Methods and systems for operating a chemical plant

Abstract

A method for operating a chemical plant (1) implemented to perform a batch process (BP) for producing a chemical product, wherein the batch process (BP) comprises a number of process steps (P1~PN), and the method includes receiving (S1) sensor data (SDQ) associated with at least one process step (PQ) of the batch process (BP), and processing the sensor data (SDQ) according to a data-driven model (DDCM) (S2) to generate a compressed digital representation (SDQ') of the sensor data (SDQ), The method includes the steps of: determining a quality measure indicating the physicochemical quality of the chemical product produced by the plant (1) as a function of the digital representation (SDQ') of the received sensor data (SDQ), and / or determining a performance index of the batch process as a function of the digital representation (SDQ') of the received sensor data (SDQ) (S3); and generating plant operation data (CD) indicating that the chemical plant (1) is in operation as a function of the determined quality measure (S4). The system (1) for producing chemical products comprises a control unit (3) associated with batch process steps (P1~PN) and controllable plant units (6, 7, 10), wherein the control unit (3) is implemented to perform the method.

Description

Technical Field

[0001] Technical Field The present disclosure particularly relates to methods and systems for operating a chemical plant or system implemented to execute a batch process. Aspects relate to batch processes for monitoring and evaluating the performance of a batch process, or the quality or product produced by a batch process. The present disclosure also relates to computer-implemented methods, computer programs, and computer-readable media embodying a method for controlling a chemical plant.

Background Art

[0002] Background Art In a batch process or batch plant, the production of multiple products is carried out using the same set of equipment or processing units, such as chemical reactors or biological reactors. In a batch process, a series of batch process steps are executed, each of which uses a processing unit or processing action according to the batch process specifications. Each processing step affects the next processing step, thereby affecting potential key performance parameters / quality. Therefore, it is very difficult at the initial stage of a batch process, that is, after only a part of the entire batch process has been executed.

Summary of the Invention

Problems to be Solved by the Invention

[0003] In some cases, batch processes are only inspected and evaluated at the laboratory scale. Before building a large-scale chemical plant, it is then desirable to predict how the industrial implementation of each batch process will function. Therefore, an object of the present disclosure is to provide a method and system for efficiently controlling or designing a chemical plant.

Means for Solving the Problems

[0004] Summary The subject matter of the attached claims addresses this objective.

[0005] Accordingly, one aspect of the present disclosure provides a method for operating a (chemical) plant implemented to perform a batch process for producing a (chemical) product, wherein the batch process comprises a number of process steps, and the method Receiving sensor data associated with at least one process step of a batch process, To generate a compressed digital representation of sensor data associated with at least one process step, the sensor data is processed according to a data-driven model, To determine a quality measure and / or performance indicator of a batch process that indicates the physicochemical quality of the (chemical) products produced by the plant, as a function of the digital representation of the received sensor data, To generate plant operation data indicating that the (chemical) plant is operating based on the determined quality measures and / or determined performance indicators, and to output the plant operation data. Includes.

[0006] Embodiment One example provides a method for operating a chemical plant implemented to perform a batch process for producing chemical products. The batch process comprises a number of (preferably sequential) process steps. The method in this embodiment is Receiving sensor data associated with at least one process step of a batch process, To generate a compressed digital representation of sensor data associated with at least one process step, the sensor data is processed according to a data-driven model (DDCM), To determine a quality measure indicating the physicochemical quality of chemical products produced by the plant and / or a performance indicator of the batch process as a function of the digital representation of the received sensor data, (For example, generating plant operation data that shows the operation of a chemical plant as a function of a determined quality measure) and outputting the plant operation data. Includes.

[0007] The embodiments presented above and below may enable reliable process control in the early stages of the process, resulting in reliable and cost-effective production, reduced waste production (e.g., due to fewer defective products), fewer resources required or used, and energy and / or time savings.

[0008] In the embodiment, "operating a chemical plant" should be interpreted as monitoring and / or controlling a (chemical) plant.

[0009] In particular, determining a function of a digital representation involves decompressing the digital representation and / or reconstructing the digital representation of at least one batch process or process step.

[0010] The generated plant operation data may include information indicating that the (e.g., chemical) plant is operational and facilitating the execution of batch processes using the (chemical) plant. In embodiments, the plant operation data includes control data and / or monitoring data of the (chemical) plant. The plant operation data may be obtained indicating that the (chemical) plant is operational and based on determined quality metrics and / or determined performance metrics. For example, a warning may be triggered if the determined quality metrics are above or below a threshold. The plant operation data may also be determined by processing sensor data and determined quality metrics (and / or determined performance metrics) according to a data-driven model to generate a compressed digital representation of the sensor data and determined quality metrics (and / or determined performance metrics), and by determining (further) batch process parameters to adjust a batch process or further (not yet executed) batch process steps, for example, based on the compressed digital representation of the sensor data and determined quality metrics (and / or determined performance metrics) in a manner similar to how the quality metrics and / or performance metrics were obtained. Such (further) batch process parameters may be or be part of the plant operation data. For example, this could allow a specific batch process step to be reconstructed from knowledge of other batch process steps, and as a result, determined or generated target variables, such as quality metrics, could be fed back into a data-driven model, which is encoded, for example, along with sensor data, to determine how (further) batch processes can be adapted, for example, through decoding. For example, if it is determined that (future or further) batch process parameters, such as temperature, should be lower than those previously set for each batch process step, the plant operation data could include an action command to lower each batch process parameter, for example, by lowering the heat by a specific temperature. The plant operation data could be output to the plant operator or operator, for example, through a user interface, so that the operator or operator can take appropriate action.However, the plant operation data can also be output to a process to automatically operate the plant (based on the plant operation data), for example, to adjust batch process parameters for further batch steps. The acquired plant operation data can be stored in a (non-temporary and / or tangible) computer-readable storage medium.

[0011] The methods and systems included enable the determination or prediction of the performance of products manufactured in a batch process or in the early stages of a batch process, i.e., after only a portion of the entire batch process has been executed. This is due to a flexible digital representation based on data-driven models that include information on the interactions and interdependencies of process steps and / or intermediate products or raw materials in a batch process, for example, through machine learning history.

[0012] Additionally or alternatively, the determination may include predicting, calculating, calculating, or determining a performance metric for a subset of multiple sequential batch process steps, in particular, a quality measure associated with one of the batch process steps, or an intermediate product, for example, a quality measure associated with one process step. It is understood that the batch process steps relating to this disclosure include the operation or use of a processing unit and may be characterized by a set of characterizing data (which may be batch process step parameters). Examples of data for characterizing the data include, for example, chemical identifiers, geometric shapes, physicochemical properties of a processing unit or medium, analytical properties of chemical pre-products, intermediate products, or final products, dimensions of a processing unit, flow rate, duration, temperature (gradient), or other observations measurable by a technical device, or sensor data relating to parameters that technically define entities involved in each batch process. Performance metrics may include data relating to the operating / operating point or status of a chemical plant by scale or preset status, information relating to other batch processes, equipment, and / or chemical plants. Plant operation data may be used, for example, to extrapolate monitored plant behavior and predict the behavior of another plant.

[0013] In embodiments of the method, at least one of the following is performed: operating a chemical plant as a function of control data, and / or comparing a determined quality measure and / or performance indicator with a target quality or performance.

[0014] By deploying a data-driven model to generate a compressed digital representation of sensor data that can be considered to parameterize the measurable characteristics of each batch process step, or the intermediate or final products of a batch process, it enables efficient evaluation of the quality and performance of a controlled chemical plant.

[0015] In a compressed digital representation of sensor data, data processing can be performed in a more efficient manner than using conventional simulation algorithms. Digital representations enable continuous and efficient monitoring of batch processes. Connecting and processing sensor data according to a data-driven model allows for the design of a real chemical plant suitable for running batch processes before conducting large-scale experiments using actual plant components.

[0016] A data-driven model (DDCM) can be configured to capture nonlinear relationships between batch process parameters, such as between batch process parameters characterizing initial resources like substrates and batch process quality measures or performance indicators. The data-driven model may be, or may include, an encoder of a multimodal variational autoencoder (MVAE), which may include an encoder of a variational autoencoder, e.g., a (sub)encoder for each mode. A data-driven model configured to capture nonlinear relationships can enable the modeling of complex tasks such as upscaling production from laboratory scale to production scale, or setting up production in different plants with different conditions, which is not possible with simpler linear models such as regression-based models, while simultaneously saving resources compared to, for example, complex simulations. Furthermore, using an autoencoder encoder can result in a compressed digital representation, such as a latent space. For example, if the data-driven model is an autoencoder encoder, the data-driven model can generate a compressed digital representation that exhibits nonlinear relationships. The quality measure can then be determined, for example, by using an autoencoder decoder as, for example, a second or further data-driven model.

[0017] The (first) data-driven model and the second data-driven model may be part of the same overall architecture, for example, the (first) data-driven model may be the encoder of a variable autoencoder, and the second data-driven model may be the decoder of a variable autoencoder. The data-driven model may be generative in that it may be configured to generate or reconstruct (new) samples of data (e.g., batch process parameters) from a compressed digital representation of the data. Alternatively or additionally, the data-driven model may be or include a general-purpose network (GAN) or a multimodal GAN.

[0018] For example, when transitioning from a laboratory setting to a production environment, various factors and variables can interact in complex and nonlinear ways, which can be captured in a compressed digital representation when a data-driven model is trained using data from both environments. As an example, in a laboratory setting, experiments may be conducted under controlled conditions that may not exist in an upscaled production environment. Changes may exist in equipment, operating conditions, material properties, and process dynamics.

[0019] A (variational) autoencoder may include at least one encoder, which may include a neural network, such as a feedforward neural network. The at least one encoder may encode input data, such as sensor data associated with at least one process step of a batch process, into a compressed digital representation. To encode the sensor data, batch process parameters that do not correspond to the sensor data, i.e., are not measured or available, as well as target variables such as quality measures and / or performance metrics, may be set to zero, thereby allowing the encoder to treat them as missing. The compressed digital representation of the sensor data may have the form of a probability distribution, e.g., a Gaussian distribution with mean and standard deviation. The probability distribution in the compressed digital representation may be sampled, thereby allowing output data, such as quality measures and / or performance metrics, to be generated, for example, by using at least one decoder of the autoencoder. A variable autoencoder may include at least one decoder, which may include a neural network, such as a feedforward neural network. By using an autoencoder, it may be possible to use a relatively small training dataset compared to other generative models. For example, good predictive performance can be achieved using a training dataset of 1500-2000 data points, which is comparable to the data required for regression.

[0020] A multimodal variational autoencoder (MVAE) can enable the processing of input from multiple different data modalities, for example, each modality may represent a step in a batch process. An MVAE may include at least one encoder for each modality, such as a feedforward neural network. Note that the encoders for modalities may collectively be referred to as the encoder or encoder portion of the MVAE in this specification. The encoders of a multimodal variational autoencoder (MVAE) can encode or transform input data, such as sensor data, into a compressed digital representation, such as a latent space (having lower dimensions, i.e., lower dimensions than the sensor data space). As described, the encoders of an MVAE may include several sub-encoders, each designed to process a different data modality. For example, in the case of a batch process, there may be one (sub)encoder for each batch process step. Each sub-encoder may be a neural network that determines mean and standard deviation vectors that can define, for example, a (multivariate) Gaussian distribution in a compressed digital representation or latent space, based on input data from its respective modality, thereby enabling the MVAE to generate (new) samples of the data. The (sub)encoders may process the input data from their respective modalities, and the encoders of the MVAE may combine the processed input data from each modality to generate a (shared) compressed digital representation or latent space, which the decoder may use to generate a new instance of the data. The combination may, for example, be by calculating the product of the individual distributions obtained by processing the input data from each modality by the (sub)encoders. The decoder of the MVAE may also include a (sub)decoder that can be configured to generate data for each modality, for example, a particular batch process step. The decoder of the MVAE may sample from the compressed digital representation. The (sub)decoder may take a sample or point from the compressed digital representation to generate output data, for example, a reconstruction of the input data in each modality.By decoding the compressed digital representation into different modalities, the MVAE may be able to generate consistent output data across all modalities. The MVAE may also be used for cross-modal generation tasks, which may enable generating data in one modality from given data in another modality. By generating (shared) compressed digital representations of different modalities, the MVAE may utilize information from all modalities to improve its performance. It may also handle scenarios where some modalities are missing by using information from other modalities to generate the output data of the missing modalities. A modality may also be referred to as a mode herein.

[0021] During the decoding phase, the MVAE may decode the (shared) digital representation back into any of the given modalities. This may enable output data in one modality to be generated using input data from another modality.

[0022] Plant operation data may be determined by including, based on the compressed digital representation, for example, generating batch process step parameters using a decoder of an autoencoder, and using the plant operation data as instructions for performing, or adjusting, or adapting a batch process or respective batch process steps for the plant or another plant.

[0023] Batch process step parameters can be, or can indicate, the chemical and physical properties of the process, such as the temperature or temperature gradient at which the corresponding batch process step is carried out, the flow rate, the rate at which the batch process step or part of the batch process step is carried out, for example the period used for the batch process step, the physico-chemical properties of the processing unit or medium used, such as the geometric shape or dimensions of the processing unit, or part of the processing unit carrying out the batch process step, or other observable values that can be measured by appropriate sensor data or analytically derived from the sensor data. Batch process step parameters can also be, or can indicate, the properties of the initial resources, pre-products, intermediate products, or final products of the batch process step (which may correspond to quality measures or performance indicators), such as the identifiers of the chemical substances, the analytical properties of the chemical pre-products, intermediate products, or final products. Such measurable batch process step parameters can correspond to sensor data or can be associated with sensor data. Note that target variables such as quality measures and / or performance indicators can be considered as batch process step parameters associated, for example, with the last step of the batch process (if at least measurable using sensors). In that case, for example, (i) the quality measure indicating the physico-chemical quality of the (chemical) product produced by the plant and / or the performance indicator of the batch process, and (ii) the compressed digital representation of the batch process step parameters of at least two process steps including at least one process step, can be the compressed digital representation of the batch process step parameters of at least two process steps including at least one (measured) process step, and of the process step including the quality measure and / or performance indicator.

[0024] The compressed digital representation may be, or contain, the latent space of an autoencoder or GAN. It may capture a nonlinear relationship between training data, such as sensor data, and the quality of the products produced by the plant. Determining or generating the compressed digital representation may be trained using training data from the plant or laboratory, which includes both process step parameters and arbitrary target variables, such as quality measures or performance metrics. Then, if measurements are performed using, for example, plant sensors or sensors associated with the part of the plant where a particular batch process step is performed, the measurement data or sensor data may be processed (e.g., compressed or encoded) into a compressed representation of the (measured) sensor data. The compressed digital representation may then be used to determine (e.g., generate) data that is not part of the (measured) sensor data, such as process parameters for other or future batch process steps, or target variables such as quality measures or performance metrics of the products produced. These may be output to, for example, a plant operator, who may take appropriate action, for example, to adjust the production process. By utilizing such a compressed digital representation, it may be possible to capture the relationships between batch process parameters of the entire batch process (essentially) and any target variables, or other target variables such as location-specific variables or changes. Thus, the compressed digital representation can be a compressed digital representation of the complete production process carried out in the plant. Consequently, omitted variable biases can be reduced or avoided (as they may be present in projection techniques, for example), enabling higher predictive accuracy even in the early stages of a batch production process, and reliable results can be obtained even after just one batch process step that allows for early adjustment of the process. As a result, for example, more reliable and cost-effective production can be achieved, waste production can be reduced (for example, due to fewer defective products), fewer resources may be required or used, and energy and / or time can be saved.Using a compressed digital representation of sensor data can enable a highly reliable and efficient production process. Furthermore, it may be possible to reconstruct specific batch process steps from knowledge of other batch process steps, and as a result, determined or generated target variables, such as quality metrics, can be fed back into a data-driven model, which is then encoded, for example, along with sensor data, via decoding, to determine how (further) batch processes can be adapted.

[0025] The compressed digital representation can be understood as generative, for example, containing a probability distribution of the input data from which (new) instances of the data can be sampled or generated. Probability distributions by different variables can be combined by expert products; in the case of MVAE, the encodings of each mode μ_i and σ_i can be considered experts. Sampling or generation can be non-deterministic. The compressed digital representation can be a latent space, which can be configured to capture the structure of the input data. New data can be sampled or generated from the latent space by converting the samples taken from the latent space back into the (higher-dimensional) input data space, regardless of the process step. The input data for each process step can be reconstructed or generated from the samples.

[0026] According to an exemplary embodiment, each process step is associated with a set of batch process step parameters, and the sensor data corresponds to at least one batch process parameter associated with at least one process step, and processing the sensor data (SDQ) according to a data-driven model (DDCM) to generate a compressed digital representation (SDQ') (S2) is as follows: S2) Using a data-driven model (DDCM), obtain a compressed digital representation of sensor data (SDQ) associated with at least one process step (PQ), wherein the data-driven model is trained to generate compressed digital representations of (i) quality measures and / or batch process (BP) performance indicators indicating the physicochemical quality of the (chemical) products produced by the plant (1), and (ii) batch process step parameters of at least two (preferably all) process steps, including at least one process step. It is, or includes, that thing.

[0027] According to an exemplary embodiment, determining a quality measure and / or performance index (S3) is or includes generating a quality measure and / or performance index based on a compressed digital representation using a second data-driven model, the second data-driven model being trained to generate a quality measure and / or performance index from a compressed digital representation of (i) a quality measure and / or a batch process (BP) performance index indicating the physicochemical quality of the (chemical) product produced by the plant (1), and (ii) batch process step parameters of at least two (preferably all) process steps including at least one process step, and determining plant operation data (CD) (S4) is based on the generated quality measure and / or generated performance index.

[0028] According to an exemplary embodiment, a data-driven model includes an encoder of an autoencoder, and a second data-driven model includes a decoder of an autoencoder, wherein the encoder is trained to encode (i) a quality measure and / or a performance indicator of a batch process (BP) indicating the physicochemical quality of the (chemical) product produced by plant (1), and (ii) batch process step parameters for at least two process steps including at least one process step into a compressed digital representation, and the decoder is trained to decode (i) a quality measure and / or a performance indicator of a batch process (BP) indicating the physicochemical quality of the (chemical) product produced by plant (1), and (ii) batch process step parameters for at least two process steps including at least one process step from the compressed digital representation.

[0029] According to exemplary embodiments, the data-driven model is multimodal, and each process step of a batch process corresponds to a mode of the multimodal data-driven model, in particular, at least one mode corresponding to a quality measure and / or performance indicator of the batch process that indicates the physicochemical quality of the (chemical) product produced by the plant. The at least one mode corresponding to the quality measure and / or performance indicator of the batch process that indicates the physicochemical quality of the (chemical) product produced by the plant may be a mode corresponding to a batch process step, in particular the last step of the batch process, or may be a further mode specifically configured for the quality measure and / or performance indicator. Having specific modes for the quality measure and / or performance indicator may be advantageous in that it can increase adaptability, for example, when the performance indicator (e.g., quantifiable metrics such as key performance indicators (KPIs)) changes (e.g., yield, quality rate, cycle time, downtime, waste), and the accuracy of generating the predicted quality measure and / or performance indicator may be improved, for example, by associating a dedicated (sub)decoder with decoding the specific mode.

[0030] According to an exemplary embodiment, the autoencoder is a multimodal variational autoencoder, where each process step of the batch process corresponds to a mode of the multimodal variational autoencoder (MVAE), and at least one further mode corresponds to a quality measure and / or a performance indicator of the batch process (BP) that indicates the physicochemical quality of the (chemical) product produced by the plant.

[0031] By utilizing multimodal approaches such as MVAE or multimodal GANs, it may be possible to precisely map or model the physical structure of a batch process, and therefore, if, for example, a new production step needs to be included to enhance the production process, it may be possible to adapt the batch process fairly easily.

[0032] Furthermore, this may enable the reconstruction of a specific batch process step from knowledge of other batch process steps, and as a result, determined or generated target variables, such as quality metrics, may be fed back into the model, for example, encoded together with sensor data, to determine how the batch process can be adapted.

[0033] According to one exemplary embodiment, at least one further mode corresponds to the location of the plant or the characteristics of the plant.

[0034] This can further enhance the production process and expand its flexibility, for example, by expanding the applicability for decoding using data from a plant or laboratory in one location, and a decoder trained on only coarse training data for a plant, or process parameters of a plant in another location, i.e., making production conversion from one plant to another or upscaling from laboratory to production easier and less expensive.

[0035] According to an exemplary embodiment, the data-driven model is further trained to generate compressed digital representations of (i) quality measures indicating the physicochemical quality of (chemical) products produced by plant (1) as produced by another plant, and / or performance metrics of batch processes (BPs) implemented on the other plant, and (ii) batch process step parameters of at least two process steps implemented on the other plant, including at least one process step. The second data-driven model is trained to generate quality measures and / or performance indicators from compressed digital representations of (i) quality measures indicating the physicochemical quality of (chemical) products produced by plant (1) as produced by other plants, and / or performance indicators of batch processes (BPs) implemented on other plants, and (ii) batch process step parameters of at least two process steps implemented on other plants, including at least one process step.

[0036] This may enable training data-driven models primarily on a single plant, such as a laboratory or research-scale plant, for example, training, fine-tuning, or pre-tuning a data-driven model with less data from a plant producing (chemical) products. Thus, a data-driven model trained on one plant can be transferred for use in another plant, thereby improving productivity in the early stages of plant operation.

[0037] According to one exemplary embodiment, a batch process carried out on plant (1) has at least one additional process step, and the batch process is then carried out on another plant, and the method is - Using a second data-driven model, determine at least one process parameter for at least one additional process step, - To provide a user, such as a plant operator, with at least one process parameter via a (graphical) user interface, and to provide plant operation data that may include at least one process parameter. It also includes.

[0038] According to one aspect of this method, the data-driven model includes a multimodal variational autoencoder (MVAE), which is implemented to receive input mode data and output reconstructed mode data. The multimodal variational autoencoder enables the processing of data from multiple modalities and the combination of them into a single digital representation. This digital representation is also called a latent space representation. An alternative model, for example, a transformer model, can be conceived as a DDCM.

[0039] The compressed digital representation of sensor data in latent space requires less data for a complete description that includes all measurable characteristics and features of the various batch process steps according to the batch process specifications and acquired sensor data. Therefore, using MVAE reduces the amount of computational and memory resources required to process and control a (chemical) plant. A suitable MVAE encoder can be considered a data-driven (compressed) model, as described above. An MVAE decoder can be considered to produce a second or data-driven model.

[0040] Preferably, the DDCM is implemented to receive input mode data and output reconstructed mode data. With respect to MVAE modes, the data refers to a modality, and each modality or mode has an associated dataset format. For example, a set of sensor data used to characterize the raw materials used in a batch process may be considered one modality. Another modality may be, for example, a specification of a batch process step relating to pressure and temperature values ​​during the execution of the batch process step. Various types of modalities can be conceived in relation to a batch process step. In embodiments, each mode data for a modality is given as a set of variables in a predetermined format for each mode / modality. Thus, the input mode data may correspond to a sensor dataset. The format may be a vector, a tensor, or other computer-readable and processable data structure. Modalities may correspond to physicochemical parameters.

[0041] In one embodiment of this method, each batch process step of a batch process is associated with one respective mode or modality. The input mode data for each mode then includes sensor data associated with at least one batch process step associated with that mode. The output mode data for each mode then includes reconstructed sensor data associated with at least one batch process step associated with that mode.

[0042] In an MVAE environment, several modalities are each received in the feature space as separate mode data or sensor datasets, and output as reconstructed modes, where information loss can occur. The latter is preferably reduced and minimized when setting up the MVAE. The MVAE provides a latent space representation or compressed digital representation of the sensor datasets received from various process steps.

[0043] In the embodiment, the DDCM has a predetermined number of modalities corresponding to the number of batch process steps. The method then further includes inputting a first set of input mode data of modes including a first number of modes, and outputting a second set of reconstructed mode data of modes including a second number of modes. The first number of modes is preferably fewer than the second number of modes.

[0044] If more mode outputs exist in the reconstruction space than the mode data input to the feature space, the additional modes with respect to the number of input modalities can be considered predictions or forecasts of a particular mode. In particular, if a batch process includes a series of batch process steps, the DDCM used is implemented to process the mode datasets of all modalities, i.e., all batch process steps. For example, to predict the "target mode," which is the result of a later process step, sensor data or mode data related only to process steps executed before the target mode or process step can be input. Thus, this method makes it possible to predict the quality or performance of the batch process implemented with respect to feature sensor data, and therefore the controlled system. As a result, control data for future process steps can be fitted as a function of the received input mode data.

[0045] In embodiments of the method, the second set of modes includes at least one mode associated with a batch process step that is performed later than the batch process step associated with a mode in the first set of modes.

[0046] In embodiments, the method includes training a DDCM using a training dataset generated by monitoring multiple parameters of a batch process to generate sensor data representing each batch process step. The MVAE may be implemented including an artificial neural network trained using the training dataset. For example, large-scale or laboratory experiments representing batch process steps may be used to generate the training dataset. For example, variations of the batch process include laboratory setups for batch process systems having different chemical plants with different production capacities for the chemical products produced, different plant sizes, different locations, or parameters characterizing raw materials from different suppliers.

[0047] This method may include measuring batch process parameters that indicate the physicochemical properties of raw materials with respect to the characteristics of sensor data and / or batch process steps, and clustering the sensor data into clusters, where each cluster is assigned to one batch process step and / or MVAE / DDCM mode. A cluster may include a set of sensor data in a predetermined format. Processing batch process steps as modes or modalities with respect to MVAE as a data-driven compression model has the advantage that modes or process steps not available as input mode data in the feature space can be reconstructed in the reconstruction space. MVAE provides a complete digital representation in latent space for the entire sensor data front-loaded with respect to modes.

[0048] In embodiments of the method, MVAE includes multiple artificial neural networks (ANNs) as encoders for receiving mode input data in feature space. Each encoder is assigned to one input mode, and at least one encoder is configured as a feedforward network, a convolutional network, or a long-term short-term neural network. Different ANN types can be used for different modalities, thereby improving the loss function of the MVAE.

[0049] In embodiments of the method, the DDCM has at least one further modality indicating the location and / or size of the chemical plant. For example, using the location of the chemical plant as a modality makes it possible to consider the influence of location on the production of chemical products. For example, even if the batch process specifications remain the same or similar, a plant in Europe may require a different configuration than a plant in the United States. Location may be, for example, sensor data.

[0050] In embodiments, the method includes predicting a quality measure and / or performance indicator for a selected batch process step based on sensor data associated with another process step in the batch process, the other process step preceding the selected process step. Preferably, the method includes generating control data as a function of the predicted quality measure and / or performance indicator. Prediction may include reconstructing the sensor data associated with the selected batch process step.

[0051] This method makes it possible to provide early indicators of quality problems when a batch process is initiated. The batch process can then be adapted or controlled to improve quality or performance. This improves the yield and output of each plant. Meanwhile, resources are saved and waste is reduced by stopping defective batch processes. The presented method also makes it possible to scale up batch process setups, for example, from laboratory or experimental scale to large-scale industrial plants. This is because scale can be used as a modality, so that DDCM can output reconstructed modal data associated with another scale with respect to the scale of the input feature space modality data.

[0052] In embodiments of the method, the batch process includes a plurality of sequential process steps, including a first process step and a final process step. Each batch process may be controlled as a function of control parameters implemented by control data. The method preferably includes adapting the control parameters of selected batch process steps as a function of predicted quality measures and / or predicted performance indicators.

[0053] This method enables efficient and reliable control of batch processes, and therefore the production of chemical products.

[0054] In an alternative embodiment, generating control data is replaced by generating an alternative batch process specification as a function of quality measures and / or performance indicators. An embodiment defines a chemical plant design suitable for implementing a batch process. The design may include a list of controls, sensors, and / or processing units and their configurations.

[0055] In another aspect, a system for producing chemical products is disclosed. The system includes a control unit and a controllable plant or processing unit associated with a batch process step. The control unit is implemented to control the plant unit in accordance with the manner or embodiment of the method disclosed above or below with respect to the examples of this disclosure.

[0056] In a further embodiment, a computer-readable medium storing computer program instructions is disclosed, and when the computer program instructions are executed by a control unit and / or processing unit according to the embodiments or models disclosed above or below in particular examples, the control unit and / or processing unit is caused to perform an operation including a method according to the embodiments or models disclosed above or below. The computer-readable medium is, in particular, a non-temporary computer-readable medium.

[0057] In embodiments, the computer program or computer program product includes program code for the computerized control unit to perform the methods and functions described above when executed on at least one computerized unit, particularly when executed on a control unit of a chemical plant. The computer program means, such as the computer program product, may be embodied as a file that can be downloaded from a memory card, USB stick, CD-ROM, DVD, or server on a network. For example, such a file may be provided by transferring the files constituting the computer program product over a wireless communication network.

[0058] Furthermore, a system for producing (chemical) products is provided, comprising a control unit associated with a batch process step and a controllable plant unit, the control unit being implemented to perform the methods described above or below.

[0059] Furthermore, an apparatus is provided that includes means for carrying out or performing the steps of the methods described above or below.

[0060] Furthermore, the provided device includes at least one processor and at least one memory that, when executed by the at least one processor, stores instructions causing the device to perform at least one of the steps of the above or below method.

[0061] Furthermore, plant operation data obtained by the aforementioned or later methods or apparatus is provided.

[0062] Furthermore, a computer program is provided which includes instructions for performing at least one step of the method described above or below.

[0063] Furthermore, possible implementations or alternative solutions of the present invention also encompass combinations of the features described above or below with respect to embodiments (not expressly mentioned herein). Those skilled in the art can also add individual or separate embodiments and features to the most basic forms of the present invention.

[0064] Other features will become apparent from the following detailed description, which will be considered in conjunction with the attached drawings. However, it should be understood that the drawings are designed for illustrative purposes only and not as limit definitions, and for that purpose, the attached claims should be referred to. Furthermore, it should be understood that the drawings are not drawn to scale and are simply intended to conceptually illustrate the structures and procedures described herein.

[0065] Detailed explanation Further embodiments, features, and advantages of this specification will become apparent from the following description and dependent claims, and in conjunction with the accompanying drawings. [Brief explanation of the drawing]

[0066] [Figure 1] This figure shows a first embodiment of a batch process. [Figure 2]This figure shows one embodiment of a chemical plant implemented to perform batch processes. [Figure 3] This figure shows an embodiment of a method for generating a digital representation of a batch process step. [Figure 4] This is a flowchart of method steps involved in an embodiment of a method for controlling a chemical plant. [Figure 5] This figure shows one embodiment of a data-driven compression model applied to a second embodiment of a batch process. [Figure 6] This is a flowchart of the steps involved in MVAE training. [Figure 7] This figure shows an embodiment of an application for MVAE prediction in a batch process. [Figure 8] This is a flowchart of the method steps involved in evaluating an MVAE application. [Figure 9] This figure shows the prediction error of MVAE as a function of the number of modes considered. [Figure 10] This figure shows the prediction error of MVAE as a function of the number of modes considered. [Figure 11] This figure shows an example of an encoder in an autoencoder. [Figure 12] This figure shows an example of a multimodal variational autoencoder. [Modes for carrying out the invention]

[0067] In drawings, similar reference numerals, unless otherwise specified, refer to similar or functionally equivalent elements.

[0068] The following explanation is intended to aid in understanding and complement the explanation provided in the summary section above this specification, and should be read in conjunction with it. Some aspects may have different terminology than those provided in the above explanation, for example. Nevertheless, those skilled in the art will understand that these terms refer to the same subject, for example, by being more specific.

[0069] Figures 1, 2, and 3 show a batch process, a chemical plant implemented to carry out the batch process, and a mechanism for generating control signals.

[0070] Figure 1 shows a first embodiment of a batch process BP. In the batch process BP of Figure 1, multiple batch process steps P1 to PN are performed. Batch process steps P1 to PN refer to a series of operations performed on a processing unit of a chemical plant over a period of time. Each of the batch process steps P1 to PN may be characterized by a batch process step specification. The batch process step specification may include, for example, a set of parameters that identify or characterize the starting product in step P1. Each batch process step specification may be considered, for example, the values ​​required for a measuring sensor or sensor data generated by a sensor that identifies the basic product or starting product. All batch process step specifications together form a batch process specification BPS. The batch process may result in the synthesis of a chemical substance, for example, as a polymer.

[0071] Figure 2 shows a chemical plant relating to System 1 for producing chemical products by a batch process controlled according to the above and below embodiments and models of a method and apparatus for generating control data CD. In this example, System 1 is implemented to synthesize polymers according to synthesis specifications which may be equivalent to the batch process specifications in Figure 1. The plant or System 1 comprises a user interface 2 and a processor 3 associated with a control unit 4, the control unit 4 being configured to receive control data CDs generated according to this disclosure. For example, synthesizing a particular polymer requires batch process steps P1-PN shown in Figure 1. In this example, the control data CDs parameterizing the batch process steps are provided from a database 5, while in other examples, the control data CDs may be provided from a server. The control unit 4 generates commands and / or control signals CS so that the batch process is carried out according to the respective batch process specifications and respective process parameters. In Figure 2, all control signals are labeled CS.

[0072] Containers 6 and 7 each contain components or raw materials of the chemical product to be produced or synthesized. Generally, there may be three or more containers. For illustrative purposes, this example shows only two containers 6 and 7 for storing chemical components. Valves 8 and 9 are associated with containers 6 and 7. Valves 8 and 9 can be controlled to dispense appropriate amounts of each component stored in containers 6 and 7 into reactor 10 as raw materials for synthesizing polymers, according to the synthesis specifications. Motors 11 associated with mixer 12 can also be controlled by control unit 4 as functions of control data CD and control signals CS, respectively. An optional heater 13 can also be controlled by appropriate control signals CS according to the synthesis and / or batch process specification BPS. Control data CD is generated to conform to the batch process specification BPS. Finally, an outlet valve 14, which is in fluid communication with reactor 10, can be controlled by control unit 4 to dispense the chemical product into container 15 or the test system.

[0073] Control data CD is generated to satisfy batch process specifications BPS, but can be adapted when the batch process is transferred to another plant or during process execution to improve, for example, its efficiency or performance. An example focusing on one batch process step PQ is shown in Figure 3. This disclosure provides, in particular, a digital representation of the characteristics of a batch process step. It is understood that each batch process step PQ may be characterized by sensor data SDQ. It is understood that sensor data is not strictly data obtained in a measurement process or from a sensor device. However, in embodiments, sensor data refers to data obtained by a physicochemical measurement process. Sensor data SDQ can be conceived to characterize the batch process step PQ as a combination of temperature, volume, and pressure values ​​in a reactor tank.

[0074] In the following, the set of sensor data characterizing such a batch process step PQ is also referred to as modality or mode data in the context of an artificial intelligence-based autoencoder. The sensor data SDQ is processed according to a data-driven compression model DDCM. This yields a digital representation SDQ' of the received sensor data SDQ. The data-driven compression model is implemented, for example, as an artificial neural network (ANN) with respect to a multimodal variational autoencoder (MVAE). One modality of such an MVAE is associated with the process step PQ and its characterizing sensor data SDQ.

[0075] In Figure 3, the digital representation SDQ plant is a latent space representation of the sensor data SDQ in the model DDCM. The data-driven model outputs a reconstructed sensor data SDQ' that points to the same process step PQ. It is separated into a feature space for the actual sensor data SDQ and a reconstructed or reconstructed space for the reconstructed sensor data SDQ'. As indicated by the dashed arrow, the digital representation is used to generate control data CD and / or modify the batch process specification BPS through the reconstructed sensor data SDQ'. For example, if the batch process step PQ is associated with the operation of heater 13 as shown in Figure 2, temperature data can be used as sensor data SDQ. Generally, sensor data is multidimensional data, not scalar. For example, the temperature field or temperature distribution in reactor 10 can constitute sensor data SDQ. In the latent space or compressed digital representation that yields SDQ', once the control data CD is converted by the control unit 4 into the respective control signals CS for heater 13, specific aspects related to the quality or performance of process step PQ, such as temperature fluctuations in the heater or reactor 10, are determined or calculated. In the example in Figure 3, one modality Q is shown. However, when controlling a chemical plant, all batch process step specifications and batch process steps are preferably taken into account through their associated sensor data when controlling each plant or system.

[0076] Figure 4 shows a flowchart of method steps involved in another embodiment of a method for controlling a chemical plant, for example, the system shown in Figure 2.

[0077] It is understood that the chemical plant is implemented to execute batch processes according to the batch process specification BPS. This can be realized with respect to a series of batch process steps as shown in Figure 1. First, in step S0, sensor data is generated by measuring process parameters of the chemical raw materials, substances, and / or basic or intermediate products of the batch process, and / or retrieving descriptive parameters. All sensor data is clustered into clusters, each cluster corresponding to a specific batch process step P1-PN. Figure 3 shows one cluster of sensor data SDQ associated with batch process step PQ.

[0078] Next, in step S1, the respective sensor datasets associated with each process step are received, for example, by a data processing device. Each process step has associated sensor data, similar to process step PQ shown in Figure 3.

[0079] Each sensor dataset associated with each batch process step is processed through a data-driven compression model in step S2. The data-driven compression model DDCM is implemented as a multimodal variational autoencoder MVAE, as will be described in more detail with respect to the following embodiments. As a result, a compressed digital representation of the sensor data is obtained in latent space. The MVAE and its latent space contain knowledge about the entire batch process, for example, through a pre-training procedure. Thus, this compressed digital representation makes it possible to reconstruct reconstructed sensor data for various modalities or batch process steps, particularly following process steps that have not yet been measured. Figure 3 shows the reconstructed sensor data SDQ' for the Qth modality or mode identified using the batch process step PQ.

[0080] In step S3, the data-driven compression model from step S2 and the received sensor data from step S1 are expanded to calculate a quality measure or performance index for one or more batch process steps. The quality measure may indicate the physicochemical quality of the intermediate products or substances involved in the batch process. For example, the viscosity of the intermediate product may be considered a quality measure. The desired properties of the final product may also be considered a quality measure. Based on the reconstructed sensor data SDQ', key performance indicators (KPIs) can be derived during the batch process. If a deviation from the target KPI occurs, provisional sensor data specifying the desired working point is generated by the data-driven compression model, and it is checked whether it is possible to move the batch process toward the desired working point. Then, corresponding control data is generated as described below. The KPI may also refer to the intermediate products rather than the entire batch process.

[0081] Next, in step S4, control data is generated or existing control data is modified to operate a chemical plant implemented to execute a batch process. The control data CD is generated as a function of a quality measure, such as viscosity. Generating the control data in step S4 may be accompanied by step S41, which sets a target quality measure, i.e., a viscosity value. In step S42, the quality measure is compared with the target quality measure. As a result of steps S1 to S4, the control data is obtained and can be stored in a computer-readable format for the processing unit and / or control unit of the chemical system. The control data is stored in memory, for example, so that the chemical plant operates accordingly, or it is executed directly. Preferably, the target quality measure is then satisfied by a batch process. The control data can be provided, for example, as a cloud service and retrieved from control units in various locations.

[0082] In step S5, the control data CD is used to operate the chemical plant more efficiently. The control data CD may specify improved operating conditions, such as operating point, temperature, pressure, flow rate, or other technical parameters, in accordance with the sensor data received in step S1.

[0083] Next, further embodiments for operating the chemical plant are presented. Figure 5 shows one embodiment of a data-driven compression model applied to another embodiment of batch process BP. Figure 5 shows a simplified batch process BP used to produce a selective catalytic reduction catalyst for a diesel particulate filter. This is shown in the first row of Figure 5. The simplified batch process comprises six process steps P1 to P6. Selective catalytic reduction (SCR) on a filter structure is used to reduce NOx to nitrogen and water and control the exhaust gas of a diesel engine. The SCR on the filter element is obtained in step P1 by providing a substrate with respect to a foam body, the foam body may be a ceramic body as the substrate. In batch process step P2, a slurry is provided, which is an emulsion containing zeolite particles. The body shown in step P1 is then coated in batch process steps P3 and P4 using two different coatings. The analytical properties of the resulting wash coat characterize the firing step P5. As a result, a ceramic-based catalyst body having SCR material is obtained, as shown by batch process step P6, and the SCR on the filter has a specific back pressure when operated and used in the exhaust flow.

[0084] Each batch process step P1 to P6 has an associated set of sensor data SD1 to SD6. For example, the sensor data of the substrate by process step P1 includes a P variable to characterize the ceramic body used. The slurry used in batch process step P2 may be characterized by Q variables, e.g., about 20 variables, which are considered sensor data SD2. The two coating processes P3 and P4 each have R variables for each coat and sensor data SD3 and SD4, respectively. R may be less than Q. The wash coat analysis associated with step P4 includes S, e.g., 10 to 20 variables, as sensor data SD5. Finally, the SCR on the filter is characterized by the back pressure under given conditions, as sensor data SD6.

[0085] Sensor data SD1 to SD6 are each considered modalities M1 to M6 with respect to the multimodal variational encoder MVAE environment. Encoders E1 to E6 each map the modality-specific sensor data sets SD1 to SD6 to a latent space representation, which is a compressed digital representation LSR of the input mode data M1 to M6. The MVAE includes decoders D1 to D6, which are implemented to reconstruct the modality or mode-specific sensor data into a reconstruction space. With respect to the MVAE, the input mode data M1 to M6 construct the feature space, while the reconstructed mode data with respect to the reconstructed sensors SD1' to SD6' each construct the reconstruction space RM.

[0086] The example shown in Figure 5 is a multivariate encoder MVAE with six modalities. Various sensor data SD1-SD6 may have different dimensions P, Q, R, R, S, and 1 for modalities M1, M2, M3, M4, M5, and M6, respectively. Regardless of the different dimensions P, Q, R, R, S, and 1, all corresponding encoders E1-E6 are projected into a latent space of the same dimension, which is then aggregated into an LSR as a result of the MVAE. The MVAE is or includes an artificial neural network with weights and connections that are preset in the training method.

[0087] An example in Figure 5 illustrates a batch process carried out on a chemical plant. In the first step P1, the starting product or substrate is characterized. For example, the sensor data or batch process parameters associated with P1 may be, for example, the chemical composition that can be measured or obtained using a spectrometer as a sensor, for example, the identifier of the product that can be obtained or acquired by a camera as a sensor, for example, the weight of the product obtained by weighing using a balance as a sensor. In the second step P2, a slurry is provided. For example, the sensor data or batch process parameters associated with P2 may be, for example, the concentration or solid content that can be measured by a hydrometer or by drying and weighing the residue, for example, the particle size distribution that can be measured by a laser diffractometer, for example, the viscosity, density, pH value, temperature, and / or fluidity that can be measured by a viscometer. Similarly, the coating steps P3 and P4 and the calcination step S5 may be similarly characterized by the associated sensor data or batch process parameters. In step S6, the final product is characterized, and for example, back pressure may be considered a target variable or measure of quality.

[0088] MVAE is initially trained on a dataset that is (essentially) complete, i.e., a dataset in which all batch process parameters, including the target variable, are measured or determined, and can therefore be considered sensor data. The measured or determined batch process parameters or sensor data, indicated as SD1-SD6, are associated with their respective process steps P1-P6 and MVAE modes M1-M6. Training may be unsupervised. Training of the MVAE can be performed by encoding SD1-SD6 into a compressed digital representation LSR, decoding or reconstructing them from the LSR as SD1'-SD6', and minimizing the difference between SD1-SD6 and SD1'-SD6'. During training, the weights of nodes in the neural network present in the encoder and decoder may be determined, which can later be used to encode or decode the LSR. To train the model more specifically, some modes may be blocked for part of the training to decode or generate SD1'-SD6' when not all of SD1-SD6 are available, for example, not all of SD1-SD6 are used. MVAEs can be trained using a loss function, along with additional components for handling multiple modalities. The loss function may include the reconstruction loss for each modality and may include a Kullback-Leibler (KL) divergence term.

[0089] In the context of MVAE, the encoder portion of the MVAE, including the (sub)encoders E1-E6, can be considered a data-driven model, and the decoder portion, including the (sub)decoders D1-D6, can be considered a second data-driven model. The encoder portion is trained to generate a compressed digital representation LSR of (i) a quality measure indicating the physicochemical quality of the chemical products produced by the plant and (ii) batch process step parameters for six process steps, with the quality measure included as a batch process parameter for the sixth process step. The decoder portion is trained in the opposite direction.

[0090] Subsequently, during production, for example, after measuring only the batch process parameters of the first process step P1 and acquiring sensor data S1, the encoder portion of the MVAE may be used to encode the sensor data S1 into LSR, and the decoder portion of the MVAE may be used to decode the process parameters of the remaining batch process steps P2 to P6, thus including the quality measure within P6. The same can be repeated, for example, after acquiring sensor data S1 and S2 for the first two process steps, or after acquiring S1 and S3 if S2 is unavailable for some reason. The reconstructed or generated batch process parameters SD2' to SD6' can be output directly as plant operation data, or they can be further processed to provide specific instructions on how to adjust future batch process steps, for example. In this way, the batch process can be reliably monitored and controlled, even in the early stages of production.

[0091] A common method for training MVAE is shown in Figure 6. Training an MVAE involves the use of training data, which could be, for example, ground truth data for each batch process BP. In training step T1, training data corresponding to sensor data SD1-SD6 shown in Figure 5 is generated. Generating training data may include monitoring batch processes corresponding to batch process specifications and collecting sensor data as shown for modes M1-M6 in Figure 5. The training data is provided in feature space.

[0092] The encoders, decoders, and autoencoders presented herein may be implemented in accordance with M. Wu and N. Goodman's “Multimodal Generative Models for Scalable Weakly-Supervised Learning”, arXiv:1802.05335, and its references, which are incorporated herein by reference.

[0093] Next, in training step T2, encoders E1-E6 receive their respective input mode training data for each mode. In training step T3, latent space data, i.e., a digital representation of the entire mode data, is generated in latent space LSR. Encoders E1-E6 may include layers characterized by weights and measurement behavior according to the type of neural network used. The weights and other parameters of the encoders and the decoders described last are optimized during the training method. The latent space representation of sensor data typically has a lower dimension than the entire mode.

[0094] In training step S4, decoders D1-D6 are used to reconstruct modes in the reconstruction space RM. Typically, encoding and decoding processes involve losses. Therefore, the reconstructed sensor datasets SD1'-SD6' may not be accurate to the input sensor datasets SD1-SD6. By comparing the input and output data, a loss function representing the similarity between the input and output data is calculated in training step T5.

[0095] The calculated loss function indicates how well the neural network or MVAE performs the mapping from input (features) to output (reconstruction) space. A smaller loss function indicates better MVAE performance. During several training runs, the weights of the multimodal encoder or decoder are fitted to minimize the loss function. This is indicated by the dashed arrows, so training steps T2-T5 are repeated. Once an acceptable loss is obtained, the MVAE can be used for a specific purpose, namely, the batch process specification on which the MVAE was trained. MVAE enables a variety of applications related to the control of a plant that performs each batch process. Note that various training algorithms and MVAE types can be used according to known techniques.

[0096] The following provides examples of the use or application of MVAE for controlling chemical plants. Figure 7 shows an embodiment of an application for MVAE to predict batch process step parameters, sensor data, and / or quality or performance indicators involved in the batch process. As an example, Figure 7 shows an MVAE with four modalities. Encoders E1-E4 receive mode data M1-M4 and generate latent spatial representations (LSR), i.e., compressed digital representations of the sensor data contained in modes M1-M4, and decoders D1-D4 generate reconstructed mode data M1'-M4'. As in the previous example, modes or modalities M1-M4 correspond to batch process steps.

[0097] The advantage of MVAE is that even if only incomplete input mode data is used, reasonable and reliable reconstructed mode data is available for each mode. For example, as shown by the dashed line representation of mode M4, only sensor data from the first three batch process steps for M1, M2, and M3 is available. Nevertheless, MVAE outputs reconstructed mode data M1', M2', M3', and M4'. Therefore, MVAE makes it possible to predict sensor data with respect to the reconstructed sensor data for the target mode M4. That is, in a batch process, the result or parameters of the last batch process step M4 can be predicted based on the mode data M1, M2, and M3, i.e., the first three batch process steps.

[0098] The plant operator can predict the target KPI based on the information from the first step (first mode M1). Uncertainty may be relatively high. However, as the batch goes through more steps (modes M2, M2, M3), the prediction becomes increasingly accurate. Between steps (modes), it is possible to react to predictions that are above or below the threshold of the selected KPI. The plant operator can run simulations of how the parameters for modes the batch has not yet gone through need to be set to move the KPI prediction towards the desired specification range. Based on the existing information and the target KPI range, decoders D1-D4 reconstruct the plant parameters, and the plant driver can validate these plant parameters if this represents a valid plant condition. For this purpose, the reconstructed data is output to a display (not shown).

[0099] Another application refers to predicting the results of process changes. For example, as indicated by the dotted modes M3a and M3b, alternative method steps M3a and M3b replace a third method step corresponding to M3. Adding or removing modes can be flexibly achieved with respect to MVAE, especially if the modes are trained in an independent manner. The results or outcomes of batch process changes, which are the digital representation presented with respect to the latent space obtained by MVAE, can be predicted. As a result, the characteristics of the modified modes / steps M2', M3' can be predicted and / or performed. Advantageously, the batch process modified accordingly may reduce the use of resources as energy, time, waste handling, or defective products / products.

[0100] Another application may involve considering the location of the production plant where the batch process under consideration is implemented. A single modality can be input or used for each location. That is, by changing the modality from a European location to Asia or America, expected outcomes or variations in specific performance or quality can be predicted. This is particularly advantageous for plant operators with distributed sites who are shifting production from one location to another. A further application of digital representations for LSR and MVAE is upscaling a given batch process from laboratory scale to industrial scale. Generally, upscaling requires considerable effort in terms of time and energy resources. Deploying the proposed application of data-driven models makes it possible to more efficiently determine new operating parameters at the modified scale.

[0101] Figures 8, 9, and 10 refer to the applicant's investigation related to the predictive capability and reliability of MVAE. Regarding Figures 8-9, a batch process representing SCR during filter production is considered. The investigation is based on synthetic data, e.g., ground truth data. The results shown in Figure 9 refer to seven common modes, and Figure 10 refers to five modes corresponding to those considered in relation to SCR production according to Figure 5, and as shown in Figure 5, include a simplified batch production process including five modes, namely M1, M2, a combination of M3 and M4, and M5. It is assumed that the target mode to be predicted is back pressure.

[0102] Figures 9 and 10 show the mean absolute error (MAE) as a function of the modes included as input sensor data. The curves in Figures 9 and 10 are obtained along the evaluation steps shown in Figure 8. In the first step C1, synthetic data is obtained as training data. For example, synthetic training data is found as ground truth data or under controllable conditions of a batch process. Figure 9 shows synthetic test data based on a four-dimensional latent space and seven mode MVAEs with mode numbers 1-7, where mode number 2 is considered the target mode. The seven mode datasets are obtained with seven nonlinear decoder functions, and the MVAEs are trained for each mode, respectively.

[0103] Therefore, the MVAE is trained so that the reconstruction loss is minimized based on the available training data (step C2).

[0104] Next, in step C3, MVAE is used and input mode data is supplied. Initially, only one mode is used, for example, sensor data indicating the substrate (P1 in Figure 5). The reconstructed modes include, in particular, mode M6 as the target mode. In the next step, two modes are provided as input data, and the target mode is reconstructed again. In Figures 9 and 10, the MAE and variance decrease as more and more modes are included as input sensor data. The mean absolute error MAE is obtained by averaging the results of the synthesized or training data used.

[0105] Referring to Figure 10, for example, inputting more modal sensor data referencing the substrate that represents a batch process step results in relatively high variance. Including further modes M2, M3, etc., which represent further batch process steps in Figure 5 reduces the variance. Including more modalities as input data for MVAE further reduces the error.

[0106] An incomplete set of sensor data with respect to the mode of MVAE, which points to a subset of the batch process steps under consideration, may be found to be sufficient to predict the target mode with respect to the reconstructed sensor dataset. Thus, the presented method and system make it possible to predict the impact of changes in the batch process, i.e., the batch process under consideration, i.e., when sensor data cannot be directly measured, on several characteristics.

[0107] The applicant's research has shown that prediction errors can be reduced by including additional modes with respect to sensor data that indicate or describe batch process steps (Figures 9 and 10). The presented use of MVAE provides flexible modifications regarding the modes used. New modes or features can be added to the model relatively easily to match real-world batch processes.

[0108] Even if this disclosure relates to embodiments and specific MVAEs as exemplary sensor data, the concept of compressed digital representation can be achieved by alternative data-driven models with machine learning aspects, such as transformer models. While the production of SCRs on filters is shown as one embodiment of a batch process, alternative batch processes, such as the synthesis of chemicals, can be envisioned.

[0109] Figure 11 shows an example of an autoencoder implemented as a feedforward neural network. Sensor data SD_N is received by an input layer 111_1 containing several nodes (represented as open circles). The sensor data is processed in several hidden layers 111_2 to an output layer 111_3, which has fewer nodes than the input layer 111_1. Thus, a node in one layer may be connected to each node in an adjacent layer, and the connections between nodes may contain weight values ​​that can be adjusted during training. The output layer 111_3 provides a compressed digital representation LSR. An example of a decoder could be an inversely structured feedforward neural network, where 111_3 corresponds to the input layer and 111_1 corresponds to the output layer.

[0110] Figure 12 shows another example of an MVAE including an encoder, which comprises several (sub)encoders E1-EN for a corresponding number of modes M1-MN, each of which may correspond to each batch process step, for example, as described above. Each (sub)encoder (which may be an inference network) determines the probability distribution of its respective mode and characterizes this probability distribution with a value or variational parameter, in this example the respective mean μ1-μ N and standard deviation σ1-σ NThis can provide the following. The distributions can be combined into a single distribution by the Expert Product (PoE) in this example. The combination can be executed within the encoder as a separate module of the encoder, for example, operably coupled to (sub)encoders M1~MN. Alternatively, prior probability distributions obtained by, for example, pre-training, pre-modeling, etc., can be input to the PoE. The PoE can combine variational parameters in a principled and efficient manner. For example, if a mode is missing during training, which may be done intentionally to train the reconstruction of missing modes, each encoder or inference network may be dropped. Thus, the parameters of E1~EN are shared across different combinations of missing inputs. The compressed digital representation LSR is or contains the combined probability distribution from which the reconstructed modes M1'~MN' can be sampled or generated through (sub)decoders D1~DN.

[0111] In this specification, it should be understood that any connection presented in the embodiments described herein is such that the components involved are operably coupled. Thus, the connection may be direct or indirect, involving any number or combination of intervening elements, and merely a functional relationship may exist between the components.

[0112] Furthermore, any of the methods, processes, and actions described or illustrated herein may be performed using executable instructions within a general-purpose or dedicated processor and stored in a computer-readable storage medium (e.g., disk, memory) executed by such a processor. The reference to “computer-readable storage medium” should be understood to include dedicated circuits such as signal processing devices and other devices.

[0113] The expressions "A and / or B" and "at least one of A or B" are considered interchangeable and mean any one of the following three scenarios: (i) A, (ii) B, or (iii) A and B. More generally, the expressions "at least one of the following list of two or more elements" and "at least one of the list of two or more elements" and similar expressions mean at least one of the elements, or at least any two or more of the elements, or at least all of them, when the list of two or more elements is joined by "and" or "or".

[0114] The article "a" should not be understood as "one," meaning that the use of the expression "an element" does not exclude the existence of further elements.

[0115] The term "comprising" should be understood in an open sense, meaning that an object "comprising element A" may also possess additional elements in addition to element A. Furthermore, the term "comprising" can sometimes be limited to "consisting of," that is, consisting of only specific elements.

[0116] All presented embodiments are illustrative, and it should be understood that any feature presented for a particular exemplary embodiment may be used in any form of itself, or in combination with any feature presented for the same or another particular exemplary embodiment, and / or in combination with any other feature not mentioned herein. In particular, the exemplary embodiments presented herein should also be understood to be disclosed in all possible combinations with each other, provided that they are technically reasonable and the exemplary embodiments are not substitutes for each other. Furthermore, it should be understood that any feature presented for an exemplary embodiment in a particular category (method / apparatus / computer program / system) may also be used in a corresponding manner in any other exemplary embodiment in any other category. It should also be understood that the presence of a feature in a presented exemplary embodiment does not necessarily mean that the feature constitutes an essential feature and cannot be omitted or replaced.

[0117] All sequences of method steps presented above are not mandatory, and alternative sequences may also be possible. Nevertheless, specific sequences of method steps illustrated in the figures should be considered as one possible sequence of method steps for each embodiment described by each figure. [Explanation of Symbols]

[0118] Reference sign: 1. A system for producing chemical substances / a chemical plant 2 Interfaces 3 processors 4 Databases 5. Control Unit 6, 7 containers 8, 9 valves 10 Reactors 11 Motor 12 Mixer 13 Heater 14 Outlet valve 15 tanks BP batch process BPS Batch Process Specification CD Plant Operation Data / Control Data CS control signal Decoder for DQ ModeQ Encoder for EQ Mode Q DDCM (Data-Driven Compression Model) LSR latent space representation M1~M1 Mode / Modality P1~PN Batch Process Steps RM Reconstructed Mode SDQ Sensor data associated with process step PQ SDQ' Reconstructed Sensor Data S0 generates sensor data S1 Receives sensor data Apply the S2 data-driven compression model. S3 Determine quality measures / performance indicators S4 Generate control data S41 Setting Target Quality Measures S42 Compare quality measures / performance indicators with targets. S5 Activate the chemical plant T1: Generate training data T2 Input mode: Enter training data. Generate T3 latent data Reconstruct the T4 mode data. Adapt encoder / decoder parameters to reduce T5 reconstruction loss (loss function). Train to minimize C1 reconstruction loss. Samples are selected from the C2 latent spatial distribution. Rebuild C3 target mode Averaging across the C4 distribution C5 Determine the Mean Absolute Error (MAE)

Claims

1. A method for operating a chemical plant (1) implemented to perform a batch process (BP) for producing a chemical product, wherein the batch process (BP) comprises a plurality of process steps (P1 to PN), and the method S1) Receiving sensor data (SDQ) associated with at least one process step (PQ) of the batch process (BP), S2: To generate a compressed digital representation (SDQ') of the sensor data (SDQ) associated with at least one process step (PQ), the sensor data (SDQ) is processed according to a data-driven model (DDCM). (S3) Determining a quality measure indicating the physicochemical quality of the chemical product produced by the plant (1) and / or a performance index of the batch process as a function of the digital representation (SDQ') of the received sensor data (SDQ), Based on the determined quality scale and / or the determined performance indicators, generate plant operation data (CD) indicating that the chemical plant (1) is in operation (S4), and output the plant operation data. Methods that include...

2. Each process step is associated with a set of batch process step parameters, and the sensor data corresponds to at least one batch process parameter associated with the at least one process step, and the sensor data (SDQ) is processed according to the data-driven model (DDCM) to generate a compressed digital representation (SDQ') (S2), S2) Using the data-driven model (DDCM), obtain the compressed digital representation of the sensor data (SDQ) associated with the at least one process step (PQ), wherein the data-driven model is trained to generate compressed digital representations of (i) a quality measure and / or a performance indicator of the batch process (BP) indicating the physicochemical quality of the chemical product produced by the plant (1), and (ii) the batch process step parameters of at least two process steps, including the at least one process step. is it that, or S2) Using the data-driven model (DDCM), obtain the compressed digital representation of the sensor data (SDQ) associated with the at least one process step (PQ), wherein the data-driven model is trained to generate compressed digital representations of (i) a quality measure and / or a performance indicator of the batch process (BP) indicating the physicochemical quality of the chemical product produced by the plant (1), and (ii) the batch process step parameters of at least two process steps, including the at least one process step. The method according to claim 1, including the following:

3. The method according to claim 2, wherein determining the quality scale and / or the performance index (S3) is to generate the quality scale and / or the performance index based on the compressed digital representation using a second data-driven model (S3), or to generate the quality scale and / or the performance index based on the compressed digital representation using a second data-driven model (S3), wherein the second data-driven model is trained to generate the quality scale and / or the performance index from the compressed digital representation of (i) the quality scale and / or the performance index of the batch process (BP) indicating the physicochemical quality of the chemical product produced by the plant (1), and (ii) the batch process step parameters of at least two process steps including the at least one process step, and determining plant operation data (CD) (S4) is based on the generated quality scale and / or the generated performance index.

4. The method according to claim 2 or 3, wherein the data-driven model includes an encoder for an autoencoder, and the second data-driven model includes a decoder for the autoencoder, wherein the encoder is trained to encode (i) the quality measure and / or the performance index of the batch process (BP) indicating the physicochemical quality of the chemical product produced by the plant (1) into the compressed digital representation, and (ii) the batch process step parameters of at least two process steps including the at least one process step, respectively, and the decoder is trained to decode (i) the quality measure and / or the performance index of the batch process (BP) indicating the physicochemical quality of the chemical product produced by the plant (1) and (ii) the batch process step parameters of at least two process steps including the at least one process step from the compressed digital representation.

5. The method according to claim 4, wherein the autoencoder is a multimodal variational autoencoder, and each process step of the batch process corresponds to a mode of the multimodal variational autoencoder, in particular, at least one mode corresponds to the quality measure indicating the physicochemical quality of the chemical product produced by the plant and / or the performance index of the batch process (BP).

6. The method according to claim 5, wherein at least one further mode corresponds to the location of the plant or the characteristics of the plant.

7. The data-driven model is further trained to generate the compressed digital representation of (i) a quality measure indicating the physicochemical quality of the chemical product produced by plant (1) as if produced by another plant, and / or a performance indicator of the batch process (BP) performed on the other plant, and (ii) the batch process step parameters of at least two process steps performed on the other plant, including the at least one process step. The method according to any one of claims 3 to 6, wherein the second data-driven model is trained to generate the quality measure and / or the performance index from the compressed digital representation of (i) the quality measure and / or the performance index of the batch process (BP) performed on the other plant, which indicates the physicochemical quality of the chemical product produced by plant (1) as produced by the other plant, and (ii) the batch process step parameters of at least two process steps performed on the other plant, which include the at least one process step.

8. The batch process performed on plant (1) has at least one additional process step, and the batch process is then performed on the other plant, and the method is - Using the second data-driven model described above, determine at least one process parameter of the at least one additional process step, - To provide at least one of the process parameters The method according to claim 7, further comprising:

9. The chemical plant is operated as a function of the plant operation data, or as a function of the control data (CD) if the plant operation data includes control data (CD) (S5). The method according to any one of claims 1 to 8, further comprising:

10. To obtain comparison results, the quality scale and / or performance indicators are compared with the target quality (S42), Based on the above comparison results, control data for operating the chemical plant is generated. The method according to any one of claims 1 to 9, further comprising:

11. Based on the quality measure and / or performance indicator, generate control data suitable for displaying on a display device at least one of the following: a warning message to an operator, a list of operating conditions for the chemical plant, an alarm, or the physicochemical quality of the intermediate products of the batch process. The method according to any one of claims 1 to 10, further comprising:

12. The received sensor data (SDQ) represents a first set of first physicochemical parameters, and the method is To generate sensor data derived from the compressed digital representation of the received sensor data (SDQ), wherein the generated sensor data represents a second set of physicochemical parameters that are at least partially different from the first set. The method according to any one of claims 1 to 11, further comprising:

13. The aforementioned data-driven model (DDCM) is implemented to receive input mode data (SD1 to SD6) and output reconstructed mode data (SD1' to SD6'), Each batch process step (P1 to P6) of the batch process (BP) is associated with one respective mode (M1 to M6). The input mode data (SDQ) for each mode (MQ) includes the sensor data (SDQ) associated with at least one batch process step (PQ) associated with each mode (MQ), The output mode data (SDQ') of each of the modes (MQ) includes reconstructed sensor data (SDQ) associated with the at least one process step (PQ) associated with each of the modes (MQ), The method according to any one of claims 1 to 12.

14. The DDCM has a predetermined number of modalities (M1 to M6) corresponding to the number of batch process steps (PN), and the method is Inputting input mode data for a first set of modes, including a first number of modes, Outputting reconstructed mode data of a second set of modes, including a second number of modes, wherein the output contains fewer modes than the second number of modes. The method according to claim 13, further comprising:

15. The method according to claim 14, wherein the second set of modes includes at least one mode associated with a batch process step (PN) that is performed later than the batch process steps (P1 to PN-1) associated with the modes in the first set of modes.

16. The DDCM is trained using a training dataset generated by monitoring multiple variables of the batch process in order to generate sensor data representing each batch process step. The method according to any one of claims 13 to 15, further comprising:

17. The variables of the batch process include the laboratory set up for the batch process, systems with different production capacities for the chemical products to be produced, different plant sizes, and different chemical plants located in different locations. The method according to claim 16, including the method described in claim 16.

18. To measure batch process parameters that indicate the physicochemical properties of the raw material with respect to the sensor data and / or the characteristics of the batch process steps (P1 to PN), Clustering the aforementioned sensor data into clusters, wherein each cluster is assigned to one batch process step and / or mode of MVAE. The method according to claim 16 or 17, further comprising:

19. The method according to any one of claims 1 to 18, wherein the sensor data includes a set of physicochemical parameters of a substance, a substance identifier, a quantity, process specifications, and batch process step specifications.

20. Predicting a quality measure and / or performance indicator of a selected batch process step (PQ) based on sensor data associated with another process step of the batch process, wherein the other process step precedes the selected process step (PQ). To generate control data as a function of the predicted quality measure and / or performance indicators. It further includes, The method according to any one of claims 1 to 19, wherein the prediction includes reconstructing sensor data associated with the selected batch process step.

21. The method according to claim 20, wherein the batch process comprises a plurality of sequential process steps (P1 to PN), including a first process step (P1) and a final process step (PN), each batch process step may be controlled as a function of control parameters implemented by the control data (CD), and the method comprises fitting the control parameters (CD) of the selected batch process step (PQ) as a function of the predicted quality measure and / or performance indicator.

22. A system (1) for producing a chemical product, comprising a control unit (3) associated with a batch process step and controllable plant units (6, 7, 10), wherein the control unit (3) is implemented to perform the method according to any one of claims 1 to 14.

23. An apparatus comprising means for carrying out or performing the steps of the method described in any one of claims 1 to 21.

24. A device comprising at least one processor and at least one memory storing instructions, wherein the instructions are An apparatus that, when executed by the at least one processor, causes the apparatus to perform the steps of the method according to at least one of claims 1 to 21.

25. Plant operation data obtained by the method according to any one of claims 1 to 21 or by the apparatus according to claim 23 or 24.

26. A computer program comprising instructions for performing at least one step of the method according to any one of claims 1 to 21.

Images