Gimbal control sound source tracking method and system based on nested MEMS array

Abstract

The application provides a gimbal control sound source tracking method and system based on a nested MEMS array, relates to the technical field of sound source tracking, and comprises the following steps: collecting a multi-channel acoustic signal, calculating a cross-correlation time delay and a phase difference, and determining an initial position coordinate; using an adaptive quantum annealing algorithm for optimization, adjusting a quantum fluctuation intensity in real time; converting the optimized position into a gimbal control parameter; and when a tracking deviation exceeds a threshold value, optimizing again to realize continuous and accurate tracking of the gimbal. The application improves the sound source positioning accuracy and tracking stability, and reduces the calculation complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to a sound source tracking technology, in particular to a gimbal control sound source tracking method and system based on a nested MEMS array. BACKGROUND

[0002] Sound source positioning and tracking technology has wide application prospects in the fields of military reconnaissance, security monitoring, intelligent conference systems, etc. Traditional sound source positioning and tracking systems usually use microphone arrays to collect sound signals and determine the sound source position through a series of signal processing algorithms, and then control the gimbal or other mechanical structures to realize tracking of the sound source. With the rapid development of micro-electro-mechanical system (MEMS) technology, MEMS microphones have gradually become the mainstream acoustic sensors in sound source positioning systems due to their small size, low power consumption, and good consistency. Nested MEMS arrays can effectively overcome the spatial resolution limitation of traditional arrays in sound source positioning and improve positioning accuracy through special geometric layout.

[0003] However, the existing sound source tracking technology based on microphone arrays still has some obvious deficiencies. First, traditional sound source positioning algorithms such as time delay estimation and beamforming have insufficient accuracy in complex acoustic environments, especially in scenes with reverberation and noise interference, resulting in large positioning errors and unstable gimbal tracking. Second, existing optimization algorithms such as gradient descent and genetic algorithms are prone to local optimal solutions when solving the sound source position, and have high computational complexity, making it difficult to meet the real-time tracking requirements. In addition, the coordination between the existing gimbal control system and the sound source positioning system is insufficient, and it is difficult to adjust the control strategy in time according to the dynamic changes of the sound source during tracking, which may lead to tracking loss when the sound source moves quickly.

[0004] To solve the above problems, there is an urgent need for a sound source tracking method that can improve sound source positioning accuracy, optimize tracking algorithm efficiency, enhance system real-time performance and robustness, to realize continuous and accurate tracking of the gimbal to the sound source. SUMMARY

[0005] The embodiments of the present application provide a gimbal control sound source tracking method and system based on a nested MEMS array, which can solve the problems in the prior art.

[0006] In a first aspect, the embodiments of the present application provide a gimbal control sound source tracking method based on a nested MEMS array, comprising:

[0007] Collecting multi-channel acoustic signals of the nested MEMS array, calculating the cross-correlation time delay information and phase difference information between adjacent array elements, and calculating the initial spatial position coordinates of the sound source according to the time delay information and the phase difference information in combination with the geometric structure of the array;

[0008] construct an optimization objective function based on the initial spatial position coordinates, perform first round of iterative optimization on the optimization objective function by using an adaptive quantum annealing algorithm, and adaptively adjust quantum fluctuation strength according to the convergence speed of each round of iteration in the optimization process to obtain the accurate spatial position of the sound source after the first round of optimization;

[0009] convert the accurate spatial position of the sound source after the first round of optimization into gimbal motion control parameters, drive the gimbal to rotate according to the gimbal motion control parameters, and simultaneously collect gimbal attitude data and current sound source position data in real time;

[0010] calculate the tracking deviation between the pointing position corresponding to the gimbal attitude data and the current sound source position; when the tracking deviation exceeds a preset deviation threshold, perform second round of iterative optimization on the current sound source position data by using the adaptive quantum annealing algorithm, wherein the quantum fluctuation strength is adaptively adjusted by inheriting the optimal parameter in the first round of optimization to obtain the corrected accurate spatial position of the sound source;

[0011] update the gimbal motion control parameters based on the corrected accurate spatial position of the sound source to realize continuous and accurate tracking of the gimbal.

[0012] calculate the cross-correlation time delay information and phase difference information between adjacent elements, and calculate the initial spatial position coordinates of the sound source according to the time delay information and the phase difference information in combination with the geometric structure of the array, including:

[0013] collect acoustic signals of adjacent elements in the MEMS array, perform frequency domain decomposition on the acoustic signals to obtain sub-signals in multiple frequency bands, perform signal-to-noise ratio analysis and coherence analysis on the sub-signals, and construct a feature vector based on the results of the signal-to-noise ratio analysis and the coherence analysis;

[0014] perform optimization processing on the feature vector by using a quasi-Newton method to obtain a weight coefficient, apply the weight coefficient to the sub-signals to obtain a weighted fusion signal, and perform time delay parameter calculation on the weighted fusion signal to obtain an initial time delay estimation value, and construct a search interval around the initial time delay estimation value;

[0015] construct an optimization objective function based on the time delay parameters in the search interval and the feature vector, perform iterative solution on the optimization objective function by using the quasi-Newton method to obtain accurate time delay information, and calculate the initial phase difference between adjacent elements according to the accurate time delay information;

[0016] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0017] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0018] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0019] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0020] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0021] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0022] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0023] constructing a phase difference correction function based on the feature vector, performing optimization calculation on the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, and substituting the geometric structure parameters of the MEMS array into the sound source position solving equation set to obtain the spatial position coordinates of the sound source.

[0024] An adaptive annealing strategy is used to solve the sound source position solving equation set, and the constraint conditions are optimized in parallel in the multi-dimensional solving space. The adaptive annealing strategy dynamically adjusts the system energy parameters according to the convergence speed of the solution, and adaptively updates the control parameters based on the gradient information.

[0025] The intermediate solution in the optimization process is probabilistically evaluated. The probability evaluation calculates the credibility of each candidate solution based on an error function, and selects the candidate solution with the highest credibility as the final spatial position coordinates.

[0026] Based on the initial spatial position coordinates, an optimization objective function is constructed, and a first round of iterative optimization of the optimization objective function is performed using an adaptive quantum annealing algorithm. In the optimization process, the quantum fluctuation strength is adaptively adjusted according to the convergence speed of each iteration, and the accurate spatial position of the sound source after the first round of optimization includes:

[0027] A target function evaluation module is constructed based on a deep neural network, and multi-scale decomposition is performed based on the output of the target function evaluation module to generate coarse-scale coordinates and fine-scale coordinates.

[0028] The environmental temperature parameters and humidity parameters are collected, and an environmental influence model is established based on the environmental temperature parameters and humidity parameters. The output result of the environmental influence model is taken as an environmental constraint term, and the environmental constraint term and the coarse-scale coordinates and fine-scale coordinates are used to construct an optimization objective function.

[0029] An adaptive annealing algorithm is used to iteratively optimize the optimization objective function. In each iteration process, the change amount and annealing temperature of the objective function are calculated, the state jump probability is determined according to the change amount and the annealing temperature, and the coarse-scale coordinates and fine-scale coordinates are updated based on the state jump probability.

[0030] The convergence speed of each iteration is input into the target function evaluation module for prediction, the weight coefficients of each term in the optimization objective function are dynamically adjusted based on the prediction result of the target function evaluation module, and the coarse-scale coordinates and fine-scale coordinates are optimized according to the weight coefficients. The environmental influence model is updated in real time, and the updated environmental influence model is re-substituted into the optimization objective function. Based on the updated optimization objective function, iterative optimization is continued;

[0031] When the coordinate difference of adjacent two iterations is less than a preset convergence threshold, the accurate spatial position of the sound source is calculated.

[0032] The adaptive quantum annealing algorithm is used to perform a second round of iterative optimization on the current position data of the sound source, wherein the quantum fluctuation strength is adaptively adjusted with the optimal parameter of the first round of optimization, and the corrected accurate spatial position of the sound source includes:

[0033] obtaining an optimal quantum fluctuation parameter set obtained through the first round of optimization, constructing a plurality of specialized agents based on the optimal quantum fluctuation parameter set, each of the specialized agents being responsible for the optimization search of a corresponding dimension parameter; constructing a dynamic mutation network based on the optimal quantum fluctuation parameter set and the search space of the specialized agents, and performing nonlinear mapping on the current parameter state based on the dynamic mutation network;

[0034] mapping the optimization parameters of the specialized agents to a quantum state space, constructing a quantum expression containing a parameter quantum state and a probability amplitude, constructing a game matrix based on the optimization results of the specialized agents, calculating the optimization benefits between adjacent agents through the game matrix, and dynamically adjusting the optimization strategy of the specialized agents according to the optimization benefits;

[0035] inputting the optimization strategy into an adaptive adjustment module, calculating the adjustment amount of quantum fluctuation intensity according to the agent weight and optimization gain in the adaptive adjustment module, and updating the quantum fluctuation parameter based on the adjustment amount;

[0036] substituting the updated quantum fluctuation parameter into the dynamic mutation network to generate a new parameter state, inputting the new parameter state into the quantum state space for optimization, and calculating the corrected accurate spatial position of the sound source based on the current optimization parameter and the quantum fluctuation parameter when the parameter state difference between adjacent two iterations is less than a preset difference threshold.

[0037] constructing a dynamic mutation network based on the optimal quantum fluctuation parameter set and the search space of the specialized agents, and performing nonlinear mapping on the current parameter state based on the dynamic mutation network includes:

[0038] constructing a dynamic mutation network containing a mutation weight matrix and a bias vector based on the optimal quantum fluctuation parameter set and the search space of the specialized agents; performing weighted summation on the current parameter state by using the mutation weight matrix and the activation function of the dynamic mutation network to obtain a preliminary mapping result;

[0039] inputting the preliminary mapping result into the activation function for nonlinear mapping to obtain the nonlinear transformation result of the parameter.

[0040] In a second aspect of the embodiment of the application, a gimbal control sound source tracking system based on a nested MEMS array is provided, which comprises:

[0041] The first unit is configured to collect multi-channel acoustic signals of the nested MEMS array, calculate the cross-correlation time delay information and phase difference information between adjacent array elements, and calculate the initial spatial position coordinates of the sound source according to the time delay information and the phase difference information in combination with the geometric structure of the array.

[0042] The second unit is configured to construct an optimization objective function based on the initial spatial position coordinates, perform first round iteration optimization on the optimization objective function by using an adaptive quantum annealing algorithm, and obtain the accurate spatial position of the sound source after the first round optimization by adaptively adjusting the quantum fluctuation strength according to the convergence speed of each round of iteration during the optimization process.

[0043] The third unit is configured to convert the accurate spatial position of the sound source after the first round optimization into gimbal motion control parameters, drive the gimbal to rotate according to the gimbal motion control parameters, and simultaneously collect gimbal attitude data and current sound source position data in real time.

[0044] The fourth unit is configured to calculate the tracking deviation between the pointing position corresponding to the gimbal attitude data and the current sound source position, and perform second round iteration optimization on the current sound source position data by using the adaptive quantum annealing algorithm when the tracking deviation exceeds a preset deviation threshold, wherein the quantum fluctuation strength is adaptively adjusted by inheriting the optimal parameter in the first round optimization to obtain the corrected accurate spatial position of the sound source.

[0045] The fifth unit is configured to update the gimbal motion control parameters based on the corrected accurate spatial position of the sound source, and realize continuous and accurate tracking of the gimbal.

[0046] In a third aspect, an electronic device is provided, including:

[0047] a processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0050] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0051] The present application has the following advantages:

[0052] By using the nested MEMS array to collect multi-channel acoustic signals, combining the cross-correlation time delay information and the phase difference information to calculate the initial position of the sound source, and then using the adaptive quantum annealing algorithm for optimization, the sound source can be quickly and accurately positioned in a complex acoustic environment, and the accuracy and robustness of the sound source positioning are improved.

[0053] Adopting the adaptive quantum annealing algorithm to iteratively optimize the objective function, and adaptively adjusting the quantum fluctuation strength according to the convergence speed in the optimization process, can effectively avoid the local optimal solution trap, accelerate the convergence speed of the algorithm, reduce the consumption of computing resources, and improve the real-time performance of the system.

[0054] By collecting the gimbal attitude data and the current position data of the sound source in real time, calculating the tracking deviation and performing secondary optimization, the system can continuously track the motion state of the sound source, adapt to the dynamic change of the position of the sound source, realize continuous and accurate tracking of the gimbal, and greatly improve the stability and adaptability of the system in practical application. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A flowchart of a gimbal control sound source tracking method based on a nested MEMS array according to an embodiment of the application is shown in

[0056] Figure 2 A comparison diagram of the accuracy of different MEMS array sound source positioning methods in various scenes according to an embodiment of the application is shown in

[0057] Figure 3 A comparison diagram of the sound source positioning accuracy of different methods according to an embodiment of the application is shown in

[0058] Figure 4 A sound source positioning optimization flowchart based on an adaptive quantum annealing algorithm according to an embodiment of the application is shown in

[0059] Figure 5 A performance comparison histogram of a dynamic variation network in different test scenes according to an embodiment of the application is shown in DETAILED DESCRIPTION

[0060] To make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0061] The technical scheme of the application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0062] Figure 1 A flowchart of a gimbal control sound source tracking method based on a nested MEMS array according to an embodiment of the application is shown in Figure 1 As shown in the figure, the method comprises:

[0063] collecting multi-channel acoustic signals of the nested MEMS array, calculating cross-correlation time delay information and phase difference information between adjacent array elements, calculating initial spatial position coordinates of the sound source according to the time delay information and the phase difference information, and combining the geometric structure of the array;

[0064] constructing an optimization objective function based on the initial spatial position coordinates, performing first round of iterative optimization on the optimization objective function by using an adaptive quantum annealing algorithm, adaptively adjusting quantum fluctuation strength according to the convergence speed of each round of iteration in the optimization process, and obtaining the accurate spatial position of the sound source after the first round of optimization;

[0065] converting the accurate spatial position of the sound source after the first round of optimization into gimbal motion control parameters, driving the gimbal to rotate according to the gimbal motion control parameters, and collecting gimbal attitude data and current sound source position data in real time;

[0066] calculating the tracking deviation between the pointing position corresponding to the gimbal attitude data and the current sound source position; when the tracking deviation exceeds a preset deviation threshold, performing second round of iterative optimization on the current sound source position data by using the adaptive quantum annealing algorithm, adaptively adjusting the quantum fluctuation strength with the optimal parameter in the first round of optimization, and obtaining the corrected accurate spatial position of the sound source;

[0067] updating the gimbal motion control parameters based on the corrected accurate spatial position of the sound source, and realizing continuous and accurate tracking of the gimbal.

[0068] In an optional embodiment, calculating cross-correlation time delay information and phase difference information between adjacent array elements, and calculating initial spatial position coordinates of the sound source according to the time delay information and the phase difference information, comprises:

[0069] collecting acoustic signals of adjacent array elements in the MEMS array, performing frequency domain decomposition on the acoustic signals to obtain sub-signals of multiple frequency bands, performing signal-to-noise ratio analysis and coherence analysis on the sub-signals, and constructing a feature vector based on the results of the signal-to-noise ratio analysis and the coherence analysis;

[0070] performing optimization processing on the feature vector by using a quasi-Newton method to obtain a weight coefficient, applying the weight coefficient to the sub-signals to obtain a weighted fusion signal, and performing time delay parameter calculation on the weighted fusion signal to obtain an initial time delay estimation value, and constructing a search interval around the initial time delay estimation value;

[0071] constructing an optimization objective function with the time delay parameters in the search interval and the feature vector, performing iterative solution on the optimization objective function by using the quasi-Newton method, obtaining accurate time delay information, and calculating the initial phase difference between adjacent array elements according to the accurate time delay information;

[0072] Based on the feature vector, a phase difference correction function is constructed, and the quasi-Newton method is used to optimize the calculation of the phase difference correction function to obtain corrected phase difference information; and based on the corrected phase difference information and the accurate time delay information, a sound source position solving equation set is established, the geometric structure parameters of the MEMS array are substituted into the sound source position solving equation set, and the spatial position coordinates of the sound source are solved.

[0073] In a specific embodiment, the method for calculating the cross-correlation time delay information and the phase difference information between adjacent elements, and calculating the initial spatial position coordinates of the sound source based on the time delay information and the phase difference information in combination with the geometric structure of the array can be implemented according to the following steps.

[0074] In the MEMS array, it is assumed that the array contains 8 microphone elements arranged in a 2x4 rectangular structure, and the distance between adjacent microphones is 10 cm. The acoustic signals collected by the microphone elements are digitized by an analog-to-digital converter, and the sampling rate is set to 16 kHz and the bit depth is set to 16 bits. The collected acoustic signals are first preprocessed, including removing direct current, pre-emphasis and windowing. The processed signals are subjected to short-time Fourier transform (STFT) for frequency domain decomposition, and the original signals are decomposed into 16 sub-signals of frequency bands, each with a bandwidth of 500 Hz.

[0075] The signal-to-noise ratio of each frequency band sub-signal is analyzed, and the ratio of signal power to noise power is calculated to evaluate the signal quality. For example, in a certain experiment, the signal-to-noise ratio of the low frequency band (0-500 Hz) is 15 dB, the signal-to-noise ratio of the middle frequency band (1000-1500 Hz) is 12 dB, and the signal-to-noise ratio of the high frequency band (4000-4500 Hz) is 8 dB. At the same time, the coherence of the sub-signals between adjacent elements is analyzed, and the normalized cross-correlation coefficient between the signals is calculated as a coherence indicator. For example, in the above frequency bands, the coherence between adjacent elements is 0.85, 0.78 and 0.62, respectively. Based on the results of the signal-to-noise ratio analysis and the coherence analysis, a feature vector is constructed. For each pair of adjacent elements, the feature vector contains the signal-to-noise ratio values and the coherence values of each frequency band, forming a 32-dimensional vector (16 frequency bands x 2 features).

[0076] The BFGS algorithm in the quasi-Newton method is used to optimize the feature vector to obtain the weight coefficients of each frequency band. In the optimization process, the initial weights are evenly distributed, the number of iterations is set to 50, and the convergence threshold is set to 10 -6 . For example, after optimization, the weight of the low frequency band is 0.25, the weight of the middle frequency band is 0.45, and the weight of the high frequency band is 0.30. These weight coefficients are applied to the sub-signals of each frequency band to obtain a weighted fusion signal by weighted summation.

[0077] The initial time delay estimation value between adjacent elements is calculated by cross-correlation analysis of the weighted fused signal. For example, in a certain experiment, the initial time delay estimation value between the two adjacent elements is 0.15 milliseconds. A search interval is constructed around this estimation value, with the interval range being [0.10 milliseconds, 0.20 milliseconds], and the search step is set to 0.001 milliseconds.

[0078] In the constructed search interval, the time delay parameter and the feature vector group are combined to construct an optimization objective function. The design of this objective function is based on the principle of maximizing the weighted cross-correlation coefficient, while considering the signal-to-noise ratio and coherence characteristics. The quasi-Newton method is used to iteratively solve this optimization objective function, with the iteration starting point set to the initial time delay estimation value and the upper limit of the iteration number set to 100 times. Through iterative optimization, the final accurate time delay information is 0.147 milliseconds.

[0079] According to the accurate time delay information and the phase information of each frequency band, the initial phase difference between adjacent elements is calculated. For example, in the 1kHz frequency band, the initial phase difference between adjacent elements is 105 degrees. Due to the influence of environmental factors and array position errors, the initial phase difference needs to be corrected.

[0080] Based on the feature vector constructed in the previous step, a phase difference correction function is constructed. This function takes into account the influence of signal-to-noise ratio and coherence of different frequency bands on the accuracy of phase estimation, and performs weighted correction on the initial phase difference. The quasi-Newton method is used to optimize the calculation of the phase difference correction function, with the iteration number set to 30 times and the convergence threshold set to 10 -5 . For example, after correction, the phase difference between adjacent elements in the 1kHz frequency band is corrected to 108.3 degrees.

[0081] According to the corrected phase difference information and the accurate time delay information, a sound source position solving equation set is established. This equation set is based on the geometric relationship between sound wave propagation time delay and element position and sound source position, as well as the relationship between phase difference and sound source direction. The geometric structure parameters of the MEMS array are substituted into the equation set, including the three-dimensional coordinates of each element and the element spacing. The equation set is solved by the least squares method to obtain the spatial position coordinates of the sound source.

[0082] In experimental verification, a sound source with a known position is placed 3 meters away from the center of the array, with an azimuth angle of 45 degrees and a pitch angle of 30 degrees. The sound source position coordinates calculated by this method are (2.12 meters, 2.12 meters, 1.5 meters), with an error of less than 5 centimeters from the actual position, and the positioning accuracy meets the expected requirements.

[0083] Figure 2 The accuracy comparison diagram of different MEMS array sound source positioning methods in various scenarios is shown in the following figure:

[0084] This figure illustrates the performance of three localization algorithms in four typical environments. The white bars represent the traditional cross-correlation method, a basic sound source localization algorithm; the bars filled with diagonal lines represent the weighted fusion method, a comprehensive localization method; and the bars filled with a grid represent the proposed solution, namely the adaptive quantum annealing optimization algorithm based on nested MEMS arrays. In low signal-to-noise ratio environments, the localization accuracies of the three methods are 68.0%, 77.0%, and 89.0%, respectively; in multi-source interference environments, the performance data are 60.0%, 69.0%, and 86.0%, respectively; in high-reverberation indoor environments, the accuracies reach 55.0%, 75.0%, and 93.0%, respectively; and in long-distance open environments, the performance of the three methods is 74.0%, 85.0%, and 94.0%, respectively. The data comparison shows that the proposed solution maintains a significant performance advantage in all test environments, especially in challenging high-reverberation indoor environments, where its accuracy is improved by 38 percentage points compared to the traditional cross-correlation method and by 18 percentage points compared to the weighted fusion method. This fully demonstrates that the adaptive quantum annealing optimization algorithm based on nested MEMS arrays has stronger robustness and adaptability in complex acoustic environments.

[0085] In one optional implementation, a phase difference correction function is constructed based on the feature vector, and the phase difference correction function is optimized using the quasi-Newton method to obtain the corrected phase difference information, including:

[0086] A phase difference correction model is established based on the feature vectors. The phase difference correction model is used as the optimization objective, and an objective function containing a regularization term is constructed. The objective function is optimized and solved using the quasi-Newton method. The optimization parameters are adaptively adjusted according to the signal characteristics in each iteration.

[0087] Error analysis is performed on the iterative results of the quasi-Newton method to calculate the relative error between two adjacent iterations. When the relative error is less than a preset error threshold, the phase difference obtained in the current iteration is output as the corrected phase difference information.

[0088] The system receives phase difference information and eigenvectors as input data. The eigenvectors typically contain multi-dimensional features such as the signal's frequency distribution, amplitude variation, and phase shift. To construct an effective phase difference correction model, the system maps the eigenvectors to a phase difference correction space and uses a polynomial model to express the relationship between the phase difference and the eigenvectors. The coefficients of this polynomial model are initialized using the least squares method to establish an initial phase difference correction model. For example, for a three-dimensional eigenvector containing frequency, amplitude, and time features, the initial model represents the phase difference as a second-order polynomial combination of these features.

[0089] When constructing the objective function, the system not only considers the fitting error between the model and the actual phase difference, but also introduces a regularization term to prevent overfitting. The objective function contains two parts: the main term is the mean square error between the predicted phase difference and the actual phase difference; the regularization term controls the complexity of the model, using the L2 norm to punish large model parameters. The objective function can be expressed in the form of the weighted sum of the square sum of the prediction error and the regularization term, where the regularization coefficient is initially set to 0.01 and then adjusted adaptively during the optimization process.

[0090] The quasi-Newton method optimization process starts from the initial point and iteratively calculates. The initial point is set based on historical data statistics or domain knowledge, for example, it can be set to a zero vector or an estimated value based on prior knowledge. In each iteration, the system calculates the gradient vector of the objective function and the approximate inverse matrix of the Hessian matrix, and updates the parameter values accordingly. Unlike the traditional Newton method, the quasi-Newton method avoids the complexity of directly calculating and inverting the Hessian matrix, but instead constructs an approximation of the Hessian matrix using current gradient information and historical iteration information. In specific implementation, the BFGS algorithm is used to update the approximation of the Hessian matrix, which can effectively capture the local curvature information of the objective function.

[0091] During the iterative optimization process, the system adaptively adjusts the step size and regularization parameter according to the signal characteristics. For signal regions with strong intensity changes, the system reduces the step size to ensure stability; while in the signal smooth region, the step size is appropriately increased to accelerate convergence. The regularization parameter is also dynamically adjusted according to the fitting error of the current iteration, when the fitting error is large, the regularization strength is reduced to allow the model to better fit the data; otherwise, the regularization strength is increased to prevent overfitting. For example, when detecting that the signal frequency changes more than 20%, the system reduces the step size coefficient from the standard value of 0.1 to 0.05 to adapt to the nonlinear changes of the signal.

[0092] To ensure the numerical stability of the algorithm, the system checks the size of the parameter update after each iteration. If an abnormal parameter change is detected (such as exceeding a pre-set threshold), the step size is automatically reduced and recalculated. At the same time, the system detects whether the objective function value is continuously decreasing, if the objective function value increases for three consecutive iterations, it is considered that the current search direction is not good, the Hessian matrix approximation is reset and the search direction is adjusted.

[0093] In the error analysis section, the system calculates the relative error of the results of the last two iterations, defined as the L2 norm of the difference between the current iteration result and the last iteration result divided by the L2 norm of the last iteration result. The pre-set relative error threshold is 0.001, when the calculated relative error is less than the threshold, it is considered that the iteration has converged, and the phase difference obtained in the current iteration is output as the final correction result. If the number of iterations reaches the maximum limit (default setting is 100 times) but still does not converge, the system selects the iteration result with the smallest objective function value as the output and records a warning message.

[0094] In a practical application case, the original phase difference data received in a certain phased array radar system has system errors. The eigenvector contains three dimensions of signal frequency (75-85 Hz), received power (-60 to -40 decibels), and temperature change (15-35 degrees Celsius). The initial phase difference error is about 15 degrees. After applying this method, the system starts from the initial all-zero parameter vector, and after 27 iterations, it reaches the convergence condition (relative error 0.0008 is less than threshold 0.001), and the final phase difference correction model reduces the original phase difference error to 2.3 degrees, significantly improving the system positioning accuracy. The entire calculation process takes about 150 milliseconds, meeting the real-time processing requirements.

[0095] The advantage of this method is that it can adaptively adjust the optimization parameters according to the signal characteristics, suitable for different working environments and signal characteristics, and has strong robustness and universality. By introducing regularization and dynamic parameter adjustment mechanism, the problem of insufficient accuracy of traditional phase difference correction method in complex environment is effectively solved.

[0096] In an optional implementation, a sound source position solving equation set is established according to the corrected phase difference information and the accurate time delay information, the geometric structure parameters of the MEMS array are substituted into the sound source position solving equation set, and the spatial position coordinates of the sound source are solved to include:

[0097] Based on the corrected phase difference information and the accurate time delay information, a sound source position solving equation set is established, the sound source position solving equation set is converted to a multi-dimensional solving space, a multi-dimensional expression containing sound source spatial coordinate components is constructed, and the multi-dimensional expression converts the solving equation set of the Euclidean space to a multi-dimensional state representation through a coordinate mapping function;

[0098] A design information encoding strategy is designed to encode the corrected phase difference information and the accurate time delay information into a combined state expression. The combined state expression uses a weighted fusion method to maintain the coupling relationship between information, and the weighted fusion method dynamically adjusts the weight coefficient based on the signal quality index;

[0099] The geometric structure parameters of the MEMS array are obtained, and constraint conditions containing array element spacing constraints and array deformation constraints are established based on the geometric structure parameters. The constraint conditions limit the feasible region of the solution through a penalty function;

[0100] An adaptive annealing strategy is used to solve the sound source position solving equation set, and the constraint conditions are optimized in parallel in the multi-dimensional solving space. The adaptive annealing strategy dynamically adjusts the system energy parameters according to the convergence speed of the solution, and adaptively updates the control parameters based on the gradient information;

[0101] A probability evaluation is performed on the intermediate solutions in the optimization process, the probability evaluation calculates the credibility of each candidate solution based on an error function, and the candidate solution with the highest credibility is selected as the final spatial position coordinate.

[0102] In a sound source positioning method based on corrected phase difference information and accurate time delay information, the specific embodiment includes the complete technical implementation process of establishing a sound source position solving equation set and solving the spatial position coordinates of the sound source.

[0103] The establishment of the sound source position solving equation set first requires mathematical expression of the corrected phase difference information and accurate time delay information. Assuming that the MEMS array is composed of 8 microphones, the collected sound signal phase difference has been corrected through previous processing, and the corrected phase difference between the i-th microphone and the reference microphone is denoted as φ i , and the corresponding accurate time delay is denoted as τ i . These information constitute the basic data of the sound source position solving equation set. When converting the sound source position solving equation set to a multi-dimensional solving space, a three-dimensional expression containing the sound source spatial coordinate components (x, y, z) is constructed. In the specific implementation, the right-angle coordinate system equation in the Euclidean space is converted into a state space representation through a nonlinear mapping function, which expresses the relationship between the sound source coordinates and the phase difference, time delay as a combination of state variables. For example, for a certain microphone pair, the state variable can be expressed as a function relationship between the sound source position vector and the microphone position vector, and through this mapping conversion, the complex nonlinear equation is easier to solve.

[0104] In the design of information coding strategy, the corrected phase difference information and accurate time delay information are encoded into a combined state expression. In actual application, for a sound source signal of a certain frequency f = 1000 Hz, it is assumed that the corrected phase difference value is [0, 0.35π, 0.42π, 0.56π, 0.71π, 0.83π, 0.92π] radian, and the corresponding accurate time delay is [0, 0.175, 0.21, 0.28, 0.355, 0.415, 0.46] milliseconds. The combined state expression can be expressed as S i = α · φ i + β · τ i , where α and β are weight coefficients. The weight coefficient dynamic adjustment mechanism is based on signal quality indicators such as signal-to-noise ratio and phase consistency. When the signal-to-noise ratio of a certain microphone is lower than 15 dB, the corresponding weight coefficient is reduced; when the phase consistency is higher than 0.85, the weight of the phase difference information is increased. In a specific scenario, the value of α can be set to 0.7 and the value of β can be set to 0.3. With the change of signal quality, α can be adjusted in the range of 0.5-0.8, and β can be adjusted in the range of 0.2-0.5.

[0105] The geometric parameters of the MEMS array involve the physical layout information of the array. In one practical application case, 8 MEMS microphones form a three-dimensional array, and the microphone coordinates are (0, 0, 0), (0.05, 0, 0), (0, 0.05, 0), (0.05, 0.05, 0), (0, 0, 0.05), (0.05, 0, 0.05), (0, 0.05, 0.05), and (0.05, 0.05, 0.05) meters, respectively. Based on these geometric parameters, constraint conditions are established, including array element spacing constraints and array deformation constraints. The array element spacing constraint ensures that the microphone spacing used in the calculation is consistent with the actual physical spacing, with a maximum allowable error of not more than 0.002 meters; the array deformation constraint considers the physical size change of the array caused by temperature change, and sets the maximum deformation coefficient to 0.001 / °C. These constraint conditions are implemented through a penalty function, which increases the value of the objective function when the solution parameters violate the constraints, guiding the optimization algorithm to search within the feasible region.

[0106] When solving the sound source position calculation equation set using the adaptive annealing strategy, the constraint conditions are optimized in parallel in the multi-dimensional solution space. The initial temperature of the annealing strategy is set to 100, the termination temperature is 0.01, and the acceptance probability at the initial temperature is 0.95. The system energy parameters are dynamically adjusted according to the convergence speed of the solution, and when the convergence speed of 10 consecutive iterations is lower than the threshold value 0.001, the system energy decay coefficient is adjusted from 0.95 to 0.9 to accelerate the convergence process. Based on the gradient information, the control parameters are adaptively updated. In each iteration, the gradient of the error function is calculated, and when the gradient size exceeds the threshold value 0.1, the search step is reduced to 0.8 times the original; when the gradient is less than the threshold value 0.01, the search step is increased to 1.2 times the original, with a maximum step of not more than 0.05 and a minimum step of not less than 0.001.

[0107] When evaluating the intermediate solutions in the optimization process, the reliability of each candidate solution is calculated based on the error function. The error function is defined as the Euclidean distance between the actual observed phase difference and the theoretical phase difference. For example, for the candidate solution (2.35, 1.42, 0.87) meters, the error between the theoretical phase difference and the actual observed phase difference is 0.032 radians, and the corresponding reliability is 98.4%; while the error of the candidate solution (2.33, 1.45, 0.82) meters is 0.048 radians, and the reliability is 97.6%. The candidate solution with the highest reliability is selected as the final result. In a complete sound source positioning process, the system initially generates 100 candidate solutions randomly, and after 500 iterations, the final sound source position coordinates are determined as (2.35, 1.42, 0.87) meters, with a positioning accuracy of ±0.02 meters, meeting the requirements of practical applications.

[0108] Figure 3A sound source positioning precision comparison diagram of different methods of embodiments of the present application is shown in the figure:

[0109] The figure shows the positioning error comparison of three different sound source positioning methods in the signal-to-noise ratio range from -5dB to 25dB. Among them, the circular mark represents the present technical solution, the square mark represents the traditional TDOA method, and the diamond mark represents the basic phase difference method. Under the condition of low signal-to-noise ratio -5dB, the positioning error of the present technical solution is 2.0m, which is significantly better than the 3.6m of the traditional TDOA method and the 4.8m of the basic phase difference method. With the improvement of signal-to-noise ratio, the positioning errors of the three methods all show a downward trend. When the signal-to-noise ratio is 0dB, the positioning errors of the three methods are 3.0m for the present technical solution, 4.5m for the traditional TDOA method, and 5.2m for the basic phase difference method. When the signal-to-noise ratio reaches 10dB, the positioning error of the present technical solution is reduced to 2.9m, while the traditional TDOA method and the basic phase difference method are 4.2m and 4.9m respectively. Under the condition of high signal-to-noise ratio 25dB, the positioning error of the present technical solution is finally reduced to 0.5m, while the traditional TDOA method and the basic phase difference method are 2.1m and 3.2m respectively. From the overall trend, the present technical solution maintains a significant performance advantage in the whole signal-to-noise ratio range, especially in the high signal-to-noise ratio condition, the positioning precision is improved more obviously, which fully proves the advancement of the scheme in the field of sound source positioning.

[0110] In an optional implementation, an optimization target function is constructed based on the initial spatial position coordinates, a first round of iterative optimization is performed on the optimization target function by using an adaptive quantum annealing algorithm, the quantum fluctuation strength is adaptively adjusted according to the convergence speed of each round of iteration in the optimization process, and a precise spatial position of the sound source after the first round of optimization includes:

[0111] A target function evaluation module is constructed based on a deep neural network, multi-scale decomposition is performed based on the output of the target function evaluation module, and coarse scale coordinates and fine scale coordinates are generated;

[0112] An environmental temperature parameter and a humidity parameter are collected, an environmental influence model is established based on the environmental temperature parameter and the humidity parameter, an output result of the environmental influence model is taken as an environmental constraint term, and the environmental constraint term and the coarse scale coordinates and the fine scale coordinates are used to jointly construct an optimization target function;

[0113] An iterative optimization is performed on the optimization target function by using an adaptive annealing algorithm, a change amount and an annealing temperature of the target function are calculated in each iteration process, a state jump probability is determined according to the change amount and the annealing temperature, and the coarse scale coordinates and the fine scale coordinates are updated based on the state jump probability;

[0114] The convergence speed of each round of iteration is input into the target function evaluation module for prediction, the weight coefficients of each term in the optimization target function are dynamically adjusted based on the prediction result of the target function evaluation module, and the coarse scale coordinates and the fine scale coordinates are optimized according to the weight coefficients; the environmental influence model is updated in real time, the updated environmental influence model is substituted into the optimization target function again, and iteration optimization is continued based on the updated optimization target function;

[0115] When the coordinate difference values of two adjacent rounds of iteration are both less than a preset convergence threshold, the accurate spatial position of the sound source is calculated.

[0116] As shown in Figure 4 The method further comprises:

[0117] After obtaining the initial spatial position coordinates of the sound source, an optimization target function is constructed for accurate positioning of the sound source. The target function includes multiple parts: a distance error term, an angle error term, and an environmental constraint term, each of which has a corresponding weight coefficient. For the distance error term, the mean square error between the theoretical distance calculated using the time difference of arrival of the sound signals received by the three microphone arrays and the actual measured distance is used as the error value; for the angle error term, the deviation of the direction vector from the sound source to the microphone array from the actual measured angle is calculated; and for the environmental constraint term, the environmental influence model is used.

[0118] In the first round of iteration optimization, the initial quantum fluctuation strength is set to 0.85, the annealing starting temperature is set to 100, and the cooling coefficient is set to 0.95. Iteration calculation is performed, and in each iteration, the quantum bits jump in the solution space through quantum fluctuation, and the target function value of the new position is calculated. If the target function value of the new position is less than the current position, the new position is accepted; if it is greater than the current position, the new position is accepted with a certain probability according to the Metropolis criterion. After 50 iterations, the convergence speed is calculated, and if the target function changes by less than 0.001 for 5 consecutive iterations, the quantum fluctuation strength is reduced to 0.8 times the original value; if the change is greater than 0.01, the quantum fluctuation strength is increased to 1.2 times the original value, but not more than 1.0. Through this way, the sound source position coordinates after the first round of optimization are obtained.

[0119] The target function evaluation module constructed by the deep neural network consists of an input layer, three hidden layers, and an output layer. The input layer receives the current coordinates, the target function value, and the convergence speed; the hidden layers have 64, 128, and 64 neurons respectively, and use the ReLU activation function; the output layer predicts the target function change trend of the next round of iteration. The network is trained using historical optimization data, the loss function is mean square error, the optimizer is Adam, and the learning rate is set to 0.001.

[0120] When performing multi-scale decomposition, the three-dimensional coordinates are split into coarse-scale and fine-scale parts. The coarse-scale coordinates represent the overall position with a precision of meters, while the fine-scale coordinates represent the small deviations with a precision of centimeters. For example, the coordinates (10.257, 15.432, 3.128) are decomposed into coarse-scale (10, 15, 3) and fine-scale (0.257, 0.432, 0.128).

[0121] The environmental impact model is constructed based on temperature and humidity parameters. Data is collected using a wireless sensor network in the temperature range of 20-35°C and humidity range of 30-80%. The model takes into account the effect of temperature and humidity on sound speed, with an increase of about 0.6 meters / second for every 1°C increase in temperature and an increase of about 0.1 meters / second for every 10% increase in humidity. At standard conditions (25°C, 50%), the sound speed is 346 meters / second. The initial weight of the environmental constraint term is set to 0.3 and is dynamically adjusted according to the amplitude of environmental changes.

[0122] When performing iterative optimization using the adaptive annealing algorithm, the initial temperature is set to 50 and the cooling coefficient is 0.92. In each iteration, the coarse-scale and fine-scale coordinates are randomly perturbed, with a perturbation range of ±0.5 meters for the coarse-scale and ±0.05 meters for the fine-scale. The value of the objective function at the new coordinates is calculated, and if it is better than the current value, it is accepted; otherwise, it is accepted with a probability p = exp(-ΔE / T), where ΔE is the increase in the objective function and T is the current temperature. The temperature is reduced by a factor of 0.92 every 10 iterations.

[0123] The weight coefficients are dynamically adjusted based on the convergence speed. When the target function evaluation module predicts a decrease in convergence speed, the weight of the fine-scale coordinates is increased from the initial value of 0.3 to 0.5; when the environmental factors are predicted to have a significant impact, the weight of the environmental constraint term is increased from 0.3 to 0.4. Real-time monitoring of environmental changes is performed, and when the temperature changes by more than 2°C or the humidity changes by more than 10%, the environmental impact model parameters are recalculated and the optimization objective function is updated.

[0124] The convergence criterion uses a coordinate difference threshold method, with a preset convergence threshold of 0.01 meters for the coarse-scale and 0.001 meters for the fine-scale. When the coordinate changes of two consecutive iterations are both less than the threshold, the optimization process is terminated and the final coordinates are output as the accurate position of the sound source.

[0125] In a real test case, the original microphone array data shows that the initial position of the sound source is approximately (12.5, 8.2, 2.4) meters, the environmental temperature is 28°C, and the humidity is 65%. After 120 iterations of optimization, the accurate position of the sound source is determined to be (12.371, 8.156, 2.387) meters, with a positioning error of less than 5 centimeters, meeting the requirements of indoor high-precision sound source positioning.

[0126] In an alternative embodiment, the adaptive quantum annealing algorithm is used to perform a second round of iterative optimization on the current sound source position data, wherein the quantum fluctuation strength is adaptively adjusted based on the optimal parameters obtained in the first round of optimization to obtain a corrected sound source accurate spatial position, including:

[0127] The optimal quantum fluctuation parameter set obtained in the first round of optimization is obtained, and a plurality of specialized agents are constructed based on the optimal quantum fluctuation parameter set, each of the specialized agents being responsible for optimization search of a corresponding dimension parameter; a dynamic variation network is constructed based on the optimal quantum fluctuation parameter set and the search space of the specialized agents, and a nonlinear mapping of the current parameter state is performed based on the dynamic variation network;

[0128] The optimization parameters of the specialized agents are mapped to a quantum state space, a quantum expression including parameter quantum states and probability amplitudes is constructed, a game matrix is constructed based on the optimization results of the specialized agents, the optimization benefits between adjacent agents are calculated through the game matrix, and the optimization strategy of the specialized agents is dynamically adjusted according to the optimization benefits;

[0129] The optimization strategy is input into an adaptive adjustment module, the adaptive adjustment module calculates the adjustment amount of the quantum fluctuation strength based on the agent weight and the optimization gain, and updates the quantum fluctuation parameters based on the adjustment amount;

[0130] The updated quantum fluctuation parameters are substituted into the dynamic variation network to generate a new parameter state, and the new parameter state is input into the quantum state space for optimization; when the parameter state difference between adjacent two iterations is less than a preset difference threshold, a corrected sound source accurate spatial position is calculated based on the current optimization parameters and the quantum fluctuation parameters.

[0131] The present application relates to a method for performing a second round of iterative optimization on current sound source position data using an adaptive quantum annealing algorithm. After the first round of iterative optimization, the system obtains an optimal quantum fluctuation parameter set, which is the best parameter combination in the first round of optimization. For example, in a certain experiment, the optimal quantum fluctuation parameter set obtained after the first round of optimization is {0.027, 0.035, 0.019}, which corresponds to the quantum fluctuation strength of X, Y and Z three spatial dimensions respectively.

[0132] Based on the obtained optimal quantum fluctuation parameter set, the system constructs multiple specialized agents, each responsible for the optimization search of a specific dimension parameter. In the case of three-dimensional spatial positioning, the system constructs three specialized agents, each responsible for the optimization of X, Y, and Z spatial coordinates. Each agent has an independent search strategy and evaluation mechanism, enabling fine-tuned optimization in the dimension it is responsible for. For example, the specialized agent for the X dimension sets the initial search range to ±0.5 meters around the current estimated position, with a search step size of 0.01 meters.

[0133] At the same time, the system constructs a dynamic mutation network based on the optimal quantum fluctuation parameter set and the search space of the specialized agents. This network adopts a multi-layer structure, including an input layer, multiple hidden layers, and an output layer. The input layer receives the current parameter state, the hidden layers perform nonlinear transformation, and the output layer generates new parameter states. The weights of each node in the network are dynamically adjusted according to the optimal quantum fluctuation parameters. For example, when the quantum fluctuation parameter is 0.027, the weights of the hidden layer nodes corresponding to the dimension in the network are randomly mutated within the range of ±0.027, achieving nonlinear mapping of the parameter space.

[0134] During the parameter optimization process, the system maps the optimization parameters of the specialized agents to the quantum state space. For each dimension parameter value, the system represents it as a superposition state of quantum bits, constructing a quantum expression containing parameter quantum states and probability amplitudes. For example, a parameter value of 10.25 in a certain dimension can be represented as the |10.25> state and assigned a certain probability amplitude. In practical applications, the system can discretize the parameter value into a binary representation of a limited number of bits, such as representing 10.25 as the binary form 1010.01, and the corresponding quantum state is represented as |1>|0>|1>|0>|.>|0>|1>.

[0135] Based on the optimization results of the specialized agents, the system constructs a game matrix. This matrix records the mutual influence of parameter changes between different agents. For example, in the matrix, element a_ij represents the degree of influence of the parameter change of the i-th agent on the optimization objective of the j-th agent. By calculating these influence values, the system can obtain the optimization benefits between adjacent agents. In a certain iteration, the system obtains the following game matrix: [[0, 0.12, 0.05], [0.08, 0, 0.15], [0.03, 0.11, 0]], representing the degree of mutual influence between dimension parameters.

[0136] According to the calculated optimal income, the system dynamically adjusts the optimization strategy of the specialized agent. When the change of a certain dimension parameter contributes more to the overall optimization goal, the search step of the corresponding agent will be reduced, and the search range will be reduced to conduct a more fine local search. Conversely, if the change of a certain dimension parameter contributes less, the search step of the corresponding agent will be increased, and the search range will be expanded to find a better solution. For example, when the X dimension parameter contributes the most to the objective function, the search step of the X dimension agent is reduced from 0.01 meters to 0.005 meters, and the search range is reduced from ±0.5 meters to ±0.2 meters.

[0137] The system inputs the adjusted optimization strategy into the adaptive adjustment module, which calculates the adjustment amount of quantum fluctuation intensity based on the agent weight and optimization gain. In practical applications, the adjustment amount of quantum fluctuation intensity can be represented as the weighted sum of the agent weight and the optimization gain. For example, if the weight of the X dimension agent is 0.4 and the optimization gain is 0.05, the adjustment amount of the X dimension quantum fluctuation intensity is 0.4 x 0.05 = 0.02. The system updates the quantum fluctuation parameter based on the calculated adjustment amount to adapt to the current optimization environment.

[0138] The system inputs the updated quantum fluctuation parameter into the dynamic mutation network to generate a new parameter state. The new parameter state reflects the current optimization stage's estimate of the sound source position. For example, in a certain iteration, the updated quantum fluctuation parameter is {0.018, 0.025, 0.014}, and after being input into the dynamic mutation network, a new sound source position estimate is obtained as (12.35, 8.72, 3.45).

[0139] The system inputs the newly generated parameter state into the quantum state space for optimization, and repeats the above process until the convergence condition is met. When the difference between the parameter states of the adjacent two iterations is less than the preset difference threshold, the system considers that the optimization process has converged, and calculates the corrected sound source accurate spatial position based on the current optimization parameter and the quantum fluctuation parameter. For example, set the difference threshold to 0.001 meters, when the Euclidean distance between the sound source position estimates of the two consecutive iterations is less than 0.001 meters, the system outputs the current estimate (12.351, 8.724, 3.448) as the corrected sound source accurate spatial position.

[0140] Through the above method, the present application realizes high-precision positioning of the sound source position, and compared with the traditional method, the positioning accuracy is improved by about 30%, and the average positioning error is reduced to within 0.015 meters, which is particularly suitable for sound source positioning tasks in complex acoustic environments.

[0141] In an alternative embodiment, a dynamic mutation network is constructed based on the optimal quantum fluctuation parameter set and the search space of the specialized agent, and the nonlinear mapping of the current parameter state based on the dynamic mutation network comprises:

[0142] constructing a dynamic mutation network containing mutation weight matrix and bias vector based on the optimal quantum fluctuation parameter set and the search space of the specialized agent; performing weighted summation on the current parameter state by using the mutation weight matrix and the activation function of the dynamic mutation network to obtain a preliminary mapping result;

[0143] inputting the preliminary mapping result into the activation function to perform nonlinear mapping to obtain a nonlinear transformation result of the parameter.

[0144] The embodiment provides a method for constructing a dynamic mutation network based on an optimal quantum fluctuation parameter set and a search space of a specialized agent, and performing nonlinear mapping on a current parameter state based on the network. The method comprises two main steps of constructing the dynamic mutation network and performing nonlinear mapping on the parameter.

[0145] In the dynamic mutation network construction stage, the system first acquires an optimal quantum fluctuation parameter set, which contains multiple high-quality parameter vectors that have been optimized by quantum calculation. These parameter vectors represent the optimal solution region in the parameter space. The system also acquires the search space information of the specialized agent, including the upper and lower bounds of the parameters, the search history trajectory, and the parameter sensitivity, etc. Based on these two parts of information, the system constructs a dynamic mutation network containing a mutation weight matrix and a bias vector.

[0146] The construction process of the mutation weight matrix is as follows: for each parameter vector in the optimal quantum fluctuation parameter set, the system calculates the similarity between it and the current parameter state. The similarity calculation can use the cosine similarity method, that is, the dot product of two parameter vectors divided by the product of their norms. For example, if the current parameter state is [0.5, 0.3, 0.7] and a certain optimal parameter vector is [0.6, 0.2, 0.8], the similarity can be calculated as (0.5*0.6+0.3*0.2+0.7*0.8) / ), obtaining a similarity value of about 0.98. The system assigns a weight to each optimal parameter vector according to the similarity value, and the higher the similarity, the greater the weight.

[0147] The system then integrates the search space characteristics of the specialized agent into the weight matrix. For example, for a search space with parameter upper and lower bounds [0, 1], the system detects whether the current parameter is close to the boundary. If a parameter value is 0.95, close to the upper bound 1, an additional constraint is added to the corresponding position in the weight matrix to make the mutated parameter less likely to exceed the boundary. In addition, the system also considers the parameter sensitivity, and gives smaller weights to sensitive parameters in the weight matrix to reduce their mutation amplitude.

[0148] For example, assuming the system has 3 parameters, the optimal quantum fluctuation parameter set contains 5 parameter vectors, and after the above calculation, a 3x3 mutation weight matrix is obtained, such as [[0.8, 0.2, 0.1], [0.1, 0.7, 0.2], [0.2, 0.3, 0.9]]. Each row of the matrix represents the mutation weight distribution of the corresponding parameter.

[0149] The construction of the bias vector is based on the search history trajectory of the specialized agent. The system analyzes the successful search direction of the agent in the past, extracts the directional preference, and forms the bias vector. For example, if the performance improves when the first parameter tends to change to a larger value in the historical search, the corresponding bias value is positive; otherwise, it is negative. Assuming that the bias vector obtained by analysis is [0.05, -0.02, 0.01].

[0150] In the parameter nonlinear mapping stage, the system first uses the constructed dynamic mutation network to perform weighted summation on the current parameter state. Specifically, the current parameter state vector is multiplied by the mutation weight matrix, and then the bias vector is added to obtain the preliminary mapping result. In the above example, if the current parameter state is [0.5, 0.3, 0.7], the preliminary mapping result is [0.5x0.8+0.3x0.2+0.7x0.1, 0.5x0.1+0.3x0.7+0.7x0.2, 0.5x0.2+0.3x0.3+0.7x0.9]+[0.05,-0.02,0.01], which is calculated to be [0.54, 0.36, 0.77].

[0151] The system selects a suitable activation function to perform nonlinear transformation on the preliminary mapping result. Common activation functions include Sigmoid function, hyperbolic tangent function, etc. In this embodiment, the system uses the Sigmoid function as the activation function, which can map the input to the legal range of parameters while maintaining appropriate nonlinear characteristics.

[0152] Taking the Sigmoid function as an example, the function form is to subtract 0.5 from the input value, multiply by 4, then perform standard Sigmoid calculation, and finally map the result to the [0, 1] interval. For the above preliminary mapping result [0.54, 0.36, 0.77], after Sigmoid function processing, the final nonlinear transformation result [0.58, 0.32, 0.85] is obtained.

[0153] The system also implements an adaptive adjustment mechanism to dynamically adjust the parameters of the mutation network according to the parameter mutation effect. When a mutation produces good results, the system will enhance the corresponding mutation mode and adjust the corresponding value in the weight matrix; otherwise, the influence of the mutation mode will be weakened. For example, if the parameters are mutated from [0.5, 0.3, 0.7] to [0.58, 0.32, 0.85] and the performance is improved, the system will increase the value of the elements in the weight matrix that contribute to this change.

[0154] Through the construction of the dynamic mutation network and the parameter nonlinear mapping process, the system can effectively utilize the optimal parameter information provided by the quantum fluctuation mechanism while maintaining the parameter space exploration ability, thereby improving the parameter optimization efficiency. Experiments show that this method improves the convergence speed by an average of 28% compared to the traditional random mutation method, and can find better solutions in complex parameter spaces.

[0155] Figure 5 The performance comparison histogram of the dynamic mutation network of the embodiment of the application in different test scenarios is as follows:

[0156] The figure shows the performance comparison of three key technical solutions in four evaluation indicators. Among them, the white column chart represents the standard mutation network, which is a traditional parameter optimization algorithm; the column chart filled with diagonal lines represents the dynamic weight matrix, which is an adaptive optimization method; and the column chart filled with grids represents the nonlinear mapping method, which is the innovative algorithm proposed in this research. In terms of complex scenario convergence, the traditional standard mutation network reaches 64.0%, the dynamic weight matrix is improved to 78.0%, and the innovative nonlinear mapping method reaches a maximum of 88.0%. For the parameter space mapping index, the three methods reach 73.0%, 84.0% and 94.0% respectively, showing that the nonlinear mapping method has obvious advantages. In terms of high-dimensional feature extraction, the standard mutation network is 60.0%, the dynamic weight matrix is improved to 75.0%, and the nonlinear mapping method is as high as 93.0%. Finally, in terms of quantum fluctuation adaptability, the standard mutation network achieves a performance of 66.0%, the dynamic weight matrix reaches 81.0%, and the nonlinear mapping method still maintains a high level of 90.0%. From the overall data, the nonlinear mapping method proposed in this research is significantly better than the two existing technical solutions in all four evaluation dimensions, followed by the dynamic weight matrix method, and the performance of the traditional standard mutation network is relatively weak, which fully proves that the newly proposed nonlinear mapping technology has outstanding advantages in improving the overall performance of the quantum optimization system.

[0157] In a second aspect of the embodiment of the application, a gimbal control sound source tracking system based on a nested MEMS array is provided, comprising:

[0158] The first unit is configured to collect multi-channel acoustic signals of the nested MEMS array, calculate cross-correlation time delay information and phase difference information between adjacent array elements, and calculate initial spatial position coordinates of a sound source according to the time delay information and the phase difference information and in combination with a geometric structure of the array.

[0159] The second unit is configured to construct an optimization objective function based on the initial spatial position coordinates, perform first-round iterative optimization on the optimization objective function by using an adaptive quantum annealing algorithm, adaptively adjust quantum fluctuation strength according to a convergence speed of each round of iteration in the optimization process, and obtain accurate spatial position of the sound source after the first-round optimization.

[0160] The third unit is configured to convert the accurate spatial position of the sound source after the first-round optimization into gimbal motion control parameters, drive the gimbal to rotate according to the gimbal motion control parameters, and simultaneously collect gimbal attitude data and current position data of the sound source in real time.

[0161] The fourth unit is configured to calculate a tracking deviation between a pointing position corresponding to the gimbal attitude data and the current position of the sound source, and perform second-round iterative optimization on the current position data of the sound source by using the adaptive quantum annealing algorithm when the tracking deviation exceeds a preset deviation threshold, wherein quantum fluctuation strength is adaptively adjusted by inheriting optimal parameters in the first-round optimization to obtain a corrected accurate spatial position of the sound source.

[0162] The fifth unit is configured to update the gimbal motion control parameters based on the corrected accurate spatial position of the sound source, and realize continuous and accurate tracking of the gimbal.

[0163] In a third aspect, an electronic device is provided, including:

[0164] a processor;

[0165] a memory for storing processor-executable instructions;

[0166] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0167] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0168] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for performing various aspects of the present application.

[0169] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A gimbal control sound source tracking method based on nested MEMS array, characterized in that, The method comprises the following steps: Collecting multi-channel acoustic signals of a nested MEMS array, calculating cross-correlation time delay information and phase difference information between adjacent array elements, and calculating initial spatial position coordinates of a sound source according to the time delay information and the phase difference information, comprising: Collecting acoustic signals of adjacent array elements in the MEMS array, performing frequency domain decomposition on the acoustic signals to obtain sub-signals of multiple frequency bands, performing signal-to-noise ratio analysis and coherence analysis on the sub-signals, and constructing a feature vector based on the results of the signal-to-noise ratio analysis and the coherence analysis; Optimizing the feature vector by using a quasi-Newton method to obtain a weight coefficient, applying the weight coefficient to the sub-signals to obtain a weighted fusion signal, and calculating an initial time delay estimation value by performing time delay parameter calculation on the weighted fusion signal and constructing a search interval around the initial time delay estimation value; Constructing an optimization objective function by using the time delay parameters in the search interval and the feature vector, and iteratively solving the optimization objective function by using the quasi-Newton method to obtain accurate time delay information, and calculating initial phase differences between adjacent array elements according to the accurate time delay information; Constructing a phase difference correction function based on the feature vector, optimizing and calculating the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, and establishing a sound source position solving equation set according to the corrected phase difference information and the accurate time delay information, substituting geometric structure parameters of the MEMS array into the sound source position solving equation set, and solving to obtain spatial position coordinates of the sound source; Constructing an optimization objective function based on the initial spatial position coordinates, performing first-round iterative optimization on the optimization objective function by using an adaptive quantum annealing algorithm, and adaptively adjusting quantum fluctuation strength according to the convergence speed of each round of iteration in the optimization process to obtain a first-round optimized accurate spatial position of the sound source; Converting the first-round optimized accurate spatial position of the sound source into gimbal motion control parameters, driving the gimbal to rotate according to the gimbal motion control parameters, and simultaneously collecting real-time gimbal attitude data and current sound source position data; Calculating a tracking deviation between a pointing position corresponding to the gimbal attitude data and the current sound source position; when the tracking deviation exceeds a preset deviation threshold, performing second-round iterative optimization on the current sound source position data by using the adaptive quantum annealing algorithm, wherein the quantum fluctuation strength is adaptively adjusted based on the optimal parameters in the first-round optimization to obtain a corrected accurate spatial position of the sound source; Updating the gimbal motion control parameters based on the corrected accurate spatial position of the sound source, and realizing continuous and accurate tracking of the gimbal.

2. The method of claim 1, wherein, Constructing a phase difference correction function based on the feature vector, optimizing and calculating the phase difference correction function by using the quasi-Newton method to obtain corrected phase difference information, comprising: Establishing a phase difference correction model based on the feature vector, taking the phase difference correction model as an optimization objective, constructing a target function containing a regularization term, and optimizing and solving the target function by using a quasi-Newton method, and adaptively adjusting optimization parameters according to signal characteristics in each iteration process. An error analysis is performed on the iteration results of the quasi-Newton method, and a relative error of adjacent two iteration results is calculated; when the relative error is less than a preset error threshold, a phase difference obtained in the current iteration is output as corrected phase difference information.

3. The method of claim 1, wherein, According to the corrected phase difference information and the accurate time delay information, a sound source position solving equation set is established, and geometric structure parameters of the MEMS array are substituted into the sound source position solving equation set to obtain spatial position coordinates of the sound source, including: According to the corrected phase difference information and the accurate time delay information, a sound source position solving equation set is established, and the sound source position solving equation set is converted to a multi-dimensional solving space, a multi-dimensional expression containing sound source spatial coordinate components is constructed, and the multi-dimensional expression converts the solving equation set in the Euclidean space to a multi-dimensional state representation through a coordinate mapping function; An information coding strategy is designed, and the corrected phase difference information and the accurate time delay information are coded into a combined state expression, the combined state expression adopts a weighted fusion manner to maintain the coupling relationship between the information, and the weighted fusion manner dynamically adjusts weight coefficients based on a signal quality index; Geometric structure parameters of the MEMS array are obtained, and constraint conditions containing array element spacing constraints and array deformation constraints are established based on the geometric structure parameters, and the constraint conditions limit the feasible region of the solution through a penalty function; An adaptive annealing strategy is used to solve the sound source position solving equation set, and the constraint conditions are optimized in parallel in the multi-dimensional solving space, the adaptive annealing strategy dynamically adjusts system energy parameters according to the convergence speed of the solution, and adaptively updates control parameters based on gradient information; A probability evaluation is performed on the intermediate solution in the optimization process, the probability evaluation calculates the credibility of each candidate solution based on an error function, and the candidate solution with the highest credibility is selected as the final spatial position coordinates.

4. The method of claim 1, wherein, Based on the initial spatial position coordinates, an optimization objective function is constructed, and an adaptive quantum annealing algorithm is used to perform first-round iterative optimization on the optimization objective function, quantum fluctuation strength is adaptively adjusted according to the convergence speed of each iteration in the optimization process, and the first-round optimized sound source accurate spatial position includes: A target function evaluation module is constructed based on a deep neural network, multi-scale decomposition is performed based on the output of the target function evaluation module, and coarse scale coordinates and fine scale coordinates are generated; Environmental temperature parameters and humidity parameters are collected, an environmental influence model is established based on the environmental temperature parameters and the humidity parameters, an output result of the environmental influence model is taken as an environmental constraint term, and the environmental constraint term and the coarse scale coordinates and the fine scale coordinates are used to construct an optimization objective function; An adaptive annealing algorithm is used to perform iterative optimization on the optimization objective function, a change amount and an annealing temperature of the target function are calculated in each iteration process, a state jump probability is determined according to the change amount and the annealing temperature, and the coarse scale coordinates and the fine scale coordinates are updated based on the state jump probability; The convergence speed of each round of iteration is input into the target function evaluation module for prediction, the weight coefficients of each term in the optimization target function are dynamically adjusted based on the prediction result of the target function evaluation module, and the coarse scale coordinates and the fine scale coordinates are optimized according to the weight coefficients; the environmental influence model is updated in real time, the updated environmental influence model is substituted into the optimization target function again, and iteration optimization is continued based on the updated optimization target function; When the coordinate difference of two adjacent iterations is less than a preset convergence threshold, the accurate spatial position of the sound source is calculated.

5. The method of claim 1, wherein, The adaptive quantum annealing algorithm is used to perform second round iteration optimization on the current position data of the sound source, wherein the quantum fluctuation strength is adaptively adjusted by inheriting the optimal parameter in the first round optimization, and the corrected accurate spatial position of the sound source is obtained, including: An optimal quantum fluctuation parameter set obtained by the first round optimization is acquired, a plurality of specialized agents are constructed based on the optimal quantum fluctuation parameter set, and each specialized agent is responsible for optimization search of a corresponding dimension parameter; a dynamic variation network is constructed based on the optimal quantum fluctuation parameter set and the search space of the specialized agent, and a current parameter state is nonlinearly mapped based on the dynamic variation network; The optimization parameters of the specialized agent are mapped to a quantum state space, a quantum expression including a parameter quantum state and a probability amplitude is constructed, a game matrix is constructed based on the optimization results of the specialized agent, the optimization benefits between adjacent agents are calculated through the game matrix, and the optimization strategy of the specialized agent is dynamically adjusted according to the optimization benefits; The optimization strategy is input into an adaptive adjustment module, the adaptive adjustment module calculates the adjustment amount of the quantum fluctuation strength according to the agent weight and the optimization gain, and updates the quantum fluctuation parameter based on the adjustment amount; The updated quantum fluctuation parameter is substituted into the dynamic variation network to generate a new parameter state, and the new parameter state is input into the quantum state space for optimization; when the parameter state difference of two adjacent iterations is less than a preset difference threshold, a corrected accurate spatial position of the sound source is calculated based on the current optimization parameter and the quantum fluctuation parameter.

6. The method of claim 5, wherein, The dynamic variation network is constructed based on the optimal quantum fluctuation parameter set and the search space of the specialized agent, and the current parameter state is nonlinearly mapped based on the dynamic variation network, including: A dynamic variation network including a variation weight matrix and a bias vector is constructed based on the optimal quantum fluctuation parameter set and the search space of the specialized agent; the current parameter state is weighted and summed by using the variation weight matrix and the activation function of the dynamic variation network to obtain a preliminary mapping result; The preliminary mapping result is input into the activation function for nonlinear mapping to obtain a nonlinear transformation result of the parameter.

7. A gimbal-controlled acoustic source tracking system based on nested MEMS arrays for implementing the method of any of the preceding claims 1-6, characterized in that, including: A first unit is configured to collect multi-channel acoustic signals of a nested MEMS array, calculate cross-correlation time delay information and phase difference information between adjacent elements, and calculate an initial spatial position coordinate of a sound source according to the time delay information and the phase difference information in combination with a geometric structure of the array; The second unit is configured to construct an optimization objective function based on the initial spatial position coordinates, perform first-round iterative optimization on the optimization objective function by using an adaptive quantum annealing algorithm, and adaptively adjust quantum fluctuation strength according to a convergence speed of each round of iteration in the optimization process to obtain a sound source accurate spatial position after first-round optimization. The third unit is configured to convert the sound source accurate spatial position after first-round optimization into a gimbal motion control parameter, drive the gimbal to rotate according to the gimbal motion control parameter, and simultaneously collect gimbal attitude data and sound source current position data in real time. The fourth unit is configured to calculate a tracking deviation between a pointing position corresponding to the gimbal attitude data and the sound source current position, and perform second-round iterative optimization on the sound source current position data by using the adaptive quantum annealing algorithm when the tracking deviation exceeds a preset deviation threshold, wherein quantum fluctuation strength is adaptively adjusted by inheriting an optimal parameter in the first-round optimization to obtain a corrected sound source accurate spatial position. The fifth unit is configured to update the gimbal motion control parameter based on the corrected sound source accurate spatial position to realize continuous and accurate tracking of the gimbal.

8. An electronic device, comprising: The computer program instructions are executed by the processor to implement the method in any one of claims 1 to 6. The computer program instructions are executed by the processor to implement the method in any one of claims 1 to 6. ​ ​ 9. A computer-readable storage medium having stored thereon computer program instructions, wherein, ​

Images