Method, system and paperboard production machine for estimating paperboard quality parameters

Abstract

The invention relates to a computer-implemented method for estimating at least one quality parameter of paperboard produced in a paperboard production subprocess (PP) of a paperboard processing line (P) by means of a data-driven module (DD) comprising a pre-processing module (PM) and a machine learning module (ML). Sensor data (D1 to D5, D1' to D5') are acquired along the processing line (P). Features (F, F') are extracted from the sensor data (D1 to D5, D1' to D5') by means of the pre-processing module (PM). The machine learning module (ML) is trained to reproduce target quality values from historical features (F'). After training, the machine learning module (ML) processes real-time features (F) and estimates the at least one quality parameter. The invention further relates to a system implementing said computer-implemented method and to a paperboard production machine comprising such a system.

Description

TECHNICAL FIELD

[0001] The present invention relates to a method and a system for estimating a paperboard quality parameter. The invention further relates to a paperboard production machine implementing such a method for estimating a paperboard quality parameter. BACKGROUND

[0002] Paperboard is continuously produced in a processing line comprising a plurality of processing steps, e.g. mixing of introduced pulp, beating, chemical treatment of the pulp, and processing of the pulp through a paperboard machine. The produced paperboard is wound on large reels, where one reel can hold approximately 40 kilometers long and approximately 7 meters wide paperboard. The processing steps can be controlled by settings, and / or described by parameters affecting the quality of the produced paperboard. Further, environmental parameters such as temperature or humidity can affect the quality. Changes in processing parameters and environmental parameters can affect the quality in a delayed and non-linear manner.

[0003] Various quality parameters describing the quality of paperboard are known in the art.

[0004] For example, a quality parameter expressed as Z-strength is known to describe the strength of paperboard along its surface normal. As another example, a quality parameter expressed as Scott bond is known to describe the in-plane strength of paperboard.

[0005] An effective range can be specified for one or more of such quality parameters. Paperboard outside this specification will be downgraded or discarded.

[0006] In order to determine these quality parameters, according to the state of the art, samples are manually taken across the width of the paperboard reel. These samples are manually or semi-manually prepared and analyzed in a laboratory. Due to the high speed of the paperboard machine, sampling is not possible while the paperboard is being wound on the reel. Thus, sampling is typically possible about every 45 minutes. Manual or semi-manual analysis takes another about 20 minutes. Thus, the status and / or trend of the quality parameters is unknown for a relatively long period, which causes a risk of a large amount of scrap in paperboard production.

[0007] There is thus a need for an improved estimation of paperboard quality. SUMMARY

[0008] It is an object of the present invention to provide an improved method for estimating at least one quality parameter of paperboard produced in a paperboard production sub-process of a processing line. This object is achieved by the computer implemented method according to claim 1.

[0009] It is another object of the present invention to provide an improved system for estimating at least one quality parameter of paperboard produced in a paperboard production sub-process of a paperboard processing line. This object is achieved by the system according to claim 12.

[0010] It is a further object of the present invention to provide a paperboard production machine implementing an improved method for estimating at least one quality parameter of produced paperboard. This object is achieved by the paperboard production machine according to claim 15.

[0011] Exemplary embodiments of the present invention are given in the dependent claims.

[0012] According to a first aspect of the present invention, a computer-implemented method is designed for estimating at least one quality parameter of paperboard produced in a paperboard production sub-process, which is a sub-process of a paperboard processing line. The paperboard processing line is implemented by a data-driven module, which comprises a pre-processing module and a machine learning module, and which comprises at least one processing step implemented by a processing unit or machine. The computer-implemented method comprises the following steps:

[0013] The computer-implemented method comprises a step, in which sensor data is acquired from the at least one processing step of the processing line and is transferred to a data repository. The data repository is designed to store the sensor data persistently. Preferably, the data repository is accessible via Internet Protocol (IP). Most preferably, the data repository is formed as a cloud storage.

[0014] The computer-implemented method comprises a further step, in which at least one historical feature is determined by the pre-processing module by evaluating at least partially historical sensor data. Such historical sensor data is acquired during at least one previous production batch of paperboard and is retrieved from the data repository.

[0015] The computer-implemented method comprises a further step, in which the machine learning module is trained to reproduce a target value of the at least one quality parameter from the at least one historical feature. The target value is determined for a previous production batch of paperboard corresponding to the historical sensor data from which the historical feature is determined, for example by a measurement in a laboratory and is transferred to the data repository, where it is associated with the respective historical sensor data.

[0016] The computer-implemented method comprises a further step, in which at least one real-time feature is determined by the pre-processing module from a current sensor data stream, which is acquired from a currently produced batch of paperboard and is retrieved from the data repository.

[0017] The computer-implemented method comprises a further step, wherein the estimation of at least one quality parameter of the produced paperboard is determined by the trained machine learning module from the at least one real-time feature and provided as an output value of the data-driven module.

[0018] The method has the advantage that the quality parameter of the produced paperboard can be estimated continuously during the production process and with less latency in relation to known methods. Thereby, the paperboard production sub-process can be controlled more precisely and the target value or range of the target quality parameter of the produced paperboard can be met more reliably. As a result, costs related to discarded or downgraded paperboard can be reduced and the efficiency of the paperboard production can be improved.

[0019] A further advantage of the method is that the sensor data is centrally stored and independent of the physical location of the machines performing the processing steps in the paperboard production. Thereby, a more comprehensive data analysis can be performed which improves the prediction accuracy of the paperboard quality estimation and helps to detect faults and defects along the paperboard production sub-process. Furthermore, the risk of data loss and data inconsistencies is reduced.

[0020] According to an embodiment of the invention, the historical sensor data is divided into a plurality of intervals, preferably 10,000 intervals or more. Events of the historical sensor data associated with error messages or inconsistencies recorded and stored in the data repository are omitted.

[0021] The embodiment has the advantage that a larger number of training samples is provided for the training of the machine learning module, wherein the training samples comprise features derived from the isolated historical sensor data and target quality values associated with the isolated historical sensor data. Thereby, by removing unreliable events, the machine learning module is trained such that it provides a more accurate and reliable estimation of the quality parameter of the current sensor data flow.

[0022] According to an embodiment of the invention, a delay group is determined for the processing step. The delay group defines a minimum latency time in which a change of a parameter in the processing step influences the outcome of the production sub-process. Current sensor data acquired within this latency time of the respective delay group counted backwards from the current production timestamp is excluded from the evaluation by the data-driven module.

[0023] The embodiment has the advantage that sensor data without potential prediction values for the estimation of the quality parameter, which are in fact noise, are removed from the training and from the prediction by the machine learning module. Thereby, the machine learning module can be trained more efficiently and the estimation accuracy and reliability can be improved.

[0024] According to an embodiment of the application, the at least one feature comprises at least one time series. The time series describes a variation of at least one parameter of the sensor data along a discrete time interval or time window. The machine learning module performs a one-dimensional (ID) convolution of the at least one time series along the time axis.

[0025] An advantage of this embodiment is that in the training of the machine learning module, parameters of the ID convolution forming a filter operating on the respective time series are adjusted, such as in order to optimize the reproduction of the target quality value. Thereby, features with a certain predictive value are extracted from the respective time series. Thereby, the machine learning module can be trained more efficiently and the estimation accuracy and reliability can be improved.

[0026] According to an embodiment of the application, for a discrete time window or time interval, at least one statistical parameter, preferably a mean value and / or a standard deviation value, is determined, such as in order to form a sample value of the at least one time series.

[0027] An advantage of this embodiment is that the statistical parameter is a more reliable representation of the process state during such a time window compared to the values of the plurality of individual samples from which the statistical parameter is determined. Thereby, the machine learning module can be trained more efficiently and the estimation accuracy and reliability can be improved.

[0028] According to an embodiment of the application, the validity and / or consistency of the current sensor data is evaluated by the preprocessing module and optionally recorded in the data repository. This enables the preprocessing module to determine events of invalidity or unreliability of the current sensor data and to remove these events from the estimation of the quality parameter. Thereby, the machine learning module can be trained more efficiently and the estimation accuracy and reliability can be improved.

[0029] According to an embodiment of the application, one quality parameter estimated by the data driven module is the Z-strength of the paperboard. The Z-strength is an established parameter that is particularly relevant for the grading of the produced paperboard. As an advantage, this embodiment reduces costs associated with discarded or downgraded paperboard and improves the efficiency of the paperboard production.

[0030] According to an embodiment of the application, one quality parameter estimated by the data driven module is the Scott bond strength of the paperboard. The Scott bond strength is an established parameter that is particularly relevant for the grading of the produced paperboard. As an advantage, this embodiment reduces costs associated with discarded or downgraded paperboard and improves the efficiency of the paperboard production.

[0031] According to an embodiment of the application, at least one range of the quality parameter is provided as an additional input to the data-driven module, wherein for each range, a probability value is determined as an output value that the respective quality parameter falls into the respective range. For a scalar quality parameter, the range can be formed as a closed, open or semi-open interval.

[0032] As an advantage, this embodiment enables a more detailed estimation of the quality parameter that better reflects its stochastic nature.

[0033] According to an embodiment of the application, for one or more probability values determined by the data-driven module for a range of the quality parameter, each probability value is compared to a probability limit assigned to the respective range and the respective quality parameter, resulting in one Boolean comparison value for each range and quality parameter. A predetermined Boolean expression is evaluated that depends on the at least one Boolean comparison value, wherein an alarm is triggered when the predetermined Boolean expression returns a logically true value.

[0034] An advantage of this embodiment is that the status of a paperboard production sub-process whose quality values can be impaired can be detected automatically, objectively and with particularly little latency.

[0035] According to an embodiment of the application, for at least one quality parameter, a first range is a set of invalid values below a valid lower limit, a second range is a set of valid values between a valid lower limit and a valid upper limit, and a third range is a set of invalid values above the valid upper limit, wherein an alarm is triggered when the probability value associated with the first range exceeds a predetermined first probability limit, or when the probability value associated with the second range falls below a predetermined second probability limit, or when the probability value associated with the third range exceeds a predetermined third probability limit.

[0036] For the sake of clarity and better understanding, an example is provided for this embodiment:

[0037] For example, for a scalar parameter q, a first range r1 = {q | q < q l} is determined as an open interval of values below the lower limit q l . A second range r2 = {q | q l ≤ q ≤ q u} is determined as a closed interval between the lower limit q1 and the upper limit q u . A third range r3 = {q | q > q u} is determined as an open interval of values above the upper limit q u .

[0038] The data driven module is designed, in particular trained, such that it provides a first output corresponding to a first probability value of the quality parameter being within the first interval, p1 = P[q e r1]. The second and third outputs correspond to probability values of the quality parameter being within the second and third intervals, respectively: p2 = P[q e r2], p3 = P[q e r3].

[0039] A conditionally predetermined Boolean expression can then be formed, such as

[0040] (p1 > θ1) or (p2 < θ2) or (p3 > θ3)

[0041] wherein θ1, θ2, θ3 represent probability limits for the first to third probability values p1, p2, p3, respectively, and "or" represents a logical (Boolean) or expression. The alarm is triggered when said predetermined Boolean expression is evaluated to be logically (Boolean) true.

[0042] An advantage of this embodiment is that the state of a paperboard production sub-process, in which the quality value can be impaired, can be detected particularly easily.

[0043] According to a second aspect of the present invention, a system comprises at least one sensor, a data repository, and a computing device.

[0044] The at least one sensor is designed to acquire sensor data in a process step of a process line comprising a paperboard production sub-process. For example, the sensor can be designed to measure a temperature of the pulp, an ambient temperature, an ambient humidity, or a concentration of one or more components of the pulp.

[0045] The data repository is designed to receive and persistently store the sensor data from the at least one sensor and to transmit the sensor data to the computing device.

[0046] The computing device comprises (i.e. implements) the computer-implemented method according to the first aspect of the present invention.

[0047] By means of said system, the advantages related to the computer-implemented method for estimating at least one quality parameter of paperboard produced in a paperboard production sub-process can be achieved, as previously explained. As a further advantage, the sensor, the data storage for forming the data repository, and the computing device for implementing such a method are readily available.

[0048] According to embodiments of the present application, the data storage is formed as a cloud storage accessible by an Internet Protocol (IP) based communication protocol stack. Cloud storage provides a number of advantages such as relative low cost, good scalability, good reliability and availability. As a further advantage, the cloud storage is decoupled from the paperboard processing line and its physical components. Thereby, a more flexible and generic data analysis is enabled which allows for an improved accuracy and reliability of the estimation of at least one quality parameter. As an example, sensor data of more than one physical instance of a paperboard processing line can be merged in the data repository formed as a cloud storage. Thus, the machine learning module can be trained such that it is less prone to random irregularities along one particular paperboard processing line.

[0049] According to embodiments of the present application, the system comprises a signaling device designed to be triggered by the computing device. The signaling device is designed to indicate, when triggered, that a manual interaction with the paperboard production sub-process is required. This embodiment enables an immediate response to a processing state which can impair a quality parameter of the produced paperboard. Thereby, the risk and / or amount of paperboard to be discarded or downgraded is reduced.

[0050] According to a third aspect of the present application, a paperboard production machine designed to produce paperboard or a semi-finished product of paperboard along a paperboard production sub-process comprises a system according to the second aspect of the present application.

[0051] By means of the system, the advantages related to the computer-implemented method for estimating at least one quality parameter of paperboard produced in a paperboard production sub-process can be achieved as explained before. In particular, the paperboard production machine according to the present application provides for a better efficiency and a better and more reliable quality of the produced paper.

[0052] Further scope of applicability of the present application will become apparent from the detailed description given hereinafter. However, it should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the application, are given by way of illustration only, since various changes and modifications within the spirit and scope of the application will become apparent to those skilled in the art from this detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0053] The present application will become more fully understood from the detailed description given hereinbelow and the accompanying drawings which are given by way of illustration and thus do not limit the present application and wherein:

[0054] Figure 1 Sensor data acquisition along a paper processing line is schematically illustrated,

[0055] Figure 2 Data driven model for paper quality parameter estimation is schematically illustrated,

[0056] Figure 3 Extraction of statistical features from sensor data is schematically illustrated,

[0057] Figure 4 Machine learning models for paper quality parameter estimation are schematically illustrated, and

[0058] Figure 5 Classification models for classifying paper quality parameters are schematically illustrated.

[0059] In all figures, corresponding components are labeled with the same reference signs. DETAILED DESCRIPTION

[0060] Figure 1 A processing line P for paper production is schematically illustrated, in which material M1, M2 is processed in sequence through a plurality of processing steps P1 to P5.

[0061] For example, the first processing step P1 consists of preparing incoming pulp. The subsequent second processing step P2 consists of beating. The subsequent third processing step P3 consists of chemically treating the beaten pulp. The subsequent fourth step P4 consists of setting parameters of a paperboard machine. The subsequent fifth processing step P5 consists of quality measurements in a laboratory. Thus, in this example, the first to fourth processing steps P1 to P4 form a production subprocess PP, while the fifth processing step P5 forms a quality assurance subprocess QA.

[0062] The material passes along the first to fourth processing steps P1 to P4 as wet pulp M1. The material for the fifth processing step P5 passes from the paperboard machine as paperboard samples M2. For example, the paperboard samples M2 are cut from the end of a reel on which continuous production of paperboard is wound and manually transferred to a remote laboratory.

[0063] The different processing steps P1 to P5 can be performed by physically different processing units or machines at physically different locations. However, it is also possible that multiple processing steps P1 to P5 are performed by a single processing unit or machine.

[0064] Irrespective of their physical implementation and location, each of the processing steps P1 to P5 is respectively associated with and respectively characterized by a specific current sensor data D1 to D5 and can thus be considered as a sensor data source.

[0065] For example, the current sensor data D5 associated with the fifth processing step P5 can be provided as Z-strength and Scott-bond measurements of the respective paperboard sample M2, i.e. as a time-discrete two-dimensional value vector sampled at a frequency of about once per hour. Other paper quality parameters such as moisture or grammage can additionally or alternatively be provided.

[0066] As another example, the current sensor data D1 associated with the first processing step P1 can be provided as categorical or numerical values describing the quality of the introduced pulp.

[0067] Further current sensor data D1 to D5 can relate to parameters of the respective processing step P1 to P4, such as settings of the beating process, or as parameters of the chemical preparation of the pulp, or as settings of the specific paper production machine designed to perform one or more of the processing steps P1 to P4.

[0068] According to the present application, the current sensor data D1 to D5 of the plurality of processing steps P1 to P5 is transferred to a central sensor data repository R. Preferably, the current sensor data D1 to D5 of all processing steps P1 to P5 is transferred to the central sensor data repository R.

[0069] In an embodiment, the data repository R is formed as a cloud storage such as e.g. a SIEMENS MindSphere storage. As an advantage, such an embodiment is independent of a dedicated data network infrastructure such as a proprietary on-premise bus. Thus, the data repository R can remain fully functional even during or after a physical relocation of the machine or processing unit. Moreover, the sensor data for the entire processing line P can be consolidated in a single location regardless of the physical location of the individual processing steps P1 to P5. For example, the building in which the laboratory for determining the quality parameters of the paperboard sample M2 in the fifth processing step P5 is located can be different from the building in which the paperboard machine performing the fourth processing step P4 is located.

[0070] Furthermore, the sensor data is available independent of the usage cycle of a specific machine or even a model of a machine, thereby providing advantages regarding the availability, consistency and security of the recorded sensor data D1 to D5.

[0071] From the sequential order of the processing steps P1 to P5 along the time axis T, a delay group T1 to T5 can be derived, wherein each processing step P1 to P5 is associated with one delay group T1 to T5.

[0072] The delay groups T1 to T5 determine a time shift of the acquisition of the current sensor data D1 to D5 relative to the production time stamp t0 defining the completion of a certain production batch. This time shift is discretized along the time axis T. For example, such a time shift can be defined as a multiple of a 15 minutes ground time interval.

[0073] More particularly, the delay groups T1 to T4 assigned to the processing steps P1 to P4 of the production subprocess PP are determined as the minimum time interval or waiting time required for variations in such processing steps P1 to P4, be it a variation of processing parameters, machine settings or parameters of the wet pulp material M1 at the input of said processing steps P1 to P4, to achieve any observed quality parameters of the produced paper batch.

[0074] When current sensor data D1 to D4 derived from the processing steps P1 to P4 of the production subprocess PP are acquired within the delay groups T1 to T4 associated with the respective processing steps P1 to P4, the current sensor data D1 to D4 are rejected for sensor data processing.

[0075] According to the present invention, the same delay groups T1 to T5 apply to all current sensor data D1 to D5 derived from the same processing steps P1 to P5.

[0076] For example, a delay group T2 = 45 minutes can be assigned to the beating process P2. The beating process P2 can be characterized by a refiner setting, a wet pulp temperature and an ambient temperature, and possibly other parameters, all of which are acquired and transmitted as sensor data D2 to the data repository R. If any such sensor data D2 value obtained at a time tx lags behind the waiting time of the associated delay group T2, i.e. tx - t0 > T2, any such sensor data D2 value obtained at a time tx will only be considered for the prediction of some quality parameters.

[0077] The delay group T5 assigned to the processing step P5 of the quality assurance subprocess QA defines a delay of the quality measurement (available as sensor data D5) relative to the production time stamp t0 and serves to align (in time) this sensor data D5 relative to the production time stamp t0.

[0078] Figure 2 The data flow through the data driven model DD for estimating at least one paper quality parameter, such as Z-strength and / or Scott-bond, is schematically illustrated. The data driven model DD comprises a pre-processing module PM and a machine learning module ML. The data driven model DD is driven by data retrieved from the central data repository R, said data comprising current sensor data D1 to D5 from the current paper production process and historical sensor data D1' to D5' acquired during the production process of previous batches and transmitted to the data repository R.

[0079] The sensor data D1 to D5 and the historical sensor data D1' to D5' are transferred to a pre-processing module PM, where features F, F' are extracted, e.g. statistical parameters. Real-time features F are extracted from the current sensor data D1 to D5 acquired from the current paper production process. Historical features F' are extracted from the historical sensor data D1' to D5' acquired before in the production of previous batches.

[0080] The pre-processing module PM checks the consistency and validity of the sensor data D1 to D5, D1' to D5'. It analyzes whether valid data is received from all merged data sources, in particular from all machines and / or processing units associated with the processing steps P1 to P5 along the processing line P. For example, the pre-processing module PM explicitly searches for missing timestamps or missing sensors along the sensor data D1 to D5, D1' to D5'. Furthermore, the pre-processing module PM can verify whether previously specified events, such as the shutdown of the entire production sub-process PP or parts thereof, occurred correctly along the processing line P. Each specific consistency check is logged as an error in case it fails.

[0081] The later use Figure 3 The extraction of features F, F' by the pre-processing module PM is explained in more detail.

[0082] The features F, F' are transferred to a machine learning model ML which estimates one or more paper quality parameters from these features F, F'. In an embodiment, the machine learning model ML can also be formed, in particular trained, to detect a violation of a specification of a specific parameter from the real-time features F and to trigger an alarm. For example, an alarm can be triggered when a parameter estimated by the machine learning model leaves a predetermined range. The later use Figure 5 Details of such a parameter classification are explained.

[0083] The machine learning model ML can operate in a training or retraining mode in which it employs the historical features F'. Optionally, in the training or retraining mode, the machine learning model ML can additionally (e.g. by means of unsupervised learning) employ the real-time features F.

[0084] The machine learning model ML can also operate in a real-time estimation (or prediction) mode in which it determines, predicts and / or classifies one or more quality parameters based on the real-time features F.

[0085] The machine learning model ML is designed to interact with a user interface UI of the software product and / or with a database DB designed to continuously store data and / or with an alarm system AS designed to emit a signal indicating that a human intervention is required in the processing line.

[0086] Figure 3 Extraction of features F of sensor data D3 associated with group delay T3 performed by pre-processing module PM (not explicitly shown in Figure 3 ) is schematically and exemplarily shown.

[0087] Sensor data D3 is acquired along a data history AT. Sensor data D3 that is more recent than the associated group delay T3 is not considered for pre-processing, as it is a priori known that this cannot influence (or predict) the quality parameter observed at production timestamp t0.

[0088] The remaining data history AT is divided into time windows AT i , i = 1, 2, 3,.... One or more statistical parameters are determined for each of the time windows AT1, AT2, AT3. Merely as an example, the mean value Figure 3 is schematically shown in which is derived from the time course of sensor data D3 in the respective time window AT1, AT2, AT3. Additionally or alternatively, each time window AT1, AT2, AT3 can derive further statistical parameters, including but not limited to standard deviation, higher order moments, or, for multivariate sensor data D3, covariance.

[0089] Instead of current sensor data D3, pre-processing module PM can also process historical sensor data D1' to D5' (not explicitly shown in Figure 3 ). Pre-processing module PM distinguishes between processing current sensor data D3 and processing historical sensor data D1' to D5', current sensor data D3 being provided by data repository R as streaming data.

[0090] For the case of streaming current sensor data D3, a single sample of real-time features F is produced, which can be formed as a feature vector comprising a plurality of scalars (e.g. statistical parameters). Along with real-time features F derived from streaming current sensor data D3, a particular error detected by pre-processing module PM (such as inconsistency) or an event detected along the process line P (such as a shutdown) can be retrieved from data repository R and communicated to machine learning module ML.

[0091] For the case of historical sensor data D1' to D5', such data is divided into a number of intervals to create a large number of samples of historical features F' used to train the machine learning model ML. As an example, historical sensor data D1' to D5' can be divided into 10000 intervals, which can be disjoint or partially overlapping. When segmenting historical sensor data D1' to D5', events including errors such as inconsistencies or irregular events such as shutdown or partial shutdown of the process line can be excluded from the generated intervals.

[0092] Furthermore, together with historical sensor data D1' to D5', ground truth data describing the quality parameters of the actual verification is retrieved from the data repository R and transferred to the machine learning module ML, enabling supervised learning.

[0093] Features F, F' can be formed by a single statistical parameter or a plurality of statistical parameters. By statistical parameters, the dimensionality of the observed sensor data D1 to D5, D1' to D5' can be reduced. Since statistical data representations are less sensitive to outliers, the robustness of the statistical data representations can be improved. Real-time features F are extracted from current sensor data D1 to D5. Historical features F' are extracted from historical sensor data D1' to D5'. All extracted features F, F' are converted into a format compatible with the input of the machine learning module ML.

[0094] Figure 4 The machine learning model ML is explained in more detail. The machine learning model ML is a mathematical model learned from processing historical sensor data D1' to D5' in a supervised learning phase. After the supervised learning phase, the trained machine learning model ML is used to estimate at least one product quality parameter.

[0095] The supervised learning is performed by minimizing a cost function, which has to be defined such that it presents a minimum value when the deviation between the estimated and the measured product quality parameter (typically defined by a distance metric such as Euclidian or Manhattan distance) is equal to zero or minimized. The minimization of the cost function is achieved by applying a stochastic gradient descent, which is a class of methods that iteratively adapt the randomly initialized parameters of the model until a certain accuracy is reached. Stochastic gradient descent methods are known from the prior art.

[0096] As an example, the machine learning model ML can be formed as a multi-layer artificial neural network comprising at least one one-dimensional (ID) convolutional layer CL as an input layer and a plurality of densely connected neuron layers NL. The layers CL, NL are arranged in a stack of layers, wherein the first (lowest) layer is the input convolutional layer CL and the last (uppermost) layer is the output neuron layer OL. Adjacent neuron layers NL can be densely connected such that each neuron in a respective upper neuron layer NL receives input from all neurons in a respective lower neuron layer NL. The connections between neurons are associated with weights that are free parameters of the artificial neural network ML.

[0097] The uppermost output neuron layer OL comprises at least one neuron that generates a model output value corresponding to a paper quality parameter value.

[0098] In a real-time estimation (or prediction) mode, the at least one model output value forms a prediction or estimation of the paper quality parameter Q.

[0099] In a training or re-training mode, the at least one model output value is compared to at least one ground truth parameter value by means of a distance measure, thereby determining an error value. The free parameters of the artificial neural network ML are adapted such that, in a stochastic sense, the error value is minimized. Supervised learning methods for adapting the free parameters of an artificial neural network ML are known in the art.

[0100] On the input side of the convolutional layer CL, the artificial neural network ML is fed with features F, F' extracted by the pre-processing module PM. These features F, F' can be formed as an input matrix IM, wherein one horizontal row of the input matrix IM corresponds to a particular time-varying parameter of the sensor data D1 to D5, D1' to D5' and wherein one vertical row of the input matrix IM corresponds to a particular time window AT i , i = 1, 2, 3,.... In other words, along its horizontal dimension, the input matrix IM reflects the change over time, while along its vertical dimension, the input matrix captures different signals measured in the process line P.

[0101] Within the ID convolutional layer CL, a convolution is applied along the time axis T, i.e., along the horizontal dimension of the input matrix IM. The ID convolutional layer CL provides a set of filters for each row of the input matrix IM.

[0102] As an example, a first convolutional filter can be provided for the first row of the input matrix IM that maintains features F, F' derived from a first sensor parameter that is part of the first sensor data D1. The first convolutional filter can be configured to access features F, F' from three consecutive time windows AT i-1 , AT i , AT i+1the three consecutive time windows are shifted along the time axis T, i.e. along the row (e.g. from ΔT i-1 , ΔT i , ΔT i+1 first to ΔT i , ΔT i+1 , ΔT i+2 then to ΔT i+1 , ΔT i+2 , ΔT i+3 etc.).

[0103] Similarly to the first convolutional filter, further convolutional filters can be provided for the same first row of the input matrix IM.

[0104] In the same way, a convolutional filter is provided within the 1D convolutional layer CL for each row of the input matrix IM, wherein the support of each convolutional filter is not necessarily limited to one row of the input matrix IM.

[0105] In the training / re-training mode, the behavior of the convolutional filters is adapted according to an error criterion to be minimized. As an example, filter coefficients of the convolutional filters, which form finite impulse response (FIR) filters, are adapted according to the error criterion. To this end, a plurality of instances of the input matrix IM is presented to the artificial neural network ML, wherein for each instance at least one output value is computed at the output layer OL and an error value is determined according to a selected distance measure.

[0106] The initial parameters of the artificial neural network ML, including the parameters of the convolutional filters in its 1D convolutional layers, are randomly initialized. Thus, for a plurality of training / re-training runs according to the prior art, the same result (i.e. the same performance of the trained artificial neural network ML) cannot be guaranteed.

[0107] To improve the stability and robustness of the training result, a plurality of convolutional filters is used, wherein for a certain number of training cycles a certain proportion of the convolutional filters is switched off. This regularization method is known from the prior art as dropout. Thereby, redundancy is retained within the plurality of trained convolutional filters. Thus, the training result is less sensitive to the random initialization and less prone to local minima of the error criterion.

[0108] For example, for a set of approximately 70 different sensors providing parameters (time series), approximately 100 convolutional filters can be used.

[0109] In an embodiment, a set of instances of the input matrix IM can be divided into training data and evaluation data. The training data and the evaluation data can also be chosen to slightly overlap. As an example, the training data can be derived from historical sensor data D1'to D5' from January 2017 to April 2019, while the evaluation data can be derived from historical sensor data D1'to D5' from February 2019 to April 2019, such that only a small subset of the training data is part of the evaluation data. As an advantage, the most recent data is subsequently used to adapt the artificial neural network ML, thereby improving its performance on the current process line P.

[0110] The machine learning model ML' can also be used for classifying the current sensor data D1 to D5 as schematically shown in Fig. 6. The machine learning model ML' can be implemented as a machine learning model of its own or as an extension of the artificial neural network ML according to Fig. 5. Figure 5 Figure 4 The machine learning model ML' can be provided at its input with an input matrix IM as shown in Fig. 6. Alternatively or additionally, the machine learning model ML' can be provided with a paper quality parameter Q estimated by the artificial neural network ML according to Fig. 5, i.e. one or more output values of the output layer OL.

[0111] The machine learning model ML' can be provided at its input with an input matrix IM as shown in Fig. 6. Alternatively or additionally, the machine learning model ML' can be provided with a paper quality parameter Q estimated by the artificial neural network ML according to Fig. 5, i.e. one or more output values of the output layer OL. Figure 4 Figure 4 The machine learning model ML' can be provided at its input with an input matrix IM as shown in Fig. 6. Alternatively or additionally, the machine learning model ML' can be provided with a paper quality parameter Q estimated by the artificial neural network ML according to Fig. 5, i.e. one or more output values of the output layer OL.

[0112] Further, the machine learning model ML' is provided with specification boundary parameters which indicate at least one range r1, r2, r3 for the respective paper quality parameter Q. For example, for a scalar paper quality parameter Q, a lower specification boundary and an upper specification boundary are provided when the scalar paper quality parameter Q is considered to be valid between the lower specification boundary and the upper specification boundary.

[0113] For the historical input data, it can be known whether the respective paper quality parameter Q is within the prescribed range r1, r2, r3, or exceeds the prescribed range r1, r2, r3 or is below the prescribed range r1, r2, r3. Thus, the machine learning model ML' can be trained by supervised learning in the same way as previously described for the artificial neural network ML.

[0114] According to this aspect of the present application, the trained machine learning model ML' determines first, second and third probability values pi, p2, p3. The first probability value pi indicates a probability that the paper quality parameter Q falls below the lower specification boundary, i.e. within the first valid range r1. The second probability value p2 indicates a probability that the paper quality parameter Q is within the second specification range r2, i.e. greater than or equal to the lower specification boundary and less than or equal to the upper specification boundary. The third probability value p3 indicates a probability that the paper quality parameter Q exceeds the upper specification boundary, i.e. within the third range r3. ​​

[0115] According to the probability values p1, p2, p3, an alarm will be triggered when the expected paper quality parameter Q is outside a defined valid range. For example, an alarm will be triggered if at least one of the following conditions is met: the first probability value p1 exceeds 0.1, or the third probability value p3 exceeds 0.1, or the second probability value p2 is below 0.85.

[0116] As an advantage, such an alarm enables to check or possibly adjust the production sub-process PP such that the defined quality parameter range r2 is reached again even if final quality data based on the board sample M2 from the laboratory is not yet available. Thereby, the production amount of board that fails to meet the defined quality standards can be reduced.

[0117] Reference signs

[0118] AS alarm system

[0119] CL 1D convolution layer

[0120] D1 to D5 real-time sensor data, sensor data

[0121] D1' to D5' historical sensor data, sensor data

[0122] average value

[0123] DB database

[0124] DD data-driven model

[0125] F real-time feature

[0126] F' historical feature

[0127] IM input matrix

[0128] Ml wet pulp, material

[0129] M2 board sample, material

[0130] ML machine learning module, artificial neural network

[0131] ML' machine learning module

[0132] NL neuron layer

[0133] OL output layer

[0134] P processing line

[0135] PM pre-processing module

[0136] PP production sub-process

[0137] P1 to P5 processing step

[0138] P1, P2, P3 first, second, third probability value

[0139] Q paper quality parameter

[0140] QA quality assurance sub-process

[0141] R data repository

[0142] r1, r2, r3 range

[0143] T time axis

[0144] T0 production timestamp

[0145] T1 to T5 delay group

[0146] ΔT data history

[0147] ΔT i , i = 1, 2, 3,... time window

[0148] UI user interface

Claims

1. A computer-implemented method for estimating at least one quality parameter of paperboard produced in a paperboard production sub-process of a paperboard processing line by means of a data-driven module, the data-driven module comprising a pre-processing module and a machine learning module, the method comprising: - acquiring sensor data from at least one processing step of the processing line and transferring the sensor data to a data repository, - determining, by the pre-processing module, at least one historical feature by at least partially evaluating historical sensor data acquired during at least one previous production batch of paperboard and retrieved from the data repository, - training the machine learning module to reproduce a target value of the at least one quality parameter from the at least one historical feature, wherein the target value is determined for the previous production batch of paperboard corresponding to the historical sensor data from which the historical feature is determined, - determining, by the pre-processing module, at least one real-time feature from a current sensor data stream acquired from a current production batch of paperboard and retrieved from the data repository, - determining, by the trained machine learning module, an estimate for the at least one quality parameter from the at least one real-time feature and providing the estimate as an output value, - in response to the output value not being within a predetermined range, adjusting the paperboard production sub-process, - wherein at least one range of the at least one quality parameter is provided as an additional input to the data-driven module, wherein for each range a probability value is determined as an output value that the respective quality parameter falls within the respective range, - wherein for one or more probability values determined by the data-driven module for a range of the at least one quality parameter, each probability value is compared to a probability limit assigned to the respective range and the respective quality parameter, resulting in one Boolean comparison value per range and quality parameter, - wherein a predetermined Boolean expression is evaluated that depends on the one Boolean comparison value, wherein an alarm is triggered when the predetermined Boolean expression returns a logical true.

2. The computer-implemented method of claim 1, wherein, The historical sensor data is divided into a plurality of intervals, wherein events of historical sensor data associated with error messages or inconsistencies are omitted.

3. The computer-implemented method of claim 2, wherein, The plurality of intervals into which the historical sensor data is divided comprises 10,000 intervals or more.

4. The computer-implemented method of claim 1, wherein, A delay group is determined for the processing step as a minimum waiting time for a change in a parameter in the processing step to affect the outcome of the production sub-process, wherein current sensor data acquired within the waiting time of the respective delay group backwards from a current production timestamp is excluded from the evaluation by the data-driven module.

5. The computer-implemented method of claim 1, wherein, The at least one feature comprises at least one time series describing a change of at least one parameter of the sensor data along a discrete time window, wherein the machine learning module performs a one-dimensional convolution of the at least one time series along a time axis.

6. The computer-implemented method of claim 4, wherein, For the discrete time window, at least one statistical parameter is determined as a sample value of the at least one time series.

7. The computer-implemented method of claim 6, wherein, The at least one statistical parameter comprises a mean value and / or a standard deviation.

8. The computer-implemented method of claim 1, wherein, The validity and / or consistency of the current sensor data is evaluated by the pre-processing module and optionally logged in the data repository.

9. The computer-implemented method of claim 1, wherein, One quality parameter is the Z-strength of the paperboard.

10. The computer-implemented method of claim 1, wherein, One quality parameter is the Scott bond strength of the paperboard.

11. The computer-implemented method of claim 1, wherein, For said at least one quality parameter, a first range is a set of invalid values below an effective lower limit, a second range is a set of valid values between an effective lower limit and an effective upper limit, and a third range is a set of invalid values above an effective upper limit, wherein an alarm is triggered when a probability value associated with the first range exceeds a predetermined first probability limit, or when a probability value associated with the second range falls below a predetermined second probability limit, or when a probability value associated with the third range exceeds a predetermined third probability limit.

12. A system comprising at least one sensor, a data repository and a computing device, - wherein, the at least one sensor is designed to acquire sensor data in a processing step of a processing line comprising a paperboard production sub-process, - wherein the data repository is designed to receive and continuously store sensor data from the at least one sensor and to transmit the sensor data to the computer, and wherein the computer implements the computer-implemented method according to claim 1.

13. The system of claim 12, wherein, The data repository is formed as a cloud storage accessible by an internet protocol based communication protocol stack.

14. The system of claim 12, further comprising an alert device, wherein, The alarm device is designed to be triggered by the computer and, when triggered, to indicate that a manual interaction with the paperboard production sub-process is required.

15. A paperboard production machine designed for producing paperboard or a semi-finished product of paperboard, the machine comprising the system according to claim 12.

Images