Agricultural management and control method and system based on Internet of Things cloud platform

Abstract

The invention discloses an agricultural management and control method and system based on an Internet of Things cloud platform, and relates to the technical field of agricultural management and control, and the method comprises the steps: carrying out the distributed deployment of edge nodes with audio collection equipment, carrying out the preprocessing of collected audio data, obtaining an environment label through combining dual-mode positioning, and forming an initial audio data set with a label; performing time-frequency transformation and normalization on the initial audio signal, extracting enhanced features through a plurality of neural networks, generating dynamic mask suppression noise in combination with a noise prior feature library, and performing reconstruction to obtain a pure pest audio signal; extracting and splicing frequency domain and time domain features of the pure pest audio, and after compression, associating environment tags to form a feature tag data set; the edge node uploads a data set to the cloud, the cloud verifies and decompresses the data set, then fuses the features and screens core features, identifies pest types and density levels through a pre-trained CNN classifier, and determines an identification result according to confidence and a feature matching degree, so that the real-time performance and reliability of agricultural pest monitoring are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural management technology, specifically to an agricultural management method and system based on an Internet of Things (IoT) cloud platform. Background Technology

[0002] With the rapid development of IoT and cloud computing technologies, agricultural management is gradually moving towards intelligence and precision. In existing technologies, agricultural pest monitoring mainly relies on manual inspections or traditional sensor networks, such as indirectly inferring pest activity through image recognition or environmental sensors. Some studies have attempted to use audio acquisition devices to capture sound signals generated by pests biting or moving, and combine them with simple signal processing algorithms, such as Fourier transform or threshold detection, for preliminary analysis. These methods can provide clues about the presence of pests to some extent.

[0003] However, existing technologies have significant limitations. Traditional sensor networks and manual inspections are inefficient, making it difficult to achieve large-scale, real-time pest monitoring. They are also susceptible to subjective factors and environmental interference. For example, image recognition performs poorly at night or under complex lighting conditions, and environmental sensors cannot directly capture the acoustic features of pests. Secondly, existing audio analysis methods typically lack effective noise suppression, resulting in low signal-to-noise ratios and inaccurate feature extraction. Most systems do not incorporate environmental tagging and multimodal data fusion technologies, making it impossible to dynamically adapt to the needs of different agricultural scenarios, and the accuracy and robustness of the recognition results are insufficient. Furthermore, existing technologies have weak cloud collaboration capabilities. The integration of edge computing and cloud analysis has not yet formed an efficient closed loop. After data is uploaded, the cloud typically only performs simple classification or storage, lacking in-depth feature optimization and dynamic verification mechanisms. At the same time, the determination of pest density levels often relies on single features or static thresholds, making it difficult to adapt to the dynamic changes of different pest species and crop growth stages. These shortcomings limit the practical application of existing technologies in precision agriculture. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an agricultural management and control method and system based on an Internet of Things (IoT) cloud platform. By distributing edge nodes to collect pest audio data, combining complex neural networks and dynamic noise suppression technology to extract pure voiceprint features, and utilizing cloud-based collaborative analysis to achieve accurate identification of pest species and density levels, this invention solves the problems of low efficiency, high noise interference, and insufficient identification accuracy of traditional methods, significantly improving the real-time performance and reliability of agricultural pest monitoring.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an agricultural management and control method based on an Internet of Things (IoT) cloud platform, comprising: Distributed deployment of edge nodes with audio acquisition devices; after preprocessing the acquired audio data, environmental labels are obtained by combining dual-mode positioning to form an initial audio dataset with environmental labels. The initial audio signal is subjected to time-frequency transformation and normalization. Enhancement features are extracted through complex neural network. Dynamic mask is generated by combining noise prior feature library to suppress noise and reconstruct a pure pest audio signal. Frequency and time domain features of pure pest audio signals are extracted and spliced ​​together. After outlier removal, dimensionality reduction and normalization optimization, the compressed data is associated with environmental labels to form a feature label dataset. Edge nodes upload feature label datasets to the cloud. The cloud verifies and decompresses the data, then fuses the features and selects core features. A pre-trained CNN classifier is used to identify pest species and density levels. The identification results are determined based on confidence and feature matching.

[0006] Furthermore, preprocessing operations are performed on the collected raw audio data: the least squares method is used to fit the linear trend term of the raw audio data, and the fitted trend term is subtracted from the raw audio data to eliminate signal drift; a pre-emphasis algorithm implemented with a first-order FIR filter is used to adjust the signal strength of the insect biting sound frequency band; the audio signal is framed and windowed to finally obtain the initial audio signal after framing; the window function used is the Hanning window.

[0007] Furthermore, the environmental labels of the region are obtained. The environmental labels include basic labels and noise labels. Using the frame number of the initial audio signal after framing as the key and the corresponding environmental label as the value, the environmental labels are associated with the initial audio signal after framing to form a labeled initial audio dataset.

[0008] Furthermore, a short-time Fourier transform is performed on the initial audio signal of each frame to obtain a complex time-frequency matrix containing amplitude and phase information; L2 norm amplitude normalization is performed on the complex time-frequency matrix to map the amplitude values ​​of all elements to the interval [0, 1]; features are extracted from the normalized complex time-frequency matrix through a complex neural network; wherein the complex neural network contains a dual-channel input layer and two levels of complex convolutional blocks to obtain advanced composite time-frequency features that fuse amplitude and phase information.

[0009] Furthermore, based on the noise type in the environmental label, a cross-regional noise prior feature library is invoked. This feature library contains complex time-frequency feature templates for agricultural environmental noise. The similarity between the advanced composite time-frequency features and the corresponding noise templates is calculated using a cosine similarity algorithm. Frequency points with similarity exceeding the corresponding similarity threshold are marked as noise frequency points, generating a binary time-frequency mask with dimensions completely consistent with the complex time-frequency matrix. Positions with values ​​of 0 in the mask correspond to masked noise frequency points, while positions with values ​​of 1 correspond to retained target sound signature frequency points. The generated binary time-frequency mask is multiplied element-wise with the normalized complex time-frequency matrix to obtain a pure complex time-frequency matrix containing only the sounds of pests biting. An inverse short-time Fourier transform is performed on the pure complex time-frequency matrix, and a moving average filtering algorithm is used to smooth the reconstructed signal, ultimately obtaining a high-quality pure pest audio signal.

[0010] Further, frequency domain feature extraction involves performing a Fourier transform on the pure insect audio signal to obtain the power spectrum, then filtering the power spectrum using a Mel filter bank, taking the logarithm of the filtering result, and performing a discrete cosine transform to obtain MFCC coefficients. Based on the MFCC coefficients, the difference between adjacent frames is calculated to obtain the first-order difference coefficients, and then the difference between adjacent frames is calculated from the first-order difference coefficients to obtain the second-order difference coefficients, ultimately forming the MFCC feature vector. Temporal feature extraction uses a sliding window method to simultaneously extract core temporal features: short-time energy, short-time zero-crossing rate, spectral entropy, and peak factor. The average value of each core temporal feature is taken according to the number of window slides to form a temporal feature vector. The MFCC feature vector and the temporal feature vector are concatenated in dimensional order to obtain the initial voiceprint feature vector.

[0011] Furthermore, a three-level optimization process is performed on the initial voiceprint feature vector: outliers are removed using the 3σ criterion, feature dimensionality is reduced using principal component analysis to obtain a multidimensional optimized voiceprint feature vector, and feature normalization is performed using the min-max normalization method; the binary data of the normalized multidimensional optimized voiceprint feature vector is compressed; and a one-to-one correspondence is established using the compressed feature data as the value and the environment label as the key to form a feature label dataset.

[0012] Furthermore, discrete labels in the environmental tags are converted into quantized feature vectors using one-hot encoding; continuous labels are converted into standard normal distribution quantized values ​​using Z-score normalization, and then concatenated to form a multidimensional label quantized feature vector; based on the attention mechanism, the multidimensional optimized voiceprint feature vector and the multidimensional label quantized feature vector are weighted by attention weights obtained through pre-training, and a multidimensional fused feature vector is output; the minimum redundancy maximum correlation algorithm is used to screen the fused feature vector, calculate the correlation coefficient between each feature dimension and the pest species and the redundancy coefficient between features, and select feature dimensions with a correlation coefficient ≥ 0.6 and a redundancy coefficient ≤ 0.3, finally obtaining a multidimensional standardized core feature vector.

[0013] Furthermore, a pre-defined lightweight CNN classifier is invoked. The network structure consists of two convolutional layers, one max-pooling layer, one fully connected layer, and one output layer. After inputting standardized core feature vectors, the output is the recognition confidence score for various pests. The recognition results include the pest species and density level. If the recognition confidence score is ≥85% and the feature matching degree is ≥80%, it is judged as a valid recognition result. If 50%≤confidence score<85% or 60%≤matching degree<80%, it is marked as a suspicious result. If the confidence score is <50% or the matching degree is <60%, it is judged as an invalid result.

[0014] An agricultural management and control system based on an Internet of Things (IoT) cloud platform includes: The edge node acquisition module is a distributed deployment of edge nodes equipped with audio acquisition devices. After preprocessing the acquired audio data, it combines dual-mode positioning to obtain environmental labels, forming an initial audio dataset with environmental labels. The signal processing module performs time-frequency transformation and normalization on the initial audio signal, extracts enhancement features through a complex neural network, generates a dynamic mask to suppress noise by combining a noise prior feature library, and reconstructs a pure pest audio signal. The feature engineering module extracts and concatenates the frequency and time domain features of the pure pest audio signal. After outlier removal, dimensionality reduction and normalization optimization, the compressed signal is associated with environmental labels to form a feature label dataset. The cloud-based analysis module allows edge nodes to upload feature label datasets to the cloud. After cloud verification and decompression, features are fused and core features are selected. A pre-trained CNN classifier is used to identify pest species and density levels, and the identification results are determined based on confidence and feature matching.

[0015] (III) Beneficial Effects This invention provides an agricultural management and control method and system based on an Internet of Things (IoT) cloud platform, which has the following beneficial effects: (1) By deploying edge nodes in a distributed manner and combining them with dual-mode positioning to obtain environmental tags, efficient and accurate audio data acquisition and preprocessing were achieved, improving the coverage and real-time performance of data acquisition and ensuring no blind spots. The sound signal of pests biting was enhanced by pre-emphasis and frame-by-frame windowing, and noise interference was suppressed. Combined with environmental tags, different agricultural scenarios were dynamically adapted, laying the foundation for subsequent noise suppression and feature extraction, which significantly improved the accuracy and reliability of pest monitoring.

[0016] (2) By using time-frequency transformation, complex neural network and dynamic noise suppression technology, the processing effect of pest audio signal is significantly improved. The complex neural network is used to extract advanced composite time-frequency features that fuse amplitude and phase, which enhances the signal representation ability. Combined with the noise prior feature library, a dynamic mask is generated to accurately separate noise and target soundprint, which greatly improves the signal-to-noise ratio. The pure pest audio signal is reconstructed by inverse time-frequency transformation, which provides a high-quality data foundation for subsequent feature extraction. This effectively solves the problems of large noise interference and inaccurate feature extraction in traditional methods, and improves the reliability of pest monitoring.

[0017] (3) By extracting the frequency and time domain features of pure insect audio and optimizing the processing, the accuracy and efficiency of insect identification are significantly improved. Multidimensional voiceprint features are formed by combining MFCC and time domain features, which enhances the representation ability of insect voiceprints. Outliers are removed by the 3σ criterion, PCA dimensionality reduction and normalization are used to optimize feature quality, reduce redundant interference, compress data and associate environmental labels, reduce transmission and storage overhead, and provide a high signal-to-noise ratio and low redundancy feature label dataset for cloud analysis, thereby improving the recognition accuracy of subsequent classifiers and the overall response speed of the system.

[0018] (4) Through cloud-based collaborative analysis, the accuracy and efficiency of pest identification are achieved. CRC-32 verification ensures the integrity of data transmission. The core features are screened by combining the attention mechanism and mRMR algorithm, which significantly improves the quality of features. A lightweight CNN classifier is used to output the pest species and density level. The identification results are verified by confidence and feature matching degree, which improves the classification accuracy. The validity of the results is dynamically determined and instructions are fed back, forming a closed-loop optimization process, which enhances the real-time performance and reliability of the system and provides efficient decision support for precision agricultural management. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the steps of the agricultural management and control method based on the Internet of Things cloud platform of the present invention; Figure 2 This is a schematic diagram of the agricultural management and control method based on the Internet of Things cloud platform of the present invention; Figure 3 This is a schematic diagram of the agricultural management and control system based on the Internet of Things cloud platform of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-2 This invention provides an agricultural management and control method based on an Internet of Things (IoT) cloud platform, comprising the following steps: Step 1: Distribute edge nodes with audio acquisition devices, collect and preprocess audio data, and combine dual-mode positioning to obtain environmental labels to form an initial audio dataset with environmental labels; Step one includes the following: Step 101: Determine the preset area division rules based on the terrain type of the area to be controlled. In plains areas, each control sub-area is divided into 50 mu (approximately 3.3 hectares), and in hilly / terraced areas, each control sub-area is divided into 30 mu (approximately 2 hectares). Distribute edge nodes in the agricultural areas to be controlled according to the preset area division rules. The audio acquisition device of each edge node faces the middle layer of the crop canopy, 20cm~30cm away from the top of the crop. The pickup angle of the audio acquisition device is set to 60°~90°, and the installation height is 1.2m~1.5m above the ground. The acquisition range of adjacent edge nodes is set with a 10%~15% overlap area to ensure no blind spots. At the same time, each edge node's audio acquisition device is equipped with a windproof and dustproof acoustic shell to reduce external environmental interference such as wind, sand and airflow during the acquisition process. Step 102: Configure audio acquisition parameters, specifically: sampling rate of 32kHz, acquisition duration of 8s / segment; acquisition trigger mode set to dual mode of timed acquisition and threshold trigger, wherein the timed acquisition cycle is once every 15 minutes, and the threshold trigger condition is that the short-time energy of the audio signal exceeds 3 times the preset baseline value; the preset baseline value is determined by pre-acquiring several sets of background audio data in an insect-free environment locally at the edge node, each set of pre-acquired data has a duration of 8s, the short-time energy of each set of data is calculated using the sliding window method, the window length is 256 points, and the average value of the short-time energy of several sets of data is taken as the preset baseline value; Step 103: Perform three rounds of preprocessing on the acquired raw audio data: First, a linear trend removal algorithm implemented using the least squares method is used to remove the DC component from the audio data. This is achieved by fitting the linear trend term of the original audio data and subtracting the fitted trend term from the original data to eliminate signal drift. Second, a pre-emphasis algorithm implemented using a first-order FIR filter is used to adjust the signal strength of the insect bite sound frequency band. The pre-emphasis coefficient is determined according to the crop type, ranging from 0.93 to 0.97. For example, 0.95 is used for gramineous crops and 0.97 for broadleaf crops. Third, the preprocessed signal is framed and windowed. The frame length is set to 30ms, the frame shift to 15ms, and the Hanning window is used to reduce inter-frame spectral leakage. Finally, the initial audio signal after frame segmentation is obtained. Step 104: Obtain the environmental labels of the area through GPS and BeiDou dual-mode positioning integrated in the edge node. The environmental labels include basic labels and noise labels. The basic labels are: area type, such as plain / hill / terraced fields; latitude and longitude range; crop type, such as rice / wheat / corn, etc. The noise labels are: main noise type, such as wind / farm machinery / stream / human voice; noise intensity level, detected by the sound level meter built into the edge node, with an average value ≤55dB within 10s indicating low intensity, 55dB < average value ≤70dB indicating medium intensity, and average value >70dB indicating high intensity; noise temporal characteristics, continuous / intermittent, determined by detecting the duration ratio of the noise signal, with a ratio ≥80% indicating continuous. Using the frame number of the initial audio signal after framing as the key and the corresponding environmental label as the value, associate the environmental label with the initial audio signal after framing to form a labeled initial audio dataset, which is stored in the local cache of the edge node.

[0022] When using this method, refer to steps 101 to 104: By deploying edge nodes in a distributed manner and combining them with dual-mode positioning to obtain environmental tags, we have achieved efficient and accurate audio data acquisition and preprocessing, improved the coverage and real-time performance of data acquisition, ensured no blind spots, enhanced the insect biting sound signal through pre-emphasis and frame-by-frame windowing processing, suppressed noise interference, and dynamically adapted to different agricultural scenarios by combining environmental tags, laying the foundation for subsequent noise suppression and feature extraction, which has significantly improved the accuracy and reliability of pest monitoring.

[0023] Step 2: Perform time-frequency transformation and normalization on the initial audio signal, extract enhancement features through a complex neural network, generate a dynamic mask to suppress noise by combining the noise prior feature library, and reconstruct a pure pest audio signal. Step two includes the following: Step 201: Obtain the labeled initial audio dataset and perform Short Time Fourier Transform (STFT) processing on each frame of the initial audio signal. The specific parameters and processing methods are as follows: the frame length is set to 960 points, the frame shift is set to 480 points, and the window function is the Hanning window. The generation rule of the Hanning window is: for each sampling point within the frame length, the window function value is calculated by subtracting 0.5 times the cosine value. The angle of the cosine value is 2π multiplied by the ratio of the sampling point index to the frame length minus 1. The STFT processing process is as follows: the time-domain audio signal of each frame is processed by the Hanning window, and then the processed signal is subjected to Fourier Transform to finally obtain a complex time-frequency matrix containing amplitude and phase information. The complex time-frequency matrix is ​​subjected to L2 norm amplitude normalization processing: the L2 norm of the complex time-frequency matrix of each frame is calculated, that is, the square root of the sum of the squares of the magnitudes of all elements in the matrix, and then each element in the matrix is ​​divided by the L2 norm of the corresponding frame to map the amplitude values ​​of all elements to the interval [0,1]. Step 202: Feature extraction and enhancement of the normalized complex time-frequency matrix are performed using a complex neural network. This neural network model is deployed based on the TensorFlow Lite framework, and the specific process and parameters are as follows: The input layer adopts a dual-channel parallel structure, receiving the real and imaginary parts of the complex time-frequency matrix respectively; the first complex convolution block is processed using 3×3 complex convolution kernels, with a number of 96, a convolution stride of 1, and a padding method of SAME; complex convolution operations are performed synchronously on the real and imaginary parts of the matrix, with the following rules: the real part of the convolution output is equal to the product of the real part of the convolution kernel and the real part of the input minus the product of the imaginary part of the convolution kernel and the imaginary part of the input; the imaginary part of the convolution output is equal to the product of the real part of the convolution kernel and the imaginary part of the input plus the product of the imaginary part of the convolution kernel and the real part of the input; the output features are activated using the complex ReLU activation function. The activation rules are as follows: positive values ​​are taken for the real and imaginary parts of the complex features, and negative values ​​are set to zero. Then, a 2×2 complex pooling layer is used for dimensionality reduction, with a pooling stride of 2 and the pooling method taking the maximum value of the element's modulus. Finally, the first-level composite time-frequency features are obtained. The second complex convolution block is processed: a 5×5 complex convolution kernel is used, with a number of 192 kernels, a convolution stride of 1, and a padding method of SAME. The first-level composite time-frequency features are subjected to complex convolution operations twice. The gradient vanishing problem is reduced by using the complex ELU activation function. The activation rules are as follows: positive features are directly retained, and the power of the feature value e minus 1 is taken as the processed feature value. Then, a 2×2 complex pooling layer is used for further dimensionality reduction, with a pooling stride of 2 and the pooling method taking the maximum value of the element's modulus. Finally, a high-level composite time-frequency feature that integrates both amplitude and phase information is obtained. Step 203: Based on the noise type in the environmental label, call the cross-regional noise prior feature library. This feature library contains complex time-frequency feature templates for typical agricultural environmental noises, specifically wind noise, farm machinery noise, stream noise, human voice, rain noise, and thunder noise. Each template is generated from several groups of pure noise at different intensities through the processing flow of step 201. The similarity between the advanced composite time-frequency feature and the corresponding noise template is calculated using the cosine similarity algorithm. The calculation process is as follows: first, calculate the sum of the products of corresponding elements of the advanced composite time-frequency feature vector and the noise template feature vector; then, calculate the square root of the sum of squares of each element of the two vectors; finally, divide the sum of the products by the product of the two square roots. The similarity value is obtained by multiplying the components. The similarity threshold is dynamically set according to the noise intensity level: the threshold for low-intensity noise is set to 65%, the threshold for medium-intensity noise is set to 70%, and the threshold for high-intensity noise is set to 75%. Frequency points with similarity exceeding the corresponding similarity threshold are marked as noise frequency points, and a binary time-frequency mask with the same dimension as the complex time-frequency matrix is ​​generated. The positions with a value of 0 in the mask correspond to the noise frequency points that are blocked, and the positions with a value of 1 correspond to the target soundprint frequency points that are retained. The dynamically generated binary time-frequency mask is multiplied element-wise with the normalized complex time-frequency matrix to finally obtain a pure complex time-frequency matrix containing only the sound of pests biting. Step 204: Perform Inverse Short-Time Fourier Transform (ISTFT) on the pure complex time-frequency matrix. The parameters of ISTFT are strictly consistent with those of STFT in Step 201. The processing procedure is as follows: Perform Inverse Fourier Transform on each frame of data of the pure complex time-frequency matrix, and then use Hanning window processing. By frame superposition and synthesis, the pure features in the time-frequency domain are reconstructed into a pure pest audio signal in the time domain. The reconstructed signal is smoothed using a moving average filtering algorithm, with the filtering window size set to 8 sampling points. The smoothing rule is as follows: For the 7th and subsequent sampling points in the signal, the average of the current point and the previous 7 sampling points is taken as the smoothed signal value; for the first 6 sampling points, the original signal value is directly retained. After the smoothing process is completed, a high-quality pure pest audio signal that can be used for subsequent feature extraction is finally obtained.

[0024] When using this method, refer to steps 201 to 204: By employing time-frequency transformation, complex neural networks, and dynamic noise suppression techniques, the processing effect of pest audio signals has been significantly improved. Complex neural networks are used to extract advanced composite time-frequency features that fuse amplitude and phase, enhancing signal representation capabilities. A dynamic mask is generated using a noise prior feature library to accurately separate noise from target acoustic signatures, significantly improving the signal-to-noise ratio. Inverse time-frequency transformation reconstructs pure pest audio signals, providing a high-quality data foundation for subsequent feature extraction. This effectively solves the problems of high noise interference and inaccurate feature extraction in traditional methods, improving the reliability of pest monitoring.

[0025] Step 3: Extract the frequency and time domain features of the pure pest audio signal and splice them together. After outlier removal, dimensionality reduction and normalization optimization, the signal is compressed and associated with environmental labels to form a feature label dataset. Step three includes the following: Step 301: Perform multi-dimensional voiceprint feature extraction on the output high-quality, pure insect audio signal, specifically including the extraction and concatenation of two types of features: frequency domain features and time domain features. Frequency domain feature extraction uses the Mel frequency cepstral coefficient (MFCC) extraction method, with parameters set as follows: 32 Mel filter banks, extraction of 16th-order MFCC coefficients, and simultaneous extraction of 1st-order and 2nd-order difference coefficients. The processing procedure is as follows: Perform a Fourier transform on the pure insect audio signal to obtain the power spectrum, then filter the power spectrum using a Mel filter bank. After taking the logarithm of the filtering result, perform a discrete cosine transform to obtain the MFCC coefficients. The first-order difference coefficients are calculated based on the MFCC coefficients to obtain the difference between adjacent frames. Then, the second-order difference coefficients are calculated based on the first-order difference coefficients to obtain the difference between adjacent frames, and finally, the MFCC feature vector is formed. The temporal feature extraction adopts the sliding window method, with a window length of 256 sampling points and a window step size of 128 sampling points. Four core temporal features are extracted simultaneously: short-time energy, short-time zero-crossing rate, spectral entropy, and peak factor. The average value of each temporal feature is taken according to the number of window slides to form a temporal feature vector. Feature concatenation: The MFCC feature vector and the temporal feature vector are concatenated in dimensional order to obtain the initial voiceprint feature vector. Step 302: Perform three-level optimization on the initial voiceprint feature vector: Outliers are removed using the 3σ criterion. The mean and standard deviation of each dimension's feature value in the initial voiceprint feature vector are calculated. Feature values ​​deviating from the mean by more than three times the standard deviation are identified as outliers and replaced using linear interpolation. The mean of the three normal feature values ​​before and after the outlier is taken. Principal Component Analysis (PCA) is used for feature dimensionality reduction. A covariance matrix is ​​constructed on the feature vector after outlier removal. The eigenvalues ​​and corresponding eigenvectors of the covariance matrix are calculated. The eigenvalues ​​are sorted from largest to smallest, and the feature vectors with a cumulative contribution rate ≥ 95% are selected as principal components. The original feature vectors are projected onto the space formed by the selected principal components to obtain a multidimensional optimized voiceprint feature vector. The min-max normalization method is used for feature normalization, mapping all feature values ​​to the [0, 1] interval to eliminate dimensional differences between different feature dimensions. Step 303: Compress the binary data of the normalized multidimensional optimized voiceprint feature vector; simultaneously add a 32-bit CRC-32 cyclic redundancy check code, the calculation range of which covers the entire compressed data, for subsequent cloud verification of data transmission integrity; using the compressed feature data as the value and the environment label generated in step 104 as the key, establish a one-to-one correspondence to form a feature label dataset; store the feature label dataset in the cache of the edge node, and simultaneously output a transmission ready signal to trigger the subsequent network transmission process; When using this method, refer to steps 301 to 304: By extracting and optimizing the frequency and temporal features of pure insect audio, the accuracy and efficiency of insect identification are significantly improved. Multidimensional voiceprint features are formed by combining MFCC and temporal features, enhancing the representation ability of insect voiceprints. Outliers are removed by the 3σ criterion, and feature quality is optimized by PCA dimensionality reduction and normalization to reduce redundant interference. Data is compressed and associated with environmental labels to reduce transmission and storage overhead. This provides a high signal-to-noise ratio and low-redundancy feature label dataset for cloud analysis, thereby improving the recognition accuracy of subsequent classifiers and the overall system response speed.

[0026] Step 4: Edge nodes upload feature label datasets to the cloud. After verification and decompression, the cloud merges features and selects core features. A pre-trained CNN classifier is used to identify pest types and density levels. The identification results are determined based on confidence and feature matching.

[0027] Step four includes the following: Step 401: The edge device uploads the generated feature label dataset to the cloud. The cloud performs CRC-32 redundancy check on the received dataset, recalculates the 32-bit CRC-32 checksum of the received data, and compares it with the checksum carried in the data. If they do not match, it is determined to be erroneous data and is discarded directly, and a retransmission command is sent to the corresponding edge node via the downlink. If they match, it is determined to be valid data. The valid data is decompressed using a decompression algorithm to restore the multidimensional optimized voiceprint feature vector and the corresponding environmental label. Step 402: Discrete environmental labels, such as region type, noise type, and crop type, are converted into quantized feature vectors using one-hot encoding. Continuous labels are converted into standard normal distribution quantized values ​​with a mean of 0 and a standard deviation of 1 using Z-score normalization. These are then concatenated to form a multidimensional label quantized feature vector. An attention-based feature fusion algorithm is used to weight the multidimensional optimized voiceprint feature vector and the multidimensional label quantized feature vector using pre-trained attention weights. This emphasizes feature dimensions strongly correlated with pest identification, suppresses invalid interference, and outputs a multidimensional fused feature vector. The minimum redundancy maximum correlation (mRMR) algorithm is used to filter the fused feature vector, calculating the correlation coefficient between each feature dimension and the pest species, as well as the redundancy coefficient between features. Feature dimensions with a correlation coefficient ≥ 0.6 and a redundancy coefficient ≤ 0.3 are selected, ultimately yielding a multidimensional standardized core feature vector. Step 403: Call the lightweight CNN classifier preset in the cloud. This classifier is customized for the voiceprint recognition scenario of agricultural pests. The core implementation parameters are as follows: The network structure consists of 2 convolutional layers, 1 max pooling layer, 1 fully connected layer, and 1 output layer; the input dimension is the same as the dimension of the multidimensional standardized core feature vector; the first convolutional layer uses a 3×3 convolutional kernel, a stride of 1, padding set to SAME, and the activation function is ReLU; after convolution, a 2×2 max pooling layer is connected, with a pooling stride of 2 and the pooling method is max pooling; the second convolutional layer uses a 3×3 convolutional kernel, a stride of 1, padding set to SAME, and the activation function is ReLU; the fully connected layer activation function... The ReLU algorithm is used; the output dimension of the output layer is the preset number of pest species, such as 10 dimensions, corresponding to 10 common agricultural pests, and the Softmax activation function is used; this classifier is pre-trained using a cross-regional pest voiceprint dataset, which contains voiceprint data of several pests; the multi-dimensional standardized core feature vector is input into a lightweight CNN classifier, and the recognition confidence of each pest is calculated through forward propagation, ranging from 0 to 100%; the category with the highest confidence is selected as the final recognition result, which includes the specific pest species, such as rice planthopper and rice stem borer, and the corresponding pest density level: low: <5 pests per square meter, medium: 5-15 pests, high: >15 pests; Step 404: The cloud compares the final recognition result with the cloud-based insect voiceprint standard feature library. The insect voiceprint standard feature library contains standard voiceprint feature templates for multiple types of insects. Each template has been calibrated multiple times through experiments, and the feature matching degree is calculated. If the recognition confidence is ≥85% and the feature matching degree is ≥80%, it is determined to be a valid recognition result, and a recognition completion confirmation signal is sent to the edge node. If 50%≤confidence <85% or 60%≤matching degree <80%, it is marked as a suspicious result, and a re-collection instruction is sent to the edge node. If the confidence is <50% or the matching degree is <60%, it is determined to be an invalid result, and a device self-check and re-collection instruction is sent to the edge node. When using this method, please refer to the content of steps 401 to 404: Cloud-based collaborative analysis enables precise and efficient pest identification. CRC-32 verification ensures data transmission integrity. Attention mechanisms and mRMR algorithms are used to filter core features, significantly improving feature quality. A lightweight CNN classifier outputs pest species and density levels. The identification results are verified by both confidence and feature matching, improving classification accuracy. The validity of the results is dynamically determined and instructions are fed back, forming a closed-loop optimization process that enhances the system's real-time performance and reliability, providing efficient decision support for precision agriculture management.

[0028] Please see Figure 3 This invention provides an agricultural management and control system based on an Internet of Things (IoT) cloud platform, comprising: an edge node acquisition module, a signal processing module, a feature engineering module, and a cloud analysis module, wherein: The edge node acquisition module is a distributed deployment of edge nodes equipped with audio acquisition devices. After preprocessing the acquired audio data, it combines dual-mode positioning to obtain environmental labels, forming an initial audio dataset with environmental labels. The signal processing module performs time-frequency transformation and normalization on the initial audio signal, extracts enhancement features through a complex neural network, generates a dynamic mask to suppress noise by combining a noise prior feature library, and reconstructs a pure pest audio signal. The feature engineering module extracts and concatenates the frequency and time domain features of the pure pest audio signal. After outlier removal, dimensionality reduction and normalization optimization, the compressed signal is associated with environmental labels to form a feature label dataset. The cloud-based analysis module allows edge nodes to upload feature label datasets to the cloud. After cloud verification and decompression, features are fused and core features are selected. A pre-trained CNN classifier is used to identify pest species and density levels, and the identification results are determined based on confidence and feature matching.

[0029] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0030] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0031] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0032] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An agricultural management and control method based on an Internet of Things (IoT) cloud platform, characterized in that: include: Distributed deployment of edge nodes with audio acquisition devices; after preprocessing the acquired audio data, environmental labels are obtained by combining dual-mode positioning to form an initial audio dataset with environmental labels. The initial audio signal is subjected to time-frequency transformation and normalization. Enhancement features are extracted through complex neural network. Dynamic mask is generated by combining noise prior feature library to suppress noise and reconstruct a pure pest audio signal. Frequency and time domain features of pure pest audio signals are extracted and spliced ​​together. After outlier removal, dimensionality reduction and normalization optimization, the compressed data is associated with environmental labels to form a feature label dataset. Edge nodes upload feature label datasets to the cloud. The cloud verifies and decompresses the data, then fuses the features and selects core features. A pre-trained CNN classifier is used to identify pest species and density levels. The identification results are determined based on confidence and feature matching.

2. The agricultural management and control method based on an Internet of Things cloud platform according to claim 1, characterized in that: Preprocessing operations are performed on the collected raw audio data: the least squares method is used to fit the linear trend term of the raw audio data, and the fitted trend term is subtracted from the raw audio data to eliminate signal drift; a pre-emphasis algorithm implemented with a first-order FIR filter is used to adjust the signal strength of the insect biting sound frequency band; the audio signal is processed by frame segmentation and windowing to finally obtain the initial audio signal after frame segmentation; the window function used is the Hanning window.

3. The agricultural management and control method based on an Internet of Things cloud platform according to claim 2, characterized in that: Obtain the environmental labels for the region. The environmental labels include basic labels and noise labels. Using the frame number of the initial audio signal after framing as the key and the corresponding environmental label as the value, associate the environmental labels with the initial audio signal after framing to form a labeled initial audio dataset.

4. The agricultural management and control method based on an Internet of Things cloud platform according to claim 1, characterized in that: A short-time Fourier transform is performed on the initial audio signal of each frame to obtain a complex time-frequency matrix containing amplitude and phase information. The complex time-frequency matrix is ​​then subjected to L2 norm amplitude normalization to map the amplitude values ​​of all elements to the interval [0, 1]. Features are extracted from the normalized complex time-frequency matrix using a complex neural network. The complex neural network contains a dual-channel input layer and two levels of complex convolutional blocks to obtain advanced composite time-frequency features that fuse amplitude and phase information.

5. The agricultural management and control method based on an Internet of Things cloud platform according to claim 4, characterized in that: Based on the noise type in the environmental label, the cross-regional noise prior feature library is called. This feature library contains complex time-frequency feature templates of agricultural environmental noise. The similarity between the advanced composite time-frequency features and the corresponding noise templates is calculated using the cosine similarity algorithm. Frequency points with similarity exceeding the corresponding similarity threshold are marked as noise frequency points, and a binary time-frequency mask with dimensions completely consistent with the complex time-frequency matrix is ​​generated. The positions with a value of 0 in the mask correspond to the noise frequency points that are blocked, and the positions with a value of 1 correspond to the target sound pattern frequency points that are retained. The generated binary time-frequency mask is multiplied element-wise with the normalized complex time-frequency matrix to obtain a pure complex time-frequency matrix that contains only the sounds of pests biting. An inverse short-time Fourier transform is performed on the pure complex time-frequency matrix, and a moving average filtering algorithm is used to smooth the reconstructed signal to obtain a high-quality pure pest audio signal.

6. The agricultural management and control method based on an Internet of Things cloud platform according to claim 1, characterized in that: Frequency domain feature extraction involves performing a Fourier transform on the pure insect audio signal to obtain the power spectrum, then filtering the power spectrum using a Mel filter bank, taking the logarithm of the filtering result and performing a discrete cosine transform to obtain MFCC coefficients, calculating the difference between adjacent frames based on the MFCC coefficients to obtain the first-order difference coefficients, and then calculating the difference between adjacent frames based on the first-order difference coefficients to obtain the second-order difference coefficients, finally forming the MFCC feature vector. Temporal feature extraction employs a sliding window method, simultaneously extracting core temporal features: short-time energy, short-time zero-crossing rate, spectral entropy, and peak factor. The average value of each core temporal feature is taken according to the number of window slides to form a temporal feature vector. The MFCC feature vector and the temporal feature vector are concatenated in dimensional order to obtain the initial voiceprint feature vector.

7. The agricultural management and control method based on an Internet of Things cloud platform according to claim 6, characterized in that: The initial voiceprint feature vector is subjected to a three-level optimization process: outliers are removed using the 3σ criterion, feature dimensionality is reduced using principal component analysis to obtain a multidimensional optimized voiceprint feature vector, and feature normalization is performed using the min-max normalization method; the binary data of the normalized multidimensional optimized voiceprint feature vector is then compressed. Using the compressed feature data as values ​​and the environment labels as keys, a one-to-one correspondence is established to form a feature label dataset.

8. The agricultural management and control method based on an Internet of Things cloud platform according to claim 1, characterized in that: Discrete environmental tags are converted into quantized feature vectors using one-hot encoding; continuous tags are converted into standard normal distribution quantized values ​​using Z-score normalization, and then concatenated to form a multidimensional tag quantized feature vector. Based on an attention mechanism, the multidimensional optimized voiceprint feature vector and the multidimensional tag quantized feature vector are weighted by pre-trained attention weights to output a multidimensional fused feature vector. The minimum redundancy maximum correlation algorithm is used to filter the fused feature vector, calculate the correlation coefficient between each feature dimension and the pest species, and the redundancy coefficient between features. Feature dimensions with a correlation coefficient ≥ 0.6 and a redundancy coefficient ≤ 0.3 are selected to finally obtain a multidimensional standardized core feature vector.

9. The agricultural management and control method based on an Internet of Things cloud platform according to claim 8, characterized in that: The system calls a pre-defined lightweight CNN classifier. The network structure consists of two convolutional layers, one max-pooling layer, one fully connected layer, and one output layer. After inputting standardized core feature vectors, it outputs the recognition confidence scores for various pests. The recognition results include the pest species and density level. If the recognition confidence score is ≥85% and the feature matching score is ≥80%, it is considered a valid recognition result. If 50%≤confidence score<85% or 60%≤matching score<80%, it is marked as a suspicious result. If the confidence score is <50% or the matching score is <60%, it is considered an invalid result.

10. An agricultural management and control system based on an Internet of Things (IoT) cloud platform, used to implement the method described in any one of claims 1 to 9, characterized in that: include: The edge node acquisition module is a distributed deployment of edge nodes equipped with audio acquisition devices. After preprocessing the acquired audio data, it combines dual-mode positioning to obtain environmental labels, forming an initial audio dataset with environmental labels. The signal processing module performs time-frequency transformation and normalization on the initial audio signal, extracts enhancement features through a complex neural network, generates a dynamic mask to suppress noise by combining a noise prior feature library, and reconstructs a pure pest audio signal. The feature engineering module extracts and concatenates the frequency and time domain features of the pure pest audio signal. After outlier removal, dimensionality reduction and normalization optimization, the compressed signal is associated with environmental labels to form a feature label dataset. The cloud-based analysis module allows edge nodes to upload feature label datasets to the cloud. After cloud verification and decompression, features are fused and core features are selected. A pre-trained CNN classifier is used to identify pest species and density levels, and the identification results are determined based on confidence and feature matching.

Images