Method and apparatus for monitoring a chemical production process

Abstract

The present disclosure is directed to a method and apparatus (100) for controlling a chemical production process, for instance by detecting one or more events associated with a chemical reaction (101) during a chemical production process. The apparatus is configured to receive at least a sequence of image data (104) representing reaction conditions of the chemical reaction, to map, at a mapping unit (110), one or more image frames of the sequence of image data to at least one lower level representation (P), to generate an event trigger (112) based on a change (ΔP) in the at least one lower level representation that is associated with a respective change in the reaction conditions of the chemical production process and to provide the event trigger in association with the sequence of image data for monitoring the chemical production process. The method and apparatus enable improved assessment of the chemical production process.

Description

[0001] Method and apparatus for monitoring a chemical production process

[0002] FIELD OF THE INVENTION

[0003] The disclosure relates to the field of monitoring a chemical production process by data- driven artificial intelligence. The disclosure specifically relates to safe and reliable monitoring chemical production processes by detecting one or more events associated with a chemical processing during a chemical production process. Moreover, the disclosure is directed to an apparatus for detecting one or more events, to a chemical equipment including an apparatus for detecting one or more events, to the use of event trigger for retrospective assessment of the chemical production process, and to a computer program.

[0004] BACKGROUND OF THE INVENTION

[0005] Chemical production processes, which typically involve one or more chemical reactions or chemical processes, are typically performed under human supervision or monitoring. The chemical production process is sometimes monitored or supervised with the aid of dedicated sensors that monitor chemical or physical process variables providing time-dependent sensor data. Assessing the chemical production process (e.g., a posteriori) implies having to check a sequence of sensor data that represents the whole chemical reaction process in search for events, or significant changes in the chemical reactions that can be associated to the quality of the chemical production process. This turns out to be particularly time consuming, especially in the case of chemical production processes including chemical reactions spanning a long time, e.g., several hours.

[0006] SUMMARY OF THE INVENTION

[0007] It is therefore an object of the present invention to provide a method and an apparatus for enabling an improved assessment of a chemical production process. This enables in turn an increase of the amount of chemical production processes that can be executed or carried out without in-situ human supervision by reducing the time necessary to assess the chemical production process.

[0008] According to a first aspect of the present invention, a method for monitoring a chemical production is disclosed. The method comprises the step of receiving at least one sequence of image data representing reaction conditions of the chemical reaction. Further, the method comprises the step of mapping one or more image frames of the at least one sequence of image data to at least one lower level representation. The method also comprises the steps of generating an event trigger based on a change in the at least one lower level representation that is associated with a respective change in the reaction conditions of the chemical production process, and providing the event trigger in association with the sequence of image data for monitoring the chemical production process.

[0009] The sequence of image data may be in particular a sequence of time-dependent image frames recorded, or otherwise obtained, during the chemical production process. Each image frame thus corresponds to a specific time step of a sequence of time steps, or, in other words, each image frame is time-stamped or associated to a specific time with respect to the chemical production process. The chemical production process involves one or more chemical reactions that cause a chemical transformation of one set of chemical substances, typically referred to as reactants or reagents, to another set of chemical substances, typically referred to as products. A chemical reaction often comprises a sequence of individual elementary reactions, and occur in a time-dependent manner at a characteristic reaction rate, which depends on the conditions of the chemical reaction, such as temperature and concentration of chemical reactants.

[0010] The chemical reaction process may also involve physical transformations of the chemical substances, which cause a change of the physical properties of a chemical substance, including its phase, shape, density, volume, etc.

[0011] The basic steps in a chemical production process are generally referred to as unit operations, and involve a physical change or a chemical transformation such as, but not limited to separation, crystallization, evaporation, filtration, polymerization, isomerization etc. The unit operations, as well as the chemical and / or physical changes of the chemical substances, are time dependent and the occurrences of unit operations and chemical and / or BASF SE 220995 physical changes can be captured by the at least one set of image data, which is provided with a sufficient sampling rate.

[0012] The lower dimensional representation may be, in particular, a representation, for instance in the form of an output parameter, having a dimension that is lower than the dimension of input data, e.g., the at least one sequence of image data, that is provided as input for the mapping step. The mapping thus transforms, via a mapping procedure, one or more image frames of the at least one sequence of image data, the image frames having a dimension “m”, into a respective output representation or parameter having a dimension “n” such that “m>n”. A data transformation process is thus carried out that transforms, for example by applying a predetermined mathematical function, the data from an image frame of the sequence of image frames into a transformed value, namely the representation or output parameter, having a lower dimension. This reduction of the dimensionality involves the transformation of data from a high-dimensional space into a low dimensional space so that the lower dimensional or lower level representation, retains meaningful properties of the input data, e.g., the at least one sequence of image data. The lower dimensional, or lower level, representation thus refers to the result of mapping the image frame, having pixels whose position is characterized by two space dimensions, to a feature space or latent space associated to n-dimensional feature vectors. Here, feature may refer to an individual measurable property or characteristic of the image frame, in particular of the pixel values of the pixels of the image frame. Examples for said properties or characteristics may include, but are not limited to, arrangements of pixels, pixel values forming a pattern, statistically determined values associates to a region or the whole image frame, etc.

[0013] Typically, events occurring in high-dimensional spaces are more challenging to identify due to the sparsity of the data and the relative distance between points becoming less meaningful. The identification of such events is thus easier in a lower-dimensional space.

[0014] In the present disclosure, the event trigger is indicative of a change in the at least one lower level representation that is associated via a mapping procedure to a given image frame or group of image frames. By providing the event trigger in association with the sequence of image data, a user is directed to that specific section or part of the sequence of image data in which an event has occurred, where the occurrence of an event in the chemical production process is given by a change on the lower level representation. Events can thus be define as variations in the chemical production process recorded in the at least one sequence of image data, which, when the frames of the sequence of image data is mapped into a lower dimension representation, produces a significant change in the lower dimension representation, for instance, a change exceeding a predetermined threshold value. BASF SE 220995

[0015] This is then consider as an event, and an event trigger associated to said identified event is generated and provided in association with the sequence of image data for monitoring the chemical production process.

[0016] The step of generating an event trigger based on the change in at the least one lower level representation includes recording the sequence of image data, generating, based on a time step of a respective event trigger, an event flag linked to the corresponding time step or image frame of the sequence of image data and compressing the sequence of image data based on the event flag and providing a time-compressed sequence of image data. For instance, the event flag indicates those points in time where an event has been detected as an event trigger. The recorded sequence of image is then provided in a compressed manner such that the image frames associated with the event trigger (those indicated by the event flag) are provided to a user in a predetermined highlighted manner, such that they are easily recognizable, therefore reducing the time needed to assess the chemical reaction. The compression is here a compression in the time domain with respect to a duration of the chemical production process.

[0017] The user can therefore selectively check the relevant parts of the sequence of image data where an occurrence of an event takes place without having to check the whole sequence corresponding to the whole chemical production process. A non-exhaustive list of events that can be detected during a chemical production process using a sequence of image data as input data include change in color, appearance or disappearance of phases, increase or decrease of turbidity, appearance of disappearance of foam or bubbles, appearance or disappearance of solid particles, changes in the fill level, formation or appearance of flames, sparks, or smoke, etc.

[0018] Thus, the method of the first aspect not only enables as improved assessment of a chemical production but also enables a safe execution of the chemical production processes which impacts the quality of the chemical product resulting from the chemical production process. This is be achieved by a proper monitoring of the chemical production process, in particular using image data from which significant features indicative of relevant events during the chemical reaction process are extracted, or otherwise identified, in a lower level representation. This represents an advantage when compared to known monitoring method where only the operating conditions in terms of sensor data is monitored and used for control.

[0019] In the following, embodiments of the method of the first aspect of the invention will be described. BASF SE 220995

[0020] In an embodiment, the step of mapping one or more image frames of the at least one sequence of image data to at least one lower level representation includes providing the one or more image frames of the sequence of image data to a feature extraction model. The feature extraction model is configured to map the image frame to at least one lower level representation and then the at least one lower level representation is generated and provided. The feature extraction model is used to reduce the number of resources needing for determining the occurrence of an event in the chemical reaction without losing important or relevant information. Applying feature extraction techniques thus enables the reduction of the dimensionality of the data that is necessary to process said input data in an effective manner. In other words, feature extraction involves creating new features that still capture the essential information from the original data, i.e., the sequence of image data, but in a more efficient way. In particular, when dealing with large datasets such as in the case of image processing, it's common to have data with numerous features, many of which may be irrelevant or redundant. Feature extraction allows forthe simplification of the data which helps algorithms to run faster and more effectively.

[0021] In another embodiment, the method includes providing the at least one lower level representation for performing an anomaly detection method. Additionally, or alternatively, the method includes providing the at least one lower level representation for performing a clustering method.

[0022] The lower level representation comprises a series of instances or values, each associated to a respective image frame or group of image frames.

[0023] In an embodiment, the anomaly detection method includes identifying lower level representation instances (e.g., values) of the lower level representation that differ in a significantly relevant manner from a majority of the lower level representation instances according to a predetermine decision algorithm, and generating the event trigger in association with the identified lower level representation instances.

[0024] Anomaly detection, also referred to as outlier detection, may include the identification of rare items, events or observations which deviate in a significantly relevant manner, from the majority of the data, and therefore do not conform to a well-defined notion of normal behavior. Normal behavior here is understood as a current or expected state of the chemical reaction. Preferably, unsupervised anomaly detection techniques are applied for anomaly detection. Alternatively, supervised or semi-supervised anomaly detection techniques can also be applied. The lower level representations obtained from different image frames or groups of image frames are used to determine the event trigger based on a change of BASF SE 220995 the lower level representation determined by the decision algorithm or any other suitably anomaly detection algorithm, for example using predetermined threshold values.

[0025] Additionally, or alternatively, the lower level representation is provided for performing a clustering method that comprises associating each lower level representation instance to one cluster of a set of clusters in accordance with a predetermined clustering algorithm and generating the event trigger in association with one or more target clusters from the set of clusters. In this embodiment, clustering techniques are additionally or alternatively applied to the different instances or values of the lower level representation. Clustering, also referred to as cluster analysis, involves grouping a set of objects or data points, in this particular case the individual values or instances of the lower level representation, in such a way that objects / data points belonging to the same group or cluster are more similar to each other that to those in other groups or clusters. The similarity is determined based on a predetermined property or variable. The lower level representations obtained from different image frames or groups of image frames are used to determine the event trigger based on a change of the lower level representation determined by a clustering algorithm. The clustering can be based on a variety of clustering methods or algorithms including, but not limited to, vector quantization using for instance k-means clustering, spectral clustering, and density-based spatial clustering of applications with noise DBSCAN, fuzzy C-means FCM. In another embodiment, a regularization term is added to the clustering-loss, which increases the probability that neighboring image frames end up in the same cluster.

[0026] In different embodiments, the feature extractor applied to the one or more image frames of the sequence of image data may result in a vector to which image intensities are reshaped, where the vector is considered as the lower level representation. Alternatively image histograms, histograms of the flow-fields of optical flow computations between image frames, features extracted from pre-trained neural networks (e.g. resnet pre-trained on image net) or any combination of the above can be used to determine the lower level representation, such as a vector.

[0027] In another embodiment, the step of mapping one or more image frames of the sequence of image data to the at least one lower level representation includes providing an initial part of the sequence of image data to train a data-driven model, providing a monitoring part of the sequence of image data to the trained data-driven model, wherein the monitoring part covers a time period following the time period of the initial part. Then, the step of generating the at least one lower level representation follows. The initial part of the sequence of image data is therefore used to train the data-driven model that will be used for evaluating another part of the sequence of image data, the so-called monitoring part. The monitoring part and BASF SE 220995 the initial part are non-overlapping parts of the sequence of image data, e.g. a given image frame does not belong to both the initial and the monitoring part.

[0028] In a particular embodiment, once an event trigger is generated, a new initial part comprising part of the sequence of image data afterthe detection of the event is provided for re-training the data driven model. The state of the chemical reaction after the detection of the event, e.g. the generation of the event trigger, is used as a new normal for the following part of the section of image data. Thus, in this embodiment, the data-driven model is adaptively retrained to a new normal, while the past normal is at least partially ignored or forgotten.

[0029] In another embodiment, the step of generating the at least one lower level representation includes determining as a lower level representation a loss value by a loss function depending on the image data provided to the data-driven model and on image data provided by the data-driven model. For instance, the data-driven model can be implemented as an autoencoder. An autoencoder refers to a type of artificial neural network primarily used for coding of unlabeled data, also referred to as unsupervised learning. An autoencoder is configured to learn two functions; an encoding function that transforms the input data, i.e. one or more image frames of the sequence of image data, and a decoding function that recreates the input data from the encoded representation. The autoencoder comprises two main parts; and encoder that maps a so-called message to a so-called code, and a decoder that reconstructs the message from the code. The autoencoder can be advantageously used for anomaly detection in the sequence of image data. By learning to replicate the most salient features in the initial part of the sequence of image data (or any other subsequent initial part following an event detection), the model is encouraged to learn to reproduce the most frequently observed characteristics. Thus, when facing anomalies, the model worsens its reconstruction performance. This reflects in the so-called loss value of the loss function L. Therefore, the fact that the loss value of the loss function for a given set of input data is higher than a predetermined threshold can be interpreted as a change in the at least one lower level representation, in this case the loss value of the lost function resulting in an event trigger.

[0030] In a particular embodiment, which may comprise any of the technical features described with respect to the previous embodiment, the sequence of image data is provided in form of a video stream or as a video file. The video file can be representative of the whole chemical reaction process, or of a part thereof. BASF SE 220995

[0031] For example, in an embodiment, a summary video is provided as a time-compressed sequence of image data. The summary video includes a subset of image frames associated to the detected event occurrences by the event flag. These image frames are provided with an event playback rate factor that is lower than a base playback factor associated to the remaining image frames. Thus, in the provided summary video, the image frames obtained at times where no event is detected are played at a higher speed than those where an event is detected and is thus timely compressed with respect to chemical reaction. The user’s attention is thus directed to those frames where a significant event has happened and an easier and less time consuming assessment of the chemical reaction is enabled. In particular, the event playback rate factor and the base playback factor are determined based on a predetermined summary video file duration. A predetermined summary video file duration can be, for example 10 minutes, 7 minutes, 5 minutes or 3 minutes. A fixed summary video file duration can be advantageous for comparing several runs of the same chemical reaction.

[0032] In another embodiment, at least part of the image frames that are not associated to the detected event occurrences) are not included in the summary video file. The elimination from the summary video files of frames that are not associated to the relevant events results in a time compression of the summary video file with respect to the duration of the chemical production process and thus the summary video file constitutes a compressed sequence of image data.

[0033] In addition to the time-compression applied to the time-compressed sequence of image data, other compression techniques can be applied thereto, for example in terms of file size (or required memory storage) or the provided summary video file or the frame resolution of the frames of the provided summary video file. The compression can also be dependent on the content of the video file, such that the image frames related to the identified events have a higher resolution than the image frames not related to the identified events.

[0034] Alternatively, or additionally, the method comprises classifying the event trigger to pre-defined event specifiers, generating, based on the time step of the event trigger, the event specifier linked to the corresponding time step of the sequence of image data, compressing the sequence of image data based on the event flag and providing a compressed sequence of image data. Here, the event trigger can be classified based on a list of predefined event specifiers that are associated to a corresponding detectable event. Some particular changes of the at least one lower level representations are characteristic of a particular event and that particular event can be highlighted in the provided compressed sequence of image data such that user can easily recognize it. Predefined event specifiers can relate BASF SE 220995 to particular important events in the chemical process such as, for instance, phase transitions, changes in turbidity, addition of material, etc. Alternatively, or additionally, pre-defined event specifiers can relate to unexpected or unwanted events, such as formation of sparks, or reactions conditions that that are outside the desirable range such as a sudden increase in pressure due to a high reaction rate or an unexpected increase in temperature, due for example an unwanted exothermic reaction. The pre-defined events can be additionally associated to an event time span. For instance, a predetermined change of color happening before or after an expected event time can be considered an unwanted or unexpected event in the chemical production process, but a normal or expected event when it happens at the expected event time span.

[0035] Further, alternatively, or additionally, the method comprises classifying event trigger to predefined event specifiers, and generating a control trigger to control the chemical reaction. Thus, in this embodiment, the method of the first aspect can be also implemented as a method for controlling the chemical production process. As a response to an occurrence of an event classified or associated to a pre-defined event specifier, control signals such as control triggers can be provided for controlling the chemical production process. The control signals can be provided to steer a well-behaved chemical production process based on the occurrence of an expected event, e.g. controlling the addition of a new reactant once an expected phase transition has been determined but can also be provided to stop an ill- behaved chemical production process based on the occurrence of an unexpected event. In an extreme case, the occurrence and identification of a potentially dangerous event, such as spark formation, can trigger the provision of a control trigger in the form of an stopcommand for halting the chemical production process (emergency stop, reactor shut-off) or at least changing the conditions in the chemical production process to lower the risk (e.g., lower the temperature in the reaction, close / open inlet / outlet valves, add neutralizing agents, etc.). This is particularly relevant when the sequence of image data is provided in form of a video stream characterizing the chemical reaction in situ. The chemical reaction can be controlled based on the detected changes of the lower level representations that are indicative of a predefined event specifiers.

[0036] In a particular embodiment, the sequence of image data is recorded during the chemical production process. In case of a continuous production process the recording monitors the production process continuously. In case of a batch production process the recording monitors the production process per batch. Preferably, the recording is started on start of the chemical process, such as providing of chemical educts to a chemical reactor, and stopped on stop of the chemical process, such as providing of chemical products as result of the chemical reaction. BASF SE 220995

[0037] In yet another embodiment, in addition to the sequence of image data, time-dependent sensor data is measured and provided by one or more sensors monitoring a process parameter of the reaction process. In this embodiment, the additional time-dependent sensor data is further used to generate the event trigger. This particular embodiment of the method comprises receiving, in addition to the at least one sequence of image data, time-dependent sensor data that is indicative of values of a process parameter of the chemical production process. Thus, the time-dependent sensor data represents the time variation of the value of a process parameter, such as, for instance, a temperature, a flow, a density, a viscosity, a volume, a turbidity, a concentration of a gas, etc. The method also includes fusing the image frames of the at least one sequence of image data with the additionally received time-dependent sensor data for a corresponding time step. In other words, both the sequence of image data and the time-dependent sensor data are referred to a common time frame, such that frames of the image data are fused with sensor data substantially obtained, within resolution constraints, at the same time. Thus, the image frames of the sequence of image data are fused, e.g., concatenated or otherwise associated, with the additional time-dependent sensor data for a corresponding time step, thereby generating fused data. The fused data is then provided as high dimensional input data for mapping to the at least one lower level representation. Changes in the value of the lower level representation are then assigned to event trigger that are provided in association with at least the sequence or image data, or, alternatively, in association to the fused data resulting from the combination of the sequence of image data and the time-dependent sensor data. As a non-limiting example, time-dependent sensor data indicative of a time-dependent temperature value, for instance in a reactor, can be advantageously fused or combined with the image data to form fuse data in the form of a heat map.

[0038] Alternatively, or additionally, the additional time-dependent sensor data is analyzed for event trigger generation, as explained above. For instance, the time-dependent sensor data is mapped to the at least one lower level representation and the event trigger is generated based on a change thereof. The sensor data can also be provided as input data to a suitable feature extraction model configured to map the sensor data to the at least one lower level representation, which, in turn, can be used for anomaly detection and / or clustering. The sequence of sensor data, in particular an initial part thereof, can be provided to train a data-driven model for generating a trained data-driven model to which a monitoring part of the sequence of sensor data. Also the sensor data can be provided in situ, as it is being recorded, or a posteriori, once the chemical reaction has ended. Predefined event specifiers can be associated to predetermined changes in the lower level representation obtained from the sequence of sensor data. BASF SE 220995

[0039] The sensor data is indicative of parameter values of one or more chemical or physical parameters associated to the chemical reaction. The sequence of image data and sensor data share a common time frame, such that both sequences can be compared in terms of time and a given value of the sequence of sensor data can be associated to at least an image frame of the sequence of image data. Sensor data may include, for instance, values obtained by a temperature sensor, a viscosity sensor, a pH sensor, a pressure sensor, a fill-level sensor, a dedicated turbidity sensor, a flow sensor, a gas concentration sensor, or any other suitable sensor for monitoring a parameter of the chemical reaction.

[0040] A second aspect of the present invention is formed by an apparatus for monitoring a chemical production process. In particular, the apparatus is also suitable for detecting one or more events associated with a chemical reaction during a chemical production process. The apparatus is configured to perform a method according to the first aspect of the invention and thus shares the advantages thereof.

[0041] In particular, the apparatus comprises an input unit for receiving the at least one sequence of image data representing the reaction conditions of the chemical production process. Additionally, the input unit can receive a sequence of time-dependent sensor data sharing the same time frame as the sequence of image data. The apparatus also comprises a mapping unit configured to map one or more image frames of the at least one sequence of image data and / or one or more values of the sequence of time-dependent sensor data to at least one lower level representation, and to generate and provide an event trigger based on a change in the at least one lower level representation.

[0042] In a particular embodiment, the apparatus also comprises an image processing unit that is configured to receive the sequence of image data and the event trigger and to generate and provide a summary video file wherein the image frames associated to a detected event are highlighted in a predetermined manner, as explained above.

[0043] In an embodiment, upon issuing an event trigger that is provided to the image processing unit, the image processing unit is configured to record the sequence of image data associated to the event, e.g. the image frames of the at least one sequence of image data representing the event and, optionally, those image frames immediately before the event and / or those image frames immediately after the event. The image processing unit can be also configured to generate, based on time step of event trigger, an event flag linked to the corresponding time step of sequence of image data; to compress the sequence of image data based on the event flag and to provide a compressed sequence of image data. BASF SE 220995

[0044] In an embodiment, the apparatus is further implemented as an apparatus for controlling a chemical production process, wherein the apparatus is configured to classify event triggers to pre-defined event specifiers associated to a corresponding detectable event, and, based thereon, to generate a control trigger to control the chemical production process.

[0045] A third aspect of the invention is formed by a chemical equipment including at least one chemical reactor and an apparatus according to the second aspect of the invention. The chemical equipment also includes one or more sensors for generating and providing the sequence of image data and / or sequence of sensor data. The sensors may include a camera unit configured to provide the sequence of image data, in particular in the form of a sequence of image frames, wherein each image frame comprises a plurality of pixels, each having a corresponding pixel value or set of pixel values. The pixel value or set of pixel values of each pixel can correspond to a brightness in a grey scale, to RGB components or any other suitable value for encoding image information on a pixel value.

[0046] In an embodiment, the chemical equipment, also referred to as chemical arrangement, comprises a plurality of camera units or optical sensor, each generating providing a respective sequence of image data. The sequences of image data may differ in the area of the reactor monitored by the respective camera unit and / or in a spectral resolution of the respective camera unit.

[0047] A fourth aspect of the present invention is formed by the use of event trigger in association with sequence of image data, in particular recorded sequence of image data and / or a time- compressed sequence of image data, the event trigger as provided according to the method according to the first aspect of the invention, for retrospective assessment of the chemical production process. Thus, the sequence of image data associated to the event trigger, which highlights or pinpoints to the frames of the sequence of image data where an event has occurred, or the time-compressed sequence of image data generated by an image processing unit using the recorded sequence of image data and the event trigger, (e.g. a summary video file) can be used for assessing, a posteriori -retrospectively- the behavior of the chemical production process, without the need to review the whole input data, thereby reducing the time needed for such review process and increasing the confidence of the assessment.

[0048] A fifth aspect of the invention is formed by a computer program that comprises instructions which, when executed by an apparatus in accordance with the second aspect of the inven- BASF SE 220995 tion, cause said apparatus to carry out the method for detecting one or more events associated with a chemical reaction during a chemical production process of the first aspect of the invention.

[0049] It shall be understood that the method described above, the apparatus described above and the computer program product described above have similar and / or identical preferred embodiments, in particular, as defined in the dependent claims.

[0050] It shall be understood that a preferred embodiment of the present invention can also be any combination of the dependent claim or above embodiments with a respective independent claim.

[0051] These and other aspects of the present invention will be apparent from and elucidated with reference to the embodiments described hereafter.

[0052] BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In the following drawings:

[0054] Fig. 1 illustrates an exemplary embodiment of a chemical equipment including an apparatus for detecting one or more events associated with a chemical reaction during a chemical production process, according to the invention;

[0055] Fig. 2 illustrates a time curve of values of a lower dimensional process parameter obtained from the sequence of image data indicative of the chemical reaction;

[0056] Fig. 3 illustrates another exemplary embodiment of a chemical equipment an apparatus for detecting one or more events associated with a chemical reaction during a chemical production process, according to the invention;

[0057] Fig. 4 and 5 each illustrate a process of compressing a video file of the chemical reaction;

[0058] Fig. 6 illustrates a flow diagram of a method according to the invention, performed by the apparatus according to Fig. 1 or Fig. 3; and

[0059] Fig. 7 and 8 each illustrate a method step of the method according to Fig. 5, wherein an image frame is mapped to a lower level representation. BASF SE 220995

[0060] DETAILED DESCRIPTION OF EMBODIMENTS

[0061] Fig. 1 illustrates a particular example of a chemical equipment, or chemical arrangement 200, in accordance with the invention. The chemical equipment 200 comprises a chemical reactor 203, exemplarily a batch reactor, into which input materials M1 , M2 are added. The input materials M1 , M2 then react in a chemical reaction 101 to produce one or more reaction products M3. This may happens in the frame of a chemical production process that can involve one or more chemical reactions performed in different reactors or process steps. The chemical reaction 101 is monitored by a sensor arrangement that comprises a camera unit 150 and which may also comprise additional sensor units 151 . The camera unit is configured to provide a sequence of image data 104 indicative of the state of the chemical reaction. The additional sensor units 151 can be configured to monitor a respective reaction parameter of the chemical reaction, such as, but not limited to, temperature, pressure, pH, flow, turbidity, filling level, density, concentration etc.

[0062] The chemical equipment 200 also comprises an apparatus 100 for monitoring and / or controlling chemical production process. The apparatus is suitable for detecting one or more events associated with the chemical reaction 101 during a chemical production process in the chemical reactor 202. This apparatus 100 is also referred to as detection device in the frame of this disclosure. Detectable events include, but are not limited to, a change in color, appearance or disappearance of phases, increase or decrease of turbidity, appearance of disappearance of foam or bubbles, appearance or disappearance of solid particle and changes in the fill level. The detectable events may differ in significance based on the time at which they occur with respect to the chemical production process. Detectable events may also include high-risk events such as generation of sparks or appearance of flames, smoke, etc.

[0063] The detection device 100 comprises an input unit or input interface 121 for receiving at least a sequence of image data 104 generated at and provided by an external camera unit 150 that is configured to monitor the chemical reaction or reactions101 taking place during the chemical production process. The camera unit 150 can monitor the chemical reaction 101 in the visible spectrum or in spectral regions different than the visible spectrum, such as in an infrared spectrum or ultraviolet spectrum.

[0064] The detection device 100 comprises a mapping unit 110 that is configured to map one or more image frames of the at least one sequence of image data 104 to at least one lower level representation P (see Fig. 2). For instance, as an example, the mapping unit 110 is BASF SE 220995 configured to map a brightness value of the pixels of each of the image frames to a respective average brightness values. The input data, i.e., the sequence of image data 104, has therefore a higher dimension than the resulting output parameter P indicative of the average brightness. The detection device 100 is further configured to analyze the lower level representation as a function of time and to generate an event trigger 112 based on a change AP in the at least one lower level representation.

[0065] The optional sensor unit or units 151 are configured to provide time-dependent sensor data, and the detection device 100 is configured to fuse the received sensor data with the sequence of image data 104 for generating fused data, which is then provided as input to the mapping unit 1 10 for generating the event trigger 112. The fused data therefore comprises not only data in form of pixel values of the pixels of the corresponding image frames, but also additional data pertaining to one or more variables characterizing the state of the chemical production process, such as temperature, pressure, flow, gas concentration, etc., which may be relevant for the determination of the event trigger.

[0066] Fig. 2 illustrates a time curve of values of a lower level representation obtained from the sequence of image data indicative of the time-varying state of the chemical production process. As an example, the lower lever representation shown in Fig. 2 is the average brightness of the pixels values of a given image frame of the sequence of image data 104. For example, in Fig. 2, the average brightness values P of the pixels of the image frames decreases in the time span labelled as te. The reduction in brightness is caused by an event 109 in the chemical reaction 101 taking place in the reactor 202, for example, a change in color 109. The reduction of the average brightness value can be considered a change in the lower level representation when it exceeds a predetermined threshold value. The event trigger 112 associated to this change AP is provided, for instance via an output unit 122 of the detection device 100, in association with the sequence of image data. For example, the event trigger 1 12 can be indicative of the time teat which the event 109 has occurred. The provision of the event trigger can be used by a user to selectively check the sequence of image data 104 at the time steps indicated by the event trigger.

[0067] The mapping unit 110 can for instance be implemented as a feature extractor model configured to map the image frames 104 to the lower level representation P. The feature extractor model provides the lower level representation, which can be used for anomaly detection or for clustering. Typically, in exemplary apparatuses where the lower level representation is used for clustering techniques, the sequence of image data is provided as a BASF SE 220995 whole after the chemical reaction has finished. Anomaly detection techniques can be however additionally used in exemplary apparatuses where the sequence of image data is provided in the form of a video stream, in particular in real time or quasi-real time and thus indicative of a current state of the chemical reaction.

[0068] The mapping unit 110 may be implemented as a data-driven model. In particular, an initial part 104a of the sequence of image data 104 is used to train the data-driven model of the mapping unit 110. For instance the initial part 104a of the sequence of image data 104 results in a lower level representation value of Pa. Then, a monitoring part 104b of the sequence of image data 104 is provided to the trained data-driven model, wherein the monitoring part 104b covers a time period following the time period of the initial part 104a. Following the example shown in Figs. 1 and 2, the reduction of the parameter value P is identified by comparing the lower level representation values obtained with the image frames of the monitoring part 104b with the values obtained with the image frames of the initial part 104a. Additionally, after the event 109 occurring during the time span te, the image frames following said event can be used to retrain the data-driven model such that the value of the lower dimensional parameter after the immediately after the event Pc is used as a new normal or reference value for identifying or determining further events in the chemical reaction.

[0069] Fig. 3 shows another embodiment of a chemical equipment 200 including an apparatus 100 for monitoring and / or controlling chemical production process. For the sake of simplicity, the same references will be used for those features of the chemical equipment 200 of Fig. 3 that have the same, or a similar, functionality as those of the chemical equipment discussed with reference to Fig. 1 .

[0070] The camera unit 150 is arranged with respect to the chemical reactor such that a field of view FOV of the camera unit covers at least a part of the reactor 202. The image frames 104 provided by the camera unit 150 are thus indicative of a state of the chemical production process at a respective different time, or in other words, the sequence of image data is indicative of the varying reaction conditions inside the reactor 202. The camera unit 150 records the chemical production process in the reactor and provides the recorded sequence of image frames in the form of a video stream, where the recorded image frames are provided “live” (as captured and internally processed) to the apparatus 100 and / or as a video file, after the chemical production process, or a part thereof, has finished.

[0071] The recording can be started on start of the chemical production process, such as when the input materials or chemical educts or reagents M1 , M2 are provided to the chemical BASF SE 220995 reactor 202, and stopped on stop of the chemical process, such when the output material or product of the chemical reaction M3 is formed or extracted from the reactor. The sequence of image data is recorded during the chemical production process, wherein in case of a continuous production process the camera unit monitors the production process continuously or in case of a batch production process the camera unit monitors the production process per batch. As explained with reference to Fig. 1 , the detection device 100 is advantageously configured to determine, using the received sequence of image data 104, an event trigger 112 indicative of a reaction event 109 in the reactor.

[0072] Furthermore, the detection device 100 comprises an image processing unit 114 that, using the event triggers 112 and the received sequence of image data 104, is configured to generate and provide a summary video file 116. In the summary video file, those image frames associated to events in the chemical reaction are highlighted in predetermined manner. Exemplary implementations of summary video files 116a, 116b provided by the image processing unit 114 are presented in Figs. 4 and 5.

[0073] Figs. 4 and 5 show a sequence of image data (video file) 104 indicative of the chemical production process, either as a whole, or of a part thereof. The differently shaded areas of the schematic representation of the video file represent respective states of the chemical production process. A transition between two different states is associated to a respective event 109a, 109b, 109c, 109d, and 109d in the chemical production process such as a change of color, formation of a phase, addition of a chemical reactant, etc. The mapping unit 1 10 is configured to determine and provide a respective event trigger 112a-112e associated to a corresponding one of the events 109a-109e detected during the chemical production process. The event triggers 112a-112e comprise respective time data indicative of a time step in the video file 104 at which the corresponding event 109a-109e occurred. The image processing unit 114 is configured to receive the video file e.g., the sequence of image data, 104 and the event triggers 112a-112e, and based thereon, to provide a summary video file 1 16a, 116b highlighting those image frames of the sequence of image data 104 corresponding to the detected events 109a-109e.

[0074] For instance, in Fig. 4, the image processing unit 114 is configured to provide a summary video file 116a only including those image frames of the sequence of image data 104 that show the corresponding event 109a-109e. The remaining image frames not corresponding to any event, or at least a significant portion thereof, are not included in the summary video file, such that the duration of the summary video file 1 16 is significantly shorter than the duration of the original video file 104. The summary video file 116a can comprise a first set of image frames obtained immediately before the start of the event, a second set of image BASF SE 220995 frames obtained during the occurrence of the event, and a third set of image frames obtained immediately after the occurrence of the event.

[0075] As shown in Fig. 5, the image processing unit 114 also receives the video file 104 and the event triggers 112a-112e generated by the mapping unit, and generates and provides a summary video file 116b. However, in the summary video file 116b, the image frames associated to the detected events 109a-109e are highlighted differently compared to the summary video file 116a of Fig. 4. In the summary video file 116b, subsets of image frames that are indicative or otherwise associated to the detected event 109a-109e in the input video file 104 have an event playback rate factor F1 that is lower than a base playback factor F2 associated to the remaining image frames, i.e., those image frames that are not representative of an event. In particular, in examples wherein the summary video file has a predetermined maximum allowed duration, the event playback rate factor F1 and the base playback factor F2 are determined based on said predetermined summary video file allowed duration.

[0076] The summary video file 116am 116b enables a quicker and more selective assessment of the chemical production process, since the user is directly presented with the image frames that show the significant events in the chemical reaction.

[0077] Returning to Fig. 3, the detection device 100 also comprises a reaction control unit 1 18 that is configured to receive the event triggers 1 12 and, based thereon, to control the chemical equipment 200 by providing control triggers 120 indicative of control instructions. For example, the reaction control unit 1 18 is configured to classify or assign event triggers to predefined event specifiers. The pre-define event specifiers or sequence of event specifiers, can be associated to a corresponding control trigger, such that, when the reaction control unit receives event triggers corresponding to the event specifier or sequence of event specifiers, the corresponding control trigger is provided. For instance, a control trigger indicative of an instruction to lower a temperature value in the reactor can be issued when a predetermined time span has lapsed after formation of bubbles has been detected as an event. In another example, a pre-defined event specifier corresponds to a risk situation such as the formation of sparks, flames or smoke. A control trigger associated to these pre-defined event specifier can be an instruction to perform an emergency stop of the chemical reaction process, or at least an instruction directed to change the conditions of the chemical reaction process in order to reduce the risk, such as lowering the temperature in the reaction, open- ing / closing valves, adding neutralizing agents for stopping a potentially dangerous chemical reaction, etc. BASF SE 220995

[0078] Additionally, the detection device 100 can comprise, or be otherwise connected to, one or more additional sensors monitoring a process parameter of the reaction process and configured to prove time-dependent sensor data, which is used to generate the event trigger. For this, the image frames of sequence of image data are fused (e.g. concatenated) with the sensor data for a corresponding time step and / or the sensor data is separately analyzed for an event trigger generation.

[0079] Fig. 6 illustrates a flow diagram of a method 500 according to the invention, in particular performed by the detection device 100 discussed with reference to Fig. 1 or Fig. 3. The method 500 is suitable for detecting one or more events 109 associated with a chemical reaction 101 during a chemical production process. The method 500 comprises, in a step 502, receiving a sequence of image data 104, in particular from a camera sensor or camera unit, the sequence of image data 104 representing reaction conditions of the chemical reaction. The method optionally comprises, as indicated by the dashed line, in a step 502a, receiving time-dependent sensor data from a sensor unit, the time-dependent sensor data being indicative of values of a process parameter of the chemical production process, such as, temperature, pressure, flow, turbidity, filling level, gas concentration, etc., and fusing this time-dependent sensor data with the sequence of image data forming fused data. The method further comprises, in a step 504, mapping one or more image frames of the sequence of image data 104 (or the fused data in case the image data has been fused with the additional time-dependent sensor data) to at least one lower level representation, LLR, for example parameter P discussed with reference to Fig. 2 above, or parameter L, discussed below with reference to Fig. 7. The method further comprises, in a step 506, generating an event trigger 112 based on a change in the at least one lower level representation, and, in a step 508, providing the event trigger in association with the sequence of image data.

[0080] In an exemplary embodiment of the method 500, the step of mapping one or more image frames of the sequence of image data 104 to at least one lower level representation can comprise providing, in a step 504a, the one or more image frames of the sequence of image data to a feature extraction model configured to map the image frame to at least one lower level representation and generating, in a step 504b, the at least one lower level representation. Additionally, or alternatively, the mapping step 504 can also comprise providing, in an step 504c, an initial part of the sequence of image data to train a data-driven model, and, in a step 504d, generating the lower level representation by providing a monitoring part of the sequence of image data to the trained data-driven model, wherein the monitoring part covers a time period following the time period of the initial part. BASF SE 220995

[0081] In exemplary alternative methods (not shown) performed by a detection apparatus that is alternatively or additionally configured to receive time-dependent sensor data from a dedicated sensor unit that monitors one or more parameters of the chemical reaction (e.g., temperature, density, flow, gas concentration, etc.), the additional time-dependent sensor data is used to generate the event trigger by fusing the image frames of the sequence of image data with the additional time-dependent sensor data for a corresponding time step and then by generating at least one lower level representation based on fused data. For example, information regarding the temperature in the reactor obtained from the sensor data provided by a temperature sensor can be combined with visual information obtained from the image data to determine whether an event has occurred. The combination of data from different sensor units may result in an increase of the dimensionality of the input data.

[0082] Additionally, or alternatively, the additional time-dependent sensor data can be directly used for event trigger generation. In this case, the time-dependent sensor data can be provided as input to a feature extraction model for determining the lower level representation, which can then be used for anomaly detection and / or clustering as explained above.

[0083] Fig. 7 and 8 each illustrate a method step of the method according to Fig. 5, wherein an image frame 104.1 is mapped to a lower level representation.

[0084] In particular, the detection device 100 (see Figs. 1 or 3) carries out a method as discussed with reference to Fig. 5. The detection device 100, processes the sequence of image data such that the image frames of the sequence are mapped to one or more lower level representation. At the beginning of th is process an image frame 104.1 is exemplarily represented by a matrix of brightness values each representing the brightness value of a pixel in the image frame 104.1. In Figs. 7 and 8, due to simplification purposes, only one image frame of an RGB color space is illustrated, corresponding to one two-dimensional matrix y (instead of tensor including three matrices, one for each of the three RGB components).

[0085] In the embodiment shown in Fig. 7, the matrix y is fed to a data driven model in the form of an autoencoder 510. The autoencoder is implemented as a convolutional neural network. In the encoding section 512 of the autoencoder, the matrix y is mapped to an input layer of the autoencoder and then convoluted with a kernel obtaining the values of another layer (one of the hidden layers). The hidden layer is again convoluted with the kernel obtaining the values of a further hidden layer and so on (depending on the number of hidden layers) until obtaining the so called latent space representation LSR, which has a lower dimension than the input matrix. This process (from initial matrix to latent space representation) can be regarded as applying feature extraction to the input matrix representing the input frame BASF SE 220995

[0086] 104.1. The latent space representation can be used as a lower level representation, for instance to control the reactor. In the decoding section 514 of the autoencoder 510 the input layer is reconstructed due to transposed convolution in order to receive the output layer. Thereafter, a loss function L can be calculated depending on the input matrix y and output matrix y’ of the autoencoder, wherein if the loss, i.e., the result of applying the loss function to the input and output matrices y, y’ (also a suitable lower level representation) is higher than a threshold value, an event trigger is determined and output. In the shown coordinate system 515, the image frame number is plotted on the x-axis, and on the y-axis, the corresponding loss values are represented. The dash line indicated the threshold value used to determine whether the lower level representation has experimented a change that can be identified with an occurrence of an event in the chemical reaction. This process can be regarded as anomaly detection and can be performed on a sequence of image data provided in situ as a stream.

[0087] When giving the autoencoder loss-feedback it can be trained due to backpropagation adjusting its node and weight values. The autoencoder is trained by inputting an initial part of the recorded video, wherein a monitoring part of the recorded video following the initial part is fed to the trained autoencoder in order to receive the event triggers.

[0088] Once an event has been detected and a corresponding event trigger has been generated, the current state of the chemical reaction can be considered as a new normal, with which the following image frames of the sequence of image data will be compared to determine a further event. In other words, the autoencoder is retrained using the image frames after the event as training data.

[0089] In the exemplary embodiment shown in Fig. 8 for the initial matrix y that represents the brightness values of different pixels of the input image frame 104.1 , a histogram 516 is generated. In the generated histogram 516, brightness value is plotted on the x-axis, and on the y-axis the number of pixels having a respective brightness value is represented. The histogram can be represented by a vector 518, which corresponds to another example of a lower level representation. This process can be regarded as feature extraction. To the vectors 518 of brightness values of a corresponding set of image frames, clustering is applied. The shown diagram 519 is simplified having two axes. If the brightness vector 518 is assigned to a special cluster 520, an event trigger is issued 112.

[0090] In particular, the use of image data allows to detect visible events in the chemical reaction directly without prior knowledge of the event. This provides a simple and effective way to BASF SE 220995 retrospectively monitor or assess the chemical reaction, and / or to control the chemical reaction by reacting to certain detected events.

[0091] The summary video file can be used for documenting the chemical reactions, for comparing different runs of the same chemical reactions, for visual assessment of the reactions, for instance for cause studies, etc. The summary video files can also be used for presentations and as appendix for knowledge transfer or process development. Having a faster way to assess a chemical reaction may also increase trust of a production team towards nonsupervised chemical processes, which may in turn increase the production yield by increasing the number of runs beyond normal production-hours.

[0092] In summary, the invention is directed to a method and apparatus for monitoring a chemical production process, for instance by detecting one or more events associated with a chemical reaction during the chemical production process. The apparatus is configured to receive a sequence of image data representing reaction conditions of the chemical reaction, to map, at a mapping unit, one or more image frames of the sequence of image data to at least one lower level representation, to generate an event trigger based on a change in the at least one lower level representation that is associated to a respective change in the reaction conditions of the chemical production process and to provide the event trigger in association with the sequence of image data for monitoring the chemical production process. The method and apparatus enable an improved assessment of the chemical production process.

[0093] For the processes and methods disclosed herein, the operations performed in the processes and methods may be implemented in differing order. Furthermore, the outlined operations are only provided as examples, and some of the operations may be optional, combined into fewer steps and operations, supplemented with further operations, or expanded into additional operations without detracting from the essence of the disclosed embodiments.

[0094] In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality.

[0095] A single unit or device may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. BASF SE 220995

[0096] Procedure steps performed by one or several units or devices can be performed by any other number of units or devices. These procedures can be implemented as program code means of a computer program and / or as dedicated hardware.

[0097] A computer program product may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium, supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

[0098] Any units described herein may be processing units that are part of a classical computing system. Processing units may include a general-purpose processor and may also include a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any other specialized circuit. Any memory may be a physical system memory, which may be volatile, non-volatile, or some combination of the two. The term “memory” may include any computer-readable storage media such as a non-volatile mass storage. If the computing system is distributed, the processing and / or memory capability may be distributed as well. The computing system may include multiple structures as “executable components”. The term “executable component” is a structure well understood in the field of computing as being a structure that can be software, hardware, or a combination thereof.

[0099] For instance, when implemented in software, one of ordinary skill in the art would understand that the structure of an executable component may include software objects, routines, methods, and so forth, that may be executed on the computing system. This may include both an executable component in the heap of a computing system, or on computer- readable storage media. The structure of the executable component may exist on a computer-readable medium such that, when interpreted by one or more processors of a computing system, e.g., by a processor thread, the computing system is caused to perform a function. Such structure may be computer readable directly by the processors, for instance, as is the case if the executable component were binary, or it may be structured to be interpretable and / or compiled, for instance, whether in a single stage or in multiple stages, so as to generate such binary that is directly interpretable by the processors.

[0100] In other instances, structures may be hard coded or hard wired logic gates, that are implemented exclusively or near-exclusively in hardware, such as within a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any other specialized circuit. Accordingly, the term “executable component” is a term for a structure that is well understood by those of ordinary skill in the art of computing, whether implemented in software, hardware, or a combination. Any embodiments herein are described with reference BASF SE 220995 to acts that are performed by one or more processing units of the computing system. If such acts are implemented in software, one or more processors direct the operation of the computing system in response to having executed computer-executable instructions that constitute an executable component. Computing system may also contain communication channels that allow the computing system to communicate with other computing systems over, for example, network.

[0101] A “network” is defined as one or more data links that enable the transport of electronic data between computing systems and / or modules and / or other electronic devices. When information is transferred or provided over a network or another communications connection, for example, either hardwired, wireless, or a combination of hardwired or wireless, to a computing system, the computing system properly views the connection as a transmission medium. Transmission media can include a network and / or data links which can be used to carry desired program code means in the form of computer-executable instructions or data structures and which can be accessed by a general-purpose or special-purpose computing system or combinations. While not all computing systems require a user interface, in some embodiments, the computing system includes a user interface system for use in interfacing with a user. User interfaces act as input or output mechanism to users for instance via displays.

[0102] Those skilled in the art will appreciate that at least parts of the invention may be practiced in network computing environments with many types of computing system configurations, including, personal computers, desktop computers, laptop computers, message processors, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, mobile telephones, PDAs, pagers, routers, switches, datacenters, wearables, such as glasses, and the like. The invention may also be practiced in distributed system environments where local and remote computing system, which are linked, for example, either by hardwired data links, wireless data links, or by a combination of hardwired and wireless data links, through a network, both perform tasks. In a distributed system environment, program modules may be located in both local and remote memory storage devices.

[0103] Those skilled in the art will also appreciate that at least parts of the invention may be practiced in a cloud computing environment. Cloud computing environments may be distributed, although this is not required. When distributed, cloud computing environments may be distributed internationally within an organization and / or have components possessed across multiple organizations. In this description and the following claims, “cloud computing” is | BASF SE | 220995 | 220995WQ01 ~ defined as a model for enabling on-demand network access to a shared pool of configurable computing resources, e.g., networks, servers, storage, applications, and services. The definition of “cloud computing” is not limited to any of the other numerous advantages that can be obtained from such a model when deployed. The computing systems of the figures include various components or functional blocks that may implement the various embodiments disclosed herein as explained. The various components or functional blocks may be implemented on a local computing system or may be implemented on a distributed computing system that includes elements resident in the cloud or that implement aspects of cloud computing. The various components or functional blocks may be implemented as software, hardware, or a combination of software and hardware. The computing systems shown in the figures may include more or less than the components illustrated in the figures and some of the components may be combined as circumstances warrant.

[0104] Any reference signs in the claims should not be construed as limiting the scope.

Claims

BASF SE220995Claims:1 . Method (500) for monitoring a chemical production process, the method comprising the steps of:- receiving (502) at least one sequence of image data (104) representing reaction conditions of the chemical production process,- mapping (504) one or more image frames (104.1) of the at least one sequence of image data to at least one lower level representation (P, L),- generating (506) an event trigger (112) based on a change (AP) in the at least one lower level representation that is associated with a respective change in the reaction conditions of the chemical production process,- providing (508) the event trigger in association with the sequence of image data for monitoring chemical production process, wherein generating the event trigger based on the change in the at least one lower level representation that is associated with a respective change in the reaction conditions of the chemical production process includes:- recording the sequence of image data;- generating, based on a time step of a respective event trigger, an event flag linked to the corresponding time step of the sequence of image data;- compressing the sequence of image data based on the event flag; and- providing a time-compressed sequence of image data, timely compressed with respect to the duration of the chemical production process.

2. The method of claim 1 , wherein generating the event trigger based on the change in the at least one lower level representation that is associated with a respective change in the reaction conditions of the chemical production process further includes:BASF SE220995- classifying event triggers to pre-defined event specifiers associated with a corresponding detectable event; generating, based on the time step of event trigger, the event specifier linked to the corresponding time step of sequence of image data; compressing the sequence of image data based on the event flag and providing a compressed sequence of image data, and / or- classifying event trigger to pre-defined event specifiers associated with a corresponding detectable event; and generating a control trigger (120) to control the chemical production process.

3. Method (500) according to claim 1 or 2, wherein mapping (504) one or more image frames of the sequence of image data to at least one lower level representation includes- providing (504a) the one or more image frames of the sequence of image data to a feature extraction model configured to map the image frame to at least one lower level representation, and- generating (504b) the at least one lower level representation.

4. Method (500) according to claim 3, further comprising,- providing the at least one level lower level representation for performing an anomaly detection method and / or a clustering method, wherein performing the anomaly detection method comprises identifying, according to a predetermined decision algorithm, lower level representation instances that differ in a significantly relevant manner from a majority of the lower level representation instances, and generating the event trigger in association with the identified lower level representation instances; and wherein performing the clustering method comprises associating each lower level representation instance to one cluster of a set of clusters in accordance with a predetermined clustering algorithm and generating the event trigger in association with one or more target clusters from the set of clusters.BASF SE2209955. Method according to any of claims 1 to 4, wherein mapping one or more image frames of the sequence of image data to the at least one lower level representation includes:- providing (504c) an initial part (104a) of the sequence of image data (104) to train a data-driven model,- providing a monitoring part of the sequence of image data to the trained data-driven model, wherein the monitoring part covers a time period following the time period of the initial part, and- generating (505d) the at least one lower level representation.

6. Method (500) according to claim 5, wherein generating the at least one lower level representation includes determining, as a lower level representation, a loss value by a loss function (L) depending on the input image data (y) provided to the data-driven model (510) and on output image data (y’) provided by the data-driven model (510).

7. Method (500) according to any one of the preceding claims, wherein the sequence of image data is provided in form of a video stream or as a video file.

8. Method according to any one of the preceding claims, wherein the sequence of image data is recorded during the chemical production process, wherein in case of a continuous production process the recording monitors the production process continuously or in case of a batch production process the recording monitors the production process per batch.

9. Method (500) according to claim 8, wherein the recording is started on start of the chemical process, such as providing of chemical educts to a chemical reactor, and stopped on stop of the chemical process, such as providing of chemical products as result of the chemical reaction.

10. Method (500) according to any one of the preceding claims, further comprising- receiving, in addition to the sequence of image data, time-dependent sensor data measured and provided by one or more sensor units, the time-dependent sensor data being indicative of values of a process parameter of the chemical production processBASF SE220995- fusing the image frames of the at least one sequence of image data with the additional time-dependent sensor data for a corresponding time step and generating at least one lower level representation based on fused data, and / or- mapping the additional time-dependent sensor data to at least one lower level representation, generating the event trigger based on a change in the at least one lower level representation that is associated with a respective change in the reaction conditions of the chemical production process and / or associated with a respective change in the values of the process parameters of the chemical production process, and providing the event trigger in association with the sequence of image data and / or sensor data for monitoring the chemical production.11 . Apparatus (100) for monitoring a chemical production process, the apparatus being configured to perform a method according to any one of the preceding claims.

12. Chemical equipment (200) including at least one chemical reactor (202) and an apparatus (100) according to claim 11 .

13. Use of event trigger in association with a sequence of image data, in particular a recorded sequence of image data and / or time-compressed sequence of image data, the event trigger as provided according to the method according to any of claims 1 to 10, for retrospective assessment of the chemical production process.

14. Computer program comprising instructions, which, when executed by an apparatus according to claim 11 , cause the apparatus of claim 11 to carry out the steps of any of the methods of claims 1 to 10.

Images