Learning-based determination of operating state of production process

Abstract

The invention relates to learning-based determination of an operating state of a production process, in particular to a computer-implemented method, a computer-implemented device, a system and a computer program product for determining an operating state of an industrial production process, the method includes: providing process data associated with a production process to at least one trained first neural network, where the first neural network includes at least one anomaly recognition model, and the process data is provided to each of the at least one anomaly recognition model; determining, based at least in part on the trained first neural network, whether the process data can be assigned to a predefined operating state; and providing a first indicator indicating whether the process data can be assigned to the operating state.

Description

Technical Field

[0001] This invention relates to a computer-implemented method, a computer-implemented apparatus, a system, and a computer program product for determining the operating state of an industrial production process, and to a corresponding training method for a neural network. Background Technology

[0002] Modern industrial production processes are mostly complex sequences of individual steps. In order to achieve optimized manufacturing processes, accurate monitoring over time is crucial. In particular, identifying and characterizing specific operational states (i.e., being able to describe the overall state of the manufacturing process as it progresses) is considered very important.

[0003] Industrial production processes can involve a series of physical and / or chemical reactions. The resulting sequence of physical and / or chemical reactions can ultimately define the potentially dynamically changing operating state.

[0004] Most commonly used methods for determining the current operating status are inefficient and / or not reliable enough, making it impossible to guarantee that the production process can be adequately characterized or monitored over time in all cases.

[0005] If the operational status cannot be reliably and accurately described, the production process will face the problem of not being able to fully determine whether the expected target state has been reached, which can lead to potential delays, errors, and poor performance. Even if it seems that the operational status can be determined manually by the operator, this necessity of manual (operator) processing still limits achievable productivity and increases the risk of human error.

[0006] For example, another type of uncertainty arises when the production process is in an operational state that cannot be categorized into any known operational state. In such cases, it is often impossible to guarantee that the operational state can be determined with sufficient accuracy.

[0007] To date, the problem of determining the operating state has largely been addressed by assuming (deterministically) a sequential execution of operating states. In such cases, the production process follows the assumption of a predefined sequence (e.g., operating states), where the final state is determined based on confirmation that specific (implementation) steps have been completed. Thus, for example, the production process can begin execution in a predefined state and then be transformed into other states through a predefined execution order. However, this approach still has the drawback of achieving accurate determination of the current operating state because it merely traces back the predefined sequence without considering real-time feedback. This also particularly ignores the ubiquitous disturbances present in the field of process automation.

[0008] Alternatively, attempts have been made to determine the running state based on fixed, predefined definitions, such as a fixed range of parameters. While this achieves state-level identification, it remains limited by its inflexibility and lack of capability to adapt to changing process conditions. Predefined definitions cannot cover all possible scenarios, which can lead to errors and inaccuracies in determining the corresponding running state.

[0009] Especially in high-dimensional spaces, determining the single value of a vector can be very challenging. In such scenarios, proportions / relationships are often interrelated. In some cases, these can be used as additional dimensions within the state space, allowing for computation or mapping, thereby reducing the overall dimensionality.

[0010] However, existing solutions have generally proven to be insufficient and inaccurate, failing to determine operational status based on process data and track it over time. The limitations of sequential processes and fixed, predefined state definitions necessitate a more advanced and intelligent system capable of adapting to dynamic process conditions, requiring real-time feedback based on available process data for accurate state detection and process automation.

[0011] In other cases, it may be necessary to know the possible sequences of different operating states. In some situations, this may fail because the unknown state has been identified, preventing further processing.

[0012] In such cases, it may be necessary to rely on the user's experience to determine the corresponding state, navigate its sequence, and reach the final or target state. Operators often refer to operating instructions or literature to understand the basic process flow and select from available subsequent states. However, this process proves time-consuming, error-prone, and inefficient (primarily in complex systems involving multiple states and transitions). Some automation functions can achieve basic management and visualization of process states, but often fail due to a lack of comprehensive support for identifying process states (i.e., determining theoretically available process states) and for determining the closest possible state when no state is currently defined.

[0013] Industry standards ISA-106 and ISA-88 can help define state-based procedures because these specifications and best practice solutions can help design and implement state-based control systems.

[0014] Systems configured to define or characterize state-based approaches are known, such as Distributed Control Systems (DCS). DCS is a commonly used automation system in process industries, capable of being configured to support state-based control strategies. DCS platforms typically provide functions for monitoring process states, managing transitions, and executing control actions based on process requirements.

[0015] Furthermore, batch control systems can be considered, typically understood as batch control systems based on the ISA-88 standard, and applicable in industries involving batch production processes. These systems enable the management and control of complex processes involving multiple states and transitions. While State Based Control (ISA106) largely focuses on continuous processes and possible transitions from one quasi-static state to another, typical batch processes are characterized by numerous distinct states following each individual process step. In some cases, long-term archiving systems, or historian databases, can be used (for process data). Process history databases are data management systems that collect and store process data. They can be used to track and document state-based process execution in writing, allowing for (post-hoc) review of the system's historical development and behavior. In some cases, Advanced Process Control (APC) systems can also be used. APC systems utilize advanced algorithms and models to optimize process performance and control. These systems can be incorporated into strategies for process management to improve process efficiency and stability. Furthermore, Manufacturing Execution Systems (MES) can be implemented. MES platforms provide real-time transparency and control over manufacturing processes. They can include state-based control capabilities to manage and track the execution of methods and workflows.

[0016] Therefore, improved operational status determination is needed. Summary of the Invention

[0017] According to a first aspect, a computer-implemented method is proposed for training a first neural network to identify the operating state of an industrial production process. The method includes: defining at least one operating state; determining at least one time interval during which the industrial production process is in an operating state to be identified; and providing a first allocation of process data associated with the production process for the operating states to be identified activated during the time interval. Further, the computer-implemented method can include: training the first neural network based on the provided first allocation, wherein the first neural network includes at least one anomaly detection model, and wherein the neural network is trained to perform the allocation of process data associated with the production process for the operating states activated during the specific time interval based on the at least one anomaly detection model.

[0018] Industrial production processes can consist of a continuous sequence of operating states. In some cases, in addition to the (predetermined) continuous sequence, operating states can also exist in a branched manner (where branches are conditionally invoked, for example).

[0019] Defining the operating state can include determining the numerical values ​​of process data so that when the process value group associated with the process data is within the target corridor, it is possible to determine that the industrial production process is in a specific target state associated with the relevant process data.

[0020] In this document, process data can be understood as all data that can be collected during the production process. This can refer to, for example, sensor data and / or calculated data (e.g., at least partially derived from sensor data). This can include, for example, the collection of temperature data, pressure data, weight data, speed data, throughput data, and / or other suitable data. In some cases, offline data, such as those indicating the characteristics of the initial product used, can also be used. In some cases, process data can include voltage and / or current.

[0021] The time interval can be selected, for example, within 10 seconds, within 30 seconds, within 1 minute, within 5 minutes, within 30 minutes, within 1 hour, within 5 hours, within 12 hours, within 24 hours, or more than 1 day. At least one time interval can include more than one time interval, such as at least two time intervals, and can provide at least one time interval whose length remains constant. In some cases, the length of at least one time interval can vary over time.

[0022] In this way, a training method for the first neural network can be provided efficiently, enabling the trained first neural network to determine whether the provided process data can, in principle, be assigned to a (predefined or known) operating state. This makes it possible to improve the monitoring of industrial production processes. In particular, this allows for the determination of the latest operating state without knowing or considering the historical context of the current operating state. In particular, this allows for targeted and efficient jumps to operating states, for example, as part of a continuous sequence of operating states.

[0023] According to one implementation, the operating state can be associated with the manufacturing of a specific product or with the initial operating state.

[0024] In this paper, "operational state" can be understood as a process step in an industrial production process. A process step is characterized by specific processing or waiting (e.g., a period of time during which no specific processing associated with the production process is performed). In other words, an operational state can be associated with the manufacture of a specific product. In such exemplary cases, the production process may, for example, involve the manufacture of substance C. This manufacture may include operational states such as "providing a first raw material," "providing a second raw material," and "mixing the first raw material with the second raw material," thereby forming substance C through mixing. The corresponding operational state can be represented by the specific numerical representation of process data.

[0025] The initial operating state can be understood as the operating state of an industrial production process that is assumed to be the first operating state when the process is initially started. In some examples, the initial operating state may include the initialization of the industrial production process. Initialization may include, for example, setting the devices to be used within the framework of the industrial production process to predetermined (initial) starting values ​​(such as liquid level, temperature, pressure, flow rate, etc.).

[0026] This can help to deterministically map industrial production processes and thereby support improved monitoring of industrial production processes.

[0027] According to another embodiment, the first neural network includes an anomaly recognition model for each operating state to be identified, and each operating state to be identified is associated with an anomaly recognition model that identifies the operating state to be identified as having no anomaly.

[0028] In this paper, the anomaly detection model can be understood as a model based on at least one artificial intelligence (e.g., in the form of at least one neural network) that is trained to identify deviations (i.e., anomalies) from at least one rated state.

[0029] In this way, multiple operating states can be identified in an efficient and reliable manner, thereby enabling improved monitoring of industrial production processes.

[0030] According to another implementation, training can include: training each anomaly detection model such that each anomaly detection model identifies the assigned running state as non-anomaly, and identifies other running states different from the assigned running state as anomaly.

[0031] An anomaly detection model can be trained such that when a running state to be identified is used for a running state for which the anomaly detection model has been trained, the anomaly detection model identifies the running state as anomaly-free. In such cases, the running state that should be identified by the anomaly detection model can be understood as the rated state used for that anomaly detection model. This rated state can be trained on good data, where good data can include process data that can be associated with a particular running state, and from the perspective of the anomaly detection model, this particular running state should be labeled "good" or "anomaly-free".

[0032] For example, when at least one anomaly detection model confirms that the collected process data does not deviate from the allocation of the process data for the operating state, the operating state can be identified as anomaly-free, wherein the anomaly detection model has been trained with respect to that operating state. Conversely, if a deviation (e.g., deviation value, new features previously unknown, etc.) is confirmed compared to the allocation of the process data for the operating state, an anomaly can be inferred, wherein the anomaly detection model has been trained with respect to the operating state.

[0033] In this way, a training method can be provided that can be used to better determine the existence of operating states.

[0034] According to another embodiment, the computer-implemented method can further include: providing a second allocation to a second neural network including a classifier, which allocates process data to a running state respectively; and training the second neural network based on the provided second allocation, wherein the second neural network is trained to allocate process data to the running state.

[0035] It can provide a second allocation of process data, enabling the process data to be preferentially and exclusively associated with the operating state.

[0036] In this way, in addition to ensuring that the collected process data can be assigned to an operating state in principle (based on the assumption that no anomalies exist), the second neural network can further assign process data to specific operating states. This can help to further improve the monitoring of at least one operating state of an industrial production process.

[0037] According to a second aspect, a computer-implemented method for determining the operating state of an industrial production process is proposed. The computer-implemented method includes: providing process data associated with the production process to at least one trained first neural network, the first neural network having been trained by the method described above, wherein the first neural network includes at least one anomaly detection model, and the process data is provided to each of the at least one anomaly detection model. Further, the computer-implemented method includes: determining, at least in part, whether the process data can be assigned to a predefined operating state based on the trained first neural network; and providing a first indicator indicating whether the process data can be assigned to an operating state.

[0038] It can provide the primary indicator as a string, message, or in other suitable manner.

[0039] Based on a trained first neural network, it is possible to improve the monitoring of industrial production processes, for example, by supporting the deterministic determination of the flow of industrial production processes in an efficient manner.

[0040] According to one implementation, determining whether process data can be assigned to one of the operating states further includes: determining that process data cannot be assigned to any of the operating states when each of at least one anomaly identification model identifies the process data as having an anomaly; or determining that process data can be assigned to one of the operating states when one of at least one anomaly identification model identifies the process data as having no anomaly; or determining that process data cannot be assigned to any of the operating states when more than one anomaly identification model is provided and when at least two anomaly identification models identify the process data as having no anomaly.

[0041] In some examples, if a state is provided with more than one anomaly detection model and at least two of the anomaly detection models identify the process data as anomaly-free, then that state can exist only temporarily. In some examples, this state can be understood as requiring retraining of at least one of the more than one anomaly detection model.

[0042] Each anomaly detection model can be identified as having an anomaly in the provided process data if there is no identification of an operating state that can be associated with the provided process data, and no anomaly detection model has been trained on it. Therefore, the provided process data cannot be assigned to any operating state for which its first neural network was trained.

[0043] If the anomaly detection model identifies the provided process data as normal, this is likely because the model has already been trained on the identification of the relevant operating state (based on the provided process data). In some examples, only a single anomaly detection model identifies the provided process data as belonging to the operating state.

[0044] In some examples, it can be determined that the provided process data can be assigned to more than one running state based on the following condition: the provided process data is similar to multiple process data used in the training of the first neural network within the framework of assigning process data to running states.

[0045] In this way, it is possible to statistically improve the determination of whether the provided process data can be assigned to the operating state of the first neural network that has been pre-trained for its recognition (due to the use of multiple anomaly recognition models).

[0046] According to another embodiment, the computer-implemented method may further include: when process data can be allocated to an operating state, providing the process data to a second neural network including a classifier, the second neural network having been trained by the above method. Further, the computer-implemented method may include: determining, via the second neural network, the allocation of process data for an operating state; and providing a second indicator indicating the allocation of process data for an operating state.

[0047] The second neural network can be logically or structurally different from the first neural network (for example, the first and second neural networks can have different depths and / or different numbers of nodes in each layer).

[0048] It can provide a second indicator as a string, a message, or in other suitable ways.

[0049] In this way, it is possible not only to determine whether the provided process data can be supplied to the (predefined) operating state, but also to subsequently determine which operating state the industrial production process is in (among multiple operating states if necessary) at a specific time interval.

[0050] According to another embodiment, the determination may further include: when a definite allocation cannot be achieved by providing process data to the first neural network and / or the second neural network, determining, at least in part, based on the calculation of metrics, the operating state in which process data can be allocated with predetermined determinism.

[0051] A metric can be understood as a tool for determining the distance from process data to (at least one) classification group. In some examples, a metric distance can be determined between process data and at least one predefined classification group. In some cases, the acquired process data can be plotted in at least a two-dimensional coordinate system, where a point cloud can be determined from the process data. A metric can, for example, include determining the metric distance between the focal points of the acquired point cloud. It should be noted that the two-dimensional case described herein is only exemplarily understood, and higher-dimensional or one-dimensional (i.e., generally n-dimensional) scenarios are also possible.

[0052] In this way, the operating status can be quantitatively determined.

[0053] According to another embodiment, the provision of the second indicator further includes: providing the second indicator to a user via a user interface; determining at least one subsequent running state that may follow the determined running state; providing the user with the determined at least one subsequent running state that may follow; accepting user input via the user interface, the user input indicating that execution should proceed to the following subsequent running state; and performing a transition of the following subsequent running state based on the accepted user input.

[0054] In some cases, the user interface can be provided as a touch-sensitive display (e.g., a touchscreen display). Additionally or alternatively, the user interface can include sound output tools (e.g., speakers) and / or sound input tools (e.g., microphones). The user interface can be configured to enable two-way communication with the user.

[0055] For example, user input can be accepted via a keyboard and / or mouse. Additionally or alternatively, accepting user input can include voice acceptance (e.g., voice-based input) and / or accepting user input via a touch-sensitive display.

[0056] In some examples, potential subsequent runtime states can be provided as a single suggestion for a potential subsequent runtime state. Alternatively, potential subsequent runtime states can also be provided within a framework that offers multiple potential subsequent runtime states. In the latter case, it is possible to accept user input that includes selecting the desired subsequent runtime state.

[0057] This enables users to interact with computer-implemented methods and allows them to exert targeted or desired influence on industrial production processes. The latter enables industrial production processes tailored to specific circumstances.

[0058] According to another embodiment, the computer-implemented method can include: analyzing the transition between two consecutive running states executed in history; and, based at least in part on the analysis, determining a prediction of a future expected transition from a given running state to a potentially following subsequent running state.

[0059] This analysis can include analyzing which nth running state followed the (n-1)th running state in the past (i.e., historically). The analysis can include determining statistical evaluations indicating the (relative) frequency with which a particular nth running state followed the (n-1)th running state in the past. The median determined by the (relative) frequency can indicate which following running state is most likely to follow the current running state.

[0060] In this way, based on historical data, a sequence of future production process operation states can be obtained. This enables improved prediction of potential subsequent operation states.

[0061] According to the third aspect, a computer program product is proposed, including instructions that, when executed by a computer, cause the computer to perform the steps of a computer-implemented method as described herein.

[0062] Computer program products (such as computer program tools) can be provided or delivered, for example, as storage media such as memory cards, USB drives, CD-ROMs, and DVDs, or as files downloadable from a network server. This can be achieved, for example, by transmitting the corresponding files, including the computer program product or computer program tool, over a wireless communication network.

[0063] According to a fourth aspect, a computer-implemented apparatus for determining the operating state of an industrial production process is proposed. The computer-implemented apparatus includes a first providing unit for providing process data associated with the production process to at least one trained first neural network, wherein the first neural network is trained by the method described above, and wherein the first neural network includes at least one anomaly detection model, and the process data is provided to each of the at least one anomaly detection model. Further, the computer-implemented apparatus includes: a determining unit for determining, at least in part, based on the trained first neural network, whether the process data can be assigned to a predefined operating state; and a second providing unit for providing a first indicator indicating whether the process data can be assigned to an operating state.

[0064] The corresponding unit, such as the first providing unit, the determining unit, and / or the second providing unit, can be implemented in hardware and / or software technologies. If implemented in hardware, the corresponding unit can be designed as an apparatus or part of an apparatus, such as a computer or a microprocessor. If implemented in software, the corresponding unit can be designed as a computer program product, function, routine, part of program code, or an executable object.

[0065] A neural network can be understood as a computer-based machine learning model. It mimics the way the human brain functions. It can be composed, for example, of interconnected artificial neurons arranged in multiple layers: an input layer, one or more hidden layers, and an output layer. Each neuron can connect to other neurons and has specific weights and thresholds. A neural network processes input data by guiding it through different layers, where each layer analyzes and transforms the data. Through training processes involving large amounts of data (e.g., as described in this paper), the network can learn to recognize samples and perform tasks such as classification, prediction, or decision-making (as described in this paper). The ability of a neural network to model the complex nonlinear relationship between input and output data enables it to generalize and respond to new, unknown inputs.

[0066] According to one embodiment, the computer-implemented apparatus may further include an execution unit for performing the steps of the computer-implemented method described above and / or another execution unit for performing the computer program product described above.

[0067] The execution unit and / or the other execution unit may include a field-programmable gate array (FPGA) and / or a central processing unit (CPU) and / or other suitable computing units.

[0068] According to the fifth aspect, a system for determining the operating state of an industrial production process is proposed. This system includes a computer-implemented apparatus as described herein and a computer program product as described herein.

[0069] In some examples, a computer-implemented apparatus can be provided for training a first neural network to identify operating states of an industrial production process. The apparatus includes: a defining unit for defining at least one operating state; a determining unit for determining at least one time interval during which the industrial production process is in an operating state to be identified; a providing unit for providing process data associated with the production process for a first assignment of the operating states to be identified activated during the time interval; and a training unit for training the first neural network based on the provided first assignment, wherein the first neural network includes at least one anomaly detection model, and wherein the neural network is trained to assign process data associated with the production process to operating states activated during specific time intervals based on the at least one anomaly detection model.

[0070] Furthermore, it is possible to provide a first and / or second neural network that has been trained as described herein.

[0071] Although the implementation methods described herein have been described in isolation from each other, it should be noted that they can also be combined arbitrarily and independently.

[0072] The implementation methods and features described for the proposed apparatus are correspondingly applicable to the proposed method, and vice versa.

[0073] Other possible implementations of the invention include combinations of features or implementation methods not explicitly mentioned in the preceding or hereinafter descriptions of the embodiments. Those skilled in the art can also add individual aspects as improvements or additions to the corresponding original forms of the invention. Attached Figure Description

[0074] Furthermore, the present invention has been described in more detail with reference to the accompanying drawings and preferred embodiments.

[0075] Figure 1 An exemplary method for determining the running status is shown; Figure 2 A flowchart illustrating an exemplary computer-implemented method for training a first neural network is shown. Figure 3 An exemplary computer-implemented method for determining the operating status of an industrial production process is shown. Figure 4 An exemplary computer-implemented apparatus for determining the operating status of an industrial production process is shown; and Figure 5 An exemplary system for determining the operating status of an industrial production process is shown. Detailed Implementation

[0076] Unless otherwise specified, in the accompanying drawings, the same or functionally equivalent elements use the same reference numerals.

[0077] Figure 1 Figure 100 illustrates an exemplary method for determining the operating status of an industrial production process.

[0078] Figure 100 shows an exemplary water mass flow 110 and an exemplary methanol mass flow 120 (both in kg / kg) as would be used in an exemplary industrial production process.

[0079] These exemplary quality streams can be used as process data to determine the operational status of industrial production processes.

[0080] For this purpose, specific data such as measurement data (sensor data) can be collected, which can then be used as process data.

[0081] In some cases, the acquired measurement data can be associated with the operating state by assigning the existence of a specific performance to a corresponding operating state. In other cases, these operating states can be predefined by the user using a computer-implemented method, where, for example, these operating states should be determined within the framework of the methods described herein.

[0082] Such pre-definition can include, for example, defining at least one time interval within which the specific performance of process data should be observed and, if necessary, correlated with a pre-defined operating state.

[0083] Next, a first neural network can be trained to identify at least one operational state (as described herein) based on process data provided for a certain time interval. In some cases, the first neural network can include pre-trained State Vector Machines (SVMs) and / or Isolation Forests (IFs), respectively (as described herein).

[0084] In one example, the presence of a first manifestation of the water mass flow 110 and the presence of a manifestation of the methanol mass flow (e.g., a first measurement) can be associated with a first operating state 130. In some cases, a trained first neural network can be used to determine whether the state of the corresponding first manifestation can actually be associated with an operating state, i.e., whether a corresponding operating state is predefined. This can be implemented, for example, based on an inverse finite element (IF). For example, an IF can be used to identify undefined operating states or deviation values. Defined operating states can be identified using an SVM.

[0085] Over time, the first performance of the water mass flow 110 will change, and may, for example, decrease to a second performance of the water mass flow 110.

[0086] At the same time (i.e., within the same (user-defined) time interval), the first manifestation of methanol mass flow 120 can also increase from the first manifestation to the second manifestation. The existence of the second manifestation of water mass flow 110 and the second manifestation of methanol mass flow 120 can be associated with the existence of the second operating state 140.

[0087] As time progresses, the water mass flow 110 may, for example, increase from the second state to the first state. Similarly, during this observation period, the ethanol mass flow 120 may also decrease from the second state back to the first state. The corresponding states of the water mass flow 110 and the ethanol mass flow 120 obtained here correspond to the state of the first operating state 130 described above.

[0088] As time progresses, the water mass flow 110 can increase from a first performance level to a third performance level, wherein the third performance level is higher than the first and second performance levels of the water mass flow 110. Within the same associated time interval, the methanol mass flow 120 can decrease from this performance level to a third performance level, wherein the third performance level is lower than the first and second performance levels.

[0089] The existence of the third manifestation of water mass flow 110 and the third manifestation of methanol mass flow 120 can be associated with the existence of the third operating state 150.

[0090] As time progresses, the water mass flow 110 can decrease from the third state back to the first state. Furthermore, within the associated time interval, the methanol mass flow 120 can also increase from the third state back to the first state. The resulting situation can then be re-associated with the existence of the first operating state 130.

[0091] Depending on some factors, a transition of running state can also be triggered and / or selected (e.g., a predetermined selection based on a possible subsequent running state).

[0092] For example, it's possible to provide relevant and ongoing operational status updates based on the user interface. The user interface can offer visual representations of operational status, transitions, and other related information. This further enhances the user-friendliness of monitoring or selecting operational status. In this case, the interface can be designed to be particularly intuitive, minimizing potential user training time and thereby improving user efficiency. Furthermore, this approach reduces the risk of potential errors.

[0093] For example, it can be based on a user assistance system to enable the selection of subsequent operating states.

[0094] In some cases, real-time monitoring of industrial production processes can be provided. In such cases, the system (e.g., the system described herein) can continuously monitor the process in real time, thereby determining its current operating status and keeping its progress up-to-date. This allows users to obtain the latest information (e.g., the latest operating status) about the industrial production process at any time.

[0095] Furthermore, it provides transition management. The system can not only display possible transitions between states, but also drive the actual transitions and the execution of corresponding subsequent operating states. Transition management specifically allows users to trigger transitions and perform other processes to change the industrial production process from its current operating state to the next following operating state (or target state).

[0096] Furthermore, it can provide a determination of the operating state. In such cases, the system can pre-set an algorithm (e.g., based on machine learning) to determine the current operating state. This can minimize, for example, the manual work associated with determining the current operating state and the errors that may arise as a result.

[0097] In some cases, the operating state may not be (explicitly) determined. In such cases, the next suitable (defined, i.e., the operating state that would be recognized by the first neural network) operating state can be identified, or a metric (e.g., a distance metric) can be determined to reach the defined operating state. Based on a specific metric, the defined operating state that is closest to the collected process data can be selected. In this case, the information thus obtained can support the user in determining the latest operating state of the industrial production process and, if necessary, characterizing it.

[0098] In some cases, management support or decision support can be provided. This is achieved for the user by highlighting possible transitions from the current operating state to the target operating state. In such cases, this helps the user make informed decisions and determine the appropriate target operating state based on available information. In some cases, the transition from the current operating state to the subsequent operating state can also be automated (e.g., through the system and / or suitable devices). In some cases, this automation can be achieved based on algorithms.

[0099] In some cases, performance prediction can be provided. Based on the executed transformations, the expected state of the process can be predicted. This enables users to obtain more targeted information about the system behavior associated with industrial production processes and make corresponding decisions. This can improve the overall efficiency and performance of the system.

[0100] In some cases, integration and automation technologies can be provided. These technologies, as described in this paper, can be seamlessly embedded into existing control environments, such as distributed control systems (DSCs) via protocols like OPC. This enables the system to interact with logically subordinate control systems and, when necessary, trigger actions critical to changing operational states. In some cases, automation technologies can enable both direct communication (e.g., through the system and / or corresponding devices (as described herein)) and communication-free actuators or sensors. This further improves functionality and efficiency, providing users with a seamless and efficient workflow.

[0101] In some cases, the system or apparatus described herein can be configured to manipulate features of connected automation technologies, such as thresholds for alarms and events in the DCS or user permissions in the DCS based on currently identified operating states.

[0102] The system can connect to program specifications or relevant documents (such as standard operating procedures), which are associated with specific program states and can provide users with information about those states. Furthermore, this allows the system to provide users with more diverse or background information.

[0103] The advantages of the present invention can be seen, for example, in the aspects described below: For operators, the increased efficiency is due to comprehensive support, integration with the DCS, and a user-friendly interface, which enable streamlined workflows and reduced time and labor costs associated with method-based process operations. This leads to increased productivity and efficiency.

[0104] Improved accuracy: Status identification, recent status assessment, and prediction capabilities minimize errors and improve process accuracy. Operators can rely on system instructions and information, thereby reducing the risk of errors and ensuring reliable process execution. In abnormal situations, the system supports operator guidance back to normal / defined / deterministic operating states.

[0105] Improved decision-making: Guidance, decision support, and visualization of potential transitions enable operators to make informed decisions when selecting target states and introducing transitions. This contributes to better decision-making and optimized process control.

[0106] Reduced training requirements: The system's intuitive user interface and automated functions reduce operator training needs. Operators can quickly familiarize themselves with the system and effectively navigate status-based process flows.

[0107] In some cases, such as within the framework of another embodiment that can be combined with the first embodiment described herein, it is possible to train a first neural network, which can then be used to identify the operating status of an industrial production process.

[0108] The training can include defining at least one operating state of an industrial production process. This at least one operating state can be associated, for example, with the production of product A or the initial operating state of the industrial production process.

[0109] Next, it is possible to define at least one time interval (e.g., for each of at least one operating state) that describes a specific time phase in which the industrial production process is in a particular operating state.

[0110] Furthermore, a training process of at least two stages can be applied. In this case, a first neural network (as described herein) can be trained in the first step to identify bias values ​​(i.e., anomalies) or new features not included in the training data. This can be done separately for all available time intervals. This can be achieved, for example, by using an isolation forest or a support vector machine. In some examples, the first neural network can also include at least one anomaly detection model that has been trained for all time intervals of the available process data.

[0111] In the second step, a second neural network (as described herein) can be trained. After training, the second neural network can be used to assign defined running states to predefined categories of running states using a classification model (e.g., a support vector machine). The second neural network can also include at least one anomaly detection model that has been trained on a particular running state. In such cases, for example, at least one anomaly detection model for a particular running state (which has been trained on that particular running state) will treat that running state as non-anomaly-free. In all other cases, at least one trained anomaly detection model will identify a running state (which has not been trained on its anomaly detection model) as anomaly-free.

[0112] The model can be trained using data from selected time intervals, which describe the defined state of an industrial production process.

[0113] Next, the trained first neural network and the trained second neural network can be used to determine at least one operating state of an industrial production process. In this case, it can first be pre-set that, with the help of the trained first neural network, the collected process data can, in principle, be assigned to an operating state (as described in this paper). If the trained first neural network determines that it cannot be assigned, it feeds back a first indicator that indicates either an unknown state exists or that it cannot be explicitly assigned. For example, this occurs when at least one anomaly detection model identifies an anomaly.

[0114] If the collected process data is not assigned to an unknown or undefined operating state, a trained second neural network can be used to determine the specific operating state to which the collected process data can be assigned. If at least one anomaly detection model is used to determine which specific operating state the collected process data can be assigned to, then the "searched" operating state is the operating state associated with the anomaly detection model that did not output any anomalies to the collected process data. If multiple anomaly detection models all identify the process data as anomaly-free, then the collected process data cannot be definitively assigned to an operating state.

[0115] In an alternative embodiment that can be combined with the first and / or second embodiments (as described herein), the identification of at least one operating state can be understood as a soft measurement model. The target value can be a defined operating state. The soft sensor can train the operating state and the uncertainty of the prediction.

[0116] During operation, soft sensors can predict the operating state. When an industrial production process is in an unknown state, uncertainty can be a measure of how far / close the state originally expressed by process data is from the known operating state.

[0117] In particular, the aspects described in this article enable state-based process operation according to standards ISA-88 and ISA-106, thereby achieving compatibility and industry compliance. This enables users to achieve determinism and reliability in industrial production processes.

[0118] Figure 2 A flowchart of an exemplary computer-implemented method 200 is shown, which is used to train a first neural network to identify the operating status of an industrial production process.

[0119] In step 210, define at least one running state.

[0120] In step 220, at least one time interval is determined in which the industrial production process is in an operating state to be identified.

[0121] In step 230, the following is implemented: providing a first allocation of process data associated with the production process for an identified operating state that is activated during the time interval.

[0122] In step 240, a first neural network is trained based on the provided first allocation, wherein the first neural network includes at least one anomaly detection model, and wherein the neural network is trained to perform the allocation of process data associated with the production process for operating states activated at specific time intervals based on the at least one anomaly detection model.

[0123] Figure 3 An exemplary computer-implemented method 300 for determining the operating status of an industrial production process is shown.

[0124] In step 310, process data associated with the production process is provided to at least one trained first neural network, wherein the first neural network has been trained by the method described above, wherein the first neural network includes at least one anomaly detection model, and the process data is provided to each of the at least one anomaly detection model.

[0125] In step 320, the determination of whether process data can be assigned to a predefined operating state is made at least in part based on the first trained neural network.

[0126] In step 330, a first indicator is provided that indicates whether process data can be assigned to the running status.

[0127] Figure 4 An exemplary computer-implemented apparatus 400 for determining the operating status of an industrial production process is shown. The computer-implemented apparatus 400 includes a first providing unit 410, a determining unit 420, and a second providing unit 430.

[0128] The first providing unit 410 is configured to provide process data associated with the production process to at least one trained first neural network, wherein the first neural network has been trained by the method described above, wherein the first neural network includes at least one anomaly detection model, and the process data is provided to each of the at least one anomaly detection model.

[0129] The determining unit 420 is configured to determine, at least in part, whether process data can be assigned to a predefined operating state based on a first trained neural network.

[0130] The second providing unit 430 is configured to provide a first indicator that indicates whether process data can be allocated to the running status.

[0131] Figure 5 An exemplary system 500 for determining the operating status of an industrial production process is shown. The system includes a computer-implemented device 510 and a computer program product 520.

[0132] The computer-implemented device 510 is capable of providing a computer-implemented device, similar to the computer-implemented device described above.

[0133] It is able to provide computer program product 520 as described in this article.

[0134] Although the invention has been described with reference to embodiments, it is possible to modify the invention in a variety of ways.

[0135] The aforementioned neural network specifically refers to computer-implemented artificial neural networks.

[0136] List of reference numerals 100 charts 110 water quality flow 120 Methanol Mass Flow 130 First Operating State 140 Second Operating State 150 Third Operating State 200 computer implementation methods 210 steps 220 steps 230 steps 240 steps 300 computer implementation methods 310 steps 320 steps 330 steps 400 computer-implemented devices 410 First Providing Unit 420 Determining Unit 430 Second Providing Unit 500 system 510 computer-implemented device 520 computer program products.

Claims

1. A computer-implemented method (200) for training a first neural network to identify the operating state of an industrial production process, the method comprising: Definition (210) at least one of the aforementioned operating states; Determine (220) at least one time interval during which the industrial production process itself is in the operating state to be identified; Provide (230) a first allocation of process data associated with the production process for the identified operating state activated during the time interval; Based on the provided first allocation, a first neural network is trained (240), wherein the first neural network includes at least one anomaly detection model, and wherein the neural network is trained to perform the allocation of process data associated with the production process for an operating state activated at a specific time interval based on at least one of the anomaly detection models.

2. The computer-implemented method according to claim 1, wherein, The operating status is associated with the manufacturing of a specific product or with the initial operating status.

3. The computer-implemented method according to any one of claims 1 or 2, wherein, The first neural network includes an anomaly recognition model for each operating state to be identified, and each operating state to be identified is associated with an anomaly recognition model that identifies the operating state to be identified as having no anomaly.

4. The computer-implemented method according to claim 3, wherein, The training further includes training each anomaly detection model such that each anomaly detection model identifies the assigned running state as non-anomaly and identifies other running states different from the assigned running state as anomaly.

5. The computer-implemented method according to any one of claims 1 to 4, the method further comprising: Provide a second neural network, including a classifier, with a second allocation that uses process data for each running state; Based on the provided second allocation, the second neural network is trained, wherein the second neural network is trained to allocate process data to the running state.

6. A computer-implemented method (300) for determining the operating state of an industrial production process, the method comprising: (310) Process data associated with the production process is provided to at least one trained first neural network, the first neural network having been trained by any one of claims 1 to 5, wherein the first neural network includes at least one anomaly detection model, and the process data is provided to each of at least one of the anomaly detection models; The determination (320) of whether the process data can be assigned to a predefined operating state is based at least in part on the first trained neural network; Provide (330) a first indicator, which indicates whether the process data can be assigned to the running state.

7. The computer-implemented method according to claim 6, further comprising determining whether the process data can be allocated to one of the operating states: When each of at least one of the anomaly detection models identifies the process data as abnormal, it is determined that the process data cannot be assigned to any of the running states. or When at least one of the anomaly identification models identifies the process data as having no anomaly, it is determined that the process data can be assigned to one of the operating states. or When more than one anomaly detection model is provided and at least two anomaly detection models identify the process data as anomaly-free, it is determined that the process data cannot be assigned to any of the operating states.

8. The computer-implemented method according to any one of claims 6 or 7, the method further comprising: When the process data can be assigned to the running state, the process data is provided to a second neural network including a classifier, the second neural network having been trained by the method according to claim 5; The second neural network determines the allocation of the process data for the operating state; A second indicator is provided, which indicates the use of the process data for the allocation of the operating state.

9. The computer-implemented method according to any one of claims 1 to 8, wherein the determination further comprises: When explicit allocation cannot be achieved by providing the process data to the first neural network and / or the second neural network, the operating state in which the process data can be allocated with predetermined determinism is determined, at least in part based on the calculation of metrics.

10. The computer-implemented method according to any one of claims 6 to 9, wherein, The provision of the second indicator also includes: The second indicator is provided to the user through the user interface; Determine at least one subsequent operating state that may follow the determined operating state; Provide the user with at least one identified subsequent running state that could potentially be followed; User input is accepted via the user interface, and the user input indicates that the subsequent running state should be executed; Based on the accepted user input, a transition to the subsequent running state is performed.

11. The computer-implemented method according to claim 10, further comprising: Analyze the transitions between two consecutive running states executed in history; Based at least in part on the analysis, predictions are made of anticipated future transitions from the determined operating state to potential subsequent operating states.

12. A computer program product comprising instructions that, when executed by a computer, cause the computer to perform the steps of a computer-implemented method according to any one of claims 1 to 11.

13. A computer-implemented apparatus (400) for determining the operating status of an industrial production process, the apparatus comprising: A first providing unit (410) is configured to provide process data associated with the production process to at least one trained first neural network, the first neural network having been trained by any one of claims 1 to 5, wherein the first neural network includes at least one anomaly detection model, and the process data is provided to each of at least one of the anomaly detection models; A determining unit (420) is configured to determine, at least in part, whether the process data can be assigned to a predefined operating state based on the first trained neural network. The second providing unit (430) is used to provide a first indicator, which indicates whether the process data can be allocated to the running state.

14. The computer-implemented apparatus of claim 13, further comprising: An execution unit is configured to perform the steps of the computer-implemented method according to any one of claims 1 to 11; and / or Another execution unit, the other execution unit being used to execute the computer program product according to claim 12.

15. A system (500) for determining the operating status of an industrial production process, the system comprising: The computer-implemented apparatus (510) according to claim 13 or 14. and The computer program product (520) according to claim 12.

Images