Sonar system for identifying or classifying known vehicles which introduce sound into water

Abstract

The invention relates to a sonar system (20) for classifying or identifying known vehicles (22) which introduce sound into water, the sonar system having the following features: - a plurality of waterborne sound transducers (24) which are each designed to convert waterborne sound (26) into a signal (28) corresponding to the sound pressure; and - a signal processing unit (30) which is designed to carry out the following steps: a) (34) processing a representation of the signals (28) in order to obtain features of the signals (28); b) (36) processing the features (35) using a model (36) trained by means of machine learning in order to identify the known vehicles (22).

Description

[0001] 2500080P10WQ 2025.103WO

[0002] Sonar system for identifying or classifying known vessels that couple sound into water

[0003] Description

[0004] The invention relates to the identification or classification of vehicles that couple sound into water, in particular watercraft such as ships or submarines, each manned and / or unmanned, or aircraft such as drones, airplanes or helicopters, provided that the (air) sound waves generated by the aircraft are coupled into the water.

[0005] The detection of vessels that couple sound into water is typically performed manually by a sonar operator. Objects are represented as (bright) points in a time-versus-bearing plot (a so-called waterfall plot). The intensity (i.e., primarily brightness) of each point in the plot corresponds to the sound intensity of the associated time / bearing pair. That is, the higher the sound intensity, the higher, and especially brighter, the intensity of the corresponding point in the plot. The sonar operator's task is to identify tracks (i.e., lines) in the plot. These tracks represent a detected object that is making noise. An experienced sonar operator can classify and, to some extent, identify these tracks based on the associated audio data using their hearing.

[0006] However, it is possible that an incorrect choice of parameters influencing the display could render very faint (i.e., quiet) objects invisible. Therefore, the multi-hypothesis test was developed for automated display evaluation. This test can detect traces in the display. However, this only works if the sonar data is evaluated across the entire frequency range. That is, the envelope of the sonar signal is analyzed. However, there are objects that emit underwater sound in specific frequency ranges. These are, for example, ships whose propellers rotate at a constant speed. Such objects can be better detected by frequency-selective evaluation and subsequent combination of the evaluations. In this case, however, the multi-hypothesis test fails because the number of resulting hypotheses becomes unmanageable.

[0007] However, the multi-hypothesis test also reaches its limits when a vessel's track breaks off or crosses. If the track is detected again after breaking off, the multi-hypothesis test cannot assign the two partial tracks to the same vessel. Similarly, the multi-hypothesis test has difficulty assigning intersecting tracks to the correct track before the intersection after the crossing has occurred.

[0008] The object of the present invention is therefore to create an improved concept for the identification and / or classification of watercraft.

[0009] 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.

[0010] Exemplary embodiments show a sonar system for identifying or classifying known vessels that couple sound into water, in particular watercraft or aircraft. The identification comprises detection and classification, with the classification being possible at a level of detail that allows not only for distinguishing classes of watercraft or aircraft, but also for determining the exact vessel. "Known" means that training data is preferably already available for the same vessel, which has advantageously been used for training the models by means of supervised learning. In this case, the model can be trained to output the exact designation of the vessels and not just to classify them. That is, within a vessel class, the individual vessels can be identified. Thus, the model can recognize the sounds of each vessel and assign them to the corresponding designation.

[0011] The sonar system comprises a multitude of underwater sound transducers, each designed to convert underwater sound into a signal corresponding to the sound pressure level. This signal is typically an electrical signal, specifically a time series of an audio signal, particularly a digitized one.

[0012] Furthermore, the sonar system includes a signal processing unit to process the signal. In particular, the signal processing unit can also perform the digitization of the signal. The signal processing unit can be a computer, for example, a CPU (central processing unit) or GPU (graphics processing unit). In any case, the signal processing unit performs the following steps:

[0013] Optionally in step aO) beamforming the signals to obtain a directional signal for each direction considered. Beamforming is also known as direction formation. The most common form of beamforming is delay-and-sum beamforming. In any case, the directional signal comprises the individual signals from the underwater transducers, specifically a summed signal of these signals. This means that at a given point in time, multiple directional signals are generated by summing the individual output signals of the underwater transducers using delay-and-sum beamforming, with each output signal being considered at a different time. Beamforming enables not only identification and classification but also tracking of the vessel or aircraft. This means that the vessel's or aircraft's path can be tracked.

[0014] a) Processing a representation of the signals, in particular the directional signals obtained in step aO), or signals prepared using standard sonar signal processing, to obtain features. The features are determined according to a computational rule, i.e., a signal analysis rule. Thus, the features of the multitude of features can include a multitude of features derived from a DEMON analysis [DEMON: Demodulation Envelope Modulation On Noise] of the directional signal. As an example of possible features from the DEMON analysis, the amplitude and / or duration of envelopes of the directional signal can be considered, to name just two of a multitude of possible features. Additionally or alternatively, the features of the multitude of features can include a multitude of features derived from a LOFAR analysis [LOFAR: Low Frequency Analysis and Recording] of the directional signal.Examples of possible features of LOFAR analysis include the distance and / or frequency of frequency lines of the directional signal, to name just two of a multitude of possible features.

[0015] As an alternative to, or in addition to, one or both of the aforementioned examples of features, the signal processing unit in step a) can use a deep artificial neural network to obtain a plurality of features from the multitude of features. Such features developed using a deep artificial neural network cannot usually be physically described by humans. However, it has been shown that features developed by artificial neural networks play a central role in the identification or classification of vehicles. Both alone and, in particular, in combination with the DEMON and / or LOFAR features, this enables robust identification or classification of the vehicles.

[0016] Step b) involves processing the features with a machine learning-trained model to identify the known vehicles. This means a machine learning model is trained to classify vehicles and, ideally, even distinguish between identical vehicles. Tests have shown that this type of vehicle identification is superior to that performed by sonar operators.

[0017] Steps a) and b) can be performed for different direction signals per observation time, in particular for each direction signal, provided step aO) has been executed, in order to obtain the direction in which the vehicles to be identified or classified are located. That is, feature extraction and analysis by the model are performed separately for the directions under consideration. In this case, not only is vehicle identification and classification possible, but also tracking, i.e., following the vehicles' path. 2500080P10WQ 2025.103WO

[0018] Vehicle identification or classification can be achieved using a model, i.e., an artificial intelligence, within the sonar data (i.e., the representation of the signals, particularly the directional signals). The model can optionally comprise multiple, particularly distinct, sub-models. A (sub-)model describes the architecture of one or more algorithms. The model can be trained by adjusting weights, i.e.,

[0019] The parameters can be optimized. Most common, well-known algorithms are suitable as models, in particular artificial neural networks, Bayesian classifiers, support vector machines (SVMs), and linear regression. Deep learning approaches, such as transformer-based artificial neural networks and / or convolutional artificial neural networks, can also be used. If the model comprises several sub-models for vehicle identification or classification, similar algorithms (e.g., different artificial neural networks) as well as different algorithms (e.g., an artificial neural network and linear regression) can be combined.

[0020] In particular, it has been shown that no special algorithms or algorithm architectures are needed to create the model for vehicle identification or classification. It is sufficient to train commercially available algorithms with appropriate data. Such algorithms are also known as COTS (Components Off-The-Shelf). It has been found that the structure—for example, the number of neurons or layers in artificial neural networks—is irrelevant for training the model, as long as a minimum level of complexity is achieved. Models that perform well in public benchmarks are particularly suitable.Furthermore, it has been shown that even simple feedforward neural networks, especially those with few layers, are already suitable for the classification and / or identification of vehicles using the features from the deep model.

[0021] Preferably, at least one (partial) model is implemented as an artificial neural network. The artificial neural network can, for example, be based on Transformer technology and / or be or include a convolutional artificial neural network. However, a classic feedforward neural network, such as a multilayer perceptron (MLP), is preferred. It has been found that a feedforward neural network is sufficient for classifying and identifying vehicles with the features from the deep model. More complex and thus computationally intensive models can therefore be avoided.

[0022] If a single model without sub-models is used, it will preferentially identify all vehicles. If multiple sub-models are used, each sub-model can, for example, identify one vehicle.

[0023] Real, recorded sonar data (especially signals or directional signals) can be used to train the model or sub-models. However, it is also possible to generate artificial sonar data.

[0024] Artificial (i.e., synthetic) training data is suitable. Artificial data can be used particularly when the amount of real training data is insufficient. Preferably, a combination of real and artificial data can be used for training. That is, artificial data can be supplemented with real sonar data. Real sonar data can be acquired during training runs, training maneuvers, or in the field. Furthermore, there are publicly accessible databases with large amounts of sonar data. Preferably, the real data is pre-labeled, for example, by a trained sonar operator. That is, the actual vehicles in the sonar data are identified beforehand. This is called labeling. This enables supervised learning of the model. Manual labeling is not necessary with artificial sonar data, since it is known where (i.e.,The system can identify when an object was simulated (at what bearing and time) and where only noise is present. Therefore, the labeling can be automated.

[0025] The idea is therefore to train a model, i.e., an artificial intelligence, using machine learning, which, using suitable features, is capable of detecting and classifying vehicles and preferably also identifying them. Such models are more reliable and accurate than the previous classification and identification by sonar operators, i.e., humans.

[0026] In exemplary embodiments, the signal processing unit is configured to perform steps a), b), and optionally aO) continuously with constantly updated signals or using different time intervals of the signals to determine a movement track of the vehicle over time, particularly in real time. This process is also referred to as tracking. That is, the signals from the underwater sound transducers are continuously sampled. This results in continuously updated signals from the underwater sound transducers and thus also updated direction signals. The entire set of updated signals, but preferably different time intervals of the signals, calculated, for example, using sliding windows, can now be used for vehicle identification. A suitable length of the time intervals, i.e., a suitable window length, is, for example, between 1 s and 10 s. The tracks are preferably available as position data and can, for example, be...displayed in a situational image and / or used for navigation, especially autonomous navigation.

[0027] It has been found that, with conventional technical support for sonar operators, the re-identification of vehicles is problematic when determining motion traces. Re-identification refers to the re-identification of a vehicle whose trace was interrupted or crossed with the trace of another vehicle. The multi-hypothesis test, for example, is not suitable for such re-identification. However, it has been shown that the model is reliably capable of such re-identification. That is, the signal processing unit is designed to obtain a re-identification of a vehicle at a later time after the motion trace has been broken.

[0028] In further embodiments, the model has at least one input node per feature. This means that the features are calculated from the direction signal, in particular a time segment thereof, using the features themselves, and fed into the model at separate input nodes.

[0029] Advantageously, the model is trained using features of the vehicles to be identified or classified (zu2500080P10WQ 2025.103WO). The training is supervised, meaning the training data is labeled. In other words, the correct vehicles are marked in the training data.

[0030] For correct classification and identification of the vehicles, it is further advantageous if the model has at least one output node for each vehicle to be identified or classified. This allows the vehicles to be distinguished from the model.

[0031] At least one additional output node can be trained to detect any vehicle. This is advantageous when sound waves from an unknown or unexpected vehicle are recorded by the underwater transducers. While identification or classification is not possible, at least detection of such a vehicle is.

[0032] In exemplary embodiments, the model incorporates or consists of an artificial neural network. Advantageously, the artificial neural network is a pure feedforward artificial neural network, in particular a multi-layer perceptron (MLP). This minimizes the complexity and required computing power (of the signal processing unit) for vehicle identification or classification. This is particularly advantageous when using the sonar system in underwater vehicles, where the available energy is limited. However, even when using the system in watercraft or other locations, it is always important to use the available energy sparingly.

[0033] The features described above, determined using an artificial neural network, can be identified using a method for classifying or identifying water-emitting vessels by analyzing sonar data with a deep learning model. A deep model is based on a multitude of hidden layers. Artificial neural networks are commonly used as deep models. Examples include transformer-based artificial neural networks, convolutional artificial neural networks, and recurrent neural networks (RNNs). Classification refers to the differentiation of vessels into different classes. For example, a Type 212A submarine can be recognized as such and distinguished from submarines of other classes.Furthermore, the identification allows for the differentiation of essentially identical vehicles within the individual classes. The deep model provides a computational formula for processing the sonar data. The result is an output signal corresponding to the sonar data input to the deep model; this signal is called the feature, or usually a multitude of features, the feature vector.

[0034] The procedure, in step a), involves providing a large number of sonar data packets, each assigned to a first category and a second category. Sonar data sets are then created from these packets, with each sonar data set containing a pair of sonar data packets from the first and second categories. That is, each pair of sonar data packets consists of one sonar data packet from the first category and one sonar data packet from the second category. A sonar data packet is defined as a segment of the sonar signals. If possible, directional sonar signals (beams) can be used. The segmentation can be achieved, for example, by applying a window to the sonar signal. Advantageously, each sonar data packet is used multiple times in different categories. Thus, a single sonar data packet can be assigned to the first and second categories multiple times.However, care should be taken to ensure that the sonar data packet pairs of the first and second categories are used only once per epoch (su). The sonar data can be real, recorded sonar data. For example, sonar data from publicly accessible databases can be used. However, it is also possible to use artificially generated sonar data. Advantageously, the sonar data, i.e., also the sonar data packets, are each time series, i.e., digitized audio files. The sonar data is passive sonar data. This means that the sounds generated by the vehicles are analyzed to determine their characteristics and to classify or identify the vehicles. The first and second sonar data packets of the sonar dataset each contain different payload data for the same vehicle to be classified or identified.The sonar data packets of the first category can be considered "positive examples." The sonar data packets of the second category can be considered "anchors." For example, the sonar data packets of the first category can be obtained by processing the sonar data packets of the second category or the associated sonar data.

[0035] In particular, the sonar data packets within the same sonar dataset are preserved through this processing. The processing can involve shifting the section of the sonar signal and / or adding noise or other interfering sounds. This processing process is also known as augmentation. The payload refers to the sounds generated by the vehicle, which allow the vehicle to be identified. These sounds can also be described as a vehicle's acoustic signature. However, the sonar data always contains other sounds as well, such as background noise.

[0036] In step b), the deep model is trained with a large number of sonar datasets to reduce, and in particular minimize, the difference between the initial values ​​of the deep model using the sonar data packages of the first category and the second category, and to increase, and in particular maximize, the difference between the initial values ​​of the deep model using the sonar data packages of the second category and one or more sonar data packages different from the first category. Such training is also known as "contrastive learning." Furthermore, it is not necessary to label the sonar data packages when assigning them to categories. Such training is also known as "self-supervised learning." An iteration is defined as training with a predetermined number of sonar datasets, or optionally a single dataset, before the deep model is optimized.The data sets that comprise the predetermined number of sonar data sets are also referred to as a "batch." The sonar data packets of one or more sonar data sets different from the sonar data set can also be referred to as third-category sonar data packets or additional sonar data packets. To reduce the amount of data required, these can be sonar data packets that are already assigned to the first and / or second category in other sonar data sets. However, it is also possible to use sonar data packets that are not used in other sonar data sets.

[0037] Advantageously, the deep model is presented with the sonar data packets of the predetermined number of sonar data sets consecutively, i.e., one after the other, and the output values ​​are stored after each presentation. It is possible to map the output values ​​to the corresponding input values, i.e., the sonar data packets. Optionally, the sonar data packets of the first and second categories can be presented alternately. It is also possible, for example, to first present the sonar data packets of the first category of data sets from the predetermined number of data sets (i.e., the data sets of a batch) and then the sonar data packets of the second category of these data sets. In principle, it is also possible, and encompassed by the term "training the deep model," to train two or even three identical deep models (with sonar data packets of the first, second, and third categories) in parallel.This saves time when training the deep model, but requires significantly more computing power.

[0038] The sonar data packets in the first and second categories each contain different data compared to the other sonar data packets. These other sonar data packets can be considered "negative examples." Each additional sonar data packet contains data from a different vehicle, or even just noise or natural sounds, compared to the sonar data packets of the sonar dataset from which the difference in the initial values ​​is to be amplified.

[0039] To reduce the difference between the output values ​​of the deep model using the first and second category sonar data packets, and to increase the difference between the output values ​​of the deep model using the second category sonar data packets and one or more sonar data packets from different sonar datasets, it is possible to determine a mathematical distance between the output values ​​that is maximized or minimized, respectively. For example, algorithms based on cosine similarity or the "InfoNCE" loss function can be used as the mathematical distance. The "AdamW" algorithm, for example, can be used as the optimization algorithm.

[0040] Thus, a deep model can be trained to generate features for the automatic classification and / or identification of vehicles. The deep model generates features to which humans cannot assign physical properties. Therefore, humans would not use these features. Tests have shown that, in conjunction with any machine learning-trained model, these features can achieve very good results for classification and identification.

[0041] The deep model can be a convolutional neural network (CNN), specifically a multitude of one-dimensional convolutional neural network modules, particularly residual networks (1d-ResNets). 1d-ResNets are especially well-suited for processing time series and have achieved very good results in this case. Furthermore, the deep model can contain one or more downsampling blocks.

[0042] Exemplary embodiments show that the signal processing unit in step b) is configured to process the features with a multitude of different, machine learning-trained models in order to identify the known vehicles independently with each of the models. Thus, one identification result is obtained per model for each known vehicle. That is, a multitude of different models are used to identify one of the vehicles. Each of the models can independently identify the known objects. In other words, the models determine their classification result independently of each other. 2500080P10WQ 2025.103WO Additionally or alternatively, the signal processing unit in step a) is configured to use a multitude of deep artificial neural networks to obtain a multitude of different sets of features.The signal processing unit can then, in step b), determine a resulting identification for each known vehicle (20) for each set of features and determine the uncertainty of each resulting identification. In step b), either a single mathematical model or the previously described plurality of mathematical models can be used. Advantageously, however, no permutation is formed; that is, each model from the plurality of mathematical models does not determine the identification for each set of features, but rather each set of features is assigned a model from the plurality of models.

[0043] The models in steps a) and / or b) can differ, for example, in terms of the type of model (e.g., CNN (convolutional neural network) or other deep learning approaches), the type of training (supervised learning, unsupervised learning, one or more hybrid forms of unsupervised and supervised learning, reinforcement learning, etc.), or the training data used (e.g., different preprocessing, different input datasets, etc.). The models (the multitude of mathematical models), i.e., the entirety of the models, can also be referred to as an ensemble (of models).

[0044] In step b), the signal processing unit now determines a resulting identification for each known vehicle based on the identification results of that specific vehicle, and calculates the uncertainty of these resulting identifications for each known vehicle. This means that the identification results of the models are evaluated and ultimately assessed for each vehicle. The uncertainty can be either model uncertainty (epistemic uncertainty) or data uncertainty (aleatory uncertainty). Model uncertainty describes a measure of the consistency of the individual identification results of the models. Data uncertainty describes a statistical measure of the uncertainty of the individual models in determining the identification result. The uncertainty can be visualized, for example, in a situational diagram, particularly in conjunction with the identified vehicle.The uncertainty allows an operator or processing unit, for example, to determine whether the identified vessel needs further examination, such as re-identification. If the identification is intended for use in an attack, such as with a torpedo, the attack might not be carried out initially if the uncertainty is too high. This is particularly true if the identification process takes place on the torpedo itself or another autonomously operating, unmanned watercraft, and the torpedo or unmanned watercraft decides independently whether to launch an attack. Furthermore, it can be determined, for example by an operator or automatically, whether the underwater sound transducers are functioning correctly or whether retraining the mathematical models for a specific vessel, especially with additional training data from that vessel, would be beneficial for more reliable identification.

[0045] Mathematically, model uncertainty and data uncertainty can be determined as follows. The parameters used are m = 1 ... M for the models used, c = 1 ... C for the different classes into which objects can be classified, H for the entropy, and x for the observation, i.e., the classification result of a model.

[0046] By way of example, the softmax vector result of the individual classification results per model m can first be determined as: p m (c|x). Classification certainty can also be referred to as normalized classification result. Other activation functions can also be used. The softmax function can be functionally assigned to the mathematical models or the classification combination unit.

[0047] The overall opinion of all models (= resulting classification result) can be determined, especially in the classification combination unit, for example as follows:

[0048]

[0049] The class with the highest classification certainty forms the resulting classification of the object, also referred to as an observation. 2500080P10WQ 2025.103WO

[0050] Based on this data, the model uncertainty (EU) can be determined as follows:

[0051]

[0052] Furthermore, data uncertainty (AU) can be determined by example as follows:

[0053]

[0054] A total uncertainty (TU) can now be determined to be:

[0055]

[0056] It should be noted again that this mathematical description is only an example. Other stochastic methods for evaluating the classification results, such as a calculation using variance, can also be used.

[0057] Similarly, a method for identifying known vehicles that couple sound into water is disclosed, comprising the following steps:

[0058] 1) Received from a variety of signals from underwater sound transducers;

[0059] 2) Processing the directional signals obtained in step 1) based on a multitude of features to obtain a multitude of features;

[0060] 3) Processing the features with a machine learning-trained model to identify the known vehicles.

[0061] Furthermore, a computer program comprising instructions that, when executed by a computer, cause it to perform the method according to one of the preceding claims using received underwater sound. The computer program can, for example, receive the underwater sound signals from the underwater sound transducers and perform the entire sonar signal processing, or, for example, perform only the signal processing relevant to the invention. In particular, the model is implemented as a sequence of instructions in the computer program.

[0062] Preferred embodiments of the present invention are explained below with reference to the accompanying drawings. These show:

[0063] Fig. 1: a schematic representation of a sonar system with a model for vehicle identification;

[0064] Fig. 2: a schematic representation of the model in an exemplary embodiment.

[0065] 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.

[0066] Fig. 1 shows a schematic diagram of a sonar system 20 for identifying and / or classifying known vessels 22 that couple sound into water. The sonar system comprises a plurality of underwater sound transducers 24, each configured to convert underwater sound 26 into a signal 28 corresponding to the sound pressure. Furthermore, the sonar system includes a signal processing unit 30 for processing the signals 28.

[0067] Fig. 2 shows a schematic diagram of the signal processing unit 30 in an exemplary embodiment. The signal processing unit includes an optional beamforming module 32 for directional signal formation to obtain a direction signal 33 for each direction considered. Furthermore, the signal processing unit includes a feature generation module 34 for processing a representation of the signals 28, in particular the direction signals, with the deep model to obtain a plurality of features 35. The features 35 are fed to a machine learning-trained model 36, in particular its input nodes 38. Advantageously, the number of input nodes 38 corresponds to the number of features 35.

[0068] Furthermore, model 36 features a large number of output nodes (here, 5 output nodes). This allows, for example, the identification of 5 different vehicles, or the identification of 4 different vehicles and the detection of other vehicles. The number of levels and the number of nodes per level can be chosen arbitrarily.

[0069] The model is illustrated as a multi-layer perceptron as an example.

[0070] The disclosed (water) sound transducers are designed for underwater use, 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. The transducers can therefore be used as underwater sound receivers. The transducer material can be a piezoelectric material, such as a piezoceramic. A plurality of underwater sound transducers, or one or more underwater sound transducers in conjunction with a signal processing unit, can be referred to as a sonar system. The transducers can be used for (passive) sonar (sound navigation and ranging). The transducers are preferably not suitable for, or are not used for, medical applications.Likewise, sound transducers are preferably not used for ultrasonic testing of materials or are not suitable for this purpose.

[0071] Although some aspects have been described in connection with a device, it is understood that these aspects also constitute a description of the corresponding process, so that a block or component of a device is also to be understood as a corresponding process step or as a feature of a process step. Similarly, aspects described in connection with or as a process step also constitute a description of a corresponding block, detail, or feature of a corresponding device. 2500080P10WQ 2025.103WO

[0072] Depending on the implementation requirements, embodiments of the invention can be implemented in hardware or 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, are stored. These signals can interact with a programmable computer system (CPU and / or GPU) in such a way that the respective method is carried out. For this reason, the digital storage medium should be computer-readable. Embodiments can 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 methods described herein is carried out.

[0073] In some embodiments, a programmable logic device (for example, a field-programmable gate array, an FPGA) can be used to perform some or all of the functionalities of the methods described herein. In some embodiments, a field-programmable gate array can interact with a microprocessor to perform one of the methods described herein. Generally, in some embodiments, the methods are performed by any hardware device. This can be general-purpose hardware such as a computer processor (CPU) or a graphics processing unit (GPU), or hardware specific to the method, such as an ASIC. Distributed execution across the CPU and GPU is also possible.

[0074] 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 of the following claims and not by the specific details presented herein by way of description and explanation of the embodiments. Reference numerals:

[0075] 20 sonar systems

[0076] 22 vehicles

[0077] 24 underwater transducers

[0078] 26 Water sound

[0079] 28 Signal

[0080] 30 Signal processing unit 32 Beamforming

[0081] 33 Directional signal

[0082] 34 Character formation

[0083] 35 Features / Feature vector 36 Model

[0084] 38 input nodes

[0085] 40 output nodes

Claims

2500080P10WQ 2025.103WO Patent claims 1. Sonar system (20) for classifying or identifying known vessels (22) that couple sound into water, comprising the following features: - a plurality of underwater sound transducers (24), each designed to convert underwater sound (26) into a signal (28) corresponding to the sound pressure; - a signal processing unit (30) configured to perform the following steps: a) (34) Processing a representation of the signals (28) to obtain features of the signals (28); b) (36) Processing the features (35) with a machine learning trained model (36) to identify the known vehicles (22).

2. Sonar system (20) according to claim 1, wherein the signal processing unit (30) is configured to perform steps a), b) and c) continuously with continuously updating signals (28) or using different time intervals of the signals to determine a movement track of the vehicle over time.

3. Sonar system (20) according to claim 2, wherein the signal processing unit (30) is configured to obtain a re-identification of the vehicle (22) at a later time after the motion track has been broken.

4. Sonar system (20) according to one of the preceding claims, wherein the model (36) has an input node (38) for each feature (35).

5. Sonar system (20) according to one of the preceding claims, wherein the model (36) is trained using features (35) of the vehicles (22) to be identified.

6. Sonar system (20) according to one of the preceding claims, wherein the model (36) has at least one output node (40) per vehicle (22) to be identified.

7. Sonar system (20) according to one of the preceding claims, wherein the model (36) comprises an artificial neural network.

8. Sonar system (20) according to one of the preceding claims, wherein the model (36) is a feed forward artificial neural network, in particular a multi-layer perceptron (MLP).

9. Sonar system (20) according to one of the preceding claims, wherein the features of the plurality of features comprise a plurality of features derived from a DEMON analysis [DEMON: Demodulation Envelope Modulation On Noise] of the directional signal.

10. Sonar system (20) according to one of the preceding claims, wherein the features of the plurality of features comprise a plurality of features derived from a LOFAR analysis [LOFAR: Low Frequency Analysis and Recording] of the directional signal.

11. Sonar system (20) according to one of the preceding claims, wherein the signal processing unit is configured to perform step aO) (32) direction formation of the signals (28) in order to obtain a direction signal (33) for each direction considered, wherein step a) and step b) are performed for different direction signals (33) per viewing time, in particular each direction signal (33), in order to obtain the direction in which the vehicle (22) to be classified is located.

12. Sonar system (20) according to one of the preceding claims, wherein the signal processing unit is configured to use a deep artificial neural network in step a) to obtain a plurality of features of the plurality of features.

13. Sonar system (20) according to claim 12,2500080P10WQ 2025.103WO - wherein the signal processing unit in step a) is configured to use a plurality of deep artificial neural networks to obtain a plurality of different sets of features; - wherein the signal processing unit in step b) is configured to determine a resulting identification for each set of features for each known vehicle (20) and to determine an uncertainty of the resulting identification for each.

14. Sonar system (20) according to one of the preceding claims, - wherein the signal processing unit in step b) (36) is configured to process the features (35) with a variety of different models (36) trained by machine learning in order to identify the known vehicles (22) independently of each other with the models in order to obtain one identification result per model per known vehicle (20); - wherein the signal processing unit is configured to determine a resulting identification per known vehicle (20) depending on the identification results of the respective known vehicle and to determine an uncertainty of the resulting identifications per known vehicle.

15. Sonar system (20) according to one of claims 13 or 14, wherein the uncertainty comprises a model uncertainty and a data uncertainty.

16. Sonar system (20) according to claim 15, wherein the model uncertainty describes a measure of the uniformity of the individual identification results of the models.

17. Sonar system (20) according to one of claims 15 or 16, wherein the data uncertainty describes a statistical measure of the uncertainty of the individual models in determining the identification result.

18. Method for identifying known vehicles (22) that couple sound into water, comprising the following steps: 2500080P10WQ 2025.103WO 1) Obtaining from a plurality of signals (28) from underwater sound transducers (24); 2) Process a representation of the signals obtained in step 1) (28) to obtain features; 3) Processing the features (35) with a machine learning trained model (36) to identify or classify the known vehicles.

19. Method according to claim 18, wherein the signal processing unit is configured to perform step 1b) direction formation of the signals (28) to obtain a direction signal (33) for each direction considered, wherein step 2) and step 3)) are performed for different direction signals (33) per viewing time, in particular each direction signal (33), to obtain the direction in which the vehicle (22) to be classified is located.

20. Computer program comprising instructions which, when the program is executed by a computer, cause it to execute the method according to claim 18 or 19.

Images