Computer system and method for evaluating physical stream sample properties

Abstract

A computer system (100), computer-implemented method and computer program product for evaluating a property of a physical stream sample is disclosed. An initial state image (ISI1) of said stream sample (S_b) at the beginning of the stream sample processing method (200) is received. A pre-trained neural network (120) determines a sample type (ST1) with a corresponding confidence value (CV1) for said stream sample (S_b). A probabilistic model (140) of the stream sample processing method predicts a target property value (PVt1) of the stream sample. A measuring result (MVe) for an actual property value (PVa1) of the processed stream sample (S_e) at the end of the sample processing method (200) is obtained. The actual property value (PVa1) is mapped backwards through the sequential probabilistic model (140) to the corresponding initial state image (ISI1). The sequential probabilistic model (140) or the neural network (120) or both are updated based on the actual property value.

Description

Computer System and Method for Evaluating Physical Stream Sample Properties Technical Field

[0001] The present invention generally relates to multi-processor systems, and moreparticularly, relates to methods, computer program products and systems for evaluating one or more properties of a physical stream sample. Background

[0002] It is advantageous to evaluate and predict properties of a physical stream sample,which are relevant parameters for optimization of a stream sample processing method used for processing said sample. A stream sample, as used herein, is a physical sample which moves through processing means performing said stream sample processing method.

[0003] US patent 5,121,467 discloses a combined neural network / expert system process anda method combining decision-making capabilities of expert systems with predictive capabilities of neural networks for improved process control. Neural networks provide predictions of measurements which are difficult to make, or supervisory or regulatory control changes which are difficult to implement using classical control techniques. Expert systems make decisions automatically based on knowledge which is well-known and can be expressed in rules or other knowledge representation forms. Sensor and laboratory data is used. In one approach, the output data from the neural network can be used by the controller in controlling the process, and the expert system can make a decision using sensor or lab data to control the controller(s). In another approach, the output data of the neural network can be used by the expert system in making its decision, and control of the processcarried out using lab or sensor data. In another approach, the output data can be used bothto control the process and to make decisions. In all approaches, the system’s neural network parameters (number of neurons, layers, etc.) are adjusted by the calibration of the inputs of the neural network via human input.

[0004] In this prior art solution, predictions are made to avoid or at least reduce physicalmeasurements for improved process control of a production process of low variability. The inputs for the neural network are already known by the system to make predictions of the final processing results. In many industrial process flow methods, the input parameters arenot known for an incoming stream sample, which requires a higher degree of flexibility compared to the prior art system. Summary

[0005] There is therefore a need to improve the prior art systems in that a higher degree offlexibility is achieved for optimizing a stream processing method without any human interaction with regard to stream samples whose properties are unknown when entering the stream processing method. This is achieved by a computer-implemented method, a computer program product and a computer system as described by the independent claims.

[0006] The computer-implemented method is provided for evaluating at least one propertyof a physical stream sample. The at least one property is a relevant parameter for optimization of a stream sample processing method. The stream sample processing method is processing said stream sample while continuously moving through a process system or continuous flow system which transforms the incoming stream sample into a different state. Examples of a stream sample can include waste material, waste water, etc. In some examples, a stream sample comprises a mix of materials which is processed according to a particular recipe. Thereby, the initial stream sample and its properties are unknown at thebeginning of the stream processing method. When the processing of the stream samplestarts, an image sensor captures images of the incoming stream sample. Such images representing the initial state of the stream sample are referred to as initial state imagesherein. In case of large amounts of material in a stream sample, multiple initial state imagesmay be captured and tagged such that these images are associated with the same sample source.

[0007] An initial state image is the received by a computer system which is adapted toexecute the herein disclosed computer-implemented method. The initial state image is characteristic for at least one property of said stream sample at the beginning of the stream sample processing method. For example, in case of a stream of waste material, the image of the received material is somehow related to a calorific value of the waste. This calorific value is however unknown at the time of receiving said initial state image. In general, properties of the physical real-world sample can be of any of the following types: physical, chemical, biological, mechanical or sensory characteristics of said sample.

[0008] The system uses a respectively pre-trained neural network to determine, based onthe initial state image, a sample type with a corresponding confidence value for said stream sample (e.g., household waste with a confidence of 80%). In one implementation, the neural network is a classification neural network and the sample type is determined according to predefined sample type classes associated with said stream sample. In another implementation, the neural network is a regression neural network and the sample type is determined as regression value.

[0009] The system further uses a sequential probabilistic model of the stream sampleprocessing method. Advantageously, the sequential probabilistic model describes the processing method as a sequence of possible events in which the probability of each event depends on the state attained in the previous event. For example, the sequential probabilistic model can be a State-Space Model or a Markov chain in which the probability of each event depends only on the state attained in the previous event. Based on the sample type, the probabilistic model predicts at least one target property value of the stream sample (e.g., the calorific value of the household waste or required temperature of combustion chamber). Further, the probabilistic model is conditioned on one or more observable control parameters which control the stream sample processing method (e.g., waste processing) in accordance with the at least one property target value. The controlparameters can be retrieved from the system (processing means) which performs the streamprocessing method. Based on such control parameters, conditional probabilities of the probabilistic model having an impact on the at least one target property value can be generated.

[0010] The at least one property target value is then written into a sample specific datastructure. The data structure may be an integral part of the computer system or it may be a remote system which is communicatively coupled with the computer system. In one implementation, the sample specific data structure may be a sample specific blockchain. Using a blockchain ensures that any modification to the sample specific data set stored in the blockchain is traceable in a reliable, secure and transparent manner.

[0011] The one or more control parameters for controlling the stream sample processingmethod are provided to respective control components used for controlling the streamprocessing method. For example, control parameters can control cranes or robots which are used to change the mix of materials in said stream sample during its processing.

[0012] At the end of the sample processing method (i.e., after the stream sample has beenprocessed), a measuring result is received for at least one actual property value of the processed stream sample. The at least one actual property value relates to the at least one predicted target property value. For example, if the target property value is a predicted calorific value of the initial stream sample, at the end of the stream processing, one or more sensors obtain data from the processed stream sample (whose state is different from the initial state) which are suitable for determining the actual calorific value of the processed stream sample.

[0013] The one or more actual property values are then mapped backwards through thesequential probabilistic model to the corresponding initial state image. For example, a back- path-operation object may be generated which incorporating details like timestamp, processing cell unit, probability, calorific value, and learning rate. Once the actual property value(s) are mapped backwards through the probabilistic model, they are associated with the respective initial state image. The at least one actual property value is then written to the sample specific data structure as new entry.

[0014] Finally, the system updates the sequential probabilistic model or the neural networkor both - the sequential probabilistic model and the neural network.

[0015] Updating the sequential probabilistic model includes updating conditional probabilitydistributions of the sequential probabilistic model based on the at least one actual property value taking into account uncertainties in the sample processing method. Updating the neural network comprises initiating automatic re-training of the neural network with additional training data for adjusting the weights of the neural network. Thereby, the additional training data are pairs of initial state images and respective actual property values collected from earlier stream samples of the determined sample type. For example, for the mapped initial state image (e.g., image of incoming household waste), the respective actual property value is associated with a label value representing the ground truth for a retraining of the neural network.

[0016] An automatic re-training of the neural network may be initiated in case the numberof collected training data pairs exceeds a predefined threshold.

[0017] Updating may depend on the confidence value provided by the neural network forthe determined sample type. For example, in case the confidence value is below a predefined minimum threshold, only the neural network is updated. In case the confidence value is above a predefined maximum threshold, only the sequential probabilistic model is updated. In case the confidence value is in a range from the predefined minimum threshold to the predefined maximum threshold, the neural network and the sequential probabilistic model are updated.

[0018] In one embodiment, a computer program product is provided for evaluating at leastone property of a physical stream sample. The computer program product comprises computer-readable instructions that can be loaded into a memory of a computing device and executed by one or more processors of said computing device. The computer instructions, when being executed by said processors, cause the computing device to perform the herein disclosed computer-implemented method.

[0019] In one embodiment, a computer system is provided for evaluating at least oneproperty of a physical stream sample. The computer system has functional modules implemented by software modules which are adapted to perform the herein disclosed computer-implemented method at runtime. In more detail, the at least one property is a relevant parameter for optimization of a stream sample processing method, wherein the stream sample continuously moves through said stream sample processing method.

[0020] The system has a first interface adapted for receiving an initial state image beingcharacteristic for the at least one property of said stream sample at the beginning of the stream sample processing method.

[0021] Further, a respectively pre-trained neural network of the system is adapted todetermine, based on the initial state image, a sample type with a corresponding confidence value (CV1) for said stream sample. The neural network may be a classification neural network which determines the sample type to predefined sample type classes associated with said stream sample. Alternatively, the neural network may be a regression neural network which determines the sample type as regression value.

[0022] Further, the system has a sequential probabilistic model of the stream sampleprocessing method adapted to predict, based on the sample type, at least one target property value of the stream sample. The probabilistic model is conditioned on one or more observable control parameters which control the stream sample processing method. In the previously mentioned example of a waste processing method, the conditional probabilities of the probabilistic model can be conditioned on the crane positions which change during the movement of a respective crane. Thereby, the crane positions are observable control parameters.

[0023] A second interface of the system is adapted to write the at least one property targetvalue into a sample specific data structure, and to provide the one or more control parameters (e.g., retrieved crane parameters) for controlling the stream sample processing method.

[0024] A third interface of the system is adapted to receive, a measuring result for at leastone actual property value of the processed stream sample at the end of the sample processing method. The at least one actual property value relates to the at least one predicted target property value.

[0025] A mapper module of the system is adapted to map the at least one actual propertyvalue backwards through the sequential probabilistic model to the corresponding initial state image.

[0026] The second interface is further adapted to write the at least one actual propertyvalue as new entry to the sample specific data structure.

[0027] An updater module of the system is adapted to update the sequential probabilisticmodel or the neural network or both. Thereby, updating the sequential probabilistic model comprises updating conditional probability distributions of the sequential probabilistic model based on the at least one actual property value taking into account uncertainties in the sample processing method. Updating the neural network comprises initiating automatic re-training of the neural network with additional training data for adjusting the weights of the neural network, wherein the additional training data are pairs of initial state images and respective actual property values collected from earlier stream samples of the determined sample type.

[0028] The updater module may be adapted to initiate automatic re-training of the neuralnetwork in case the number of collected training data pairs exceeds a predefined threshold. Further, the updater module may depend on the confidence value provided by the neural network for the determined sample type in that:- In case the confidence value is below a predefined minimum threshold, only the neuralnetwork is updated;- In case the confidence value is above a predefined maximum threshold, only thesequential probabilistic model is updated; and- In case the confidence value is in a range from the predefined minimum threshold to thepredefined maximum threshold, the neural network and the sequential probabilistic model are updated.

[0029] Further aspects of the invention will be realized and attained by means of theelements and combinations particularly depicted in the appended claims. It is to be understood that both, the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention as described. Short description of the figures

[0030] FIG.1 shows a block diagram of a computer system for evaluating at least one property of a physical stream sample in accordance with an embodiment; FIG.2 is a simplified flowchart of a computer-implemented method for evaluating at least one property of a physical stream sample in accordance with an embodiment; FIG.3 is a simplified illustration of an embodiment of a waste example scenario with a stream processing method adapted to process incoming waste units; FIG.4A illustrates pseudo code for an algorithm to update model states based on actual values obtained from true measurements according to an example embodiment; FIG.4B illustrates pseudo code for an algorithm to retrain models according to an example embodiment;FIG.4C illustrates pseudo code for a generalized algorithm to update model states based on actual values obtained from true measurements according to an example embodiment;FIG. 4D illustrates pseudo code for a generalized algorithm to retrain models according to anexample embodiment; and FIG.5 is a diagram that shows an example of a generic computer device and a generic mobile computer device, which may be used with the techniques described herein. Detailed description

[0031] FIG. 1 shows a block diagram of an example embodiment of a computer system 100for evaluating at least one property of a physical stream sample. FIG.2 is a simplified flowchart of an exemplary computer-implemented method 1000 for evaluating at least one property of said physical stream sample. Method 1000 may be executed by system 100. Therefore, the following description of the system 100 in FIG.1 is provided in view of the method 1000 of FIG.2 and reference numbers of both figures are referred to.

[0032] The example discussed in detail in the following is based on a scenario where thestream sample consists of waste wherein the incoming waste can be of different waste types. A person skilled in the art is able to transfer the disclosure of the herein discussed waste example to other domains (i.e., other stream sample types). For example, the stream sample may be waste water, chemical effluent, air pollution, slurry, liquid waste. However, the herein disclosed conceptual approach is applicable to any physical stream sample processed by processing means which can perform respective stream sample processing methods. In the example, an incoming stream sample S_b is received in its original state as an input to stream sample processing method (SPM) 200. The stream sample continuously moves through SMP 200 while being processed. The physical stream sample has one or more properties with at least one property being a relevant parameter for optimization. In the waste example, the received stream sample is associated with a respective calorific value. SPM 200 can be adapted to optimize this calorific value. For this purpose, the incoming waste S_b may be received, e.g., in a so-called waste bunker, and is then moved through arespective waste processing apparatus which is adapted to process the waste and eventuallytransform the originally received stream sample S_b into one or more transformed states (stream sample(s) S_i). The goal of this processing may be an increased calorific value of the stream sample S_e at the end of SPM 200.

[0033] An image sensor 210 (e.g., a RGB camera, depth sensor, thermal imager,multispectral or hyperspectral imager, ultraviolet imager, monochrome camera, polarimetric imager, or X-ray sensor) is placed such that the sensor 210 captures an initial state image ISI1 of the received stream sample S_b before being processed. This initial state image is characteristic for the at least one property of the stream sample S_b in its original state (i.e.,before being processed by SPM 200). For example, a camera sensor may be mounted abovethe location where the incoming waste is unloaded, and can capture initial state images of the incoming waste from a bird’s eye view. System 100 is communicatively coupled with image sensor 210 via a first interface 110 which is adapted to receive 1100 the initial state image ISI1 from sensor 210.

[0034] In a real-world waste processing method, waste of the same type is typicallydelivered by a plurality of trucks which may even arrive at different gates. In this scenario, multiple initial state images captured at respective gates (gate images) with the same type of waste delivered by different trucks may be assigned by system 100 to a common “unloaded object” until the last truck of the plurality of trucks has been unloaded.

[0035] System 100 has a pre-trained neural network (NN) 120 adapted to determine 1200,based on a particular received initial state image ISI1, a sample type ST1 with a corresponding confidence value CV1 for said stream sample S_b. In the example, each gateimage is classified by NN 120. NN 120 can be pre-trained by using an expert-labeled initialtraining set such that each gate Image is classified into, for example, five waste classes (e.g.,household waste, waste wood, scrap metal, packaging waste, construction waste). Theneural network may be implemented as a classification neural network. The sample type is then determined according to predefined sample type classes associated with said stream sample. The sample type classes may be text strings (e.g., household waste, construction waste, etc.) but can also be numerical values (e.g., ‘1’ for household waste, ‘2’ for construction waste, etc.). In an alternative implementation, the neural network may be a regression neural network and the sample type is determined as regression value. In this implementation, a sample type may be associated with an interval of regression values (e.g., [0.50|1.49] for household waste, [1.50|2.49] for construction waste, etc.)

[0036] In the multiple truck scenario, the class of the whole truckload (i.e., an overall class ofthe multiple truck loads) can be determined by the system using a majority vote.

[0037] In the multi-truck example, each truckload classification may create a new wasteunit. In other words, each truckload can be considered as a new waste unit. The new waste unit is the classified with a respective waste class. In this example, a waste unit corresponds to the stream sample S_b which is moved through the waste processing system, and is defined by the properties of its state. The waste unit has a property “calorific value” with an initial value. The waste unit’s calorific value may be modeled as a Gaussian distribution: N(μ, σ2) where μ is the mean calorific value estimated indirectly by the waste class, and σ2is the variance.

[0038] The output of NN 120 – the determined waste class value ST1 in the example – isprovided to a sequential probabilistic model (PM) 140 of the stream sample processing method 200. SPM 140 of system 100 is adapted to predict 1300, based on the sample type ST1, at least one target property value PVt1 of the stream sample, and observe one or more control parameters CP* controlling the stream sample processing method 200 in accordancewith the at least one property target value PVt1. In the example, each class is associatedwith a calorific value Ci assigned to it by experts. For example, this value can be the mean of a Gaussian distribution. It is assumed that the determined waste class value corresponds to the sample type household waste. PM 140 may translate the sample type ST1 into a Gaussian distribution with a predefined mean value for the initial calorific value as characteristic property value associated with the respective initial state image ISI1. System 100 writes 1400 the at least one property target value PVt1 into a sample specific data structure 300, and provides 1500 the one or more control parameters CP* for controlling the stream sample processing method 200. System 100 may use a second interface (not shown) adapted to communicate with control elements of SPM 200 and the sample specific data structure 300. The second interface may also be the same as the first interface 110. The data structure 300 may be an integrated component of system 100 or it may be a remote data structure accessible by system 100. In the example scenario, control parameters CP* may include instructions for executing pick and drop operations by respective cranes during waste processing to achieve the property target value PVt1 (a target calorific value) at the end of SPM 200.

[0039] PM 140 can describe the processing method as a sequence of possible events inwhich the probability of each event depends on the state attained in the previous event. Forexample, PM 140 can be a Markov chain in which the probability of each event depends only on the state attained in the previous event. In another example as demonstrated in more detail for the waste example scenario, PM 140 may be implemented as a State-Space-Model (SSM) to model a forward path with the waste units’ calorific values as hidden states (i.e., states that cannot be directly observed). In this example of the stream sample processing method being be a waste processing method which can be performed by a waste incineration plant, the control parameters can be movements performed by cranes in the plant. The probabilities of the probabilistic model are then conditioned on such crane movements. Thereby, the crane movements can be directly retrieved from the plant. That is, the crane movements are given by the plant and are used to generate the conditional probabilities (of the probabilistic model) which represent which waste units are affected by the crane movements and how they affect respective calorific values. In more detail, in the SSM implementation, each waste unit Wi at time t is characterized by the following state:- Calorific value mean (μi,t): The average calorific value of the waste unit i at time t.- Calorific value variance (σ2i,t): The variance of the calorific value of the waste unit i attime t.- Physical mass in kg (Mi,t): The physical mass of the waste unit i in kilograms at time t.- Position (xi,t, yi,t): The coordinates of the waste unit i within the processing means at timet.

[0040] Thereby, the initial state distribution describes the probability P of each waste unitstarting in a given state. This probability P is derived from the Neural Network 120 that classifies an image of the waste from the truck into one of for example five predefined classes. Each of these classes has an associated initial mean and variance for the new waste unit’s calorific value. NN 120 processes an image from the truck and outputs probabilities for each of the five classes. Each class si has an associated initial mean μi,0 and variance σ2i,0 for the calorific value: P(si | image) = NN(image)i Where:- NN(image)i: The probability output by the Neural Network that the waste unit belongs toclass i given the image.- μi,0: Initial mean calorific value for class i.- σ2i,0: Initial variance of the calorific value for class i.After classification, each new waste unit is assigned an initial state Si,0based on the class probabilities:Where:- Si,0: Initial state of waste unit i.- μi,0: Initial mean calorific value.- σ2i,0: Initial variance of the calorific value.- Mi,0: Initial mass of the waste unit.- (xi,0, yi,0): Initial position of the waste unit.

[0041] The crane’s pick and drop operations are observable events. The observationsinclude a crane’s operations and occasional true measurements. A pick operation is specified by the crane’s position (xp, yp) and the amount of weight picked, denoted by Mpick. A drop operation is specified by the crane’s position (xd, yd) and the amount of weight dropped, denoted by Mdrop. For example, sensors Si1 to Sik may be associated with respective cranes and provide measurement values MVi (pick, drop, location, weight) which characterize corresponding intermediate states of the stream sample S_i during waste processing. For each pick and drop operation, there is a probability distribution describing how the intermediate state of affected waste units S_i changes, based on the location and weight of a respective crane’s sensor. Every pick and drop operation may also be stored in the sample specific data structure 300 (e.g., a blockchain storage) for integrity reasons.

[0042] The SSM forward path model allows to track the waste unit S_i through the entirewaste-to-energy conversion process via the pick and drop operation as implemented in SPM 200. The following notation is used further down below:- μcrane: Mean calorific value of the waste being picked or dropped by the crane.- σ2crane: Variance of the calorific value of the waste being picked or dropped by the crane.- μoperation: Combined mean calorific value after a pick or drop operation.- Σi: Covariance matrix describing the spread and correlation between μi,t+1, σ2i,t+1, andMi,t+1.- μtrue: True mean calorific value obtained from burning the waste unit.

[0043] When a crane executes a pick operation and returns the waste that is picked, system100 tracks the weight of the waste amount in the crane (e.g., measured value MVi). The crane then performs a drop operation and updates the states of the waste units. When different types of waste from different waste units are combined via respective pick and drop operations, the influence of the crane operations on surrounding waste units can becalculated by using a 2D Gaussian distribution model. Thereby, G in formula F1 correspondsto the probability that a waste unit is included in a pick or drop operation:Where:- (xi, yi): Coordinates of a waste unit.- (xc, yc): Coordinates of the crane’s position.- σx and σy: Standard deviations representing the spread of the crane’s influence.

[0044] The state transition model of the SSM in formula F2 describes how the state of eachwaste unit evolves over time due to crane operations. For each waste unit Wi, the state transition due to a pick operation can be modeled as follows:Where:- Σj G(xj , yj | xp, yp)μj,t: The linear combination of the mean calorific values of all affectedwaste units j.- Σj G(xj , yj | xp, yp)(σ2j,t + (μj,t − μpick)2): The linear combination of the variances of allaffected waste units j.- G(xj , yj | xp, yp)Mpick: The Gaussian influence factor for the pick operation centered at (xp,yp).- Σi: Covariance matrix for waste unit i.

[0045] For each waste unit Wi, the state transition due to a drop operation can be modeledas described in formula F3:Where:- Σj G(xj , yj | xd, yd)μj,t: The linear combination of the mean calorific values of all affectedwaste units j.- Σj G(xj , yj | xd, yd)(σ2j,t + (μj,t – μdrop)2): The linear combination of the variances of allaffected waste units j.- G(xj , yj | xd, yd)Mdrop: The Gaussian influence factor for the pick operation centered at(xd, yd).- Σi: Covariance matrix for waste unit i.

[0046] It is to be noted that the state of the system in the SSM at time t is the collection ofstates of all individual waste units: St= {Si,t| i = 1, ... ,N}, where N is the total number of waste units.

[0047] At the end of SPM 200, one or more further sensors Se1 to Sem are used to obtainone or more measuring results MVe for at least one actual property value PVa1 of the nowprocessed stream sample S_e. The measuring result(s) MVe characterizes the final state ofthe stream sample S_e. Thereby, the at least one actual property value PVa1 relates to the at least one predicted target property value PVt1. In other words, actual property value and target property value are comparable. In the waste example scenario, the actual propertyvalue is the calorific value which can be derived from the energy output of a waste incineration plant at the final stage of SPM 200. The calorific value C is calculated as: C = E / m where E is the measured energy output and m is the measured mass of the waste. That is, the true calorific value measurement of a waste unit when it gets burned during the burning process also belongs to the observations of the SSM. System 100 receives 1600 such measuring results MVe and determines the actual calorific value PVa1. The measuring result is received via a third interface (not shown) which can also be the same as the first interface 110. The at least one actual property value PVa1 is then written 1800 as new entry to the sample specific data structure 300. At this stage, sample specific data structure 300 has a record history characterizing the incoming waste unit S_b, the intermediate waste unit(s) S_i and the final waste unit S_e where all intermediate steps to get from S_b to S_e are modeled by PM (SSM) 140.

[0048] A mapper module 130 (illustrated by dash dotted arrows) maps 1700 the at least oneactual property value PVa1 backwards through the sequential probabilistic model 140 to the corresponding initial state image ISI1. By tracking the pick and drop operations for eachwaste unit, the waste units in the hopper (of the waste incinerator), which contributed to themeasured energy output, can be identified. For each waste unit, the sequence of pick and drop operations can be reversed to find the state sequence z = (s1, s2, …, sT), that is most likely to produce the actual target value, by using an iterative Kalman gain algorithm, as illustrated in FIG., 4A. This allows to map the actual property values from the hopper back to the initial state images of the initial waste units.

[0049] When a waste unit is burned, its true calorific value mean, μtrue is obtained. Thisinformation is used to update the states of the affected waste units, thereby improving the estimates of their respective states. This process involves smoothing, where past states are adjusted based on the new information to maximize the likelihood of the observed outcome. Thereby, states are updated based on the true measurements (actual measurement values).

[0050] The algorithm 1 in FIG. 4A illustrates in pseudo code 401 how such updates areachieved. Given a true (actual) measurement μtrue from burning a waste unit, the affected waste units’ states are updated iteratively, starting from the time they were created (e.g., when new trash is unloaded by trucks). The weight / proportion of each contributing wasteunit in the final burned waste unit is calculated using the 2D Gaussian distribution (cf. above description for the Gaussian Influence Calculation). Where:- μtrue: True calorific value obtained from burning the waste unit.- burned waste: The waste unit that has been burned.- contributing units: The waste units that contributed to the burned waste unit.- μest: Estimated mean calorific value of the contributing waste units.- ωj : The weight / proportion of waste unit j in the final burned waste unit, calculated usingthe 2D Gaussian distribution centered at the crane’s position.- Kj : Kalman gain, which determines the weight given to the new measurement.- Wj.μt+1: Updated mean calorific value of waste unit j.- Wj.σ2t+1: Updated variance of the calorific value of waste unit j.- Wj.Mt+1: Updated mass of waste unit j.

[0051] An updater module 150 (illustrated by dotted arrows) of system 100 finally updates1900 the sequential probabilistic model 140 or the neural network 120 or both. The decision of what to update can depend on the confidence value CV1 provided by NN 120 for the determined sample type ST1 or can be based on the discrepancy between estimated and actual target value. For example, in case the confidence value is below a predefined minimum threshold, only the neural network may be updated. In case the confidence value is above a predefined maximum threshold, only the sequential probabilistic model may be updated. In case the confidence value is in a range from the predefined minimum thresholdto the predefined maximum threshold, the neural network and the sequential probabilisticmodel may be updated.

[0052] FIG. 4B illustrates pseudo code 402 for an algorithm 2 which implements the aboveupdating approach including the retraining the neural network. One aim of algorithm 2 is to improve the accuracy of the Neural Network, wherein the update is based on the true calorific values obtained through the iterative Kalman gain smoothing process. If thediscrepancy between the estimated mean calorific value corresponding to the sample type ST1 provided by the Neural Network and the updated value is above a certain threshold, anew training dataset for retraining is created. For example, a lookup table may be used toretrieve such mean value which corresponds to the respective sample type. If it is below the threshold, the associated initial mean and variance is adjusted for the respective class. Where:- μtrue: True calorific value obtained from the final state update.- classtrue: The class corresponding to the true calorific value.- μclass: Mean calorific value for the respective class, looked up from a table.- σ2class: Variance for the respective class, looked up from a table.- image: Image of the waste unit used for classification.- threshold: The threshold value for determining whether to retrain the Neural Network.- α: Learning rate, an optional parameter set externally, typically between 0 and 1.If the discrepancy δ is below the threshold, the class parameters can be updated using a fixed learning rate α as follows:- Update the class mean: μclass ← (1 − α)μclass + αμtrue- Update the class variance: σ2class ← (1 − α)σ2class + α(μtrue − μclass)2

[0053] The updater 150 may further apply predefined restrictions to ensure accuracy of theadjustment (e.g., a waste unit can never be off by more than 2000 kcal). Such predefined restrictions are typically set by experts in the respective application field of SPM 200.

[0054] The updater may also update surrounding waste units (cf., row 2 in algorithm 1 ofFIG.4A). A surrounding waste unit is a waste unit which is in physical proximity to a previously processed waste unit. Thereby, the surrounding waste units are iteratively updated in the crane operation’s location using the probability of being picked or dropped as described by the 2D Gaussian Distribution for Crane Influence of formula F1.

[0055] The entire SPM 200 can thus be traced back to the initial waste units (e.g., truck loadsreceived at respective gates) and update the corresponding gate images with new labels.Thereby, a new dataset with updated labels and gate images of the initial waste units iscreated which can serve as additional training data for NN 120.

[0056] Once the number of such additional training data pairs exceeds a predefinedthreshold, the updater module 140 can initiate automatic re-training of NN 120 and thereby update weights of the NN model, for example, in accordance with algorithm 2 in FIG.4B.

[0057] In other words, when updating 1920 NN 120, the updater module 150 initiatesautomatic re-training of the neural network with additional training data (ISI*|PVa*) for adjusting the weights of the neural network. Thereby, the additional training data are pairs of initial state images ISI* and respective actual property values PVa* (waste types associated with the actual calorific values) collected from earlier stream samples of the determined sample type ST1.

[0058] FIG. 3 is a simplified illustration of an embodiment of the waste example scenariowith a stream processing method SPM 390 adapted to process an incoming waste unit 302 (delivered by truck 301) by applying to the incoming waste unit pick operations P and drop operations D via crane 303, and finally burning the processed waste unit in a waste incineration unit 304 where the energy output 305 and the mass of the processed waste unit are provided as measurement results to determine the actual calorific value CVa of the processed waste unit.

[0059] Camera sensor 307 captures the initial state image 308 of the incoming waste unit302 which is provided to pretrained neural network NN 320 to classify image 308 by waste types WT1 to WT5. In the example, NN 320 is a regression neural network where the outputs are numerical values. Each waste type is associated with an interval of output values(illustrated by a pair of up- and down-arrows) and a mean value (illustrated by a dashedhorizontal line) representing the calorific value associated with the respective waste type. In the example, image 308 is classified as WT1 (indicated by dotted background pattern).

[0060] Based on waste type WT1, the sequential probabilistic model SSM 340 predicts atarget calorific value CVt for waste unit 302 at the end of SPM 390 processing. After SPM 390 processing, the actual calorific value CVa for the processed waste unit is processedbackwards by SSM 340 as described in detail above. As a result, the output (regression) value for the waste type WT1 determined for image 308 can be adjusted in accordance with the actual calorific value CVa. The adjusted output value is the ground truth of a new training data pair (adjusted output|image 308) for retraining NN 320.

[0061] It is to be noted, that the herein disclosed approach is entirely self-controlled. Nomanual adjustments by any operator needs to made. The system automatically adapts to any changes applied during the stream sample processing as these changes are affecting the measured actual property values. The actual property values are then mapped back to the initial state image of the respective incoming stream sample through a backward path of the probabilistic model. The probabilistic model and the neural network are then automatically adjusted to match the actual property values for the processed stream sample. After such adjustment, which may include the retraining of the neural network, a newly received stream sample which may be very similar to an earlier received stream sample, can be classified into a sample type which is different from the sample type determined for the earlier received stream sample.

[0062] As already mentioned earlier, the herein disclosed approach for evaluating propertiesof stream samples is applicable to any physical sample which moves through processing means performing a respective stream sample processing method. The previously example described in detail is directed evaluating the calorific value of waste streams of waste segments in Waste-to-Energy application scenarios. To demonstrate the broad applicability of the herein disclosed and claimed concept, further application scenarios are described in the following. The additional examples include example scenarios in the field of wastewater treatment, hydropower, chemical processing and air pollution control. The concept enables comprehensive tracking and optimization of stream sample units, even when direct measurements are sparse. The same principles apply across a wide range of domains.

[0063] In many industrial and environmental processes, key target properties of streamsample units are of great importance — for example:- the calorific value of waste streams in Waste-to-Energy,- the pollutant concentration in water or air treatment,- the chemical composition in chemical processing, or- the energy efficiency of hydropower turbine flow.

[0064] However, these target properties are typically not directly observable during moststages of the process. At system entry, only indirect features (such as images, spectra, or sensor data) are available. True measurements of the target property only become availableoccasionally and typically after significant processing — for example:- as shown in detail in the previously described Waste-to-Energy example when a wastesegment is burned and calorific value is measured, or- when treated water or air is sampled and pollutant concentration is analyzed, or- when the chemical yield of a batch is tested.

[0065] Despite this partial observability, it is possible to model both the forward evolution ofthe target property of each stream sample unit and to perform a backward correction when true measurements of actual property values become available. This enables the system to:- Maintain a probabilistic belief about the target property of each stream sample unit as itflows through the process.- Update these beliefs continuously via a state-space model, driven by known processoperations and control parameters.- Perform a Kalman update when true measurements of the target property becomeavailable, retrospectively correcting prior state estimates and improving future predictions.- Learn and improve the initial state estimation model (typically a Neural Network), whichassigns an initial probabilistic belief about each stream sample unit based on indirect observations (e.g., initial state image) at system entry.- Represent and monitor the global state of the system — that is, the collection of allindividual stream sample unit states at any time.

[0066] In the additional example scenarios, different models of different stream sampleprocessing methods are described using different control parameters and different target properties. For example:- In a wastewater treatment scenario, the pollutant concentration of water segments ismodeled.- In a Hydropower scenario: the flow quality or energy efficiency of turbine flow segmentsis modeled.- In a chemical processing scenario, the concentration or yield of chemical batch units ismodeled.- In an air pollution control scenario, the pollutant concentration of air flow segments ismodeled.

[0067] The herein disclosed general and flexible state-space modeling approach is applicableto all these scenarios. In short, an initial state estimation model is used to assign probabilistic beliefs at system entry (the pre-trained neural network determining, based on an initial state image, a sample type with a corresponding confidence value for a stream sample). It is to be noted that even a randomly initialized neural network can be used. However, when using a randomly initialized state estimation model, the system would take a while (i.e., multiple iterations) until it becomes stable in that it provides accurate and reliable predictions. That is, pre-training of the model with labelled training data shortens the initial calibration process of the system. Stream sample units are represented as probabilistic states. A state transition model governs forward evolution through process operations (a sequential probabilistic model of the stream sample processing method predicts at least one target property value of the stream sample, wherein the probabilistic model is conditioned on one or more control parameters which control the stream sample processing method). A Kalman update is used for backward correction based on true measurements (a measuring result for at least one actual property value of the processed stream sample is received at the end of the sample processing method, wherein the at least one actual property value relates to the at least one predicted target property value. The at least one actual property value is mapped backwards through the sequential probabilistic model to the corresponding initialstate image). An iterative refinement of the state estimation model is performed based oncorrected outcomes (the sequential probabilistic model or the neural network (120) or both are finally updated accordingly).

[0068] As each stream sample unit flows through the process (the respective stream sampleprocessing method), its true target property remains unobservable in most stages. However, the system uses the sequential probabilistic model to model the forward evolution of the target property distribution by tracking how process operations affect each unit over time.

[0069] In one implementation, the sequential probabilistic model can be implemented by astate-space model, where each stream sample unit is represented by a probabilistic state vector. The state vector evolves step by step, driven by the known sequence of process operations and associated control parameters.

[0070] The forward path enables the system to:- Propagate a probabilistic belief about each unit’s target property throughout the process- even in the absence of direct measurements.- Explicitly model the effect of process operations such as flow control, pick / drop, mixing,heating, or chemical reactions.- Support later backward correction, when occasional true measurements of the targetproperty become available.

[0071] The same forward modeling structure applies across domains (e.g., from Waste-to-Energy to hydropower, chemical processing, wastewater treatment, and air pollution control).

[0072] A person skilled in the art can easily adapt the formulas describing staterepresentations for stream sample units presented above in the detailed description with regard to the Waste-to-Energy example scenario to other domains and properties already mentioned in the above summary section.

[0073] The evolution of units of a stream sample through a process including multiplecontrolled operations is modeled such that each unit Ui of the stream sample at time t is characterized by a joint state vector:where:- μi,t: True mean of the target property distribution of unit Ui at time t.- σ2i,t: True variance (or other suitable parameterization of dispersion) of the targetproperty distribution of unit Ui at time t, representing the physical heterogeneity of the unit.- Qi,t: Physical quantity of the stream unit, such as mass, volume, flow rate, or otherrelevant domain-specific quantity.- yi,t): Position in physical space or process space, depending on the domain.- Other properties: Additional optional properties depending on the domain (e.g.,timestamp, source identifier, batch identifier, process stage).

[0074] The target property of each unit Ui is represented as a probability distribution whoseparameters evolve over time as the process operates on the unit. For example, in many applications, the target property is modeled as a Gaussian distribution:due to the central limit effect and practical tractability. However, other distributions (e.g., log-normal, beta, empirical distributions) can be used depending on the nature of the process and target property. The specific meaning of the target property, Qi,t, (xi,t, yi,t), and any additional properties depends on the application domain.

[0075] With regard to the above-mentioned examples scenarios, a person skilled in the artcan identify the following examples. For the previously described detailed waste management example the following parameters were used:- Target property: calorific value (energy content per unit mass), μi,t and σ2i,t.- physical mass in kg.- (x, y): bunker position.

[0076] For the Hydropower (water / turbine control) example, the following parameters canbe used:- Target property: effective energy conversion efficiency, pollutant concentration (e.g.,mercury, sulphur), or flow quality parameter.- Qi,t: water flow volume (liters per time segment) or total flow mass.- (x, y): turbine stage, process stage, or time coordinate.

[0077] For the chemical processing example, the following parameters can be used:- Target property: concentration of target chemical component, or pollutantconcentration (e.g., mercury, sulphur), or reaction yield.- Qi,t: batch volume in liters or mass in kg.- (x, y): reactor stage or position in process line.

[0078] For the air pollution control example, the following parameters can be used:- Target property: pollutant concentration (e.g., mercury, sulphur, NOx, particulatematter).- Qi,t: air flow volume or mass flow.- (x, y): duct segment, time coordinate, or monitoring point in the air flow system.

[0079] For the wastewater treatment example, the following parameters can be used:- Target property: pollutant concentration (e.g., chemical oxygen demand (COD), heavymetals, mercury).- Qi,t: water volume in liters or flow rate.- (x, y): treatment stage, reactor position, or pipeline segment.

[0080] During the forward path, typically, only certain components of the state vector Si,t areobservable through process operations and sensors. The following variables are examples of typically observable parameters, depending on the domain:- Control parameters co: All control variables applied by the process operation at time t,including: oProcess coordinates (e.g., spatial position (xo, yo), temporal stage)o Actuator settings (e.g., valve openings, crane pick / drop commands, pumpspeeds, rotational speeds, pressures, temperatures). oManipulated physical quantity Qo (e.g., mass picked, volume transferred, flowrate).- Physical quantities of sample units:o Position or process coordinate (xi,t, yi,t) if tracked.o Physical quantity Qi,t (e.g., mass, volume, flow), observable through weightsensors, flow meters, etc.- Timestamps: Time at which each observation or operation occurs.

[0081] The target property of interest (e.g., calorific value, pollutant concentration, reactionyield), represented by μi,t and σ2i,t, is typically not directly observable during the forwardpath. The system maintains a probabilistic belief about the target property, which is updatedvia the state transition model using the influence kernel based on observed co and the statesof the relevant units.

[0082] Occasional true measurements of the target property for the backward path becomeavailable only at specific observation points in the stream sample processing method (e.g., when a waste unit is burned, when turbine performance is measured, when a chemical batch is analyzed). These measurements trigger updates in the backward path and are used to refine the estimates of μi,tand σ2i,t.

[0083] With regard to the above example scenarios, the following examples of observedmeasurements / measured actual property values may be used:- Waste management:o Observed during forward path: control parameters co = ((xo, yo), cranepick / drop command, Qo), where (xo, yo) is the crane position and Qothe picked or dropped mass. oTrue measurement in backward path: calorific value after burning.- Hydropower:o Observed during forward path: control parameters co = ((xo, yo), turbinespeed, valve settings, Qo), where (xo, yo) corresponds to turbine stage and Qothe manipulated water flow rate. oTrue measurement in backward path: turbine energy efficiency or pollutantconcentration at output.- Chemical processing:o Observed during forward path: control parameters co = ((xo, yo), temperature,pressure, chemical dosage, Qo), where (xo, yo) indicates the reactor stage or pipeline segment and Qo is the volume or mass transferred. oTrue measurement in backward path: concentration or reaction yield fromproduct sampling.- Air pollution control:o Observed during forward path: control parameters co = ((xo, yo), fan speed,valve positions, duct flow rate), where (xo, yo) specifies the location of thecontrolled fan, valve, or duct segment, and Qo is of the is the mass flow at that location.o True measurement in backward path: pollutant concentration at monitoringpoints.- Wastewater treatment:o Observed during forward path: control parameters co = ((xo, yo), valveopenings, applied pressure / temperature, Qo), where (xo, yo) refers to the treatment stage or reactor position, and Qois the flow rate. oTrue measurement in backward path: pollutant concentration in treatedwater.

[0084] The state transition model implemented by the sequential probabilistic modeldescribes how the state of each stream sample unit evolves over time as a result of process operations (applied by the stream sample processing method), which are applied through a set of externally controlled parameters.

[0085] Each process operation at time t can be denoted as: ot = (op_type, co), where:- op_type indicates the operation type (e.g., pick, drop, flow control, mixing, compression,heating, etc.).- co is a vector of control parameters applied by the operation. The control parameters caninclude: oProcess coordinates, e.g., spatial position (xo, yo) or temporal stage.o Actuator settings, e.g., flow rate, rotational speed, temperature set point,pressure set point. oManipulated physical quantity Qo, e.g., mass picked, volume mixed.o Other domain-specific control variables.

[0086] Some examples of Process Operations for some application scenarios are:- In the Waste-to-Energy example, process operations correspond to pick and dropoperations performed by a crane, with control parameters such as crane position and picked mass.- In hydropower, process operations may correspond to controlling water flow throughturbine stages, with control parameters such as rotation speed, pressure, or valve position.- In chemical processes, process operations may represent mixing, heating, separation, orreaction stages, with control parameters such as temperature, flow rate, or chemical dosage.

[0087] The influence of a process operation on the state of each stream sample unit Ui canbe modeled via an influence kernel G.- The influence kernel calculates how the operation ot (through co) interacts not only withunit Uiitself, but also with relevant neighboring units Uj.- These interactions between units, weighted by G, collectively determine the new state ofUiafter the operation.

[0088] Mathematically, the state transition probability is defined as:where:- F(・) is the probabilistic function that maps the prior state of Ui, the influence ofneighboring units, and the control parameters to the new state. F(・) can be a Gauss function or any other probabilistic function.- The sum over j aggregates neighboring units weighted by G.

[0089] The influence kernel G(Uj | co, Sj,t) defines how strongly unit Uj contributes to thestate update of unit Ui. It can depend on:- The distance between Uj and the process coordinates in co.- The current state Sj,t (e.g., physical quantity Qj,t, position).- The type of operation.

[0090] An example for an influence kernel is a spatial Gaussian kernel centered at (xo, yo):

[0091] Other types of kernels include:- Flow-based kernels (based on flow topology).- Time-based kernels (delayed causal influence).- Attention-like kernels dynamically computed from the current unit states.

[0092] If the process operation is local (i.e., acts only on Ui), the influence kernel is definedas a degenerate kernel such that:In this case, the state update depends only on Ui’s prior state and the applied control parameters co, and no neighboring unit contributes.

[0093] For each unit Ui, the new state Si,t+1 is computed as:The transition function in a concrete implementation of function F(・).- For the update of the various components of the state vector (corresponding toestimates for the components of formula F4), the target property distribution can be denoted as:Thereby, the new target property distribution of Ui is a weighted aggregation of contributions from neighboring units (or itself if local).- The update of a physical quantity can be denoted as:where ΔQi(co) depends on the control parameters and the influence kernel.An example for pick / drop reads:(F12)- The update for positions of process coordinates (if applicable) can be denoted as:- Other properties, if applicable, are updated analogously.

[0094] To summarize, the new state of each stream sample unit results from:- Its own prior state.- The redistribution or aggregation of target property and physical quantities driven by theprocess operation.- The interactions with other units mediated by the influence kernel (which may be localor distributed).- The applied control parameters of the respective process operation.

[0095] In the following, the Waste-to-Energy Scenario is described as a concrete exampleinstantiation of the above-described general state transition model. This example illustrates how a skilled person can instantiate the claimed models in any domain of stream sample processing method-based applications based on the influence kernel G, control parameters co, and state update equations.

[0096] In the example, the process includes crane-based pick and drop operations acting onwaste units within a bunker. The same general state transition model can similarly be instantiated in other domains (such as the above shown example scenarios for hydropower, chemical processing, wastewater treatment), with different types of operations and kernels.

[0097] In the Waste-to-Energy Scenario, the influence kernel G(Sj,t, co) corresponds to thespatial 2D Gaussian defined by formula F1.

[0098] The control parameters co include:- Crane position (xc, yc).- Manipulated physical quantity: picked or dropped mass Mpick or Mdrop.

[0099] The update of the target property distribution (μi,t+1, σ2i,t+1) and physical quantityMi,t+1 matches the general formulas F10:- The update of the mass Mi,t+1 is:where Moperationis either Mpickor Mdrop, and the sign corresponds to the type of operation. This shows that the Waste-to-Energy model is fully covered as a specific example of the general state transition model, with:- A spatial kernel G.- Control parameters co = (xc, yc, Moperation).- Standard target property update formulas from the general model.

[0100] For each waste unit Wi, the state transition due to a pick operation can bemodeled as shown in formula F2, and the state transition due to a drop operation can be modeled as shown in formula F3.

[0101] When a stream sample unit enters the stream sample processing method, itstarget property of interest (e.g., calorific value, pollutant concentration, reaction yield) is typically not directly observable or measurable. At this point, the herein disclosed approach assigns an initial probabilistic belief about the target property, which serves as the starting point for the forward modeling of state transitions. Indirect sensor data (such as images, spectra, or process metadata) is typically available at the entry point. The initial belief about the target property can therefore be inferred through statistical estimation, typically using a machine learning model trained on respective historical data.

[0102] A respectively Neural Network NN is used to estimate the initial probabilisticstate of the target property for each stream sample unit. The NN takes as input the available sensor data and outputs a probabilistic description of the target property distribution. Two approaches can be used:- Classification-based estimation: The NN performs classification over predefined classessi, where each class is associated with a known target property distribution.- Direct regression-based estimation: The NN directly outputs a target propertydistribution in the form of mean μi,0 and variance σ2i,0, without an intermediate class. Both approaches produce an initial state representation of the stream sample unit as a probability distribution. In an embodiment where the Neural Network output is a classification-based estimation, the probability distribution can be denoted as:Where:- NN(input)i: Probability that the stream sample unit belongs to class si, given the inputsensor data.- input: Sensor data available at entry (domain-specific: images, spectra, sensor readings,etc.).

[0103] Classes determined by the neural network are then mapped to the targetproperty. Each class si has a known prior target property distribution:- Initial mean μi,0 of the target property.- Initial variance σ2i,0 of the target property.Based on the class probabilities P(si | input), either the most likely class can be selected or a mixture distribution can be computed.

[0104] In an embodiment where the Neural Network is implemented as a regressionneural network, the NN can be trained to directly output: μi,0, σ2i,0 = NN(input) In this embodiment, the NN directly predicts the mean and variance of the target property distribution, bypassing the class step.

[0105] No specific architecture is required for the initial state estimation model. Thesequential probabilistic model (state-space model) can integrate the output of any supervised learning model. Possible architectures include:- Deep Neural Networks (DNNs) with fully connected layers.- Convolutional Neural Networks (CNNs) for image-based data (e.g.,YOLO, ResNet).- Recurrent Neural Networks (RNNs), LSTM networks, or Transformers for time-seriessensor data.- Linear regression or Support Vector Machines (SVMs) for simple cases.

[0106] The choice of model typically depends only on the nature of the availablesensor data - the overall state-space framework is independent of the specific architecture.The Neural Network can be trained in a supervised learning setting, using standard loss functions. For example, for classification-based estimation cross-entropy loss may be used. For direct regression mean squared error (MSE) may be used, or negative log-likelihood ifmodeling both μ and σ2.

[0107] The initial training (pre-training) of the NN can be performed on a superviseddataset including input sensor data (e.g., images, spectra, sensor time series, process metadata) and corresponding labeled target property values or classes. For example:- For Waste-to-Energy, the training data can include images of waste loads from trucks,labeled with known calorific values (or mapped to classes with associated calorific value distributions).- For hydropower, the training data can include water quality measurements withcorresponding energy efficiency or pollutant concentration labels.- For chemical processing, the training data can include sensor readings from rawmaterials with corresponding chemical concentration labels.

[0108] In accordance with formula F4, each new stream sample unit Ui is assigned aninitial state vector:Where:- μi,0: Initial mean of the target property (predicted directly or mapped from class).- σ2i,0: Initial variance of the target property.- Qi,0: Initial physical quantity of the unit (mass, volume, flow rate, etc.).- (xi,0, yi,0): Initial position or process coordinate.- other properties: Other domain-specific initial state variables.

[0109] In the Waste-to-Energy example scenario:- The NN is trained on images of waste from the truck.- The classes si correspond to predefined waste classes (e.g., biomass, plastics, mixedwaste, high moisture waste, metals).- Each class si is associated with a prior calorific value distribution (μi,0, σ2i,0).- The initial state becomes:

[0110] The same mechanism applies to other stream sample processing domains. Forexample:- In hydropower, the NN can be trained on upstream water quality and flowmeasurements.- In chemical processing, the NN can be trained on sensor readings from incoming rawmaterial batches.- In air pollution control, the NN can be trained on inlet gas composition measurements.The result is always an initial probabilistic belief over the target property of each stream sample unit, represented by the assigned (μi,0, σ2i,0).

[0111] The initial training (pre-training) of the Neural Network typically uses asupervised learning setup based on an initial labeled dataset from the respective domain. This dataset includes sensor data (images, spectra, time series, metadata, etc.) and corresponding labeled target property values or class assignments, obtained through laboratory analysis or expert assessment.

[0112] Thereby, manual labeling is useful for initial calibration, but not strictlyrequired. The system can also be initialized with a small seed dataset or even random initialization. As the system operates, it continuously performs backward path corrections whenever actual target property measurements become available. These corrections automatically generate high-quality labeled training data including the corresponding sensorinput and the corrected target property value. No human intervention is required for thislabeling process: the system itself provides the supervision signal through its backward pathmechanism. This enables fully automated fine-tuning of the Neural Network based solely on the system’s own feedback loop.

[0113] During the forward path, the target property of each stream sample unit ismodeled as a probabilistic belief (distribution), because the target property cannot be directly observed or measured during processing. However, at certain points in the process, actual measurement values of the target property are obtained for an aggregated stream segment (e.g., when a waste unit is burned, when turbine performance is measured, when chemical batch is analyzed). Such actual measurement values are used to perform a backward correction: the states of the individual contributing stream sample units are updated retrospectively to maximize consistency with the observed true outcome.

[0114] In the forward path, the mean and variance of the target property areupdated via the physical process model (mixing of distributions), according to the statetransition function and influence kernel - without any access to true (actual) measurements.

[0115] In the backward path, a true measurement of the target property becomesavailable. Now, the measurement variance and the prior variance of each contributing unit are statistically fused. This can be performed via a Kalman gain, which optimally balances prior belief and measurement based on their respective uncertainties.

[0116] Given a true measurement μtrue of an actual target property value with knownmeasurement variance σ2meas, the affected contributing units are identified, and their state beliefs can be updated using the following Kalman update.

[0117] Influence weights: The contribution of each unit to the observed truemeasurement is modeled via an influence kernel G, similar to that used in the forward path. This determines how much each unit has contributed to the observed measurement.

[0118] Kalman update: For each contributing unit Uj, the update of the targetproperty mean and variance is performed as follows:(F16) Where:- ωj : Normalized influence weight of unit Uj on the observed segment (computed viakernel G), expressing the fraction of the observed segment attributed to unit Uj.- σ2meas: Known measurement variance of the target property.- σ2j,t: Prior variance of the target property of unit Uj.- Kj : Kalman gain, determining the relative trust in the measurement vs. the prior belief.

[0119] Thereby, units that contributed more strongly to the observed segment (ωjlarge) will receive a stronger update (higher Kj). Units with lower prior uncertainty (σ2j,tsmall) will be corrected less strongly (lower Kj). The updated variance reflects the reduced uncertainty after incorporating the measurement. The physical quantity Qj,t+1is reduced to account for the fact that a portion of unit Ujhas now been consumed in the measured segment.

[0120] FIG. 4C shows pseudo code 403 for an algorithm 1a to update states based ontrue actual measurements in the Backward Path, which is a generalized version of algorithm 1 shown in FIG.4A. For algorithm 1a, the following definitions apply:- μtrue: True actual target property value measured from the observed stream segment.- σ2meas: Known variance of the true measurement.- observed segment: The stream segment or unit on which the true measurement wasperformed.- contributing units: The set of stream sample units that contributed to the observedsegment.- ωj : Influence weight of unit Uj on the observed segment (computed via kernel G).- Kj : Kalman gain determining the weight of the update.- Uj.μt+1: Updated mean of the target property for unit Uj.- Uj.σ2t+1: Updated variance of the target property.- Uj.Qt+1: Updated physical quantity (typically reduced to reflect that some portion wasconsumed in the observed segment).

[0121] In practical applications, the measurement variance σ2meas is typicallyestimated based on known sensor uncertainty, and does not need to be measured directly for each individual observation. This ensures proper Bayesian fusion of prior belief and measurement in the Kalman gain update.

[0122] In the Waste-to-Energy (WTE) example, the variance of the calorific valuemeasurement can be modeled as proportional to the observed mean: σ2meas= (α · μtrue)2where α is a domain-specific coefficient representing the relative accuracy of the calorific value sensor.- Example: α = 0.05 → 5% rela^ve standard devia^on → σ2meas = (0.05 · μtrue)2- Example: α = 0.10 → 10% rela^ve standard devia^on → σ2meas = (0.10 · μtrue)2

[0123] There is no need to change σ2meas for every individual measurement – unlessthe sensor system provides a dynamic confidence estimate (which is rare). Typically, a fixed and reasonable value of α, tuned for your system, is sufficient. The measurement variance σ2measis thus treated as a known parameter in the Kalman update, ensuring consistent and statistically grounded correction of the stream unit states. The update proceeds recursively through the causal chain of prior operations, adjusting earlier states to improve global consistency with the true observation.

[0124] The retraining of the neural network has already been described in thesummary, and in detail for the Waste-to-Energy example scenario in the context of algorithm 2 in FIG.4B. A skilled person understands from FIG.4B that the algorithm can begeneralized independent of the application scenario. Such generalization is shown in theform of algorithm 2a shown in FIG.4D.

[0125] The Neural Network used for initial state estimation can be improved overtime, through supervised fine-tuning, based on the corrected target property values obtained via the backward path (based on the actual (measured) target values).

[0126] In the herein disclosed approach, retraining requires no manual labeling: thenew training data is fully automatically labeled by the backward path, based on true (actual) process measurements.

[0127] For each stream sample unit, the following is available:- Input features (images, spectra, sensor data, etc.) - known from the forward path.- Corrected target property value μtrue (with uncertainty σ2true) - provided by the backwardpath. The result is an automatically generated dataset of (input features, target property) pairs.

[0128] Retraining is performed as standard supervised finetuning of the alreadytrained model:- If the model is a classification-based model, the corrected target property μtrue is mappedto the appropriate class label. The corrected target property may deviate from the originally estimated target property value dependent on the actually measured target value.- If the model is a regression-based model, the corrected μtrue and σ2true can be useddirectly for training.

[0129] In the following, the generalized retraining procedure (algorithm2a) for theclassification-based model is described in more detail. The following definitions apply:- μtrue: Corrected target property value (posterior mean) from backward smoothing.- σ2true: Estimated variance (uncertainty) of the corrected target property value, from theKalman backward path.- input features: Input data used by the NN for initial state estimation (e.g., image,spectra, sensor readings).- classtrue: Class corresponding to μtrue.- μclass: Current prior mean for this class.- σ2class: Current prior variance for this class.- δ: Difference between corrected value and class prior mean.- α: Learning rate for smooth class updates.

[0130] Using algorithm 2a, the Neural Network is continuously fine-tuned duringsystem operation in any stream sample processing domain. The training labels are generatedfully automatically by the backward path - no human intervention is required. The NN isenabled to adapt over time to new process conditions, new stream types, or changing input distributions. The same retraining principle applies to both classification and regression models.

[0131] In case of having a plurality of individual stream sample units, the state of theentire system at time t is the collection of the states of all individual stream sample units:St = {Si,t | i = 1, .. . ,Nt} (F17)where Nt is the total number of stream sample units present in the system at time t. Eachunit state Si,tis recursively updated via process operations in the forward path, and via a Kalman update in the backward path when true (actual) measurements of the target property become available.

[0132] A skilled person will recognize that the claimed approach is applicable to awide range of stream sample processing domains. The herein described example domains are exemplary only and the skilled person is able to identify further domains without additional burden based on the features of the herein disclosed approach which is summarized in the following: 1. Initial State Estimation: The target property of each stream sample unit is initially unobservable and is estimated via a Neural Network (or another equivalent estimation model), based on external features (e.g., images, sensor data).2. State-Space Model (Sequential Probabilistic Model): Each stream sample unit Ui has astate Si,t comprising: -Target property distributionσ2i,t)- Physical quantity Qi,t- Process coordinatesyi,t) (if applicable)- Other properties (domain-specific)A State-Space Model is always sequential. That is, in a probabilistic State-Space Model state transitions occur sequentially. 3. State Transition Model (Forward Path): The transitions of each unit are driven by process operations with externally controlled parameters, and mediated by an influence kernel that models interactions between units. This allows flexible modeling of a wide range of process dynamics (pick / drop, flow control, mixing, heating, etc.). 4. State Correction and Backward Path: When true measurements (actual values) of the target property become available at observation points, the states of contributing units are updated via a recursive Kalman update, fusing prior beliefs and measurement uncertainty. 5. Iterative Refinement of Initial State Estimation Model: The initial state estimation model (e.g., Neural Network) is continuously improved based on corrected target property values obtained through the backward path, enabling better predictions for future stream sample units. 6. System State: The state of the entire system at any time t is the collection of all individual stream sample unit states {Si,t}.

[0133] By modeling the state-space system for each individual stream sample unit, itis possible to accurately capture the complex dynamics and interactions resulting from process operations. The herein disclosed approach supports both online state propagation and delayed state correction, enabling system optimization across a wide range of process domains (e.g., Waste-to-Energy, Hydropower, Chemical Processing, Wastewater Treatment, Air Pollution Control).

[0134] FIG. 5 is a diagram that shows an example of a generic computer device 900and a generic mobile computer device 950, which may be used with the techniques described here. Computing device 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. Generic computer device 900 may correspond to the computer system 100 of FIG.1. Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smart phones, and other similar computing devices. For example, computingdevice 950 may include the data storage components and / or processing components of agent devices as shown in FIG.1. The components shown here, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the inventions described and / or claimed in this document.

[0135] Computing device 900 includes a processor 902, memory 904, a storagedevice 906, a high-speed interface 908 connecting to memory 904 and high-speed expansion ports 910, and a low speed interface 912 connecting to low speed bus 914 and storage device 906. Each of the components 902, 904, 906, 908, 910, and 912, are interconnected using various busses, and may be mounted on a common motherboard or in other manners as appropriate. The processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or on the storage device 906 to display graphical information for a GUI on an external input / output device, such as display 916 coupled to high speed interface 908. In other implementations, multiple processors and / or multiple buses may be used, as appropriate, along with multiple memories and types of memory. Also, multiple computing devices 900 may be connected, with each device providing portions of the necessary operations (e.g., as a server bank, a group of blade servers, or a multi-processor system).

[0136] The memory 904 stores information within the computing device 900. In oneimplementation, the memory 904 is a volatile memory unit or units. In another implementation, the memory 904 is a non-volatile memory unit or units. The memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.

[0137] The storage device 906 is capable of providing mass storage for thecomputing device 900. In one implementation, the storage device 906 may be or contain a computer-readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid-state memory device, or an array of devices, including devices in a storage area network or other configurations. A computer program product can be tangibly embodied in an information carrier. The computer program product may also contain instructions that, when executed, perform oneor more methods, such as those described above. The information carrier is a computer- ormachine-readable medium, such as the memory 904, the storage device 906, or memory on processor 902.

[0138] The high-speed controller 908 manages bandwidth-intensive operations forthe computing device 900, while the low-speed controller 912 manages lower bandwidth- intensive operations. Such allocation of functions is exemplary only. In one implementation, the high-speed controller 908 is coupled to memory 904, display 916 (e.g., through a graphics processor or accelerator), and to high-speed expansion ports 910, which may accept various expansion cards (not shown). In the implementation, low-speed controller 912 is coupled to storage device 906 and low-speed expansion port 914. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet) may be coupled to one or more input / output devices, such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.

[0139] The computing device 900 may be implemented in a number of differentforms, as shown in the figure. For example, it may be implemented as a standard server 920, or multiple times in a group of such servers. It may also be implemented as part of a rack server system 924. In addition, it may be implemented in a personal computer such as a laptop computer 922. Alternatively, components from computing device 900 may be combined with other components in a mobile device (not shown), such as device 950. Each of such devices may contain one or more of computing device 900, 950, and an entire system may be made up of multiple computing devices 900, 950 communicating with each other.

[0140] Computing device 950 includes a processor 952, memory 964, aninput / output device such as a display 954, a communication interface 966, and a transceiver 968, among other components. The device 950 may also be provided with a storage device, such as a microdrive or other device, to provide additional storage. Each of the components 950, 952, 964, 954, 966, and 968, are interconnected using various buses, and several of the components may be mounted on a common motherboard or in other manners as appropriate.

[0141] The processor 952 can execute instructions within the computing device 950,including instructions stored in the memory 964. The processor may be implemented as a chipset of chips that include separate and multiple analog and digital processors. The processor may provide, for example, for coordination of the other components of the device950, such as control of user interfaces, applications run by device 950, and wireless communication by device 950.

[0142] Processor 952 may communicate with a user through control interface 958and display interface 956 coupled to a display 954. The display 954 may be, for example, a TFT LCD (Thin-Film-Transistor Liquid Crystal Display) or an OLED (Organic Light Emitting Diode) display, or other appropriate display technology. The display interface 956 may comprise appropriate circuitry for driving the display 954 to present graphical and other information to a user. The control interface 958 may receive commands from a user and convert them for submission to the processor 952. In addition, an external interface 962 may be provided in communication with processor 952, so as to enable near area communication of device 950 with other devices. External interface 962 may provide, for example, for wired communication in some implementations, or for wireless communication in other implementations, and multiple interfaces may also be used.

[0143] The memory 964 stores information within the computing device 950. Thememory 964 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or units, or a non-volatile memory unit or units. Expansion memory 984 may also be provided and connected to device 950 through expansion interface 982, which may include, for example, a SIMM (Single In-Line Memory Module) card interface. Such expansion memory 984 may provide extra storage space for device 950, or may alsostore applications or other information for device 950. Specifically, expansion memory 984may include instructions to carry out or supplement the processes described above, and may include secure information also. Thus, for example, expansion memory 984 may act as a security module for device 950, and may be programmed with instructions that permit secure use of device 950. In addition, secure applications may be provided via the SIMM cards, along with additional information, such as placing the identifying information on the SIMM card in a non-hackable manner.

[0144] The memory may include, for example, flash memory and / or NVRAMmemory, as discussed below. In one implementation, a computer program product is tangibly embodied in an information carrier. The computer program product contains instructions that, when executed, perform one or more methods, such as those describedabove. The information carrier is a computer- or machine-readable medium, such as thememory 964, expansion memory 984, or memory on processor 952, that may be received, for example, over transceiver 968 or external interface 962.

[0145] Device 950 may communicate wirelessly through communication interface966, which may include digital signal processing circuitry where necessary. Communication interface 966 may provide for communications under various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. Such communication may occur, for example, through radio-frequency transceiver 968. In addition, short-range communication may occur, such as using aBluetooth, WiFi, or other such transceiver (not shown). In addition, GPS (Global PositioningSystem) receiver module 980 may provide additional navigation- and location-relatedwireless data to device 950, which may be used as appropriate by applications running on device 950.

[0146] Device 950 may also communicate audibly using audio codec 960, which mayreceive spoken information from a user and convert it to usable digital information. Audio codec 960 may likewise generate audible sound for a user, such as through a speaker, e.g., in a handset of device 950. Such sound may include sound from voice telephone calls, may include recorded sound (e.g., voice messages, music files, etc.) and may also include sound generated by applications operating on device 950.

[0147] The computing device 950 may be implemented in a number of differentforms, as shown in the figure. For example, it may be implemented as a cellular telephone 980. It may also be implemented as part of a smart phone 982, personal digital assistant, or other similar mobile device.

[0148] Various implementations of the systems and techniques described here canbe realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0149] These computer programs (also known as programs, software, softwareapplications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object-oriented programming language, and / or in assembly / machine language. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term “machine-readable signal” refers to any signal used toprovide machine instructions and / or data to a programmable processor.

[0150] To provide for interaction with a user, the systems and techniques describedhere can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0151] The systems and techniques described here can be implemented in acomputing device that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front endcomponents. The components of the system can be interconnected by any form or mediumof digital data communication (e.g., a communication network). Examples of communication networks include a local area network (“LAN”), a wide area network (“WAN”), and the Internet.

[0152] The computing device can include clients and servers. A client and server aregenerally remote from each other and typically interact through a communication network.The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship with each other.

[0153] A number of embodiments have been described. Nevertheless, it will beunderstood that various modifications may be made without departing from the spirit and scope of the invention.

[0154] In addition, the logic flows depicted in the figures do not require theparticular order shown, or sequential order, to achieve desirable results. In addition, other steps may be provided, or steps may be eliminated, from the described flows, and other components may be added to, or removed from, the described systems. Accordingly, other embodiments are within the scope of the following claims.

Claims

1. Claims 1. A computer-implemented method (1000) for evaluating at least one property of a physical stream sample, wherein the at least one property is a relevant parameter for optimization of a stream sample processing method, wherein the stream sample moves through processing means performing said stream sample processing method, the computer- implemented method comprising: receiving (1100) an initial state image (ISI1) being characteristic for the at least one property of said stream sample (S_b) at the beginning of the stream sample processing method (200); by a respectively pre-trained neural network (120), determining (1200), based on the initial state image (ISI), a sample type (ST1) with a corresponding confidence value (CV1) for said stream sample (S_b); based on the sample type (ST1), predicting (1300), by a sequential probabilistic model (140) of the stream sample processing method, at least one target property value (PVt1) of the stream sample, wherein the probabilistic model is conditioned on one or more control parameters which control the stream sample processing method; writing (1400) the at least one property target value (PVt1) into a sample specific data structure (300), and providing (1500) the one or more control parameters (CP*) for controlling the stream sample processing method (200); receiving (1600) a measuring result (MVe) for at least one actual property value (PVa1) of the processed stream sample (S_e) at the end of the sample processing method (200), wherein the at least one actual property value (PVa1) relates to the at least one predicted target property value (PVt1);mapping (1700) the at least one actual property value (PVa1) backwards through the sequential probabilistic model (140) to the corresponding initial state image (ISI1), and writing (1800) the at least one actual property value (PVa1) as new entry to the sample specific data structure (300);updating (1900) the sequential probabilistic model (140) or the neural network (120) or both, wherein: updating (1910) the sequential probabilistic model (140) comprises updating conditional probability distributions of the sequential probabilistic model (140) based on the at least one actual property value taking into account uncertainties in the sample processing method (200); and updating (1920) the neural network (120) comprises initiating automatic re- training of the neural network with additional training data (ISI*|PVa*) for adjusting the weights of the neural network, wherein the additional training data are pairs of initial state images (ISI*) and respective actual property values (PVa*) collected from earlier stream samples of the determined sample type (ST1).

2. The method of any of claim 1, wherein the sequential probabilistic model (140) describes the processing method as a sequence of possible events in which the probability of each event depends on the state attained in the previous event.

3. The method of any of the previous claims, wherein the sequential probabilistic model (140) is a Markov chain in which the probability of each event depends only on the state attained in the previous event.

4. The method of any of the previous claims, wherein the sample specific data structure is a sample specific blockchain.

5. The method of any of the previous claims, wherein properties of the physical real-world sample comprise one or more of: physical, chemical, biological, mechanical or sensory characteristics of said sample.

6. The method of any of the previous claims, wherein, in case the neural network is a classification neural network, the sample type is determined according to predefined sample type classes associated with said stream sample, or, in case the neural network is a regression neural network, the sample type is determined as regression value.

7. The method of any of the previous claims, wherein automatic re-training of the neural network is initiated in case the number of collected training data pairs exceeds a predefined threshold.

8. The method of any of the previous claims, wherein updating depends on the confidence value provided by the neural network for the determined sample type, characterized in that: In case the confidence value is below a predefined minimum threshold, only the neural network is updated; In case the confidence value is above a predefined maximum threshold, only the sequential probabilistic model is updated; and In case the confidence value is in a range from the predefined minimum threshold to the predefined maximum threshold, the neural network and the sequential probabilistic model are updated.

9. A computer program product that, when loaded into a memory of a computing device and executed by at least one processor of the computing device, executes the steps of the computer-implemented method according to any one of the previous claims.

10. A computer system (100) for evaluating at least one property of a physical stream sample, wherein the at least one property is a relevant parameter for optimization of a stream sample processing method (200), wherein the stream sample continuously moves through said stream sample processing method, the system comprising: a first interface adapted for receiving an initial state image (ISI1) being characteristic for the at least one property of said stream sample (S_b) at the beginning of the stream sample processing method (200); a respectively pre-trained neural network (120) adapted to determine, based on the initial state image (ISI), a sample type (ST1) with a corresponding confidence value (CV1) for said stream sample (S_b); a sequential probabilistic model (140) of the stream sample processing method adapted to predict, based on the sample type (ST1), at least one target propertyvalue (PVt1) of the stream sample, wherein the probabilistic model is conditioned on one or more control parameters which control the stream sample processing method; a second interface adapted to write the at least one property target value (PVt1) into a sample specific data structure (300), and to provide the one or more control parameters (CP*) for controlling the stream sample processing method (200); a third interface adapted to receive, a measuring result (MVe) for at least one actual property value (PVa1) of the processed stream sample (S_e) at the end of the sample processing method (200), wherein the at least one actual property value (PVa1) relates to the at least one predicted target property value (PVt1); a mapper module (130) adapted to map the at least one actual property value (PVa1) backwards through the sequential probabilistic model (140) to the corresponding initial state image (ISI1); the second interface further adapted to write the at least one actual property value (PVa1) as new entry to the sample specific data structure (300); and and an updater module (150) adapted to update the sequential probabilistic model (140) or the neural network (120) or both, wherein: updating the sequential probabilistic model (140) comprises updating conditional probability distributions of the sequential probabilistic model (140) based on the at least one actual property value (PVa1) taking into account uncertainties in the sample processing method (200); and updating the neural network (120) comprises initiating automatic re-training of the neural network with additional training data (ISI*|PVa*) for adjusting the weights of the neural network, wherein the additional training data are pairs of initial state images (ISI*) and respective actual property values (PVa*) collected from earlier stream samples of the determined sample type (ST1).

11. The system of claim 10, wherein the sequential probabilistic model (140) describes the processing method as a sequence of possible events in which the probability of each event depends on the state attained in the previous event.

12. The system of claim 10 or 11, wherein the sequential probabilistic model (140) is a Markov chain in which the probability of each event depends only on the state attained in the previous event.

13. The system of any of claims 10 to 12, wherein, in case the neural network is a classification neural network, the sample type is determined according to predefined sample type classes associated with said stream sample, or, in case the neural network is a regression neural network, the sample type is determined as regression value.

14. The system of any of claims 10 to 13, wherein the updater module (140) is adapted to initiate automatic re-training of the neural network in case the number of collected training data pairs exceeds a predefined threshold.

15. The system of any of claims 10 to 14, wherein the updater module (160) depends on the confidence value provided by the neural network for the determined sample type, characterized in that: In case the confidence value is below a predefined minimum threshold, only the neural network is updated; In case the confidence value is above a predefined maximum threshold, only the sequential probabilistic model is updated; and In case the confidence value is in a range from the predefined minimum threshold to the predefined maximum threshold, the neural network and the sequential probabilistic model are updated.

Images