Computer-implemented method for training a base model with further training data

Abstract

The invention relates to a computer-implemented method comprising the following steps: a) training a base model (20) used for processing sensor data, which base model was generated on the basis of an original training data set (28) by means of machine learning, with further training data (22), in order to obtain a modified model (24); b) testing the modified model (24) with a new training data set (26) on the basis of the original training data set (28) and the further training data (22) in order to obtain a test result (30) of the modified model (24); c) deciding, on the basis of the test result (30) of the modified model (24), whether the basic model (20) or the modified model (24) is used.

Description

[0001] Computer-implemented method for training a base model with additional training data

[0002] Description

[0003] The invention relates to the updating of (mathematical) models generated by machine learning for processing sensor signals, in particular sensor signals from maritime sensors, using further training data.

[0004] Models trained using machine learning are also referred to as artificial intelligence (AI). These models are initially trained with a dataset and, if the model delivers sufficiently good results, are used according to their intended purpose. After the initial training, however, the model is often updated with further training data. This makes it possible to adapt the model to, for example, changed conditions. However, updating the model can also lead to a decrease in its overall quality. This can occur, for example, due to poor training data or over-adaptation of the model.While a degraded model in consumer electronics may at most affect the user experience, a degraded model used to process sensor data for controlling autonomous machines (especially autonomous vehicles) or to obtain environmental information from machines can cause significant damage.

[0005] Especially when maritime sensor signals, such as sonar signals or signals from other sensors used on watercraft, are processed by the model, no compromise should be made regarding the model's quality. For example, autonomous watercraft may experience degraded collision detection, or manned underwater vehicles may receive incorrect environmental information while diving. This can lead to life-threatening injuries, such as those caused by collisions. The object of the present invention is therefore to provide an improved concept for training mathematical models to process sensor signals using machine learning.

[0006] The problem is solved by the subject matter of the independent patent claims. Further advantageous embodiments are the subject matter of the dependent patent claims.

[0007] Exemplary embodiments show a computer-implemented method with the following steps: a) Using a basic mathematical model, trained by machine learning using an initial training dataset (28), to process sensor data and provide information for controlling a platform in a maritime environment based on the sensor data. That is, a finished, ready-to-use mathematical model exists. A model describes the architecture of one or more algorithms. The model is used on a platform in a maritime environment. Examples of platforms in a maritime environment include watercraft, e.g.manned submarines, unmanned underwater vehicles (AUV - autonomous underwater vehicle, ROV - remotely operated underwater vehicle), manned ships, unmanned surface vehicles (USV - unmanned surface vehicle, ASV - autonomous surface vehicle), offshore wind turbines, drilling platforms or aircraft over the water or coast.

[0008] After the model has already been used, step b) trains the base model with additional training data to obtain a modified model. This means that a fully trained base model is already available and receives an update with the additional training data. This update can be performed particularly when suitable, labeled training data is available, i.e., data containing the result corresponding to the sensor data. The sensor data could be data from sensors used to create a situational awareness picture of a watercraft. This further training allows the mathematical model to keep its capabilities up-to-date. For example, this is necessary if certain input and / or output information was not yet known at the time of training with the original training data.For example, the development of new watercraft could be mentioned, on which the mathematical model can be trained with the update. Improving the model by acquiring additional training data beyond the original training dataset can also be achieved in this way.

[0009] Step c) follows, testing the modified model with a new training dataset based on the original training dataset and the additional training data to obtain a test result for the modified model. This means that the training data used to train the base model so far are combined to obtain the test result. This step is unusual, as a separate test dataset, independent of the training data, is typically used to determine the quality of the base model. If a validation dataset is also available, it can be used for testing as well. The modified model should be tested on as much of the available data as possible, ideally all of it. It should be noted that this is supervised learning, meaning that the data available for use must produce the correct output (label) of the model.This labeling often has to be done manually.

[0010] Step d) follows: Decide, based on the test result of the modified model, whether to use the base model or the modified model. For example, the test result of the existing base model can be compared with the test result of the modified model to decide whether to continue using the base model or the modified base model. That is, the modified model can be used if the test result is equal to or better than a test result of the base model. Preferably, the test result of the base model is generated at least on the basis of the original training data. That is, on all data with which the base model was trained. Preferably, this includes the original training data as well as any training data used for previous updates of the base model. Ideally, the base model is tested with the same data with which the modified model is tested.with the new training dataset. The idea is to re-examine the quality of a modified model already in use, rather than assuming that it has automatically improved due to the larger test dataset. It is particularly advantageous if the quality of the modified model is at least identical to, and ideally better than, the quality of the previously used base model. To allow an update of the base model even with only a small amount of training data, the training data is also included in the test. This method limits the statement about the predictive ability of the base model or the modified model. However, it can, for example, determine whether the modified model recognizes the old data with the same accuracy as the original base model.

[0011] Examples show that in step d), the training data is extended with additional training data if the base model continues to be used, and steps b) to d) are executed after the extension until, based on the test results, a decision is made to use the modified model. That is, new training data is collected until the training of the modified model performs better than the base model. It should be noted that the modified model can receive further updates. Accordingly, the modified model then becomes the base model for the next update.

[0012] If the modified model is used after training, the new training data becomes part of the original training data in some implementation examples. That is, if the modified model receives an update, the (new) original training data also includes the new training data used for training the modified model (then the base model).

[0013] In exemplary embodiments, the (basic) model is designed to process maritime sensor signals.

[0014] In exemplary embodiments, the (basic) model is configured to process received radar signals or received (active or passive) sonar signals (e.g., as output signals from underwater transducers) and to detect, and in particular classify, objects recognizable in the signals. In other words, the (basic) model can automatically generate a situational awareness picture for a vessel or at least propose one. This proposal can be reviewed by a human operator. Specifically, the (basic) model can detect a sea mine as an object in a received sonar signal. Furthermore, the (basic) model can issue an indication of the detection of a hostile object. In particular, knowing the hostile object, the (basic) model can also recommend a suitable weapon for engaging it.

[0015] In further embodiments, the (basic) model is designed to control an autonomous underwater vehicle based on sensor data. That is, the (basic) model can automatically create a situational image of the autonomous underwater vehicle and, based on this image, output control signals to influence the autonomous underwater vehicle, for example, to control a propulsion and control unit of the autonomous underwater vehicle.

[0016] Similarly, a data processing unit is disclosed which is configured to perform the steps of the computer-implemented method according to one of the preceding claims.

[0017] Furthermore, analogously, a computer program comprising instructions that, when the program is executed by a computer, cause it to perform the steps of the described computer-implemented procedure is disclosed.

[0018] Preferred embodiments of the present invention are explained below with reference to the accompanying drawing. It shows:

[0019] Fig. 1 : a schematic block diagram of a computer-implemented procedure for updating a trained (base) model.

[0020] Before exemplary embodiments of the present invention are explained in detail below with reference to the drawings, it should be noted that identical, functionally equivalent or equivalent elements, objects and / or structures in the different figures are provided with the same reference numerals, so that the description of these elements shown in different exemplary embodiments is interchangeable or can be applied to one another.

[0021] Fig. 1 shows a schematic block diagram of a computer-implemented method for updating a base model 20, generated by machine learning with an initial training dataset, using new training data 22. In step a), a mathematical base model, trained by machine learning using an initial training dataset (28), is used to process sensor data and, depending on the sensor data, to provide information for controlling a platform in a maritime environment. In step b), the base model 20 is trained with further training data 22 to obtain a modified model 24. In step c), the modified model 24 is tested with a new training dataset 26 based on the original training dataset 28 and the further training data 22 to obtain a test result 30 of the modified model.In step d), based on test result 30 of the modified model 24, a decision is made as to whether the base model 20 or the modified model 24 will be used to process the next sensor data. This decision can optionally be made based on a test result 32 of the base model. For example, the modified model 24 will be used if test result 30 of the modified model is better than test result 32 of the base model.

[0022] If the base model 20 continues to be used, optional case A occurs. If the modified model 24 continues to be used, optional case B occurs. In case A, the additional training data 22 is extended by new additional training data 22', and steps a) to c) are executed after the extension until a decision is made, based on the test result, to use the modified model. In case B, the new training data 22 becomes part of the original training data 28. That is, the new original training data now includes the old original training data 28 as well as the additional training data 22 if an update of the modified model 24 is to be performed.

[0023] The disclosed (underwater) sound transducers are designed for use underwater, particularly in the sea. The transducers can convert underwater sound into an electrical signal (e.g., voltage or current) corresponding to the sound pressure, the (received) underwater sound signal. Furthermore, it is possible for the transducers to convert an applied electrical voltage into underwater sound. The electrical voltage can follow a predefined pattern and then be referred to as the (transmitted) sonar signal, while the underwater sound resulting from the sonar signal to be transmitted is referred to as the (transmitted) sonar signal. Examples of sonar signals are a chirp (frequency-modulated signal) or, as a special case of the chirp, a sweep (linearly frequency-modulated signal). The transducers can therefore be used as underwater sound receivers and / or underwater sound transmitters.The transducers can be made of a piezoelectric material, such as a piezoceramic, to serve as the sensor material. A plurality of underwater transducers, or one or more underwater transducers in conjunction with a signal processing unit, can be referred to as a sonar system. The transducers can be used for (active and / or passive) sonar (sound navigation and ranging). The transducers are preferably not suitable for, or are not used for, medical applications. Likewise, the transducers are preferably not used for, or are not suitable for, the ultrasonic testing of materials.

[0024] Although some aspects have been described in connection with a device, it is understood that these aspects also constitute a description of the corresponding method, such that a block or component of a device is also to be understood as a corresponding method step or as a feature of a method step. Analogously, aspects described in connection with or as a method step also constitute a description of a corresponding block, detail, or feature of a corresponding device. Exemplary embodiments of the invention are preferably implemented in software. The implementation can be carried out using a digital storage medium, for example, a magnetic or optical storage medium, on which electronically readable control signals, e.g.,A computer program is stored that can interact with a programmable computer system (CPU and / or GPU) in such a way that the respective procedure is carried out. For this reason, the digital storage medium should be computer-readable. Exemplary embodiments may therefore include a data carrier that has electronically readable control signals capable of interacting with a programmable computer system in such a way that one of the procedures described herein is carried out.

[0025] The embodiments described above merely illustrate the principles of the present invention. It is understood that modifications and variations of the arrangements and details described herein will be obvious to other people skilled in the art. Therefore, it is intended that the invention be limited only by the scope of protection set forth in the following claims and not by the specific details presented herein by way of description and explanation of the embodiments.

[0026] Reference symbol list:

[0027] 20 Base model

[0028] 22 more training data points 22' new additional training data points

[0029] 24 modified model

[0030] 26 new training data set

[0031] 28 original training dataset

[0032] 30 Test result 32 Test result of the basic model

Claims

Patent claims 1. Computer-implemented method comprising the following steps: a) Using a basic mathematical model trained by machine learning on an original training dataset (28) to process sensor data and provide information for controlling a platform in a maritime environment, depending on the sensor data; b) Training the basic model (20) with further training data (22) to obtain a modified model (24); c) Testing the modified model (24) with a new training dataset (26) based on the original training dataset (28) and the further training data (22) to obtain a test result (30) of the modified model (24); d) Deciding, based on the test result (30) of the modified model (24), whether to use the basic model (20) or the modified model (24) for processing the next sensor data.

2. Computer-implemented method according to claim 1, wherein in step d) the modified model (24) is used if the test result (30) is equal to or better than a test result (32) of the base model.

3. Computer-implemented method according to claim 2, wherein the test result (32) of the base model is generated based on the original training data (28).

4. Computer-implemented method according to one of the preceding claims, wherein in step d) the further training data (22) are extended by new further training data (22') when the basic model (20) is continued and steps b) to d) are carried out after the extension until, based on the test result (30), a decision is made to use the modified model (24).

5. Computer-implemented method according to one of the preceding claims, wherein the new training data (22) become part of the original training data (28) when the modified model (24) is used in step d).

6. Computer-implemented method according to one of the preceding claims, wherein the basic model (20) is configured to process maritime sensor signals.

7. Computer-implemented method according to one of the preceding claims, wherein the basic model (20) is configured to process received radar signals or received sonar signals and to perform a detection, in particular also classification, of objects recognizable in the signals as information for controlling the platform.

8. Computer-implemented method according to claim 7, wherein the basic model (20) is configured to detect and / or classify sea mines as objects in the sensor data.

9. Computer-implemented method according to one of claims 7 or 8, wherein the basic model (20) is configured to detect and / or classify watercraft, in particular underwater vehicles, as objects in the sensor data.

10. Computer-implemented method according to one of claims 7 to 9, wherein the basic model (20) is configured to issue an indication of the detection of a detected hostile object and, in particular, to issue a recommendation for a suitable weapon to combat the hostile object.

11. Computer-implemented method according to one of the preceding claims, wherein the basic model (20) is configured to control an unmanned underwater vehicle based on the sensor data.

12. Data processing unit configured to execute the steps of the computer-implemented method according to any one of the preceding claims.

13. Computer program comprising instructions that are executed during the execution of the Programs by a computer cause it to execute the steps of the method according to any one of claims 1 to 11.

14. Use of a mathematical model trained using the method according to any one of claims 1 to 11 for controlling an autonomous underwater vehicle based on sensor data.

Images