Ultrasonic sensor unit
The ultrasonic sensor unit uses IQ mixing and 1D CNN for efficient object classification with reduced computational resources, addressing accuracy and cost challenges in the vehicle sector.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2025-12-01
- Publication Date
- 2026-07-16
Smart Images
Figure EP2025084893_16072026_PF_FP_ABST
Abstract
Description
[0001] R.414801
[0002] - 1 -
[0003] Description
[0004] title
[0005] Ultrasonic sensor unit
[0006] State of the art
[0007] The present invention relates to an ultrasonic sensor unit and a vehicle.
[0008] Currently, a wide variety of solutions exist for classifying objects using ultrasonic sensors. Due to the increasing number of ultrasonic sensors in use, as well as the rising performance and quality requirements for ultrasonic sensors, the need for innovative and robust ultrasonic systems is constantly growing.
[0009] The continuous reduction of weight in the vehicle sector to reduce fuel consumption, as well as increasing competition, is creating cost pressure, leading to a greater demand for cheaper and more efficient vehicle components.
[0010] Disclosure of the invention
[0011] The ultrasonic sensor unit according to the invention, with the features of claim 1, has the advantage over known units that the ultrasonic sensor unit can classify an object in the vicinity of the ultrasonic sensor unit, achieving very high correct classification rates. It can be particularly advantageous that the ultrasonic sensor unit, especially the classification, is particularly efficient and requires only a small number of model parameters or FLOPS. Furthermore, a plurality of processing hardware can preferably be used, since, due to the reduced computational effort, a plurality of logic units or R.414801
[0012] - 2 -
[0013] Computing units are being considered. A particularly advantageous effect has been observed when using processors specifically suited for time and frequency transformation and CNN architectures.
[0014] This is achieved according to the invention by the fact that the ultrasonic sensor unit has an ultrasonic element which is configured to emit and / or receive at least one air-based ultrasonic wave, wherein the ultrasonic sensor unit has and / or is connected to a logic unit, wherein the logic unit is configured to provide a digital ultrasonic signal based on the received ultrasonic wave, wherein the logic unit is configured to transfer the digital ultrasonic signal into a complex baseband by means of IQ mixing, wherein the logic unit is configured to identify at least one object in the vicinity of the ultrasonic element in the complex baseband, wherein the logic unit is configured to extract a section from the baseband when the object is present in the baseband, wherein the logic unit is configured toto apply at least one convolution operation using a specific 1D CNN module to the section in order to determine at least one feature of the section, wherein the logic unit is configured to provide a classification probability of the object based on the feature of the section.
[0015] Thus, the ultrasonic sensor unit is capable of performing signal preprocessing, which enables the transfer of the acquired sensor signal to the baseband via IQ mixing for lossless reduction of the data rate during preprocessing, as well as the removal of object backflows from the time signal. Furthermore, the ultrasonic sensor unit is capable of extracting low-threshold features and determining the size of the convolution kernel in the first CNN layer as a function of the transmit signal frequency. Preferably, the ultrasonic sensor unit is configured to perform an efficient convolution operation in a specific 1D CNN module with learnable downsampling, with N-fold repetition of successive 1D depthwise convolution, pointwise convolution, and strided convolution.Preferably, the ultrasonic sensor unit is configured to transfer the extracted features into a classification result using a classifier.
[0016] Class probabilities to be calculated using a fully connected layer. R.414801
[0017] - 3 -
[0018] The dependent claims describe preferred embodiments of the invention.
[0019] Preferably, the logic unit is configured to extract low-threshold features in the section using at least one 1D convolution kernel. An advantage of this embodiment is that not all sensor signals are subjected to the convolution operation, thus saving resources on the ultrasonic sensor unit and / or the logic unit.
[0020] Preferably, the logic unit is configured to adjust the width of the 1D convolution kernel as a function of the frequency of the emitted ultrasound wave and / or the detection rate of the ultrasound element. An advantage of this embodiment is that the detection accuracy of the convolution kernel can be adjusted depending on the received or emitted ultrasound wave.
[0021] Preferably, the logic unit is configured to apply normalization and convolution with a step size to the low-threshold features in order to perform dimensionality reduction or subsampling of the low-threshold features. An advantage of this embodiment is that, unlike conventional pooling operations, adaptation to the domain or the training dataset is permitted.
[0022] Preferably, the logic unit is configured to perform a channel-based 1D convolution in depth for the convolution operation on the section. An advantage of this embodiment is that, unlike conventional convolution layers, it is not necessary to learn a kernel for each input channel for each output channel; instead, there is exactly one kernel and, correspondingly, one output channel for each input channel.
[0023] Preferably, the logic unit is configured to apply layer normalization to a result of the 1D convolution. An advantage of this embodiment is that the feature vectors can be used to normalize the feature vectors of the objects through layer normalization. R.414801
[0024] - 4 -
[0025] Preferably, the logic unit is configured to apply a point convolution to a result of the layer normalization. An advantage of this embodiment is that the channel dimensions can be reduced by means of the point convolution, thus enabling the creation of feature vectors.
[0026] Preferably, the logic unit is configured to apply a ReLU or GELU activation function to a result of the point convolution. An advantage of this embodiment is that non-linearities can be represented using the ReLU or GELU activation function. A GELU activation function can be a Gaussian Error Linear Unit activation function. A ReLU activation function can be a Rectified Linear Unit activation function.
[0027] Preferably, the logic unit is configured to apply a second point-by-point convolution to a result of the ReLU or GELU activation function. An advantage of this embodiment is that complex features can be extracted with reduced computational effort.
[0028] Preferably, the logic unit is configured to transfer the result of the second point convolution to the 1D convolution or to provide the result. An advantage of this embodiment is that the accuracy of the class probability can be increased by repeatedly iterating the 1D convolutional block and preferably by downsampling.
[0029] Preferably, the logic unit is configured to apply at least one linear layer to a result of the second point convolution in order to provide the classification probabilities. An advantage of this embodiment is that class-specific probability values can be calculated using the softmax activation function to identify a type of object.
[0030] For example, the signal preprocessing using IQ mixing includes the architecture of the 1D Convolutional Neural Network (CNN), a front end, an encoder, and a classifier. In signal processing, the digitized converter signal of an object backscatter detected by the sensor is first processed using IQ mixing (R.414801).
[0031] - 5 -
[0032] The time signal is transferred into the complex baseband. This allows the data rate to be significantly reduced without losing information in the relevant bandwidth. Accordingly, the time signal can be transmitted efficiently from the sensor to a central processing unit in the vehicle for further calculations.
[0033] Subsequently, the sections containing object backscatter are preferably extracted from the complete time signal (windowing). This can be done using a sliding window approach or by considering conventionally available echo points. The extracted sections can then be processed directly in the complex baseband or, particularly advantageously, as a reconstructed, real-valued time signal for classification by a neural network. Further signal processing measures for feature extraction or the input of additional features alongside the time signal into the classification network are also possible.
[0034] In the frontend, the first processing step in the CNN can be the extraction of low-level features using particularly wide 1D convolutional kernels. Instead of, for example, a time-frequency transformation for feature extraction, a relevant filter bank is preferably learned over the training dataset. The width of the convolutional kernel should preferably be chosen depending on the frequency of the transmit signal and the sampling frequency, so that, analogous to the window width of a Fourier transform, all relevant frequency components can be extracted. This is followed by dimensionality reduction using learnable downsampling, in contrast to more common pooling operations, which do not allow adaptation to the domain or training dataset. The downsampling is preferably implemented using a 1D strided convolutional layer with a stride of at least s = 2, and the feature vectors are further normalized using layer normalization or batch normalization.
[0035] The encoder efficiently extracts high-level features through a combination of different convolutional operations and implements dimensionality reduction using machine learning downsampling. Depending on the complexity of the dataset, the convolutional block and dimensionality reduction can be repeated as often as needed, thereby increasing the depth of the CNN.
[0036] The convolutional block preferably comprises a depthwise 1D convolution with a wide kernel with k > 5, preferably k = 7. In contrast to conventional convolution layers, an individual convolution block is used for each input channel.
[0037] - 6 -
[0038] After learning the convolutional kernel, an equal number of input and output channels result for a number of trainable parameters that simply corresponds to K multiplied by the number of input and output channels, respectively. Layer normalization or batch normalization is then preferably applied to normalize the feature vectors. The block depth is preferably increased using parameter-efficient 1x1 pointwise convolutions with a depth d > 2, preferably d = 4. This is preferably followed by a ReLU or GELU activation function to model non-linearity. Finally, the channel dimension is preferably reduced again using pointwise convolution. A so-called residual or skip connection preferably adds the input to the output of the convolutional block, enabling good convergence during training even with very deep networks. Finally, a downsampling layer consisting of 1D strided convolutions is preferably applied again.
[0039] Preferably, after N iterations of the convolutional block and downsampling, the extracted feature vectors in the classifier are concatenated and placed in at least one linear layer to calculate class-specific probability values using the softmax activation function. The number of neurons preferably corresponds to the number of classes.
[0040] Another aspect of the invention relates to a vehicle which has an ultrasonic sensor unit as described above and below.
[0041] Brief description of the drawings
[0042] Exemplary embodiments of the invention are described in detail below with reference to the accompanying drawings. The drawing shows:
[0043] Figure 1 shows an ultrasonic sensor unit according to one embodiment,
[0044] Figures 2 to 5 are a flowchart illustrating the operation of the ultrasonic sensor unit according to one embodiment.
[0045] Figure 6 shows a vehicle according to one embodiment. R.414801
[0046] - 7 -
[0047] Embodiments of the invention
[0048] Preferably, all elements, units and / or steps in all figures are labelled with the same reference symbols.
[0049] Figure 1 shows an ultrasonic sensor unit 10 according to one embodiment. The ultrasonic sensor unit 10 has an ultrasonic element 12, which is configured to emit and / or receive at least one air-based ultrasonic wave, wherein the ultrasonic sensor unit 10 has and / or is connected to a logic unit 14, wherein the logic unit 14 is configured to provide a digital ultrasonic signal based on the received ultrasonic wave, wherein the logic unit 14 is configured to convert the digital ultrasonic signal into a complex baseband by means of IQ mixing, wherein the logic unit 14 is configured to identify at least one object 16 in the vicinity of the ultrasonic element 12 in the complex baseband, wherein the logic unit 14 is configured to extract a section from the baseband if the object 16 is present in the baseband, and wherein the logic unit 14 is configured toat least one convolution operation is applied to the section using a specific 1D CNN module to determine at least one feature of the section, wherein the logic unit 14 is configured to provide a classification probability of the object 16 based on the feature of the section.
[0050] Figure 2 shows a flowchart 200 illustrating the operation of the ultrasonic sensor unit 10 according to one embodiment. The transducer 202 receives an ultrasonic wave. The analog-to-digital converter 204 converts the received ultrasonic wave into a digital signal. The baseband conversion 206 converts the digital signal from the analog-to-digital converter 204 into a baseband signal. Preferably, windowing or temporal clipping 208 can be performed in the baseband. The clipped segments can then be loaded into a 1D CNN module 210. The 1D CNN module 210 can determine the probability of classifying an object 16 in the vicinity of the ultrasonic sensor unit 10 by means of a variety of operations, in particular convolution operations. The classification probabilities can be output in block 212.
[0051] - 8 -
[0052] Figure 3 shows a flowchart 300 illustrating the operation of the ultrasonic sensor unit 10 according to one embodiment. Block 302 represents a time-snipped signal in the baseband. This signal 302 is loaded into a frontend 304. Downsampling 306 takes place from the frontend 304. The downsampled signal is then loaded into a 1D CNN module and an encoder 308. The steps of the encoder 308 can be repeated multiple times, in particular by means of a loop 310. The encoder 308 can include a 1D convolutional block 312 as well as a downsampling step 314. The encoder 308 is followed by at least one linear layer 316, which can provide the classification probabilities 318.
[0053] Figure 4 shows a flowchart 400 illustrating the operation of the ultrasonic sensor unit 10 according to one embodiment. The flowchart 400 clarifies the operation of the 1D convolutional block 312 of Figure 3. Here, a step of deep 1D convolution 402 can be performed, particularly based on the values of the downsampling 306. More preferably, a layer normalization 404 can be applied to the results of the deep 1D convolution 402. More preferably, a point convolution 406 can be applied to the results of the layer normalization 404. Preferably, a GELU 408 can be applied to the results of the point convolution 406. More preferably, a second point convolution 410 can be applied to the results of the GELU 408.
[0054] Figure 5 shows a flowchart 500 illustrating the operation of the ultrasonic sensor unit 10 according to one embodiment. A normalization 502 can be applied to the results of the 1D convolution block 312. Preferably, a convolution with a step size 504 can be applied to the results of the normalization 502.
[0055] Figure 6 shows a vehicle 100 according to one embodiment. The vehicle 200 preferably has an ultrasonic sensor unit 10, as described above and below.
Claims
R.414801 - 9 - Claims 1. Ultrasonic sensor unit (10) comprising an ultrasonic element (12) which is configured to emit and / or receive at least one air-based ultrasonic wave, - wherein the ultrasonic sensor unit (10) has and / or is connected to a logic unit (14), - wherein the logic unit (14) is configured to provide a digital ultrasound signal based on the received ultrasound wave, - wherein the logic unit (14) is configured to transfer the digital ultrasound signal into a complex baseband by means of IQ mixing, - wherein the logic unit (14) is configured to identify at least one object (16) in the vicinity of the ultrasound element (12) in the complex baseband, - wherein the logic unit (14) is configured to extract a section from the baseband when the object (16) is present in the baseband, - wherein the logic unit (14) is configured to apply at least one convolution operation to the section using a specific 1D CNN module in order to determine at least one feature of the section, - wherein the logic unit (14) is configured to provide a classification probability of the object (16) based on the feature of the section.
2. Ultrasonic sensor unit (10) according to claim 1, wherein the logic unit (14) is configured to extract low-threshold features in the section using a 1D convolution kernel.
3. Ultrasonic sensor unit (10) according to claim 2, wherein the logic unit (14) is configured to adjust the width of the 1D convolution kernel as a function of the frequency of the emitted ultrasonic wave and / or the detection rate of the ultrasonic element (12). R.414801 - 10 - 4. Ultrasonic sensor unit (10) according to one of the preceding claims, wherein the logic unit (14) is configured to apply normalization and convolution with a step size to the low-threshold features in order to perform a dimensional reduction of the low-threshold features.
5. Ultrasonic sensor unit (10) according to one of the preceding claims, wherein the logic unit (14) is configured to perform a channel-based 1D folding into depth for the folding operation on the section.
6. Ultrasonic sensor unit (10) according to claim 5, wherein the logic unit (14) is configured to apply a layer normalization to a result of 1D convolution.
7. Ultrasonic sensor unit (10) according to claim 6, wherein the logic unit (14) is configured to apply a point convolution to a result of layer normalization.
8. Ultrasonic sensor unit (10) according to claim 7, wherein the logic unit (14) is configured to apply a ReLU or a GE LU activation function to a result of point convolution.
9. Ultrasonic sensor unit (10) according to claim 8, wherein the logic unit (14) is configured to apply a second point convolution to a result of the ReLU or GELU activation function.
10. Ultrasonic sensor unit (10) according to claims 9 and 5, wherein the logic unit (14) is configured to transfer the result to the 1D folding or to provide the result based on a result of the second point folding.
11. Ultrasonic sensor unit (10) according to claim 10, wherein the logic unit (14) is configured to apply a linear layer to the result of the second point convolution to provide the classification probabilities. R.414801 - 11 - 12. Vehicle (100) comprising an ultrasonic sensor unit (10) according to one of the preceding claims.