Novel acoustic signal characteristic frequency identification method based on deep learning algorithm

Abstract

The invention discloses a novel acoustic signal characteristic frequency identification method based on a deep learning algorithm, and the method comprises the steps: carrying out the identification of the characteristic frequency of an acoustic signal through innovative combination with a deep algorithm; providing a broadband ultrasonic transducer as an acoustic signal receiver; performing feature extraction on the acoustic signal received by the broadband receiver to form a small sample feature data set; the deep learning algorithm performs small sample learning (FSL) based on a feature data set, converges an optimal solution based on a genetic optimization algorithm, and autonomously generates feature distribution of acoustic signals in a wider frequency band range, thereby realizing accurate identification of acoustic signal feature frequencies. According to the method, training of a deep learning algorithm is completed through a small sample feature data set, the recognition of the acoustic signal feature frequency has broadband benefits in combination with the algorithm, meanwhile, the limitation of the 6dB working frequency band of a traditional acoustic transducer is broken through, and the method has good frequency recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of acoustics, and in particular to a novel acoustic signal characteristic frequency recognition method based on a deep learning algorithm. Background Art

[0002] In the field of acoustics, target recognition refers to the detection, tracking, imaging, identification and classification of targets based on their characteristics through active or passive equipment supplemented by deep learning algorithms, sensitive sensor devices, visual image processing and other comprehensive approaches. It has important practical significance in many fields such as biological protection, economic development and people's livelihood.

[0003] With the rapid development of modern sensing technology and deep learning algorithms, the research and application of target detection and recognition and classification technology have become more mature. R2 Sonic Co., Ltd. (CN116794661A) provides a method for multi-signal messages and deconstructing matching multi-signal echoes; Harbin Engineering University provides a sound source passive positioning system based on the cross-correlation function matching of received signals in the deep-sea convergence zone and its positioning method (CN116840842A). However, most of the existing research focuses on the positioning of targets, the detection of the presence of targets, the imaging of targets, and the classification of targets. In contrast, the research on active target characteristic frequency recognition is relatively scarce.

[0004] The defects of existing technologies and research include: existing technologies and research lack an identification method with wide-band working efficiency and targeting the characteristic frequency of acoustic signals; traditional methods based on ultrasonic transducers are restricted by the frequency bandwidth of the transducer itself, have a single identification range, poor device flexibility and a complex design process; target frequency identification based on traditional methods relies on empirical judgment and cannot meet the needs of modern intelligent detection and identification equipment; existing methods require large sample data to train deep learning algorithms, and have low detection accuracy.

[0005] Therefore, it is of innovative significance to provide a new acoustic signal characteristic frequency recognition method based on deep learning algorithm. Summary of the invention

[0006] In order to optimize the shortcomings and defects of the prior art, the present invention proposes a new acoustic signal characteristic frequency recognition method based on a deep learning algorithm, including: innovatively combining a deep learning algorithm with a wide-band ultrasonic transducer to achieve high-accuracy recognition of the characteristic frequency of the acoustic signal in a wide frequency range, while broadening the effective working frequency band of the ultrasonic transducer.

[0007] Deep learning algorithms include but are not limited to feedforward neural network algorithm (MLP), recurrent neural network algorithm (RNN), long short-term memory network algorithm (LSTM), generative adversarial network algorithm (GAN), and autoencoder.

[0008] The deep learning algorithm converges to the optimal solution based on the optimization algorithm, which includes but is not limited to gradient descent, RMSProp, AdaGrad, Adam, genetic algorithm, and simulated annealing algorithm.

[0009] The broadband ultrasonic transducer serves as an acoustic signal receiver, and includes a single broadband ultrasonic transducer unit and an arrayed broadband ultrasonic transducer.

[0010] The deep learning algorithm is trained based on the feature dataset to perform small sample learning (FSL) and generate the characteristic distribution of acoustic signals in a wider frequency band.

[0011] The training of deep learning algorithms includes training based on a single feature dataset and training based on a multi-feature dataset.

[0012] Perform data scale transformation on feature data sets, including but not limited to normalization, standardization, data scaling, and scale translation.

[0013] The characteristic values ​​in the characteristic data set include at least one of the signal oscillation attenuation period characteristics, the signal Fourier spectrum characteristics, the signal pulse characteristics, the signal energy and power characteristics, the signal phase characteristics, and the signal's own voltage and current characteristics.

[0014] Preferably, the broadband ultrasonic transducer is a broadband ultrasonic transducer based on a discrete ultrasonic transducer array.

[0015] Preferably, the deep learning algorithm is an MLP algorithm optimized based on a genetic optimization algorithm.

[0016] Preferably, the deep learning algorithm trained using the feature data set outputs a deep learning result that stores the global optimal solution, and the optimal solution learning result is used to accurately identify the signal characteristic frequency.

[0017] Compared with the prior art, the novel acoustic signal characteristic frequency recognition method based on deep learning algorithm provided by the present invention has the following beneficial effects, including:

[0018] The method achieves high-accuracy recognition of characteristic frequencies of acoustic signals, requires only a small number of discrete ultrasonic transducers as a whole, has low device costs, and does not require the development of new processes.

[0019] The method is combined with a deep learning algorithm, so that the recognition of characteristic frequencies of acoustic signals has a wide-band working benefit, effectively overcoming the defect of the traditional method of detecting a single frequency band.

[0020] The method effectively optimizes the problem of inaccurate recognition of deep learning algorithms under the premise of small sample training data, improves the problem that traditional methods are highly dependent on experience judgment in the field of target detection and recognition, and is simple to operate and has a fast response speed.

[0021] The method can make the ultrasonic transducer work in a state that breaks through the -6dB working frequency band limit, effectively broadens the effective working frequency band range of the ultrasonic transducer, and cleverly optimizes the shortcomings of traditional ultrasonic transducers, hydrophones and their arrays that have a single working frequency band and poor adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings:

[0023] Figure 1 This is a flow chart of a novel acoustic signal characteristic frequency recognition method based on a deep learning algorithm of the present invention;

[0024] Figure 2 The topological structure of the MLP neural network algorithm used in the embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the structure of a discrete ultrasonic transducer unit used in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to illustrate the present invention in more detail, the technical solution of the present invention is further described below in conjunction with preferred embodiments and drawings. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0027] In the present invention, it is necessary to understand that the term "small amount" is to illustrate the advantage that this embodiment only requires a small sample ultrasonic transducer rather than an array; it is necessary to understand that the term "optimal" is to illustrate the output index of the deep learning algorithm in this embodiment; it is necessary to understand that the term "connection" includes electrical connection and mechanical connection; it can also be a direct connection or an indirect connection through an intermediate structure; it can be an integral connection or a partial connection. Therefore, the terms cannot be understood as limiting the present invention. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0028] Example:

[0029] Reference Figure 1 This embodiment provides a novel acoustic signal characteristic frequency recognition method based on a deep learning algorithm, comprising:

[0030] Innovatively combining deep learning algorithms with broadband ultrasonic transducers, the characteristic frequencies of acoustic signals can be accurately identified in a wide frequency range with high accuracy, while broadening the effective working frequency band of the ultrasonic transducer. This embodiment uses a discrete ultrasonic transducer array as a broadband ultrasonic transducer, and each ultrasonic transducer unit has a large aspect ratio of at least 6 times. During the recognition process, the acoustic signal is received by the discrete ultrasonic transducer array and output and stored in the form of an electrical signal. Features of the stored electrical signals are extracted to form a feature data set, and this embodiment extracts the voltage features of the electrical signals. The characteristic frequencies of acoustic signals are identified using a deep learning algorithm trained with the feature data set. In this embodiment, reference is made to Figure 2 , a lightweight MLP algorithm is selected, and the acoustic signal is an underwater acoustic signal.

[0031] Reference Figure 3 , which is a schematic diagram of a single unit structure in the discrete ultrasonic transducer array of this embodiment. This embodiment preferably uses PZT-5H, metal copper Au, conductive silver paste material and silicon Si substrate. In this embodiment, a total of four discrete ultrasonic transducers are prepared, namely P Ⅰ -P Ⅳ , and the U Ⅰ -U Ⅳ The combined operating frequency band is distributed in the range of 5.5MHz-7.5MHz.

[0032] The U Ⅰ -U Ⅳ In this embodiment, the process of sending and receiving acoustic signals includes the following steps:

[0033] T1. The sending and receiving process of acoustic signals takes place in a test tank filled with water;

[0034] T2. Preferably, ten groups of sound source probes are selected as acoustic signal transmitters, respectively named S Ⅰ -S X The S Ⅰ -S X The operating frequency is distributed in 0.5MHz-3.5MHz, the S Ⅰ -S X are excited by the same excitation source and emit ultrasonic signals;

[0035] T3, the U Ⅰ -U Ⅳ Sense and receive S described in T1 Ⅰ -S X The emitted ultrasonic signal is finally output in the form of a time-domain continuous electrical signal through the piezoelectric effect and stored in a storage unit.

[0036] In this embodiment, a recognition test process of a novel acoustic signal characteristic frequency recognition method based on a deep learning algorithm includes:

[0037] P1. Extract features from the stored electrical signals and store the feature values ​​in the form of feature data sets. After normalization and scaling, the feature data sets are classified into training sets and test sets. The training set is used to train the deep learning algorithm, and the test set is used to verify the accuracy of the new acoustic signal feature frequency recognition method based on deep learning.

[0038] P2. Import the data set for training and testing. Subsequently, a multi-layer perceptron model is constructed using the back propagation algorithm. The perceptron model architecture of this embodiment includes three hidden layers, each of which has a predetermined number of nodes. Configure the parameters for model training, including the maximum number of iterations, the error threshold, and the learning rate. In this experiment, the maximum number of iterations is 1000, the error threshold is set to 1e-6, and the learning rate is 1e-2.

[0039] This embodiment uses a genetic algorithm to iteratively evolve a series of solutions to optimize the performance of the neural network. The genetic generation, population size and mutation strategy are adjusted to ensure efficient exploration of the solution space. In the embodiment, the genetic generation is 100, the population size is 10, the mutation strategy includes non-uniform mutation, and the parameters are [2,100,3]. After convergence, the optimal model parameters are decoded from the genetic representation. These parameters are seamlessly integrated into the neural network architecture, enhancing its predictive ability. Then, the model is trained using the training data set, and the performance is monitored through an adaptive training window. After the training is completed, the effect of the model is evaluated by simulating on the training and test data sets. The root mean square error index is calculated to measure the prediction accuracy and ensure the robustness and reliability of the model. The optimization trajectory is further visualized through the evolution curve, illustrating the convergence behavior of the genetic algorithm. Further, by comparing the predicted value with the actual value, the predictive performance of the model is demonstrated, and an analysis of the model effect and generalization ability is provided.

[0040] After the algorithms in P3 and P2 are trained and converged to the optimal solution, they are further used to identify underwater acoustic signals of unknown operating frequencies. In this embodiment, underwater acoustic signals with operating frequencies of 1.00 MHz and 2.39 MHz are tested, and recognition results with accuracy rates of 97.36% and 97.07% are obtained.

[0041] The embodiment is based on a wideband ultrasonic transducer array composed of four discrete ultrasonic transducers, and obtains a small sample feature data set outside the -6dB operating frequency band of the ultrasonic transducer. Finally, the deep learning algorithm trained with small sample data is used to accurately recognize the characteristic frequency of the acoustic signal, with a recognition accuracy of at least 90%.

[0042] Finally, it should be explained that the above-described embodiments are only preferred examples for clearly illustrating the present invention, but this is not a limitation on the implementation scheme of the present invention. Ordinary technicians in the relevant field should understand that the technical features in the above schemes can be combined arbitrarily, and other different forms of modifications or equivalent replacements of some technical features can be made on the basis of the above specific implementation schemes. It is impossible to list all implementation schemes here. Therefore, all modifications, improvements, equivalent replacements, etc. derived from the technical schemes of the present invention within the spirit and principles of the present invention should be within the technical scope of the present invention.

Claims

1. A new acoustic signal characteristic frequency recognition method based on deep learning algorithm, comprising: The innovative combination of deep learning algorithms and wide-band ultrasonic transducers can achieve high-accuracy recognition of the characteristic frequencies of acoustic signals in a wide frequency range, while broadening the effective operating frequency band of the ultrasonic transducer.

2. According to a novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 1, it is characterized in that: The deep learning algorithms include but are not limited to feedforward neural network algorithm (MLP), recurrent neural network algorithm (RNN), long short-term memory network algorithm (LSTM), generative adversarial network algorithm (GAN), and autoencoder.

3. According to the novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 1, it is characterized in that: The deep learning algorithm converges to the optimal solution based on an optimization algorithm, and the optimization algorithm includes but is not limited to gradient descent, RMSProp, AdaGrad, Adam, genetic algorithm, and simulated annealing algorithm.

4. The novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 1 is characterized in that: The broadband ultrasonic transducer serves as an acoustic signal receiver, and includes a single broadband ultrasonic transducer unit and an arrayed broadband ultrasonic transducer.

5. The novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 1 is characterized in that: The deep learning algorithm performs small sample learning (FSL) based on feature data set training and generates feature distribution of acoustic signals in a wider frequency band.

6. The novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 5 is characterized in that: The training of the deep learning algorithm also includes training based on a single feature data set and training based on a multi-feature data set.

7. The novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 5 is characterized in that: The feature data set requires data scale transformation, including but not limited to normalization, standardization, data scaling, and scale translation.

8. The novel acoustic signal characteristic frequency recognition method based on deep learning algorithm according to claim 5 is characterized in that: The characteristic values ​​in the characteristic data set include at least one of the signal oscillation attenuation period characteristics, the signal Fourier spectrum characteristics, the signal pulse characteristics, the signal energy and power characteristics, the signal phase characteristics, and the signal's own voltage and current characteristics.

Images