Identification and / or analysis of the operating conditions of industrial processes

Abstract

1. A computer-implemented method (100) for identifying and / or analyzing an operating state (2) of an industrial process (1) being performed in an industrial plant, the method comprising the steps of: obtaining (110) at least one time series (3a) of measurements of at least one process variable (3) of the industrial process (1); determining (120) from the at least one time series (3a) at least one operating state (2) of the industrial process (1) for at least one time point in the time series (3a) using a predetermined classification logic (4); and computing at least one statistic (2a) calculated over a plurality of operating states (2), at least one sequence (2b) of the operating states (2), and at least one duration of the at least one operating state (2). a combination of at least one operating state (2) and measurements from a time series (3a) relating to the operating state (2), to determine (130) one or more of: a duration (2c); at least one quantity of interest (5) further characterizing the operating state (2) of the industrial process (1); a goodness of fit (6a) of at least one predetermined model (6) to explain the time series of measurements (3a); and a goodness of fit (3b) of at least one measurement of the time series (3a) as a training example for training at least one machine learning model (1); and / or a step of training (140) a model (6) describing the behavior of the industrial process (1) at the operating state (2) determined by the classification logic (4).

Description

[Technical Field]

[0001] The present invention relates to the monitoring and analysis of industrial processes carried out in industrial plants. [Background technology]

[0002] The state and behavior of industrial processes, such as chemical processes, are characterized by a large number of process variables. It is a frequent task to predict the future evolution of a process or to derive insights from the past behavior of a process by analyzing time series of process variables. For this purpose, machine learning models trained on historical data can be used.

[0003] It has been found that for many industrial processes, time series of measurements of process variables tend to be highly heterogeneous in nature, which makes analyzing the data, and in particular training machine learning models on the data, difficult. Summary of the Invention

[0004] [Objective of the Invention] It is therefore an object of the present invention to facilitate the analysis of time series of measurements of process variables and the training of machine learning models on such data.

[0005] This object is achieved by a computer-implemented analysis method according to the first independent claim and by a further computer-implemented training method according to the second independent claim. Further advantageous embodiments are detailed in the respective dependent claims.

[0006] [DISCLOSURE OF THE INVENTION] The present invention provides a computer-implemented method for identifying and / or analyzing the operating state of an industrial process running in an industrial plant, the state and / or behavior of which is characterized by a plurality of process variables, examples of which include temperature, pressure, voltage, amperage, fill level, substance concentration, and mass flow rate.

[0007] At least one time series of measurements of at least one process variable of the industrial process is obtained. For example, these measurements can be retrieved from an online monitoring of the industrial process. However, the measurements can also be obtained, for example, from a plant historian.

[0008] From the at least one time series, at least one operating state of the industrial process is determined using a predetermined classification logic, i.e., at least a point in time and / or a stretch of time in the time series is annotated with at least one operating state of the industrial process. Examples of operating states include: The process is stopped. The process is running at full capacity. The process is running at partial capacity. Industrial plants are operating in cleaning and / or maintenance mode. A process is currently using energy source X from several available energy sources.

[0009] "At least one operating state" means that, for example, if an industrial process changes from one operating state to another at a particular point in time, this can be interpreted as the end of the previous state and the beginning of a new state, and the coarser-grained state and the finer-grained sub-states of this coarser-grained state can all be determined in one go.

[0010] For example, the distinction between whether a process is stopped or running at all can already be made according to simple indications in process variables such as the rotational speed of a pump, compressor or other rotating equipment, or power consumption.

[0011] The operating conditions thus determined and / or combinations of conditions and measurements may then be utilized to determine one or more quantities of interest desired in connection with the industrial process and / or the description of the behavior of the industrial process by one or more models.

[0012] For example, the duration of at least one operating state may be evaluated. In another example, at least one sequence of operating states may be evaluated. Also, at least one statistic calculated across multiple operating states may be evaluated. For example, the statistic may be related to a particular characteristic of each operating state. Examples of the statistic include the mean, median, and / or standard deviation of the characteristic. In one example, the characteristic of each operating state may be the duration of the operating state.

[0013] The sought quantity determined in this way can, in one example, be at least one quantity of interest that further characterizes the operating state of the industrial process. For example, such a quantity can be a prediction of at least one process variable and / or a prediction of the operating state into which the process will evolve. In this respect, the operating state of the industrial process has proven to be very important information in addition to the time series data itself. For example, one and the same time series of measurements can mean different things depending on the operating state of the process.

[0014] In a simple example, in a thin film coating system where a workpiece is brought into a vacuum chamber and material is deposited thereon by sputtering, when the system is idle and material is not being sputtered, the pressure in the vacuum chamber should be the base pressure of the vacuum (e.g., 10 -6 In contrast, when there is a sputtering gas (such as argon) in the chamber and material is being actively deposited, the pressure should be much higher (e.g., 10 -3 mbar). Therefore, 10 -6A low pressure of 10 mbar means that everything is OK, while the same low pressure in the active sputtering state means that the sputtering gas supply is not working. -3 A high pressure of mbar means that everything is OK, whereas the same high pressure in idle state means that there is likely a leak in the vacuum chamber.

[0015] In another example, the rate of a chemical reaction in an industrial process may depend on the amount of educts in the batch and / or their initial temperature. For example, if the batch is only 1 / 4 of its maximum volume, the educts will mix more quickly and it will take less time to heat them to the required reaction temperature. Therefore, what evolution in a time series of measurements is considered "normal" depends on the size of the batch.

[0016] In another example, a power plant may have very well-defined operating states, such as a startup phase, a steady state, an increase or decrease in output power, and a shutdown phase. The behavior of the power plant is governed and described by very well-defined laws and models in each state. For example, in the steady state, where many derivatives of quantities with respect to time are zero, the description of the behavior is fairly simple. The increase or decrease in output power can be described using a model with differential equations for quantities affected by dynamics. Startup and shutdown are more complex, and not all the data required for a closed description of the dynamics using differential equations may be available. Here, a machine learning model can use the available data to model the behavior in a way that is sufficiently accurate for the application at hand.

[0017] It has also been found that the fitness of at least one given model to explain a time series of measurements also depends on the operating conditions. Two very frequently used types of models are: Models based on a-priori assumptions about industrial processes, such as the physical or chemical laws that govern them; and · Trained machine learning models that do not require such assumptions.

[0018] A model based on any assumption is only valid as long as the assumption is fully satisfied, and whether the assumption is satisfied may depend on the operating conditions of the process. For example, when an extract from a chemical process is charged into a reaction vessel and heated toward a temperature at which the extract will begin to react, the temperature development within the vessel is primarily governed by the transfer of heat from the vessel's heater to the extract mixture. The extract mixture acts as a thermal mass whose inertia slows the temperature rise. However, once the activation temperature of the reaction begins, this model is no longer valid. Rather, the reacting extract mixture may act as a heat sink (if the reaction is endothermic) or as an additional heat source (if the reaction is exothermic). The model needs to be expanded to account for the new heat sink or heat source.

[0019] In another example, at least one extract may be introduced into the process using a multiple parallel pump or compressor arrangement. In this case, the process dynamics may depend on how many and which pumps or compressors are used. The changed dynamics may require an adaptation of the model used, or a change to an entirely new model based on a revised set of assumptions.

[0020] A trained machine learning model will generalize to unseen data in training only to the extent that the data still belongs to the same domain and / or distribution as the data the model was trained on. If a model is trained only on data related to a particular operating state of an industrial process, data related to a different operating state may fall outside the distribution.

[0021] Also, depending on how heterogeneous the operating conditions are, it has been found that it may not be practical to capture the behavior of an industrial process under all of these operating conditions in one single machine learning model. Rather, it may be appropriate to use and train separate models for different operating conditions. This means that a model trained using all data may not perform as well as a model trained on and applied to data from one specific operating condition.

[0022] Similarly, the suitability of at least one measurement of a time series as a training example for training at least one machine learning model may be a quantity of interest derived using operating conditions and / or sequences or statistics of operating conditions.

[0023] For example, if a machine learning model is intended to predict the behavior of an industrial process under normal operating conditions and some training examples were obtained by chance when the industrial plant was in a special cleaning state, using these data for training may "poison" the model, such that the model no longer provides sufficiently accurate predictions for the normal operating state. Thus, training samples related to the cleaning state are not suitable as training samples for a model that describes the normal operating state of the industrial process.

[0024] As an alternative or in combination with determining a specific desired quantity, a model describing the behavior of the industrial process at the operating state determined by the classification logic can be trained based on the same inputs that can be used to determine the desired quantity. That is, the model is specific to this particular operating state, and the operating state determined by the classification logic and / or the segmentation of the time series into a sequence of operating states triggers the creation of this model. The model can then be used, for example, to predict future behavior of the industrial process based on historical data. It can also be used, for example, to determine whether current behavior of the industrial process is normal or abnormal.

[0025] In a particularly advantageous embodiment, the determined operating state is part of a multi-level hierarchy of operating states, in which at least one operating state has substates. For example, the multi-level hierarchy of possible operating states can be preset according to the application at hand. For the same process, different subdivisions into the multi-level hierarchy may be appropriate for different purposes. That is, the operating state may first be obtained at a first, rather abstract level, and then more concrete or specific substates of this abstract first state may be determined. For example, different levels of the hierarchy may differ in how much analysis is required to determine a specific state. More abstract states, which may already be sufficient for many applications, may be determined fairly quickly, while determining one specific substate may require much more analysis, which takes more time.

[0026] Also, the level of a priori knowledge required to determine a particular state may depend on the level of the hierarchy. Only little a priori knowledge may be required to determine more abstract states, while more detailed a priori knowledge may be required to determine fine-grained substates. This means that the amount and level of detail of available a priori knowledge, or the availability of a priori knowledge in the first place, may determine how fine-grained an operational state can be determined in a multi-level hierarchy.

[0027] In a further particularly advantageous embodiment, determining the at least one operating state comprises: determining a first operational state at a first level of the multi-level hierarchy (the "first level operational state"); limiting the determination of a second operating state at a next level of the multi-level hierarchy ("second-level operating state") to operating states available as sub-states of the first operating state according to the multi-level hierarchy; It may comprise:

[0028] In this way, inconsistencies between operational states determined at different levels of a multi-level hierarchy can be avoided. For example, a substate of a first level state of stopped may never be used if the first level state is determined to be in fact an operational state. Also, determining operational states incrementally from one level to the next may save time because the number of possibilities is more limited, and the accuracy of the ultimately obtained operational state may be improved.

[0029] In a further particularly advantageous embodiment, in the process of determining at least one operating state, the time series is divided into segments during which the industrial process remains in a respective operating state. For each such segment, a respective operating state is determined. Determining whether the industrial process remains in one single operating state—whatever this operating state—for a duration of time is quicker and more reliable than directly determining the operating state for multiple points in time during that duration and then comparing these operating states. Furthermore, the longer the duration the industrial process remains in the same operating state, the more effort is saved by only having to determine a specific operating state once before changing to another operating state.

[0030] One exemplary method for segmenting at least one time series is to cluster the measurements. For example, time series k-means clustering or Toeplitz Inverse Covariance-Based Clustering (TICC) may be used. For example, measurements that belong to the same cluster may be considered to be associated with the same operating state. Another method is to detect change points that indicate transitions between operating states and / or to detect steady states. These and other sources of knowledge may all be pooled.

[0031] In a further particularly advantageous embodiment, determining at least one operating state is based at least in part on a probability of transitioning the industrial process from one predetermined operating state to one of several possible next operating states. In this way, further a priori knowledge about which states are likely to transition to which other states can be utilized. For example, if an industrial process comprises multiple stages corresponding to multiple operating states, it is much more probable that these operating states will continue in the intended order and / or dynamics than that one such operating state will suddenly transition to a state in which the industrial process is stopped and shut down, discarding the work and materials already expended.

[0032] In a further particularly advantageous embodiment, at least one quantity of interest further characterizing the operating state of the industrial process indicates whether the industrial process is in an abnormal state. As mentioned above, whether one and the same time series of measurements indicates an abnormality may depend on the operating state. Also, the sequence of operating states, the duration the operating states last, and statistics calculated over multiple operating states may themselves serve as aggregation products of the time series that indicate whether an abnormality exists. For example, if filling or emptying a vessel takes much longer than expected, this may indicate that a pipe is clogged or a pump is not functioning.

[0033] In a further particularly advantageous embodiment, the at least one quantity of interest further characterizing the operating state of the process comprises a pattern and / or signature whose presence in a process variable of the industrial process indicates that the industrial process is in this particular operating state. That is, after it is determined that a particular operating state S exists, a rule of the type "if a constellation of measurements x, y, and z exists, this indicates that operating state S exists" can be generated. Such a rule can, in turn, be used to classify operating states in the future. However, such a rule can also be used to explain and / or analyze the behavior of an industrial process and / or to identify the root cause of a problem. For example, if it is found that the operating state of a process changes when the temperature at a particular location in a plant exceeds a constant value for more than 20 minutes, and there is no physical reason for such a dependency, something that should not be temperature-dependent may be temperature-dependent. For example, there may be a temperature-dependent fault in a capacitor, a semiconductor, or a solder joint on a circuit board.

[0034] Using the rules so determined, insights into the operating state of the industrial process can be gleaned in a way that is both human understandable and machine processable for data analysis, machine learning tools, or even direct control of the industrial process. It is a well-known problem that such insights into the process exist only in the minds of process operators and are not documented in a machine processable format, or even at all.

[0035] In a further particularly advantageous embodiment, the industrial plant is a power plant, a hydrocarbon refinery, a chemical production plant, or a waste incineration plant. Although these types of plants operate very differently, they share the commonality that the transition from one operating state to the next is not entirely under the control of the plant operator, but rather can be triggered by influences that are not entirely controllable by the operator. For example, a gas-fired power plant may be configured to run on biological digester gas. The chemical composition of this gas is not always constant, and a change in this chemical composition can cause combustion to enter another distinct regime corresponding to the plant's new operating state. The same applies to waste incineration plants, where the calorific value of the waste is constantly changing and cannot be directly measured. Similarly, when hydrocarbons are processed in a refinery, even if the nominally same type of crude oil is always used, the exact composition can change without the possibility of measuring it in real time. In a chemical production plant, the processed extract may always remain nominally the same, but the initial temperature of the extract may vary. For example, extracts may be delivered in conventional tank trucks, so that their temperature is close to ambient temperature. As previously mentioned, initiation of the intended reaction may depend on the local temperature within the extract mixture. A temperature difference of 40-50°C between winter and summer will affect the time it takes to initiate the intended reaction and transition from a "reaction ready" operating state to a "reaction running" operating state.

[0036] The present invention also provides a computer-implemented method for training classification logic for use in the above-described methods.

[0037] During the method, at least one historical time series of at least one process variable of the industrial process is provided, and ground truth regarding the existence of particular operating states at particular times in the historical time series and / or the probability of transitions between operating states is provided, which ground truth may be provided at any level of granularity.

[0038] Classification logic to be trained is provided, which comprises at least one model whose behavior is characterized by a set of model parameters. For example, the classification logic may comprise a function that can be fitted to data by varying the parameters and / or a trainable machine learning model.

[0039] Using this classification logic, operational states and / or transitions between operational states at specific times are determined from the historical time series. Discrepancies between the operational states and / or transitions thus determined, on the one hand, and the corresponding ground truth, on the other hand, are then rated by a predetermined loss function. For example, if the ground truth specifies a certain operational state and / or transition at a specific time and the classification logic predicts the operational state and / or transition for the same time, this discrepancy can be directly measured as a deviation (difference) from the ground truth. However, even if the prediction, on the one hand, and the ground truth, on the other hand, do not relate to the same time (or a continuum), the concept of discrepancy can still be measured. For example, if it is not physically possible for the predicted operational state to become the ground truth operational state within a time interval that separates the respective times, there is an apparent discrepancy between the prediction and the ground truth.

[0040] The model parameters are optimized with the goal that further processing of the historical time series will result in an improved rating by the loss function. The method may use any kind of available a priori knowledge as ground truth. For example, this a priori knowledge may include insights provided by plant operators or alarms or events collected during the execution of the industrial process.

[0041] In a particularly advantageous embodiment, at least one model comprises a hidden Markov model for modeling sequences of operating states and / or transitions between operating states. This model starts from the assumption that transitions between states are governed by a set of rules that are valid without being explicitly known. For example, the trained model may specify that certain operating states shall occur in a particular sequence and that there is no "backtracking" through this sequence in the opposite direction.

[0042] In a further particularly advantageous embodiment, at least one model comprises a decision tree that identifies time series behaviors that indicate particular operating states. In this way, signatures of certain operating states in the time series of measurements can be discovered. These signatures can be used to detect the presence of the respective operating states in the future.

[0043] Since they are computer-implemented, the methods may be embodied in the form of software. Accordingly, the present invention also relates to one or more computer programs having machine-readable instructions that, when executed by one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the above-described methods. Examples of computing instances include virtual machines, containers, or serverless execution environments in the cloud. The present invention also relates to machine-readable data carriers and / or downloadable products having one or more computer programs. Downloadable products are, for example, digital products having computer programs that can be sold in online shops for immediate fulfillment and downloading to one or more computers. The present invention also relates to one or more computing instances having one or more computer programs and / or machine-readable data carriers and / or downloadable products.

[0044] In the following, the invention will be illustrated using figures, without any intention of limiting the scope of the invention. [Brief explanation of the drawings]

[0045] [Figure 1] FIG. 1 is an exemplary embodiment of a method 100 for identifying and / or analyzing an operating state 2 of an industrial process 1 being performed in an industrial plant. [Figure 2] FIG. 2 is an exemplary analysis of a time series 3 a of measurements of a process variable 3 . [Figure 3] FIG. 3 is an exemplary statistical analysis of a time series 3a for operating states 2 and transitions between them. [Figure 4] FIG. 4 is an exemplary embodiment of a method 200 for training classification logic 4. DETAILED DESCRIPTION OF THE INVENTION

[0046] FIG. 1 is a schematic flow chart of one embodiment of a method 100 for identifying and / or analyzing an operating state 2 of an industrial process 1 being performed in an industrial plant.

[0047] In step 110, at least one time series 3a of measurements of at least one process variable 3 of the industrial process 1 is obtained.

[0048] In step 120, at least one operating state 2 of the industrial process 1 is determined from the at least one time series 3a using a predetermined classification logic 4 for at least one time point in the time series 3a.

[0049] According to block 121, the determined operating state 2b may be part of a predetermined multi-level hierarchy of operating states 2, in which at least one operating state 2 has sub-states 2', 2'', 2'''.

[0050] In particular, according to block 121a, a first operating state 2 may be determined at a first level of the multilevel hierarchy. According to block 121b, determining a second operating state 2', 2'', 2''' at a next level of the multilevel hierarchy may in this case be limited to operating states available as substates of the first operating state 2 according to the multilevel hierarchy.

[0051] According to block 122, the time series 3a may be divided into segments 3c-3i during which the industrial process 1 remains in a respective operating state 2. According to block 123, a respective operating state 2 may then be determined for each such segment 3c-3i.

[0052] In particular, according to block 122a, the dividing into segments 3c to 3i may comprise one or more of clustering, change point detection, and steady state detection.

[0053] According to block 124, determining at least one operating state 2 may be based at least in part on a probability for the industrial process 1 to transition from one predetermined operating state 2 to one of several possible next operating states 2.

[0054] In step 130, at least one statistic 2a calculated over a plurality of operating states 2; at least one sequence 2b of operating states 2; a duration 2c during which at least one operating state 2 continues; a combination of at least one operating state 2 with measurements from a time series 3a relating to this operating state 2; From one or more of at least one quantity of interest 5 further characterizing the operating state 2 of the industrial process 1; the goodness of fit 6a of at least one predetermined model 6 for explaining the time series 3a of measurements; a goodness of fit 3b of at least one measurement of a time series 3a as a training example for training at least one machine learning model 1; One or more of the following is determined:

[0055] As an alternative or in combination with determining a specific desired quantity, in step 140, a model describing the behavior of the industrial process at the operating state determined by the classification logic can be trained based on the same inputs that can be used to determine the desired quantity. That is, the model is specific to this particular operating state, and the operating state determined by the classification logic and / or the segmentation of the time series into a series of operating states triggers the creation of this model. The model can then be used, for example, to predict future behavior of the industrial process based on historical data. It can also be used, for example, to determine whether current behavior of the industrial process is normal or abnormal.

[0056] FIG. 2 illustrates how an exemplary time series 3 a of measurements of a process variable 3 can be analyzed for operating conditions 2 .

[0057] In the example shown in FIG. 2, there is a first operating state 2 in which the industrial process 1 is stopped, and a second operating state 2# in which the industrial process 1 is running. The second operating state 2# has three sub-states 2′, 2″, and 2′″. These sub-states 2′, 2″, and 2′″ are correlated with the value of the process variable 3. In the example shown in FIG. 2, the running industrial process 1 is determined to be: When the value of process variable 3 is greater than 500, the first substate 2' is reached. When the value of process variable 3 is between 300 and 500, it is in second substate 2''; and · When the value of process variable 3 falls below 300, it is in the third substate 2'''.

[0058] The time series 3a is divided into segments 3c to 3i each relating to one single operating state 2, and an operating state 2 is then determined for each such segment 3c to 3i.

[0059] FIG. 2 also includes an inset comparing the probability p(2) that the industrial process 1 is in each of states 2, 2', 2'', and 2''', at any given point in time in the time series 3a, for different operating states 2, 2', 2'', and 2''', respectively. It is noteworthy that the probability for state 2''' is particularly low, which is a clue that this state 2''' may be an abnormal state.

[0060] 2 also illustrates how probable transitions between operating states are. A transition from operating state 2'' to one of operating states 2 and 2''' is probable (p+). A transition from one of operating states 2 and 2''' to operating state 2', and a transition from operating state 2' to operating state 2'' are very probable (p++).

[0061] Figure 3 illustrates a statistical analysis of a time series 3a of measurements of a process variable 3. In the example shown in Figure 3, there are four operating states A, B, C, and D. The time series 3a comprises samples taken at discrete times.

[0062] 3a shows, for operating states A, B, C, and D, the number N(2) of times that each operating state A, B, C, or D was determined to exist. In this example, operating state A is by far the most frequent, followed by operating state B. Operating states C and D are much less frequent, providing clues that they may be abnormal.

[0063] Figure 3b shows the number of such transitions N(→) encountered in time series 3a for a subset of possible transitions between operating states A, B, C, and D. It is noteworthy that the transitions from operating state C to operating state B and from operating state A to operating state C are the least frequent. In contrast, the transition from operating state C to operating state A is the most frequent.

[0064] FIG. 4 is a schematic flow chart of one embodiment of a method 200 for training classification logic 4 for use in the method 100 described above.

[0065] In step 210, at least one historical time series 3a of at least one process variable 3 of the industrial process 1 is provided.

[0066] In step 220, a ground truth 2* is provided regarding the existence of particular operating states 2 at particular times in the historical time series 3a and / or regarding the probability of transitions between operating states 2.

[0067] In step 230, a classification logic 4 is provided. The classification logic 4 comprises at least one model whose behavior is characterized by a set of model parameters 4a.

[0068] According to block 231, the at least one model 4 may comprise a hidden Markov model for modeling the sequence of operating states 2 and / or the transitions between operating states 2.

[0069] According to block 232 , the at least one model 4 may comprise a decision tree that identifies behavior of the time series 3 a that indicates a particular operating state 2 .

[0070] In step 240, the classification logic 4 determines from the historical time series 3a the operating states 2 and / or the transitions between operating states 2 at particular times.

[0071] In step 250, a predefined loss function 7 rates the discrepancy between the determined operating states 2 and / or transitions on the one hand and the corresponding ground truth 2* on the other hand.

[0072] In step 260, the model parameters 4a are optimized with the goal that further processing of the historical time series 3a will result in an improved rating 7a according to the loss function 7. The final optimized state of the model parameters 4a is labeled with the reference character 4a*. These parameters 4a* define the fully optimized state 4* of the model 4.

[0073] [List of references] 1. Industrial Processes 2, 2# Industrial process 1 operating status 2', 2'', 2''' Substates in a multilevel hierarchy 2* Ground truth for operating state 2 2a Statistics calculated across multiple operating conditions 2 2b Operation state 2 sequence 2c Duration of operating state 2 3 Process Variables 3a Time series of values ​​of process variable 3 3b Goodness of fit of measurements as training examples 3c~3i Segment of Time Series 3a 4. Classification logic 4* Classification logic 4 trained state 4a Parameters characterizing the behavior of classification logic 4 4a* Fully optimized state of parameter 4 5. Quantity to focus on 6. Model to explain time series 3a 6a Goodness of fit of Model 6 7 Loss Function 7a Rating by loss function 7 100 Method for identifying and / or analyzing operating conditions 2 110 Get time series 3a 120 Determine operating state 2 from time series 3a 121 Using Operational State 2 in a Multilevel Hierarchy 121a Determines the operational state 2 at the first level of the hierarchy 121b Restricting the second operating state to available sub-states 122 Divide time series 3a into segments 3c-3i 122a Special method for determining segments 3c-3i 123 Determine the operating state 2 for each segment 3c to 3i 124 Consider the probability of transition between operational states 2 130 Determine quantities 5, 6a, and 3b from inputs 2, 2a, 2b, 2c, and 3a 200 Methods for training classification logic 4 210 Provide historical time series 3a 220 Provide ground truth2* 230 Provide classification logic 4 231 Using Hidden Markov Models 232 Using Decision Trees 240 Use classification logic 4 to determine operating state 2 250 Loss function 7 rates the discrepancy with the ground truth 2* 260 Optimize model parameter 4a A, B, C, D Specific operating conditions2

Claims

1. A computer-implemented method (100) for identifying and / or analyzing an operating condition (2) of an industrial process (1) being carried out in an industrial plant, comprising: - acquiring (110) at least one time series (3a) of measurements of at least one process variable (3) of said industrial process (1); - determining (120) from said at least one time series (3a) at least one operating state (2) of said industrial process (1) for at least one time point in said time series (3a) using a predetermined classification logic (4); and ·the below described: At least one statistic (2a) calculated over a plurality of operating states (2); At least one sequence (2b) of operating states (2), the duration (2c) of at least one operating state (2), and a combination of at least one operating state (2) and measurements from said time series (3a) relating to this operating state (2); From one or more of at least one quantity of interest (5) further characterizing the operating state (2) of the industrial process (1); the goodness of fit (6a) of at least one predetermined model (6) to explain said time series (3a) of measurements; the goodness of fit (3b) of at least one measurement of said time series (3a) as a training example for training at least one machine learning model (1); determining 130 one or more of: and / or training (140) a model (6) describing the behavior of the industrial process (1) in the operating conditions (2) determined by the classification logic (4); A method (100) comprising:

2. 2. The method (100) of claim 1, wherein the determined operational state (2) is part of a multi-level hierarchy (121) of operational states (2), in which at least one operational state (2) has sub-states (2', 2'', 2''').

3. The determining (120) of the at least one operating state may further comprise: - determining (121a) a first operating state (2) at a first level of said multi-level hierarchy; - restricting (121b) said determination of second operating states (2', 2'', 2''') at the next level of said multi-level hierarchy to operating states available as sub-states of said first operating state (2) according to said multi-level hierarchy; The method (100) of claim 2, comprising:

4. The determining (120) of the at least one operating state (2) comprises: Dividing (122) the time series (3a) into segments (3c-3i) during which the industrial process (1) remains in each operating state (2); - determining (123) for each such segment (3c-3i) said respective operating state (2); The method (100) of any one of claims 1 to 4, comprising:

5. The method (100) of claim 4, wherein the dividing (122) into segments (3c-3i) comprises one or more of clustering, change point detection, and steady state detection (122a).

6. 6. The method (100) of any one of claims 1 to 5, wherein determining (120) the at least one operating state (2) is based at least in part on a probability (124) for the industrial process (1) to transition from one predetermined operating state (2) to one of several possible next operating states (2).

7. 7. The method (100) according to any one of claims 1 to 6, wherein at least one quantity of interest (5) further characterizing the operating state (2) of the industrial process (1) indicates whether the industrial process (1) is in an abnormal state.

8. 8. The method (100) according to any one of claims 1 to 7, wherein at least one quantity of interest (5) further characterizing the operating state (2) of the industrial process (1) comprises a pattern and / or signature whose presence in the process variables (3) of the industrial process (1) indicates that the industrial process (1) is in this particular operating state (2).

9. The method (100) according to any one of claims 1 to 8, wherein the industrial plant (1) is a power plant, a hydrocarbon refinery, a chemical production plant or a waste incineration plant.

10. A computer-implemented method (200) for training a classification logic (4) for use in a method (100) according to any one of claims 1 to 9, comprising: - providing (210) at least one historical time series (3a) of at least one process variable (3) of said industrial process (1); - providing (220) a ground truth (2*) regarding the existence of specific operating states (2) at specific times in the historical time series (3a) and / or regarding the probability of transitions between operating states (2); - providing (230) a classification logic (4) having at least one model whose behavior is characterized by a set of model parameters (4a); - determining (240) from the historical time series (3a) by the classification logic (4) operational states (2) and / or transitions between operational states (2) at specific times; a step (250) of rating the discrepancies between the determined operating states (2) and / or transitions on the one hand and the corresponding ground truth (2*) on the other hand by means of a predetermined loss function (7); optimizing (260) the model parameters (4a) with the goal that further processing of historical time series (3a) will result in an improvement in the rating (7a) according to the loss function (7); A method (200) comprising:

11. 11. The method (200) of claim 10, wherein the at least one model (4) comprises a Hidden Markov Model (231) for modeling a sequence of operating states (2) and / or transitions between operating states (2).

12. 12. The method (200) according to claim 10 or 11, wherein at least one model (4) comprises a decision tree (232) that identifies a behavior of the time series (3a) that is indicative of a particular operating state (2).

13. One or more computer programs comprising machine-readable instructions that, when executed on one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the method (100, 200) of any one of claims 1 to 12.

14. A non-transitory machine-readable data carrier and / or download product comprising one or more computer programs according to claim 13.

15. One or more computers and / or computing instances comprising one or more computer programs according to claim 13 and / or comprising a machine-readable data carrier and / or downloadable product according to claim 14.

Images