Dynamic self-adaptive mosquito wing vibration frequency acoustic trapping system and method

Abstract

The invention relates to the technical field of mosquito acoustic trapping, in particular to a dynamic self-adaptive mosquito wing vibration frequency acoustic trapping system and method, and the system comprises a collection module, a correlation module, a first processing module, a second processing module, a third processing module and a correction module. The method comprises the following steps: acquiring mosquito wing vibration frequency data, preprocessing to construct a frequency distribution model, generating a multivariate feature data set in combination with environment variables, extracting features by using deep learning and a recurrent neural network, performing weighted convergence, setting common and special scene branch optimization frequency adjustment, and generating a correction value based on region division and a frequency correction curve. And the final frequency is output by fusing the multi-branch predicted value. According to the invention, dynamic frequency adjustment, complex scene adaptation and long-acting luring can be realized, and the trapping precision and engineering applicability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mosquito trapping, in particular to a dynamic adaptive mosquito wing vibration frequency acoustic trapping system and method. BACKGROUND

[0002] With the increasing demand for mosquito control, acoustic trapping technology has gradually attracted attention in the field of mosquito control. Mosquitoes have a significant tendency to respond to specific frequency sounds, especially in the 300Hz to 800Hz range, which covers the key areas of their response to bat ultrasonic waves and wing vibration frequency. In addition, the wing vibration frequency of female mosquitoes before blood sucking is about 450Hz, while male mosquitoes show the best response to a frequency of about 362.5Hz when courting, which provides a theoretical basis for the design of acoustic trapping systems.

[0003] However, existing acoustic trapping technology mainly relies on a single frequency or fixed combination signal, and does not fully consider the dynamic behavior changes of mosquitoes and their adaptability to environmental factors in actual application. For example, existing technology is usually based on static frequency output design and lacks real-time tracking capability for dynamic changes in mosquito wing vibration frequency. Although some schemes attempt to combine flower fragrance and other chemical attractants to enhance effectiveness, they fail to achieve intelligent linkage of fragrance release and acoustic signals, and the lack of fragrance persistence limits their long-term use effectiveness. In addition, existing devices have certain limitations in indoor and outdoor scene adaptability, such as the lack of egg removal function in water application scenarios. Specifically, patent CN119067269A discloses a mosquito trapping device integrating acoustic and chemical attraction, which achieves trapping by combining multi-band acoustic waves with volatile aromatic substances. Although this scheme improves trapping efficiency, there is still room for improvement: it fails to dynamically adjust the frequency of acoustic waves according to the behavior characteristics of mosquitoes; it does not optimize the acoustic signal and fragrance release strategy for different scenarios; and its adaptability in complex environments needs to be improved.

[0004] Therefore, it is necessary to design a dynamic adaptive mosquito wing vibration frequency acoustic trapping system and method to address the shortcomings of current technology in dynamic frequency adjustment, multi-scene adaptation, and long-term attraction mechanism. SUMMARY

[0005] In view of this, the present application proposes a dynamic adaptive mosquito wing vibration frequency acoustic trapping system and method, aiming to solve the problem that existing acoustic trapping technology in current technology lacks real-time dynamic tracking and adaptive adjustment capability of mosquito behavior, and also has deficiencies in intelligent linkage of acoustic signals and chemical attractants, fragrance persistence, and multi-scene adaptability.

[0006] The present application proposes a dynamic adaptive mosquito wing vibration frequency acoustic trapping method, comprising: Multi-band data of wing vibration frequency of mosquitoes is acquired, and the multi-band data is preprocessed to construct a frequency distribution model; A basic frequency sequence is generated according to the frequency distribution model, and the basic frequency sequence is time-series associated with environmental variables, while environmental characteristics of a corresponding time period are collected to form a multi-feature data set; The multi-feature data set is input into a deep learning network to extract local features of the frequency signal, and the local features are input into a recurrent neural network, an adaptive weight mechanism is introduced in the feature sequence output by the recurrent neural network, the correlation of each time step feature is calculated to obtain a weight distribution, and the time step features are weighted and aggregated according to the weight distribution to obtain a global feature representation; Two independent branch processing units are set, and the global feature representation is input into the corresponding branch processing unit for frequency optimization adjustment to obtain ordinary scene branch prediction values and special scene branch prediction values; The multi-feature data set is divided into several hierarchical regions according to environmental variables, and a frequency correction curve of each region is constructed based on the relationship between historical frequency data and actual frequency response, the environmental variables of the target time period are input into the frequency correction curve of the target region and its adjacent regions to obtain regional correction results, and the regional correction results of adjacent regions are fused by inverse distance weighting to obtain a frequency correction value; The final frequency output value is obtained by weighted fusion of the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value.

[0007] Further, when pre-processing the multi-band data to construct the frequency distribution model, it includes: Interval normalization is performed on each frequency band data on the frequency axis to unify the frequency grid, linear interpolation is performed on the missing data at the corresponding frequency points, and repeated or abnormal points are removed, and boundary value completion is performed according to the maximum response range to ensure data continuity, to obtain optimized data of each frequency band; The optimized data of each frequency band is added at each unified frequency point, and the sum is obtained along the frequency grid to obtain the frequency distribution model.

[0008] Further, when generating a basic frequency sequence according to the frequency distribution model, and time-series associating the basic frequency sequence with environmental variables, while collecting environmental characteristics of a corresponding time period to form a multi-feature data set, it includes: The frequency response is acquired within the historical period, and the frequency-environment monotonic correspondence relationship is established according to the frequency distribution model, and the basic frequency sequence of each time is calculated on the frequency grid based on reverse lookup and piecewise linear interpolation; Data representing environmental variables are acquired, angle data are de-rolled and numerically encoded in sine and cosine, and the basic frequency sequence is time-aligned according to a unified sampling step; Collect and normalize environment features on the same time axis as the historical period, including temperature, humidity, air pressure and light intensity; Based on the sliding time window, the environment features in the window are taken as the input, and the actual frequency response corresponding to the window forward step is taken as the label to construct a multivariate feature dataset, and the scene labels of ordinary scenes and special scenes are labeled according to the environment level threshold; Among them, the derived features include environment vector components, environment change rate, environment stability index and light intensity change rate.

[0009] Further, when the multivariate feature dataset is input into the deep learning network to extract the local features of the frequency signal, it includes: Stack the multivariate feature dataset in the time dimension according to the sliding time window and concatenate the environment features and the basic frequency sequence in the feature dimension to form a one-dimensional time series tensor; Adopt a one-dimensional time series convolution deep learning network, use causal padding in the first layer, use multi-scale convolution kernels and optional empty convolution in the middle layer, and the kernel length of the multi-scale convolution kernel is one or more of 3, 5 and 7; Nonlinear activation is set after each convolution layer in turn, and residual connection is used to suppress deep degradation, and finally the local features are output.

[0010] Further, the correlation of each time step feature is calculated to obtain an adaptive weight distribution, and the time step features are weighted and aggregated according to the weight distribution to obtain a global feature representation, which includes: Input the local features into the stacked recurrent neural network to model the long-term dynamic dependence and obtain a hidden state sequence; Calculate the correlation score based on the adaptive weight mechanism, and obtain the weight distribution through Softmax normalization; According to the weight distribution, the time step features are weighted and aggregated to obtain a context vector, and the context vector is projected through a nonlinear projection to obtain a global feature representation.

[0011] Further, set two independent branch processing units, input the global feature representation into the corresponding branch processing unit for frequency optimization adjustment to obtain the ordinary scene branch prediction value and the special scene branch prediction value, which includes: When setting two independent branch processing units, it includes an ordinary scene branch processing unit and a special scene branch processing unit; Each branch processing unit comprises at least two fully connected layers and a nonlinear activation, and random deactivation is arranged between the layers to suppress overfitting; during the training phase, only the corresponding branch processing unit is updated based on the scene label, and the other branch processing unit is frozen, and during the prediction phase, the global feature representation is respectively input into the two branch processing units in a forward direction to simultaneously generate the ordinary scene branch prediction value and the special scene branch prediction value, and the quantile loss is used as a robust loss function to constrain the training error.

[0012] Further, the multivariate feature data set is divided into several hierarchical regions according to the environmental variable, and the frequency correction curve of each region is constructed based on the relationship between the historical frequency data and the actual frequency response, including: The environmental circumference is divided into regions with equal angular width according to the environmental variable as the angle variable, and the regions are extended to form a hierarchical structure according to the frequency level. In each region, a monotonic constraint regression fitting is performed to obtain the frequency correction curve, with the historical frequency data as the independent variable and the actual frequency response as the dependent variable.

[0013] Further, the environmental variable of the target period is input into the frequency correction curve of the target region and its adjacent regions to obtain the regional correction result, and the frequency correction value is obtained by inversely distance weighted fusion of the regional correction results of the adjacent regions, including: Determine the target region and its adjacent regions corresponding to the environmental variable of the target period, and substitute the environmental variable of the target period into the corresponding frequency correction curve to obtain multiple regional correction results. According to the angle between the central axis of each region and the environmental variable, the inverse distance weight is calculated, and the regional correction results are weighted and summed to obtain the frequency correction value.

[0014] Further, according to the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value, the final frequency output value is obtained by weighted fusion, including: The performance indicators of the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value are calculated, and the subjective weight and the objective weight are determined according to the performance indicators, the combined weight is obtained by weight normalization and non-negative constraint, and the sum of all weights in the combined weight is 1. The final frequency output value is calculated according to the combined weight and the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value.

[0015] Compared with the prior art, the application has the beneficial effects that: by acquiring multi-band data of mosquito wing vibration frequency and preprocessing, the system can not only construct a complete frequency distribution model, but also eliminate noise and abnormal data, so that the basic frequency sequence is closer to the real vibration characteristics of mosquitoes. Combined with environmental variables to form a multivariate feature dataset, the system can consider the influence of temperature, humidity, light and other environmental factors on mosquito behavior at the same time, and realize fine analysis of mosquito dynamic behavior, which significantly improves the scientificity and accuracy of frequency signal generation and solves the problem that traditional fixed frequency trapping technology is difficult to cope with behavior changes. Secondly, the method uses a mechanism combining deep learning network and recurrent neural network to realize intelligent optimization of frequency signal. By extracting local features of frequency signal through deep learning network, the system can capture tiny vibration patterns and instantaneous frequency changes; the recurrent neural network introduces an adaptive weight mechanism in the time sequence, which can calculate the correlation of each time step feature and weight and aggregate the key features to obtain global feature representation. This process can effectively identify the influence of behavior rules and environmental interference factors, so that the predicted frequency signal has real-time performance and global optimization capability, greatly improving the pertinence and effectiveness of acoustic trapping. Finally, the method realizes multi-scene adaptive optimization through two independent branch processing units and regional correction mechanism. The ordinary scene branch and the special scene branch can generate predicted values for regular indoor and outdoor environments and special environments (such as water surface or complex terrain) respectively, ensuring that reasonable acoustic frequency output can be obtained in different scenes. At the same time, the regional frequency correction curve based on historical frequency data and actual response can be used to finely correct each environmental level, and the correction results of adjacent regions are fused through inverse distance weighting to realize dynamic adaptation of environmental variables. This not only enhances the stability of the system in complex environments, but also prolongs the long-term use effect of the acoustic trapping device, avoiding the decline of trapping efficiency caused by frequency deviation.

[0016] On the other hand, the application also provides a dynamic adaptive mosquito wing vibration frequency acoustic trapping system, comprising: The acquisition module is configured to acquire multi-band data of mosquito wing vibration frequency and preprocess the multi-band data to construct a frequency distribution model; The correlation module is configured to generate a basic frequency sequence according to the frequency distribution model, and time series correlate the basic frequency sequence with environmental variables, while collecting environmental characteristics in the corresponding period to form a multivariate feature dataset; The first processing module is configured to input the multivariate feature dataset into a deep learning network to extract local features of the frequency signal, and input the local features into a recurrent neural network to introduce an adaptive weight mechanism in the feature sequence output by the recurrent neural network, calculate the correlation of each time step feature and obtain a weight distribution, weight and aggregate the time step features according to the weight distribution, and obtain a global feature representation; The second processing module is configured to set two independent branch processing units, input the global feature representation into the corresponding branch processing units respectively for frequency optimization adjustment, and obtain the normal scene branch prediction value and the special scene branch prediction value; The third processing module is configured to divide the multivariate feature data set into a plurality of hierarchical regions according to the environmental variables, construct a frequency correction curve of each region based on the relationship between the historical frequency data and the actual frequency response, input the environmental variables of the target period into the frequency correction curve of the target region and its adjacent regions to obtain the region correction result, and obtain the frequency correction value by inversely distance weighted fusion of the region correction results of the adjacent regions; The correction module is configured to obtain the final frequency output value by weighted fusion of the normal scene branch prediction value, the special scene branch prediction value and the frequency correction value.

[0017] It can be understood that the dynamic adaptive mosquito wing vibration frequency acoustic trapping system and method in each of the above embodiments have the same beneficial effects, which will not be described again. BRIEF DESCRIPTION OF DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments and are not meant to limit the present application. Furthermore, the same reference numerals are used throughout the various drawings to designate the same or similar parts. In the drawings: Figure 1 A flow chart of a dynamic adaptive mosquito wing vibration frequency acoustic trapping method provided by an embodiment of the present application; Figure 2 A flow chart of a dynamic adaptive mosquito wing vibration frequency acoustic trapping method provided by an embodiment of the present application; Figure 3 A functional block diagram of a dynamic adaptive mosquito wing vibration frequency acoustic trapping system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in detail with reference to the drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0020] As Figures 1-2As shown, in some embodiments of the present application, the present embodiment provides a dynamic adaptive mosquito wing beat frequency acoustic trapping method, comprising: Step S100, obtaining multi-band data of mosquito wing beat frequency, and preprocessing the multi-band data to construct a frequency distribution model.

[0021] Specifically, when preprocessing the multi-band data to construct the frequency distribution model, it includes: interval normalization of each frequency band data on the frequency axis to unify the frequency grid, linear interpolation of missing data at the corresponding frequency point and elimination of repeated or abnormal points, setting constraint conditions according to the maximum response range at the frequency boundary and completing the boundary value to ensure data continuity, obtaining optimized data of each frequency band; adding each optimized data at each unified frequency point, and obtaining the frequency distribution model by traversing the sum along the frequency grid.

[0022] It can be understood that through the standardization of the frequency axis and the data reconstruction, the unified modeling of the mosquito wing beat multi-band signal is realized. Due to the influence of individual differences, flight posture, environmental noise and other factors, the original collected data usually has problems such as uneven frequency bands, missing points or repeated points, which is difficult to be directly used for feature analysis. Therefore, the method first performs interval normalization on the frequency axis, that is, maps each frequency band data to a unified frequency grid, thereby eliminating the frequency resolution difference caused by different sampling devices or time periods, and providing a unified scale basis for subsequent data fusion. Secondly, the method realizes accurate completion and purification of data through linear interpolation and abnormal data elimination mechanism. For the missing data points that may occur in the multi-band sampling process, the linear relationship between adjacent frequency points is used for interpolation reconstruction, which can effectively restore the continuity of the missing frequency band; and by detecting and eliminating repeated points or abnormal response values, the deviation caused by high noise or abnormal measurement on the overall frequency model is avoided. This process is equivalent to establishing a noise suppression and data correction constraint mechanism in the frequency domain, making the constructed frequency model more stable and reliable. Thirdly, setting the maximum response range constraint condition at the frequency boundary and completing the boundary, is to ensure the continuity and physical reasonableness of the data in the full frequency range. Since the response value of the edge frequency band is usually weak or missing in actual sampling, by setting the boundary constraint based on the maximum response range, the range of interpolation or completion value can be effectively limited, so that the model is more consistent with the real mosquito wing beat frequency spectrum characteristics. This constraint process is mathematically reflected as a constraint on the boundary conditions of the frequency function, and in the sense of signal processing, it ensures the integrity and smoothness of the frequency spectrum energy distribution. Finally, by superimposing and summing each optimized multi-band data at the unified frequency point, the result obtained by traversing along the frequency grid is the frequency distribution model. The model essentially reflects the energy distribution characteristics of mosquito wing beat signal in different frequency intervals, and is an important basis for subsequent analysis of mosquito species, behavior characteristics and dynamic frequency variation law.

[0023] It can be seen that based on the signal processing methods of frequency normalization, data interpolation completion, boundary constraint and spectral superposition summation, a continuous, stable and quantifiable mosquito wing vibration frequency distribution model is constructed, thereby providing reliable theoretical support and data basis for subsequent dynamic frequency prediction and acoustic trapping optimization.

[0024] Step S200, generating a basic frequency sequence according to the frequency distribution model, and time series correlating the basic frequency sequence with the environmental variables, while collecting the environmental characteristics of the corresponding period to form a multivariate feature dataset.

[0025] Specifically, when generating a basic frequency sequence according to the frequency distribution model, and time series correlating the basic frequency sequence with the environmental variables, while collecting the environmental characteristics of the corresponding period to form a multivariate feature dataset, it includes: obtaining the frequency response in the historical period and establishing the frequency-environment monotone corresponding relationship according to the frequency distribution model, and calculating the basic frequency sequence of each time on the frequency grid based on reverse lookup and piecewise linear interpolation; obtaining data representing environmental variables, de-rotating the angle data and numerically encoding with sine and cosine, and time aligning the basic frequency sequence with the environmental variables according to the uniform sampling step; collecting the environmental characteristics on the same time axis as the historical period and normalizing the environmental characteristics, the environmental characteristics including temperature, humidity, air pressure and light intensity; constructing a sample based on a sliding time window, taking the environmental characteristics, the basic frequency sequence and the derived characteristics in the window as input, taking the actual frequency response corresponding to the forward step of the window as label, and constituting a multivariate feature dataset, and labeling the scene labels of ordinary scene and special scene according to the environmental grade threshold; wherein the derived characteristics include environmental vector components, environmental change rate, environmental stability index and light intensity change rate.

[0026] It can be understood that the dynamic coupling relationship between the wing vibration frequency of mosquitoes and external environmental variables is revealed through frequency-environment correlation modeling. The wing vibration frequency of mosquitoes is often influenced by factors such as temperature, humidity, air pressure, and light, showing nonlinear characteristics that change over time. Therefore, the method obtains the frequency response data of mosquitoes within a historical period and establishes a monotonic correspondence between frequency and environmental variables by combining the frequency distribution model that has been constructed. By using the reverse lookup and piecewise linear interpolation method, the response at different time points is interpolated on the frequency grid, so that the frequency change trend can be continuously and smoothly mapped onto the time axis, thereby generating a basic frequency sequence with time correlation. The essence of this process is to compress the multi-dimensional frequency response space into a function sequence that can be dynamically tracked over time, providing a continuous frequency baseline for subsequent feature fusion and prediction. Second, in order to achieve high-precision time synchronization between frequency and environmental variables, the method introduces a time series and numerical processing mechanism for environmental variables. By obtaining representative environmental variables (such as temperature, humidity, air pressure, and light), and de-circularizing angle data (such as solar elevation angle or wind direction angle) with periodic characteristics, and then using sine and cosine encoding for numericalization, the discontinuity caused by angle jumps can be effectively avoided. Aligning the environmental variables and the basic frequency sequence in time according to a uniform sampling step ensures consistency between the environmental input and the frequency output at the same time, providing a stable input data structure for subsequent machine learning models. Third, to enhance the robustness and feature expression ability of the model, the method introduces a sliding time window mechanism in the time series dimension. By sliding the window, training samples for consecutive periods can be constructed, and the environmental features, basic frequency sequence, and derived features extracted from the environmental data within the window form the input feature vector. Among them, the derived features include environmental vector components (used to describe the comprehensive directional changes of multiple environmental variables), environmental change rates (reflecting short-term fluctuation trends), environmental stability indicators (used to measure the stationarity of environmental fluctuations), and light intensity change rates (reflecting the energy changes of day-night dynamics). By aggregating these features in the time window, the system can better capture the dynamic evolution law of frequency changes with the environment. Finally, in order to realize scene-level adaptive modeling, the method labels the samples according to the environmental level threshold. By analyzing the distribution characteristics of environmental variables at a certain threshold, the data samples are divided into ordinary scenes and special scenes (such as high humidity, high temperature, or low light environment), so that the subsequent model can learn different frequency response patterns for different scenes, thereby enhancing the generalization ability and environmental adaptability of the system.

[0027] It can be seen that, in the frequency-environment dynamic mapping, time series synchronization, feature enhancement and scene classification, a multi-feature dataset containing frequency baseline, environmental input and derived indicators is constructed, which provides a complete, continuous and physically relevant input basis for subsequent deep learning models, and realizes the structured and intelligent modeling of the mosquito wing vibration frequency prediction process.

[0028] In step S300, the multi-feature dataset is input into a deep learning network to extract local features of the frequency signal, and the local features are input into a recurrent neural network. An adaptive weight mechanism is introduced in the feature sequence output by the recurrent neural network, the correlation of each time step feature is calculated to obtain a weight distribution, and the time step features are weighted and aggregated according to the weight distribution to obtain a global feature representation.

[0029] Specifically, when the multi-feature dataset is input into the deep learning network to extract the local features of the frequency signal, it includes: stacking the multi-feature dataset in the time dimension according to the sliding time window and channel splicing the environmental features and the basic frequency sequence in the feature dimension to form a one-dimensional time series tensor; a one-dimensional time series convolution deep learning network is used, causal padding is used in the first layer, multi-scale convolution kernels and optional dilated convolution are used in the middle layer, and the kernel length of the multi-scale convolution kernel is one or more of 3, 5 and 7; nonlinear activation is set after each convolution layer in turn, and residual connection is used to suppress deep degradation, and finally the local features are output.

[0030] Specifically, when the correlation of each time step feature is calculated to obtain an adaptive weight distribution, the time step features are weighted and aggregated according to the weight distribution to obtain a global feature representation, including: inputting the local features into a stacked recurrent neural network to model long-term dynamic dependencies to obtain a hidden state sequence; based on the adaptive weight mechanism, the correlation score is calculated, and the weight distribution is obtained by Softmax normalization; the context vector is obtained by weighting and aggregating the time step features according to the weight distribution, and the global feature representation is obtained by nonlinear projection of the context vector.

[0031] It can be understood that based on the principle of deep learning-based time sequence feature extraction, the local pattern recognition of mosquito wing vibration frequency signal is realized through multi-scale convolution structure. Specifically, after stacking the constructed multi-element feature dataset according to the sliding time window, the environmental features and the basic frequency sequence are spliced in the channel to form a continuous one-dimensional time sequence tensor in the feature dimension. The time sequence tensor retains the local change characteristics of the time sequence and the spatial correlation of the environmental factors in the structure, providing a high-dimensional input basis for subsequent convolution operations. By introducing a one-dimensional time sequence convolution network, causal padding is used to ensure that the convolution calculation of each time step only depends on its historical information, thereby maintaining the time causality and preventing future information from being leaked. In the middle layer, multi-scale convolution kernels (kernel length 3, 5, 7) and optional dilated convolution are used to extract short-term changes, periodic oscillations and long-period modulation features in different receptive field ranges, thereby enhancing the sensitivity of the model to frequency changes at different time scales. The residual connection structure relieves the gradient vanishing problem in deep network training through an identity mapping path, making the local feature extraction process more stable and effective. Finally, the local features extracted by the convolution layer represent the frequency distribution pattern of the mosquito wing vibration signal in the time segment and the environmental response characteristics. Secondly, in order to further capture the dynamic evolution rule of the frequency signal at a long time scale, the local features extracted by the convolution are input into a stacked recurrent neural network (RNN) for sequence modeling. The recurrent neural network has a memory and state transfer mechanism, which can retain the dependency relationship between the information before and after in the time dimension. Through the structure design of multi-layer stacking, the model can simultaneously learn short-term fluctuations and long-term trends, and realize dynamic response modeling of mosquito wing vibration frequency under complex environmental changes. This process is equivalent to establishing a nonlinear state transfer model in the time sequence of the frequency signal, providing dynamic memory support for global feature expression. Thirdly, the adaptive weight mechanism (Attention mechanism) is introduced in the output stage of the recurrent neural network to realize the importance evaluation of the time step features. Specifically, by calculating the correlation scores between the features of each time step and normalizing them through the Softmax function, the weight distribution of each time step is obtained. The weight size reflects the contribution of the time step feature to the overall frequency change pattern, thereby realizing the "feature focusing" in the time dimension. Then, all the time step features are weighted and aggregated according to the weight distribution to obtain the context vector (Context Vector), which can be regarded as a global summary of the information in the entire time window in the semantic level. Finally, the context vector is converted into a high-dimensional embedding representation through nonlinear projection to obtain the final global feature representation (Global Feature Representation).The global feature not only retains the detailed texture information in the local convolutional feature, but also integrates the time sequence dependence modeled by the recurrent neural network and the key moment information screened by the attention mechanism, thereby forming a feature representation with global consistency and semantic expressiveness.

[0032] It can be seen that the multi-level extraction and pattern capture of local features are realized by the multi-scale time sequence convolutional network; the long-term dependence of time series is modeled by the recurrent neural network, and the dynamic change rule of the frequency signal is mastered; the weight distribution and information focusing in the time dimension are realized by the adaptive attention mechanism, and the most representative dynamic features are extracted. The intelligent representation process of the frequency signal from local to global and from static to dynamic is effectively realized, so that the model can accurately capture the time-varying characteristics of the mosquito wing vibration frequency under different environmental conditions, and provide high-quality feature input and decision basis for the subsequent adaptive acoustic trapping strategy.

[0033] In step S400, two independent branch processing units are set, and the global feature representation is respectively input into the corresponding branch processing unit for frequency optimization adjustment to obtain the ordinary scene branch prediction value and the special scene branch prediction value.

[0034] Specifically, when the two independent branch processing units are set, the global feature representation is respectively input into the corresponding branch processing unit for frequency optimization adjustment to obtain the ordinary scene branch prediction value and the special scene branch prediction value, including: when the two independent branch processing units are set, including the ordinary scene branch processing unit and the special scene branch processing unit; each branch processing unit includes at least two fully connected layers and a nonlinear activation, and random deactivation is set between layers to suppress overfitting; in the training stage, only the corresponding branch processing unit is updated based on the scene label, and the other branch processing unit is frozen, and in the prediction stage, the global feature representation is respectively input into the two branch processing units to generate the ordinary scene branch prediction value and the special scene branch prediction value at the same time, and the quantile loss is used as a robust loss function to constrain the training error.

[0035] It can be understood that the difference of the wing vibration frequency adjustment law of mosquitoes in different environmental scenarios is learned through the multi-branch parallel modeling principle. Due to the significant difference in the wing vibration frequency response mechanism of mosquitoes in ordinary environment and special environment (such as high humidity, high temperature, low light or strong wind conditions), if a single model is used for unified prediction, it is easy to cause feature learning deviation and decline of model generalization performance. Therefore, the method designs two independent branch processing units: ordinary scene branch processing unit and special scene branch processing unit, which correspond to different environmental state intervals. The essence of this structure is a conditional modeling mechanism, which realizes the independent adaptation of the model to different environmental sub-distributions through branch differentiation. Secondly, the multi-layer full connection structure (MLP) and the nonlinear activation function are combined in each branch processing unit to enhance the nonlinear mapping ability of the model. The full connection layer is used for dimension transformation and feature interaction of global feature representation, and the nonlinear activation (such as ReLU or GELU) gives the network nonlinear expression ability in high-dimensional space, so that the model can capture complex frequency adjustment relationship. In addition, the dropout mechanism is introduced between layers to effectively suppress the overfitting phenomenon and improve the robustness of the model under different environmental samples. Thirdly, in the training mechanism of the model, the method adopts a scene label driven selective parameter update strategy. Specifically, during the training phase, when the input sample is labeled as an ordinary scene, only the ordinary scene branch is activated and its parameters are updated by back propagation, while the parameters of the special scene branch are frozen. Conversely, when the input sample belongs to a special scene, only the special branch parameters are updated and the ordinary branch remains unchanged. This selective training mechanism can prevent interference between different scene feature distributions, so that each branch network focuses on the feature rule learning of its own scene, thereby realizing a clearer decision boundary and stronger scene adaptability. This idea belongs to “distribution-conditioned learning” in neural network training theory, which effectively improves the generalization performance of the model under environmental heterogeneity. Fourthly, in the inference stage, that is, the prediction stage, the system inputs the global feature representation into the two branch processing units and outputs the ordinary scene branch prediction value and the special scene branch prediction value respectively. This design enables the model to still evaluate the frequency response trend under different environmental conditions in unknown or mixed scenarios, providing a comparable basis for the final output based on environmental weight fusion. Finally, in the optimization objective, the method adopts the quantile loss function as the robust loss function to constrain the training error of the model. Unlike the traditional mean square error, the quantile loss can reduce the influence of abnormal samples on the overall model when the frequency distribution has skewness or noise interference, so that the prediction result is more stable and has stronger noise resistance.This robust optimization principle guarantees the reliability and interpretability of the frequency prediction results under environmental uncertainty.

[0036] In step S500, the multivariate feature dataset is divided into several hierarchical regions according to the environmental variables, the frequency correction curves of each region are constructed based on the relationship between the historical frequency data and the actual frequency response, the environmental variables of the target period are input into the frequency correction curves of the target region and its adjacent regions to obtain the regional correction results, and the frequency correction value is obtained by inversely distance weighted fusion of the regional correction results of the adjacent regions.

[0037] Specifically, when the multivariate feature dataset is divided into several hierarchical regions according to the environmental variables, and the frequency correction curves of each region are constructed based on the relationship between the historical frequency data and the actual frequency response, it includes: dividing the environmental circumference into regions with equal angular width with the environmental variable as the angle variable, and extending the regions according to the frequency level to form a hierarchical structure; in each region, the historical frequency data is used as the independent variable and the actual frequency response is used as the dependent variable to obtain the frequency correction curve by monotonic constraint regression fitting.

[0038] Specifically, when the environmental variables of the target period are input into the frequency correction curves of the target region and its adjacent regions to obtain the regional correction results, and the frequency correction value is obtained by inversely distance weighted fusion of the regional correction results of the adjacent regions, it includes: determining the target region and its adjacent regions corresponding to the environmental variables of the target period, and substituting the environmental variables of the target period into the corresponding frequency correction curves to obtain multiple regional correction results; the inverse distance weight is calculated according to the included angle between the axis of each region and the environmental variable, and the regional correction results are weighted and summed to obtain the frequency correction value.

[0039] It can be understood that the present technology takes the environmental variable as the division basis to construct a "hierarchical regionalization" feature space. Specifically, the environmental variables (such as temperature and humidity, air pressure, wind speed, sound wave propagation parameters, etc.) are regarded as multi-dimensional input and projected into an environmental circumference model with angle variable as the main axis. The circumference is evenly divided into several regions with equal angular width to represent different types of environmental feature intervals; at the same time, it is divided into multiple rings along the radial direction according to the frequency level, thereby forming a hierarchical regional structure with angle dimension (environmental type) and radius dimension (frequency level). This design enables the model to simultaneously realize local fine modeling and global distribution expression between different environmental categories and frequency levels, which facilitates capturing the influence law of environmental changes on frequency response.

[0040] Secondly, within each partition, the system performs a monotonic constrained regression fitting to generate a frequency correction curve based on the correspondence between historical frequency data and actual frequency response. The introduction of monotonic constraint ensures that the frequency correction curve maintains physical reasonableness and interpretability between independent variables (historical frequency) and dependent variables (actual response), avoiding non-physical oscillations caused by noise or local anomalies. This regression process is equivalent to establishing an adaptive frequency correction model for each environmental level, describing the stable deviation relationship between historical frequency input and actual frequency output in that environment.

[0041] When performing real-time correction, the system first determines the target region and its adjacent regions according to the environmental variables of the target period. Then, the environmental variables of the target period are substituted into the frequency correction curves corresponding to these regions, obtaining the correction results of multiple regions. This step realizes multi-source prediction of frequency correction, i.e., referring to the frequency response characteristics of the target region and its adjacent environment at the same time, to enhance the continuity and robustness of the model.

[0042] Finally, based on the inverse distance weighting (IDW, Inverse Distance Weighting) principle, the region correction results are fused. Specifically, the angle between the target environmental variable and the central axis of each region is calculated, and the inverse ratio of the angle is used as the weight, giving higher weight to regions closer to the target environment, thus achieving a locally continuous and smooth correction surface in mathematics. The frequency correction value obtained after weighted summation not only maintains the dominant role of the target region characteristics, but also takes into account the gradual change characteristics of the surrounding environment, significantly improving the smoothness and accuracy of the correction.

[0043] Step S600, obtaining the final frequency output value according to the weighted fusion of the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value.

[0044] Specifically, when obtaining the final frequency output value according to the weighted fusion of the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value, it includes: calculating the performance indicators of the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value respectively, and determining the subjective weight and the objective weight according to the performance indicators, obtaining the combined weight by weight normalization and non-negative constraint, and the sum of all weights in the combined weight is 1; calculating the final frequency output value according to the combined weight and the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value.

[0045] It can be understood that due to the difference in frequency response mode under different scenarios, a single branch model (such as only ordinary scene or only special scene prediction) is difficult to maintain accuracy and robustness in complex environment. Therefore, the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value obtained based on the environment correction model are used at the same time. The ordinary scene branch is used to represent the frequency trend under normal environment, the special scene branch is optimized to compensate for extreme or disturbance conditions, and the frequency correction value reflects the correction effect of external environmental variables on the frequency characteristics. Through the fusion of the three, the double-layer robust structure of “scene self-adaptation + external environment compensation” is realized at the model level.

[0046] Secondly, in the fusion process, instead of directly averaging multiple prediction results, a weighted optimization mechanism combining subjective and objective weights is introduced. The subjective weight is usually derived from the designer's or prior experience in the training stage, and is used to reflect the importance of each branch in different scenarios; the objective weight is automatically calculated based on the model performance indicators (such as prediction error, stability, correlation coefficient, confidence interval, etc.), and is used to reflect the actual performance of each branch in the current period. By combining subjective and objective weights, the data-driven dynamic adjustment capability can be introduced while maintaining the controllability of the model experience.

[0047] In order to ensure the mathematical consistency and interpretability of the fusion process, further normalization and non-negative constraint are performed on the weights. The normalization operation ensures that the sum of all weights is equal to 1, so that the fusion result has numerical comparability and physical meaning; the non-negative constraint prevents the weight from appearing reverse cancellation effect, and ensures that the contribution of each branch is reflected in the final result in a positive form. The combined weight obtained after constrained optimization can stably reflect the relative contribution of different prediction sources in the current environment.

[0048] Finally, according to the combined weight, the ordinary scene branch prediction value, the special scene branch prediction value and the frequency correction value are weighted and summed to output the final frequency output value. This fusion process is essentially a linear weighted integration model based on performance adaptive adjustment, which can automatically balance the influence of multi-source prediction in different scenarios. For example, when in stable environment, the weight of ordinary scene branch will dominate; while in high noise or special interference conditions, the weights of special scene branch and frequency correction term will dynamically increase, realizing adaptive compensation.

[0049] It can be seen that through the subjective and objective weight optimization and constraint fusion mechanism based on performance indicators, the frequency prediction results and environment correction results are dynamically weighted and integrated, so that more stable and accurate frequency prediction values can be output in different environments and scene conditions. This fusion strategy not only improves the generalization ability and scene adaptability of the model, but also ensures the physical reasonableness and algorithm robustness of the prediction results.

[0050] In the above embodiment, by acquiring multi-band data of mosquito wing vibration frequency and preprocessing, the system can not only construct a complete frequency distribution model, but also eliminate noise and abnormal data, so that the basic frequency sequence is closer to the real vibration characteristics of mosquitoes. Combined with environmental variables to form a multivariate feature dataset, the system can consider the influence of temperature, humidity, light and other environmental factors on mosquito behavior at the same time, and realize the fine analysis of mosquito dynamic behavior, which significantly improves the scientificity and accuracy of frequency signal generation, and solves the problem that traditional fixed frequency trapping technology is difficult to cope with behavior changes. Secondly, the method uses a mechanism combining deep learning network and recurrent neural network to realize intelligent optimization of frequency signal. By extracting local features of frequency signal through deep learning network, the system can capture tiny vibration patterns and instantaneous frequency changes; the recurrent neural network introduces an adaptive weight mechanism in the time series, which can calculate the correlation of each time step feature and weight and converge the key features to obtain global feature representation. This process can effectively identify the influence of behavior rules and environmental interference factors, so that the predicted frequency signal has real-time performance and global optimization ability, greatly improving the pertinence and effectiveness of acoustic trapping. Finally, the method realizes multi-scene adaptive optimization through two independent branch processing units and regional correction mechanism. The ordinary scene branch and the special scene branch can generate predicted values for regular indoor and outdoor environments and special environments (such as water surface or complex terrain) respectively, ensuring that reasonable acoustic frequency output can be obtained in different scenes. At the same time, based on the historical frequency data and the actual response, a regional frequency correction curve is constructed to finely correct each environmental level, and the correction results of adjacent regions are fused through inverse distance weighting to realize dynamic adaptation of environmental variables. This not only enhances the stability of the system in complex environments, but also prolongs the long-term use effect of the acoustic trapping device, avoiding the decline of trapping efficiency caused by frequency deviation.

[0051] In another preferred mode based on the above embodiment, as shown in Figure 3 The embodiment provides a dynamic adaptive mosquito wing vibration frequency acoustic trapping system, comprising: The acquisition module is configured to acquire multi-band data of mosquito wing vibration frequency, and preprocess the multi-band data to construct a frequency distribution model; The correlation module is configured to generate a basic frequency sequence according to the frequency distribution model, and time series correlate the basic frequency sequence with environmental variables, while collecting environmental characteristics in the corresponding period to form a multivariate feature dataset; The first processing module is configured to input the multi-element feature data set into a deep learning network to extract local features of the frequency signal, and input the local features into a recurrent neural network, introduce an adaptive weight mechanism in a feature sequence output by the recurrent neural network, calculate the correlation of features at each time step and obtain a weight distribution, weight and aggregate the time step features according to the weight distribution, and obtain a global feature representation; The second processing module is configured to set two independent branch processing units, input the global feature representation into the corresponding branch processing units for frequency optimization adjustment, and obtain a normal scene branch prediction value and a special scene branch prediction value. The third processing module is configured to divide the multi-element feature data set into a plurality of hierarchical regions according to environmental variables, construct a frequency correction curve of each region based on a relationship between historical frequency data and an actual frequency response, input an environmental variable of a target period into the frequency correction curve of a target region and adjacent regions to obtain a region correction result, and obtain a frequency correction value by inversely distance weighted fusion of the region correction results of the adjacent regions. The correction module is configured to obtain a final frequency output value by weighted fusion of the normal scene branch prediction value, the special scene branch prediction value and the frequency correction value.

[0052] It can be understood that the dynamic adaptive mosquito wing vibration frequency acoustic trapping system and method in each of the above embodiments have the same beneficial effects, which will not be described again.

[0053] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks

[0055] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0057] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the same. Even though the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently, and any modification or replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A dynamic adaptive acoustic trapping method for mosquito wing vibration frequency, characterized in that, include: Multi-band data of mosquito wing vibration frequency were acquired, and the multi-band data were preprocessed to construct a frequency distribution model. A basic frequency sequence is generated based on the frequency distribution model, and the basic frequency sequence is correlated with environmental variables over time. At the same time, environmental characteristics of the corresponding time period are collected to form a multivariate feature dataset. The multivariate feature dataset is input into a deep learning network to extract local features of frequency signals. The local features are then input into a recurrent neural network. An adaptive weighting mechanism is introduced into the feature sequence output by the recurrent neural network to calculate the correlation of features at each time step and obtain the weight distribution. The time step features are then weighted and converged according to the weight distribution to obtain a global feature representation. Two independent branch processing units are set up, and the global feature representation is input into the corresponding branch processing units for frequency optimization and adjustment to obtain the branch prediction value of ordinary scene and the branch prediction value of special scene. The multivariate feature dataset is divided into several hierarchical regions according to environmental variables. Frequency correction curves for each region are constructed based on the relationship between historical frequency data and actual frequency response. Environmental variables for the target time period are input into the frequency correction curves corresponding to the target region and its adjacent regions to obtain regional correction results. Frequency correction values ​​are obtained by fusing the regional correction results of adjacent regions through inverse distance weighting. The final frequency output value is obtained by weighted fusion of the predicted values ​​of the ordinary scenario branch, the predicted values ​​of the special scenario branch, and the frequency correction value.

2. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 1, characterized in that, Preprocessing multi-band data to construct a frequency distribution model includes: The frequency data of each frequency band is normalized at intervals along the frequency axis to unify the frequency grid. Linear interpolation is performed on the missing data at the corresponding frequency points and duplicate or outlier points are removed. Constraints are set at the frequency boundaries according to the maximum response range and boundary values ​​are filled to ensure data continuity, thereby obtaining optimized data for each frequency band. At each unified frequency point, the optimized data are summed, and the summation is obtained by traversing the frequency grid to obtain the frequency distribution model.

3. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 2, characterized in that, When generating a basic frequency sequence based on a frequency distribution model, and then correlating the basic frequency sequence with environmental variables over time, while simultaneously collecting environmental characteristics for the corresponding time periods to form a multivariate feature dataset, the dataset includes: The frequency response is acquired within a historical period and a frequency-environment monotonic correspondence is established based on the frequency distribution model. The basic frequency sequence at each moment is obtained on the frequency grid based on reverse lookup and piecewise linear interpolation. Data representing environmental variables is acquired, the angle data is de-looped and quantified using sine and cosine encoding, and time-aligned with the fundamental frequency sequence using a uniform sampling step size. Environmental features were collected on the same timeline as historical periods and normalized. These environmental features included temperature, humidity, air pressure, and light intensity. Samples are constructed based on a sliding time window. The environmental features, basic frequency sequences, and derived features within the window are used as inputs, and the actual frequency response corresponding to the window's forward step length is used as a label to form a multivariate feature dataset. Scene labels for ordinary and special scenes are labeled according to environmental level thresholds. Among them, derived features include environmental vector components, environmental change rate, environmental stability index, and light intensity change rate.

4. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 3, characterized in that, When inputting a multivariate feature dataset into a deep learning network to extract local features of a frequency signal, the following are included: The multivariate feature dataset is stacked in the time dimension by a sliding time window, and the environmental features and the fundamental frequency sequence are concatenated in the feature dimension to form a one-dimensional time series tensor. A deep learning network with one-dimensional temporal convolution is used, and causal padding is used in the first layer. Multi-scale convolution kernels and optional dilated convolution are used in the middle layers. The kernel length of the multi-scale convolution kernel is one or more of 3, 5, and 7. After each convolutional layer, nonlinear activation is applied sequentially, and deep degradation is suppressed through residual connections, ultimately outputting local features.

5. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 4, characterized in that, When calculating the correlation of features at each time step and obtaining an adaptive weight distribution, and then weighting and converging the features at each time step according to the weight distribution to obtain the global feature representation, the process includes: By modeling long-term dynamic dependencies through stacked local feature inputs into a recurrent neural network, a sequence of hidden states can be obtained. The relevance score is calculated based on an adaptive weighting mechanism, and the weight distribution is obtained by Softmax normalization. The context vector is obtained by weighted aggregation of the time step features based on the weight distribution, and then the context vector is obtained by nonlinear projection to obtain the global feature representation.

6. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 5, characterized in that, Two independent branch processing units are set up, and the global feature representation is input into the corresponding branch processing unit for frequency optimization adjustment. When obtaining the branch prediction values ​​for ordinary scenes and special scenes, the following are included: When setting up two independent branch processing units, there are two types of processing units: a normal scene branch processing unit and a special scene branch processing unit. Each branch processing unit includes at least two fully connected layers and nonlinear activation, and random deactivation is set between layers to suppress overfitting. During the training phase, the parameters of the corresponding branch processing unit are updated only based on the scene label while the other branch processing unit is frozen. During the prediction phase, the global feature representation is fed forward into the two branch processing units respectively to generate the prediction values ​​of ordinary scene branches and the prediction values ​​of special scene branches simultaneously. Quantile loss is used as a robust loss function to constrain the training error.

7. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 6, characterized in that, When dividing a multivariate feature dataset into several hierarchical regions based on environmental variables, and constructing frequency correction curves for each region based on the relationship between historical frequency data and actual frequency response, the following steps are taken: Using environmental variables as angular variables, the environmental circumference is divided into regions of equal angular width, and the outer extension of the region is divided into several rings according to frequency level to form a hierarchical structure. Within each region, a frequency correction curve is obtained by performing monotonic constrained regression fitting with historical frequency data as the independent variable and the actual frequency response as the dependent variable.

8. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 7, characterized in that, When obtaining regional correction results by inputting environmental variables for the target time period into the frequency correction curves corresponding to the target area and its adjacent areas, and obtaining frequency correction values ​​by fusing the regional correction results of adjacent areas through inverse distance weighting, the following steps are taken: Determine the target region and its adjacent regions corresponding to the environmental variables of the target time period, and substitute the environmental variables of the target time period into the corresponding frequency correction curves to obtain correction results for multiple regions. The inverse distance weights are calculated based on the angle between the central axis and environmental variables in each region, and the frequency correction values ​​are obtained by weighted summation of the regional correction results.

9. The dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in claim 8, characterized in that, When obtaining the final frequency output value by weighted fusion of the normal scenario branch prediction value, the special scenario branch prediction value, and the frequency correction value, the following are included: Performance indices are calculated for the predicted values ​​of branches in ordinary scenarios, predicted values ​​of branches in special scenarios, and frequency correction values. Subjective weights and objective weights are determined based on the performance indices. Weight normalization and non-negativity constraints are used to obtain combined weights. The sum of all weights in the combined weights is 1. The final frequency output value is calculated based on the combined weights, the predicted values ​​of the branches in the normal scenario, the predicted values ​​of the branches in the special scenario, and the frequency correction value.

10. A dynamic adaptive mosquito wing vibration frequency acoustic trapping system, used in applying the dynamic adaptive mosquito wing vibration frequency acoustic trapping method as described in any one of claims 1-9, characterized in that, include: The acquisition module is configured to acquire multi-band data of mosquito wing vibration frequency and preprocess the multi-band data to construct a frequency distribution model. The association module is configured to generate a basic frequency sequence based on the frequency distribution model, associate the basic frequency sequence with environmental variables over time, and collect environmental features for the corresponding time period to form a multivariate feature dataset. The first processing module is configured to input a multivariate feature dataset into a deep learning network to extract local features of frequency signals, input the local features into a recurrent neural network, introduce an adaptive weight mechanism into the feature sequence output by the recurrent neural network, calculate the correlation of features at each time step and obtain the weight distribution, and perform weighted aggregation of the time step features according to the weight distribution to obtain a global feature representation. The second processing module is configured to set up two independent branch processing units, and input the global feature representation into the corresponding branch processing units for frequency optimization and adjustment to obtain the branch prediction value of ordinary scene and the branch prediction value of special scene. The third processing module is configured to divide the multivariate feature dataset into several hierarchical regions according to environmental variables, construct frequency correction curves for each region based on the relationship between historical frequency data and actual frequency response, input the environmental variables of the target time period into the frequency correction curves corresponding to the target region and its adjacent regions to obtain the regional correction results, and obtain the frequency correction value by fusing the regional correction results of adjacent regions through inverse distance weighting. The correction module is configured to obtain the final frequency output value by weighted fusion of the normal scene branch prediction value, the special scene branch prediction value and the frequency correction value.

Images