Target detection and identification method based on bat sonar echo Doppler compensation bionics

Abstract

The invention belongs to the technical field of sonar, and discloses a bat sonar echo Doppler compensation bionics-based target detection and identification method, which comprises the steps of transmitting a sonar signal to a target; establishing a target motion model; receiving an echo signal; performing narrowband filtering on the echo signal to obtain a time-frequency diagram; performing time-frequency analysis on the time-frequency graph to estimate Doppler frequency offset; based on the frequency offset, dynamically adjusting the subsequently emitted sonar frequency; meanwhile, generating a data set based on the compensated echo amplitude; constructing a lightweight expansion convolutional neural network model; and inputting the data set into the network model to complete target identification. Through the bionic bat Doppler compensation mechanism, the recognition spectrum range is effectively narrowed, the detection sensitivity of the moving target is improved, and the performance requirement for signal processing hardware is lowered.

Description

Technical Field

[0001] This invention belongs to the field of sonar technology, specifically relating to a target detection and recognition method based on bat sonar echo Doppler compensation biomimetic. Technical Background

[0002] In the current field of robotics and automation, active sensing systems, especially sonar-based systems, are central to achieving environmental detection, target tracking, and identification. These systems typically emit sound waves of specific frequencies into the environment and then obtain information such as the target's distance, speed, and orientation by receiving and analyzing the echo signals reflected back from the target. Such systems perform well when dealing with static or low-speed moving targets. However, when the detected target moves relative to the robot at a higher speed, the Doppler effect occurs, causing a frequency shift in the echo signal. When the target speed is high, the resulting Doppler shift can be very significant, potentially causing the echo frequency to exceed the receiver's preset optimal operating frequency range. This can not only lead to signal loss and target tracking failure but also make subsequent signal processing and feature extraction extremely difficult. Therefore, to address the challenges posed by moving targets, a Doppler shift compensation mechanism simulating that of bats is proposed. These bats possess an auditory system with extremely high sensitivity within a narrow frequency range, enabling them to accurately identify prey targets. This region is known as the "fovea centralis." To ensure that the echo signal from the prey falls precisely into this highly sensitive fovea centralis, the bats actively and dynamically adjust the frequency of their emitted sound waves. This biological mechanism offers a new technological approach: instead of passively receiving broadband echoes, bats actively adjust their emission frequency to stabilize the echo frequency. This effectively narrows the recognition spectrum, improving the sensitivity for detecting moving targets. Furthermore, with the rapid development of artificial intelligence, technologies such as deep learning have demonstrated enormous potential in processing complex signals and pattern recognition, providing a powerful computational framework and entirely new solutions for realizing this biologically inspired complex dynamic frequency control and intelligent signal analysis.

[0003] The current common approach in robot perception systems, which uses fixed-frequency transmission and wideband reception, while simple in design, suffers from inherent limitations that severely restrict overall system performance, accuracy, and cost when dealing with high-speed dynamic targets. First, the most direct problem is noise interference and signal-to-noise ratio degradation. To capture echoes generated by the high-speed movement of a target with significant frequency shifts, the receiver must have a very wide receiving bandwidth. However, this increased bandwidth means the system will receive more environmental noise, thermal noise, and other electromagnetic interference across a wider frequency band. Second, in fixed-frequency transmission mode, wideband echoes require a high-bandwidth signal acquisition module at the receiver, and the massive amounts of data generated by the wideband echo signal necessitate real-time processing by a high-speed signal processor, ultimately leading to higher hardware costs and power consumption.

[0004] Based on the above analysis, there is an urgent need for an innovative target detection and recognition method. This method should fundamentally overcome the drawbacks of broadband reception and provide a stable, high-signal-to-noise ratio, and key-feature-distinct ideal input signal for subsequent signal processing algorithms. Therefore, by drawing on and simulating the Doppler shift compensation biological mechanism used by bats in echolocation, a novel biomimetic intelligent sensing method is developed. This lays a solid technical foundation for realizing a high-precision, robust, and low-cost high-speed target detection and recognition system for robots. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to propose a target detection and recognition method based on bat sonar echo Doppler compensation biomimetic.

[0006] To achieve the above objectives, the present invention employs the following technical methods.

[0007] A target detection and recognition method based on bat sonar echo Doppler compensation biomimetic includes the following steps:

[0008] Step 1: Transmit a sonar signal to the target at an initial frequency of 30 kHz;

[0009] Step 2: Establish a model of forearm swing motion with different heights and body shapes when walking at different speeds and with different arm swing amplitudes;

[0010] Step 3: Receive sonar echo signals that generate stable echo frequencies under different motion models;

[0011] Step 4: Input the echo signal into a narrowband filter for filtering to obtain the time-frequency diagram of the signal;

[0012] Step 5: Perform time-frequency analysis on the time-frequency diagram to estimate the Doppler frequency offset, and dynamically adjust the frequency of subsequent sonar transmissions based on the frequency offset.

[0013] Step 6: Extract the compensated echo amplitude features to generate a dataset for training and recognition, and randomly divide the dataset into training set, validation set and test set according to the proportion.

[0014] Step 7: Construct a lightweight dilated convolutional neural network model;

[0015] Step 8: Input the training set into the lightweight dilated convolutional neural network model to obtain the target recognition model, and input the test set into the recognition model to obtain the target speed recognition result;

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] (1) The target detection and recognition method based on bat sonar echo Doppler compensation bionics of the present invention, compared with the conventional method of using fixed transmission sonar signals, adopts a dynamic adjustment transmission frequency mode. After the echo signal passes through the narrowband filter, the random noise introduced is less and the frequency spectrum range is narrowed, thus having stronger anti-interference ability, more sensitive detection ability, and effectively reducing the difficulty of subsequent recognition algorithms.

[0018] (2) The target detection and recognition method based on bat sonar echo Doppler compensation bionics of the present invention reduces the requirements for echo signal processing hardware. Specifically, it can use a low-bandwidth digital-to-analog converter and reduce the computing power requirements of the signal processor, thereby helping to reduce hardware resource occupation, improve system compactness and integration, and achieve low-cost target detection and recognition.

[0019] (3) The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic of the present invention constructs a lightweight dilated convolutional neural network model. This model combines different convolution strategies. First, depthwise separable convolution is used to obtain a larger receptive field through dilated convolution while maintaining low computational cost. After depthwise separable convolution and dilated convolution, standard convolution is used to further refine local features and supplement detailed information, so as to achieve a good balance between "high accuracy and low computational cost" in image classification tasks. Attached Figure Description

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] Figure 1 This is a flowchart of a target detection and recognition method based on bat sonar echo Doppler compensation biomimetic.

[0022] Figure 2 It is the echo time-frequency diagram under dynamic transmission frequency mode;

[0023] Figure 3 It is the echo time-frequency diagram under the fixed transmission frequency mode;

[0024] Figure 4 This is the echo amplitude diagram under dynamic transmission frequency mode;

[0025] Figure 5 This is a diagram of the lightweight dilated convolutional neural network model structure;

[0026] Figure 6 This is a comparison chart of the recognition accuracy of a lightweight dilated convolutional neural network model and a traditional convolutional neural network model; Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0028] like Figure 1 As shown, the target detection and recognition method based on bat sonar echo Doppler compensation biomimetic includes the following steps:

[0029] Step 1: Transmit a sonar signal to the target at an initial frequency of 30 kHz;

[0030] Sub-step 1.1: Because bats primarily target the constant-frequency portion of their emission frequency when performing Doppler compensation, this invention uses a constant-frequency signal for biomimetic research. Therefore, the constant-frequency signal emitted by a bat can be expressed as:

[0031]

[0032] Where A represents the signal amplitude, f c T represents the carrier frequency and T represents the pulse width.

[0033] Sub-step 1.2: Assuming the total number of pulses transmitted within a certain time period is N, the transmission signal corresponding to the nth pulse can be expressed as:

[0034]

[0035] Where rect(·) is the rectangular pulse function, Tr is the pulse repetition period, and f cn Let be the carrier frequency of the nth pulse;

[0036] Step 2: Establish a model of forearm swing motion with different heights and body shapes when walking at different speeds and with different arm swing amplitudes;

[0037] Sub-step 2.1: The human motion process constructed based on the Boulic model is manifested as the motion of joint connections. Assuming the fixed motion duration is 1 second, specifically, the spatial positions of the 17 feature points in the Boulic model can be parameterized by mathematical methods, and different human motion models can be obtained by changing the human walking speed, height, forearm size and forearm swing amplitude.

[0038] Step 3: Receive sonar echo signals that generate stable echo frequencies under different motion models;

[0039] Sub-step 3.1: Establish the target echo model. By analyzing the information carried by the target echo, obtain the target's velocity and distance. The initial time is t = 0. Let the initial radial distance between the target and the sonar be R0. The target moves in a curved path with a radial velocity v in the direction of sonar movement. Assuming this direction is positive, the radial distance at time t is R0. n =R0-vt, when the bat emits a signal at time t, the time delay τ of receiving the reflected echo from the target is:

[0040] τ=2R n / c (3)

[0041] The echo signal of the nth pulse within the observation duration can be expressed as:

[0042]

[0043] Where, σ n This represents the scattering cross section of the forearm corresponding to the nth pulse, and the nth Doppler frequency shift is f. dn =2vf cn / c;

[0044] Sub-step 3.2 establishes an echo signal model that conforms to the actual environment:

[0045] y(n,t)=x(n,t)+s(t) (5)

[0046] Where s(t) is Gaussian white noise;

[0047] Step 4: Input the echo signal into a narrowband filter for filtering to obtain the time-frequency diagram of the signal;

[0048] Sub-step 4.1: In order to reduce the influence of noise from other frequencies, the frequency range is determined based on the known desired frequency and bandwidth, and the parameters of the narrowband filter are determined accordingly. After filtering, a relatively pure original echo signal is obtained.

[0049] Sub-step 4.2 involves performing time-frequency analysis on the filtered signal. Under the dynamic transmission frequency mode provided by this invention, compensation is made for the Doppler frequency shift caused by target motion, stabilizing the echo signal frequency within a narrow frequency band. Figure 2 As shown, with Figure 3 Compared to the fixed transmission frequency mode shown, this mode effectively suppresses the introduction of out-of-band noise, narrows the recognition bandwidth, and thus enhances the quality of the echo signal and improves the sensitivity of target detection.

[0050] Step 5: Perform time-frequency analysis on the time-frequency diagram to estimate the Doppler frequency offset, and dynamically adjust the frequency of subsequent sonar transmissions based on the frequency offset.

[0051] Sub-step 5.1, Initial velocity measurement (first sonar transmission cycle)

[0052] During the first transmission cycle, the sonar uses the initial transmission frequency f. c1 When detecting a target on a swing arm, based on the Doppler effect principle, the Doppler frequency shift f produced by the target... d1 for:

[0053]

[0054] Where v1 is the swing arm velocity measured in the first measurement, c is the speed of signal propagation in air (sound speed c = 340 m / s), and f can be extracted through time-frequency analysis at the sonar receiver. d1 Then, the swing arm speed v1 measured in the first measurement can be obtained by reverse deduction;

[0055] Meanwhile, the actual echo frequency f measured for the first time e1 It is obtained by adding the transmission frequency and the Doppler frequency shift:

[0056]

[0057] Because the initial transmission frequency was not compensated, f e1 It may deviate from the frequency of a specific region, and subsequent steps are needed to adjust the transmission frequency to achieve compensation.

[0058] Sub-step 5.2: Dynamic transmission frequency adjustment (second sonar transmission cycle)

[0059] By drawing inspiration from the Doppler compensation mechanism in bats, and stabilizing the echo frequency within a specific region of the fovea of ​​hearing, the desired frequency f can be defined. target A highly extended frequency characterization region (f) is formed around it. min, f max ), that is, a specific region, where f min f is the minimum frequency in this region. max Let f be the maximum frequency in this region.target The value at the midpoint of a specific frequency range can be represented as:

[0060] f target =(f max +f min ) / 2 (8)

[0061] The core objective of the second transmission cycle is to adjust the transmission frequency f c2 This makes the echo frequency f of the second measurement. e2 To get as close as possible to the desired frequency, since the swing arm speed v2 of the second measurement cannot be directly measured before launch, but the swing arm speed has the characteristic of continuity between adjacent moments, the speed v1 of the first measurement is used as the estimate of the second echo, and thus the compensated launch frequency is obtained.

[0062] According to the Doppler effect, the expected echo frequency of the second measurement should satisfy f. e2 =f target The desired echo frequency is related to the transmission frequency and the target velocity as follows:

[0063]

[0064] By reversing the above formula, we obtain the second optimized transmission frequency f. c2 for:

[0065]

[0066] Using f c2 The transmission frequency of the second cycle is used again to extract the second actual Doppler frequency shift f through signal processing at the receiving end. d2 Therefore, the actual swing arm speed v2 in the second calculation is:

[0067]

[0068] Therefore, the actual echo frequency f of the second echo is e2 for:

[0069]

[0070] Because of the adjacent v1 and v2 The difference is not significant, f e2 It will be significantly close to f target This achieves the first compensation of the echo frequency to the frequency of a specific region;

[0071] Sub-step 5.3: Iterative compensation (nth sonar transmission cycle, n≥2)

[0072] To maintain the stability of the echo frequency, an iterative strategy is adopted: the transmission frequency f of the nth transmission is... cnBased on the actual velocity v in the (n-1)th iteration n-1 Calculations are performed to ensure that each transmission frequency adjustment is based on the latest speed information and adapts to the dynamic changes in the swing arm speed;

[0073] The nth time to optimize the transmission frequency f cn The formula is:

[0074]

[0075] The corresponding nth velocity v n for:

[0076]

[0077] The true echo frequency f of the nth time en The formula is:

[0078]

[0079] Through this iterative process, f can be made en Continuously stable at f target Nearby, real-time compensation for dynamic changes in the swing arm speed is achieved;

[0080] Step 6: Extract the compensated echo amplitude features to generate a dataset for training and recognition, and randomly divide the dataset into training set, validation set and test set according to the proportion.

[0081] Sub-step 6.1, in dynamic transmission frequency mode, performs amplitude analysis on the filtered signal, such as... Figure 4 As shown, the forearm swing frequency and amplitude differ at different walking speeds, and these differences are directly reflected in the time-domain characteristics of the amplitude diagram, such as the peak amplitude, the time interval between peaks, and the fluctuation of the amplitude envelope. Therefore, the walking speed can be identified based on the variation of the forearm echo signal amplitude over time.

[0082] Sub-step 6.2 involves generating a forearm swing dataset for model training by adjusting parameters such as human walking speed, height, forearm size, and forearm swing amplitude. Three labels are assigned to the human walking speed, with baseline values ​​of 0.9 m / s, 1.2 m / s, and 1.5 m / s. To simulate accuracy errors in actual equipment testing, a fluctuation of ±0.1 m / s is introduced around these baseline values, thus setting the actual speed ranges for the three categories to 0.8–1.0 m / s, 1.1–1.3 m / s, and 1.4–1.6 m / s, respectively. Furthermore, values ​​within each speed range are taken at 0.1 m / s intervals. The height parameter in each speed label ranges from 1.45m to 1.9m, with an interval of 0.05m. The forearm size parameter in each speed label is defined as three categories: thin, medium, and thick, with four specific size values ​​set under each category. The forearm swing amplitude parameter in each speed label is defined as three amplitude levels: low, medium, and high, with a specific coefficient set under each level. Through the combination of the above parameters, a total of 1080 unique simulation samples are generated for each speed label, and a total of 3240 samples are generated for the three speed labels. Finally, classification labels are set for the three types of motion speeds according to the sequence numbers 1 to 3.

[0083] Sub-step 6.3: In each class of dataset, randomly select amplitude diagrams generated by forearm swing and combine them into training set, validation set and test set in a ratio of 7:2:1;

[0084] Step 7: Construct a lightweight dilated convolutional neural network model;

[0085] Sub-step 7.1 converts the amplitude map dataset into a single channel to prevent the inclusion of three-channel images and the resulting information redundancy. The image size is unified to [224, 244], and data augmentation operations are performed only on the training set. Specific transformations include: random rotation, random translation, random scaling, and vertical flipping.

[0086] Sub-step 7.2: Establish the input layer to receive single-channel grayscale images;

[0087] Sub-step 7.3 establishes a depthwise separable convolution module, which includes 3×3 channel group convolution, batch normalization, ReLU activation, 1×1 point convolution, batch normalization, ReLU activation and 2×2 max pooling. While ensuring feature extraction capability, it significantly reduces the amount of computation and parameters, and achieves network lightweighting.

[0088] Sub-step 7.4: Establish an expanded convolution module with 3×3 convolution, dilation rate of 2, batch normalization, ReLU activation and 2×2 max pooling. The 3×3 convolution with dilation rate of 2 is equivalent to the receptive field of a 5×5 convolution. Without increasing the kernel size and computational cost, the receptive field is expanded to capture long-distance dependencies in the image.

[0089] Sub-step 7.5 establishes a standard convolution module, consisting of 3×3 convolution, batch normalization, ReLU activation, and 2×2 max pooling. This, combined with the preceding convolution, forms a progressive feature learning chain of "lightweight feature extraction - global receptive field expansion - local feature refinement," balancing efficiency and accuracy.

[0090] Sub-step 7.6: After multiple feature extractions, a fully connected layer maps the high-dimensional features to an output dimension equal to the number of categories. Then, a Softmax activation function is used to generate the predicted probability of each category. Finally, a classification loss layer is used to calculate the loss value between the predicted probability distribution and the true category label of the image. The backpropagation mechanism guides the network parameter update, and finally the classification result is output.

[0091] Sub-step 7.7 generates a lightweight dilated convolutional neural network model, such as... Figure 5 As shown; Step 8: Input the training set into the lightweight dilated convolutional neural network model to obtain the target recognition model, and input the test set into the recognition model to obtain the target speed recognition result;

[0092] Sub-step 8.1 employs the Adam optimizer with an initial learning rate of 0.001. The learning rate is scheduled to decay in stages (decreasing to 50% of the current value every 10 epochs), the batch size is 32, and the maximum number of training epochs is 20. After each training epoch, the model performance is evaluated using a validation set, and the training state is dynamically adjusted.

[0093] Sub-step 8.2 inputs the enhanced training set into the lightweight dilated convolutional neural network model, updates the parameters through backpropagation, monitors the overfitting risk with the validation set accuracy as an indicator, and finally saves the trained model.

[0094] Sub-step 8.3 uses an independent test set with randomly divided images (which did not participate in parameter updates or data augmentation during training) to evaluate the model performance and obtain the classification results;

[0095] The effects of the present invention will be further described below with reference to simulation experiments.

[0096] 1. Simulation conditions

[0097] This invention fixes the duration of the human walking process to 1 second, simulates a sonar signal with an initial transmission frequency of 30KHz, with each signal pulse having a repetition period of 20 milliseconds, a pulse width of 5 milliseconds, a fixed number of pulses of 52, and a fixed sampling frequency of 100KHz.

[0098] 2. Simulation Content and Result Analysis

[0099] The training set, validation set, and test set used in the simulation experiment of this invention are three types of moving targets constructed according to step 6 of this invention. The training set, validation set, and test set are randomly divided in a ratio of 7:2:1. Data augmentation is performed on the training set to improve the model's generalization ability. The training set has a total of 2268 samples, with 756 samples in each class; the validation set has a total of 648 samples, with 216 samples in each class; and the test set has a total of 324 samples, with 108 samples in each class.

[0100] The three moving targets are 0.9 m / s, 1.2 m / s, and 1.5 m / s;

[0101] Input the test set into the trained lightweight dilated convolutional neural network model to obtain the classification results;

[0102] from Figure 6 It can be seen that the lightweight dilated convolutional neural network model proposed in this invention is significantly better than the traditional convolutional neural network method in terms of recognition accuracy. At the same time, its accuracy curve is more stable under different test conditions, showing excellent robustness. Even in a harsh environment with a low signal-to-noise ratio of -15dB, the model's recognition accuracy can still be maintained above 95%.

[0103] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A target detection and recognition method based on bat sonar echo Doppler compensation biomimetic, comprising the following steps: Step 1: Transmit a sonar signal to the target at an initial frequency of 30 kHz; Step 2: Establish a model of forearm swing motion with different heights and body shapes when walking at different speeds and with different arm swing amplitudes; Step 3: Receive sonar echo signals that generate stable echo frequencies under different motion models; Step 4: Input the echo signal into a narrowband filter for filtering to obtain the time-frequency diagram of the signal; Step 5: Perform time-frequency analysis on the time-frequency diagram to estimate the Doppler frequency offset, and dynamically adjust the frequency of subsequent sonar transmissions based on the frequency offset. Step 6: Extract the compensated echo amplitude features to generate a dataset for training and recognition, and randomly divide the dataset into training set, validation set and test set according to the proportion. Step 7: Construct a lightweight dilated convolutional neural network model; Step 8: Input the training set into the lightweight dilated convolutional neural network model to obtain the target recognition model, and input the test set into the recognition model to obtain the target speed recognition result.

2. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 1, characterized in that, Step 1 includes the following sub-steps: Sub-step 1.1: Because bats primarily target the constant-frequency portion of their emission frequency when performing Doppler compensation, this invention uses a constant-frequency signal for biomimetic research. Therefore, the constant-frequency signal emitted by a bat can be expressed as: Where A represents the signal amplitude, f c T represents the carrier frequency and T represents the pulse width. Sub-step 1.2: Assuming the total number of pulses transmitted within a certain time period is N, the transmission signal corresponding to the nth pulse can be expressed as: Where rect(·) is the rectangular pulse function, Tr is the pulse repetition period, and f cn Let be the carrier frequency of the nth pulse.

3. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 2, characterized in that, Step 2 includes the following sub-steps: Sub-step 2.1: The human motion process constructed based on the Boulic model is manifested as the motion of joint connections. Assuming a fixed motion duration of 1 second, the spatial positions of the 17 feature points in the Boulic model can be parameterized by mathematical methods, and different human motion models can be obtained by changing the human walking speed, height, forearm size, and forearm swing amplitude.

4. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 3, characterized in that, Step 3 includes the following sub-steps: Sub-step 3.1: Establish the target echo model. By analyzing the information carried by the target echo, obtain the target's velocity and distance. The initial time is t = 0. Let the initial radial distance between the target and the sonar be R0. The target moves in a curved path with a radial velocity v in the direction of sonar movement. Assuming this direction is positive, the radial distance at time t is R0. n =R0-vt, when the bat emits a signal at time t, the time delay τ of receiving the reflected echo from the target is: τ=2R n / c The echo signal of the nth pulse within the observation duration can be expressed as: Where, σ n This represents the scattering cross section of the forearm corresponding to the nth pulse, and the nth Doppler frequency shift is f. dn =2vf cn / c; Sub-step 3.2 establishes a model that conforms to the echo signal generated in the actual environment as follows: y(n,t) = x(n,t) + s(t) Where s(t) is Gaussian white noise.

5. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 4, characterized in that, Step 4 includes the following sub-steps: Sub-step 4.1: In order to reduce the influence of noise from other frequencies, the frequency range is determined based on the known desired frequency and bandwidth, and the parameters of the narrowband filter are determined accordingly. After filtering, a relatively pure original echo signal is obtained. Sub-step 4.2 involves performing time-frequency analysis on the filtered signal. In the dynamic transmission frequency mode provided by this invention, the Doppler frequency shift caused by the target motion is compensated, so that the echo signal frequency is stabilized within a narrow frequency band. Compared with the fixed transmission frequency mode, this mode effectively suppresses the introduction of out-of-band noise, narrows the recognition bandwidth, and thus enhances the quality of the echo signal and improves the sensitivity of target detection.

6. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 5, characterized in that, Step 5 includes the following sub-steps: Sub-step 5.1, Initial velocity measurement (first sonar transmission cycle) During the first transmission cycle, the sonar uses the initial transmission frequency f. c1 When detecting a target on a swing arm, based on the Doppler effect principle, the Doppler frequency shift f produced by the target... d1 for: Where v1 is the swing arm velocity measured in the first measurement, c is the speed of signal propagation in air (sound speed c = 340 m / s), and f can be extracted through time-frequency analysis at the sonar receiver. d1 Then, the swing arm speed v1 measured in the first measurement can be deduced from the reverse. Meanwhile, the actual echo frequency f measured for the first time e1 It is obtained by adding the transmission frequency and the Doppler frequency shift: Because the initial transmission frequency was not compensated, f e1 It may deviate from the frequency of a specific region, and subsequent steps are needed to adjust the transmission frequency to achieve compensation. Sub-step 5.2: Dynamic transmission frequency adjustment (second sonar transmission cycle) By drawing inspiration from the Doppler compensation mechanism in bats, and stabilizing the echo frequency within a specific region of the fovea of ​​hearing, the desired frequency f can be defined. target A highly extended frequency characterization region (f) is formed around it. min f max ), that is, a specific region, where f min f is the minimum frequency in this region. max Let f be the maximum frequency in this region. target The value at the midpoint of a specific frequency range can be represented as: f target =(f max +f min ) / 2 The core objective of the second transmission cycle is to adjust the transmission frequency f c2 The echo frequency f of the second measurement e2 To get as close as possible to the desired frequency, since the swing arm speed v2 of the second measurement cannot be directly measured before launch, but the swing arm speed has the characteristic of continuity between adjacent moments, the speed v1 of the first measurement is used as the estimate of the second echo, and thus the compensated launch frequency is obtained. According to the Doppler effect, the expected echo frequency of the second measurement should satisfy f. e2 =f target The desired echo frequency is related to the transmission frequency and the target velocity as follows: By reversing the above formula, we obtain the second optimized transmission frequency f. c2 for: Using f c2 The transmission frequency of the second cycle is used again to extract the second actual Doppler frequency shift f through signal processing at the receiving end. d2 Therefore, the actual swing arm speed v2 in the second calculation is: Therefore, the actual echo frequency f of the second echo is e2 for: Since the adjacent v1 and v2 are not much different, f e2 It will be significantly close to f target This achieves the first compensation of the echo frequency to the frequency of a specific region; Sub-step 5.3: Iterative compensation (nth sonar transmission cycle, n≥2) To maintain the stability of the echo frequency, an iterative strategy is adopted: the transmission frequency f of the nth transmission is... cn Based on the actual velocity v in the (n-1)th iteration n-1 Calculations are performed to ensure that each transmission frequency adjustment is based on the latest speed information and adapts to the dynamic changes in the swing arm speed; The nth time to optimize the transmission frequency f cn The formula is: The corresponding nth velocity v n for: The true echo frequency f of the nth time en The formula is: Through this iterative process, f can be made en Continuously stable at f target Nearby, it enables real-time compensation for dynamic changes in the swing arm speed.

7. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 6, characterized in that, Step 6 includes the following sub-steps: Sub-step 6.1: In dynamic transmission frequency mode, perform amplitude analysis on the filtered signal. Since the forearm swing frequency and forearm swing amplitude differ at different walking speeds, these differences will be directly reflected in the time domain characteristics of the amplitude diagram, such as the peak size of the amplitude, the time interval between peak occurrences, and the fluctuation of the amplitude envelope. Therefore, the walking speed of the human body can be identified based on the variation law of the amplitude of the forearm echo signal over time. Sub-step 6.2 involves generating a forearm swing dataset for model training by adjusting parameters such as human walking speed, height, forearm size, and forearm swing amplitude. Three labels are assigned to the human walking speed, with baseline values ​​of 0.9 m / s, 1.2 m / s, and 1.5 m / s. To simulate accuracy errors in actual equipment testing, a fluctuation of ±0.1 m / s is introduced around these baseline values, thus setting the actual speed ranges for the three categories to 0.8–1.0 m / s, 1.1–1.3 m / s, and 1.4–1.6 m / s, respectively. Furthermore, values ​​within each speed range are taken at 0.1 m / s intervals. The height parameter in each speed label ranges from 1.45m to 1.9m, with an interval of 0.05m. The forearm size parameter in each speed label is defined as three categories: thin, medium, and thick, with four specific size values ​​set under each category. The forearm swing amplitude parameter in each speed label is defined as three amplitude levels: low, medium, and high, with a specific coefficient set under each level. Through the combination of the above parameters, a total of 1080 unique simulation samples are generated for each speed label, and a total of 3240 samples are generated for the three speed labels. Finally, classification labels are set for the three types of motion speeds according to the sequence numbers 1 to 3. Sub-step 6.3: In each class of dataset, randomly select amplitude diagrams generated by forearm swing and combine them into training set, validation set and test set in a ratio of 7:2:

1.

8. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 7, characterized in that, Step 7 includes the following sub-steps: Sub-step 7.1 converts the amplitude map dataset into a single channel to prevent the inclusion of three-channel images and the resulting information redundancy. The image size is unified to [224.244], and data augmentation operations are performed only on the training set. The specific transformations include: random rotation, random translation, random scaling, and vertical flipping. Sub-step 7.2: Establish the input layer to receive single-channel grayscale images; Sub-step 7.3 establishes a depthwise separable convolution module, which includes 3×3 channel group convolution, batch normalization, ReLU activation, 1×1 point convolution, batch normalization, ReLU activation and 2×2 max pooling. While ensuring feature extraction capability, it significantly reduces the amount of computation and parameters, and achieves network lightweighting. Sub-step 7.4: Establish an expanded convolution module with 3×3 convolution, dilation rate of 2, batch normalization, ReLU activation and 2×2 max pooling. The 3×3 convolution with dilation rate of 2 is equivalent to the receptive field of a 5×5 convolution. Without increasing the kernel size and computational cost, the receptive field is expanded to capture long-distance dependencies in the image. Sub-step 7.5 establishes a standard convolution module, consisting of 3×3 convolution, batch normalization, ReLU activation, and 2×2 max pooling. This, combined with the preceding convolution, forms a progressive feature learning chain of "lightweight feature extraction - global receptive field expansion - local feature refinement," balancing efficiency and accuracy. Sub-step 7.6: After multiple feature extractions, a fully connected layer maps the high-dimensional features to an output dimension equal to the number of categories. Then, a Softmax activation function is used to generate the predicted probability of each category. Finally, a classification loss layer is used to calculate the loss value between the predicted probability distribution and the true category label of the image. The backpropagation mechanism guides the network parameter update, and finally the classification result is output. Sub-step 7.7 generates a lightweight dilated convolutional neural network model.

9. The target detection and recognition method based on bat sonar echo Doppler compensation biomimetic as described in claim 8, characterized in that, Step 8 includes the following sub-steps: Sub-step 8.1 employs the Adam optimizer with an initial learning rate of 0.

001. The learning rate is scheduled to decay in stages (decreasing to 50% of the current value every 10 epochs), the batch size is 32, and the maximum number of training epochs is 20. After each training epoch, the model performance is evaluated using a validation set, and the training state is dynamically adjusted. Sub-step 8.2 inputs the enhanced training set into the lightweight dilated convolutional neural network model, updates the parameters through backpropagation, monitors the overfitting risk with the validation set accuracy as an indicator, and finally saves the trained model. Sub-step 8.3 uses an independent test set with randomly divided images (which did not participate in parameter updates or data augmentation during training) to evaluate the model performance and obtain the classification results.

Images