Training a predictive model to predict undesired events during process execution

Abstract

A method (100) for training a predictive model (1) for predicting the likelihood of at least one predefined undesirable event occurring during execution of a process (2) using training samples (3), each training sample (3) comprising data characterizing a state of the process (2), the method (100) comprising obtaining (110) training samples (3) representing states of the process (2) that do not cause the undesirable event, labeling these training samples with a predefined low likelihood that the undesirable event occurs, and labeling states of the process (2) that have an increased likelihood of causing the undesirable event based at least in part on the process model (2a) and a set of predefined rules (2b) that define in which states of the process (2) an increased likelihood of the undesirable event occurring is determined. and labeling these training samples (4) with the increased likelihood; providing (130) the training samples (3, 4) to a predictive model (1) to be trained in order to obtain from the predictive model (1) a prediction (5) of the likelihood of an undesirable event occurring at a state of the process (2) represented by each sample (3, 4); rating (140) a difference between the prediction (5) and the label of each sample (3, 4) by a predefined loss function (6); and optimizing (150) parameters (1 a) characterizing the behavior of the predictive model (1) such that when a prediction (5) is made for the further samples (3, 4), the rating (6 a) by the loss function (6) is more likely to improve.

Description

[Technical Field]

[0001] The present invention relates to improving the execution of processes on industrial plants and / or electrical networks so as to reduce the likelihood of at least one undesirable event occurring. [Background technology]

[0002] Industrial plants and electrical networks for carrying out industrial processes are protected by interlock conditions. Such interlock conditions are very often used on equipment to protect against operator errors, such as grounding an electrical circuit while simultaneously being energized by a power source (thereby shorting the power source to ground). For the execution of processes in a plant or network, there are more safety interlock rules that impose certain limits on the plant's state or other variables. For example, if a pressure gauge in a vacuum chamber housing material processing equipment registers an increase in pressure above that which is tolerable for the system, the equipment may be turned off to prevent damage to the system.

[0003] Safety interlock rules often prescribe drastic responses to unsafe conditions. For example, a plant or part of it may be brutally disconnected from its power supply, regardless of how tedious or difficult it may be to get the process up and running again. For example, turning off the heat in a chemical reactor may cause its liquid contents to cool and solidify. In this case, using the reactor again would require laborious manual removal with a hammer and chisel at best and be impossible at worst. Triggering a safety interlock event may therefore be a major hindrance. Summary of the Invention

[0004] [Objective of the present invention] It is therefore an object of the present invention to reduce the occurrence of safety interlock events by enabling the prediction of their occurrence and / or by actively avoiding such events based on such prediction.

[0005] [Disclosure of the Invention] The present invention provides a method for training a predictive model to predict the likelihood of at least one predetermined undesirable event occurring during the execution of a process. The process may be an industrial process carried out in an industrial plant, such as a chemical process that converts one or more extractives into one or more products by at least one chemical reaction. The process may also be an electrical process in an electrical network, such as the operation of a power grid. The method uses training samples having data characterizing the state of the industrial process.

[0006] During the course of the method, training samples are obtained that represent process states that do not cause undesired events. Operation without undesired events is the normal operating state of a plant or network, and therefore there is an abundant supply of such training samples. The training samples are labeled with a preset low probability that the undesired event will occur, which depends on the scale by which this probability is measured. For example, if the probability is measured as a probability, it can be set to a very low non-zero value rather than zero to avoid divide-by-zero errors at run time.

[0007] The fortunate rarity or absence of undesirable events during normal plant or network operation is coupled with the adverse condition that few, if any, training samples are available for process conditions that trigger such undesirable events. That is, the set of training samples is strongly unbalanced with respect to training samples that represent normal operation. To train a predictive model to reliably predict the likelihood of undesirable events, a more balanced set of training samples is needed.

[0008] To this end, further training samples are obtained based at least in part on the process model and a set of predetermined rules that define at which states of the process there is an increased likelihood that an undesired event will occur, and these further training samples represent states of the process that have an increased likelihood of causing the undesired event and are therefore labeled with this increased likelihood.

[0009] A process model may be a monolithic model of the process as a whole, but may also be composed of sub-models at any desired level of granularity. For example, different sub-models may predict the future evolution of different aspects of the state of the process, which may correspond to different sub-units of a plant or network. For clarity, only one process model will be referred to below.

[0010] In particular, the process model may be particularly configured to predict the future evolution of the state of the process based on at least one current and / or past state of the process. The future state thus predicted may then be checked, for example, against predetermined rules for the occurrence of an undesired event. If the predicted evolution of the process leads, according to the predetermined rules, to a state that causes an undesired event, the current and / or past states from which this predicted evolution was determined may be considered to be the cause of the undesired event.

[0011] Preferably, the process model comprises: Machine learning models, and / or Simulation models, and / or A surrogate approximation of this simulation model may be provided.

[0012] Which type of process model is most advantageous depends on how much knowledge is available about the internal structure of the process (or any part of it). For example, if the process is a "black box," a machine learning model can be trained on inputs to this "black box" and outputs obtained in response, without having to delve deeply into the internal structure of the process. If the internal structure of a process is well known and understood, a simulation model of the process based on knowledge about the process can be used.

[0013] Simulation models can predict the future evolution of a process with high fidelity. However, this comes at the cost that the simulation models can be expensive to compute. Complex calculations can take hours or even longer. However, to avoid undesirable events during real-time execution of a process, being able to obtain results faster is more important than the highest level of fidelity.

[0014] Here, surrogate approximations of simulation models can save a great deal of computation time. At the expense of a small sacrifice in fidelity, computation can be accelerated to the point where it can be much faster than real time. This, in turn, makes it possible to compute multiple scenarios and explore multiple possible paths a process could take, depending on the actions currently being taken. This is somewhat similar to lossy compression of audio and video data, which removes some of the quality but reduces the bandwidth requirements so that the data can be streamed in real time.

[0015] Using a process model and a set of rules, a previously insufficient set of training examples for conditions that may trigger undesired events can be augmented to any degree to arrive at a more balanced set of training examples. This, in turn, enables supervised training of predictive models. In this context, utilizing a set of predetermined rules provides the additional advantage that these rules provide a level of abstraction from the exact internal behavior of the actual process, which is not always known. For example, in an industrial plant, some equipment may be purchased and used as a "black box" with abstract technical specifications without access to its internal structure. Large power grids are composed of many subnetworks operated by different operators, and the internal complexity of these subnetworks is hidden behind abstract specifications. Just as the Earth can be abstracted into a point mass for some astronomical calculations, a power plant may ultimately be reduced to small quantities such as maximum power output and power output slew rate.

[0016] During supervised training, training samples are provided to a predictive model to be trained. The predictive model then outputs a prediction of the likelihood of an undesired event occurring at the process state represented by each sample. The difference between the prediction thus obtained, on the one hand, and the label of each sample, on the other hand, is rated by a predetermined loss function (or "cost function"). Examples of such loss functions include cross-entropy and log-likelihood.

[0017] The parameters characterizing the behavior of the predictive model are optimized so that the loss function is likely to improve its rating when predictions are made on further samples. Training may be stopped in response to any suitable stopping criterion, such as achieving a certain prediction accuracy on the training samples, a number of training epochs, or convergence (i.e., no further changes to the parameters). For example, in a neural network predictive model, the parameters comprise the weights by which the inputs to each neuron are summed to that neuron's activation.

[0018] The predetermined rules may comprise, for example, safety interlock rules that specify under what circumstances a safety interlock event should be triggered. Many such safety interlock rules trigger a safety interlock event in response to a particular alarm being issued within the plant and / or in response to a particular state variable of the plant and / or process exceeding a particular pre-set threshold. In electrical networks, sub-networks are typically configured to disconnect from each other if the current between them rises above a hard limit. If the respective condition is met, this necessarily leads to the occurrence of an event; i.e., the probability of the event occurring is then maximum (i.e., 1 on a probability scale).

[0019] Thus, in particularly advantageous embodiments, the undesired event comprises a safety interlock event that forces at least a partial shutdown and / or shutdown of the process and / or the industrial plant or electrical network in which the process is running. The ability to predict these events, which were previously dealt with reactively, allows them to be dealt with proactively before they actually occur. For example, the occurrence of these events can be avoided and / or the consequences can be mitigated.

[0020] For other types of undesirable events, the rules that link process states to an increased likelihood of the event occurring may be more lenient. For example, each rule may have a certain amount of "penalty points," and these "penalty points" occur if the process state satisfies multiple rules at once. In this case, the undesirable event may be coupled with a condition, for example, that at least a threshold amount of "penalty points" has occurred. In such a situation, no rule by itself stipulates that the event will occur under certain conditions, but each rule that is satisfied increases the likelihood of the event.

[0021] The likelihood of an undesired event occurring may be measured in any suitable manner. For example, the likelihood may be measured in terms of the probability that this event will occur, which is an easily interpretable measure. For example, a safety requirement may specify that the probability of a particular event must be below a certain threshold.

[0022] To avoid the occurrence of undesired events in real-time process execution, another concept of likelihood may also be advantageous. Likelihood may be measured in terms of the closeness of a process state to a state that causes the undesired event. This provides direct guidance on what can be done to reduce the likelihood of the undesired event. This is somewhat similar to how the risk of drone operations is measured in terms of the proximity of operations to critical airspace or ground facilities.

[0023] In a further advantageous embodiment, the method further comprises determining, based at least in part on predetermined rules, which variables characterizing the state of the process have an influence on the likelihood of an undesired event occurring. These variables and / or processing results obtained from these variables can then be included in a training sample, i.e., the state variables can be pre-filtered to reduce the complexity of the predictive model.

[0024] The training samples may comprise, for example, any kind of state variable of a process. For example, in a chemical process, the state variable may comprise temperature, concentration, or other physical property of a substance, pressure, or mass flow rate. In an electrical process, the state variable may comprise voltage, current, power, temperature, or the switching state of a switchable connection.

[0025] In a further advantageous embodiment, at least one statistical moment and / or time series of at least one state variable of the process is included in the training samples. In particular, some types of predictive models, such as recurrent neural networks and transformer networks, are particularly adapted to process samples comprising time series and to learn directly from trends evident in such time series. Statistical moments of the state variables can be included in the training samples in order to compress the information of the state variables.

[0026] In a further particularly advantageous embodiment, the predictive model obtains a prediction of the likelihood of an undesired event occurring at the end of a given time window based on samples within this time window. In this way, a time horizon for causality can be established, and it is then clear how much time remains for any corrective action aimed at avoiding the undesired event.

[0027] In a further particularly advantageous embodiment, the behavior of a trained predictive model is approximated by a surrogate model that is computationally less expensive to evaluate than the trained predictive model. In this way, some of the accuracy of the prediction is traded in for obtaining the prediction more quickly. In particular, as mentioned above, when predictions can be obtained faster than real time, multiple candidate actions to avoid the occurrence of undesirable events can be tested using the predictive model. That is, "what-if" scenarios can be analyzed, and the best action can then be implemented on the process.

[0028] The ultimate goal of the trained predictive model is to improve process performance and proactively react to and avoid impending occurrences of undesirable events. The present invention also provides methods for performing a process on at least one industrial plant or in at least one electrical network.

[0029] During the method, one or more samples representing a state of the process are provided to a trained predictive model and / or its surrogate approximation to obtain a prediction of the likelihood of an undesirable event occurring in the state of the process represented by the one or more samples.

[0030] The prediction is tested against at least one predetermined criterion, which may comprise, for example, a threshold probability of the occurrence of the undesired event or a threshold proximity in parameter space of the current state of the process to the condition that triggers the undesired event.

[0031] In response to the criteria being met, an alarm is issued to a process operator and / or process execution is modified with the goal of reducing the likelihood of the undesired event occurring. Such modifications may involve degradation of execution. For example, the production rate of product may be reduced or completely halted, or some load may be removed from the electrical network. Of course, it is most desirable to run the process without degradation, but degradation is usually much more graceful than a safety interlock event occurring. For example, temporarily removing less product from a chemical process is much less detrimental than having to clear solidified material from a reaction vessel with a hammer and chisel, and removing some load from an electrical network is much better than a complete loss of power in the network.

[0032] In a further particularly advantageous embodiment, to modify the process execution, samples representing multiple candidate states of the process that differ from the current state of the process are provided to the predictive model. This provides the probability of an undesired event occurring for the candidate states. That is, options for states toward which the process can be moved can be explored in a "what-if" manner. The process execution can then be steered toward the candidate state with the least probability of an undesired event as the target state.

[0033] Directing a process may comprise, for example, changing setpoint values ​​provided to lower level controllers of an industrial plant or electrical network that executes the process, and / or enabling or disabling equipment or interconnections within the plant or network.

[0034] In many industrial plants or electrical installations where processes are being run, many Internet of Things (IoT) devices are involved in this operation. These devices communicate with IoT edge systems for process control purposes. These edge systems have some processing power to evaluate predictive models, but not enough processing power to train such predictive models. However, the edge systems can delegate this task to cloud platforms that have more processing power.

[0035] Thus, in a further particularly advantageous embodiment, samples representative of the state of a process are collected by an edge system at an industrial plant or other site involved in the execution of the process. The samples are provided by the edge system to a cloud platform. Based on these samples, a predictive model is trained and / or updated on the cloud platform. The cloud platform also creates a surrogate approximation for the trained and / or updated predictive model and / or updates to such existing approximation. The surrogate approximation and / or updates are provided back to the edge system, where the surrogate approximation is evaluated to obtain a prediction of the likelihood of an undesirable event occurring.

[0036] The method may be computer-implemented in whole or in part. Accordingly, the present invention also relates to one or more computer programs having machine-readable instructions that, when executed on one or more computers and / or computational instances, cause the one or more computers to perform one of the above-described methods. In this context, virtualization platforms, hardware controllers, network infrastructure devices (such as switches, bridges, routers, or wireless access points), and end devices in a network (such as sensors, actuators, or other industrial field devices) capable of executing machine-readable instructions shall also be considered as computers.

[0037] Therefore, the present invention also relates to a non-transitory storage medium and / or a downloadable product having one or more computer programs. A downloadable product is a product that can be sold in an online shop for immediate fulfillment by download. The present invention also provides one or more computers and / or computing instances having one or more computer programs and / or one or more non-transitory machine-readable storage media and / or a downloadable product.

[0038] In the following, the invention will be illustrated using figures without intending to limit the scope of the invention. [Brief explanation of the drawings]

[0039] [Figure 1] 1 is an exemplary embodiment of a method 100 for training a predictive model. [Figure 2] 2 is an exemplary embodiment of a method 200 for performing a process. [Figure 3] 1 is an exemplary arrangement of an edge system and a cloud platform for keeping a surrogate approximation of a predictive model updated. DETAILED DESCRIPTION OF THE INVENTION

[0040] 1 is a schematic flowchart of an exemplary embodiment of a method 100 for training a predictive model 1 for predicting the likelihood of at least one predetermined undesirable event occurring during the execution of a process 2.

[0041] In step 110, training samples 3 are obtained. These training samples 3 represent states of the process 2 that do not cause undesired events. They are labeled with a preset low probability that the undesired event will occur.

[0042] In step 120, further training samples 4 are obtained based at least in part on the process model 2a and a set of predetermined rules 2b that define at which states of the process 2 there is an increased likelihood that an undesired event will occur. These further training samples 4 represent states of the process that have an increased likelihood of causing an undesired event, and therefore they are labeled with this increased likelihood.

[0043] In step 130, the training samples 3, 4 are provided to the predictive model 1 to be trained. The predictive model 1 then outputs a prediction 5 of the likelihood of an undesirable event occurring at the state of the process 2 represented by each sample 3, 4.

[0044] In step 140, the difference between the prediction 5 and the label of each sample 3, 4 is rated by a predetermined loss function 6. The loss function 6 results in a rating 6a.

[0045] In step 150, parameters 1a characterizing the behavior of the predictive model 1 are optimized so that the ratings 6a from the loss function 6 are likely to improve when predictions 5 are made on further samples 3, 4. The final trained state of parameters 1a is labeled with reference character 1a*. These final optimized parameters 1a* characterize the behavior of the fully trained predictive model labeled with reference character 1*.

[0046] In step 160, the behavior of the trained predictive model 1* is approximated by a surrogate model 1**, which is computationally less expensive to evaluate than the trained predictive model 1*.

[0047] According to block 111, it may be determined, based at least in part on predetermined rule 2b, which of the variables characterizing the state of process 2 have an influence on the likelihood that an undesired event will occur. These variables and / or processing results obtained from these variables may then be included in training sample 3 according to block 112 and in training sample 4 according to block 123.

[0048] According to blocks 113 or 124, respectively, at least one statistical moment and / or time series of at least one state variable of process 2 may be included in training sample 3 or 4, respectively.

[0049] According to block 121, the process model 2a may be specifically configured to predict the future evolution of the state of the process 2 based on at least one current and / or past state of the process 2.

[0050] According to block 122, the process model 2a: Machine learning models, and / or Simulation models, and / or A surrogate approximation of this simulation model may be provided.

[0051] According to block 131, the predictive model 1 may generate, based on samples 3, 4 within a given time window, a prediction 5 of the likelihood that an undesirable event will occur at the end of this time window.

[0052] FIG. 2 is a schematic flow chart of an exemplary embodiment of a method 200 for performing a process 2 on at least one industrial plant or in at least one electrical network.

[0053] In step 210, one or more samples 7 representing states of process 2 are provided to a trained predictive model 1* and / or its surrogate approximation 1**. The trained predictive model 1* or surrogate approximation 1** then outputs a prediction 5 of the likelihood of an undesirable event occurring in the states of process 2 represented by the one or more samples.

[0054] In step 220, the prediction 5 is tested against at least one predetermined criterion 8.

[0055] If criterion 8 is met (truth value 1), an alarm may be output to operator 16 of process 2 in step 230. Alternatively, or in combination, execution of process 2 may be modified in step 240 with the goal of reducing the likelihood of an undesired event occurring.

[0056] In accordance with block 211, samples representative of the state of process 2 may be collected by an edge system 10 at an industrial plant or other site involved in the execution of process 2. In step 250, the edge system 10 may then provide the samples to the cloud platform 11.

[0057] In step 260, the cloud platform 11 may then train and / or update the predictive model 1 based on the samples 7 received from the edge system 10. The cloud platform 11 may then, in step 270, create a surrogate approximation 1** for the trained and / or updated predictive model 1* and / or updates to such approximation 1**.

[0058] This surrogate approximation l**, and / or its updates, may then be provided back to the edge system 10 in step 280. On the edge system 10, this new and / or updated surrogate approximation l** may then be evaluated to obtain a prediction 5, according to block 212.

[0059] FIG. 3 illustrates an exemplary arrangement of edge system 10 and cloud platform 11 that may be used to perform method 200.

[0060] 3, an industrial plant performing a process 2 includes sensors 12a, actuators 12b, and controllers 12c. The sensors 12a and actuators 12b generate samples 7 of data indicative of the state of the process 2. These samples are captured and pre-processed by an edge data manager 14 of the edge system 10. They are passed to a process control system 13, which then provides control actions 9 to the actuators 12b and controller 12c to react to the state of the process 2 reflected in the samples 7.

[0061] When determining the control action 9, the process control system 13 also considers the prediction 5 output by the surrogate approximation 1** regarding the likelihood of the occurrence of the undesired event. That is, the process control system 13 may guide the process 2 by the control action 9 to avoid the occurrence of the undesired event. The process control system 13 also communicates with an operator 16 of the industrial plant.

[0062] The samples 7 are passed to the cloud platform 11 where the trained predictive model 1* resides. They are used to further train this model 1*. After an update to the predictive model 1*, the cloud platform 11 also generates a corresponding update to the surrogate approximation 1**. This update is passed back to the edge system 10.

[0063] The cloud platform 11 also includes an operational safety interface 15 for communicating with a plant operator 16 . The inventions described in the claims of the present application as originally filed are set forth below. [C1] 1. A method (100) for training a predictive model (1) for predicting the likelihood of at least one predetermined undesirable event occurring during execution of a process (2) using training samples (3), each training sample (3) comprising data characterizing a state of the process (2), the method (100) comprising: obtaining (110) training samples (3) representing states of the process (2) that do not cause the undesired event, and labeling the training samples with a predetermined low probability that the undesired event will occur; obtaining (120) further training samples (4) representing states of the process (2) that have an increased likelihood of causing the undesired event based at least in part on a process model (2a) and a set of predetermined rules (2b) that define at which states of the process (2) the undesired event has an increased likelihood of occurring, and labeling the training samples (4) with the increased likelihood; providing (130) the training samples (3, 4) to the predictive model (1) to be trained in order to obtain from the predictive model (1) a prediction (5) of the likelihood of the undesired event occurring at a state of the process (2) represented by each training sample (3, 4); rating (140) the difference between said prediction (5) and said label of each said training sample (3, 4) by a predetermined loss function (6); optimizing (150) parameters (1 a) characterizing the behavior of the predictive model (1) so that when predictions (5) are made on further samples (3, 4), the ratings (6 a) of the loss function (6) are likely to improve; A method (100) comprising the steps of: [C2] The method (100) according to C1, wherein the undesired event comprises a safety interlock event that forces at least a partial stop and / or shutdown of the process (2) and / or an industrial plant or electrical network in which the process (2) is running. [C3] The likelihood of the undesired event occurring is a measure of the probability that said undesired event will occur; and / or a measure of the proximity of the state of the process (2) to a state that causes the undesired event, The method (100) according to C1 or C2, wherein the measurement is performed. [C4] The method (100) according to any one of C1 to C3, wherein the process model (2a) is specifically configured (121) to predict the future evolution of the state of the process (2) based on at least one current and / or past state of the process (2). [C5] The process model (2a) is Machine learning models, and / or Simulation models, and / or a surrogate approximation of said simulation model; The method (100) according to any one of C1 to C4, comprising (122) [C6] determining (111) which of the variables characterizing the state of the process (2) have an influence on the likelihood of the undesired event occurring, based at least in part on the predetermined rules (2b); including (112, 123) said variables and / or processing results obtained from said variables in said training samples (3, 4); The method (100) according to any one of C1 to C5, further comprising: [C7] A method (100) according to any one of C1 to C6, further comprising including (113, 124) at least one statistical moment and / or time series of at least one state variable of said process (2) in said training samples (3, 4). [C8] A method (100) according to any one of C1 to C7, further comprising obtaining (131) by the predictive model (1) a prediction (5) of the likelihood of the undesirable event occurring at the end of a predetermined time window based on samples (3, 4) within the time window. [C9] A method (100) according to any one of C1 to C8, further comprising approximating (160) the behavior of the trained predictive model (1*) using a surrogate model (1**) that is computationally less expensive to evaluate than the trained predictive model (1*). [C10] A method (200) for carrying out a process (2) on at least one industrial plant or in at least one electrical network, comprising: providing (210) one or more samples (7) representing states of the process (2) to a trained predictive model (1*) and / or a surrogate approximation (1**) thereof to obtain a prediction (5) of the likelihood of an undesired event occurring in the state of the process (2) represented by the one or more samples; testing (220) said prediction (5) against at least one predetermined criterion (8); In response to the criterion (8) being met, issuing an alarm (230) to an operator (16) of the process (2) and / or modifying (240) the execution of the process (2) with the goal of reducing the likelihood of the undesired event occurring; A method (200) comprising the steps of: [C11] Modifying 240 the execution of the process includes: providing (241) samples (7) representing a plurality of candidate states of the process (2) different from a current state of the process (2) to the predictive model (1), thereby obtaining a probability of occurrence of the undesired event for the candidate states; directing (242) the execution of the process (2) towards a candidate state that has the least likelihood of occurrence of the undesired event as a target state; The method (200) according to C10, comprising: [C12] collecting (211) samples (7) representative of the state of the process (2) by edge systems (10) at industrial plants or other sites involved in the execution of the process (2); providing (250) the samples (7) to a cloud platform (11) by the edge system (10); training and / or updating (260) the predictive model (1) on the cloud platform (11) based on the samples (7) obtained from the edge system (10); generating (270) a surrogate approximation (1**) for the trained and / or updated predictive model (1*) and / or updates to the surrogate approximation (1**) by the cloud platform (11); providing (280) the surrogate approximation (1**) and / or the update to the surrogate approximation to the edge system (10) by the cloud platform (11); evaluating (212) the surrogate approximation (1**) to obtain the prediction (5) of the likelihood of the undesired event occurring on the edge system (10); The method (200) according to C11, further comprising: [C13] A computer program comprising machine-readable instructions that, when executed by one or more computers, cause the one or more computers to perform the method (100, 200) of any one of C1 to C12. [C14] A non-transitory machine-readable storage medium and / or download product having a computer program according to C13. [C15] One or more computers having a computer program as described in C13 and / or a non-transitory machine-readable storage medium and / or download product as described in C14. [Explanation of symbols]

[0064] 1. Prediction model 1*... fully trained state of model 1 1**...Surrogate approximation of the trained model 1* 1a…Parameters characterizing the behavior of prediction model 1 1a*...Parameter 1a fully optimized 2. Process 2a...Process Model 2b…Rules about increased likelihood of undesirable events 3...Training samples that do not have an increased probability of undesired events 4...Training samples with increased probability of unwanted events 5. Predicting the likelihood of undesirable events 6...Loss function 6a…Rating by loss function 7...Samples acquired during execution of process 2 8... Criteria for Prediction 5 9...Control Actions 10. Edge Systems 11. Cloud Platform 12a...Sensor 12b...Actuator 12c...controller 13...Process control system for process 2 14. Edge Data Manager 15...Driving safety interface 16...Industrial plant operator 100...Method for training a predictive model 1 110...Get training sample 3 111...Determine variables that affect the likelihood of an undesired event 112...Include influential variables in training sample 3 113…Include statistical moments in training sample 3 120…Get 4 more training samples 121...Select process model 2a to predict future states 122...Select a specific type of Process Model 2a 123...Include influential variables in further training sample 4 124… statistical moments included in further training samples 4 130…Provide training samples 3 and 4 to prediction model 1 131...Get a forecast at the end of the time frame 140…Rate the difference between prediction 5 and the labels of samples 3 and 4 150...Optimize parameter 1a 160...Create a surrogate approximation 1** for the trained model 1* 200...Method for executing process 2 210…Provide sample 7 to trained predictive model 1* 211...Collecting samples by edge system 10 212...Evaluate the surrogate approximation 1** on the edge system 10 220…Test your predictions against criterion 8 230: Output an alarm to operator 16 240...Fix process execution 241…Provide candidate state samples to prediction model 1 242...Guide the process to avoid undesirable events 250…Providing sample 7 to cloud platform 11 260...Training and / or updating a predictive model 1 270...Create a surrogate approximation 1** 280...Providing updates to Edge System 10

Claims

1. 1. A method (100) for training a predictive model (1) for predicting the likelihood of at least one predetermined undesirable event occurring during execution of a process (2) using training samples (3), each training sample (3) comprising data characterizing a state of the process (2), the method (100) comprising: obtaining (110) first training samples (3) representing states of the process (2) that do not cause the undesired event, and labeling the first training samples with a pre-set low probability that the undesired event will occur; obtaining (120) second training samples (4) representing states of the process (2) that have an increased likelihood of causing the undesired event based at least in part on a process model (2a) and a set of predetermined rules (2b) that define at which states of the process (2) the undesired event has an increased likelihood of occurring, and labeling the second training samples (4) with the increased likelihood; providing (130) the first and second training samples (3, 4) to the predictive model (1) to be trained to obtain from the predictive model (1) a prediction (5) of the likelihood of the undesired event occurring at a state of the process (2) represented by each of the first and second training samples (3, 4); rating (140) the difference between said prediction (5) and said label of said training sample (3, 4) of each of said first and second training samples by a predetermined loss function (6); optimizing (150) parameters (1 a) characterizing the behavior of the predictive model (1) so that when a prediction (5) is made on the first and second training samples (3, 4), the rating (6 a) according to the loss function (6) is likely to improve; A method (100) comprising the steps of:

2. 2. The method (100) of claim 1, wherein the undesirable event comprises a safety interlock event that forces at least a partial shutdown and / or stop of the process (2) and / or an industrial plant or electrical network in which the process (2) is running.

3. The likelihood of the undesired event occurring is a measure of the probability of said undesired event occurring, and / or a measure of the proximity of the state of the process (2) to a state that causes the undesired event, The method (100) of claim 1 or 2, wherein the temperature is measured.

4. 2. The method (100) of claim 1, wherein the process model (2a) is configured (121) to predict a future evolution of the state of the process (2) based on at least one current and / or past state of the process (2).

5. The process model (2a) is machine learning models, and / or simulation models, and / or a surrogate approximation of the simulation model; The method (100) of claim 1, comprising (122)

6. determining (111) which of the variables characterizing the state of the process (2) have an influence on the likelihood of the undesired event occurring, based at least in part on the predetermined rules (2b); including (112, 123) said variables and / or processing results obtained from said variables in said first and second training samples (3, 4); The method (100) of claim 1 further comprising:

7. 10. The method (100) of claim 1 or 6, further comprising including (113, 124) at least one statistical moment and / or time series of at least one state variable of the process (2) in the first and second training samples (3, 4).

8. 2. The method (100) of claim 1, further comprising obtaining (131) by the predictive model (1) a prediction (5) of the likelihood of the undesirable event occurring at the end of a predetermined time window based on the first and second training samples (3, 4) within the time window.

9. 2. The method (100) of claim 1, further comprising approximating (160) the behavior of the trained predictive model (1*) with a surrogate model (1**) that is computationally less expensive to evaluate than the trained predictive model (1*).

10. A method (200) for carrying out a process (2) on at least one industrial plant or in at least one electrical network, comprising: providing (210) one or more samples (7) representative of a state of the process (2) to a predictive model (1*) trained according to the method of claim 1 and / or a surrogate approximation (1**) thereof to obtain a prediction (5) of the likelihood of an undesired event occurring in the state of the process (2) represented by the one or more samples; - testing (220) said prediction (5) against at least one predetermined criterion (8); - in response to said criterion (8) being met, issuing an alarm (230) to an operator (16) of said process (2) and / or modifying (240) the execution of said process (2) with the goal of reducing the likelihood of said undesired event occurring; A method (200) comprising the steps of:

11. Modifying 240 the execution of the process includes: providing (241) samples (7) representing a plurality of candidate states of the process (2) different from a current state of the process (2) to the predictive model (1), thereby obtaining a probability of occurrence of the undesired event for the candidate states; directing (242) the execution of the process (2) towards a candidate state that has the least likelihood of occurrence of the undesired event as a target state; The method (200) of claim 10, comprising:

12. collecting (211) samples (7) representative of the state of the process (2) by edge systems (10) at industrial plants or other sites involved in the execution of the process (2); providing (250) the samples (7) to a cloud platform (11) by the edge system (10); training and / or updating (260) the predictive model (1) on the cloud platform (11) based on the samples (7) obtained from the edge systems (10); generating (270) a surrogate approximation (1**) for the trained and / or updated predictive model (1*) and / or updates to the surrogate approximation (1**) by the cloud platform (11); providing (280) the surrogate approximation (1**) and / or the updates to the surrogate approximation to the edge system (10) by the cloud platform (11); evaluating (212) the surrogate approximation (1**) to obtain the prediction (5) of the likelihood of the undesired event occurring on the edge system (10); The method (200) of claim 11 further comprising:

13. A computer program comprising machine-readable instructions that, when executed by one or more computers, cause the one or more computers to perform the method (100, 200) of claim 1 or 10.

14. A non-transitory machine-readable storage medium having the computer program of claim 13.

15. One or more computers comprising a computer program according to claim 13.

16. 15. One or more computers comprising the non-transitory machine-readable storage medium of claim 14.

Images