multiple historical process cycles, resulting in a predetermined threshold of + / - 5%. Since an 8% deviation exceeds the predetermined threshold, it is determined that there is a deviation in the data. Information associated with the state of an industrial system is not limited to current or historical data; deviations can be detected from information associated with the future state of the industrial system.

[0029] When a deviation exceeding a predetermined threshold is detected, the impact of that deviation is calculated. The impact of a deviation is the potential effect that the deviation would have on the system if it continued to function without intervention. For example, the impact of a deviation may be a decrease in output or an increase in energy consumption. Therefore, the impact of a deviation is also called the effect of the deviation, the possible effect, the outcome, or the result / consequence. If there is no impact associated with a deviation, that deviation may be irrelevant to the operator.

[0030] Furthermore, the impact of the deviation helps in evaluating the causes of the deviation, such as component blockages. By evaluating the causes of the deviation, measures to mitigate or resolve the problem can be determined, and the mitigation measures are linked to the corresponding problems that need to be mitigated. The link between the problem and the measures is based on prior knowledge of the processes and components of the industrial system, such as process know-how and the configuration of the industrial system. Therefore, the prior knowledge is dependent on the specific industrial system and may include information obtained from industrial system experts such as operators and component manufacturers.

[0031] The action calculated to have the greatest impact on the problem causing the deviation is selected and presented to the operator along with the calculated impact. The selected mitigation action is the most appropriate course of action for the operator, and the operator does not need to understand and analyze all the information related to the industrial system used during decision support. Furthermore, the impact calculation is the most appropriate analysis the operator needs when deciding whether to implement one of the selected mitigation control actions. This is especially important in scenarios where multiple deviations are detected, multiple mitigation control actions are selected, and the operator needs to decide which mitigation control action to implement based on experience and technical knowledge when optimizing the performance of the industrial system.

[0032] For example, components in an industrial system have an optimal performance range. Exceeding this range degrades component performance, leading to decreased process efficiency. In this example, deviations from information related to the predicted future state of the system are detected based on comparison with historical data associated with the component. The impact of the deviation is assessed as exceeding the optimal performance range in two process cycles. The problem is determined to be an increase in the amount of by-products related to future process cycles due to changes in the quality of input resources. Control actions related to this problem include mitigation actions such as adjusting the configuration of input resources, slowing down the process, and rescheduling future process cycles. The impact of the selected mitigation action, such as decreased process efficiency due to the input configuration change and increased time required to complete future process cycles, is calculated. Both the selected mitigation action and its impact are provided to the operator, who then decides to take mitigation action, such as rescheduling future process cycles.

[0033] Traditional industrial systems lack correlation and impact analysis, preventing operators from taking the latest corrective actions to optimize them. However, by introducing decision support methods and systems, operators are provided with the most relevant impact analyses in addition to pre-determined corrective actions. Operators can quickly, accurately, and confidently decide whether to implement recommended corrective actions and optimize industrial systems based on real-time and forecasting information.

[0034] Figure 1 shows a flowchart illustrating how to determine measures to improve the operation of an industrial system. Specifically, Figure 1 shows a computer implementation method 100 compatible with processes 400, 500, 600, 800, 900, and / or 1000. Figure 1 illustrates Method 100, a process that assists the operator of an industrial system, in which corrective actions are proposed to the operator to minimize the trend of detected deviations in the performance of the industrial system. Implementing these proposed corrective actions is aimed at mitigating adverse impacts on the operation of the industrial system related to productivity, performance quality, output quality, sustainability, efficiency, and / or other performance aspects.

[0035] For each individual state value representing the performance of an industrial system, predictive data, historical data, and / or real-time data from various data sources are evaluated to assess the potential deviation of the state value compared to the expected state value. Method 100 can be customized based on individual state values ​​and can reliably identify deviations using different combinations of predictive data, historical data, and / or real-time data. Once the deviation is identified, its impact is calculated, and the relevant problems and corrective actions can be correctly identified. These corrective actions are presented to users of the industrial system's user interface, such as operators, who can review the conclusions and recommended actions and decide whether to implement corrective actions to mitigate the impact of the deviation. An example of a system for implementing Method 100 is shown in Figure 2.

[0036] Method 100 includes steps 102, 104, 106, 108, 110, and 112. Step 102 includes obtaining reference data for a first period. Step 104 includes determining whether a deviation exists in the reference data. Step 106 includes evaluating the impact of the deviation on the industrial system. Step 108 includes selecting one or more mitigation control actions from a plurality of predetermined control actions. Step 110 includes calculating the impact of performing each control action. Step 112 includes outputting a display of the impact.

[0037] Step 102 involves obtaining reference data that represents one or more states of the industrial system. The reference data is data related to the active processes of the industrial system and is actively monitored to ensure the industrial system operates in an optimal state. The reference data is associated with data collected from one or more sensors of the industrial system, which will be described in more detail below. Once obtained, the reference data can be combined, refined, or otherwise processed using standard data processing methods that enable the reference data to be evaluated in a more reliable manner in Step 104, such as data joining techniques including many-to-many fully joined computational methods to form a multidimensional dataset, internal join computational methods that retain only the intersections of the joined data, or data cleaning techniques such as filtering.

[0038] For example, the reference data may include real-time data collected from one or more sensors in an industrial system. In another example, the reference data may include historical state values ​​associated with historical data accessed from a database or other storage medium. In yet another example, the reference data may include predicted future state values ​​based on current or historical state values, or predicted future state values ​​based on the output of a trained machine learning model configured to predict the future performance of an industrial system, such as the trained machine learning model of Method 600, which is described in detail below.

[0039] The reference data includes at least one of the examples above, and the composition of the reference data is tailored to specific state values ​​being monitored to determine whether the industrial system is operating optimally. State values ​​indicate the state and / or performance of the industrial system and include values ​​associated with the progress of processes occurring in the industrial system, such as progress completion rate; values ​​associated with the availability of the industrial system, such as batch duration and phase / cycle duration; values ​​related to the quality performance of the industrial system, such as the quality of output to the industrial system, average temperature, mass and / or energy amount for a particular process; values ​​related to the productivity of the industrial system, such as process speed, safety indicators, or performance indicators of equipment and / or components; and the efficiency of the industrial system, such as energy recovery rate and percentage, or resource consumption.

[0040] The reference data optionally includes the reference state values ​​described above, and / or data collected from one or more sensors in the industrial system used to calculate each state value. For example, the reference data includes state values ​​related to a particular monitored state value, or data used to calculate or obtain the included state values. This provides reference data that includes all potential variables that may affect the monitored state value, improving the accuracy and reliability of deviation detection, as described in step 104 below.

[0041] One or more sensors include physical sensors, motion sensors, proximity sensors, time-based sensors, virtual sensors, detectors, and / or transducers. Furthermore, reference data includes constants such as time-independent values, variables such as time-dependent values, or a combination of the two.

[0042] Step 104 includes determining whether a deviation exists in the reference data, for example, determining whether a deviation exists that exceeds a predetermined threshold in a comparison between the reference data and the expected value of the reference data, where the predetermined threshold is associated with a process in an industrial system. A deviation is a variation in information associated with the state of the industrial system when compared to an expected value, where the expected value can be based on a historical or predictive trend based on multiple data points, or on a single data point associated with a specific condition and / or value. For example, a deviation is an outlier such as a global outlier, a contextual outlier, or a collective outlier, or a value that exceeds the range of the expected value.

[0043] Deviation detection is based, for example, on a data sequence or real-time data measurement within a predetermined window of a certain number of recent measurements, or within a predetermined time window. Depending on the type of deviation, different predetermined thresholds are required to determine whether a deviation exists or not, such as thresholds based on processed measurements, such as estimated differences in reference data from the measurements. For example, the thresholds are based on process expertise such as recipe parameters, instrument specifications, and / or historical performance.

[0044] In one example where the reference data includes state values ​​associated with components within an industrial system, the predetermined threshold used for deviation detection is a constant value based on the safe and efficient operating range of the components within the industrial system. In another example where the reference data includes state values ​​associated with components within an industrial system, the predetermined threshold used is variable, and deviation detection is based on historical and / or predicted data trends. Deviation detection enables the detection of deviations specific to particular components and processes, even when the reference data is within a standard safe range, and allows for early detection of deviations even when the overall processing time is within standard operating time, such as batch processing that takes longer than expected for a particular component or longer than historically for that particular component. Therefore, determining whether deviations exist enables process and component-specific early detection, improving the efficiency of the industrial system.

[0045] Step 106 includes evaluating the impact of the deviation on the industrial system. For example, if a deviation exceeding a predetermined threshold exists, as determined in Step 104, the impact of the deviation on one or more outputs of the industrial system is evaluated based on the difference between the performance of the industrial system at a first time point in time when the deviation exists and the predicted performance of the industrial system at a first time point in time when the deviation does not exist. The first time point is either the present or a future time point.

[0046] If the first time point is a future time point, the predictive performance of the industrial system at the future time point in which the deviation exists is based on the predicted future state of the industrial system, which is based on the state of the industrial system contained in the reference data. Predictive performance is obtained using machine learning processes such as Method 600 described later, or mathematical processes such as averaging and trend prediction. If the first time point is the present time, the performance of the industrial system is represented by the current state of the industrial system associated with data collected in real time or during the current period from one or more sensors of the industrial system.

[0047] The predictive performance of the industrial system at a first point in time is a mathematical projection of the industrial system's state, based on the current state of the industrial system if the deviation is constant, and / or on the historical state of the industrial system if the deviation is time-dependent (e.g., increases over time). This mathematical projection is determined using a standard mathematical process such as the mean over a given time window, or a trend prediction method such as linear regression. Alternatively, the current and / or past performance of the industrial system may be used, or a virtual performance of the industrial system generated based on theoretically optimal performance and data collected from one or more sensors of the industrial system may be used.

[0048] Step 108 includes selecting one or more mitigation control actions from a plurality of predetermined control actions, each predetermined control action representing a predetermined operation for modifying one or more processes of an industrial system, and the selection includes identifying a problem, and then identifying one or more control actions based on the identified problem. Mitigation actions are a set of predetermined actions performed by the system to address the identified problem.

[0049] From a set of predetermined potential problems, the problem causing the deviation is identified based on an assessment of the impact of the deviation. Furthermore, based on the identified problem causing the deviation, one or more control actions are identified from a set of predetermined control actions. The identified one or more control actions are selected to mitigate the impact of the deviation on one or more outputs of the industrial system. The set of predetermined problems, the impacts arising from the set of predetermined problems, and the set of predetermined control actions are based on the process know-how and configuration of the industrial system and are stored in a corrective action database. The corrective action database is a database associated with the industrial system and may be local to the industrial system or not, such as an edge-based approach or a cloud database. Each problem is linked to an impact. Each problem is also linked to a control action in a one-to-one, one-to-many, or many-to-one relationship. The corrective action database is described in more detail below, referring to the corrective action database 220 of the operational support system 200.

[0050] For example, a heat recovery unit in an industrial system has a limited capacity for the amount of heat it can store. If more heat is recovered than is used, the unit will eventually reach its maximum storage capacity. In this case, the deviation detection in step 104 determines that a deviation exists in the reference data, and the evaluation in step 106 assesses, based on historical trends, that the deviation is caused by more heat being recovered than usual. The mitigation control action selected in step 108 is to reschedule production to maximize the potential of heat recovery using thermal storage, based on the calculated difference in energy efficiency during operation when the heat recovery unit is at maximum capacity and when it is not at maximum capacity.

[0051] Step 110 includes calculating the impact of performing each mitigation control action, for example, calculating the impact of performing each mitigation control action on one or more outputs of the industrial system. The impact calculation considers one or more types of impacts based on the potential consequences of the impacts related to the industrial system. The impact types and / or impact values ​​may include production measures, operational measures, environmental measures, safety measures, or efficiency measures. Continuing with the heat recovery unit example described above, the impact of rescheduling production to utilize the stored heat is calculated, which includes cascading effects such as the rescheduling of subsequent processes and the energy used by the system during idle time.

[0052] For example, the calculated impacts include the impacts on output related to changes in process time, changes in output quality and / or quantity, changes in resources required to maintain output quality and / or quantity, changes in sustainability, changes in energy consumption, and / or changes in environmental footprint. If the calculated impacts include multiple impact types, the total impact can be calculated by applying weights to each impact type. Weighting the impact types yields adjusted final impacts, such as adjusted total impacts for each deviation type, each process in the industrial system, each component in the industrial system, each operation of focus in the industrial system (e.g., water conservation during drought), and / or adjusted total impacts for each industrial system. Optionally, one or more selected control actions may be prioritized or reordered based on their overall impact, for example, by using the process of Method 400.

[0053] Step 112 includes outputting an impact display, for example, by outputting an impact display to the user interface of one or more mitigation control measures, so that the user of the user interface can select an operation to initiate one or more mitigation control measures to modify one or more processes in an industrial system, such as by using a visual display based on the example in Figure 3.

[0054] For example, the user interface displays the overall impact and / or at least one calculated type of impact (e.g., deviation, problem, and / or the type of impact most relevant to one or more mitigation control measures). The user of the user interface uses technical know-how related to the operation of the industrial system to decide whether to initiate one or more mitigation control measures, take no action, or initiate another measure to modify one or more processes of the industrial system with the aim of mitigating the impact of the deviation on the performance of the industrial system. Preferably, the user's decision is saved and used to improve the reliability and / or accuracy of step 104, step 106, step 108, or step 110, for example, by using one of the steps of method 500.

[0055] If control actions are prioritized, an operation that modifies one or more processes in an industrial system is the highest-priority mitigation control action, and if selected, it brings about a change in one or more processes in the industrial system, thereby mitigating the impact of the deviation on the industrial system. Alternatively, an operation that modifies one or more processes in an industrial system may not be the highest-priority mitigation control action.

[0056] Optionally, step 106 may be performed in combination with step 110 to assess the effects of the deviation, which includes calculating the effects of the deviation. Alternatively, or in addition, step 110 may be performed in combination with step 106 to calculate the effects of performing each relaxation control action, which may include calculating the effects of the deviation based on the assessment of the effects of the deviation.

[0057] Preferably, method 100 is performed for one unique state value. For example, for multiple state values, method 100 is repeated in series or in parallel, and a predetermined threshold in step 104 depends on a particular state value. An example of using the steps of method 100 for a single state value is detailed below. Alternatively, method 100 can be performed for multiple state values, such as when a set of state values ​​are associated, and a single deviation classifier can be used for each state value in the set of state values.

[0058] For a single state value, step 102 involves obtaining predicted state values, historical state values, and additional data from multiple data sources for use in evaluating possible deviations. Depending on the single state value, various types of evaluations are selected, such as multiple trained machine learning models. In this case, the input is a combination of real-time data sources, historical data, and estimated data. The result of such an analysis could be, for example, "inefficient heating rate in the wort kettle." In step 104, deviations are detected based on a data sequence or real-time measurements (within the most recent measurement window, or measurements within the last few minutes). Depending on the type of deviation, individual deviation classifiers require different measurements or features. Features are pre-processed measurements, such as estimated variance from measured values. Optionally, simultaneously, a graphical representation of the symptoms is created, as shown in Figure 3. Such a representation could be, for example, a mapping of time series, batch performance, or other relevant performance data showing the deviations.

[0059] When a deviation is triggered, in step 106 the impact of the affected single state value is calculated, and in step 108 the problem is identified. This analysis helps the plant operator determine whether it is better to continue operating according to the current plan or whether there is a better method from an economic standpoint. The impact calculation can be based on efficiency calculations of how much the operating time and energy costs will increase or how much the production rate will decrease compared to a viable alternative (e.g., a clean-in-place (CIP) process). A CIP process involves stopping one or more relevant components, processes, and / or pieces of equipment in an industrial system to clean the components (e.g., cleaning filters without removing them). Such knowledge is based, for example, on equipment specifications, product costs, resource costs, and the state of the system (e.g., degree of fouling). For example, if the degree of fouling is constant, the equipment will use more power. The average power allocation for this particular process in an industrial system is 20% of the total allocation. The allocation of additional operations per hour is 5% per hour. The average additional allocation required for CIP is 15%. Therefore, the effect of the deviation is that more power is used, and performing CIP is considered the most efficient action for operation exceeding 3 hours.

[0060] Once a problem is identified, in step 108, corrective actions are identified based on process know-how and plant configuration, such as using the corrective action database 220, which is described in detail below with reference to Figure 2. The impact of the action execution is also determined, such as using step 110. Multiple control actions are stored in the corrective action database, and the action that has the greatest impact on correcting the problem is selected. The result of such an action identifier is, for example, "Clean the heat exchanger if possible." The appropriate set of actions for a particular problem is usually known in advance. In this case, the possible set of actions is known based on the detected deviation, i.e., on a rule basis. In the case of heat exchanger fouling described above, it is known to skilled people, such as operators of industrial systems, that it is recommended to perform CIP on the unit if the deviation of the heat transfer coefficient is too low. Another case is when there is a buffer tank between three lines (P, C1, and C2). One line (P) is a producer that fills the buffer tank, and the other two lines (C1 and C2) perform downstream processing of the product that removes the product from the buffer tank (e.g., UHT processing). If the input and output flows of a tank are not balanced, the tank will either fill up or become empty. Three appropriate actions when there is a risk of the tank becoming full are: a. reschedule the process to increase the product rate of the downstream process (e.g., C1) if possible; b. reschedule the process to increase the product rate of the downstream process (e.g., C2) if possible; or c. readjust the production schedule to lower the production rate of the upstream process P. Another example is a heat recovery unit with limited heat storage capacity. When the amount of heat recovered exceeds the amount of heat used, the unit reaches its maximum storage capacity. In this case, the solution is, as in the former example, to readjust the production schedule to maximize the potential of heat recovery by utilizing the stored heat. It is highly desirable that steps 108 and / or 110 be process know-how-based analyses based on operational rule-based models.

[0061] In step 112, the above conclusions are applied to each deviation state value. Optionally, the results from the previous step may be input into a prioritization algorithm that ranks the issues the operator should focus on. The prioritization is a plant objective-based sort that takes into account weighted combinations of production, environment, safety, and other measures. For example, suppose there are three result sets "C" for observed problems. C1: The first value based on production losses due to longer batches rather than time-optimized batches. C2: The second highest value, based on product quality adjustments due to reduced yield. This signifies the overuse of expensive raw materials. C3: A third metric based on sustainability, energy consumption, and environmental footprint. For each result, apply the weighting index "I" to calculate the total cost C. T Therefore, the result C T teeth

number

[0062] Optionally, the reference data in step 102 includes data collected over a first period from one or more sensors of the industrial system, which represents the historical state of the industrial system. In addition, or instead, the reference data may include real-time data collected by one or more sensors of the industrial system, which represents the current state of the industrial system. In addition, or instead, the reference data may include predicted state values ​​of the industrial system, which are related to the predicted performance of the industrial system in a second period and are based on the data collected over the first period, and which represent the future state of the industrial system. Beneficially, the different configurations of the reference data allow method 100 to be applied based on monitoring real-time values, historical trends, and / or future predicted values. This adaptability enables an operational support system applicable to a wide range of industrial systems, including industrial systems where predictions are unreliable due to complex processes or where real-time data collection is not possible.

[0063] If the reference data includes data collected over a first period, the determination of whether a deviation exists may optionally include evaluating the data collected over the first period (where the data collected over the first period is time-dependent) and identifying one or more trends and / or time-dependent trends in the evaluated data collected over the first period during the first period.

[0064] Optionally, data collected from one or more sensors in the industrial system associated with the reference data may be collected over a predefined time frame, such as 10 minutes, the duration of a batch, and / or the duration of one or more process cycles in the industrial system, or in real time. Alternatively or additionally, data collected from one or more sensors in the industrial system may include a data sequence of a predetermined size, such as 10 data points or the most recent 100 data points collected from a particular sensor.

[0065] Preferably, step 104 is performed by a deviation classifier. The deviation classifier is a first-principles model and / or one or more trained machine learning models, each trained machine learning model configured to detect one or more deviation types that exceed a predetermined threshold. For example, the one or more trained machine learning models are supervised machine learning models, unsupervised machine learning models, or semi-supervised machine learning models, examples of which include support vector machines, neural networks, decision tree-based algorithms, set-based algorithms, generalized likelihood ratio detectors, autoencoders, Bayesian networks, K-nearest neighbors algorithms, density-based spatial clustering for noisy applications (DBSCAN) algorithms, or cumulative sum detectors. From the one or more trained machine learning models, a trained machine learning model is selected based on one or more states and / or state values ​​associated with the reference data. Additionally or alternatively, the deviation classifier may be configured to output the presence or absence of a deviation based on an input containing reference data representing one or more states of an industrial system. Optionally, a predetermined threshold is determined by the deviation classifier.

[0066] In one embodiment, the deviation classifier is based on an unsupervised autoencoder neural network.

[0067] An autoencoder is a feedforward, non-recurrent neural network comprising an input layer, three hidden dense layers, and an output layer. The input and output layers have the same number of nodes so that the autoencoder outputs a reconstruction of the input data. The autoencoder is trained to learn a first transformation from a high-dimensional input space to a lower-dimensional space (represented by the intermediate layers of the three hidden dense layers) and a second transformation from the lower-dimensional space to the high-dimensional input space. By training the autoencoder with data related to the normal operation of an industrial system, the autoencoder learns how to reconstruct such normal operation data with high confidence and abnormal operation data (i.e., deviations) with low confidence. Thus, the reconstructions obtained from the autoencoder are used to classify whether the state value of the reference data is a deviation.

[0068] As mentioned earlier, autoencoders are trained with data representing "normal" operating conditions, such as data collected from one or more sensors in an industrial system when the system is operating optimally. During training, the autoencoder learns how to reconstruct the interactions between various variables and how to reconstruct them into their original variables (as outputs). When a process or component of an industrial system is affected by a problem, the interactions between variables are affected, such as changes in pressure, temperature, or other values ​​related to state values. As a result, the error in the autoencoder's output increases because the autoencoder can no longer reconstruct the input (affected data) based on the relationships (normal operating data) it learned during training. By monitoring the reconstruction error, we can obtain an indicator that the input data is deviating from the normal operating data, such as the error increasing as the equipment deteriorates. The probability distribution of the reconstruction error can be used to identify whether the data points are deviating, and by extension, whether the reference data is deviating.

[0069] In the implementation described here, the training data consists of one or more datasets, each consisting of a separate file that is a 1-second snapshot of the infrared signal recorded at specific intervals related to the process's output quality. Each file consists of 20,000 points with a sampling rate set to 10 kHz. The file name indicates the date and time the data was collected. Each record (row) in the data file is a data point. A large interval between timestamps (as shown in the file name) indicates that the experiment will resume on the next business day. Assuming that the process degradation occurs gradually over time as the input composition to the process deteriorates, one data point every 10 minutes is used for subsequent analysis. Each 10-minute data point is aggregated using the mean absolute value of the infrared photon sensor recordings across the 20,000 data points in each file and merged to generate a single data frame.

[0070] In this example, the goal is to identify output quality degradation based on a snapshot of the infrared signal and select predictive mitigation measures (such as allocating different resources to the process or repeating the process until output quality improves) to avoid output quality control failures. A dataset consisting of data collected from four infrared photon sensors is used to train the autoencoder. As the process is repeatedly run and the equipment operates until output quality is completely degraded, data from the first process cycle of the operation is used as training data representing normal behavior associated with high-quality output. The remainder of the dataset, based on the time until output degradation occurs, is used as test data to evaluate whether various methods can detect deviations before complete output degradation. After loading the infrared data, the index is converted to a datetime format and sorted by index in chronological order for access for training.

[0071] Standard splitting is performed on the training data to separate the test portion. Splitting is unnecessary when the autoencoder is updated with new training data, such as when a new dataset is added to the previous training data, as the original test data can be used. Standard data preprocessing, such as Scikit-learn's "MinMaxScaler," is performed on both the training and test data to scale the data for optimal training so that the data fits within the range [0,1].

[0072] As mentioned earlier, the autoencoder compresses infrared sensor readings into a low-dimensional representation, thereby capturing the correlations and interactions between various variables that influence the infrared sensor readings over time. In the neural network with the three hidden layers described above, the middle hidden layer has fewer nodes than the surrounding layers. More specifically, the first hidden layer consists of 15 nodes, the middle hidden layer has 5 nodes, and the third hidden layer has 15 nodes. The mean squared error is used as the loss function, and the autoencoder is trained using the "Adam" optimizer, which is a stochastic gradient descent method based on adaptive estimation of first and second moments. 5% of the training data is isolated after each epoch to validate the training (validation_split=0.05). This helps determine whether the autoencoder is training correctly (e.g., whether it is converging to a global minimum or underfitting).

[0073] The distribution of calculated losses within the training set (such as the mean absolute error (average absolute difference) between actual and predicted values) is collected and used to identify a suitable predetermined threshold for identifying deviations. This ensures that the predetermined threshold is set above the "noise level," that flagged deviations are statistically significant beyond the noise background, and therefore relevant to the operator of the industrial system. From the calculated loss distribution, a predetermined threshold of 0.4 is identified, for example. Next, the loss of the test set is calculated to determine when the output exceeds the predetermined threshold. Similar metrics are calculated for the training set, and all data are integrated into a single data frame. A standard plot of reconstructed loss against time may optionally be plotted to visualize the autoencoder output.

[0074] Alternatively or additionally, data associated with one or more processes in an industrial system, and / or data associated with one or more components of an industrial system, may be used to determine the presence or absence of deviation. For example, process and / or component-specific inputs may be used to select a trained machine learning algorithm, to preprocess reference data before inputting it into the deviation classifier, and / or as input to the deviation classifier.

[0075] Optionally, steps 106 and 108 may be interchangeable. For example, the problem determination in step 106 may further include using a rule-based model associated with one or more processes in an industrial system, and may include analyzing a number of predetermined control actions in step 108 based on the rule-based model.

[0076] Optionally, step 112 may further include generating a visual representation of the deviation effect by determining a relevant graphical interpretation of the deviation effect from a list of graphical interpretations, based on mapping the data associated with the deviation effect to one or more state values ​​obtained or calculated from reference data.

[0077] Figure 2 shows a high-level system architecture diagram relating to this disclosure for determining actions to improve the operation of an industrial system. Specifically, Figure 2 shows an operational support system 200 configured to perform method 100 and aimed at determining actions to improve the operation of an industrial system.

[0078] The operation support system 200 comprises a decision module 202a and a feedback module 202b. The decision module 202a comprises a deviation detection unit 204, an impact analysis unit 206, and an action identification unit 208. The deviation detection unit 204 acquires reference data 210 based on real-time data 212 or trend data 214. The impact analysis unit 206 optionally acquires resource data 216 and / or equipment data 218. The action identification unit 208 acquires data from a corrective action database 220 that stores a plurality of predetermined problems 222 and a plurality of predetermined control actions 224. The feedback module 202b comprises feedback data 226, which can be acquired from an operation interpretation unit 228. The decision module 202a and the feedback module 202b communicate with the user interface 230. Optionally, the operational support system 200 may further include a priority unit 232 that acquires graphical problem data 234, problem symptom data 236, impact status value data 238, and / or corrective action data 240.

[0079] The deviation detection unit 204 obtains reference data 210, for example, by performing step 102 of method 100. The reference data 210 includes either real-time data 212 and / or trend data 214. The real-time data 212 is associated with data collected by one or more sensors of the industrial system and indicates the current state of the industrial system. The trend data 214 is associated with data collected at a first point in time by one or more sensors and indicates the past state of the industrial system, or it is associated with data collected at a second point in time by one or more sensors and indicates the future state of the industrial system. For example, the trend data 214 includes historical data and / or forecast data (e.g., forecast future data obtained using method 600 detailed below). Optionally, the deviation detection unit 204 may obtain inputs that include forecast data, values ​​collected from one or more sensors of the industrial system, and calculated state values ​​indicating the state of the industrial system.

[0080] The deviation detection unit 204 determines whether a deviation exists in the reference data 210, for example, by performing step 104 of method 100. The deviation detection unit 204 identifies deviations based on data sequences from trend data 214, or real-time measurements from real-time data 212, etc. The deviation detection unit 204 consists of one or more deviation classifiers, each configured to determine whether a deviation exists. The one or more deviation classifiers are based on first-principles models and / or machine learning models such as support vector machines, neural networks, decision tree-based methods, set-based methods, generalized likelihood ratio (GLR) detectors, and cumulative sum detectors. For example, the deviation detection unit 204 is a model trained on both normal data, where no deviation exists or where deviations exist but are below a threshold, and failure scenario data, where deviations exist and exceed a threshold. The threshold is a predetermined threshold, and each deviation classifier has a corresponding threshold.

[0081] The impact analysis unit 206 obtains the presence or absence of a deviation from the deviation detection unit 204 and evaluates the impact of the deviation on one or more outputs of the industrial system based on the difference between the performance of the industrial system when a deviation exists at the first time point and the predicted performance of the industrial system when no deviation exists at the first time point, using step 106 of method 100, for example. The impact analysis unit 206 calculates the result of the affected state value of the industrial system.

[0082] The action identification unit 208 obtains an assessment from the impact analysis unit 206 and selects one or more mitigation control actions from the corrective action database 220, for example, by performing the action using step 108 of method 100. The selection includes identifying a problem from a plurality of predetermined problems 222 based on the assessment from the impact analysis unit 206 and identifying one or more control actions from a plurality of predetermined control actions 224 based on the corresponding problem. The one or more control actions are selected to mitigate the impact of the deviations detected and assessed by the impact analysis unit 206.

[0083] The corrective action database 220 is a systematic collection of predetermined problems, the consequences arising from those problems, and corresponding predetermined control actions accessible to database users associated with the industrial system, such as the operational support system 200 and operators. Each predetermined problem is linked to an impact and is associated with at least one predetermined control action based on the process know-how and rules associated with the industrial system's configuration. Therefore, the corrective action database 220 is specific to the industrial system. The links between predetermined problems and predetermined control actions are static, based on past know-how obtained from database users, or dynamic, changing in response to operational decisions and / or feedback from users of the industrial system.

[0084] For example, an operator of an industrial system lists all potential problems that may occur in the industrial system, such as potential problems related to each process or component of the industrial system, and the impacts resulting from each problem. The operator then lists each control action that mitigates or resolves each listed problem. The listed problems and listed control actions are organized in the corrective action database 220 as multiple predefined problems 222 and multiple predefined control actions 224, each with its associated impact. Alternatively, a list of all control actions that may be performed in the industrial system is generated. One or more problems affected by each control action are also listed. The potential impacts of each problem are further listed or stored along with the related problems. The corrective action database 220 is constructed by organizing the listed control actions into multiple predetermined control actions 224 and multiple predetermined problems 222 associated with them, where each problem is associated with one or more impacts on the performance, state, or processes of the industrial system. In addition, or alternatively, each control action is associated with one or more impacts.

[0085] Further details regarding the issues, impacts, and potential control measures are described in numerous technical documents, including "Brewery utilities (manual of good practice)," EBC Technology and Engineering Forum, Getranke-Fachverlag, Hans Carl, Nurnberg (1997); "Clean-in Place for Biopharmaceutical Processes," DA Seiberling, Drugs and the Pharmaceutical Sciences 173, pp.302-313 (2007); and "Handbook of Milk Powder Manufacture (2nd Edition)," J Pisecky, GEA Process Engineering A / S p.222 (2012).

[0086] By understanding or acquiring impacts, such as the impact of deviations determined using the impact analysis unit 206, the corrective action database 220 can be accessed, and relevant issues can be retrieved from a plurality of predetermined problems 222 (i.e., by identifying issues with relevant impacts similar to or matching the determined impact). Since the corrective action database 220 is configured to link each predetermined problem to one or more control actions from a plurality of predetermined control actions 224, the relevant control actions can then be retrieved from the corrective action database 220. When a new process, component, or link between a problem and a control action becomes available in the industrial system, the corrective action database 220 is updated with the new predetermined problem, control action, and optionally related impacts.

[0087] The user interface 230 obtains one or more mitigation control actions from the action identification unit 208, and also obtains the calculation results of the impact of each mitigation control action on the output of the industrial system, as performed using step 110 of method 100. The user interface 230 provides the user of the user interface with a display of the impact of performing one or more mitigation control actions, as performed using step 112 of method 100. The display of impacts includes calculation results obtained from the impact analysis device, such as impact status value data 238. This analysis helps the user of the user interface, such as an operator of an industrial system plant, to determine whether it is better to continue operations according to the current plan or whether there is a better method from an economic standpoint, and thus helps the user of the user interface to select an operation to initiate one or more mitigation control actions to modify one or more processes of the industrial system.

[0088] Optionally, the user interface 230 may obtain one or more mitigation control actions from the action identification unit 208 via the priority unit 232. The priority unit 232 is configured to prioritize one or more mitigation control actions based on the impact of each action, and obtains graphical problem data 234, problem symptom data 236, impact status value data 238, and / or corrective action data 240, and uses the obtained data to prioritize one or more control actions. The graphical problem data 234 includes mapping of relevant data, such as relevant time series data, batch performance data, or other relevant data related to the performance and deviation of the industrial system, in order to generate a graphical representation of the impact of the detected deviation. An example of the graphical representation is shown in Figure 3.

[0089] The operation interpretation unit 228 depends on the type and / or amount of feedback data 226 required. For example, the operation interpretation unit 228 is configured to interpret whether one of the selected control actions has been implemented, or whether an alternative control action has been implemented, started, and / or completed. The alternative control action does not include a course of action within a given time frame. In addition, or alternatively, the operation interpretation unit 228 is configured to label the selected control action, such as whether the control action is relevant, whether the problem was the cause of the deviation, and / or whether the deviation is relevant.

[0090] In an exemplary application of the operation support system 200, based on real-time data 212 and trend data 214, the batch occupancy time in the first industrial process is predicted to be 15 minutes longer than usual. The minimum and maximum occupancy times are predetermined to be 105 minutes and 135 minutes, respectively, based on technical know-how. The average historical values ​​included in the trend data 214 support the above. The impact analysis unit 206 assesses that the occupancy time affects the productivity and output quality of the first industrial process. Based on the equipment data 218, the capacity associated with the first industrial process is 2000 hectoliters. Based on the resource data 216, the downgrade of the product associated with the first industrial process has an impact of 100 weighted units per hectoliter.

[0091] The impact analysis unit 206 calculates two impacts. The first is the productivity loss due to the additional time occupied by the first industrial process, and since the resource data 216 for occupation is 30,000 weighted units per hour, the impact is 7,500 weighted units. The second is that the decrease in quality yield causes process iteration, resulting in batches associated with the first industrial process being mixed with other batches with lower quality yields, and consequently the margin for high-quality batches is lost, with a total impact of 20,000 weighted units. The impact analysis unit 206 evaluates these two impacts together as 27,500 weighted units.

[0092] The action identification unit 208 analyzes the reason using, for example, real-time data 212 of input / output flows from several predetermined problems 222. This data is analyzed in conjunction with the corresponding historical values ​​in trend data 214. The input / output flows are identified as not being the cause of the longer-than-usual occupancy time. Another problem among the several predetermined problems 222, the heating rate, is calculated using the real-time data 212 and, based on a comparison of the actual heating rate with the expected head-up rate based on volume, configuration, equipment information, thermal surface, and steam supply characteristics obtained from resource data 216 and equipment data 218, is determined to be lower than the historical rate of the first industrial process from trend data 214. The main reason is determined to be related to the heating rate. Using the corrective action database 220, "Review steam valve settings in machine settings" is selected as the corresponding control action from several predetermined control actions 224. The priority unit 232 retrieves not only data related to the heating rate, with a calculated impact weight of 27,500 units, but also other issues related to the calculated impact. Next, the priority unit 232 sorts the issues according to the calculated impact. Then, the user of the user interface 230 determines that the proposed issues and / or actions are relevant, and this is registered by the operation interpretation unit 228. As a result, one or more predetermined thresholds of the deviation detection unit 204 are readjusted, or an alternative action for the same issue in the corrective action database 220 (e.g., "increase steam supply setpoint") is proposed.

[0093] Preferably, the operation support system performs one of the steps of method 100, 400, and / or 500 to support the display of the visual representation 300 in Figure 3.

[0094] Figure 3 shows an illustrative representation of the impact of a problem on the performance of an industrial system in the user interface of an industrial system. Specifically, Figure 3 shows an exemplary visual representation 300 displayed to the user interface user in order to graphically convey the impact of an identified problem on the industrial system.

[0095] The visual display 300 includes a user interface display 302, a graphical figure 304, impact information 306 related to the evaluation of the impact of the deviation, a first icon 308 related to the first control operation, a second icon 310 related to the second control operation, and a third icon 312 related to feedback data.

[0096] Display 302 is a display of all or part of the user interface, directly or indirectly connected to the industrial system, and accessible to users of the user interface, such as operators of the industrial system. Display 302 is configured to provide users of the user interface with a graphical figure 304 and impact information 306. The graphical figure 304 and impact information 306 relate to the detected deviation and the assessment of the impact of that deviation on the industrial system. Since the graphical figure 304 is determined to be the most relevant graphical representation of the impact of the deviation, the graphical figure 304 and impact information 306 may optionally be the same data display representation.

[0097] User interfaces, such as the user interface 230 in Figure 2, are further configured to receive data related to user selections. User selections correspond to a first icon 308, a second icon 310, or a third icon 312. The first icon 308 is associated with a first control action, which is, for example, a mitigation control action accompanied by the highest priority control action, which, when performed, reduces the evaluated impact of the deviation. The second icon 310 is optional and is associated with a second control action, which is, for example, a second-highest priority mitigation control action, or a control action associated with not performing the highest priority control action (e.g., a control action that, when performed, does not result in a change to one or more related processes in an industrial system). The third icon 312 represents an option to provide the user with feedback related to the control action of the first icon 308, such as whether the mitigation control action of the first icon 308 is appropriate with respect to impact information 306, and the feedback is based on the user's technical knowledge and information provided from the visual representation 300.

[0098] Figure 4 shows a flowchart illustrating how to select one or more mitigation control actions based on their impact on the performance of the industrial system. Specifically, Figure 4 shows a computer implementation method 400 that is compatible with processes 100, 500, 600, 800, 900, and / or 1000.

[0099] Figure 4 shows Method 400, which is a process for selecting and prioritizing one or more mitigation control actions. This process can be performed in addition to or instead of step 108 of Method 100. First, one or more mitigation control actions are selected from a plurality of predetermined control actions. In this example, the selection consists of a two-step process. The first is to identify the problems that may be causing the deviations detected in the state of the industrial system, and the second is to identify at least one control action that is known to solve or mitigate the identified problems. Next, the impact of each mitigation control action is calculated, and the results of performing each mitigation control action in the industrial system are determined. These calculation results can be used to prioritize the control actions. An example of a system on which Method 400 is performed is shown in Figure 2.

[0100] Method 400 includes steps 402, 404, 406, 408, and 410. Step 402 includes selecting one or more mitigation control actions from a plurality of predetermined control actions. Step 404 includes identifying one problem from a plurality of predetermined problems that cause deviations. Step 406 includes identifying one or more control actions based on the identified problem. Step 408 includes calculating the impact of each mitigation control action on the industrial system. Step 410 includes prioritizing one or more mitigation control actions.

[0101] Step 402 involves selecting one or more mitigation control actions from a plurality of predetermined control actions, each predetermined control action representing a predetermined operation to modify one or more processes of an industrial system. Selecting one or more mitigation control actions is a technical knowledge-based process, where a problem related to the industrial system is known to result in one or more changes to one or more processes of the industrial system. Similarly, a change to one or more processes of an industrial system can be associated with at least one predetermined problem. Thus, the selection of one or more mitigation control actions is based on history and prior knowledge about how to modify one or more processes of the industrial system, such as mitigating the undesirable effects of a detected deviation.

[0102] Step 404 involves identifying one problem from a set of predetermined problems that cause a deviation, and identifying the problem causing the deviation from a set of predetermined problems is based on an assessment of the impact of the deviation. The set of predetermined problems are stored in a corrective actions database and include problems related to processes, equipment, environment, resources, and / or operations of the industrial system. Each problem is linked to at least one control action in a one-to-one, many-to-one, or one-to-many relationship.

[0103] Step 406 involves identifying one or more control actions from a set of predetermined control actions based on the identified problem causing the deviation, the identified one or more control actions being selected to mitigate the impact of the deviation on one or more outputs of the industrial system. The control actions represent operations that, once initiated and / or executed, bring about a change in one or more processes of the industrial system to mitigate the deviation.

[0104] Step 408 includes calculating the impact of each mitigation control action on the industrial system based on calculating one or more impact values ​​from a plurality of impact values, including any of the following: production indicators, operational indicators, environmental indicators, safety indicators, or efficiency indicators. One or more impact values ​​are combined to form a total impact value calculated for each mitigation control action by weighting each impact value based, for example, on a specific process and / or component of the industrial system. The weighted impact values ​​can then be summed, multiplied, averaged, or mathematically combined using other standard methods. The weighting of the impact values ​​yields an adjusted final impact, such as the adjusted total impact for each deviation type, each process of the industrial system, each component of the industrial system, each focus operation of the industrial system (e.g., water conservation during drought), and / or for each industrial system. Optionally, Step 408 may be adjusted to a specific state value, such as a state value associated with a detected deviation (e.g., a deviation detected in Step 404).

[0105] Step 410 includes prioritizing one or more mitigation control actions, for example, based on which mitigation control action is calculated to have the greatest impact on mitigating deviations or improving the performance of the industrial system, as performed in step 408. Optionally, step 410 may not be performed, for example, if only one mitigation control action is selected and / or appropriate.

[0106] Beneficially, Method 400 provides users of the user interface, such as operators of industrial systems, with the most relevant information and relevant data to quickly and efficiently decide whether to perform operations that modify one or more processes in the industrial system, and helps optimize the performance of the industrial system with up-to-date and appropriate operational recommendations.

[0107] Figure 5 shows a flowchart illustrating how to improve the selection of one or more mitigation control actions based on feedback data. Specifically, Figure 5 shows a computer implementation method 500 that is compatible with processes 100, 400, 600, 800, 900, and / or 1000.

[0108] Figure 5 shows Method 500, which is a process for improving Method 100 and / or Method 400 and / or Operation Support System 200. This process includes determining and storing whether a selected mitigation control action has been performed, for example, whether an operator of an industrial system initiated a mitigation control action. This information can be used to improve deviation detection and analysis. An example of a system for implementing Method 500 is shown in Figure 2.

[0109] Method 500 includes steps 502 and 504, and one or more of steps 506, 508, or 510. Step 502 includes determining whether a mitigation control action has been performed. Step 504 includes storing feedback data related to the determination. Step 506 includes using the feedback data to further determine whether there is a deviation. Step 508 includes using the feedback data to further identify the problem. Step 510 includes using the feedback data to further identify a control action. Step 512 includes using the feedback data to tune the deviation classifier to improve the accuracy of the deviation determination.

[0110] Step 502 includes determining whether a mitigation control action has been performed, such as the selected mitigation control action of Method 400. Determining whether a mitigation control action has been performed includes determining whether the mitigation control action has been initiated, is in progress, or has been completed. This determination can be made based on data collected from one or more sensors in the industrial system, or based on input from the operator of the industrial system.

[0111] For example, a user of the user interface checks whether they initiated the selected control action. In another example, the user of the user interface checks whether the proposed problem and / or control action is related to the deviation and / or behavior of the industrial system. Alternatively, if the selected control action is not related, or is not the best course of action for the operator at that time, no user input is required, and the operator of the industrial system will either initiate another action, such as manual readjustment, or will not perform the control action. Therefore, in this example, it is determined that the mitigation control action was not performed.

[0112] Step 504 includes saving feedback data related to the decision. The feedback data consists of any relevant data, such as data used to perform step 502, e.g., user input or data collected from one or more sensors. Optionally, saving the feedback data may also include labeling the selected control action, such as whether the control action is relevant, whether the relevant problem is relevant (e.g., a cause of the deviation) or not (e.g., not a cause of the deviation), and / or whether the deviation is relevant.

[0113] Step 506 includes using feedback data to further determine whether there is a deviation. For example, if the feedback data indicates that the deviation is irrelevant, a predetermined threshold associated with that deviation is adjusted to prevent the detection of irrelevant deviations. Further determination of whether there is a deviation includes determining whether there is a new deviation in the new reference data, which would result in a more reliable determination, such as by performing step 104 of method 100.

[0114] Step 508 includes further identifying the problem using the feedback data. Further identifying the problem includes adjusting the association between the problem and the calculated impact of the deviation, for example, by incorporating further proposed problems and their relevant impacts on the industrial system, or by adjusting the weighting applied to specific impact values. Further identifying includes identifying new problems based on the new deviation, for example, by performing step 106 of Method 100 and / or step 406 of Method 400.

[0115] Step 510 includes further identifying control actions using feedback data. Further identification of control actions includes adjusting how problems and control actions are linked or associated in the corrective action database. For example, based on feedback data, alternative control actions can be linked to previously detected problems. Further identification includes identifying new control actions based on new problems, such as by performing step 106 of method 100 and / or step 408 of method 400.

[0116] Step 512 includes updating at least one deviation classifier based on feedback data and / or using the feedback data as further input to the deviation classifier. Optionally, step 512 may be contained within step 506, and adjusting a given threshold may include updating the relevant deviation classifier.

[0117] By incorporating the process of Method 500 into, for example, Method 100 or the operational support system 200, a feedback loop is obtained for continuously improving operational recommendations, enabling precise, reliable, and finely tuned processes for specific processes, components, or operations of an industrial system.

[0118] To accurately and reliably apply the operational decision support detailed above (refer to Figures 1-5) to predict future deviations based on reference data related to the predicted future performance of an industrial system, it is necessary to accurately and reliably predict the future performance of the industrial system. Figures 6-11 detail performance prediction methods and / or systems for predicting system performance using a combination of trained machine learning models, predictive models, and continuous learning loops. Including a trained machine learning model provides accurate predictions based on historical data related to the industrial system. Including a predictive model updates the predictions based on real-time data related to the industrial system. The continuous learning loop continuously improves the predictions or adjusts them to more closely approximate the measured future performance of the industrial system.

[0119] Figure 6 shows a flowchart illustrating a method for predicting the system performance of an industrial system. Specifically, Figure 6 shows computer implementation method 600, which is compatible with process methods 100, 400, 500, 800, 900, and / or 1000.

[0120] Figure 6 illustrates Method 600, a process for predicting how one or more processes in an industrial system will develop or change in the future. Each state value, such as the state values ​​described above with respect to Method 100, has a dedicated prediction process. This process uses a combination of trained machine learning outputs and real-time data to enhance explainability and employs an embedded learning loop to increase reliability. Thus, Method 600 is an explainable and reliable process for predicting the system performance of an industrial system, and can be used, for example, for monitoring industrial systems or for supporting decision-making by industrial system operators. An exemplary system for implementing Method 600 is shown in Figure 7.

[0121] Method 600 includes steps 602, 604, 606, and 608. Step 602 includes obtaining a first state value from a trained machine learning model. Step 604 includes obtaining real-time system information, including data collected from a set of sensors in an industrial system. Step 606 includes determining a second state value using a predictive model with the first state value and the real-time system information. Step 608 includes updating the trained machine learning model.

[0122] Step 602 involves obtaining a first state value from the trained machine learning model, which predicts the performance of the industrial system in a second period based on the industrial system's performance in a first period. The industrial system's performance in the first period is determined using data collected from a set of sensors in the industrial system during the first period. The first state value of the system is associated with the predicted performance of the industrial system in the second period and indicates the future state of the industrial system.

[0123] For example, a trained machine learning model uses historical data related to an industrial system to predict the future performance of that system. The trained machine learning model is specialized for a particular state value and is therefore trained to output one or more specific state values. The input data corresponds to a first state value, and different state values ​​from different trained machine learning models use the same or different input data, depending on the output state value. Alternatively, a single trained machine learning model may be configured to predict multiple first state values. The input data is collected over a first period from a sensor set of the industrial system and includes data indicating the historical state of the industrial system, data related to one or more relevant processes of the industrial system, data related to one or more relevant components of the industrial system such as equipment, and / or data related to the trained machine learning model, such as model configuration specifications. Furthermore, feedback data is input to the trained machine learning model to improve the reliability of the output obtained.

[0124] Step 604 includes obtaining real-time system information, which includes data collected from a sensor set of the industrial system. For example, the real-time system information indicates the current state of the industrial system and includes data collected by the sensor set of the industrial system. The real-time system information indicates the current state of the industrial system, and the current state value is calculated based on data collected by the sensors during the current period. In addition to the data collected by the sensors of the industrial system, the real-time system information consists of data from various sources, depending on the target state value targeted by Method 600.

[0125] Examples of real-time data sources include programmable logic controllers (PLCs) in process units, experimental equipment related to one or more processes in an industrial system, upstream data related to resource availability and operational scheduling, batch progress estimates related to process cycle duration, output quality and output volume metrics, and other calculated or measured values ​​related to the industrial system.

[0126] Step 606 involves determining a second state value using a predictive model, which uses the first state value and real-time system information to determine the second state value. The first state value obtained in Step 602 serves as a starting point for tracking the target characteristics and / or features of the industrial system and predicting the future performance of the industrial system. The real-time system information is used to fine-tune or adapt the prediction to bring it closer to the future state of the industrial system.

[0127] For example, for a specific process in an industrial system, such as a focus process unit and its operating characteristics, comprehensive sensor and actuator values ​​are collected from the process unit PLC. Furthermore, available quality data is collected from various lab devices that can mirror data at the PLC level. Upstream data influencing how the unit behaves against given process requirements is also collected. Additionally, other computational data (such as virtual sensors for quality measurement) are input into the predictive model. For batch operations, post-batch values ​​indicating the industrial system's performance are also used in the predictive process for the current batch. The batch progress is also input into the predictive model to evaluate the quality of the predictors relative to actual values. Based on this acquired real-time system information, the predictive model modifies the first state value, taking into account current influencing factors on the process that may be too variable, unpredictable, or rare to incorporate into the training data of the trained machine learning model in step 602. Beneficially, using a predictive model provides more process-specific and reliable predictions than a more generalized approach using only a trained machine learning model.

[0128] Step 608 involves updating the trained machine learning model, for example, by calculating a third state value for the system using data collected from the sensor set of the industrial system during the second period, where the third state value of the system relates to the performance of the industrial system during the second period and indicates the state of the industrial system during the second period. The trained machine learning model is then adjusted to reduce the difference between the second state value determined by the predictive model and the third state value of the system.

[0129] The third state value is obtained from real-time system information or calculated using additional data related to the system's performance during the second period. The third state value represents an unpredicted measured state value that can be compared to a predicted state value, such as the second or first state value. By adjusting the trained machine learning model based on the difference between the second and third state values, the entire prediction system is updated so that the predicted performance better matches the future measured performance of the industrial system. Method 900 is an exemplary process for updating a trained machine learning model.

[0130] Preferably, the first period is prior to the second period, for example, the first period is the present and / or past period, and the second period is a future period. For example, one or more process cycles of an industrial system obtained from a series of process cycles occur within the second period.

[0131] Optionally, a third state value may be used to obtain further first state values ​​from the trained machine learning model, and method 600 is repeated as performed in method 800. The third state value may be used directly or obtained from storage, and the second and third state values ​​may optionally be stored for use as input to and / or feedback data to the trained machine learning model or predictive model.

[0132] When systematic quality labeling is performed, the calculation of the third state value is repeated at predetermined time intervals over a second period. For example, the second state value is assigned the first quality label, and the third state value is assigned the second quality label. Subsequently, tuning the trained machine learning model involves determining whether the first quality label matches the second quality label. Updating the trained machine learning model is performed remotely or on-site, and the updating of the trained machine learning model may optionally be further based on comparisons of multiple second state values ​​and multiple third state values.

[0133] In an example where multiple third state values ​​are used to update the machine learning model trained in step 608, step 608 further includes filtering the multiple third state values ​​and determining whether the filtered multiple third state values ​​meet a predetermined criterion, which is associated with the number of filtered third state values. For example, the multiple third state values ​​are used to update the trained machine learning model only if there are values ​​of sufficient quality to update the trained machine learning model without introducing inaccuracies or, for example, bias. If the predefined criterion is not met, the update of the trained machine learning model is canceled, and the trained machine learning model is returned to its state before the update, for example, the state of the trained machine learning model at the time the first state value in step 602 was obtained. In addition, or alternatively, if the predefined criterion is not met, step 608 may further include generating one or more first state values ​​using an alternative model, such as a model that averages the multiple filtered third state values.

[0134] The trained machine learning model is, for example, an artificial narrow-field intelligence (AVI) model configured to match a specific first state value, or, for example, an artificial general intelligence (AGI) model configured to predict a wider range of future performance of an industrial system. Preferably, the trained machine learning model in step 602 is one of several trained machine learning models, and each state value associated with a specific component or process of the industrial system corresponds to a specific trained machine learning model. For example, the several trained machine learning models include any of a linear multivariable regression model, a polynomial regression model, a random forest regression model, a neural network model, or an expert model, and step 602 includes receiving a focus indicator from among several focus indicators (each focus indicator is associated with a component, process, input, or output of the industrial system) and selecting a trained machine learning model from among several trained machine learning models based on the focus indicator.

[0135] Optionally, inputs to trained machine learning models and / or predictive models may include information relating to the components, processes, inputs, or outputs of an industrial system, and / or data collected by a set of sensors in the industrial system (including any of the following: temperature data, humidity data, pressure data, consumption data, process speed, or operating conditions of the industrial system).

[0136] In an exemplary implementation of Method 100, model design specifications, such as sensor and actuator data, historical data, product recipe data, equipment design specification data, and process know-how-based specifications for determining the model type (e.g., gray-box model), are used to train a representation of a specific process behavior (e.g., filtration time or process thermal efficiency) represented by state values. Depending on the nature of the state values ​​in the problem, one of the following modeling techniques is used: a linear multivariable regression model, a polynomial regression model, a random forest regression model, a neural network model, and / or an expert model, which is a manual linear or nonlinear model defined by an expert.

[0137] To fit one or more trained machine learning models (optionally gray-box models), the formula f is defined by a state value y, a specified input variable / feature x, and a parameter p.

number

[0138] The input variables / features can include upstream data of the target process associated with the state value y (related state values, settings, process variables, etc.) and historical data (past state values, settings, process variables, etc.). All of the above data is stored in a central database on the edge device.

[0139] Model parameters are fitted using standard mathematical methods that either minimize the least-squares error or fit the error of the expected target value compared to the model output value, using the target value of the state value at the prediction point (usually the final batch value of the state value) and the value of the input variable. Since the trained machine learning model is desirable to be nonlinear, numerical iteration is necessary. An example, though not limited to this, is "gradient descent optimization." However, when fitting a model with least-squares error as the objective function, any optimization method can be used arbitrarily. For linear models, an analytical solution can be obtained.

[0140] During model tuning, data such as missing data in batches (gaps between data) and faulty state values ​​due to process state and / or data processing issues are filtered based on quality. If the tuning algorithm does not have enough data, the model fitting for the current state value y is canceled for faulty or poor-quality data, the model tuning is terminated, and / or the prediction for the next state value is fitted / tuned instead. A fallback strategy is used, such as ignoring the fitted model using predefined criteria and using a model based on average state values ​​fitted by a real-time data processing algorithm.

[0141] The formulas, inputs used, and parameter values ​​are stored in a database and associated with a specific state value y in which the model is used. This state value may refer to a specific recipe or process.

[0142] Another example of a gray-box model is a hybrid model, which combines known formulas with input variables / features and parameters, and uses a black-box layer to estimate parameter values ​​from the same or additional input variables / features. Examples of black-box models include neural networks, random forests, and polynomial regression models. To fit a black-box model, no process expertise / know-how is required; for example, standard deep learning algorithms can be used to automatically find input features / variables from upstream and historical data of the relevant process.

[0143] For example, a black-box model may include a data filtering process, such as removing bad ranges from the entire dataset if batch data is bad due to missing data / gaps within a batch, or if state values ​​are bad / defective based on process state and / or data processing issues. The best fit of the selected model is determined using the currently selected inputs / features. The selected black-box model may optionally have setup-related parameters. For example, a neural network model may require specifications such as the number of layers, the number of neurons / layer nodes, and the activation model, while a random forest (random decision forest) model may require information such as the number of depths, the number of estimators, and the tuning algorithm specification. Setup-related parameters are stored in a database and associated with specific state values. Preferably, cross-validation is performed on the model and data to analyze the goodness of fit. Based on this analysis, predefined criteria are used to ignore the fitted model and use the aforementioned fallback strategy. In some black-box model types, such as when the model is a linear regression model or a random forest model, tuning determines the most important features / inputs. Using this decision, features are selected based on criteria, and then the model is refitted with the new feature selection until all features meet the criteria. The formulas, inputs to be used, and parameter values ​​are stored in a database, referencing the assigned state values ​​on which the model should be used. These state values ​​preferably represent specific recipes and / or processes.

[0144] The trained machine learning model and associated parameters are used as a starting point to track designed characteristics and predict their future behavior in real time. To do this for a specific process unit under consideration and its operating characteristics, a comprehensive set of sensor and actuator values ​​is collected from the unit PLC. In addition, available quality data is collected from various lab devices or mirrored at the PLC level. Furthermore, upstream data influencing the unit's behavior for a specific recipe is also collected. In addition, other computational data (such as virtual sensors for quality measurement) are also input into the predictive model. For batch operations, post-batch state values ​​are used in the predictive process for the current batch. The batch progress state is input into the predictive model to evaluate the predictor's quality against actual values.

[0145] All predicted values ​​are saved, and the user interface is updated with the latest values. Periodic quality checks are performed on each state value to improve the quality and reliability of predictions. Results are saved along with snapshots of the relevant input values, and any necessary updates to the trained machine learning model and / or the training associated with the machine learning model are performed manually (in the cloud) by experts. Feedback is based on systematic sampling of the quality of state values. Sampling is the manual labeling of the quality of state values ​​by a lead operator. Sampling is based on a fixed / variable frequency of a specific process or production batch to cover representative quality samples of prediction quality.

[0146] The format or type of data used in trained machine learning models and / or predictive models varies depending on the live evaluation of the trained machine learning models and / or predictive models. In examples involving live evaluation, some values ​​within the real-time system information are constants, while others depend on variables / time. Considering the trained machine learning models in the working examples above, it is further related to the mash tun process. [Total Occupancy Time: Mashtan] =a0*[Progress]+a1*[Sequence Steps]+a2*"Malt Load"+a3*[Average Heating Rate]

[0147] The above predictions use a multivariable linear model. PLC data (obtained from PLC storage, collected by different PLC protocols and OPC) and internal live calculations (obtained from RAM storage) are real-time time-dependent values / data representing sequence steps and average heating rate, respectively. Upstream data (obtained from database storage), post-batch / historical state value data, and experimental equipment data (obtained from database storage) are constant values / data. In this example, the malt load is the upstream value. Batch progress estimation is real-time time-dependent values / data (stored in RAM memory or database [8]) that constitute the prediction itself. a0, a1, a2, and a3 are tuned parameters for the prediction model.

[0148] In the example of tuning a pre-trained machine learning model, the input data is structured as vectors / arrays. Target values ​​are provided and not calculated for tuning the model. The following table shows an example dataset configuration (total sample size M, total batch size B, total input / feature size N (including progress estimates)). [Table 1]

[0149] Example 1. Dataset for predicting the total occupancy time of a mashed tan: [Table 2]

[0150] Example 2. A dataset for predicting the progress of mashed tan: [Table 3]

[0151] The typical input and output values ​​of trained machine learning models and component algorithms are floating-point numbers and are generally not limited to their possible values. Only the state values ​​associated with progress (target or predicted values) are limited to a range of 0-100%.

[0152] In expert-based models, all real-time system information is essential for both tuning and live evaluation. If certain data is unavailable, the algorithm will either stop tuning or perform live evaluation. In black-box models, the tuning algorithm in the example above performs feature selection independently of the expert analysis, so all data available in the process unit is optional. Data / inputs are available during live evaluation; otherwise, predictions will stop or be canceled.

[0153] The training dataset for a trained machine learning model should ideally include multiple batches / runs to gain a more general understanding of how the process works. For example, at least three batches are needed for statistical calculations and algorithm tuning. This is because the dataset for each batch consists of multiple time samples across the entire batch, such as a sample frequency every 10 seconds. For example, for state value prediction tuning with a batch duration of 2 hours, using 10 inputs / features, and with three batches of ideal (unfiltered) data, the estimated data volume would be as follows: [Batch number] * [Batch period] * [Sample frequency] * ([Number of inputs] + 1) =3*2*60*(60 / 10)*(10+1) = 23,760 data points, or a dataset of 2,160 rows and 11 columns (10 inputs and 1 target value)

[0154] Examples of data preprocessing steps performed in Method 100 include: selecting batch data (sensor values, logic / automation values, setpoints, status values) from the relevant process (e.g., mash tan in the example above) (all data not relevant to the batch is removed); finding the batch index / number within the current dataset and attaching the same batch ID / number to all data (setpoints and status values) from the upstream process, where the upstream data is set as a constant value for each individual batch, which in the example above would be the mill and the malt hopper of the mill; creating a progress column which is a linear interpolation of 0-1 over time between the start and end of the batch and optionally a target value or additional input / predictor; creating a target value column containing the target value for each batch, where for status value prediction, the target value is a constant value of the final status value of the batch, and for progress prediction, the target value is a time-dependent value of 0-1 for the batch based on the relative time above; selecting features / inputs (columns). For example, in the black box example, multiple tuning interactions occur based on a feature selection algorithm (thresholds for the importance of inputs / features), and some columns are excluded from the dataset; in the expert model example, only specified features / inputs are included in addition to the target values ​​(specified in the database for the combination of state values ​​and recipes); rows containing NaN in any of the columns are removed, and the quality of the dataset (row number and batch number) is checked. If the quality is too poor, the algorithm is stopped; or K-fold or other types of data splitting (training set / test set). Preferably, this is included in the tuning and evaluation algorithm.

[0155] Optionally, real-time system information is stored in a database and / or RAM memory, and method 100 may further include database connections such as database queries and processes. The first, second, and third state values, as well as associated calculated values, are stored in RAM memory and / or local storage for recovery purposes. PLC data such as sensor, logic, and automation values ​​are stored in the PLC, and the values ​​are read by various standard means based on the industrial system's communication protocol.

[0156] Figure 7 shows a high-level system architecture diagram relating to this disclosure for predicting the system performance of an industrial system. Specifically, Figure 7 shows a performance prediction system 700 configured to perform method 600 and aimed at predicting state values ​​that indicate the future state of an industrial system.

[0157] The performance prediction system 700 comprises a static prediction unit 702, a dynamic prediction unit 704, and a learning unit 706. The static prediction unit 702 includes a trained machine learning model 708, to which historical data 710, process data 712, component data 714, model specification data 716, etc., are input. The dynamic prediction unit 704 comprises a prediction model 718 and real-time data 720. The real-time data 720 includes data from multiple sources such as the process unit PLC 722 and experimental equipment 724, as well as data such as upstream data 726, batch progress estimates 728, post-batch state values ​​730, and / or calculated state values ​​732. The dynamic prediction unit 704 may optionally further include stored data, static prediction data 734, obtained from the static prediction unit. Furthermore, the dynamic prediction unit 704 outputs the prediction model 718 as dynamic prediction data 736. The learning unit 706 includes feedback data 738 and optionally a learning data labeling unit 740.

[0158] The static prediction unit 702 and the dynamic prediction unit 704 work together for accurate and reliable predictions specific to site, process, component, and / or state values. For example, the static prediction unit 702 is configured for offline design and tuning, and is maintained and / or updated offsite by machine learning experts and tuning algorithms. The dynamic prediction unit 704 is configured to be site-specific. Beneficially, the combination of offsite and onsite prediction units allows for the use of larger training data from multiple processes, etc., for the trained machine learning model 708. The training data can then be subjected to supervised data cleaning and processing methods.

[0159] Furthermore, the impact of updating the trained machine learning model 708 can be analyzed before implementing the trained machine learning model 708 for monitoring industrial systems. Next, the on-site prediction model 718 transforms the first state value 742 from a more general future prediction to a real-time customized future prediction by combining the first state value 742 with real-time data 720 to output a second state value 744. The second state value 744 may optionally be stored as dynamic prediction data 736.

[0160] The dynamic prediction data 736 may optionally be output to a user interface 746, allowing a user of the user interface, such as an industrial system operator, to select an operation to modify one or more processes of the industrial system based on a second state value 744. Alternatively, or additionally, the dynamic prediction data 736 may be used by the learning unit 706 to update a trained machine learning module 708. For example, a third state value 748 is calculated by the learning data labeling unit 740 based on data collected from a sensor set of the industrial system. The third state value 748 is associated with the performance of the industrial system during the period associated with the second state value 744 and is stored as feedback data 738 with or without the second state value 744.

[0161] Optionally, data associated with the third state value 748 and the second state value 744 may be stored as feedback data 738 and / or as the difference between the third state value 748 and the second state value 744. The trained machine learning module 708 is adjusted to reduce the difference between the second state value 744 and the third state value 748. Optionally, the third state value 748 may also be input to the static prediction unit 702 and / or to the dynamic prediction unit 704 in combination with historical data 710, for example, the third state value 748 may be included in real-time data 720. Beneficially, using the third state value 748 as real-time data 720 allows the dynamic prediction unit 704 to be updated efficiently, and the continuously updated performance prediction system 700 improves the accuracy and reliability of the second state value 744. Even more beneficially, updating the static prediction unit 702 with the third state value 748 provides a continuous learning loop, and the performance prediction system 700 improves over time.

[0162] Figure 8 shows a flowchart illustrating a method for updating a trained machine learning model configured to predict the performance of an industrial system based on data collected from a sensor set of the industrial system. Specifically, Figure 8 shows a computer implementation method 600 that is compatible with the processes of methods 100, 400, 500, 600, 900, and / or 1000.

[0163] Figure 8 shows Method 800, which is a process that runs a continuous learning loop to further predict the performance of an industrial system, where the further predictions are more accurate and / or reliable predictions indicating the future state of the industrial system, and Method 600 is repeated to form a continuous monitoring process of the industrial system. An example of a system for implementing Method 800 is shown in Figure 7.

[0164] Method 800 includes steps 802, 804, 806, and 808. Step 802 includes calculating a third state value using data collected from a sensor set of an industrial system. Step 804 includes tuning a trained machine learning model using the third state value. Step 806 includes obtaining a fourth state value from the trained machine learning model. Step 808 includes determining a fifth state value using the third and fourth state values. Step 810 includes updating the trained machine learning model.

[0165] Step 802 includes calculating a third state value for the system using data collected from the sensor set of the industrial system during the second period. The third state value of the system relates to the performance of the industrial system during the second period and indicates the measured state of the industrial system during the second period. For example, step 802 includes determining the state of the industrial system at a future point in time, which can then be compared with a previously obtained predicted state of the industrial system for the same future point in time.

[0166] Step 804 involves tuning the trained machine learning model to reduce the difference between a second state value determined by the predictive model and a third state value of the system. For example, step 804 involves using the third state value as input to the trained machine learning model to obtain more accurate and reliable output predictions from the trained machine learning model. In another example, step 804 involves using the third state value as additional training data for the third state value to obtain more accurate and reliable output predictions. In yet another example, step 804 involves modifying the mathematical processes, parameters, or hyperparameters of the trained machine learning model based on the difference between the third and second state values, for example, by an offsite machine learning expert or a tuning algorithm.

[0167] Step 806 involves obtaining a fourth state value for the system from the trained machine learning model, which predicts the performance of the industrial system in a further future period based on the performance of the industrial system in a new current period. The performance of the industrial system in a new current period is determined using data collected during the current period from a set of sensors in the industrial system. For example, the fourth state value is obtained based on the third state value in step 802. The fourth state value of the system is associated with the predicted performance of the industrial system in a new future period that occurs after the period in which the third state value was calculated, and therefore indicates the future state of the industrial system.

[0168] Step 808 includes determining a fifth state value of the system using a predictive model configured to predict the performance of the industrial system at a new future point in time, based on the fourth state value of the system and real-time system information that optionally includes the third state value from step 802. For example, the fourth state value is adjusted using relevant values ​​such as recently measured data and the calculated third state value to form the fifth state value. Thus, the fifth state value is an updated and more reliable prediction of the new future state of the industrial system.

[0169] Step 810 is an optional additional step, which includes updating the trained machine learning model. Updating the trained machine learning model includes calculating a sixth state value corresponding to a new third state value using data collected from the sensor set of the industrial system, and adjusting the trained machine learning model to reduce the difference between the fifth state value determined by the predictive model and the calculated sixth state value of the system. For example, step 810 is an iteration of steps 802 and 804 at a new future point in time. Preferably, as the industrial system process progresses, method 800 is continuously repeated at systematic time intervals in a cycle of continuously updating and improving the predictions of the trained machine learning model and / or the predictive model.

[0170] Figure 9 shows a flowchart illustrating how quality labeling is used when updating a trained machine learning model configured to predict the performance of an industrial system. Specifically, Figure 9 shows a computer implementation method 900 that is compatible with the processes of methods 100, 400, 500, 600, 800, and / or 1000. For example, method 900 is performed between steps 802 and 804 of method 800.

[0171] Figure 9 illustrates Method 900, a process that applies quality labels to predicted state values ​​for use in updating a trained machine learning model and determines whether the available data is of sufficient quality to be used for updating the trained machine learning model. An example of a system for performing Method 900 is shown in Figure 7.

[0172] Method 900 includes steps 902, 904, 906, 908 or 910, and 912. Step 902 includes assigning a first quality label and a second quality label to a second state value and a third state value. Step 904 includes filtering a plurality of third state values. Step 906 includes determining whether the filtered plurality of third state values ​​meet a predefined criterion. Step 908 includes canceling the update of the trained machine learning model. Step 910 includes continuing the update of the trained machine learning model. Step 912 includes outputting the second state value to the user interface. The second and third state values ​​refer to predicted state values, as well as calculated and / or measured state values, such as the second and third state values ​​of Methods 600 and 800.

[0173] Step 902 includes assigning a first quality label and a second quality label to a second and a third state value. The first quality label indicates the quality of the second state value. For example, the first quality label is assigned to the second state value based on the difference between the second and third state values. The first quality label may represent binary labels such as relevance, good or bad, and / or correct or incorrect, or non-binary labels such as the percentage difference between the second and third state values ​​and / or the industrial system operator's opinion on the quality of the second state value. The second quality label indicates the quality of the third state value. For example, the third state value is calculated based on data collected from a sensor set of the industrial system. If the calculated value and / or the data collected from the sensor set falls below a quality threshold or is erroneous and / or abnormal, the third state value is assigned a second quality label indicating a low-quality third state value. Optionally, the first quality label may be assigned only to the second state value, or the second quality label may be assigned only to the third state value.

[0174] Step 904 includes filtering a plurality of third state values, for example, filtering each of the plurality of third state values ​​based on a second quality label assigned to each third state value. Alternatively, step 904 may be performed before step 902, in which the plurality of third state values ​​are filtered, and then the second quality label is assigned to the filtered third state values. The third state values ​​and / or the data associated with the third state values ​​are filtered based on quality. The data associated with the third state values ​​includes data collected from one or more sensors in the industrial system. Quality is assessed based on a number of factors, such as whether the third state value is bad / poor based on missing data in the batch (gaps in the data) or the state of the process and / or problems in data processing.

[0175] Step 906 involves determining whether a filtered set of third state values ​​meets a predefined criterion. The predefined criterion depends on quality labels and quality ratings. For example, the predefined criterion is a predefined number of third state values ​​within the filtered set of third state values. If the data quality is poor, the tuning algorithm, or for example, a machine learning expert, may not have enough data to accurately and / or efficiently update the trained machine learning model of Method 600. For example, updating a trained machine learning model with a small amount of data can lead to bias and inaccuracies.

[0176] Step 908 involves canceling updates to a trained machine learning model. For example, fitting a trained machine learning model based on a filtered set of third state values ​​is canceled for the current target third state value. Optionally, updates to one or more trained machine learning models may be canceled, in which case one or more trained machine learning models may not be tuned, or may be reset to a previous state if an update has been initiated. Alternatively, updates, fitting, or other tuning of one or more trained machine learning models may continue for different target third state values.

[0177] Step 910 includes continuing to update the trained machine learning models. Continuing to update the trained machine learning models includes initiating the fitting and / or tuning of at least one machine learning model. Furthermore, continuing to update the trained machine learning models includes tuning one or more trained machine learning models to a third state value with a different target.

[0178] Step 912 includes outputting a second state value to the user interface. Step 912 is optional; otherwise, steps 906, 908, and / or 910 may be repeated. Optionally, step 912 may further include outputting one or more filtered third state values ​​corresponding to the second state value to the user interface, so that the user of the user interface can compare the second state value with the corresponding third state value. Furthermore, the second state value and optionally the third state value are used to select one or more control actions so that the user of the user interface can select an operation to modify the process of the industrial system, as detailed in Method 1000.

[0179] Beneficial in this regard, by deciding whether to update or cancel the trained machine learning model, only high-quality data is used as input and / or training data for the trained machine learning model. Thus, Method 900 improves the accuracy and reliability of the trained machine learning model because continuous learning cycles do not cause degradation of the trained machine learning model, and data unsuitable for training the trained machine learning model, such as data associated with transient fluctuations in measurements or inaccurate sampling, does not affect the trained machine learning model.

[0180] Figures 10A-10B show flowcharts illustrating methods for automated monitoring of industrial systems and generating operational recommendations for those systems. Specifically, Figures 10A-10B show computer implementation method 1000, which is compatible with processes 100, 400, 500, 600, 800, and / or 900.

[0181] Figures 10A-10B illustrate Method 1000, which is a process for predicting state values ​​that indicate the future state of an industrial system, determining whether there are any deviations associated with the predicted state values ​​based on those predicted state values, and selecting one or more mitigation control actions to reduce the impact of the deviations. Exemplary systems for implementing Method 1000 are shown in Figures 2, 7, and 11.

[0182] Method 1000 includes steps 1002, 1004, 1006, 1008, 1010, 1012, 1014, 1016, 1018, 1020, and 1022. Step 1002 includes obtaining a first state value from a trained machine learning model. Step 1004 includes obtaining real-time system information, including data collected by a set of sensors in the industrial system. Step 1006 includes determining a second state value using a predictive model. Step 1008 includes determining whether a deviation exists using the second state value. Step 1010 includes evaluating the impact of the deviation. Step 1014 includes evaluating the differences in the performance of the industrial system. Step 1016 includes identifying the problem causing the deviation based on the evaluation. Step 1018 includes identifying one or more control actions based on the problem. Step 1020 includes prioritizing one or more mitigation control actions. Step 1022 includes outputting a visual display.

[0183] Step 1002 includes obtaining a first state value from the trained machine learning model, such as by performing step 602 of method 600. The trained machine learning model predicts the performance of the industrial system over a second period based on the performance of the industrial system over a first period. The performance of the industrial system over the first period is determined using data collected from a set of sensors of the industrial system during the first period. Furthermore, the first state value of the system is associated with the predicted performance of the industrial system over the second period and indicates the future state of the industrial system.

[0184] Step 1004 includes obtaining real-time system information, including data collected by a set of sensors in the industrial system, such as by performing step 604 of method 600, the real-time system information indicating the current state of the industrial system and including data collected by a set of sensors in the industrial system.

[0185] Step 1006 includes determining a second state value using a predictive model configured to predict the performance of the industrial system over a second period based on a first state value of the system and real-time system information, such as by performing step 606 of method 600.

[0186] Step 1008 includes determining, using a second state value, whether there is a deviation exceeding a predetermined threshold in the predictive performance of the industrial system during the second period, such as by performing step 104 of method 100, where the predetermined threshold is associated with a process in the industrial system.

[0187] Step 1010 includes, if a deviation exceeding a predetermined threshold exists, performing step 106 of method 100, and evaluating the impact of the deviation on one or more outputs of the industrial system based on the difference between the predictive performance of the industrial system in the second period and the predictive performance of the industrial system when no deviation exists.

[0188] Step 1012 includes selecting one or more mitigation control actions from a plurality of predetermined control actions, such as by performing step 108 of method 100, where each predetermined control action represents a predetermined operation for modifying one or more processes of an industrial system.

[0189] Step 1014 includes evaluating the difference between the real-time performance of the industrial system and the performance of the industrial system over a first period, such as by performing steps 402 and 404 of Method 400.

[0190] Step 1016 includes determining the problem from a set of predetermined problems causing the deviation, based on the evaluated difference, such as by performing step 404 of method 400.

[0191] Step 1018 includes identifying one or more control actions from a set of predetermined control actions based on the problem causing the deviation, such as by performing step 406 of method 400, the identified one or more control actions that mitigate the difference between the predicted performance of the industrial system and the real-time performance of the industrial system in a second period.

[0192] Step 1020 includes prioritizing one or more mitigation control actions based on calculating the impact on the industrial system if each mitigation control action is performed, such as by performing step 410 of method 400.

[0193] Step 1022 includes outputting a visual display, such as outputting a display of the effects of performing one or more mitigation control actions to the user interface, by performing step 112 of method 100. Beneficially, the visual display allows the user of the user interface to select an operation to initiate one or more mitigation control actions in order to modify one or more processes in the industrial system.

[0194] Optionally, step 1022 may further include collecting user feedback on visual representations and associated relaxation control actions. The feedback is used to manually tune the trained machine learning model and / or predictive model, such as by adjusting hyperparameters, feature selection, and / or model type. Collection is preferably automated. For example, if user ratings associated with visual representations are too low, the trained machine learning model and / or predictive model disable predictions and associated ratings / tuning in the user interface.

[0195] Beneficially, Method 1000 enables the detection, correction, or mitigation of deviations from estimated current state values ​​or predicted future state values ​​while the impact is minimized or before problems begin to occur. Thus, Method 1000 results in an optimized industrial system in which predictive monitoring can potentially be configured as an early warning system for equipment or process inefficiencies. Furthermore, Method 1000 enables updating operations based on changing conditions, process variations, and / or equipment failures occurring in real time, and / or equipment failures occurring as a result of future processes.

[0196] Figure 11 shows a high-level system architecture diagram in accordance with this disclosure for automated monitoring of an industrial system and generation of operational recommendations for the industrial system. Specifically, Figure 11 shows a semi-automated industrial system 1100 configured to perform method 1000 and aimed at monitoring the industrial system 1100 by predicting state values ​​that indicate the future state of the industrial system and recommending operations for the industrial system 1100 based on the predicted state values.

[0197] The industrial system 1100 comprises an input resource 1102, a first sensor 1104, an operating unit 1106, a process unit 1108, a second sensor 1110, an output 1112, a third sensor 1114, data collected by one or more sensors 1116, a monitoring module 1118, a performance prediction system 1120, an operation support system 1122, a corrective action database 1124, an edge-based cloud server 1126, a user interface 1128, an operation recommendation 1130, and a control action 1132.

[0198] The monitoring module 1118 obtains a first state value from the performance prediction system 1120, which is an arbitrary performance prediction system 700 in Figure 7. The performance prediction system 1120 includes a trained machine learning model configured to predict the performance of the industrial system 1100 in a second period based on the performance of the industrial system 1100 in a first period. The performance of the industrial system 1100 in the first period is determined using data collected by one or more sensors 1116 during the first period, and the set of sensors includes a first sensor 1104, a second sensor 1110, and / or a third sensor 1114. The first state value of the system is associated with the predicted performance of the industrial system 1100 in the second period and indicates the future state of the industrial system 1100. The monitoring module 1118 further obtains real-time system information. The real-time system information indicates the current state of the industrial system 1100 and includes data collected by one or more sensors 1116.

[0199] The monitoring module 1118 uses the performance prediction system 1120 to determine a second state value of the industrial system 1100. The performance prediction system 1120 is configured to predict the performance of the industrial system 1100 in a second period based on the first state value of the system and real-time system information. Using the second state value, the monitoring module 1118 determines whether there is a deviation exceeding a predetermined threshold in the predicted performance of the industrial system in the second period. The predetermined threshold is associated with the process of the process unit 1108.

[0200] If a deviation exceeding a predetermined threshold exists, the monitoring module 118 evaluates the impact of the deviation on one or more outputs 1112 of the industrial system based on the difference between the predicted performance of the industrial system in the second period and the predicted performance of the industrial system when no deviation exists. Subsequently, the monitoring module 1118 selects one or more mitigation control actions from a plurality of predetermined control actions, each predetermined control action representing a predetermined operation to modify one or more processes of the industrial system.

[0201] The selection of one or more control actions is performed by the operational support system 1122 and includes evaluating the difference between the real-time performance of the industrial system and the performance of the industrial system over a first period. This evaluation of the difference is based on one of the following: data collected from a first sensor 1104 that monitors the input resources required by the process unit 1108; data collected from a second sensor 1110 that monitors one or more processes of the process unit 1108; or data collected from a third sensor 1114 that monitors outputs related to the process unit 1108.

[0202] The operational support system 1122 determines the problem by identifying one problem from a set of predetermined problems that cause the deviation, based on the evaluated difference. The set of predetermined problems are stored in the corrective action database 1124. Subsequently, based on the problem that caused the deviation, one or more control actions are identified from a set of predetermined control actions, and the identified one or more control actions mitigate the difference between the predicted performance of the industrial system and the real-time performance of the industrial system in the second period.

[0203] The operation support system 1122 is further configured to prioritize one or more mitigation control actions based on the impact that the execution of each mitigation control action will have on the industrial system. Next, an operation recommendation 1130 representing the impact of executing one or more mitigation control actions is output from the monitoring module 118 to the user interface 1128. Based on the operation recommendation 1130, the user of the user interface 1128 can select an operation in the operation unit 1106 and initiate one or more mitigation control actions to modify one or more processes in the process unit 1108.

[0204] The user interface 1128 is configured to analyze the operation recommendation 1130 and the subsequent operation selection in the operation unit 1106, for example, by performing method 500. This analysis is used to update the corrective action database 1124 and is used by the cloud server 1126 to update the monitoring module 1118. Optionally, the cloud server 1126 may be a local server and / or computing device.

[0205] Figure 12 shows an example of a computing system for orchestrating the organization, deployment, and / or updating of policy management rules. Specifically, Figure 12 shows a block diagram of one embodiment of the computing system according to an exemplary embodiment of the present disclosure.

[0206] The computing system 1200 can be configured to perform any operation disclosed herein, for example, any operation described with reference to methods 100, 400, 500, 600, 800, 900, 1000, the operation support system 200 in Figure 2, and / or the performance prediction system 700 in Figure 7. The computing system includes one or more computing devices 1202. The computing devices 1202 of the computing system 1200 include one or more processors 1204 and memory 1206. For example, one or more processors 1204 may be one or more arbitrary general-purpose processors configured to execute an instruction set. For example, one or more processors 1204 may be one or more general-purpose processors, one or more field-programmable gate arrays (FPGAs), and / or one or more application-specific integrated circuits (ASICs). In one embodiment, one or more processors 1204 includes one processor. Alternatively, one or more processors 1204 includes multiple operably connected processors. One or more processors 1204 are communicatively coupled to memory 1206 via address bus 1208, control bus 1210, and data bus 1212. Memory 1206 may be random access memory (RAM), read-only memory (ROM), persistent storage devices such as hard drives, or erasable programmable read-only memory (EPROM). The computing device 1202 further comprises an I / O interface 1214 communicatively coupled to address bus 1208, control bus 1210, and data bus 1212.

[0207] Memory 1206 can store information accessible by one or more processors 1204. For example, memory 1206 (e.g., one or more non-temporary computer-readable storage media, memory devices) may include computer-readable instructions (not shown) that can be executed by one or more processors 1204. Computer-readable instructions may be software written in any suitable programming language, or they may be implemented in hardware. In addition, or alternatively, computer-readable instructions may be executed in separate logical and / or virtual threads on one or more processors 1204. For example, memory 1206 may store instructions (not shown) that, when executed by one or more processors 1204, cause one or more processors 1204 to perform operations such as any operations and functions that constitute the computing system 1200 as described herein. In addition, or alternatively, memory 1206 may store data (not shown) that can be acquired, received, accessed, written, manipulated, created, and / or stored. This data may include, for example, the data and / or information described herein in connection with this disclosure. In some implementations, the computing device 1202 can retrieve data from one or more memory devices located away from the computing system 1200, and / or store the data in those memory devices.

[0208] The computing system 1200 further comprises a storage unit 1216, a network interface 1218, an input controller 1220, and an output controller 1222. The storage unit 1216, the network interface 1218, the input controller 1220, and the output controller 1222 are communicatively coupled to a central control unit via an I / O interface 1214.

[0209] The storage unit 1216 is a computer-readable medium containing one or more programs, and may be a temporary or non-temporary computer-readable medium. The one or more programs, when executed by one or more processors 1204, include instructions that cause the computing system 1200 to perform the method steps of this disclosure. Alternatively, the storage unit 1216 is a temporary computer-readable medium. The storage unit 1216 is a persistent storage device such as a hard drive, a cloud storage device, or other suitable storage device.

[0210] Optionally, the industrial system 1100 and processing unit 1108 are directed to, but are not limited to, any or combination thereof, of the following: mixers and other feed preparation equipment, atomizers, gas dispersers, drying chambers, fluidized beds and other drying equipment, spray freezing equipment, pneumatic evaporators and crystallizers, centrifuges and separators, cooling equipment and heat pumps, compressors, homogenizers, granulators, cyclones and / or filters, heaters, pumps, blowers, condensers, valves, stationary washing and disinfection systems, powder processing systems, brewing systems, cooling equipment, refrigerators, distillation and fermentation systems, discharge control systems, agricultural machinery, filling, bottling and packaging systems, food processing systems, ice makers, membrane filtration systems, milking systems, liquid processing systems, pumps and valves, tablet presses, vacuum systems, bioreactor systems, perfusion and feeding systems, sterilization systems, culture medium recovery systems, harvesting systems, scaffolding equipment designs, interconnecting ducts and / or piping, and other related auxiliary equipment, as well as related processes, process parameters, control systems, bioprocess sensors, etc.

[0211] The network interface 1218 may be a Wi-Fi module, a network interface card, a Bluetooth module, and / or other suitable wired or wireless communication device. In one embodiment, the network interface 1218 is configured to connect to a network such as a local area network (LAN), a wide area network (WAN), the internet, or an intranet.

[0212] It should be noted that, as stated above, the orchestration of the method of this disclosure may, to some extent, involve processing input data and generating output data. This processing of input data and generation of output data may be implemented in hardware or software. For example, certain electronic components may be used in a control module or similar or related circuit to implement the functions related to the method of this disclosure as described above. Alternatively, one or more processors operating according to instructions may implement the functions related to the method of this disclosure as described above. In this case, it is within the scope of this disclosure that such instructions may be stored in one or more temporary or non-temporary processor-readable storage media (e.g., magnetic disks or other storage media) or transmitted to one or more processors via one or more signals embodied in one or more carrier waves.

[0213] This disclosure is not limited in scope by the specific embodiments described herein. In fact, various other embodiments and modifications of this disclosure, in addition to those described herein, will be apparent to those skilled in the art from the foregoing description and accompanying drawings. Such other embodiments and modifications are therefore intended to be included within the scope of this disclosure. Furthermore, although this disclosure is described herein in the context of at least one specific implementation for at least one specific purpose and in at least one specific environment, those skilled in the art will understand that its usefulness is not limited thereto and that this disclosure may be beneficially implemented for any number of purposes and in any number of environments. Accordingly, the following statements should be interpreted in light of the entire scope and spirit of this disclosure as described herein.

[0214] Numbered description of the invention 1. A method implemented in a computer for determining actions to improve the operation of an industrial system, the method comprising: Obtaining reference data that represents one or more states of an industrial system. The reference data is associated with data collected from one or more sensors in the industrial system. To determine whether a deviation exceeding a predetermined threshold exists within the reference data. The predetermined threshold is associated with a process in an industrial system; If a deviation exceeding a predetermined threshold exists, the impact of the deviation on one or more outputs of the industrial system is evaluated based on the difference between the performance of the industrial system at a first time point where the deviation exists and the predicted performance of the industrial system at a first time point where the deviation does not exist; Selecting one or more mitigation control actions from a set of predetermined control actions. Each predetermined control action represents a predetermined operation to modify one or more processes in an industrial system, and the selection steps include: Identifying the problem causing the deviation from a given set of problems, based on the evaluated impact of the deviation; Identifying one or more control actions from a set of predetermined control actions based on the problem identified as causing the deviation. The identified one or more control actions are selected to mitigate the impact of the deviation on one or more outputs of the industrial system; Calculate the impact that the execution of each mitigation control action has on one or more outputs of the industrial system; and To output a representation of the effects of performing one or more mitigation control actions to the user interface, and to allow the user of the user interface to select an operation to initiate one or more mitigation control actions in order to modify one or more processes in an industrial system. 2. The method described in Description 1, wherein the reference data includes the following: Data collected over a first period from one or more sensors in an industrial system. The data collected over the first period indicates the historical state of the industrial system; or Real-time data collected by one or more sensors in an industrial system. Real-time data indicates the current state of the industrial system; or Predicted state value of the industrial system. The predicted state value is associated with the predicted performance of the industrial system in the second period and is based on data collected over the first period, indicating the future state of the industrial system. 3. The method described in Description 2, for determining whether a deviation exists, includes the following: Evaluate the data collected over the first period. The data collected over the first period is time-dependent; Identify one or more trends and / or time-dependent trends over a first period in evaluated data collected over a first period. 4. One of the methods described in 1 to 3, further comprising prioritizing one or more mitigation control actions based on the calculated impact of performing each mitigation control action. 5. The operation described in Description 4, which modifies one or more processes in an industrial system, is a highest-priority mitigation control action, and if selected, it brings about a change in one or more processes in the industrial system, thereby mitigating the impact of the deviation on the industrial system. 6. An operation that modifies one or more processes in an industrial system using the method described in 4 is not a highest-priority mitigation control action. 7. A method of any of the descriptions 1 to 6, further comprising determining whether one or more mitigation control actions have been performed. 8. The method of description 7, further comprising storing feedback data, the feedback data being associated with a determination of whether any of one or more mitigation control actions have been performed. 9. The method described in 8, wherein the feedback data is used to further determine whether a deviation exists. 10. The method described in 8 or 9, wherein the feedback data is used to further identify the issues causing the deviation. 11. Any method described in 8 to 10, wherein the feedback data is used to further identify one or more control actions from a plurality of predetermined control actions. 12. Any method described in 1 through 11, further including: Update the reference data. The updated reference data will include further predicted state values ​​from the industrial system's user interface; To determine whether there is a new deviation between the further predicted state value and the updated measurement performance of the industrial system, based on updated data collected from one or more sensors of the industrial system; If a deviation exists, calculate the effect of the new deviation; and Select one or more further control actions based on the impact of the aforementioned new deviation. 13. The determination of whether or not there is a deviation is performed by a deviation classifier using one of the methods described in 1 to 12. 14. The method described in 13, wherein the deviation classifier is based on a first-principles model. 15. The method described in 13 or 14, wherein the deviation classifier is one or more trained machine learning models. 16. The method of description 15, wherein the one or more trained machine learning models are any of the following: a support vector machine, a neural network, a decision tree-based algorithm, a set-based algorithm, a generalized likelihood ratio detector, or a cumulative sum detector. 17. A method according to description 15 or 16, further comprising selecting a trained machine learning model from the one or more trained machine learning models based on one or more states of an industrial system associated with acquired reference data. 18. A method described in any of descriptions 15 to 17 wherein the deviation classifier is configured to output whether or not there is a deviation based on an input which includes reference data indicating one or more states of the industrial system. 19. Data collected from one or more sensors of the industrial system is collected over a predetermined time frame by any of the methods described in 1 to 18. 20. The method of description 19, wherein the predefined time frame is 10 minutes. 21. Data collected from one or more sensors in an industrial system is collected in real time using any of the methods described in 1 to 20. 22. A method according to any of descriptions 1 to 21, wherein data collected from one or more sensors in an industrial system includes a data sequence of a predetermined size. 23. Any method described in 1 to 22, further comprising generating a visual representation of the effects of deviation by determining a graphical interpretation relating to the effects of deviation from a list of graphical interpretations, based on mapping data related to the effects of deviation to predicted state values. 24. Any method described in 1 to 23, wherein the problem is determined using a rule-based model relating to one or more processes in an industrial system, and comprises analyzing a set of predetermined control actions based on the rule-based model. 25. Any method described in 1 to 24, wherein the impact of the deviation includes multiple impact values, which include any of the following: production indicators, operational indicators, environmental indicators, safety indicators, or efficiency indicators. 26. The method of description 25, wherein the prioritization of one or more control actions includes ordering one or more control actions based on a weighted combination of multiple influence values. 27. The method described in 26, wherein the weighted combination is adjusted to match the predicted state value. 28. Any method described in 1 to 27 further comprises initiating at least one of one or more mitigation control actions. 29. One of the methods described in 1 to 28, which visually displays the effect of the deviation to the user This further includes displaying the information on the turret's display. 30. One of the methods described in descriptions 1 to 29, wherein a plurality of predetermined control actions and a plurality of predetermined problems are stored in a corrective action database, and each problem is assigned to one or more predetermined control actions. 31. Any method described in 1 to 30, wherein data collected over a past period from one or more sensors in an industrial system is used to determine whether or not there is a deviation. 32. Any method described in 1 to 31 is used to determine whether or not there is a deviation in data associated with one or more processes in an industrial system. 33. Any method described in 1 to 32, wherein data associated with one or more components of an industrial system is used to determine whether or not there is a deviation. 34. A system comprising one or more processors and memory that stores instructions, when executed by one or more processors, causing the system to perform one of the steps described in 1 to 33. 35. A computer-readable medium containing computer-executable instructions that, when executed by one or more processors, perform any of the steps described in 1-33. 36. A computer-readable medium as defined in Statement 35, wherein the medium is non-transient. 37. A method implemented on a computer for predicting the system performance of an industrial system, comprising: Obtaining the first state value of a system from a trained machine learning model; The trained machine learning model predicts the performance of the industrial system over a second period based on the performance of the industrial system over a first period, and the performance of the industrial system over the first period is determined using data collected from the industrial system's sensor set during the first period; and The first state value of the system is associated with the predicted performance of the industrial system over a second period, indicating the future state of the industrial system; To acquire real-time system information. Real-time system information indicates the current state of an industrial system and includes data collected by the industrial system's sensor set; Determining a second state value of the system using a predictive model configured to predict the performance of the industrial system over a second period based on the first state value of the system and real-time system information; and Updating a trained machine learning model includes: To calculate a third state value of the system using data collected from the sensor set of the industrial system during the second period. The third state value of the system is associated with the performance of the industrial system during the second period and indicates the state of the industrial system during the second period; and Adjusting a trained machine learning model to reduce the difference between a second state value determined by the predictive model and a third state value of the system. 38. The method described in description 37, wherein the first period precedes the second period. 39. The method described in description 37 or 38, wherein the third state value is used to obtain a further first state value from the trained machine learning model. 40. One or more process cycles from a series of process cycles in an industrial system occur within a second period, by any of the methods described in 37 to 39. 41. Any method described in 37 to 40, further comprising storing a second state value and a third state value. 42. A method according to any of descriptions 37 to 41, wherein the calculation of the third state value is repeated at predetermined time intervals over a second period. 43. In any of the methods described in 37 to 42, the first quality label is assigned to the second state value and the second quality label is assigned to the third state value. 44. The method described in 43, which includes tuning a trained machine learning model to determine whether a first quality label matches a second quality label. 45. The update of the trained machine learning model is performed remotely using one of the methods described in 37 to 44. 46. ​​Any method described in 37 to 45, wherein the updating of the trained machine learning model is further based on a comparison of multiple second state values ​​with multiple third state values. 47. The method described in description 46, further comprising filtering a plurality of third state values. 48. A method according to Description 47, further comprising determining whether a plurality of filtered third state values ​​satisfy a predetermined criterion, the predetermined criterion being related to the number of filtered third state values. 49. If the method described in Description 48 fails to meet the specified criteria, the update of the trained machine learning model is canceled and the trained machine learning model is restored to its state before the update. 50. A method of description 47 or 48, further comprising generating one or more first state values ​​using an alternative model if a predetermined criterion is not met. 51. The method described in Description 50, wherein the alternative model is based on averaging one or more stored third state values. 52. One of the methods described in 37 to 51, wherein one or more first state values ​​obtained from a trained machine learning model are associated with a component or process of an industrial system. 53. One of the methods described in 37 to 52, wherein one or more first state values ​​obtained from a trained machine learning model are associated with the inputs and / or outputs of an industrial system. 54. The method described in 37 to 53, wherein the trained machine learning model is one of several trained machine learning models. 55. The method described in 54, wherein the multiple trained machine learning models include any of the following: linear multivariable regression models, polynomial regression models, random forest regression models, neural network models, or expert models. 56. The method described in 54, further including: Receiving a focus indicator from among multiple focus indicators. Each focus indicator is associated with a component, process, input, or output of an industrial system; and The process of selecting a trained machine learning model from multiple trained models based on a focus indicator. 57. Any method described in 37 to 56, wherein the sensor set of the industrial system includes one or more virtual sensors. 58. Any method described in 37 to 57, wherein the sensor set of the industrial system includes one or more actuators. 59. A method according to any of descriptions 37 to 58, wherein the real-time system information further includes information relating to components, processes, inputs, or outputs of an industrial system. 60. The method described in Description 59, wherein the information relating to a component, process, input, or output of an industrial system is a prediction based on data collected from one or more virtual sensors. 61. The method described in Description 59 or 60, wherein the information relating to a component, process, input, or output of an industrial system is a prediction based on one or more outputs from a trained machine learning model. 62. A method according to any of descriptions 37 to 61, wherein the data collected by a sensor set of an industrial system includes any of the following: temperature data, humidity data, pressure data, consumption data, process rate, or operating conditions of the industrial system. 63. Any method described in 37 to 62, wherein the second state value includes information associated with any of the following: a quality indicator of output from an industrial system, a quality indicator of input to an industrial system, a duration indicator of a process cycle in an industrial system, a productivity indicator of a process in an industrial system, a performance indicator of a component of an industrial system, or the state of a process in an industrial system. 64. Real-time system information is obtained from a storage database by any of the methods described in 37 to 63. 65. The method described in 64, wherein real-time system information is stored using RAM and / or PLC data storage. 66. The method described in 64 or 65, wherein the storage database resides on an edge device or a cloud server. 67. Using any of the methods described in 37 to 66, data collected from the sensor set of the industrial system during the third period is input to a trained machine learning model. 68. A method according to any one of descriptions 37 to 67, wherein the trained machine learning model is configured to predict the performance of an industrial system over a third period, based on further data relating to one or more processes in the industrial system. 69. A method according to any of descriptions 37 to 68, wherein the trained machine learning model is configured to predict the performance of the industrial system over a third period, based further on data relating to one or more components of the industrial system. 70. A method according to any of descriptions 37 to 69, wherein the predictive model is a multivariable linear model. 71. A method according to any of descriptions 37 to 70, wherein a trained machine learning model predicts the performance of an industrial system over a third period using a second state value, the third period being after the second period. 72. One of the methods described in 37 to 71, wherein a trained machine learning model predicts the performance of an industrial system over a third period using a third state value, the third period being after a second period. 73. The method of description 71 or 72, further comprising obtaining a further first state value from a trained machine learning model and generating a further second state value using a predictive model, the further second state value being associated with the performance of an industrial system over a third period. 74. Any method described in 37 to 73, further comprising generating a trained machine learning model. 75. Any method described in 37 to 74, wherein the first state value is time-dependent and / or variable. 76. Any method described in 37 to 75, wherein the first state value is independent of time and / or constant. 77. One of the methods described in 37 to 76, wherein the first state value is the first of a plurality of first state values. 78. The second state value is output to the user interface of the industrial system by any of the methods described in 37 to 77. 79. A system comprising one or more processors and memory that stores instructions that, when executed by one or more processors, cause the system to perform any of the steps described in 37 to 78. 80. A computer-readable medium storing computer-executable instructions that, when executed by one or more processors, perform any of the steps described in 37 to 79. 81. A computer-readable medium as described in 80, wherein the medium is non-temporary. 82. A computer-implemented method for automated monitoring of an industrial system and subsequent generation of operational recommendations for one or more processes of the industrial system, comprising: To obtain a first state value from a trained machine learning model. Here, The trained machine learning model predicts the performance of the industrial system over a second period based on the performance of the industrial system over a first period, and the performance of the industrial system over the first period is determined using data collected from the industrial system's sensor set during the first period; and The first state value of the system is associated with the predicted performance of the industrial system over a second period, indicating the future state of the industrial system; To acquire real-time system information. Real-time system information indicates the current state of an industrial system and includes data collected by the industrial system's sensor set; Determining a second state value of the system using a predictive model configured to predict the performance of the industrial system over a second period, based on the system's first state value and real-time system information; Using a second state value, determine whether there is a deviation greater than a predetermined threshold in the predicted performance of the industrial system during the second period. The predetermined threshold is associated with the processes of the industrial system. If a deviation greater than a predetermined threshold exists, the impact of the deviation on one or more outputs of the industrial system is evaluated based on the difference between the predicted performance of the industrial system in a second period and the predicted performance of the industrial system when no deviation exists; Selecting one or more mitigation control actions from a set of predetermined control actions. Each predetermined control action represents a predetermined operation to modify one or more processes in an industrial system, and the selection includes: To evaluate the difference between the real-time performance of an industrial system and the performance of the industrial system over a first period; The process of determining one problem from several given problems that cause a deviation, based on the evaluated differences; Identifying one or more control actions from a set of predetermined control actions based on the problem causing the deviation. The identified one or more control actions mitigate the difference between the predicted performance of the industrial system and the real-time performance of the industrial system in a second period. Prioritizing one or more mitigation control actions based on the calculation of the impact each mitigation control action has on the industrial system; and To display the effects of performing one or more mitigation control actions in the user interface, and to allow the user of the user interface to select an operation to initiate one or more mitigation control actions in order to modify one or more processes in an industrial system. 83. A system comprising one or more processors and memory that stores instructions that, when executed by one or more processors, cause the system to perform the steps described in 82. 84. A computer-readable medium containing computer-executable instructions that, when executed by one or more processors, perform any of the steps described in 82. 85. A computer-readable medium as described in 84, wherein the medium is non-temporary.

Claims

1. A method implemented in a computer for determining actions to improve the operation of an industrial system, Obtain reference data that represents one or more states of an industrial system, and the reference data is associated with data collected from one or more sensors in the industrial system. In comparing the reference data with the expected value of the reference data, it is determined whether there is a deviation greater than a predetermined threshold, and the predetermined threshold is associated with the process of the industrial system. If a deviation greater than a predetermined threshold exists, the impact of the deviation on one or more outputs of the industrial system is evaluated based on the difference between the performance of the industrial system with the deviation at a first time point and the predicted performance of the industrial system when no deviation exists at the first time point. Select one or more mitigation control actions from a plurality of predetermined control actions, each predetermined control action representing a predetermined operation to modify one or more processes in an industrial system, and the selection is, Identifying the problem causing the deviation from a given set of problems based on the impact of the evaluated deviation; and Based on the problem identified as causing the deviation, one or more control actions are identified from a set of predetermined control actions, and the identified one or more control actions are selected to reduce the impact of the deviation on one or more outputs of the industrial system. Includes, Calculate the impact that the execution of each mitigation control action has on one or more outputs of the industrial system. A method comprising outputting a display of the effects of performing one or more mitigation control actions to a user interface, and enabling a user of the user interface to select an operation to initiate at least one of the one or more mitigation control actions in order to modify one or more processes in an industrial system.

2. The method according to claim 1, wherein the reference data is Data collected over a first period from one or more sensors of the industrial system, wherein the data collected over the first period is data indicating the past state of the industrial system, or Real-time data collected by one or more sensors of the industrial system, wherein the real-time data is data indicating the current state of the industrial system, or The predicted state value of the industrial system, wherein the predicted state value is associated with the predicted performance of the industrial system in a second period and is based on data collected over the first period, and the predicted state value is data indicating the future state of the industrial system. Methods that include...

3. The method according to claim 2, wherein determining whether the deviation exists is: The data collected over the first period is evaluated, and the data collected over the first period is time-dependent, To identify one or more trends and / or time-dependent trends over the first period in the evaluated data collected over the first period. Methods that include...

4. A method according to any one of claims 1 to 3, further comprising prioritizing one or more mitigation control actions based on the calculated effects of performing each mitigation control action.

5. A method according to any one of claims 1 to 4, further comprising determining whether one of one or more mitigation control actions has been performed.

6. A method according to claim 5, further comprising storing feedback data, wherein the feedback data is associated with determining whether any of one or more mitigation control actions have been performed.

7. The method according to claim 6, wherein the feedback data is Further determination of whether a deviation exists, Further identification of the problems causing deviations from the aforementioned multiple predetermined problems, or A method used to further identify one or more control actions from the plurality of predetermined control actions.

8. A method according to any one of claims 1 to 7, further, The reference data is updated, and the updated reference data includes further predicted state values ​​from the user interface of the industrial system. Based on updated data collected from one or more sensors in the industrial system, it is determined whether there is a new deviation between the further predicted state value and the updated measurement performance of the industrial system. If a deviation exists, calculate the effect of the new deviation. A method comprising selecting one or more further control actions based on the effects of the new deviation.

9. A method according to any one of claims 1 to 8, wherein the determination of whether or not there is a deviation is performed by a deviation classifier, the deviation classifier is one or more trained machine learning models.

10. A method according to claim 9, further comprising selecting a trained machine learning model from one or more trained machine learning models based on one or more states of an industrial system associated with the acquired reference data.

11. A method according to any one of claims 1 to 10, further comprising generating a visual representation of the effects of deviation by determining a relevant graphical interpretation relating to the effects of deviation from a list of graphical interpretations, based on mapping data associated with the effects of deviation to predicted state values.

12. A method according to any one of claims 1 to 11, wherein the determination of the problem is made using a rule-based model associated with one or more processes in an industrial system, and the method comprises analyzing the plurality of predetermined control actions based on the rule-based model.

13. A method according to any one of claims 1 to 12, wherein the effect of the deviation includes a plurality of effect values, which include any of a production indicator, an operational indicator, an environmental indicator, a safety indicator, or an efficiency indicator.

14. A system comprising one or more processors and a memory that stores instructions that, when executed by the one or more processors, cause to perform any of the steps in claims 1 to 13.

15. A computer-readable medium containing computer-executable instructions that, when executed by one or more processors, perform any of the steps in claims 1 to 13.

Images