Bat pulse parameter automatic calibration method and device based on adaptive BO-NM hybrid optimization

Abstract

This invention discloses an automatic calibration method and apparatus for bat pulse parameters based on adaptive BO-NM hybrid optimization, mainly addressing the problems of existing technologies such as reliance on manual setting of detection parameters, limited pulse boundary accuracy, and low automation of multi-dimensional parameter extraction. The scheme includes: preprocessing the original recording data to obtain a smooth envelope signal and root mean square value; defining a four-dimensional parameter vector and constructing a multi-objective fitness function; obtaining the globally optimal parameter configuration through alternating execution of Bayesian global search and simplex local refinement, and calculating an adaptive threshold to drive a finite state machine to complete preliminary pulse detection; dynamically refining the start and end boundaries based on the energy ratio, anchoring the CF main frequency within each pulse interval, and bidirectionally tracking the frequency trajectory to perform iFM-CF-tFM three-component decoupling, outputting a multi-dimensional acoustic parameter feature set matrix. This invention can achieve automatic calibration of detection parameters, improve the accuracy of pulse detection and decoupling, and can be used for bat bio-sonar research and biomimetic radar parameter mapping.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of signal processing and intelligent optimization, and specifically relates to an adaptive calibration method and system for pulse detection parameters of bat echolocation signals, which can be used in bat bio-sonar systems and biomimetic radar. Background Technology

[0002] Bat bio-sonar systems possess extremely high detection accuracy and environmental adaptability, serving as important biomimetic models for artificial sensing systems such as radar and sonar. Bat echolocation signals exhibit high species specificity and behavioral correlation, and their acoustic parameters contain rich ecological and biomimetic information. Accurate extraction of these parameters is crucial for revealing bat echolocation strategies, understanding their target detection and tracking mechanisms, and guiding radar anti-jamming waveform design. However, pulse detection and parameter extraction of bat signals face multiple challenges: large signal intensity variations, coexistence of strong and weak pulses, and difficulty in balancing fixed thresholds; blurred component boundaries in the CF-FM composite modulation structure, requiring precise frequency tracking and segmentation; and the coupling of multiple detection parameters, making it difficult to manually determine the optimal configuration. Therefore, a method with intelligent parameter calibration and automatic multi-dimensional feature extraction capabilities is urgently needed to support research on bio-sonar mechanisms and the transformation of biomimetic sensing technologies.

[0003] Current research has focused on pulse detection, component segmentation, and parameter extraction of bat acoustic signals, with key technical solutions including:

[0004] For pulse detection, threshold-based pulse recognition methods determine the start and end positions of pulses by comparing the short-time energy or spectral peak of the signal with a preset threshold. These methods are highly dependent on manual intervention, requiring repeated trial and error adjustments of the threshold parameters. The signal-to-noise ratio and dynamic range vary significantly under different experimental conditions. Fixed-threshold methods often require resetting parameters for each data segment, resulting in low processing efficiency and difficulty in reproducing results due to differences in operator experience.

[0005] For component segmentation of CF-FM composite modulated calls, manual segmentation is intuitive but time-consuming and has poor repeatability. The peak frequency method separates CF and FM components by setting a peak frequency percentage threshold and using high-pass and low-pass filtering, but requires high-quality axial recording, and its performance degrades significantly when the recording angle deviates. The pseudo-Wigner-Ville distribution method tracks instantaneous frequencies and segments based on modulation rate thresholds, which is highly adaptable to recording conditions, but is sensitive to noise. Instantaneous frequency estimation is prone to errors under low signal-to-noise ratio conditions, and the lack of anchor references results in limited accuracy in locating the boundaries of the gradual transition region. The fractional Fourier transform method combined with empirical mode decomposition can separate nonlinear components, but its computational complexity is high. Furthermore, all of the above methods require manual setting of key parameters and mainly focus on the component segmentation stage, lacking a systematic ability to extract pulse detection and temporal dynamic features.

[0006] A patent document with publication number CN120675663B discloses a method for detecting and countering unmanned aerial vehicles (UAVs) in order to address intelligent signal perception and automatic parameter optimization. This method deploys a multi-dimensional perception network for instantaneous scanning across the entire frequency band, simultaneously acquiring radio signal noise characteristics, terahertz band material characteristics, and Doppler effect characteristics. It then combines this with a camouflage sample library from an adversarial generative network for in-depth comparison, providing valuable insights into multi-dimensional signal feature fusion detection. However, because this method targets UAV target detection scenarios, the physical characteristics of the signal processing object differ significantly from those of bat ultrasound. Furthermore, it does not address the decoupling problem of the CF-FM composite modulation structure unique to biological sonar signals, making it difficult to apply to the refined extraction of bat pulse parameters.

[0007] Patent document WEIEP4396730B1 discloses a system and method for automatically constructing stochastic deep neural network architectures. It employs automatic search for the most relevant stochastic patterns, utilizes heterogeneous, irregular, and misconfiguration confidence as intermediate representations of the deep neural network, and regularizes it using divergence metrics including Renyi divergence. This provides a useful approach for automated parameter inference and model structure search. However, because this method is designed only for general deep learning architecture search problems, and its optimization objective is the network structure rather than fine-grained calibration in a continuous parameter space, it lacks the ability to fuse prior knowledge of acoustic signal processing. Therefore, it cannot be directly applied to the global optimization of coupling parameters in bat pulse detection.

[0008] In summary, the existing technology has the following key shortcomings:

[0009] First, the detection parameters rely on manual setting and lack automatic calibration capabilities. There are complex coupling relationships between multiple detection parameters, making it difficult to find the globally optimal configuration through manual parameter tuning. Furthermore, the differences in parameter selection by different operators make it difficult to reproduce the processing results, which cannot support the batch processing of large-scale data and cross-laboratory comparative studies.

[0010] Secondly, the accuracy of pulse boundary detection is limited. Existing methods use fixed energy thresholds for pulse detection, which is difficult to adapt to the large dynamic range of signal amplitude changes. Strong pulses are prone to premature threshold triggering and boundary shifting forward, while weak pulses are prone to missed detection or boundary shifting backward due to insufficient signal-to-noise ratio. The pulse termination determination does not consider the differences in envelope attenuation characteristics of pulses of different intensities, resulting in systematic errors in boundary localization. Although multidimensional signal fusion detection technology has made progress in scenarios such as UAV detection, the physical characteristics of its signal processing objects are significantly different from those of bat ultrasound, and it cannot solve the special requirement of consistent boundary localization of heterogeneous intensity pulses in biological sonar signals.

[0011] Third, the automation level of multidimensional parameter batch extraction is low. Existing methods generally lack the ability to finely decouple the CF-FM component structure, and cannot automatically and in batches extract the key acoustic parameters of each modulation component, making it difficult to support high-throughput processing of large-scale data and systematic research on biomimetic radar parameter mapping. Summary of the Invention

[0012] The purpose of this invention is to address the shortcomings of the prior art by proposing an automatic calibration method and apparatus for bat pulse parameters based on adaptive BO-NM hybrid optimization, so as to improve the efficiency of acoustic parameter calibration and the decoupling accuracy of pulse detection and CF-FM components, and automatically output a complete feature set of acoustic parameters in multiple dimensions such as time domain and frequency domain.

[0013] The technical approach to achieving the objective of this invention is as follows: by configuring coupling parameters such as the globally optimal automatic configuration threshold coefficient and the FM deviation threshold, the efficiency of acoustic parameter calibration is improved; by performing energy adaptive boundary correction and bidirectional frequency tracking segmentation, the decoupling accuracy of pulse detection and CF-FM components is improved; and by automatically outputting a complete feature set of acoustic parameters in multiple dimensions such as time domain and frequency domain through pulse parameter calibration and detection.

[0014] Based on the above ideas, the technical solution of the present invention includes:

[0015] 1. A method for automatic calibration of bat pulse parameters based on adaptive BO-NM hybrid optimization, characterized in that it includes:

[0016] (1) Obtain the original bat ultrasound recording data and preprocess it to obtain the smooth envelope signal and the root mean square value of the envelope;

[0017] (2) Define a four-dimensional parameter vector to be optimized and feasible region boundary constraints based on the smooth envelope signal and the root mean square value of the envelope.

[0018] And construct a multi-objective weighted fitness function;

[0019] (3) Using the fitness function as the optimization objective, set the exponentially decaying dynamic exploration coefficient, use this coefficient and the Gaussian process surrogate model to perform a Bayesian global search in the feasible region, and record the optimal parameter configuration in this stage;

[0020] (4) Using the optimal parameter configuration as the initial point, normalize it to the unit space and perform local refinement of the Nelder-Mead simplex to obtain the optimal parameter configuration for the simplex stage;

[0021] (5) Select the one with the better fitness value among the two optimal parameter configurations in the above two stages and update its global optimal parameter configuration;

[0022] (6) Calculate the adaptive detection threshold by multiplying the threshold coefficient obtained by the global optimal parameter configuration with the root mean square value of the envelope, and drive the finite state machine to complete the initial pulse detection by combining the minimum pulse width constraint and the minimum pulse interval constraint.

[0023] (7) The detection threshold is dynamically adjusted according to the energy ratio of each pulse, and the start and end boundaries of each pulse obtained from the preliminary detection are refined by energy self-adaptation to obtain a high-precision pulse time interval;

[0024] (8) Anchor the CF main frequency in each pulse time interval and track the instantaneous frequency trajectory bidirectionally along the time axis, and use the FM frequency deviation threshold as the judgment criterion to complete the decoupling of the iFM-CF-tFM three components;

[0025] (9) Summarize the decoupling results of the pulse time interval and the three components of iFM-CF-tFM, and automatically output the complete feature set matrix of multi-dimensional acoustic parameters.

[0026] Furthermore, in step (6), the initial pulse detection is completed by using an adaptive detection threshold combined with minimum pulse width constraints and minimum pulse interval constraints to drive a finite state machine, which includes:

[0027] 6a) Define the three states of a finite state machine: silent, impulsive, and interval.

[0028] 6b) In the silent state, continuously monitor the envelope value of the current sampling point to wait for the arrival of a valid pulse; if the envelope value of the current sampling point exceeds the adaptive detection threshold and the interval with the end position of the previous pulse is greater than the minimum pulse interval constraint, then switch to the pulse in-pulse state and record the pulse start candidate point.

[0029] 6c) In the pulse state, track the duration of the current pulse and accumulate the pulse duration; if the envelope value falls below the threshold, check whether the pulse duration meets the minimum pulse width constraint. If it does, it is confirmed as a valid pulse and transferred to the interval state; if it does not meet the constraint, it is judged as a false alarm and returns to the silent state.

[0030] 6d) In the interval state, block brief disturbances immediately following the valid pulse to prevent the same pulse from being triggered repeatedly; if the interval between the current sampling point and the confirmed pulse end position exceeds the minimum pulse interval constraint, return to the silent state.

[0031] 6e) Repeat steps 6b) and 6d) one sample point at a time until all signals have been processed and a preliminary pulse list is output.

[0032] Furthermore, in step (8), the decoupling of the iFM-CF-tFM three components is completed using the FM frequency deviation threshold as the criterion, which includes:

[0033] 8a) Perform a short-time Fourier transform on each pulse time interval, search for the maximum power spectrum in the prior frequency interval of CF, and determine the dominant frequency of CF and its time position as anchor points.

[0034] 8b) Extract the peak frequency frame by frame from the anchor point along the time axis. When the deviation from the CF main frequency exceeds the FM frequency deviation threshold, determine the iFM end position and search for the maximum bandwidth point within the analysis window above the power threshold to determine the iFM start frequency.

[0035] 8c) Extract the peak frequency frame by frame from the anchor point along the time axis. When the deviation exceeds the FM frequency deviation threshold, determine the tFM start position and continue tracking to the pulse end boundary to determine the tFM termination frequency.

[0036] 8d) Output the time interval and frequency parameters of the iFM segment, CF segment, and tFM segment of each pulse.

[0037] 2. An automatic calibration device for bat pulse parameters based on adaptive BO-NM hybrid optimization, characterized in that it comprises:

[0038] The signal preprocessing module is used to perform bandpass filtering, Hilbert transform, and short-time Fourier transform on the raw bat ultrasound recording data, and output a smooth envelope signal and root mean square value of the envelope.

[0039] The parameter intelligent calibration module is used to define the four-dimensional parameter vector to be optimized and the fitness function. It outputs the globally optimal parameter configuration by dynamically exploring the coefficient Bayesian optimization and the normalized spatial soft boundary simplex method in stages, combined with the perturbation restart mechanism.

[0040] The high-precision pulse detection module is used to drive a finite state machine to complete the initial pulse detection using the globally optimal parameter configuration, and to perform start and end boundary refinement by dynamically adjusting the detection threshold according to the energy ratio.

[0041] The CF-FM component segmentation module is used to anchor the CF main frequency and track the instantaneous frequency trajectory bidirectionally within each pulse time interval, thereby completing the decoupling of the iFM-CF-tFM three components.

[0042] The multidimensional parameter extraction module is used to summarize the decoupling results of the pulse time interval and the three components of iFM-CF-tFM, and automatically output the complete feature set matrix of multidimensional acoustic parameters.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] 1. This invention achieves globally optimal automatic configuration of detection parameters through a collaborative mechanism of BO global search, NM local refinement, and perturbation restart to escape local extrema, without the need for manual intervention throughout the process. Therefore, it can significantly improve the efficiency of acoustic parameter extraction and the results are completely reproducible.

[0045] 2. This invention uses an energy-adaptive boundary detection mechanism to dynamically adjust the detection threshold based on the pulse energy, making the positioning accuracy of the start and end boundaries of strong and weak pulses more consistent. This effectively solves the systematic error problems such as the shift of the end boundary caused by the envelope tail of strong pulses and the boundary positioning drift caused by the high fixed threshold of weak pulses.

[0046] 3. This invention uses a two-way frequency trajectory with the peak power of CF as the anchor point for tracking and uses the global optimal FM frequency deviation threshold as the accurate judgment criterion to locate the boundaries of the three components iFM, CF, and tFM. Therefore, it can independently extract parameters such as frequency, bandwidth, duration, and frequency modulation slope of each component, providing a complete and accurate bio-sonar feature mapping for biomimetic radar waveform design. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the automatic calibration method for bat pulse parameters based on adaptive BO-NM hybrid optimization of the present invention.

[0048] Figure 2 This is a sub-flowchart of the pulse energy adaptive boundary correction method in the present invention;

[0049] Figure 3 This is a block diagram of the automatic calibration device for bat pulse parameters based on adaptive BO-NM hybrid optimization of the present invention;

[0050] Figure 4 The graph shows the comparison results of different indicators for pulse detection and parameter optimization using the present invention and existing manual threshold method and fixed parameter method.

[0051] Figure 5 The graph shows the comparison results of different indicators for pulse detection and parameter optimization using the present invention and existing manual threshold method and fixed parameter method. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of the present invention.

[0053] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

[0054] Example 1: Automatic calibration method for bat pulse parameters based on adaptive BO-NM hybrid optimization.

[0055] In bat echolocation signal processing, accurate pulse detection and parameter extraction require the rational configuration of multiple coupled detection parameters. Due to the complex nonlinear interactions between these parameters, forming a multi-peaked, non-convex fitness surface, a single optimization algorithm struggles to balance global search and local convergence. This example employs a phased, alternating approach of Bayesian optimization and the Nelder-Mead simplex method, combined with energy-adaptive boundary detection and bidirectional frequency tracking techniques, to achieve intelligent optimization of detection parameters and fully automated extraction of multidimensional acoustic parameters.

[0056] Reference Figure 1 The implementation steps of this embodiment include the following:

[0057] Step 1, Signal Input and Preprocessing.

[0058] (1.1) Receive raw bat echolocation data, which includes bat echolocation signals and environmental noise, with a sampling rate of This embodiment is provided, but is not limited to ;

[0059] (1.2) Design a Chebyshev Type I bandpass filter to extract the target frequency band signal and suppress out-of-band noise:

[0060] This embodiment sets the target frequency band of the Chebyshev Type I bandpass filter. The passband boundary frequency is The stopband boundary frequency is The passband ripple is 0.1 dB;

[0061] Determine the minimum filter order that meets the specifications using the Chebyshev Type I order estimation function. Based on this, filter coefficients were designed with the target species' CF main frequency as the center frequency. The original signal is filtered to obtain the filtered signal. :

[0062] ,

[0063] in, For pulse envelope, The instantaneous frequency of the pulse;

[0064] (1.3) Filtered signal Perform Hilbert transform to obtain analytic signal :

[0065] ,

[0066] in, For Hilbert transform operators, The imaginary unit;

[0067] (1.4) Calculate the instantaneous amplitude envelope of the analytic signal ;

[0068] (1.5) Apply a moving average window to smooth the instantaneous amplitude envelope to obtain a smoothed envelope signal:

[0069] ,

[0070] in, The length of the window;

[0071] (1.6) Calculate the root mean square value of the envelope. As an adaptive threshold reference benchmark.

[0072] Step 2: Define the parameter space and fitness function.

[0073] The performance of bat pulse detection is highly dependent on the coordinated configuration of four coupled parameters: threshold coefficient, minimum pulse width, minimum pulse interval, and FM frequency deviation threshold. To provide subsequent optimization algorithms with a clear search space and quantifiable optimization objectives, this step encapsulates these four parameters into a unified parameter vector to be optimized. Furthermore, by constructing a comprehensive fitness function, it integrates three quality indicators—pulse number deviation, physiological rationality of duration, and detection stability—into a single scalar, providing a comparable and directional evaluation basis for subsequent Bayesian optimization and simplex search. This step includes:

[0074] (2.1) Defined by threshold coefficient Minimum pulse width constraint Minimum pulse interval constraint and FM frequency deviation threshold The four-dimensional parameter vector to be optimized And set the upper bound of the feasible region for each component. and the lower realm ;

[0075] (2.2) Calculate the normalized deviation term used to quantify the number of tests. :

[0076] ,

[0077] in, To detect the number of pulses, The a priori expected number of pulses;

[0078] (2.3) Calculate the proportional temporal anomaly used to quantify the deviation of pulse duration from physiological constraints. :

[0079] ,

[0080] in, The number of pulses whose duration exceeds the physiologically reasonable range;

[0081] (2.4) Calculate the statistical stability term used to quantify the statistical dispersion of the detection results. :

[0082] ,

[0083] in, This is the duration sequence of all detected pulses. To prevent division by zero;

[0084] (2.5) Adjust the above quantity deviation items according to the weighting coefficients. Time-domain anomalies and statistical stability term The weighted sum of these three terms yields the comprehensive fitness function:

[0085] ,

[0086] in, , , These are the weighting coefficients for the quantity deviation term, the time-domain anomaly term, and the statistical stability term, respectively.

[0087] Step 3: Perform Bayesian optimization on the dynamic exploration coefficients.

[0088] Bayesian optimization approximates the fitness surface using a Gaussian process surrogate model, achieving efficient global search of samples in a parameter space where evaluation is costly. Fixed exploration coefficients struggle to balance initial wide-area sampling with later fine-grained convergence. This step introduces exponentially decaying dynamic exploration coefficients, enabling the optimization to focus on global exploration in the early stages and automatically shift to local utilization in the later stages, thereby obtaining a high-quality Bayesian stage-optimal solution within a limited iteration budget. This step includes:

[0089] (3.1) Calculate the exploration coefficients for the current iteration. :

[0090] ,

[0091] in, This is the initial exploration coefficient. To minimize the exploration coefficient, The attenuation rate, This represents the current iteration number. This represents the total number of iterations.

[0092] Early stage of iteration Approaching The acquisition function focuses on exploring regions of high uncertainty; as iterative progresses... It decays exponentially and approaches [a certain value]. The acquisition function gradually shifts towards utilizing known high-fitness regions;

[0093] (3.2) Construct a Gaussian process surrogate model using existing evaluation points, improve the acquisition function based on the expected value, and use the current exploration coefficients. Balancing regulation with exploration and utilization, and selecting the next evaluation point in the parameter space;

[0094] (3.3) Perform fast impulse detection at the candidate evaluation points and calculate the fitness value. Add the new evaluation points and their fitness values ​​to the observation set to update the posterior distribution of the surrogate model.

[0095] (3.4) Repeat steps (3.1) to (3.3) until the preset number of iterations is completed, and record the optimal parameter configuration for this stage. .

[0096] Step 4: Refine the normalized space using a soft-boundary simplex.

[0097] Bayesian optimization excels at global localization but has limited local convergence accuracy. The Nelder-Mead simplex method can quickly achieve high-precision local refinement within the neighborhood of its optimal solution. However, simplex search is prone to exceeding the feasible region boundary in the original parameter space, leading to invalid evaluations. This step normalizes the parameters to unit space and introduces a Sigmoid soft boundary projection, implicitly constraining parameter feasibility without significantly interfering with the search trajectory, thus ensuring the stability and effectiveness of the simplex local search. This step includes:

[0098] (4.1) Normalize the parameter vector to the interval [0,1] according to the upper and lower bounds of the feasible region: ;

[0099] (4.2) Perform Sigmoid soft boundary projection on the normalized parameters to obtain the projected Sigmoid parameter vector. :

[0100] ,

[0101] in, For steepness parameters, The upper bound of the feasible region. This is the lower bound of the feasible region.

[0102] (4.3) will The amplitude is limited to the [0,1] interval, and finally denormalized to the original parameter space. ;

[0103] (4.4) Configure the optimal parameters in the Bayesian optimization stage in the normalized space. For the initial simplex, perform a Nelder-Mead search, with the parameter set to: maximum number of iterations. Function tolerance Parameter tolerance The search results are then inversely normalized to the original parameter space after Sigmoid projection to obtain the optimal parameter configuration for the simplex stage. .

[0104] Step 5: Update the globally optimal parameters and perform a perturbation restart check.

[0105] The Bayesian optimization and simplex refinement, performed alternately in stages, may repeatedly converge to the same local extremum due to the proximity of initial points during multiple iterations. To actively escape local extrema and ensure global convergence, this step updates the global optimum and introduces a perturbation restart judgment. This step includes:

[0106] (5.1) Compare the fitness value of the optimal parameter configuration obtained in the current round with the recorded global optimal parameter configuration. If the fitness value of the current round is better, update the global optimal parameter configuration. Otherwise keep constant;

[0107] (5.2) Monitor the update status of the global optimal parameter configuration between consecutive rounds. If the global optimal parameter configuration is not updated for two consecutive rounds, it is determined that the optimization has stalled.

[0108] (5.3) Determine if the current optimization is stalled:

[0109] If the optimization is stalled, apply a random perturbation to the current globally optimal parameter configuration to obtain the perturbed parameter configuration. :

[0110] ,

[0111] Among them, the amplitude of the disturbance Based on a value of 0.1 and varying with the number of stalled rounds The parameters are configured linearly after the disturbance. As the initial parameter configuration for the next round;

[0112] If the optimization is not stalled, then configure the current globally optimal parameters. As the initial parameter configuration for the next round;

[0113] (5.4) Repeat steps (5.1) to (5.3) until all preset optimization rounds are completed. .

[0114] Step 6: Perform pulse detection based on a finite state machine.

[0115] After parameter optimization, the globally optimal parameter configuration needs to be applied to actual pulse detection. A finite state machine, using an adaptive threshold as the criterion, performs state transitions for each sampling point of the envelope signal. Combined with minimum pulse width and minimum interval constraints to filter false alarms, efficient and controllable preliminary pulse timing localization can be achieved. This implementation includes the following steps:

[0116] (6.1) Calculate the adaptive detection threshold :

[0117] ,

[0118] in, The threshold coefficient in the globally optimal parameter configuration. This represents the root mean square value of the smoothed envelope signal;

[0119] (6.2) Define the three states of the finite state machine: silent, intra-pulse, and interval, denoted as follows: , , Perform state transition determination for each sample point of the smooth envelope signal:

[0120] exist In this state, the envelope value of the current sampling point is continuously monitored to wait for the arrival of a valid pulse: if the envelope value of the current sampling point exceeds the adaptive detection threshold... Furthermore, the interval between the current position and the end position of the previous pulse is greater than the minimum pulse interval constraint. Then transfer to State and record the candidate pulse start point; otherwise, maintain. The state remains unchanged;

[0121] exist In this state, track the duration of the current pulse and accumulate the pulse duration: if the envelope value falls below the threshold, check whether the pulse duration meets the minimum pulse width constraint. If the condition is met, it is confirmed as a valid pulse and transferred to... If the condition is not met, it is judged as a false alarm and a response is returned. If the envelope value is above the threshold, then keep it. The state remains unchanged;

[0122] exist In this state, a brief disturbance immediately following a valid pulse is masked to prevent the same pulse from being triggered repeatedly: if the interval between the current sampling point and the end position of the confirmed pulse exceeds the minimum pulse interval constraint. Then return Otherwise, maintain the status quo.

[0123] (6.4) Repeat step (6.3) for each sampling point until all signals are processed and a preliminary pulse list is output.

[0124] Step 7: Perform energy adaptive boundary correction on the preliminary pulse list.

[0125] The pulse boundaries obtained from the initial detection are determined based on a uniform threshold. For strong pulses, the end boundary is prone to shifting due to envelope tailing, and for weak pulses, the boundary is prone to drifting due to insufficient signal-to-noise ratio, resulting in a systematic positioning error related to pulse energy. To make the boundary positioning accuracy of pulses of different intensities more consistent, this step performs energy-adaptive boundary correction on the initial pulse list.

[0126] Reference Figure 2 The implementation of this step includes:

[0127] (7.1) Calculate the energy ratio of each pulse. ,in For the first The peak envelope of each pulse, This is the root mean square value of the envelope;

[0128] (7.2) Dynamically determine the detection threshold ratio based on the energy ratio: If the energy ratio If the value exceeds the preset threshold, the detection threshold ratio is increased linearly. base value With upper limit The energy ratio between them; otherwise, the baseline value is maintained; this embodiment is set, but not limited to , ;

[0129] (7.3) Perform initial boundary correction and backtrack from the initial detection starting point to find the point where the positioning envelope first falls below the preset peak value as the precise pulse start position;

[0130] (7.4) Search along the time axis from the peak position to find the dynamic threshold proportion of envelope values ​​lower than the peak value. Alternatively, the pulse may be precisely terminated when the attenuation exceeds a preset amplitude within a short time window, thus correcting the termination boundary.

[0131] (7.5) Calculate the pulse duration based on the corrected start and end boundaries, and then determine its duration:

[0132] If the pulse duration after refinement is less than the minimum protection pulse width, then pulse correction protection is executed. After restoring the pulse interval to the initial detection boundary, step 8 is executed.

[0133] Otherwise, proceed directly to step 8.

[0134] Step 8: Perform CF-FM segmentation based on bidirectional frequency tracking.

[0135] After determining the high-precision pulse time interval, it is necessary to further decouple the iFM, CF, and tFM components precisely to extract the frequency and duration parameters of each segment. The implementation steps include:

[0136] (8.1) Perform a short-time Fourier transform on each pulse time interval, search for the power spectral maximum within the CF prior frequency interval, and determine the first pulse. CF main frequency of one pulse and its time and location As an anchor point;

[0137] (8.2) From the anchor point Extract the peak frequency frame by frame along the time axis and compare it with the CF main frequency. The deviation exceeds the FM frequency deviation threshold. The iFM start frequency is determined by searching for the maximum bandwidth point within the analysis window above the power threshold.

[0138] (8.3) From the anchor point The peak frequency is extracted frame by frame along the time axis and compared with the CF main frequency. The deviation exceeds the FM frequency deviation threshold. The starting position of tFM is determined at the pulse end boundary point, and the tFM termination frequency is determined by continuing to track until the pulse ends.

[0139] (8.4) Output the time interval and frequency parameters of the iFM segment, CF segment and tFM segment of each pulse.

[0140] Step 9: Extract and output multidimensional parameters.

[0141] To form a systematic acoustic parameter set for subsequent bio-sonar mechanism analysis and biomimetic radar parameter mapping, this step summarizes the acoustic parameters obtained in steps 7 and 8, outputting a complete feature set matrix, which includes:

[0142] (9.1) Extract the time-domain parameters obtained in steps 7 and 8 to form a time-domain parameter matrix. :

[0143] ,

[0144] in, The total pulse duration For iFM segment duration, For CF segment duration, For the duration of the tFM segment, The pulse interval;

[0145] (9.2) Extract the frequency domain parameters obtained in step 8 to construct a frequency domain parameter matrix. :

[0146] ,

[0147] in, For CF main frequency, and These are the start and cutoff frequencies of the iFM, respectively. and These are the start and cutoff frequencies of tFM, respectively. For iFM bandwidth, For tFM bandwidth;

[0148] (9.3) Extract the energy domain parameters obtained in step 7 and construct the energy parameter matrix. :

[0149] ,

[0150] in, The peak amplitude of the pulse;

[0151] (9.4) Summarize the above parameters for all pulses and output the complete feature set matrix. :

[0152] ,

[0153] in, The total number of pulses, with each row representing a complete 13-dimensional acoustic parameter set for one pulse.

[0154] Example 2: Automatic calibration device for bat pulse parameters based on adaptive BO-NM hybrid optimization.

[0155] Reference Figure 3This embodiment includes: a signal preprocessing module 1, a parameter intelligent calibration module 2, a high-precision pulse detection module 3, a CF-FM component segmentation module 4, and a multi-dimensional parameter extraction module 5. The parameter intelligent calibration module 2 includes: a parameter space definition submodule 21, a multi-objective fitness evaluation submodule 22, a dynamic exploration coefficient Bayesian optimization submodule 23, a soft boundary NM optimization submodule 24, and an adaptive perturbation restart control submodule 25; the high-precision pulse detection module 3 includes: a finite state machine detection submodule 31 and an energy adaptive boundary correction submodule 32; the CF-FM component segmentation module 4 includes: a CF frequency anchoring submodule 41 and a bidirectional frequency trajectory tracking submodule 42.

[0156] The working principle of the entire device is as follows:

[0157] The signal preprocessing module 1 is used to acquire the original bat ultrasonic recording data and perform bandpass filtering, Hilbert transform and smoothing on it, and output the smoothed envelope signal and the root mean square value of the envelope to the parameter intelligent calibration module 2.

[0158] The parameter intelligent calibration module 2 receives the smoothed envelope signal and root mean square value of the envelope from the signal preprocessing module 1. It completes parameter optimization through a phased alternating execution of dynamic exploration coefficient Bayesian optimization and normalized spatial soft boundary simplex method. Specifically: the parameter space definition submodule 21 constructs a four-dimensional parameter vector to be optimized and feasible region boundary constraints, consisting of threshold coefficients, minimum pulse width constraints, minimum pulse interval constraints, and FM frequency deviation thresholds, based on the smoothed envelope signal and root mean square value from the signal preprocessing module 1. It then outputs the four-dimensional parameter vector to be optimized and feasible region boundary constraints to the multi-objective fitness evaluation submodule 22. The multi-objective fitness evaluation submodule 22 constructs a weighted fitness function that integrates quantity deviation terms, time-domain anomalies, and statistical stability terms based on the four-dimensional parameter vector to be optimized and feasible region boundary constraints. It then outputs the fitness function to the dynamic exploration coefficient BO submodule 23. The dynamic exploration coefficient BO submodule 23 uses a Gaussian process surrogate model and exponentially decaying exploration coefficients to optimize the parameters based on the fitness function. A global search is performed within the row domain, and the optimal parameter configuration for the Bayesian stage is output to the soft boundary NM optimization submodule 24. The soft boundary NM optimization submodule 24 is used to set the optimal parameter configuration for the Bayesian stage as the initial point, perform Sigmoid soft boundary projection and Nelder-Mead local refinement, and output the optimal parameter configuration for the simplex stage to the perturbation restart control submodule 25. The perturbation restart control submodule 25 is used to compare the fitness values ​​of the optimal parameter configuration for the Bayesian stage from the dynamic exploration coefficient BO submodule 23 and the optimal parameter configuration for the simplex stage from the soft boundary NM optimization submodule 24, and update the global optimal parameter configuration with the better one. It also monitors the update of the global optimal parameter configuration between consecutive rounds. When it is detected that the global optimal parameter configuration has not been updated for several consecutive rounds, a random perturbation is applied to the current global optimal parameter configuration, and the perturbed parameter configuration is output and returned to the dynamic exploration coefficient BO submodule 23 for the next round of optimization. After completing all preset optimization rounds, the global optimal parameter configuration is output to the high-precision pulse detection module 3.

[0159] The high-precision pulse detection module 3 performs pulse detection and boundary refinement based on the globally optimal parameter configuration. Specifically: the finite state machine detection submodule 31 calculates the adaptive detection threshold based on the threshold coefficient, minimum pulse width constraint, and minimum pulse interval constraint from the globally optimal parameter configuration obtained from the intelligent parameter calibration module 2, performs a sample-point-by-sample state transition determination on the smooth envelope signal, and outputs a preliminary pulse list to the energy adaptive boundary correction submodule 32; the energy adaptive boundary correction submodule 32 calculates the energy ratio of each pulse based on the preliminary pulse list from the finite state machine detection submodule 31, dynamically adjusts the boundary detection threshold, performs energy adaptive refinement on the start and end boundaries of each pulse, transmits the output high-precision pulse time interval to the CF-FM component segmentation module 4, and transmits the output energy domain parameters to the multidimensional parameter extraction module 5.

[0160] The CF-FM component segmentation module 4 is used to perform iFM-CF-tFM three-component decoupling on pulses based on high-precision pulse time intervals. Specifically, the CF frequency anchoring submodule 41 is used to perform short-time Fourier transform on each pulse time interval from the high-precision pulse detection module 3, search for the maximum power spectrum value within the CF prior frequency interval, determine the CF main frequency and its time position as anchor points, and output anchor point information to the bidirectional frequency trajectory tracking submodule 42. The bidirectional frequency trajectory tracking submodule 42 is used to track the instantaneous frequency trajectory frame by frame from the anchor point along the time axis, based on the anchor point information from the CF frequency anchoring submodule 41. The boundary position between iFM and tFM is determined using the FM frequency deviation threshold as the judgment criterion, and the time intervals and frequency parameters of the iFM segment, CF segment, and tFM segment of each pulse are output to the multidimensional parameter extraction module 5.

[0161] The multidimensional parameter extraction module 5 is used to summarize the high-precision pulse time interval, energy domain parameters, and time interval and frequency parameters of the three components of iFM-CF-tFM, and output a complete feature set matrix containing acoustic parameters in three dimensions: time domain, frequency domain, and energy domain.

[0162] It should be noted that the above functional modules can be implemented, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as program instruction products. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The program instructions can be stored in a computer-readable and writable storage medium, or transferred from one computer's readable and writable storage medium to another.

[0163] In this embodiment, direct coupling or communication connections between modules can be achieved through indirect coupling or communication connections via interfaces, devices, or modules. The functional modules and sub-modules in this embodiment can be integrated into a single processing unit, or each module can exist physically independently, or two or more modules can be integrated into a single processing unit. When the integrated components are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.

[0164] The effects of this invention can be further illustrated by the following simulation results:

[0165] I. Simulation Conditions

[0166] The hardware environment is as follows: CPU (Intel Core i9-13900K), 64GB of memory.

[0167] The software environment is: Windows 11, MATLAB R2023b.

[0168] Bat recording data acquisition conditions: sampling rate 1MHz, quantization bit depth 16bit, target species is horseshoe bats, CF frequency search range 84-90kHz, recording duration 10-60s / s, signal-to-noise ratio range 10-40dB.

[0169] Parameter configuration: Threshold coefficient search range Minimum pulse width search range ms, minimum interval search range ms, FM frequency deviation from threshold search range Hz. Bayesian optimization iterates 8 times per round, simplex optimization iterates 30 times per round, and the total number of optimization rounds is 3 rounds.

[0170] II. Simulation Content and Results

[0171] Simulation 1: Under the above simulation conditions, pulse detection was performed on the original bat recording data using the present invention and existing manual thresholding and fixed parameter methods. The detection accuracy was compared, and the results are as follows: Figure 4 ,in:

[0172] Figure 4 a) A graph showing the comparison of pulse detection accuracy under different signal-to-noise ratios;

[0173] Figure 4 b) A comparison of the average boundary deviation index under different signal-to-noise ratios.

[0174] from Figure 4As shown in a) and 4b), the pulse detection accuracy of the method of the present invention reaches 97.3%, which is significantly better than the 89.6% of the manual threshold method and the 85.2% of the fixed parameter method. Furthermore, regarding boundary positioning error, the average boundary deviation of the method of the present invention is 0.12 ms, much smaller than the 0.45 ms of the manual threshold method and the 0.63 ms of the fixed parameter method. In other words, the detection accuracy and boundary positioning accuracy of the method of the present invention are superior to the manual threshold method and the fixed parameter method under various signal-to-noise ratio conditions, and the performance gain is more significant with increasing signal-to-noise ratio.

[0175] Simulation 2: Under the above simulation conditions, the original bat recording data was calibrated using the present invention and a single optimization algorithm. The convergence performance of the present invention and the single optimization algorithm was compared, and the results are as follows: Figure 5 .

[0176] from Figure 5 As can be seen, the method of the present invention reaches the global optimum in the 6th round of parameter calibration, and the convergence speed is significantly faster than the average of 12 rounds of pure Bayesian optimization and the average of 15 rounds of pure simplex method, and avoids the problem of existing methods being prone to getting trapped in local extrema.

[0177] Simulation results show that the automatic calibration method and device for bat pulse parameters based on adaptive BO-NM hybrid optimization of the present invention, through the phased alternation of dynamic exploration coefficient Bayesian optimization and normalized spatial soft boundary simplex method, combined with energy adaptive boundary correction and bidirectional frequency tracking CF-FM segmentation technology, can achieve the global optimal automatic configuration of detection parameters and the fully automatic extraction of multi-dimensional acoustic parameters. It effectively solves the problems of parameter calibration relying on manual methods, limited boundary detection accuracy, and low degree of automation in batch extraction in the prior art, and provides support for the analysis of biological sonar mechanisms and the mapping and use of biomimetic radar parameters.

Claims

1. An automatic calibration method for bat pulse parameters based on adaptive BO-NM hybrid optimization, characterized in that, include: (1) Obtain the original bat ultrasound recording data and preprocess it to obtain the smooth envelope signal and the root mean square value of the envelope; (2) Define a four-dimensional parameter vector to be optimized and feasible region boundary constraints based on the smooth envelope signal and the root mean square value of the envelope, and construct a multi-objective weighted fitness function; (3) Using the fitness function as the optimization objective, set the exponentially decaying dynamic exploration coefficient, use this coefficient and the Gaussian process surrogate model to perform a Bayesian global search in the feasible region, and record the optimal parameter configuration in this stage; (4) Using the optimal parameter configuration as the initial point, normalize it to the unit space and perform local refinement of the Nelder-Mead simplex to obtain the optimal parameter configuration for the simplex stage; (5) Select the one with the better fitness value among the two optimal parameter configurations in the above two stages and update its global optimal parameter configuration; (6) Calculate the adaptive detection threshold by using the product of the threshold coefficient obtained by the global optimal parameter configuration and the root mean square value of the envelope, and drive the finite state machine to complete the initial pulse detection by combining the minimum pulse width constraint and the minimum pulse interval constraint. (7) The detection threshold is dynamically adjusted according to the energy ratio of each pulse, and the start and end boundaries of each pulse obtained from the preliminary detection are refined by energy self-adaptation to obtain a high-precision pulse time interval; (8) Anchor the CF main frequency in each pulse time interval and track the instantaneous frequency trajectory bidirectionally along the time axis, and use the FM frequency deviation threshold as the judgment criterion to complete the decoupling of the iFM-CF-tFM three components; (9) Summarize the decoupling results of the pulse time interval and the three components of iFM-CF-tFM, and automatically output the complete feature set matrix of multi-dimensional acoustic parameters.

2. The method according to claim 1, characterized in that, The process of acquiring and preprocessing the raw bat ultrasonic recording data in (1) includes: 1a) Perform Chebyshev type I bandpass filtering on the raw recording data and extract the target frequency band signal with the target species CF main frequency as the center frequency; 1b) Perform Hilbert transform on the filtered signal to obtain the analytic signal, take its instantaneous amplitude envelope and smooth it with the moving average to obtain the smoothed envelope signal; 1c) Calculate the root mean square value of the smoothed envelope signal as an adaptive threshold reference.

3. The method according to claim 1, characterized in that, In step (2), a four-dimensional parameter vector to be optimized and feasible region boundary constraints are defined based on the smooth envelope signal and the root mean square value of the envelope, and a multi-objective weighted fitness function is constructed, which includes: 2a) Defined by threshold coefficient Minimum pulse width constraint Minimum pulse interval constraint and FM frequency deviation threshold The four-dimensional parameter vector to be optimized And set the upper bound of the feasible region for each component. and the lower realm ; 2b) Calculate the normalized deviation term used to quantify the number of tests. : , in, To detect the number of pulses, The a priori expected number of pulses; 2c) Calculate the proportional temporal anomaly used to quantify the deviation of pulse duration from physiological constraints. : , in, The number of pulses whose duration exceeds the physiologically reasonable range; 2d) Calculate the statistical stability term used to quantify the statistical dispersion of the test results. : , in, This is the duration sequence of all detected pulses. To prevent division by zero; 2e) Adjust the above quantity deviation items according to the weighting coefficients. Time-domain anomalies and statistical stability term The weighted sum of these three terms yields the comprehensive fitness function: , in, , , These are the weighting coefficients for the quantity deviation term, the time-domain anomaly term, and the statistical stability term, respectively.

4. The method according to claim 1, characterized in that, The process (3) utilizes exponentially decaying dynamic exploration coefficients and a Gaussian process surrogate model to perform a Bayesian global search within the feasible region, which includes: 3a) Calculate the exploration coefficients for the current iteration. : , in, This is the initial exploration coefficient. To minimize the exploration coefficient, The attenuation rate, This represents the current iteration number. This represents the total number of iterations. 3b) Construct a Gaussian process proxy model using existing evaluation points, improve the acquisition function based on expectations, and adjust the trade-off between exploration and utilization using the current exploration coefficients, and select the next evaluation point in the parameter space; 3c) Perform fast impulse detection at candidate evaluation points and calculate fitness values. Add the new evaluation points and their fitness values ​​to the observation set to update the posterior distribution of the surrogate model. 3d) Repeat steps 3a) to 3c) until the preset number of iterations is completed, and record the optimal parameter configuration for this stage.

5. The method according to claim 1, characterized in that, The process in (4) involves normalizing the optimal parameter configuration to a unit space for local refinement of the Nelder-Mead simplex, which includes: 4a) Normalize the parameter vector to the interval [0,1] according to the upper and lower bounds of the feasible region. ; 4b) Perform a Sigmoid soft-boundary projection on the normalized parameters to obtain the projected Sigmoid parameter vector. : , in, For steepness parameters, This is the upper bound of the feasible region. This is the lower bound of the feasible region. 4c) will The amplitude is limited to the [0,1] interval, and finally denormalized to the original parameter space. ; 4c) In the normalized space, with the optimal parameter configuration of the Bayesian optimization stage as the initial simplex, perform Nelder-Mead search, and then project the search results onto the Sigmoid and back-normalize them to the original parameter space to obtain the optimal parameter configuration of the simplex stage.

6. The method according to claim 1, characterized in that, The globally optimal parameter configuration for updating the fitness value of the better candidate in (5) includes: 5a) Compare the fitness value of the optimal parameter configuration obtained in the current round with the recorded global optimal parameter configuration. If the fitness value of the current round is better, update the global optimal parameter configuration; otherwise, keep the global optimal parameter configuration unchanged. 5b) Monitor the update status of the global optimal parameter configuration between consecutive rounds. If the global optimal parameter configuration is not updated for two consecutive rounds, it is determined that the optimization has stalled. 5c) Determine if the current optimization is stalled: If the optimization is stalled, a random perturbation is applied to the current global optimal parameter configuration. The perturbation amplitude is based on 0.1 and increases linearly with the number of stall rounds. The perturbed parameter configuration is then used as the initial parameter configuration for the next round. If the optimization is not stalled, the current globally optimal parameter configuration will be used as the initial parameter configuration for the next round. 5d) Determine if all preset optimization rounds have been completed: If not completed, the Bayesian global search in (3) and the simplex local refinement in (4) will be re-executed with the initial parameter configuration obtained in 5c), and then 5a) will be returned to enter the next round of update judgment; If all preset optimization rounds have been completed, the globally optimal parameter configuration will be output.

7. The method according to claim 1, characterized in that, The preliminary pulse detection in (6) is achieved by using an adaptive detection threshold combined with minimum pulse width constraints and minimum pulse interval constraints to drive a finite state machine. This includes: 6a) Define the three states of a finite state machine: silent, impulsive, and interval. 6b) In the silent state, continuously monitor the envelope value of the current sampling point to wait for the arrival of a valid pulse; if the envelope value of the current sampling point exceeds the adaptive detection threshold and the interval with the end position of the previous pulse is greater than the minimum pulse interval constraint, then switch to the pulse in-pulse state and record the pulse start candidate point. 6c) In the pulse state, track the duration of the current pulse and accumulate the pulse duration; if the envelope value falls below the threshold, check whether the pulse duration meets the minimum pulse width constraint. If it does, it is confirmed as a valid pulse and transferred to the interval state; if it does not meet the constraint, it is judged as a false alarm and returns to the silent state. 6d) In the interval state, block brief disturbances immediately following the valid pulse to prevent the same pulse from being triggered repeatedly; if the interval between the current sampling point and the confirmed pulse end position exceeds the minimum pulse interval constraint, return to the silent state. 6e) Repeat steps 6b) and 6d) one sample point at a time until all signals have been processed and a preliminary pulse list is output.

8. The method according to claim 1, characterized in that, The process in (7) involves dynamically adjusting the detection threshold based on the energy ratio of each pulse and performing energy-adaptive refinement on the start and end boundaries of each pulse obtained from the initial detection. This includes: 7a) Calculate the energy ratio of each pulse. ,in For the first The peak envelope of each pulse. This is the root mean square value of the envelope; 7b) Dynamically determine the detection threshold ratio based on the energy ratio: when the energy ratio is greater than the preset threshold, the detection threshold ratio increases linearly with the energy ratio between the base value and the upper limit value; otherwise, it remains at the base value. 7c) Search backward from the initial detection starting point to locate the point where the envelope first falls below the preset percentage of the peak value as the precise starting position of the pulse; 7d) Search along the time axis from the peak position, and determine the precise end position of the pulse when the envelope value is lower than the dynamic threshold ratio of the peak value or the decay exceeds the preset amplitude within a short time window; 7e) If the pulse duration after refinement is less than the minimum protection pulse width, the initial detection boundary shall be retained.

9. The method according to claim 1, characterized in that, The decoupling of the iFM-CF-tFM three components in (8) is completed using the FM frequency deviation threshold as the criterion, which includes: 8a) Perform a short-time Fourier transform on each pulse time interval, search for the maximum power spectrum in the prior frequency interval of CF, and determine the dominant frequency of CF and its time position as anchor points. 8b) Extract the peak frequency frame by frame from the anchor point along the time axis. When the deviation from the CF main frequency exceeds the FM frequency deviation threshold, determine the iFM end position and search for the maximum bandwidth point within the analysis window above the power threshold to determine the iFM start frequency. 8c) Extract the peak frequency frame by frame from the anchor point along the time axis. When the deviation exceeds the FM frequency deviation threshold, determine the tFM start position and continue tracking to the pulse end boundary to determine the tFM termination frequency. 8d) Output the time interval and frequency parameters of the iFM segment, CF segment, and tFM segment of each pulse.

10. An automatic calibration device for bat pulse parameters based on adaptive BO-NM hybrid optimization, characterized in that, include: The signal preprocessing module is used to perform bandpass filtering, Hilbert transform, and short-time Fourier transform on the raw bat ultrasound recording data, and output a smooth envelope signal and root mean square value of the envelope. The parameter intelligent calibration module is used to define the four-dimensional parameter vector to be optimized and the fitness function. It outputs the globally optimal parameter configuration by dynamically exploring the coefficient Bayesian optimization and the normalized spatial soft boundary simplex method in stages, combined with the perturbation restart mechanism. The high-precision pulse detection module is used to drive a finite state machine to complete the initial pulse detection using the globally optimal parameter configuration, and to perform start and end boundary refinement by dynamically adjusting the detection threshold according to the energy ratio. The CF-FM component segmentation module is used to anchor the CF main frequency and track the instantaneous frequency trajectory bidirectionally within each pulse time interval, thereby completing the decoupling of the iFM-CF-tFM three components. The multidimensional parameter extraction module is used to summarize the decoupling results of the pulse time interval and the three components of iFM-CF-tFM, and automatically output the complete feature set matrix of multidimensional acoustic parameters.

11. The apparatus according to claim 10, characterized in that, The intelligent parameter calibration module includes: The parameter space definition submodule is used to construct a four-dimensional parameter vector to be optimized and feasible region boundary constraints; The multi-objective fitness evaluation submodule is used to construct a weighted fitness function that integrates quantitative bias, temporal anomaly, and statistical stability terms. The dynamic exploration coefficient Bayesian optimization submodule is used to perform a global search using a Gaussian process surrogate model and exponentially decaying exploration coefficients; The soft boundary NM optimization submodule is used to implement soft boundary projection and perform Nelder-Mead local refinement through the Sigmoid function; The adaptive disturbance restart control submodule is used to monitor the optimization stagnation state and apply random disturbances to escape local extrema.

12. The apparatus according to claim 10, characterized in that: The high-precision pulse detection module includes: The finite state machine detection submodule is used to perform state transition determination on a sample-by-sample point basis with adaptive detection threshold and validity constraints, and output a preliminary pulse list. The energy adaptive boundary correction submodule is used to dynamically adjust the boundary detection threshold according to the energy ratio of each pulse and to refine the start and end boundaries. The CF-FM component segmentation module includes: The CF frequency anchoring submodule is used to search for the power spectrum maxima within the CF prior frequency range to determine the CF dominant frequency and time position. The bidirectional frequency trajectory tracking submodule is used to track the instantaneous frequency trajectory bidirectionally along the time axis from the CF anchor point, and determine the boundary positions of iFM and tFM based on the FM frequency deviation threshold.

Images