Fish school feeding intensity analysis method based on electromagnetic wave spectrum component decoupling optimization

Abstract

The application discloses a fish feeding intensity analysis method and system based on electromagnetic wave spectrum decoupling optimization. The method first collects monitoring and reference channel radar signals of a target water area and carries out direct current preprocessing, then converts time domain echoes into frequency domain spectrum fingerprints, and divides a Doppler frequency domain space into water flow low frequency interference and equipment and biological coexistence space; then, common mechanical strong interference peaks are removed by subtracting spatial reference spectrum, and through multi-peak value search and space-time stability verification, pure biological activity spectrum is deeply decoupled and purified; further, energy density, normalized spectrum entropy and energy dispersion ratio are extracted in the pure spectrum to construct a multi-dimensional feature vector; finally, time sliding window and accumulation judgment logic are introduced to eliminate accidental jumping and other pseudo-target interference, and accurate fish feeding intensity analysis results are output. The application effectively removes complex environmental and mechanical background noise, and realizes all-weather, non-contact structured semantic precise mapping of fish feeding intensity.

Description

Technical Field

[0001] This invention belongs to the fields of smart fisheries, modern aquaculture information monitoring, and radar signal processing technology. It relates to a method and system for analyzing fish feeding intensity based on electromagnetic wave spectrum component decoupling optimization. Specifically, it relates to a method and system for analyzing fish feeding intensity in aquaculture scenarios based on spatiotemporal collaborative sensing of dual-channel millimeter-wave radar and electromagnetic wave signal spectrum component decoupling optimization. Background Technology

[0002] With the rapid development of intensive, large-scale, and factory-style recirculating aquaculture (RAS), all-weather, non-contact real-time monitoring and quantitative characterization of fish feeding behavior has become a core technological approach to achieve precise automatic feeding, reduce feed conversion ratio (FCR), and avoid secondary water pollution in smart fisheries.

[0003] Currently, biological behavior monitoring in aquaculture scenarios mainly relies on computer vision recognition technology and traditional single-channel microwave / millimeter-wave radar detection technology. However, in the actual physical environment of modern high-density intensive aquaculture, the above-mentioned existing technologies all have significant performance bottlenecks and application defects: First, conventional computer vision recognition methods (such as monitoring systems based on the YOLO deep learning object detection network architecture) are extremely dependent on lighting conditions. They cannot function properly in indoor environments with no light, at night, or in low light. At the same time, the high-power oxygenators in intensive aquaculture tanks operate year-round, resulting in a large number of high-density, non-periodic bursting bubbles, water mist, and highly turbid scattering bodies on the water surface. This environmental noise can easily cause large-area occlusion and false targets in images, significantly reducing the accuracy of fish identification and feeding recall rates of visual algorithms. Furthermore, the high energy consumption of hardware computing power also limits its low-cost deployment on edge embedded platforms.

[0004] Secondly, while traditional single-channel millimeter-wave radar threshold detection methods offer advantages such as all-weather operation and independence from ambient light, in complex factory aquaculture workshops, the physical vibrations and fixed-frequency broadband noise generated by rotating machinery such as water pumps and aerators directly translate into high-intensity Doppler frequency shift components. Because only a single-channel radar is used for vertical or oblique illumination monitoring, the fixed-frequency, narrow-band, strong Doppler interference generated by these rotating machines often experiences severe overlap and spectral masking effects in the frequency domain with the dynamic surface scattering spectrum caused by fish feeding. Under this single-channel architecture, simply relying on whether the radar intensity exceeds a fixed or dynamic gain threshold can easily lead to the mechanical operation being mistaken for "fish feeding," resulting in frequent false positives and mis-triggered events.

[0005] In summary, existing single-channel radar monitoring systems, lacking a physical spatial reference, cannot effectively separate and decouple "mechanical vibration / water flow background interference" from "pure biological feeding disturbance" at the algorithmic level. Therefore, achieving accurate semantic representation of fish feeding with ultra-low latency and high signal-to-noise ratio in extremely intensive aquaculture scenarios with strong mechanical noise and multiple bubble interference is a pressing technical challenge in the field of intelligent fishery equipment. Summary of the Invention

[0006] To achieve the above objectives, this invention proposes a method and system for analyzing fish feeding intensity in aquaculture scenarios based on spatiotemporal collaborative sensing of dual-channel millimeter-wave radar and decoupling optimization of electromagnetic wave signal spectral components.

[0007] The technical solution adopted by the method of the present invention is: a method for analyzing the feeding intensity of fish groups based on electromagnetic wave spectral component decoupling optimization, characterized by comprising the following steps: Step 1: Collect signals from the monitoring channel radar and reference channel radar deployed in the target aquaculture area to obtain the raw monitoring intermediate frequency signal. and the original reference intermediate frequency signal ; Step 2: For the above and After preprocessing, the DC-free monitoring differential signal is obtained. Differential signal with reference ; Step 3: Separately... and The time-domain echo is converted into a frequency-domain spectral fingerprint reflecting the energy density distribution with frequency, thus obtaining the real-time monitoring spectral fingerprint Fk. Mon Background fingerprint Fk of the reference environment Ref ; Step 4: Based on the frequency unit index Using the same frequency boundary rules for each and The Doppler frequency domain space is divided into two physical characteristic spaces: the low-frequency interference space of hydrodynamics and the coexistence space of equipment and biological activities. Step 5: Based on spatial reference spectrum subtraction, perform spectral fingerprint offsetting in the same frequency domain for the space where the device and biological activity coexist, physically eliminating strong mechanical Doppler interference peaks common to both the monitoring channel and the reference channel; physically isolate the low-frequency interference space of hydrodynamics directly through frequency band stopbands; finally, summarize the frequency domain components after processing the two spaces to obtain the residual spectral fingerprint after spectral subtraction and isolation. ; Step 6: Extract the residual spectral fingerprint after spectral subtraction. By performing multi-peak search and spatiotemporal stability verification, a pure biological activity spectrum, after removing all environmental and mechanical background, is decoupled and purified. ; Step 7: For the pure biological activity spectrum Fk, extract three key physical characterization parameters to construct a multidimensional feature vector V = [ , H, S], Here, H is the energy density parameter, H is the normalized spectral entropy parameter, and S is the energy dispersion ratio parameter. Step 8: Set V = [ The algorithm introduces a time-dimensional accumulation judgment logic. By setting a time-sliding observation window W, it eliminates false targets caused by accidental leaps of fish or sudden environmental changes in a single frame, and outputs the analysis results of fish feeding intensity.

[0008] Preferably, in step 1, the monitoring channel radar is vertically aligned with the water surface of the feeding activity area to scan and acquire the raw monitoring intermediate frequency signal, which includes fish feeding disturbances, static background, and mechanical point source interference. The reference channel radar scans a non-feeding reference area far from the active feeding area and free from fish activity. It maintains the same parameter configuration and synchronous sampling as the monitoring channel radar to acquire the original reference intermediate frequency signal, which includes a purely static background and mechanical point source interference. .

[0009] Preferably, the step 2 described above... and Preprocessing involves establishing buffer queues with a sliding window length of N for each of the two original intermediate frequency signals and calculating the statistical mean of the complex-domain signals within the queues. A mean subtraction operation is then performed to eliminate the zero-Doppler DC component generated by the tank walls and the static water surface. Subsequently, a Hanning window is applied to suppress spectral energy leakage, resulting in the DC-free monitoring differential signal. Differential signal with reference Where N is a preset value; The mean subtraction operation will subtract the time-domain complex signal input in the current frame. Subtracting the statistical mean within window N completes the mean subtraction operation, thereby eliminating the zero Doppler DC component generated by the aquaculture tank wall and the static water surface, resulting in the intermediate-state de-DC time-domain signal. Subsequently, Hanning windowing is applied to this de-DC time-domain signal to suppress spectral energy leakage. At this point, the final time-domain result output by the windowed signal is compared with the corresponding dynamic background estimate. Perform difference subtraction in the complex field; where... , This is the dynamic background estimate for the current frame. The background estimate is the value of the previous frame, and β is a preset smoothing factor that satisfies the interval [missing information]. ; The specific formula for the Hanning window addition process is as follows:

[0010] ; ; In the formula, The index of discrete sampling points in the time domain and satisfying ; and These are the time-domain complex signals of the monitoring channel and the reference channel after performing the differential subtraction in the current frame, respectively. The length is the same as the length of the cache queue. Strictly consistent discrete Hanning window function; the final obtained and That is, the monitoring differential signal and the reference differential signal after DC removal are output to the subsequent frequency domain spectral fingerprint conversion step.

[0011] Preferably, in step 3, the... and Perform 512-point Fast Fourier Transform (FFT) to convert the time-domain echo into a frequency-domain spectral fingerprint that reflects the energy density distribution with frequency.

[0012] Preferably, in step 4, the frequency unit index To obtain the frequency domain discrete spectral point indexes after performing Fast Fourier Transform (FFT) on the preprocessed monitoring differential signal and the reference differential signal, and satisfying the following conditions: If the frequency unit index satisfy Then and The corresponding amplitude components are divided into a low-frequency interference space for hydrodynamics; if the frequency unit index satisfy Then and The corresponding amplitude components are divided into spaces where equipment and biological activities coexist.

[0013] Preferably, in step 5, the energy ratio of the two-channel spectral fingerprints in the low-frequency interference space is calculated, and the spatial propagation and hardware gain cutoff factor are established. , of which molecules With denominator The accumulation intervals are all frequency unit indices 0 ≤ k < 15; within the coexistence space 16 ≤ k ≤ 255, the background spectral fingerprint of the reference environment is used. As a dynamic background barrier, it monitors spectral fingerprints. Perform adaptive spatial spectral subtraction to obtain the residual spectral fingerprint after spectral subtraction. , where α is the preset lower limit gain compensation factor of the spectrum.

[0014] Preferably, in step 6, the local energy maxima are located using a sliding window maximum value filtering algorithm, and their frequency indices are recorded. Peak amplitude and full width at half maximum (FWHM) Simultaneously maintain a length of The feature tracking queue is used to calculate the maximum point in a continuous sequence. Standard deviation of center frequency within a frame and the coefficient of variation of amplitude ;in, This is a preset value; when it meets the requirements... and When the residual narrowband spike is determined to be mechanical residual interference at a fixed frequency, the frequency range is locked and a mechanical interference mask range is established. Finally, the mechanical interference mask area is utilized. left edge point and right edge point The energy value is linearly interpolated to reconstruct and replace the mask interval. The background baseline energy within the spectrum is used to decouple and purify the pure biological activity spectrum after removing all environmental and mechanical backgrounds. .

[0015] Preferably, in step 7... This is used to characterize the overall intensity of biological disturbance to aquaculture water bodies caused by fish populations, among which The increase in the value is positively correlated with the intensity of the fish's tail-wagging competition; at the same time, the preprocessed original reference intermediate frequency signal is added to the frequency unit index. Low-frequency integral energy of the interval And through the formula Mapping calculation of dynamic electromagnetic attenuation factor ,in The background energy of a pre-defined static standard water body is obtained by collecting radar signals in a pure, still water environment without fish or mechanical operation and calculating the corresponding low-frequency integral mean; finally, the formula is used... The corrected energy density parameter is obtained by performing exponential logarithmic adaptive gain compensation on the energy density parameter. This is used to physically eliminate electromagnetic media attenuation disturbances caused by large dynamic fluctuations in low-frequency water flow. ;in The energy probability density mapping at each frequency point is used to quantify the disorder and randomness of water body fluctuations. , used to quantitatively describe the degree of energy dispersion broadening along the Doppler frequency domain axis; where, when When the frequency is less than the preset dispersion threshold, it indicates that there are extremely high peaks in the spectrum and the energy in other frequency bands is weak, which is judged as "point-source mechanical residual vibration" with highly concentrated energy; if When the value is greater than or equal to the preset dispersion threshold, it indicates that the overall spectrum is raised and there are no specific abrupt narrowband strong spikes, which is determined to be "area source biological community activity" with energy bursting across the entire frequency band.

[0016] Preferably, in step 8, if the feature parameters extracted in real time simultaneously satisfy the energy density within the continuous observation period... And normalized spectral entropy If the condition is met, it is considered a calm state; among which... , The preset static judgment threshold; Before performing spectral subtraction and neighborhood linear interpolation reconstruction, if the feature tracking queue identifies stable narrowband Doppler spike components and utilizes the original monitoring spectral fingerprint... The ratio of the total energy across the entire frequency band to the maximum peak energy is used to calculate the initial energy distribution ratio before decoupling, which satisfies... This indicates the presence of highly concentrated, strong mechanical peaks in the original spectrum, suggesting that the water body is currently in the operation state of an oxygenator or water pump rotating machinery; among which... The preset threshold for narrowband mechanical interference identification; If the pure biological activity spectrum is reconstructed through decoupling purification, spectral subtraction, and linear interpolation within the mask interval... The corresponding characteristic parameters simultaneously satisfy the modified energy density. Normalized spectral entropy And the energy distribution after decoupling is higher than that of decoupling. This indicates that the residual energy after spectral subtraction is significantly increased, the fluctuation disorder is extremely high, and it exhibits broadband surface source dispersion broadening characteristics on the Doppler frequency axis, and within the time-sliding observation window... If the percentage of valid frames meeting the above conditions exceeds a preset threshold, then occasional mutation false targets are immediately eliminated, and the fish are identified as feeding in a school of fish; , , This is the preset food intake activation threshold.

[0017] The technical solution adopted by the system of the present invention is: a fish feeding intensity analysis system based on electromagnetic wave spectral component decoupling optimization, comprising: One or more processors; A storage device, configured to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors are caused to implement the fish school feeding intensity analysis method based on decoupling optimization of electromagnetic wave spectrum components.

[0018] Compared with the prior art, the beneficial effects of the present invention include: (1) Complete decoupling of physical mechanical interference and biological activity characteristics is achieved, and the typical mechanical noise rejection ratio reaches 28.7 dB.

[0019] The present invention overcomes the drawback that the conventional single-channel radar detection system lacks a physical space reference benchmark, and innovatively proposes a spatial spectrum subtraction and neighborhood linear interpolation reconstruction algorithm based on spatio-temporal collaborative sensing of dual channels (monitoring channel and reference channel). Experimental data show that in the extremely complex intensive aquaculture environment where an oxygen generator (50W) and a water pump (30W) operate year-round and the water body contains a large number of non-periodically ruptured bubbles, the present invention can accurately lock fixed-frequency and narrow-band strong Doppler interference peaks caused by rotating machinery, and achieve a typical mechanical noise rejection ratio as high as 28.7dB. The overall signal-to-noise ratio (SNR) of the system is improved by more than 15dB compared with the conventional un-decoupled single-channel algorithm, fundamentally eliminating the spectral masking effect caused by mechanical vibration or drastic fluctuation of water flow.

[0020] (2) The present invention significantly improves the semantic representation accuracy and robustness of fish school feeding behavior, and the recognition accuracy of feeding scramble state reaches 98.1%.

[0021] By introducing a multi-dimensional characterization vector of normalized spectral entropy H and energy dispersion ratio S (i.e., SDR), the present invention can accurately distinguish between "point-source mechanical residual vibration" (energy is highly concentrated at a single frequency point, S<10 before decoupling) and "areal-source biological population activity" (energy erupts in the whole frequency band, Doppler frequency shift is randomly broadened, H>5.5 and S>40 in the feeding state). Combined with the joint constraint of the sliding time window integral decision matrix, false targets caused by accidental jumping of fish schools or single-frame environmental mutation are completely eliminated. Actual measurement results in an intensive high-density aquaculture environment show that the recognition accuracy of feeding behavior of the present invention is greatly increased from about 60% of the conventional method to 94.8%, and is as high as 98.1% in the dual-channel cooperative working mode, with a recall rate reaching 96.4%.

[0022] (3) The present invention has an adaptive electromagnetic attenuation gain compensation capability, and completely eliminates the influence of high turbidity and high-density bubbles on behavior quantification.

[0023] The present invention fully utilizes the low-frequency interference space of hydrodynamics where the frequency unit index satisfies 0<k<15, and extracts the low-frequency spectrum energy integral value in real time Dynamic mapping is used to calculate the electromagnetic wave attenuation factor caused by high-density bursting bubbles and turbidity in the current water body. Furthermore, an exponential gain compensation correction is applied to the core energy density characteristics. This mechanism effectively corrects the propagation loss of electromagnetic waves in medium-to-high flow velocities and multi-bubble water bodies, ensuring the spatiotemporal consistency of feature vectors and the high robustness of semantic mapping under different aquaculture densities and water quality conditions.

[0024] (4) The algorithm of this invention has low complexity and the single-frame processing delay at the edge is only 7.5 ms, which perfectly supports all-weather ultra-real-time closed-loop feeding control.

[0025] This invention, based on a clear physical dual-channel spectral subtraction and frequency domain reconstruction logic, avoids the strong dependence of conventional computer vision solutions (such as YOLO and other deep learning object detection networks) on high-performance graphics cards (GPUs) and their failure in low-light or foggy environments. The algorithm proposed in this invention has a single-frame full-processing latency (including dual-channel 512-point FFT, spatial spectral subtraction, feature extraction, and multi-dimensional semantic decision matrix) of only 7.5ms, far lower than the frame acquisition cycle (50ms) of millimeter-wave radar. The system has low hardware cost and minimal computing power overhead, and can run directly in a closed loop on low-power advanced embedded microprocessors (MCUs) or edge computing platforms, providing a highly feasible closed-loop decision-making device for fully automated and precise feeding control of circulating water in factories. Attached Figure Description

[0026] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0027] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the dual-channel radar spatial collaborative sensing deployment in an embodiment of the present invention; Figure 3 This is a flowchart of the entire process of the mechanical interference component decoupling and stripping algorithm based on spatial reference spectrum subtraction in this embodiment of the invention. Figure 4 This is a schematic diagram illustrating the frequency unit interval division and feature region definition of the dual-channel Doppler bispectral space in an embodiment of the present invention; Figure 5 This is a 3D waterfall spectrum diagram of the original monitoring channel radar echo containing strong rotating mechanical interference when dual-channel decoupling is not performed in this embodiment of the invention. Figure 6This is a Doppler 3D waterfall spectrum of pure biological activity (fish feeding behavior) successfully decoupled and purified after spatial spectrum subtraction and neighborhood reconstruction in step 3 of this invention. Figure 7 This is a state mapping logic diagram of the classification decision of the multidimensional anti-interference feature vector V based on spatiotemporal joint constraints in an embodiment of the present invention within a time-sliding observation window. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] Please see Figure 1 This embodiment provides a method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization, including the following steps: Step 1: Collect signals from the monitoring channel radar and reference channel radar deployed in the target aquaculture area to obtain the raw monitoring intermediate frequency signal. and the original reference intermediate frequency signal ; In one implementation, please see Figure 2 The monitoring channel radar is vertically aligned with the water surface of the feeding activity area to scan and acquire raw monitoring intermediate frequency signals that include fish feeding disturbances, static background, and mechanical point source interference. The reference channel radar is used to scan a non-feeding reference area far from the active feeding area and where there is no fish activity, to obtain the raw reference intermediate frequency signal containing pure static background and mechanical point source interference. Please see Figure 4 .

[0030] Step 2: For the above and After preprocessing, the DC-free monitoring differential signal is obtained. Differential signal with reference ; Please see Figure 3 In one implementation, buffer queues with a sliding window length of N are established for the two original intermediate frequency signals, and the statistical mean of the time-domain complex domain signals within the queues is calculated. A mean subtraction operation is performed to eliminate the zero Doppler DC component generated by the aquaculture tank wall and the static water surface. Subsequently, a Hanning window is applied to suppress spectral energy leakage, resulting in the DC-free monitoring differential signal. Differential signal with reference Where N is a preset value; the mean subtraction operation is to subtract the time-domain complex signal input in the current frame. With the corresponding dynamic background estimate Perform difference subtraction in the complex field; where... , This is the dynamic background estimate for the current frame. The background estimate is the value of the previous frame, and β is a preset smoothing factor that satisfies the interval [missing information]. .

[0031] In one implementation, a background estimation method based on exponentially weighted moving average (EWMA) is used to establish time-varying dynamic background models for the time-domain echoes of the monitoring channel and the reference channel, respectively. The adaptive iterative update formulas for the background echo models of the two channels are as follows: ; ; in, and These are the dynamic background estimates for the monitoring channel and the reference channel in the current t-th frame, respectively. and For the previous one The dynamic background estimate of the frame. and Given the current t-th frame, the input is a time-domain complex signal sequence, where β is a preset smoothing factor and satisfies the weighting interval. .

[0032] Among them, the original monitoring intermediate frequency signal of the feeding active area was collected. The system first dynamically monitors the peak amplitude, and adaptively reduces the intermediate frequency gain if it approaches the saturation voltage. Then, a sliding window buffer queue is established for the monitoring channel, the time-domain statistical mean within the queue is calculated, and an adaptive iterative formula based on exponentially weighted moving average (EWMA) is used to independently establish a time-varying dynamic background model for the monitoring channel. The monitoring signal of the current frame is subtracted from the background model in the complex domain to filter out zero-Doppler DC components such as the aquaculture tank wall and static water surface. Finally, the signal is smoothed by a Hanning window to output the DC-free monitoring differential signal. .

[0033] Within the same sampling period, the original reference intermediate frequency signal is simultaneously acquired in a reference area completely free of fish activity. Similarly, a strictly consistent sliding window buffer queue is established and its mean is subtracted. The same adaptive iterative formula is then used to independently build a time-varying dynamic background model for the reference channel. The reference signal of the current frame is subtracted from the reference background model in the complex domain to remove the scattered waves from the fixed assets in the reference area. After smoothing with a Hanning window of the same type, the DC-free reference differential signal is output. .

[0034] The monitoring differential signal is obtained by performing a differential subtraction in the complex domain between the time-domain input signal of the current t-th frame and the corresponding time-varying dynamic background estimate. Differential signal with reference This method filters out static scattering noise caused by fixed objects such as the walls and supports of aquaculture tanks from the physical time domain level, thereby improving the significance of echoes from moving targets.

[0035] In one implementation, the specific formula for adding a window to the Hanning window is as follows:

[0036] ; ; In the formula, The index of discrete sampling points in the time domain and satisfying ; and These are the time-domain complex signals of the monitoring channel and the reference channel after performing the differential subtraction in the current frame, respectively. The length is the same as the length of the cache queue. Strictly consistent discrete Hanning window function; the final obtained and That is, the monitoring differential signal and the reference differential signal after DC removal are output to the subsequent frequency domain spectral fingerprint conversion step.

[0037] In one implementation, the receive gain closed-loop calibration is performed during system operation by monitoring the raw intermediate frequency signal of the monitoring channel radar in real time through a sliding observation window. peak amplitude The full-scale voltage threshold of the analog-to-digital converter (ADC) is When the system detects that the sampled data meets the voltage out-of-bounds criterion Determining that the current feeding area is at risk of saturation clipping due to violent fish jumping or strong mechanical reflections, the system's main control unit dynamically reduces the intermediate frequency gain of the microwave front-end. This is to prevent high-frequency harmonic distortion and spectral uplift caused by time-domain signal truncation.

[0038] Step 3: For the above and Perform Fast Fourier Transform (FFT) to convert the time-domain echo into a frequency-domain spectral fingerprint reflecting the energy density distribution with frequency, thus obtaining the real-time monitoring spectral fingerprint Fk. Mon Background fingerprint Fk of the reference environment Ref ; In one implementation, the dual-channel frequency domain spectral fingerprint independent conversion performs a Fast Fourier Transform (1D-FFT) on the two independently output time-domain differential signals from the aforementioned steps, mapping the time-domain echoes into a frequency domain spectral fingerprint reflecting the energy density distribution with frequency, wherein: the monitoring differential signal After FFT transformation, the real-time monitoring spectral fingerprint is output. Reference differential signal After FFT transformation, the background spectral fingerprint of the reference environment is output. The discrete spectral points (frequency unit indices) of the frequency domain axes of both are uniformly represented as: .

[0039] Step 4: Based on the frequency unit index Using the same frequency boundary rules for each and The Doppler frequency domain space is divided into two parts: the low-frequency interference space of hydrodynamics and the space where equipment and biological activities coexist.

[0040] Please see Figure 4 In one implementation, the monitoring spectral fingerprints processed in parallel and independently in the aforementioned steps are... With reference spectral fingerprint This approach integrates three stages—homogeneous spatial alignment, adaptive spectral subtraction hedging, and neighborhood mask reconstruction—to achieve deep decoupling between spatial hedging and mechanical noise. The specific integration flow is as follows: (1) First stage: Homogeneous feature space partitioning and alignment In order to enable the data of two independent channels to be quantized and compared under the same physical feature dimension, this embodiment is based on the frequency unit index. The value of affects the monitoring spectrum. With reference spectrum Implementing identical boundary rules for delineation, and uniformly aligning the definition to two layers of space: one is... The low-frequency interference space of hydrodynamics; secondly The space where equipment and biological activities coexist.

[0041] (2) Second stage: Dual-channel data subtraction and integration (adaptive spatial spectrum subtraction) Under the aligned spatial dimension, the two data streams are formally integrated through physical subtraction. First, the low-frequency interference space of water flow is utilized ( The overall energy ratio of the two-channel spectral fingerprints is used to calculate the propagation and gain cutoff factors between the two channels. : ; Furthermore, in the space of coexistence ( Within ) the reference spectrum As a background barrier for dynamic environments, the monitoring spectrum Perform differential subtraction hedging. In this step, the two data streams are combined into one, and the residual spectral fingerprint after the output spectrum is subtracted is integrated. : ; In the formula, This is the preset lower limit gain compensation factor for the spectrum. Through this subtraction integration, the strong Doppler clutter spikes common to both channels, caused by mechanical rotation, are physically offset and eliminated.

[0042] (3) Third stage: Narrowband mask of residual mechanical noise and linear interpolation are integrated into the two-channel data fusion. Based on this, to eliminate residual fixed-frequency mechanical components due to minor hardware differences, the system undergoes deep purification by incorporating spatiotemporal stability verification: establishing a length of The temporal feature tracking queue is determined by the standard deviation of the center frequency of local maxima points within consecutive frames. and amplitude variation coefficient The joint convergence determination accurately identifies the remaining fixed-frequency mechanical residual indexes that have not been completely removed. and half-peak full width Subsequently, a mechanical interference mask area is automatically established. Finally, using the mask interval The true energy values ​​at the left and right edge points are linearly interpolated at two points to reconstruct and "flatten" all background energy within the interval.

[0043] After the above three stages of integration and deep decoupling, the two original radar data streams were finally completely free of environmental and mechanical background noise, and the combined output was a pure biological activity spectrum. This is used for extracting the multidimensional feature vector of fish feeding in subsequent steps.

[0044] Step 5: Based on the subtraction of the spatial reference spectrum, physically remove the strong mechanical Doppler interference peaks common to both channels to obtain the residual spectral fingerprint after spectral subtraction. ; Please see Figure 5 In one implementation, the energy ratio of the two-channel spectral fingerprints within the low-frequency interference space is calculated, and a formula for spatial propagation and hardware gain cutoff factor is established. The cumulative intervals of the numerator and denominator are both frequency unit indices 0 ≤ k < 15; within the coexistence space 16 ≤ k ≤ 255, the background spectral fingerprint of the reference environment is used. As a dynamic background barrier, it monitors the spectral fingerprint. Perform adaptive spatial spectral subtraction to obtain the residual spectral fingerprint after spectral subtraction. , where α is a preset lower limit gain compensation factor for the spectrum, which physically eliminates the strong mechanical Doppler interference peaks shared by the two channels.

[0045] Step 6: Extract the residual spectral fingerprint after spectral subtraction. By performing multi-peak search and spatiotemporal stability verification, a pure biological activity spectrum, after removing all environmental and mechanical background, is decoupled and purified. ; Please see Figure 6 In one implementation, the residual spectral fingerprint after spectral subtraction... In this process, the sliding window maximum value filtering algorithm is used to locate local energy maxima and record their frequency indices. Peak amplitude and full width at half maximum (FWHM) Simultaneously maintain a length of The time-series feature tracking queue is used to calculate the maximum point in continuous Standard deviation of center frequency within a frame and the coefficient of variation of amplitude ,in The preset frame length.

[0046] When the standard deviation of the center frequency is satisfied (i.e., center frequency fluctuation less than 0.5 frequency units) and amplitude variation coefficient At that time, the residual narrowband spike was determined to be mechanical residual interference at a fixed frequency, and then the frequency of the maximum point was used as the index. Centered on, with half the peak and full width To lock the frequency range by radius, establish a mechanical interference mask range. .

[0047] Finally, the mechanical interference mask area is utilized. left edge point and right edge point Perform two-point linear interpolation on the true spectral energy value at that location to reconstruct and replace the mask interval. The energy at each internal frequency point is used to "smooth out" the originally towering residual mechanical noise peaks, thereby decoupling and purifying the pure biological activity spectrum after eliminating all environmental and mechanical background. .

[0048] Step 7: For the pure biological activity spectrum Fk, extract three key physical characterization parameters in parallel to construct a multidimensional feature vector V = [ [, H, S], where E is the energy density parameter, H is the normalized spectral entropy parameter, and S is the energy spread ratio parameter; In one implementation, the energy density parameter is extracted. : The sum of the pure biological activity spectrum In frequency unit index The amplitude within the interval is used to obtain the energy density parameter. This is used to characterize the overall intensity of biological disturbance to aquaculture water bodies caused by fish populations, among which The increase in the value is positively correlated with the intensity of the fish's tail-wagging competition, providing a baseline amplitude scale for feeding intensity; simultaneously, the preprocessed original reference intermediate frequency signal is added to the frequency unit index. Low-frequency integral energy of the interval And through the formula Mapping calculation of dynamic electromagnetic attenuation factor ,in The background energy of a pre-defined static standard water body is obtained by collecting radar signals in a pure, still water environment without fish or mechanical operation and calculating the corresponding low-frequency integral mean; finally, the formula is used... The corrected energy density parameter is obtained by performing exponential logarithmic adaptive gain compensation on the energy density parameter. This is used to physically eliminate electromagnetic media attenuation disturbances caused by large dynamic fluctuations in low-frequency water flow. In one implementation, the normalized spectral entropy parameter H is calculated using the formula... Calculate the Shannon spectral entropy, where This is used to map the energy probability density at each frequency point, quantifying the disorder and randomness of water body fluctuations. Specifically, when the water surface is calm or experiences only a single fixed-frequency mechanical vibration, energy is concentrated at individual frequencies, resulting in a highly uneven probability distribution. Approaching a minimum of 0, it exhibits low entropy characteristics; when fish densely compete for food and wag their tails, a large amount of disordered movement causes the spectral line broadband to homogenize and rise, at which point... Approaching the maximum value of 1, it exhibits high entropy characteristics; this correspondence is used to distinguish between high-energy regular disturbances and chaotic feeding behavior; In one implementation, the energy distribution ratio parameter is extracted. Using formulas The ratio of total energy to the maximum peak energy of the spectrum is calculated to quantitatively describe the degree of energy dispersion broadening along the Doppler frequency domain axis. A larger value indicates a flatter spectral energy distribution and a higher degree of broadening; this is specifically distinguished by setting a preset dispersion threshold: when When the frequency is less than the preset dispersion threshold, it indicates that there are extremely high peaks in the spectrum and the energy in other frequency bands is weak, which is judged as "point-source mechanical residual vibration" with highly concentrated energy. When the value is greater than the preset dispersion threshold, it indicates that the overall spectrum is raised and there are no specific abrupt narrowband strong spikes, which is determined to be "surface source biological community activity" with energy bursting across the entire frequency band. Please see Figure 7 The energy distribution ratio S is used as the core criterion to distinguish between "point source mechanical residual vibration" with highly concentrated energy and "area source biological gregarious activity" with energy bursts across the entire frequency band.

[0049] Step 8: Construct a multidimensional semantic decision matrix; Set V = [ [H, S] Introducing a time-dimensional accumulation judgment logic, this paper eliminates false targets caused by accidental fish jumping or sudden environmental changes in a single frame by setting a time-sliding observation window W. It establishes boundary decision boundaries for three states—"calm water surface," "equipment operation," and "fish feeding"—within a multi-dimensional feature space, and sets feature judgment threshold arrays. Activate concurrent state retrieval; output the fish feeding intensity results.

[0050] In one implementation, the multidimensional feature vector V = [ extracted in parallel] [H, S] Introducing a time-dimensional accumulation judgment logic, by setting a time-sliding observation window W, false targets caused by accidental leaps of fish or sudden environmental changes in a single frame are eliminated. The specific mapping decision logic is as follows: Calm State Determination: If, within a continuous observation period, the feature parameters extracted in real time simultaneously satisfy the energy density... And normalized spectral entropy This indicates that there is no effective energy rise in the Doppler frequency domain and the spectral line distribution is highly regular. Therefore, it is determined that there are only weak background noise fluctuations in the current aquaculture water body. The state latch is executed, and a structured semantic message "water surface calm" is output. , The preset static judgment threshold; Equipment operating status determination: Before performing spectral subtraction and neighborhood linear interpolation reconstruction, if the feature tracking queue identifies stable narrowband Doppler peak components and uses the original monitoring spectral fingerprint... The ratio of the total energy across the entire frequency band to the maximum peak energy is used to calculate the initial energy distribution ratio before decoupling, which satisfies... If this indicates the presence of highly concentrated strong mechanical peaks in the original spectrum, it is determined that the current water body is in the operating state of an oxygenator or water pump rotating machinery. A structured semantic message indicating "equipment operation" is output, and a dynamic stripping mask based on the reference environment's background spectral fingerprint is maintained. The preset threshold for narrowband mechanical interference identification; Fish feeding status determination: The pure biological activity spectrum after decoupling purification, spectral subtraction, and linear interpolation of the mask interval is used to reconstruct the spectrum. The corresponding core feature parameters simultaneously satisfy the modified energy density. Normalized spectral entropy And the energy distribution after decoupling is higher than that of decoupling. This indicates that the residual energy after spectral subtraction is significantly increased, the fluctuation disorder is extremely high, and it exhibits broadband surface source dispersion broadening characteristics on the Doppler frequency axis, and within the time-sliding observation window... If the percentage of valid frames meeting the above conditions exceeds a preset threshold, then the occasional mutation pseudo-target is immediately eliminated, triggering and outputting a "fish feeding" structured semantic message. , , This is the preset food intake activation threshold.

[0051] In one implementation, a time-based accumulation and judgment mechanism is introduced. The system dynamically maintains a sliding observation window W covering the past K frame observation periods. For each frame feature vector within the window W, state classification is performed, and the fish feeding judgment criteria (i.e., simultaneously satisfying the criteria) are statistically analyzed in real time. ) effective frame count ; Calculate the ratio of the activation probability of states within the window. Set the state activation ratio threshold to When the data meets the judgment criteria Upon that time, a standard "fish feeding" structured semantic message is immediately triggered and output; Simultaneously, if the system determines that it is currently in oxygen generator operation, the algorithm synchronously activates a weak signal parallel tracking branch to uncover the initial weak micro-Doppler features of the fish population that are masked by strong noise. The residual spectral entropy... The acquisition method is as follows: directly extract the mask region after step 6. Pure biological activity spectrum after linear interpolation reconstruction Using the formula ; Calculation, where The normalized energy probability density of the reconstructed spectrum; since the mechanical dead peaks have been flattened, if there is a slight disturbance from the fish, the originally flattened interpolation baseline and surrounding frequency bands will be elevated with white noise, resulting in a significant increase in the residual spectral entropy value. The residual energy dispersion ratio The method to obtain it is: using the formula Calculate the ratio of the total integrated energy to the maximum amplitude of the reconstructed pure biological activity spectrum; since the strong mechanical spikes have been suppressed by interpolation, the denominator... When the fish return to normal baseline, if there is scattered and weak feeding activity in the school, the spectral lines will show broad diffusion, and the molecules will be larger, thus... Significantly increased. When the characteristic parameter simultaneously satisfies the residual spectral entropy And the residual energy distribution ratio At that time, it was determined that the current water body was under severe masking of strong noise, resulting in weak feeding behavior of fish (among which...). (Activate the low-order threshold for the preset weak signal); at this time, the system follows the weighted formula: ; Dynamic calculation of weak feeding confidence score ,in These are adaptive weighting coefficients; ultimately, the system uses the calculated confidence scores... Concurrent output of compound semantic decision instructions, the specific mapping rule is as follows: If The system outputs the compound semantic message "equipment is operating and accompanied by weak feeding," and controls the actuator to switch to the "pulse-type micro-feeding" command to test the fish's aggregation level; if The system outputs a compound semantic message of "equipment is running and accompanied by clear feeding", controls the actuator to switch to the "intermittent half-power regular feeding" instruction, and sends a structured data message of "noise-masked feeding activation anomaly warning" to the main control terminal.

[0052] The invention will be further illustrated below through specific experiments.

[0053] The hardware architecture and scenario setup of this experimental system: The physical experiment in this embodiment is based on a circular fiberglass tank (diameter D = 2.0m, water depth H = 1.2m) in a modern factory-style recirculating aquaculture (RAS) workshop. The fish species cultured is California bass, and the stocking density is 40. Above the water body, a 50W airflow surge-type aerator and a 30W circulating water pump operate year-round, with a large number of high-density, non-periodic bursting bubbles on the water surface. The core sensing hardware of the system uses two 60GHz frequency-modulated continuous wave (FMCW) millimeter-wave radars based on the Infineon BGT60TR13C chip: Monitoring Channel Radar (Mon): rigidly and vertically deployed 0.8m above the active water surface in the feeding area using a stainless steel cantilever bracket, with the beam center directly facing the core area where fish are feeding; Reference Channel Radar (Ref): rigidly and vertically deployed 0.8m above the non-feeding reference water surface far from the active feeding area and where there is no fish activity, used to capture purely static background, tank wall reflections, and the same source of rotating mechanical physical vibration noise from the aerator and water pump. The radar microwave front-end configuration parameters are: starting frequency... The sweep bandwidth B = 4GHz, the number of sampling points for a single linear frequency modulated pulse (Chirp) N = 512, the pulse repetition period PRT = 50s, and the frame period is set to 50ms (i.e., 20 frames of spectral fingerprints are collected per second).

[0054] The experimental procedure is as follows: Step 1: Dual-channel signal synchronous acquisition and adaptive preprocessing. The two radars achieve frame synchronous acquisition through an external unified hardware trigger source. The full-scale voltage threshold of the analog-to-digital converter (ADC) is... Receiver gain closed-loop calibration: At the moment of feeding, the leaping of the fish causes a significant increase in water surface scattering. The system sliding window detects the raw intermediate frequency (IF) signal of the monitoring channel. Mon The peak amplitude reached A max = 0.88V. Out-of-bounds condition triggered. The unit immediately transmits the radar's programmable intermediate frequency gain via the SPI bus. The distortion was reduced from the initial 30dB to 24dB, perfectly avoiding clipping distortion of the time-domain signal; time-domain differential processing: a smoothing factor β was set to 0.95. The system utilizes an updated formula. ; An adaptive iterative two-channel time-varying dynamic background model is employed. The time-domain complex signal input in the current t-th frame is subtracted from the dynamic background in the complex domain, successfully eliminating the zero Doppler DC component caused by fixed scattering bodies such as the steel wall of the aquaculture tank and the cantilever support, and outputting a monitoring differential signal. Differential signal with reference .

[0055] Step 2: Dual-channel Doppler bispectral spatial construction and region segmentation. Hanning windows are applied to the two differential signals, and a 512-point Fast Fourier Transform (FFT) is performed to obtain the real-time monitoring spectral fingerprint. Background fingerprint Fk of the reference environment Ref Based on the spatial characteristics of the Doppler frequency domain, the following characteristic regions are defined: Low-frequency interference space of hydrodynamics: the frequency unit index interval is 0≤k<15 (corresponding to the low-frequency band of Doppler frequency shift, mainly reflecting the macroscopic flow velocity of the water body and aperiodic bubble disturbance); Coexistence space of equipment and biological activities: the frequency unit index interval is 16≤k≤255 (corresponding to the mid-high frequency band of Doppler frequency shift, the overlapping area of ​​mechanical vibration peaks and fish feeding behavior).

[0056] Step 3: Decoupling and stripping of mechanical interference components based on spatial reference spectrum subtraction, spatial gain cutoff calibration: Calculate the energy ratio of the two channels in the low-frequency interference space. Based on actual measurements and calculations, the spatial propagation and hardware gain cutoff factor γ = 1.05 for the current frame; adaptive spatial spectrum subtraction: within the coexistence space 16 ≤ k ≤ 255, using... Subtraction is performed as a dynamic barrier: (Set the lower limit compensation factor α = 0.01). At this point, the strong Doppler mechanical spike of the oxygen generator at frequency unit 80 (Bin 80) was reduced by 28.7 dB; Spatiotemporal stability verification and neighborhood reconstruction: a narrow-band spike was found in the residual spectrum, and its index was located. Read the feature tracking queue of historical frames with a length of M = 20, and calculate the standard deviation of the center frequency of the maximum point. The amplitude variation coefficient CV = 0.021 < 0.05. The system confirmed that it was residual fixed-frequency mechanical interference from the oxygen generator, locking the mask interval W = [78, 82]. Edge points were extracted. and The spectral energy is then used for linear interpolation and reconstruction filling. At this point, mechanical interference spikes are completely eliminated, and the pure biological activity spectrum Fk is decoupled and purified.

[0057] Step 4: Multidimensional Semantic Feature Extraction of Pure Biological Activity Spectrum. For the pure biological activity spectrum Fk, multidimensional semantic indicator extraction is performed in parallel: Energy density parameter: The energy integral across the entire frequency band is... Simultaneously, low-frequency energy integrals are extracted. Input the preset benchmark Calculate the dynamic electromagnetic attenuation factor An adaptive gain compensation correction is applied to the energy density to obtain... Normalized spectral entropy parameter: used to calculate the normalized probability density. Substituting into the Shannon entropy formula, we get This high entropy value reflects the disordered and scattered fluctuations on the water surface when fish are competing for food; the energy distribution ratio parameter is calculated as follows: It is demonstrated that energy exhibits overtone broadband broadening of surface-source biological community activity on the frequency domain axis. A highly robust multidimensional anti-interference eigenvector V = [68755, 6.2, 52.4] is constructed.

[0058] Step 5: Semantic state mapping and decision output under spatiotemporal joint constraints. The system maintains a sliding observation window W covering the past K = 10 frame observation periods (i.e., a delay of 7.5ms × 10 = 75ms). The decision threshold array is set as follows: , , , proportion threshold Upon verification, all parameters of the current feature vector V far exceed the judgment thresholds (68755>= 30000, 6.2>= 5.0, 52.4>= 35). The system counted the number of valid frames that met the feeding judgment criteria in the past 10 frames. The activation probability ratio P = 9 / 10 = 0.9. Since 0.9 >= 0.8, the decision layer immediately triggers and sends a structured semantic message "fish swarm feeding aggressively" to the external PLC automated feeder, driving the feed spraying mechanism to start feeding at full power, thus realizing high-precision, low-latency non-contact intelligent aquaculture closed-loop control.

[0059] In the experiment, the present invention was further compared and analyzed with the traditional single-channel radar thresholding scheme and the traditional computer vision (YOLOv8) scheme. The results are shown in Table 1 below. Table 1

[0060] As can be seen from Table 1, the fish feeding intensity analysis method based on electromagnetic wave spectrum component decoupling optimization proposed in this invention demonstrates significant technological advancements and beneficial effects compared to traditional single-channel radar thresholding schemes and traditional computer vision (YOLOv8) schemes in terms of anti-interference capability, recognition accuracy, and all-weather edge deployment. Specifically, these are reflected in: (1) Strong background noise and mechanical interference suppression capability: Traditional single-channel radar schemes lack spatial offset references, resulting in a suppression ratio of 0.0 dB for strong clutter caused by rotating machinery such as aerators and pumps, leading to a static background and barrel wall echo residual rate as high as 18.4%. However, this invention achieves a strong mechanical noise suppression ratio of up to 28.7 dB through dual-channel differential subtraction and spatiotemporal joint decoupling, reducing the static background and barrel wall echo residual rate to 0.5%. This directly increases the signal-to-noise ratio (SNR) of complex water body signals from 6.2 dB to 21.5 dB, achieving a key performance leap of over 15 dB, clearing the background clutter for subsequent extraction of weak biological signals.

[0061] (2) Accurate identification and high recall of fish feeding behavior in extremely harsh environments: In aquaculture sites, there are often visual obstructions and interferences such as high-density water mist and bursting bubbles from aerators. Traditional computer vision solutions (YOLOv8) are limited by optical visibility, and their accuracy in identifying feeding is only 78.5%, with a recall rate as low as 45.2% for concealed weak feeding behavior; single-channel radar, due to noise submersion, has a weak feeding recall rate of only 12.3%. This invention utilizes the physical properties of electromagnetic waves penetrating water mist bubbles, and through decoupled multi-dimensional feature vector extraction, significantly increases the accuracy of fish feeding status identification to 98.1%; in particular, by activating the weak signal parallel tracking branch and decoupling the residual features, the capture recall rate of weak feeding behavior, which is originally easily masked by strong noise, is increased to 96.4%, fundamentally solving the pain point of missing small-scale stealing and slight probing behavior in the early stages of industrial sites.

[0062] (3) Advantages of low-latency, all-weather edge MCU deployment: Traditional computer vision solutions have extremely high computing power requirements, with a single-frame full-process processing latency of up to 45.8 ms, and are heavily reliant on expensive GPU computing power. They are completely inoperable in environments with no light at night or in heavy rain and fog. Although single-channel radar solutions have lower latency (4.2 ms), their accuracy cannot meet the standards for industrial applications. This invention performs feature decoupling and dimensionality reduction directly in the frequency domain, without the need for complex temporal reconstruction of images or large model inference. The single-frame full-process processing latency is only 7.5 ms, demonstrating extremely high computational efficiency and lightweight characteristics, perfectly supporting the adaptive deployment of low-cost edge MCUs. At the same time, due to the adoption of a microwave active detection architecture, it completely gets rid of the constraints of lighting and weather conditions, and has the ability to operate continuously and stably with high precision in all weather conditions. This provides extremely solid underlying hardware and algorithm support for industrial-grade precision intelligent aquaculture closed-loop material control.

[0063] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0064] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization, characterized in that, Includes the following steps: Step 1: Acquire radar signals from the monitoring channel and the reference channel to obtain the raw monitoring intermediate frequency signal. and the original reference intermediate frequency signal ; Step 2: For the above and The monitoring differential signal is obtained after preprocessing. Differential signal with reference ; Step 3: Separately... and The time-domain echo is converted into a frequency-domain spectral fingerprint reflecting the energy density distribution with frequency, thus obtaining the real-time monitoring spectral fingerprint Fk. Mon Background fingerprint Fk of the reference environment Ref ; Step 4: Based on the frequency unit index , respectively and The Doppler frequency domain space was divided into two parts: the low-frequency interference space of hydrodynamics (A) and the coexistence space of equipment and biological activities (B). Step 5: Based on the subtraction of the spatial reference spectrum, perform spectral fingerprint offsetting for A to remove strong mechanical Doppler interference spikes; perform physical isolation directly for B; summarize the frequency domain components after the two spatial processing to obtain the residual spectral fingerprint. ; Step 6: Extract the residual spectral fingerprint after spectral subtraction. Multi-peak search and spatiotemporal stability verification were performed to decouple and purify the pure spectrum of biological activity. ; Step 7: Targeting the pure biological activity spectrum Extract the multidimensional feature vector V; Where V = [ , H, S], Here, H is the energy density parameter, H is the normalized spectral entropy parameter, and S is the energy dispersion ratio parameter. This is used to characterize the overall intensity of biological disturbance to aquaculture water bodies caused by fish populations, among which The increase in the value is positively correlated with the intensity of the fish's tail-wagging competition; at the same time, the preprocessed original reference intermediate frequency signal is added to the frequency unit index. Low-frequency integral energy of the interval And through the formula Mapping calculation of dynamic electromagnetic attenuation factor ,in The background energy of a pre-defined static standard water body is obtained by collecting radar signals in a pure, still water environment without fish or mechanical operation and calculating the corresponding low-frequency integral mean; finally, the formula is used... The corrected energy density parameter is obtained by performing exponential logarithmic adaptive gain compensation on the energy density parameter. This is used to physically eliminate electromagnetic media attenuation disturbances caused by large dynamic fluctuations in low-frequency water flow. ;in The energy probability density mapping at each frequency point is used to quantify the disorder and randomness of water body fluctuations. , used to quantitatively describe the degree of energy dispersion broadening along the Doppler frequency domain axis; where, when When the value is less than the preset dispersion threshold, it indicates the presence of extremely high peaks in the spectrum and weak energy in other frequency bands, which is determined to be point-source mechanical residual vibration with highly concentrated energy; if When the value is greater than or equal to the preset dispersion threshold, it indicates that the overall spectrum is raised and there are no specific abrupt narrowband strong spikes, which is determined to be a surface-source biological community activity with energy bursting across the entire frequency band. Step 8: Introduce V into the time dimension of the cumulative judgment logic, eliminate false targets by setting a time sliding observation window W, and output the fish feeding intensity analysis results.

2. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to claim 1, characterized in that: In step 1, the monitoring channel radar is vertically aligned with the water surface of the feeding activity area to scan and acquire the raw monitoring intermediate frequency signal, which includes fish feeding disturbances, static background, and mechanical point source interference. ; The reference channel radar scans a non-feeding reference area far from the active feeding area and where there is no fish activity. It maintains the same parameter configuration and synchronous sampling as the monitoring channel radar to acquire the raw reference intermediate frequency signal containing pure static background and mechanical point source interference. .

3. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to claim 1, characterized in that: In step 2, for the and Preprocessing involves establishing buffer queues with a sliding window length of N for each of the two original intermediate frequency signals and calculating the statistical mean of the complex-domain signals within the queues. A mean subtraction operation is then performed to eliminate the zero-Doppler DC component generated by the tank walls and the static water surface. Subsequently, a Hanning window is applied to suppress spectral energy leakage, resulting in the DC-free monitoring differential signal. Differential signal with reference Where N is a preset value; The mean subtraction operation will subtract the time-domain complex signal input in the current frame. Subtracting the statistical mean within window N completes the mean subtraction operation, thereby eliminating the zero Doppler DC component generated by the aquaculture tank wall and the static water surface, resulting in the intermediate-state de-DC time-domain signal. Subsequently, Hanning windowing is applied to this de-DC time-domain signal to suppress spectral energy leakage. At this point, the final time-domain result output by the windowed signal is compared with the corresponding dynamic background estimate. Perform difference subtraction in the complex field; where... , This is the dynamic background estimate for the current frame. The background estimate is the value of the previous frame, and β is a preset smoothing factor that satisfies the interval [missing information]. ; The specific formula for the Hanning window addition process is as follows: ； ； In the formula, The index of discrete sampling points in the time domain and satisfying ; and These are the time-domain complex signals of the monitoring channel and the reference channel after performing the differential subtraction in the current frame, respectively. The length is the same as the length of the cache queue. Strictly consistent discrete Hanning window function; the final obtained and That is, the monitoring differential signal and the reference differential signal after DC removal are output to the subsequent frequency domain spectral fingerprint conversion step.

4. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to claim 1, characterized in that: In step 3, for the and Perform 512-point Fast Fourier Transform to convert the time-domain echo into a frequency-domain spectral fingerprint that reflects the energy density distribution with frequency.

5. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to claim 1, characterized in that: In step 4, the frequency unit index To obtain the frequency domain discrete spectral point indices after performing Fast Fourier Transform on the preprocessed monitoring differential signal and the reference differential signal, and satisfying the following conditions: ; If the frequency unit index satisfy Then and The corresponding amplitude components are divided into a low-frequency interference space for hydrodynamics; if the frequency unit index satisfy Then and The corresponding amplitude components are divided into spaces where equipment and biological activities coexist.

6. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to claim 1, characterized in that: In step 5, based on the subtraction of the spatial reference spectrum, spectral fingerprint offsetting is performed on A in the same frequency domain dimension to physically remove the strong mechanical Doppler interference spikes common to both the monitoring channel and the reference channel; for B, physical isolation is directly achieved through the frequency band stopband; finally, the frequency domain components after the two spatial processing are summarized to obtain the residual spectral fingerprint after spectral subtraction and isolation. ; Calculate the energy ratio of the two-channel spectral fingerprints in the low-frequency interference space, and establish the spatial propagation and hardware gain cutoff factor. , of which molecules With denominator The accumulation intervals are all frequency unit indices 0 ≤ k < 15; within the coexistence space 16 ≤ k ≤ 255, the background spectral fingerprint of the reference environment is used. As a dynamic background barrier, it monitors spectral fingerprints. Perform adaptive spatial spectral subtraction to obtain the residual spectral fingerprint after spectral subtraction. , where α is the preset lower limit gain compensation factor of the spectrum.

7. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to claim 1, characterized in that: In step 6, the pure biological activity spectrum, after removing all environmental and mechanical background, is decoupled and purified. The sliding window maximum value filtering algorithm is used to locate local energy maxima and record their frequency indices. Peak amplitude and full width at half maximum (FWHM) Simultaneously maintain a length of The feature tracking queue is used to calculate the maximum point in a continuous sequence. Standard deviation of center frequency within a frame and the coefficient of variation of amplitude ;in, This is a preset value; when it meets the requirements... and When the residual narrowband spikes are determined to be mechanical residual interference at a fixed frequency, the frequency range is locked and a mechanical interference mask range is established. Finally, the mechanical interference mask area is utilized. left edge point and right edge point The energy value is linearly interpolated to reconstruct and replace the mask interval. The background baseline energy within the spectrum is used to decouple and purify the pure biological activity spectrum after removing all environmental and mechanical backgrounds. .

8. The method for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization according to any one of claims 1-7, characterized in that: In step 8, if the feature parameters extracted in real time during the continuous observation period simultaneously satisfy the energy density... And normalized spectral entropy If the condition is met, it is considered a calm state; among which... , The preset static judgment threshold; Before performing spectral subtraction and neighborhood linear interpolation reconstruction, if the feature tracking queue identifies stable narrowband Doppler spike components and utilizes the original monitoring spectral fingerprint... The ratio of the total energy across the entire frequency band to the maximum peak energy is used to calculate the initial energy distribution ratio before decoupling, which satisfies... This indicates the presence of highly concentrated, strong mechanical peaks in the original spectrum, suggesting that the water body is currently in the operation state of an oxygenator or water pump rotating machinery; among which... The preset threshold for narrowband mechanical interference identification; If the pure biological activity spectrum is reconstructed through decoupling purification, spectral subtraction, and linear interpolation within the mask interval... The corresponding characteristic parameters simultaneously satisfy the modified energy density. Normalized spectral entropy and the energy distribution ratio after decoupling And in the time-sliding observation window If the percentage of valid judgment frames meeting the judgment criteria exceeds a preset threshold, it indicates a significant increase in residual energy after spectral subtraction, extremely high degree of fluctuation disorder, and broadband surface source dispersion broadening characteristics on the Doppler frequency domain axis. In this case, the occasional mutation false target is immediately eliminated, and the fish is judged to be in a feeding state. , , This is the preset food intake activation threshold.

9. A system for analyzing fish feeding intensity based on electromagnetic wave spectral component decoupling optimization, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the fish feeding intensity analysis method based on electromagnetic wave spectral component decoupling optimization as described in any one of claims 1 to 8.

Images