Device and automated method for evaluating sensor measurements and use of the device
By combining neural networks and least squares regression, the problem of initial parameter estimation dependence in sensor data evaluation is solved, achieving more accurate, faster and more reliable data evaluation, which is suitable for complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SIEMENS AG
- Filing Date
- 2021-03-23
- Publication Date
- 2026-06-23
Smart Images
Figure CN115398442B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an apparatus and automated method, such as a computer-implemented method, for evaluating sensor measurements, such as those in spectroscopy. The invention also relates to applications of this apparatus. Background Technology
[0002] Evaluation of sensor data is typically achieved through signal processing in a digital computer. Signal processing is often necessary because analog sensor hardware does not directly provide the desired parameters or information. This signal processing can range from simple linear temperature correction to spectral analysis or even complex image or video analysis.
[0003] Known systematic data evaluation methods are model-based data evaluation. Here, a computational model, also called a model function, exists for the behavior of sensor hardware, including the relevant physical environment, and depends on parameters of interest. This is used to evaluate the (digital) data (=measurement data) provided by the sensor hardware using least-squares regression. The parameter estimates obtained form or provide the sensor output values (=sensor output signals), and these parameter estimates are (in the least-squares sense) parameters that provide the best simulation of the measurements.
[0004] In general, nonlinear regression is implemented, and in special cases, linear regression is implemented when the model is linear with respect to all parameters. Mathematically, this corresponds to model inversion, which is typically performed using suitable, known iterative methods. The advantage of least squares regression is that it is asymptotically efficient under general assumptions. This means that this method of parameter determination is statistically optimal. There is no unbiased method that can provide smaller errors in parameter estimation. Therefore, model-based data evaluation using nonlinear regression is a systematic approach that can be appropriately used wherever sufficiently accurate sensor hardware or physical models exist or can be created relatively easily.
[0005] The least squares method is a standard mathematical approach for compensation calculations. It involves determining a function for a set of data points that approximates the direction of these points as closely as possible and thus summarizes them as well as possible. The most common function is a straight line, which is then called the fitted line. For this method to be applied, the function must contain at least one parameter. These parameters are then determined by the method such that when the function is compared to these data points and the squares of the distances between the function values and the data points are squared, the sum of these squared distances becomes as small as possible. This sum of squared distances is called the residual sum of squares. The distance vector between the function values and the data points is called the residual. Therefore, the residual sum of squares is the square of the absolute value of the (Euclidean) vector of the residuals.
[0006] This method is typically used to examine real-world data, such as physical data. This data usually contains unavoidable measurement errors and fluctuations. Assuming that the measured values are close to the "true values" on which they are based and that there is a relationship between these measurements, this method can be used to find a function that describes this relationship as accurately as possible. Conversely, this method can also be used to test different functions and thereby describe unknown relationships within this data.
[0007] However, there are problems with the computation of model inversion. Iterative methods (Levenberg-Marquardt or similar methods) are usually used, which require relatively good initial parameter estimates in order to determine the global optimum, i.e., the best possible fit between the model and the measurement vectors.
[0008] So far, heuristics have been used in least squares regression to determine initial values. This could be methods derived from a specific model using expert knowledge, or general heuristics such as Nelder-Mead. However, in the case of complex problems, the methods mentioned later can also fail.
[0009] Another approach to sensor data evaluation is to consider data evaluation based on KI (KI = Artificial Intelligence). Here, artificial neural networks, etc., are used to achieve this data evaluation through machine learning (keyword: "Reinforcement Learning"). A major challenge here is ensuring reliability. Thus, if the KI obtains measurement data outside its training range, it is often impossible to reliably predict how the KI will perform. Consequently, error analysis or certification for safety-critical tasks cannot be easily implemented.
[0010] However, there are problems that can only be solved by KI and no other signal processing methods are available, such as in tasks from the field of image-based object recognition. Summary of the Invention
[0011] The objective of this invention is to describe a solution for an improved evaluation of sensor measurements.
[0012] The task is accomplished by a device according to the invention for evaluating sensor measurements or by an automated method according to the invention for evaluating sensor measurements. Advantageous extensions and design schemes are the subject of different embodiments. Other features, uses, and advantages of the invention will become apparent from the following description.
[0013] This invention combines artificial intelligence using neural networks to evaluate sensor measurements with least squares (LS) regression. It is assumed below that a sufficiently accurate computational model (i.e., a model function) exists for evaluating sensor data, as is required, especially for least squares regression. Using this computational model, any number of scenarios corresponding to actual measurements can be simulated. This model function is described by parameters forming a parameter vector. Now, random parameter vectors are generated from the parameter space, and the model function is evaluated. These parameter vectors form the ground truth.
[0014] Now, a neural network designed to predict these parameter vectors is trained together with the result vector of the model function. This is a classic reinforcement learning approach for KI-based data evaluation. In a specific measurement task, for each real measurement, the measurement vector (composed of measured values) is first determined using KI to identify an initial parameter estimate vector. This initial parameter estimate vector is used as the initial value for subsequent least squares regression. If the least squares regression converges and the residuals meet specified criteria, such as a norm less than a specified limit, the data evaluation is marked as "successful". The parameter estimates of the LS regression (or a selection of these parameter estimates) form one or more sensor output signals. If the data evaluation is unsuccessful, an error message (= "unsuccessful") is output.
[0015] For the purposes of this invention, "neural network" refers to any device suitable for machine learning.
[0016] This invention claims a device for evaluating sensor measurements, the device having:
[0017] - A sensor, wherein, in order to evaluate the sensor's sensor measurements, a model function suitable for least squares regression is provided, which can be defined by a parameter vector, wherein at least one parameter of the parameter vector forms the sensor's output signal;
[0018] and
[0019] - A computation and evaluation unit having a neural network and a least-squares regression module for estimating the parameter vector based on real-world sensor measurements, wherein the neural network is trained using the parameter vector and associated sensor measurements, and the computation and evaluation unit is configured as follows:
[0020] °- For the sensor measurements taken using this sensor, at least one parameter estimation vector is determined using a trained neural network as the input parameter for the least squares regression of this least squares regression module.
[0021] °- Discontinue the least squares regression if the convergence criteria are met when performing it, and
[0022] °- Output at least one parameter from the final determined parameter vector in the least squares regression with minimum squared error as the sensor output signal.
[0023] A sensor, also known as a detector, measurement transducer, or inductor, is a technological component that can qualitatively detect or quantitatively assess the physical or chemical properties of its environment (such as heat, temperature, humidity, pressure, sound field magnitude, brightness, acceleration, pH, ionic strength, or electrochemical potential) and / or material properties as measurement parameters. These parameters are detected by means of physical or chemical effects and converted into assessable electrical signals.
[0024] In an extended embodiment, the calculation and evaluation unit may have an evaluation module downstream of the least squares regression module. This evaluation module is configured to determine a success state of the evaluation based on the least squares regression residuals, information about the interruption state of the least squares regression, and at least one other piece of information about the least squares regression, and output this success state as a separate sensor output signal, where the success state can be either "successful" or "unsuccessful." The success state is a binary parameter.
[0025] In another design, the evaluation module can be configured to additionally consider at least one parameter of the finally determined parameter vector and / or these sensor measurements in order to determine the success state.
[0026] In another implementation, the evaluation module can be configured to determine the quality information of the evaluation based on the residuals of the least squares regression, information about the interruption state of the least squares regression, and at least one other piece of information about the least squares regression, and output the quality information as another sensor output signal.
[0027] This quality information, also known as "Quality of Sensing," is a continuous, non-negative scalar variable.
[0028] Preferably, the quality information is the Euclidean norm of the residual or the dimensionless normalized Euclidean norm of the residual.
[0029] In another embodiment, the evaluation module is configured to set the success status to "success" when the quality information is below a specified quality threshold.
[0030] In another implementation, the evaluation module can be configured to select a signal from multiple least-squares regression signals, for example, the signal with the lowest squared error in the case where the success state is "success".
[0031] The present invention also claims protection for the application of the apparatus according to the present invention for evaluating chromatograms in gas chromatography.
[0032] The present invention also claims protection for the application of the device according to the invention for spectral evaluation in spectroscopy.
[0033] The present invention also claims protection for the application of the device according to the invention for spectral evaluation of time series.
[0034] This application is used, for example, to evaluate measured voltage / current, ultrasonic vibration, etc., and for checking or determining the condition of technical equipment or instruments.
[0035] The present invention also claims protection for applications of the device according to the invention for analyzing audio data, such as speech.
[0036] The present invention also claims protection for the application of the device according to the invention for identifying objects in image data, such as in the identification of automated components in production.
[0037] This invention also claims an automated method for evaluating sensor measurements:
[0038] - To evaluate these sensor measurements, a model function suitable for least-squares regression, constrained by a parameter vector, is provided, wherein the sensor output signal is formed by at least one parameter of this parameter vector.
[0039] and
[0040] - This includes a neural network and a least-squares regression module that estimate the parameter vector based on real-world sensor measurements. The neural network is trained using the parameter vector and relevant sensor measurements.
[0041] The measured sensor values are used to determine at least one parameter estimation vector as input parameters for the least squares regression module using a trained neural network. The least squares regression is interrupted if the convergence criterion is met during its implementation, and at least one parameter from the last determined parameter vector in the least squares regression with the least squares error is output as the sensor output signal.
[0042] In an extended scheme, the success state of the evaluation can be determined based on the residuals of the least squares regression, information about the interruption state of the least squares regression, and at least one other piece of information about the least squares regression, and the success state can be output as another sensor output signal, wherein the success state can be "successful" or "unsuccessful".
[0043] In an extended scheme, to determine the success state, at least one parameter of the finally determined parameter vector and / or these sensor measurements may be additionally considered.
[0044] In another design, based on information about the interruption state of least squares regression and at least one other piece of information about least squares regression, the quality information of the evaluation can be determined and output as another sensor output signal.
[0045] The combination of artificial intelligence using neural networks as the initial estimator and LS regression as a "refinement" with the final test has the following advantages:
[0046] 1. If the regression is "successful," then the data assessment provides a high degree of certainty in providing the correct parameter values. That is, these parameter values are identical to the physical true values, except for noise.
[0047] 2. Compared to pure LS regression: This method is more robust because it does not require separate initial value estimation.
[0048] 3. Compared to pure LS regression: This method is faster because LS fitting only requires a small number of additional iterations.
[0049] 4. Compared to pure KI: This method is validated, meaning that the result parameters are checked using a (validated) model to see if these result parameters match the measurement vector.
[0050] 5. Compared to pure KI: This method provides more accurate results. Compared to LS regression, which is an asymptotically efficient estimator, KI typically cannot achieve very high accuracy in estimating these parameters alone.
[0051] Other features and advantages of the invention will become apparent from the following description of embodiments with reference to the illustrative drawings. Attached Figure Description
[0052] in:
[0053] Figure 1 A block diagram of a device for evaluating sensor measurements is shown.
[0054] Figure 2 A flowchart is shown for a method to evaluate sensor measurements. Detailed Implementation
[0055] Figure 1 A block diagram of an apparatus for evaluating sensor measurements 1.1 is shown. Sensor 1 generates sensor measurements 1.1, which are used as input signals for calculation and evaluation unit 2. To evaluate the sensor measurements 1.1 of sensor 1, a model function suitable for least squares regression is provided, which can be defined by a parameter vector. At least one parameter of the parameter vector forms the sensor output signal 3. Sensor output signal 3 and other sensor output signals are output and presented on display unit 4.
[0056] The calculation and evaluation unit 2, such as a computer, has a neural network 2.1 that estimates the parameter vector based on the truly determined sensor measurement 1.1, and a least squares regression module 2.2. The neural network 2.1 is trained using the parameter vector and the associated sensor measurement 1.1. The calculation and evaluation unit 2 is configured, i.e., designed and programmed, to determine at least one parameter estimation vector as input parameters for the least squares regression module 2.2 using the trained neural network 2.1 for the sensor measurement 1.1 measured by sensor 1. If the convergence criterion is met during the implementation of the least squares regression, the least squares regression is interrupted, and one or more parameters of the finally determined parameter vector are output as the sensor output signal 3.
[0057] The convergence criterion could be, for example, a specified threshold below the sum of squares of the residuals from least squares regression, a specified maximum number of iterations, or a specified maximum time.
[0058] The calculation and evaluation unit 2 also has an evaluation module 2.3, which is downstream of the least squares regression module 2.2. Inputs to the evaluation module 2.3 include, for example, the residuals of the final model evaluation of the least squares regression in the least squares regression module, information about the interruption status of the least squares regression, and at least one other piece of information about the least squares regression. Based on these inputs, the evaluation module 2.3 determines the success status of the evaluation and outputs this success status as another sensor output signal 3, where the success status can be "successful" or "unsuccessful". In determining the success status, at least one parameter of the final determined parameter vector of the least squares regression and / or the sensor measurement 1.1 may be additionally considered.
[0059] Evaluation module 2.3 can also be configured to determine the quality information of the evaluation based on the residuals of the least squares regression, information about the interruption state of the least squares regression, and at least one other piece of information about the least squares regression, and output this quality information as another sensor output signal 3. This quality information (also referred to as "Quality of Sensing") is a continuous, non-negative scalar variable. This quality information can be, for example, the Euclidean norm of the residuals or the dimensionless normalized Euclidean norm of the residuals of the selected least squares regression.
[0060] The evaluation module can also be configured to set the success status to "success" when the quality information is below a specified quality threshold.
[0061] A variation of the above determination of this quality information is that the range of the Euclidean norm is restricted. If, for example, the relevant information is known to lie within a certain range of the measurement vector, then the measurement vector can be selectively chosen and the deviation between the model and the measurement can only be examined within that range. The range of relevant information can be output as an "auxiliary parameter" from the previous model evaluation.
[0062] Another variation involves applying weighting factors (->vectors) to the residual before forming the Euclidean norm. This results in a "soft" selection compared to "hard" masking (the former case). These weighting factors are also used as auxiliary variables in the output of the previous evaluation of the model function.
[0063] Another variant specifies that the algorithm is associated with the interrupted states of least squares regression, for example via a heuristic rule framework that additionally evaluates certain interrupted state events negatively.
[0064] The described device can primarily be used for:
[0065] 1. Evaluation of chromatograms in gas chromatography. Here, good initial parameter values for least squares regression are very important because there are multiple local minima in LS regression tasks due to multiple peaks, and only one of these local minima is the correct global optimum, i.e., the convergence of typical LS regression, i.e., its algorithm is not robust.
[0066] 2. Spectral evaluation in high-resolution spectroscopy, such as spectroscopy based on tunable lasers. Here, good initial values for the parameters used in LS regression are also very important because there are multiple local minima due to spectral fingerprints, and only one of these local minima is the correct global optimum, meaning that the convergence of typical LS regression algorithms is not robust.
[0067] 3. Spectral evaluation of time series, such as measured voltage / current, ultrasonic vibrations, etc., for condition checking or determination of technical equipment or instruments. In the spectral data of time series of physical signals, resonances (i.e., peaks) are often present. These resonances typically follow specific patterns of how each peak may have overtones. Because multiple fundamental resonances may exist, the resulting spectrum can appear very complex. If a general model is to be adapted, the fundamental resonant frequencies must first be known. Identifying these fundamental resonant frequencies is a difficult problem, complicated by noise and the possibility of other interfering signals. In short, if a model is to be adapted, and in which the parameters affecting these fundamental resonant frequencies are to be adapted, initial estimates of these parameters are absolutely necessary for successful LS. An example is the condition monitoring of the current of a motor of unknown size and speed.
[0068] 4. Analysis of audio data, such as speech. Here, the KI of the neural network can perform speech recognition and determine other parameters used for speech synthesis. The physical model here is a module suitable for speech synthesis. This physical model should be programmable so that the speech spoken by the speaker can be simulated sufficiently accurately with the help of other parameters. The verification step is carried out by comparing the measurement with the synthesized signal.
[0069] 5. Object recognition in image data. In industrial applications, there are often well-developed models of objects of interest (CAD, etc.). That is, a scenario with appropriate variations in (disturbing) background can be simulated a priori. Parameters such as the orientation of one or more objects are used to calculate the parameters of the model and their orientation in space. A neural network is trained to estimate at least these parameters. Then, LS regression is used to improve the estimate and then validation is performed. LS regression operating on image data is not fundamentally different from (one-dimensional) nonlinear regression. The model is compared point-by-point with measurements (the recorded images), and then the mean squared error is calculated.
[0070] In the case of image data analysis, other more suitable test criteria can also be used, such as weighting the model and measurements before calculating the squared difference. This weighting function can then more strongly weight the ranges of the effective signal that have strong amplitudes or contain the sought-after "information." This can be used to suppress interference outside the range of interest and thus reduce the rejection rate.
[0071] Figure 2 A flowchart is shown of an automated, for example computer-implemented, method for evaluating sensor measurements. In a first step 101, to evaluate these sensor measurements, a model function suitable for LS regression, definable by a parameter vector, is provided, wherein the sensor output signal is formed by at least one parameter of the parameter vector. In a second step 102, a neural network and a least-squares regression module are provided to estimate the parameter vector based on truly determined sensor measurements, wherein the neural network was trained in the previous step 100 using the parameter vector and the associated sensor measurements.
[0072] In the third step 103, at least one parameter estimation vector is determined using a trained neural network based on the measured sensor values as input parameters for the least squares regression module. In the following fourth step 104, least squares regression is performed, and the regression is terminated if the convergence criterion is met. Then, in the fifth step 105, at least one parameter of the finally determined parameter vector is output as the sensor output signal.
[0073] In step 6, 106, the success status of the evaluation is determined based on the residuals of the least squares regression, information about the interruption status of the least squares regression, and at least one other piece of information about the least squares regression. In step 7, 107, the success status is output as another sensor output signal, wherein the success status can be "successful" or "unsuccessful".
[0074] In order to determine the success state in step 6 106, at least one parameter of the finally determined parameter vector and / or these sensor measurements may be additionally considered.
[0075] In step 8, 108, the quality information of the evaluation is determined based on the information about the interruption state of least squares regression and at least one other piece of information about least squares regression, and in step 9, 109, the quality information is output as another sensor output signal.
[0076] Although the present invention has been further illustrated and described in detail by way of embodiments, the present invention is not limited to the disclosed examples, and other variations can be derived by those skilled in the art without departing from the scope of protection of the present invention.
[0077] List of reference numerals
[0078] 1. Sensor
[0079] 1.1 Sensor Measurement Values
[0080] 2. Calculation and Evaluation Unit
[0081] 2.1 Neural Networks
[0082] 2.2 Least Squares Regression Module
[0083] 2.3 Evaluation Module
[0084] 3. Sensor output signal
[0085] 4 Display Units
[0086] 100 Training Steps
[0087] 101 First step (providing model functions)
[0088] 102 Second step (providing neural network and LS regression modules)
[0089] 103 Third step (determining the parameter estimation vector)
[0090] 104. Step Four (Interrupting LS Regression)
[0091] 105 Fifth Step (Output Parameters)
[0092] 106. Step Six (Confirm Success Status)
[0093] 107. Step Seven (Output Success Status)
[0094] 108 Step 8 (Determining Quality Information)
[0095] 109 Step 9 (Output quality information).
Claims
1. An apparatus for evaluating sensor measurements (1.1), said apparatus having: - A sensor (1) having a model function that is adapted to least squares regression for evaluating sensor measurements (1.1) of the sensor (1) and can be defined by a parameter vector, wherein at least one parameter of the parameter vector forms the sensor output signal (3). and - A calculation and evaluation unit (2), the calculation and evaluation unit having a neural network (2.1) and a least squares regression module (2.2) for estimating the parameter vector based on real-determined sensor measurements (1.1), wherein the neural network (2.1) is trained using randomly selected parameter vectors and relevant sensor measurements determined by means of the model function, and the calculation and evaluation unit is configured as follows: °- Based on the sensor measurements (1.1) measured using the sensor (1), at least one parameter estimation vector is determined using a trained neural network (2.1) as the input parameter for the least squares regression of the least squares regression module (2.2). °- Discontinue the least squares regression if the convergence criterion is met during its implementation, and °- Output at least one parameter from the final determined parameter vector in the least squares regression with the least squares error as the sensor output signal. Its features are, The calculation and evaluation unit (2) has an evaluation module (2.3) downstream of the least squares regression module (2.2), the evaluation module being configured to: determine the success state of the evaluation based on the residuals of the least squares regression, information about the interruption state of the least squares regression, and at least one parameter of the last determined parameter vector of the least squares regression, and / or based on the sensor measurement (1.1), and output the success state as another sensor output signal (3), wherein the success state can be "success" or "unsuccess".
2. The device according to claim 1, wherein The evaluation module (2.3) is configured to determine the quality information of the evaluation based on the residuals of the least squares regression, information about the interruption state of the least squares regression, and at least one other information of the least squares regression, and output the quality information as another sensor output signal (3).
3. The device according to claim 2, wherein The quality information is the Euclidean norm of the residual or the dimensionless normalized Euclidean norm of the residual.
4. The device according to claim 2 or 3, wherein The evaluation module (2.3) is configured to set the success status to "success" when the quality information is lower than the specified quality threshold.
5. The apparatus according to any one of claims 1 to 3, wherein the apparatus is used to evaluate chromatograms in gas chromatography.
6. The apparatus according to any one of claims 1 to 3, wherein the apparatus is used for spectral evaluation in spectroscopy.
7. The device according to any one of claims 1 to 3, wherein the device is used for spectral evaluation of time series.
8. The device according to any one of claims 1 to 3, wherein the device is used to analyze audio data.
9. The device according to any one of claims 1 to 3, wherein the device is used to identify an object in image data.
10. An automated method for evaluating sensor measurements (1.1): - Wherein, in order to evaluate the sensor measurement (1.1), a model function (101) suitable for least squares regression and definable by a parameter vector is provided, wherein the sensor output signal (3) is formed by at least one parameter of the parameter vector. and - This includes (102) a neural network and a least-squares regression module that estimate the parameter vector based on truly determined sensor measurements, wherein the neural network is trained (100) using randomly selected parameter vectors and relevant sensor measurements determined by the model function, and The measured sensor values are determined by means of a trained neural network (103) at least one parameter estimation vector as input parameters for the least squares regression of the least squares regression module. The least squares regression is interrupted (104) if the convergence criterion is met when the least squares regression is performed, and at least one parameter from the last determined parameter vector in the least squares regression with the least squares error is output as the sensor output signal (3) (105). Its features are, Based on the residuals of the least squares regression, information about the interruption state of the least squares regression, and at least one parameter of the last determined parameter vector of the least squares regression, and / or based on the sensor measurement (1.1), the success state of the evaluation is determined (106) and the success state is output as another sensor output signal (107), wherein the success state can be "success" or "unsuccess".
11. The method of claim 10, wherein Based on information about the interruption state of the least squares regression and at least one other piece of information about the least squares regression, determine (108) the quality information of the evaluation and output the quality information as another sensor output signal (109).