Methods for detectors, detectors

Abstract

The invention takes the starting point from a method for a detector, which is provided for non-destructively detecting a detection signal of an object (16) arranged inside a test base (14). It is proposed to assign the detection signal in at least one method step to a detection result based on a machine learning process, thereby outputting the detection result and / or using it for adjusting the detector.

Description

Technical Field

[0001] A method for a detector and related to a detector.

Background Art

[0002] A method for a detector, in which the detector is provided for non-destructively detecting a detection signal of an object disposed inside an inspection base has already been proposed.

Summary of the Invention

[0003] The present invention starts from a method for a detector, in which the detector is provided for non-destructively detecting a detection signal of an object disposed inside an inspection base.

[0004] It is proposed to assign the detection signal to a detection result based on a machine learning process in at least one method step of the method, thereby outputting the detection result and / or using it for adjusting the detector. The method preferably includes an operation stage. In the operation stage, in at least one detection step of the method, the detection signal is detected by a sensor unit of the detector. In the operation stage, in at least one evaluation step of the method, the detection signal is evaluated by an evaluation unit. The evaluation unit is preferably an arithmetic unit of the detector, or alternatively, an external arithmetic device having a data link with the detector, particularly a data link connected by radio waves, particularly an Internet server. In the operation stage, the detection result is preferably output from an output unit of the detector and / or from an external output device, such as a smartphone, a tablet, etc., to a user of the detector. Additionally or alternatively, in the operation stage, by using the detection result by an arithmetic unit of the detector in an adjustment step of the method, the settings of the detector are changed, particularly repeating the detection step with settings of the detector, particularly settings of the sensor unit, that match the object and / or the background.

[0005] The method preferably includes a teaching step. In the teaching step, in the data acquisition step of the method, example values ​​(so-called training data) for the detection signal are collected. The example values ​​can be detected by the sensor unit of a detector, detected by the sensor unit of another detector, generated by simulation, read from a protocolized data bank, and / or similarly. In the learning step of the teaching step, the example values ​​are processed by a machine learning process to form a model that maps the detection signal to the detection result. The machine learning process is preferably carried out by a learning unit. The learning unit is preferably an external computing device, particularly as described above, and alternatively, the computing unit of the detector or other external computing device, such as a private server. Optionally, the method includes a preparation step, in which, after the teaching step, the model generated by the machine learning process is transmitted to the detector and filed in the memory of the evaluation unit, particularly the memory of the detector's computing unit. Preferably, in the use step, the evaluation unit performs an evaluation step using the model generated by the machine learning process. The teaching phase may be completed before the usage phase, or it may overlap in time with the usage phase, in which case the example values ​​on which the model is based are supplemented, in particular, with measured values ​​of the detection signals detected during the usage phase.

[0006] The machine learning process is an algorithm based on the field of machine learning, preferably starting from example values ​​to create a model. The machine learning process preferably includes a neural network and / or a classification method. The machine learning process may include supervised learning and / or unsupervised learning. The method preferably includes an assignment step, in which at least one evaluation parameter is assigned to an example value of the detection signal. The detection result may be identical to the evaluation parameter, may include multiple evaluation parameters, or may be determined depending on the evaluation parameter. The evaluation parameter for a single example value, and in particular the evaluation parameter as a whole, indicates, in the case of supervised learning during the learning process, what detection result the model should map the single example value to. In the case of unsupervised learning, the evaluation parameter is later assigned to the structure of the example value determined by the learning process. The assignment step may be performed manually or may be performed automatically during the course of the data acquisition step. The evaluation parameter may describe, for example, the properties of the inspection substrate, the type of object, the size of the object, the depth of the object within the inspection substrate, etc. The evaluation parameters may be known, for example, if they are detected on a known inspection substrate containing an object for which one example value of the detection signal is known. The inspection substrate and / or object may be determined, for example, by readjusting the detection conditions by the detector in a laboratory or by simulating the detection conditions based on a computer, and / or by reading from the construction plan. The evaluation parameters may be detected, for example, by other sensors, particularly invasive sensors, and / or sensors that cannot be incorporated into the detector due to limited installation space, and / or external sensor devices containing sensors more accurate than the detector's sensor unit.

[0007] The detector is configured to move along the inspection substrate, in particular to contact it and / or move along the inspection substrate, during the operation phase, in order to detect an object. "Non-destructive" should be understood in particular to mean that the inspection substrate and the object do not permanently change, in particular not to be damaged. In particular, the detector detects the object without invading it. Preferably, the sensor unit generates a signal in the detection step and transmits it into the inspection substrate, in particular as an electromagnetic wave or electromagnetic field, in which case the sensor unit detects the change in the signal by the inspection substrate and / or the object, in particular back scattering, as the detection signal. Alternatively or additionally, the detector is configured to detect a signal emanating from an object as the detection signal. The detector is configured in particular to detect objects such as metallic objects, non-metallic objects, electrical wires, in particular low-voltage wires, single-phase AC or DC current wires, wooden girders, metal beams or plastic pipes, in particular water-filled or unwater-filled plastic pipes, etc. The inspection substrate is, for example, the walls, ceilings, floors or fixtures of a building.

[0008] The configuration according to the present invention is advantageous because it allows for detailed evaluation of the detection signal. In particular, it is advantageous because it is possible to reliably extract from the detection signal auxiliary information that is contained in the detection signal and relates only to the mere presence of an object, or auxiliary information that relates only to the absence of an object. By utilizing the auxiliary information, the operation of the detector can be made advantageously simple, and / or, particularly with just one measurement by the detector, a large amount of information can be advantageously provided to the user of the detector. Furthermore, it is advantageous because it is possible to obtain advantageous and reliable repeatability of the detection results, and in particular because it is possible to reduce the dependence of the detection results on the distance the detector travels along the inspection substrate.

[0009] Furthermore, we propose classifying example values ​​for a detected signal into different groups, particularly during the teaching phase of the machine learning process already discussed, depending on the clustering of those example values. Preferably, the teaching phase includes a preprocessing step in which the learning unit edits the example values ​​before the learning step. In the preprocessing step, the learning unit advantageously performs unsupervised learning to classify the example values ​​into multiple groups. Advantageously, the learning unit classifies the example values ​​into at least two different groups in the preprocessing step. The learning unit determines the multiple groups and the assignments of the example values ​​to these groups, preferably using a clustering method. Preferably, in the preprocessing step, the learning unit extracts at least one feature, advantageously multiple features, from each example value to be processed. The feature may be, for example, a physical quantity, a statistical quantity, or an extracted code number of the example value. If an example value has similar values ​​for this feature / all features, the learning unit preferably assigns the example value to the same group. If an example value has different values ​​for this feature / one of the multiple features, the learning unit preferably allocates the example value to a different group. In this process, whether the values ​​of multiple features are similar or different is determined by the clustering method used. The configuration according to the present invention allows similar example values ​​with respect to the measured and extracted features to be processed together. This is particularly advantageous because all example values ​​within the same group can be assigned to the same evaluation parameter with a low error risk. It is also advantageous because it reduces the time required to perform the assignment step. Furthermore, example values ​​detected under unknown detection conditions can be assigned the evaluation parameter of example values ​​in the same group detected under known detection conditions.

[0010] Furthermore, to perform classification into different groups, we propose that the machine learning process uses nearest neighbor classification, particularly in the preprocessing step. Particularly advantageous for performing classification into different groups, the learning unit uses the K-Nearest Neighbor Algorithm (KNN). Specifically, the learning unit assigns one instance value to one group, and the assignment depends on which group a number of adjacent instances values ​​have been assigned to. Adjacent instances are a predetermined number of instances that have the shortest distance to the instance value to be classified in a parameter space spanned by multiple features. The distance between two instances in the parameter space spanned by multiple features can be determined using the Euclidean metric, the Manhattan metric, or other metrics. The number of adjacent instances used can be set by the operator of the learning unit and / or determined using an optimization algorithm. Optionally, in the case of majority voting of adjacent instances, the voting proportion of the instances is weighted, particularly by the distance between each adjacent instance value and the instance value to be classified. Alternatively, the example values ​​are classified into different groups using a block classifier, distance classifier, polynomial classifier, or other classifier. The configuration according to the present invention is advantageous because it can reliably classify multiple example values ​​into multiple groups.

[0011] Furthermore, we propose that during the teaching phase of the machine learning process, specific evaluation parameters, particularly those already discussed, be assigned to the specific example values ​​for the detection signals already discussed, group by group. Preferably, at least one example value is selected from each group as a representative example for that group. A representative example is an example value whose evaluation parameters are known from readjustment, simulation, or other sources. Preferably, the learning unit assigns the evaluation parameters, or a set of evaluation parameters, assigned to the representative example to all example values ​​within the group of the representative example. The representative example may be added to the remaining example values ​​in the data acquisition step, or it may be determined after classifying the remaining example values ​​into multiple groups. Optionally, the learning unit uses the representative example and / or domain knowledge as control points for nearest neighbor classification. Alternatively, the number of groups is automatically determined, for example, by variations in initial cluster centers and minimum total distance. The configuration according to the present invention is advantageous because it shortens the duration of the assignment step. In particular, it is advantageous because a large number of example values ​​whose evaluation parameters were not individually detected can be used for the machine learning process.

[0012] Furthermore, we propose that in the machine learning process, particularly in the teaching phase already discussed, the dataset consisting of example values ​​for the detection signal consists of at least substantially user data. "Substantial" should be understood as more than 30%, preferably more than 60%, particularly preferably more than 90%, and even more preferably more than 99%. Preferably, the dataset includes representative examples of the group in addition to user data. "User data" should be understood as detection signals detected by the user using a detector or additional detectors in the usage phase, particularly detected in real detection situations, i.e., in situations that are not readjusted or simulated detection situations. Preferably, the dataset is formed by collecting user data with a large number of detectors. In particular, user data is automatically transmitted to the learning unit via the detector's data interface. The configuration according to the present invention is advantageous because it reduces the time cost and expense of readjusting and / or simulating detection situations. Furthermore, the example values ​​are advantageous because they include many realistic detection situations and do not include only ideal detection situations that cannot be readjusted in the laboratory, particularly in terms of complexity and / or diversity. In particular, it is advantageous because it allows for the creation of robust models through machine learning processes.

[0013] Furthermore, we propose that in at least one method step, the detector's operating mode be automatically selected depending on a machine learning process. The detector, in particular the sensor unit, includes at least one operating mode provided for a specified detection situation and at least one other operating mode provided for other detection situations. The basis for distinguishing the multiple detection situations is, for example, differences in the material or structure of the inspection substrate and / or differences in the type of object being detected. The operating modes are distinguished, for example, by the signal parameters of the sensor element used by the sensor unit and / or the signal transmitted, such as intensity, frequency, duration, etc. Preferably, in one detection situation, the sensor unit detects an initial detection signal that is evaluated by an evaluation unit to form an initial result. Preferably, the detector's evaluation unit, in particular the calculation unit, performs all subsequent detection steps in the same detection situation in the selected operating mode by selecting one of the operating modes based on the initial result. Preferably, the memory of the detector's evaluation unit, in particular the calculation unit, contains a table for selecting the operating mode by assigning multiple operating modes to the initial result and reading them out by the evaluation unit. Alternatively, the detection result includes the operating mode provided for the detection situation as an evaluation parameter, which is learned during the machine learning process. The configuration according to the present invention is advantageous because it enables simple and intuitive operation of the detector. In particular, it eliminates the need for user adjustment of the detector. In particular, it is advantageous because it reduces the number of necessary operating elements of the detector and / or the number of menu items on the detector's graphic surface. In particular, it is advantageous because it reduces the risk of user error in operating the detector.

[0014] Furthermore, we propose that in at least one method step, a zero reference for the detection signal be automatically selected depending on a machine learning process. The zero reference is a measurement of the detection signal when no detectable object is present on the inspection substrate. Preferably, the evaluation unit determines whether an object is present on the inspection substrate by evaluating the difference between the detection signal and the zero reference. In standard cases, the zero reference depends on the material and structure of the inspection substrate. In particular, different operating modes of the detector have different zero references. One or more zero references are preferably part of a model created by a machine learning process. In particular, in the assignment step, one zero reference is assigned to each example value, and especially to each group of example values. Alternatively, multiple zero references are filed as a table in the memory of the calculation unit and read by the calculation unit depending on the selected operating mode. The configuration according to the present invention eliminates the need for the user of the detector to make a separate determination of the zero reference.

[0015] In particular, it is advantageous because the detector can be used quickly for object detection after power is applied. It is also advantageous because it reduces the risk of operator error with the detector.

[0016] Furthermore, when the detector assigns the measured value of the detection signal to one group of example values ​​of the detection signal, in particular, as previously discussed, we propose that in at least one method step, it outputs the evaluation parameters of the one group, in particular, as previously discussed. The evaluation unit preferably assigns the evaluation parameters to the measured value using a model. The output unit preferably outputs at least whether an object is detected or not. The output unit preferably outputs evaluation parameters in addition to the information on whether an object is detected or not. For example, the output unit outputs the material or structure of the inspection substrate as evaluation parameters. For example, the output unit outputs the type of object detected, the size of the object detected, in particular the maximum extent of the object detected, parallel to the surface of the inspection substrate, the distance between the object detected and the surface of the inspection substrate, in particular the distance between the edge of the object detected and the current measurement point of the detector, parallel to the surface of the inspection substrate, and / or similar matters as evaluation parameters. The configuration according to the present invention is advantageous because it can provide the user with a lot of information, especially in just one measurement.

[0017] Furthermore, we propose a detector comprising at least one sensor unit, particularly one of those previously discussed, for detecting the detection signal, and at least one arithmetic unit, particularly one of those previously discussed, for carrying out the method according to the present invention. The sensor unit includes at least one sensor element. Optionally, the sensor unit includes multiple sensor elements, particularly in different configurations. For example, for the sensor elements of the sensor unit, the radar includes a narrowband radar, particularly a narrowband radar and / or ultra-wideband radar in the frequency range of 2.4 GHz to 2.4835 GHz, an inductive sensor, a capacitive sensor, an AC sensor, particularly a 50 Hz AC sensor and / or a 60 Hz AC sensor, etc. The “arithmetic unit” should be understood in particular as a unit comprising an information input unit, an information processing unit, and an information output unit. Advantageously, the arithmetic unit has at least a processor, memory, input / output means, other electrical components, an operating program, a tuning routine, a control routine, and / or a calculation routine. Preferably, the components of the arithmetic unit are arranged on one common circuit board and / or, advantageously, in one common housing. Preferably, the detector includes an output unit for outputting the detection result. The output unit includes, for example, a display, microphone, and vibration alarm to output the detection results. Preferably, the detector includes a data interface. The data interface may have wired interface elements and / or wireless interface elements, particularly Bluetooth® interface, WLan interface, Ethernet interface, etc. The data interface is provided for outputting detection results to an external output device, for transferring detection signals to an external computing device, and / or for receiving models from an external computing device. The configuration according to the present invention allows for the use of a detector that is advantageously easy to operate and / or advantageously provides a large amount of information.

[0018] Furthermore, the methods and / or detectors according to the present invention are not limited to the uses and embodiments described above. In particular, the detectors and / or methods according to the present invention may have quantities different from those of the individual components, parts, units, and method steps mentioned herein in order to satisfy the functions described herein. Also, within the range of values ​​described herein, values ​​within the limits mentioned herein should also be considered as disclosed and available for use at will.

[0019] Further advantages will become apparent from the following description with reference to the drawings. The drawings illustrate one embodiment of the present invention. The drawings, the detailed description of the invention, and the claims contain a number of combinations of constituent elements. It is also appropriate for those skilled in the art to consider these constituent elements individually and combine them into meaningful further combinations. [Brief explanation of the drawing]

[0020] [Figure 1] This is a schematic diagram of the detector according to the present invention. [Figure 2] This is a flowchart outlining the method according to the present invention. [Figure 3] This is a schematic diagram showing the classification of a dataset of real-world values ​​during the machine learning process of the method according to the present invention. [Modes for carrying out the invention]

[0021] Figure 1 shows a detector 12. The detector 12 is shown positioned on the surface of an inspection substrate 14 (e.g., a wall). The detector 12 is provided to non-destructively detect the detection signal of an object 16 located inside the inspection substrate 14. The detector 12 is preferably formed to be hand-holdable, preferably with one hand, and especially operable with the same hand. Particularly advantageous is that the detector can reach 7000 cm⁻¹. 3 Less than 5000cm is advantageous. 3 Less than 3000cm is particularly advantageous. 3It has a total volume of less than . The detector 12 includes a sensor unit 24 for detecting a detection signal. The sensor unit 24 preferably includes at least one sensor element, which is provided to radiate an electromagnetic signal into the inspection substrate 14 and to detect the component of the radiated signal that has been scattered back from the object 16 as a detection signal. Alternatively or additionally, the sensor unit 24 includes a capacitive sensor and / or an inductive sensor for detecting the detection signal. The detector 12 includes a calculation unit 26 for evaluating the detection signal to obtain a detection result.

[0022] The detector 12 preferably includes an output unit 28 for outputting detection results. The output unit 28 preferably includes a display for showing the detection results. Preferably, the detector 12 includes at least one data interface 32 for data exchange with an external computing unit and / or external output device. The detector 12 preferably includes an energy supply unit 34 for supplying electrical energy to the sensor unit 24, computing unit 26, output unit 28 and / or data interface 32. The energy supply unit 34 is provided in particular for receiving at least one battery and / or accumulator. The detector 12 preferably includes at least one operating element 30 for operating the detector 12, in particular for initiating detection by the detector 12, such as a switch, key, slide controller, etc. The detector 12 includes a housing 36, inside which at least the sensor unit 24 is located. Preferably, the computing unit 26, data interface 32 and / or energy supply unit 34 are located inside the housing 36. Preferably, the output unit 28 and / or operating element 30 are located on the side of the housing 36 or housed inside the housing 36. An external computing unit and / or external output unit is preferably formed independently of the detector 12 (not shown herein).

[0023] Figure 2 shows a flowchart of method 10 for detector 12. Method 10 preferably includes a detection step 40 in which the sensor unit 24 detects a detection signal. Method 10 includes an evaluation step 42 in which the detection signal is evaluated depending on a machine learning process for the detection result. Method 10 particularly includes an output step 44 for outputting the detection result. Method 10 preferably includes an adjustment step 46 in which the arithmetic unit 26 changes the settings of detector 12, particularly the settings of sensor unit 24 and / or arithmetic unit 26, depending on the detection result.

[0024] Method 10 preferably includes an operation phase 38. The operation phase 38 is advantageously carried out completely, alternatively partially, using detector 12. The operation phase 38 preferably includes the detection step 40, the evaluation step 42, the output step 44 and / or the adjustment step 46. In particular, the evaluation step 42 is carried out by the arithmetic unit 26. Alternatively, the evaluation step 42 is carried out particularly by the arithmetic device already mentioned or by a further external arithmetic device. The output step 44 is preferably carried out by the output unit 28. Alternatively, the output step 44 is carried out by an external output device. In particular, in an alternative configuration, the data interface 32 sends the detection signal to the further external arithmetic device, receives the detection result from this external arithmetic device, and / or sends the detection result to the external output device.

[0025] Method 10 preferably includes an instruction phase 18. The instruction phase 18 is preferably carried out by an external arithmetic device, alternatively by the arithmetic unit 26. In the instruction phase 18, a model is created by a machine learning process for mapping the detection signal to a detection result. The model is filed in the memory of the arithmetic unit 26 of detector 12 in a preparation step 58. In particular, the arithmetic unit 26 uses the model in the evaluation step 42 to convert the measured value of the detection signal detected by the sensor unit 24 into a detection result.

[0026] The teaching step 18 preferably includes a data acquisition step 48. In the data acquisition step 48, the example values 22 (see FIG. 3) to be processed for creating a model are detected. Advantageously, the method 10 includes a collection step 56 of collecting the measured values of the detection signals detected by the sensor unit 24 as the example values 22. Preferably, the data interface 32 transmits the measured values detected by the sensor unit 24, particularly collectively or individually, to an external computing device in the collection step 56. Particularly advantageously, the external computing device collects the measured values from a number of detectors, particularly within the framework of the Internet-of-Things (IoT) concept. The data set consisting of the example values 22 of the detection signals consists at least substantially of user data. The user data are the measured values of the detection signals detected by the user using the detector 12 in the usage step 38. Preferably, in the data acquisition step 48, the manufacturer of the detector 12 additionally determines example values 22 as representative examples. The representative examples are determined, for example, by the detection by the detector 12 including different objects 16 in replicas of various inspection bases 14, and / or created by the simulation of the detection step 40 including different objects 16 in various inspection bases 14. Preferably, the data set includes more, at least 10 times more, advantageously at least 100 times more, particularly advantageously at least 1000 times more user data than the representative examples.

[0027] The teaching step 18 preferably includes a preprocessing step 50. In the preprocessing step 50, the example values ​​22 are evaluated with respect to at least one feature 60,62, preferably multiple features 60,62 (see Figure 3). At least one feature 60,62 is preferably provided as a discrimination criterion, and based on this criterion, the external computing unit determines whether two example values ​​22 are the same or different. The features 60,62 of the example values ​​may be, for example, the intensity of the detected signal, the ratio of the scattered and returned component to the transmitted component of the detected signal, the mean value of the detected signal, the extreme value of the detected signal, the duration of amplitude modulation caused by an object in the detected signal, the range of variation of the detected signal, etc. The example values ​​22 for the detected signal are classified into different groups 20 by the external computing unit, depending on the clustering of these example values ​​22. The external computing unit preferably evaluates the clustering of the example values ​​22 based on the detected features 60,62. In particular, the external computing unit determines whether the distribution of the example values ​​22 has clusters within a parameter space spanned by at least one feature 60,62. The number of dimensions of the parameter space is equal to the number of different features 60,62 per example value 22. The distribution of example values ​​22 within a two-dimensional parameter space spanned by one feature 60 and the other feature 62 is illustrated in Figure 3. Preferably, the external computing unit assigns example values ​​22 that form one cluster to the same group 20. Preferably, the external computing unit assigns example values ​​22 that form different clusters to different groups 20. In Figure 3, the example values ​​22 are illustrated by dividing them into 10 groups 20. The external computing unit uses nearest neighbor classification to perform the classification of example values ​​22 into different groups 20.

[0028] Method 10 preferably includes an assignment step 52. In the assignment step 52, at least one evaluation parameter is assigned to each group of example values ​​22 for the detection signal. Preferably, at least one evaluation parameter is assigned to each group 20. Optionally, multiple evaluation parameters are assigned to at least one group 20. Two different groups 20 are distinguished from each other by at least one evaluation parameter assigned to them. At least one evaluation parameter is processed by a machine learning process, in particular as an intermediate result to be learned or as a final result to be learned. Preferably, a group 20 is assigned a base parameter as an evaluation parameter that describes or characterizes the inspection base 14. The base parameter includes values ​​such as the thickness and / or quantity of concrete, lightweight construction, dry construction, especially gypsum board and / or wooden boards in lightweight or dry construction, brick walls, especially the type of stone in brick walls, and the presence or absence of underfloor or wall heating. Preferably, a group 20 is assigned a parameter as an evaluation parameter that describes or characterizes the type of object 16. Parameters relating to the type of object include, for example, values ​​for metal, nonmetal, conductive, nonconductive, magnetic, nonmagnetic, low-voltage cable, single-phase AC current cable, especially cables between 110V and 230V, multi-phase AC current cable, three-phase current cable, wooden beam, metal beam, plastic pipe, water-filled pipe, especially fresh water pipe, unfilled pipe, and especially sewage pipe. Preferably, group 20 is assigned a depth parameter as an evaluation parameter that describes or characterizes the distance from the surface of the inspection substrate 14 to the object 16. The depth parameter can be presented as distance continuously or in regions. Preferably, group 20 is assigned a shape parameter as an evaluation parameter that describes or characterizes the shape or extent of the object 16. The shape parameter indicates, for example, the diameter of the object 16. The assignment step 52 is performed by an external computing device using representative examples.The evaluation parameters of a representative example can be detected by the manufacturer of the detector 12 using a sensor device installed on the replica, independent of the detector 12, and then read from the datasheet to the replica, or read from the simulation program, and filed in the memory of an external computing device. Preferably, the external computing device determines at least one representative example for each group 20, reads its evaluation parameters from the memory of the external computing device, and assigns these evaluation parameters to all example values ​​22 within the same group 20 as representative examples. If there are multiple representative examples with different evaluation parameters within the same group 20, the external computing device optionally divides this group 20 into further subgroups.

[0029] The teaching stage 18 includes, in particular, the learning step 54. In the learning step 54, the external computing unit creates a model using a machine learning process. Specifically, the external computing unit uses the example values ​​22 as input values ​​for the model and the evaluation parameters assigned to the example values ​​22 as output values ​​for the model. The detection results are determined in part by the evaluation parameters as a whole, or depending on the evaluation parameters. In the preparation step 58, the external computing unit transmits the model to the computing unit 26 of the detector 12 via the data interface 32.

[0030] In the operation stage 38, the calculation unit 26 of the detector 12 utilizes a model to automatically select the operating mode of the detector 12. Preferably, in the detection step 40, the sensor unit 24 detects an initial detection signal which will be evaluated for initial results in the evaluation step 42. Depending on the initial detection signal, the calculation unit 26 automatically selects a zero reference to the detection signal in the adjustment step 46, depending on the machine learning process, by using the model to determine particularly fundamental parameters. Depending on the initial detection signal, the calculation unit 26 improves the sensitivity of the sensor unit 24 to the object 16 and / or avoids oversaturation of the sensor unit 24 in the adjustment step 46, by using the model to determine particularly object type parameters, depth parameters and / or shape parameters.

[0031] In the output step 44 during the usage stage 38, the detector 12 outputs at least one evaluation parameter to the same group 20 of example values ​​22 of the detected signal, if the model assigns the measured value of the detected signal to this group 20. The output unit 28 preferably outputs whether an object 16 has been detected. The output unit 28 preferably outputs parameters relating to the type of object and base parameters. The output unit 28 optionally outputs depth parameters and / or shape parameters. [Explanation of symbols]

[0032] 12 detectors 14. Inspection infrastructure 16 Object 18 Teaching Stages 20 Groups of actual values 22 Examples 24 Sensor Units 26 arithmetic units

Claims

1. A method for a detector to detect an object (16) inside an inspection board (14) by radiating an electromagnetic signal into the inspection board (14) and based on a detection signal which is the returned signal, Based on a model obtained by machine learning to correlate the value of at least one evaluation parameter indicating at least one of the object (16) and the inspection substrate (14) with the value indicated by the detection signal, an evaluation of the object (16) is performed from the detection signal. Output the results of the above evaluation, In the aforementioned machine learning, multiple example values ​​(22) corresponding to the values ​​of multiple detection signals are classified into multiple groups (20) by clustering. Each of the aforementioned plurality of groups (20) is associated with the value of at least one evaluation parameter, A method wherein the result of the evaluation includes the value of at least one evaluation parameter associated with one of a plurality of groups (20), to which one of a plurality of instance values ​​(22) corresponding to the value indicated by the detection signal is assigned.

2. In the machine learning, a zero reference is assigned to each of the plurality of groups (20) or each of the plurality of example values ​​(22), The method according to claim 1, wherein the zero reference is the value indicated by the detection signal when the object (16) is not present in the inspection substrate (14).

3. Based on the detection signal, select one of a plurality of operating modes set for a plurality of detection conditions, and operate the detector based on the selected one operating mode. The method according to claim 2, wherein each of the plurality of operating modes is a different zero-reference based mode.

4. The method according to any one of claims 1 to 3, characterized in that the nearest neighbor classification method is used in the machine learning to classify the plurality of example values ​​(22) into the plurality of groups (20).

5. The method according to any one of claims 1 to 3, wherein the plurality of example values ​​(22) include a value indicated by at least one of the detection signals.

6. The method according to any one of claims 1 to 3, characterized in that the at least one evaluation parameter includes at least one base parameter, at least one object parameter, at least one depth parameter, and / or at least one shape parameter.

7. A detector comprising at least one sensor unit (24) for detecting the detection signal and at least one arithmetic unit (26) for carrying out the method according to any one of claims 1 to 3.

Images