Radiometric measuring device

Abstract

Radiometric measuring instrument (1), comprising: - a number n of sensors (2_1 to 2_n), wherein each sensor (2_i) of the number n of sensors (2_1 to 2_n) is configured to provide associated sensor data (x i ) to generate a total of n sensor data (x1, ...,x n ) is generated using the number n of sensors (2_1 to 2_n), - a measurement unit (4) designed to calculate a number m of measurement values ​​(ŷ1, ...,y) m ) depending on the number n of sensor data (x1, ...,x n ) based on values ​​of a number d of parameters (θ1, ..., θ d to calculate, and - a learning unit (5), wherein the learning unit (5) is trained to perform based on training data ( ( xt 1 ( i ) ,..., xtn ( i ) ) ; ( ys 1 ( i ) ,..., ysm ( i ) ) ) the values ​​of the number d of parameters (θ1, ..., θ d to calculate - where the number n of sensors (2_1 to 2_n) is selected from the set of sensors consisting of: - at least one radiometric sensor, in particular a radiometric sensor designed to generate sensor data in the form of a count rate or radiation intensity data, and / or a radiometric sensor designed to generate sensor data in the form of information about radiometric spectra, - at least one sensor designed to generate sensor data in the form of temperature data, - at least one sensor designed to generate sensor data in the form of acceleration data, - at least one sensor designed to generate sensor data in the form of speed data, - at least one sensor designed to generate sensor data in the form of position data, - at least one sensor, in particular in the form of an ultrasonic sensor or laser sensor, which is designed to generate sensor data in the form of information about a load height profile, and - at least one sensor designed to generate sensor data in the form of humidity data, - where at least one measurement variable is selected from the set of measurement variables consisting of: - Fill level, - Positions and / or thicknesses of individual material layers, - Density, - Delivery rate, in particular total delivery mass, - Throughput, especially mass flow rate, and - Material composition, in particular content levels, - wherein the learning unit (5) is trained to extract data from the training data ( ( xt 1 ( i ) ,..., xtn ( i ) ) ; ( ys 1 ( i ) ,..., ysm ( i ) ) ) a number n of training sensor data ( xt 1 ( i ) ,..., xtn ( i ) ) to extract and a number m of associated setpoints ( ys 1 ( i ) ,..., ysm ( i ) ) the number m of measured quantity values ​​(y1, ..., y m ) to extract, and - wherein the measurement unit (4) is configured to provide a number m of training values ( yt 1 ( i ) ,..., ytm ( i ) ) the number m of measured variable values ​​(y1, ...,y m ) depending on the number n of training sensor data ( xt 1 ( i ) ,..., xtn ( i ) ) to calculate, and - wherein the learning unit (5) is trained to, based on the number m of setpoints ( ys 1 ( i ) ,..., ysm ( i ) ) the number m of measured quantity values ​​(y1, ..., y m ) and the number m of training values ( yt 1 ( i ) ,..., ytm ( i ) ) the number m of measured quantity values ​​(ŷ1, ..., y m) the values ​​of the number d of parameters (θ1, ..., θ d to calculate.

Description

[0001] The invention relates to a radiometric measuring device.

[0002] In process measurement technology, radiometric measuring devices, which include radiometric sensors for radiation measurement, are frequently used to measure process variables or material properties, for example for level measurement, density measurement, etc.

[0003] DE10 2017 219 843 A1 discloses an example of machine learning, in which measured values ​​are calculated based on a number of sensor data using a measurement calculation unit.

[0004] US patent 2003 / 0227382A1 discloses a measurement setup for various radiometric measuring instruments, which includes a measurement quantity calculation unit and a learning unit.

[0005] The invention is based on the objective of providing a radiometric measuring device that enables reliable determination of values ​​of a number of quantities to be measured, even in complex radiometric measurement problems.

[0006] The invention solves this problem by means of a radiometric measuring device according to claim 1.

[0007] The radiometric measuring device has a number n of sensors, where n is a natural number greater than zero. The number n can, for example, be in a range between one and 12.

[0008] Each sensor, of number n, is configured to generate its associated sensor data, resulting in a total of n sensor data being generated by n sensors. The sensor data can be, for example, digital data with a resolution between 8 bits and 64 bits. The digital sensor data is generated continuously at a predefined repetition rate. For example, every 100 ms, the data from all n sensors is generated simultaneously or with a known temporal relationship to each other.

[0009] The radiometric measuring instrument further comprises at least one measurement unit configured to calculate a number m of measurement values ​​or process measurements as a function of the number n of sensor data, based on the values ​​of a number d of parameters. Specifically, the number d of parameters are parameters of a hypothesis function h, which approximates a measurement function f. The sensor data may also be time-shifted. The number m is a natural number greater than zero and can, for example, be in a range between 1 and 4. The number d is a natural number greater than zero and can, for example, be in a range between 1 and 500. The measurement values ​​represent, for example, a corresponding number of process variables to be measured.

[0010] The radiometric measuring device further includes a learning unit, which is trained to calculate the values ​​of the number d of parameters based on training data.

[0011] The training unit, the number of sensors, and the measurement unit can be located together or separately. For example, the number of sensors and the measurement unit can be located at the point of measurement, i.e., in the field, while the training unit can be located separately. The training unit can be implemented, for example, by a powerful computer that calculates the number d of parameters based on the training data. These parameters are then provided to the measurement unit, for example, by transmitting them from the training unit to the measurement unit via a data network. Of course, the training unit, the measurement unit, and, if necessary, the number n of sensors can be integrated into a single physical device.

[0012] According to one embodiment, the measurement calculation unit has at least one feature extraction unit configured to extract feature data from the number n of sensor data, in particular based on the values ​​of the number d of parameters, wherein the at least one measurement calculation unit is configured to calculate the number m of measurement values ​​from the feature data based on the values ​​of the number d of parameters.

[0013] The number n of sensors is selected from the set of sensors consisting of: at least one radiometric sensor, in particular a radiometric sensor configured to generate sensor data in the form of a count rate or radiation intensity data, and / or a radiometric sensor configured to generate sensor data in the form of information about radiometric spectra, at least one sensor configured to generate sensor data in the form of temperature data, at least one sensor configured to generate sensor data in the form of acceleration data, at least one sensor configured to generate sensor data in the form of velocity data, at least one sensor configured to generate sensor data in the form of position data, and at least one sensor, in particular in the form of an ultrasonic sensor or laser sensor, configured toto generate sensor data in the form of information about a load height profile, and at least one sensor designed to generate sensor data in the form of humidity data.

[0014] The learning unit is trained to extract training sensor data from the training data and to extract the corresponding target values ​​of the number m of measured variable values ​​associated with the training sensor data, wherein the measured variable calculation unit is trained to calculate training values ​​of the number m of measured variable values ​​as a function of the training sensor data, and wherein the learning unit is trained to calculate the values ​​of the number d of parameters based on the target values ​​of the number m of measured variable values ​​and the training values ​​of the number m of measured variable values.

[0015] The at least one measured variable is selected from the set of measured variables consisting of: fill level, positions and / or thicknesses of individual material layers, density, delivery rate, in particular total conveyed mass, throughput, in particular mass flow, and material composition, in particular content parameters.

[0016] In complex radiometric measurement problems, the relationships between sensor data and measured values ​​cannot always be described analytically with absolute precision. Therefore, these relationships must currently be approximated using suitable physical and mathematical models. However, the measurement model is often unknown or too simplistic, a suitable sensor fusion model is not known, and / or the measurement model is too complex. As a result, the overall quality of the measured values ​​or process values ​​calculated in the measurement system suffers.

[0017] According to the invention, this problem is solved by the radiometric measuring device automatically establishing relationships between sensor data and measured values ​​using artificial intelligence methods, such as "machine learning" or "deep learning", without explicit knowledge of an analytical measurement equation based on training data.

[0018] This increases measurement accuracy by reducing systematic errors. The quality and accuracy of the calculated measured values ​​are improved because the radiometric measuring instrument learns the entirety of all process influences using the training data. This includes not only known influences but also, in particular, unknown disturbances whose existence, for example, is not known during commissioning. Therefore, the radiometric measuring instrument can even contain more expert knowledge than a human operator and process expert in the field. This leads to a significant improvement in the accuracy of the measured values.

[0019] Furthermore, previously unsolvable measurement tasks can now be solved. In contrast to classical measurement systems in radiometric process measurement technology, the invention does not require a pre-existing explicit measurement model. Instead, the radiometric measuring device creates the measurement model itself using artificial intelligence. The physical process to be measured can thus be treated as a black box whose internal logic is unknown and is learned and organized by the radiometric measuring device itself. As a result, the invention allows measurement tasks for novel, previously unfamiliar, or unsolvable physical processes to be represented. This offers a significant advantage, as new markets and problem areas can be accessed without expert knowledge or with only rudimentary expertise in the respective domain. The radiometric measuring device according to the invention acquires the expert knowledge itself.

[0020] Furthermore, the radiometric measuring device according to the invention requires less memory compared to conventional solutions of similar performance. When comparing the amount of information processed and learned during the learning phase according to the invention with the number of model parameters, the memory requirement of the radiometric measuring device according to the invention is comparatively small. If one wanted to achieve similar measurement performance with conventional measuring devices (if this is even possible), extensive calibration tables would be necessary, which would exceed any memory limitations of embedded systems.

[0021] Furthermore, the radiometric measuring device according to the invention enables shorter cycle times and higher update rates. The computational steps performed in the radiometric measuring device according to the invention for processing the sensor data into the measured values ​​during the measurement process are largely based on methods of linear algebra, specifically matrix-vector multiplications. These operations are executed significantly more efficiently on suitable hardware than classical procedural mathematical algorithms of radiometric process measurement technology. This results in shorter cycle times of the measuring system and thus, from the customer's perspective, higher update rates in the process control system.

[0022] Furthermore, the radiometric measuring device according to the invention enables shorter design times. Since only minimal expert knowledge from the respective field is required, customized solutions for novel measurement tasks can be developed more quickly. The associated measuring systems can be designed and customized much more easily. This saves time and reduces costs.

[0023] Furthermore, the radiometric measuring device according to the invention enables simpler operation. Calibration of the radiometric measuring device according to the invention is very simple. Instead of determining complex, application-specific calibration parameters as with conventional measuring devices, only training data needs to be recorded according to the invention. The radiometric measuring device then calibrates itself using the learning unit. This standardizes and significantly simplifies the calibration process for the operator across all measurement applications.

[0024] The invention is described in detail below with reference to the drawings. These show: Fig. 1. A highly schematic block diagram of a radiometric measuring device according to the invention. Fig. 2. Highly schematic, a block diagram of the internal structure of an embodiment of a measurement quantity calculation unit of the in Fig. 1 radiometric measuring device shown, Fig. 3. Highly schematic, a block diagram of the internal structure of another embodiment of a measurement quantity calculation unit of the in Fig. 1 radiometric measuring device shown, and Fig. 4 to the highest schematically a block diagram of the in Fig. 1 radiometric measuring device shown in a learning mode.

[0025] Fig. Figure 1 shows a highly schematic block diagram of a radiometric measuring device 1 according to the invention.

[0026] The radiometric measuring device 1 has a number n of sensors 2_1 to 2_n, wherein each sensor 2_i of the number n of sensors 2_1 to 2_n is configured to measure associated sensor data x i to generate a total of n sensor data x1, ...,x n generated by means of the number n of sensors 2_1 to 2_n.

[0027] The radiometric measuring instrument 1 further comprises a measurement quantity calculation unit 4, which is designed to calculate a number m of measurement quantity values ​​y1, ...,y m depending on the number n of sensor data x1,...,x n based on values ​​of a number d of parameters θ1,...,θ d to calculate.

[0028] Referring to Fig. 4 The radiometric measuring device 1 has a learning unit 5, wherein the learning unit 5 is trained to perform based on training data xt1(i),...,xtn(i);ys1(i),...,ysm(i) the values ​​of the number d of parameters θ1, ..., θ d to calculate.

[0029] The radiometric measuring device 1 according to the invention calculates input quantities in the form of the sensor values ​​x1, ...,x n , which may also be time-delayed, in output variables in the form of process measurements or measured variable values ​​y1, ...,y m um.

[0030] The conversion depends on the model parameters θ1, ..., θ d from which, initially unknown, are learned using learning unit 5 via so-called machine learning. Recorded training data, also called learning data, is used, which can be derived from real recorded data during operation and / or from simulation data.

[0031] Machine learning means that the radiometric measuring device artificially generates knowledge from experience. The radiometric measuring device learns from examples and can generalize these after the learning phase is complete. This means that it doesn't simply memorize the examples, but rather the radiometric measuring device recognizes patterns and regularities in the training data. In this way, it can also evaluate unknown data (learning transfer).

[0032] The radiometric measuring device primarily employs learning techniques from so-called supervised learning, in which the device learns a measurement function from given pairs of inputs and outputs. During the learning process, the correct measured values ​​for a number n of sensor data are provided, for example, based on a reference measurement or a simulation.

[0033] Formally, the radiometric measuring device approximates a measuring function. f:(x1,...,xn)↦(y1,...,ym), which n input variables or sensor data (x1, ...,x n ) on m output variables or measured variable values ​​(y1, ...,y m ) maps, through a suitable hypothesis function hθ:(x1,...,xn)↦(y^1,...,y^m), which the n sensor data (x1, ..., x n ) on m estimates (ŷ1, ...,ŷ m ) for the (y1, ..., y m ) maps and depends on the model parameters θ := (θ1, ..., θ d ) is.

[0034] In this process, each of the d individual model parameters θ is considered. i understood one of the following three things: - a mathematical object, in particular - a number - a vector - a function - a parameterized piece of program logic or source code - a piece of program logic or source code generated by a code generator

[0035] The model parameters (θ1, ..., θ) dThe learning algorithm learns from training data. More precisely, training data consists of l (where l, for example, lies in a range between 10). 5 and 10 7 lies, in particular l = 10 6 ) Training pairs (xt(1),ys(1)),...,(xt(l),ys(l)), which, for example, each have the dimension n + m and each consist of a complete set of input data or training sensor data. xt(i):=(xt1(i),...,xtn(i)) plus associated target values ys(i):=(ys1(i),...,ysm(i)) The number m of measured variable values ​​consists of, where i = 1, ..., l. The setpoint values (ys1(i),...,ysm(i)) are also referred to as training labels.

[0036] Based on the training sensor data (xt1(i),...,xtn(i)) The measurement unit 4 calculates the parameters (θ1, ..., θ). d ) dependent training values (yt1(i),...,ytm(i)):=hθ(xt1(i),...,xtn(i)) the number m of measured variable values ​​(ŷ1, ...,y) m Learning unit 5 is designed to perform calculations for i = 1, ..., l based on the target values. (ys1(i),...,ysm(i)) and the training values (yt1(i),...,ytm(i)) the values ​​of the parameters (θ1, ..., θ d to calculate.

[0037] The calculation of the model parameters (θ1, ..., θ) d ) can be performed iteratively multiple times. That is, it is performed with random starting parameters (θ1, ..., θ). d ) started. These are then iteratively improved by repeatedly checking all (yt1(i),...,ytm(i)) based on the respective current (θ1, ..., θ d ) are calculated and then new, improved (θ1, ..., θ) are derived from them. d), until a predefined measure of quality is reached (for example, the minimum of a cost function). Furthermore, only a subset of the total l data records can be processed in each iteration step. (yt1(i),...,ytm(i)), a so-called mini-batch, can be used to create new (θ1, ... ,θ d ) to calculate. This means that several iterations may be needed to consider all the training data, a so-called training epoch.

[0038] The learning algorithm is performed, for example, once when the radiometric measuring device is put into operation or repeatedly in real time during the operation of the radiometric measuring device, e.g. through additional reference measurements.

[0039] Fig. Figure 2 shows a highly schematic block diagram of the internal structure of an embodiment of a measurement quantity calculation unit 4 of the in Fig. 1 radiometric measuring device shown 1.

[0040] The measurement calculation unit 4 has an optional feature extraction unit 3, which is configured to extract data from the number n of sensor data x1, ...,x n , in particular based on the values ​​of the number d of parameters θ1, ..., θ d , to extract feature data FD.

[0041] The measurement unit 4 further includes an artificial intelligence (AI) unit 6, which is trained to determine the number m of measurement values ​​y1, ...,y m from the feature data FD based on the values ​​of the number d of parameters θ1, ..., θ d to calculate.

[0042] From the "raw" sensor data or measurement data x1,...,x nFirst, suitable features are extracted and transformed to generate the most meaningful input data possible for AI unit 6. In particular, one or more of the following techniques are used for this purpose: Principal Component Analysis (PCA), Discriminant Analysis, Statistical Normalization, Polynomial Transformation, Exponential Transformation, Logarithmic Transformation.

[0043] It goes without saying that feature extraction can also be omitted, so that the AI ​​unit 6 can use the unprocessed, raw sensor data x1, ...,x n used.

[0044] Depending on the measurement application, AI Unit 6 calculates either a continuous output signal (regression method) or a discrete output signal (classification method). It is implemented using an AI model from one of the following four categories: 1.) Models that use metrics or suitable similarity functions to compare the input values ​​with the stored training data in a single-stage or multi-stage manner and then assign them the output values ​​of those training data that are in some way “obvious” or similar. • This could be, for example, one of the following two AI models: ◯ k-Nearest Neighbor Classification ◯ k-Nearest Neighbor Regression • Metrics or similarity functions used may include, in particular: p-norm Minkowski Distance ◯ Kullback-Leibler Divergence 2.) Models that calculate thresholds from the training data, against which the given input values ​​are then compared in multiple stages, usually recursively, to determine the corresponding output values. • This could be, for example, one of the following two AI models: ◯ Decision Tree Classification ◯ Decision Tree Regression 3.) Models that estimate transition probabilities from the training data and combine these (possibly in multiple stages) additively and multiplicatively using Bayes' theorem to estimate a univariate or multivariate probability distribution on the output values ​​for given input values. The output values ​​with the highest probabilities are then assigned to the input values. • This could be, for example, one of the following two AI models: ◯ Bayes Classifier, especially Naive Bayes Classifier ◯ Bayesian Network Classifier 4.) Models which, using methods of linear algebra, apply so-called activation functions in one or more stages to linear combinations and / or convolutions of the transformed or untransformed input values ​​in order to then calculate the output values. • This could be, for example, one of the following two AI models: ◯ Multiclass Support Vector Machine (SVM) with One-Vs-One or One-Vs-All ▪ Kernel functions used may include, in particular: ◆ Polynomial Kernel ◆ Gaussian RBF Kernel ◆ Laplace RBF Kernel ◆ Sigmoid Kernel ◆ Hyperbolic Tangent Kernel ◆ Bessel Kernel ◆ Anova Kernel ◆ Linear Splines Kernel ◯ Artificial Neural Network (ANN) and / or Deep Neural Network (DNN) • Activation functions used may include, in particular: ◯ Identity ◯ Sigmoid ◯ Hyperbolic Tangent ◯ ReLu ◯ Softmax ◯ Signum

[0045] Referring to Fig. 3, which is a highly schematic block diagram of the internal structure of a further embodiment of a measurement quantity calculation unit 4 of the in Fig. As shown in section 1, a radiometric measuring device 1 represents, optionally a number of individual AI units from the 6 categories above can be combined into a more powerful overall model using ensemble learning. Fig. Three feature extraction units and three downstream AI units operate in parallel, with an ensemble combiner combining the respective data. For example, bagging or boosting can be used as an ensemble learning technique.

[0046] The model parameters (θ1, ..., θ) d The AI ​​unit(s) 6 are determined from the training data using machine learning, for example, using one of the following techniques: • by minimizing metrics once or repeatedly or by maximizing similarity functions. This can include, in particular, one or more of the following: ◯ Entropy ◯ Gini Impurity ◯ Variance p-norm Minkowski Distance ◯ Kullback-Leibler Divergence • By minimizing a cost function once or repeatedly, which depends on the chosen AI model and whose function arguments consist of the training data and the model parameters. The minimization may be subject to certain mathematical constraints (restriction of the search domain), which may also depend on the training data and / or model parameters. The cost function is minimized with respect to the model parameters using mathematical optimization methods and techniques, in particular one or more of the following: ◯ Backpropagation ◯ Gradient Descent based Method ◯ Stochastic Gradient Descent based Method (for example, AdaGrad, RMSProp or Adam) ◯ Gauss-Newton Method ◯ Quasi-Newton Method Linear Programming Quadratic Programming

[0047] Minimizing a cost function can involve maximizing a value function, in particular a maximum likelihood function or a maximum a posteriori probability function, especially by changing the mathematical sign.

[0048] To prevent overfitting, improve the ability to transfer learning, and thus increase the performance of the AI ​​unit(s) 6, additional regularization techniques can be used in the learning process, such as: • p-norm penalty terms (L1, L2, etc.) • Dropout • Batch Normalization

[0049] Unless otherwise defined, all AI terminology is to be understood according to standard academic literature on AI and machine learning. See in particular: 1. Bishop, Christopher M.: “Pattern Recognition and Machine Learning” 2. Mitchell, Tom M.: „Machine Learning“ 3. Russell, Stuart J. and Norvig, Peter: „Artificial Intelligence : A Modern Approach" 4. Richard O. Duda and Hart, Peter E. and David G. Stork: „Pattern Classification" 5. Aggarwal, Charu C.: „Neural Networks and Deep Learning: A Textbook“

Claims

[1] Radiometric measuring device (1), comprising: - a number n of sensors (2_1 to 2_n), wherein each sensor (2_i) of the number n of sensors (2_1 to 2_n) is configured to provide associated sensor data (x i ) to generate a total of n sensor data (x1, ...,x n ) is generated using the number n of sensors (2_1 to 2_n), - a measurement unit (4) designed to calculate a number m of measurement values ​​(ŷ1, ...,y) m ) depending on the number n of sensor data (x1, ...,x n ) based on values ​​of a number d of parameters (θ1, ..., θ d to calculate, and - a learning unit (5), wherein the learning unit (5) is trained to perform based on training data ((xt1(i),...,xtn(i));(ys1(i),...,ysm(i))) the values ​​of the number d of parameters (θ1, ..., θ d to calculate - where the number n of sensors (2_1 to 2_n) is selected from the set of sensors consisting of: - at least one radiometric sensor, in particular a radiometric sensor designed to generate sensor data in the form of a count rate or radiation intensity data, and / or a radiometric sensor designed to generate sensor data in the form of information about radiometric spectra, - at least one sensor designed to generate sensor data in the form of temperature data, - at least one sensor designed to generate sensor data in the form of acceleration data, - at least one sensor designed to generate sensor data in the form of speed data, - at least one sensor designed to generate sensor data in the form of position data, - at least one sensor, in particular in the form of an ultrasonic sensor or laser sensor, which is designed to generate sensor data in the form of information about a load height profile, and - at least one sensor designed to generate sensor data in the form of humidity data, - where at least one measurement variable is selected from the set of measurement variables consisting of: - Fill level, - Positions and / or thicknesses of individual material layers, - Density, - Delivery rate, in particular total delivery mass, - Throughput, especially mass flow rate, and - Material composition, in particular content levels, - wherein the learning unit (5) is trained to extract data from the training data ((xt1(i),...,xtn(i));(ys1(i),...,ysm(i))) a number n of training sensor data (xt1(i),...,xtn(i)) to extract and a number m of associated setpoints (ys1(i),...,ysm(i)) the number m of measured quantity values ​​(y1, ..., y m ) to extract, and - wherein the measurement unit (4) is configured to provide a number m of training values (yt1(i),...,ytm(i)) the number m of measured variable values ​​(y1, ...,y m ) depending on the number n of training sensor data (xt1(i),...,xtn(i)) to calculate, and - wherein the learning unit (5) is trained to, based on the number m of setpoints (ys1(i),...,ysm(i)) the number m of measured quantity values ​​(y1, ..., y m ) and the number m of training values (yt1(i),...,ytm(i)) the number m of measured quantity values ​​(ŷ1, ..., y m ) the values ​​of the number d of parameters (θ1, ..., θ d to calculate. [2] Radiometric measuring device (1) according to claim 1, characterized by , that - the measurement calculation unit (4) has at least one feature extraction unit (3) which is configured to extract from the number n of sensor data (x1, ...,x n ), in particular based on the values ​​of the number d of parameters (θ1, ..., θ d ), to extract feature data (FD), - wherein the at least one measurement calculation unit (4) is configured to determine the number m of measurement values ​​(ŷ1, ...,y) m ) from the feature data (FD) based on the values ​​of the number d of parameters (θ1, ..., θ d to calculate.

Images