DDGS feed detection method and system based on wireless transmission

Abstract

The invention discloses a DDGS feed detection method and system based on wireless transmission, relates to the technical field of intelligent Internet of Things monitoring devices, and aims to solve the problems of incomplete data acquisition, unstable communication link, large sensor data error and inaccurate abnormality positioning in existing DDGS storage yard monitoring. According to the method, channel attenuation is predicted through RSSI and bit error rate data, and signal transmission is optimized in combination with self-adaptive frequency hopping and an MIMO antenna; a blind source separation algorithm is used for inhibiting a multipath effect, and the data accuracy is improved; inertial navigation and UWB distance measurement are fused to realize high-precision positioning, and a thermodynamic diagram is generated to mark an abnormal region; meanwhile, the edge gateway adopts a BP neural network to correct sensor errors, so that the measurement reliability is improved; according to the invention, the monitoring precision, the stability and the real-time performance are obviously improved, and the safety management level of the DDGS storage yard is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent Internet of Things monitoring devices, and more particularly to a method and system for detecting DDGS feed based on wireless transmission. Background Art

[0002] With the rapid development of information technology and wireless communication, emerging technologies such as the Internet of Things and big data are constantly penetrating all walks of life. The agricultural and feed production fields have also entered the era of intelligent monitoring, promoting the popularization of unattended early warning systems, providing strong technical support for product quality and warehouse safety management, and laying a foundation for industry upgrading.

[0003] At present, the general survey of DDGS feed in China generally adopts the method of stacking storage. Since DDGS feed has strong thermal conductivity and hydrophilicity, the stacked materials are easily invaded by insects and ants during storage, and the large stacking density is likely to cause safety hazards such as mildew and spontaneous combustion. To reduce risks, various production and use manufacturers often take measures such as manual rat repelling and regular warehouse turning. However, there are construction difficulties in laying conventional cable-type temperature and humidity sensors in the storage yard, and non-contact detectors cannot accurately reflect the internal environmental parameters of the stacked body. Therefore, on-site detection mainly relies on manual inspection with a 2-meter temperature detector, which has a large labor intensity and a limited coverage range, resulting in incomplete data collection and untimely response. In addition, due to factors such as metal equipment, uneven distribution of the stacked materials themselves, humidity, and dust in the DDGS storage yard, the signal is extremely vulnerable to severe attenuation during transmission, resulting in unstable communication links. At the same time, when using a distributed collection point layout, the phase interference caused by multipath propagation makes the sampled data of the sensor have obvious deviations and cannot accurately reflect the temperature, humidity, and PH3 concentration. Secondly, GPS is easily blocked inside the storage yard, and the accuracy of temperature, humidity, and PH3 sensors is easily interfered in harsh environments, which all affect the accurate positioning and judgment of abnormal situations. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention discloses a method and system for detecting DDGS feed based on wireless transmission, aiming to solve the problems raised in the background art.

[0005] The present invention adopts the following technical solutions:

[0006] A method for detecting DDGS feed based on wireless transmission, comprising the following steps:

[0007] Step 1, based on the real-time collected RSSI and bit error rate data, the detection node predicts the channel attenuation trend through a long short-term memory network. If the predicted signal quality is lower than the preset threshold, the wireless working frequency band is switched through an adaptive frequency hopping mechanism, and the real-time signal is weighted and synthesized through a MIMO antenna array;

[0008] Step 2: Based on the optimized wireless channel data in Step 1, the detection node extracts the main path signal through the blind source separation algorithm, and calculates the phase offset of the signals on different paths in combination with the detector position topology data to perform phase compensation on the received signal;

[0009] Step 3: When the detected GPS signal strength is lower than the preset threshold, the detection node generates the absolute position of the node by fusing the inertial navigation module and the UWB ranging data of adjacent nodes using the federated filtering algorithm, and outputs the thermal maps of temperature, humidity and PH3 concentration distribution;

[0010] Step 4: The detection node samples data through the PH3 sensor at a preset sampling period, and transmits the collected data and the position information output in Step 3 to the edge gateway through the wireless channel. The edge gateway constructs an error correction model using the BP neural network, and performs real-time compensation on the PH3 sensor data according to the dust concentration and humidity data in the current environment, and outputs the calibrated PH3 data;

[0011] Step 5: If the PH3 concentration of the same detection node exceeds the preset threshold for three consecutive samples and the temperature and humidity conform to the fermentation anomaly characteristics, the concentration diffusion path is traced back in reverse through the gradient descent algorithm, and the abnormal core area is marked in combination with the position data output in Step 3, and the abnormal data is output to the warning control host;

[0012] Step 6: The warning control host uploads the local collected data to the cloud. The cloud aggregates the data of multiple storage yards through the federated learning framework to train the anomaly diagnosis model, and feeds back the updated model to the detection node;

[0013] Step 7: Collect the battery power and current communication load data of each detection node, allocate the transmission power through the game theory power allocation model. If the node power is lower than 20% and it is in a non-abnormal area, switch to the LoRa communication mode;

[0014] Step 8: Periodically detect the health status of the detection node, including battery power, sensor data drift and wireless communication status; if any data exceeds the preset abnormal threshold, mark it as a faulty node and start the topology reconstruction of adjacent nodes, and synchronously report the maintenance instruction.

[0015] Preferably, the working steps of Step 2 include:

[0016] Step S1: Based on the optimized wireless channel data in Step 1, the detection node receives and stores the wireless signal containing the multipath effect to form the observation signal matrix X, X = [x1(t), x2(t),..., x n (t)] T ; where n represents the number of receiving antennas; T represents the number of sampling points;

[0017] Step S2: Preprocess the observed signal matrix X through a blind source separation algorithm, construct a cost function J(W) based on the generalized information maximization method to achieve the statistical independence of each independent component. The formula expression of the cost function is:

[0018] J(W) = G{G(WX white )}-E + β||W|| 2 (1)

[0019] In formula (1), W is the demixing matrix, and X white is the signal matrix after pre-whitening processing, which is used to reduce signal correlation; G is the objective function, which is used to enhance signal sparsity control; β is the regularization coefficient, which is used to prevent numerical overflow; E is the eigenvector matrix of the received observed signal matrix X;

[0020] Step S3: Based on the cost function J(W), iteratively optimize the demixing matrix W through the gradient descent method to obtain the optimal independent component matrix S = WX, and select the main path signal S m ;

[0021] Step S4: Based on the three-dimensional coordinates P ref =(x ref ,y ref ,z ref ) of the detector, construct a multipath channel propagation model, and define the propagation distance d i =(x i ,y i ,z i ) of the signal propagating to the scattering point P i as:

[0022]

[0023] In formula (2), (x i ,y i ,z i ) are the spatial coordinates of the reflection point of path i; (x0, y0, z0) are the coordinates of the signal emission source; k is the path correction factor, which is used to consider the signal diffraction effect; θ i is the incident angle of path i;

[0024] Step S5: Use the propagation model to calculate the phase shift of each path signal The calculation formula is:

[0025]

[0026] In formula (3), γ represents the signal wavelength, and d ref is the reference position coordinate; is the non - linear phase compensation weight, used to correct the interference effect of signals in different paths;

[0027] Step S6: Considering the attenuation characteristics of signals in each path, calculate the comprehensive phase using the generalized optimal weight method The calculation formula is:

[0028]

[0029] In formula (4), w i is the path loss weight; is the path loss exponent correction factor, where ∈ is the loss adjustment coefficient; N is the number of received paths; in the path loss weight w i the path loss value L i The calculation formula is:

[0030]

[0031] In formula (5), n, m are path attenuation exponents; δ is the correction factor;

[0032] Step S7: During the real - time processing, the detection node compensates the received signal S according to the phase offset observed to obtain the compensated signal S comp The formula expression of the compensation operation is:

[0033]

[0034] In formula (5), is the phase compensation factor, is the remote channel correction coefficient, where ρ is the channel non - linear correction parameter.

[0035] A DDGS feed detection system based on wireless transmission, comprising: an early warning control host and a distributed wireless detection and rat - repelling integrated machine;

[0036] The early warning control host includes a host computer system, a first wireless communication module, a channel evaluation and frequency hopping module, and a remote management and control interface; the distributed wireless detection and rat repellent integrated machine includes a second wireless communication module, an inertial navigation module, a temperature and humidity detection module, a PH3 sensing module, a dust sensing module, an ultrasonic rat repellent module, a MIMO antenna array, and a lithium battery module; the output ends of the inertial navigation module, the temperature and humidity detection module, the PH3 sensing module, and the dust sensing module are connected to the host computer system of the early warning control host through the second wireless communication module; the output end of the host computer system is connected to the inertial navigation module, the temperature and humidity detection module, the PH3 sensing module, the dust sensing module, the ultrasonic rat repellent module, the MIMO antenna array, and the lithium battery module of the distributed wireless detection and rat repellent integrated machine through the first wireless communication module.

[0037] Based on the above technical solutions, the positive and beneficial effects of the present invention are as follows:

[0038] 1. Aiming at the problems of strong dependence on manual inspection and limited detection coverage, the present invention realizes the accurate identification and real-time monitoring of abnormal areas through the deployment of distributed wireless detection nodes and the combination of the gradient descent algorithm to track the concentration diffusion path, avoiding the lagging discovery of abnormalities caused by too long manual sampling intervals in traditional methods, and improving the inspection efficiency and accuracy.

[0039] 2. Aiming at the problem of serious signal attenuation inside the DDGS yard, the present invention uses the LSTM network to predict the channel attenuation trend, combines the adaptive frequency hopping mechanism to optimize the wireless communication link, and enhances the signal reception quality through the MIMO antenna array, thereby reducing the interference of the stacking environment on the wireless signal, improving the communication stability, and solving the problems of unstable data transmission and high packet loss rate in the existing methods.

[0040] 3. Aiming at the phase interference problem caused by multipath propagation, the present invention uses the blind source separation algorithm to extract the main path signal and combines the phase compensation technology to correct the phase offset of signals on different paths to improve the accuracy of the received signal, thereby reducing the sensor measurement error caused by signal interference and making the temperature, humidity, and PH3 concentration data more reliable.

[0041] 4. Aiming at the problem of low positioning accuracy affected by GPS signal occlusion, the present invention combines the inertial navigation module with the UWB ranging fusion algorithm and optimizes the node position data through the federated filtering algorithm, enabling the detection node to still be accurately positioned even in an environment without GPS signals, thereby improving the positioning accuracy of abnormal areas and avoiding misjudgment.

[0042] 5. Aiming at the problem that temperature and humidity and PH3 sensors are vulnerable to interference in harsh environments, the present invention adopts a BP neural network error correction model, uses dust concentration and humidity data to compensate the measured values of the sensors in real time, improves the accuracy of the data, avoids monitoring errors caused by environmental changes, and improves the ability to judge abnormal situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the steps of a method for detecting DDGS feed based on wireless transmission according to the present invention;

[0044] Figure 2 It is a framework diagram of a DDGS feed detection system based on wireless transmission according to the present invention;

[0045] Figure 3 It is a principle framework diagram of step S1 of the present invention;

[0046] Figure 4 It is a working method framework diagram of step S5 of the present invention;

[0047] Figure 5 It is a working method framework diagram of step S6 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0049] In the embodiment, as Figure 1 shown: The steps of a method for detecting DDGS feed based on wireless transmission are as follows: Step 1. Based on the real-time collected RSSI and bit error rate data, the detection node predicts the channel attenuation trend through a long short-term memory network. If the predicted signal quality is lower than the preset threshold, the wireless working frequency band is switched through an adaptive frequency hopping mechanism, and the real-time signal is weighted and synthesized through a MIMO antenna array; as Figure 3As shown in the figure, the specific working method is as follows: Based on the real-time collected RSSI and bit error rate data, the detection node first takes the collected channel state data as the feature vector and inputs it into the long short-term memory network. During the time series prediction process, a channel attenuation model is constructed to predict the future signal quality. During the model training and prediction process, the weights are optimized through historical data to obtain the predicted signal attenuation trend curve. If the prediction result shows that the signal quality is lower than the preset threshold, an adaptive frequency hopping mechanism is adopted. Through the energy detection and interference recognition algorithms, the working frequency band is analyzed and selected in real time, and the software-defined radio interface is used for spectrum switching. Based on the multi-channel data received after switching, through the MIMO antenna array, the maximum ratio combining method is used to perform weighted synthesis and spatial filtering on the signals of each channel, and the synthesized signal is output as the optimal channel data.

[0050] Step 2: Based on the optimized wireless channel data in Step 1, the detection node extracts the main path signal through the blind source separation algorithm, and combines the detector position topology data to calculate the phase offset of the signals of different paths to perform phase compensation on the received signal. The specific working steps are as follows:

[0051] Step S1: Based on the optimized wireless channel data in Step 1, the detection node receives and stores the wireless signal containing the multipath effect to form the observation signal matrix X, X = [x1(t), x2(t),..., x n (t)] T ; where n represents the number of receiving antennas; T represents the number of sampling points;

[0052] Step S2: Preprocess the observation signal matrix X through the blind source separation algorithm, construct the cost function J(W) based on the generalized information maximization method to achieve the statistical independence of each independent component. The formula expression of the cost function is:

[0053] J(W) = G{G(WX white )} - E + β||W|| 2 (1)

[0054] In formula (1), W is the demixing matrix, X white is the signal matrix after pre-whitening processing, which is used to reduce signal correlation; G is the objective function, which is used to enhance signal sparsity control; β is the regularization coefficient, which is used to prevent numerical overflow; E is the eigenvector matrix of the received observation signal matrix X;

[0055] Step S3: Based on the cost function J(W), iteratively optimize the demixing matrix W through the gradient descent method to obtain the optimal independent component matrix S = WX, and select the main path signal S m ;

[0056] Step S4: Based on the three-dimensional coordinates P of the detector ref=(x ref ,y ref ,z ref ), construct a multipath channel propagation model, and define the propagation distance d of the signal propagating to the scattering point P i =(x i ,y i ,z i ) as: i is:

[0057]

[0058] In formula (2), (x i ,y i ,z i ) are the spatial coordinates of the reflection point of path i; (x0, y0, z0) are the coordinates of the signal emission source; k is a path correction factor used to consider the signal diffraction effect; θ i is the incident angle of path i;

[0059] Step S5. Calculate the phase offset of each path signal using the propagation model The calculation formula is:

[0060]

[0061] In formula (3), γ represents the signal wavelength, d ref is the reference position coordinate; is the non-linear phase compensation weight used to correct the interference effect of signals on different paths;

[0062] Step S6. Considering the attenuation characteristics of each path signal, calculate the comprehensive phase using the generalized optimal weight method The calculation formula is:

[0063]

[0064] In formula (4), w i is the path loss weight; is the path loss exponent correction factor, where ∈ is the loss adjustment coefficient; N is the number of received paths; among them, the path loss weight w i In, the path loss value L i The calculation formula is:

[0065]

[0066] In formula (5), n, m are the path attenuation exponents; δ is the correction factor;

[0067] Step S7. During the real-time processing, the detection node adjusts the received signal S according to the phase offset observed ​Perform a compensation operation to obtain the compensated signal S comp , and the formula expression of the compensation operation is:

[0068]

[0069] In formula (5), is the phase compensation factor, is the remote channel correction coefficient, where ρ is the channel nonlinear correction parameter.

[0070] Step 3: When the GPS signal strength is monitored to be lower than the preset threshold, the detection node generates the absolute position of the node by fusing the inertial navigation module and the UWB ranging data of adjacent nodes, and outputs the thermal maps of temperature, humidity and PH3 concentration distribution; among them, the detection node first collects the acceleration data a k , angular velocity data w k and the relative distance data z k obtained by the adjacent node through the UWB ranging module; construct the state vector X k =[p k ,v k T , where p k represents the node position, v k represents the speed, and T represents the number of sampling points; and set the state transition matrix F, control input matrix B and process noise covariance Q, and use the prediction formula and prediction error covariance to obtain the prior state estimate of each node and predict each local node; the prediction formula is:

[0071]

[0072] The prediction error covariance is:

[0073]

[0074] In formulas (7) and (8), u k-1 represents the acceleration input obtained by the IMU, X k-1|k-1 represents the posterior state estimate at the previous moment; t k represents the process noise; P k-1|k-1 is the error covariance matrix of X k-1|k-1 ; then, in the local update stage, combine the observation vector z k measured by the inertial navigation module and the UWB ranging module, define the observation matrix H and the measurement noise covariance R, calculate the Kalman gain and update the local posterior state estimate and posterior error covariance matrix, and the calculation formula of the Kalman gain is:

[0075] ​

[0076] In formula (9), H k is the observation matrix, which is used to map the state vector to the measurement space; R k is the covariance; q k is the measurement noise; for the N nodes participating in the federated filtering, through the state estimates and corresponding covariances of each local node, a weighted optimal fusion method is used to obtain the global state estimate The formula expression is:

[0077]

[0078] In formula (10), is the local posterior state estimate of the i-th node; is the corresponding error covariance; N is the total number of received paths; finally, the first three-dimensional components of the global state estimate are extracted as the absolute position of the node, and the temperature, humidity, and PH3 concentration data corresponding to each node are associated with the absolute position of the node through the Kriging interpolation method to generate a thermal map of the DDGS yard environment parameter distribution.

[0079] Step 4: The detection nodes perform data sampling through the PH3 sensors according to a preset sampling period, and transmit the collected data and the position information output in Step 3 to the edge gateway through a wireless channel. The edge gateway constructs an error correction model using a BP neural network, performs real-time compensation on the PH3 sensor data according to the dust concentration and humidity data in the current environment, and outputs the calibrated PH3 data; among them, the error correction model includes an input layer, a feature normalization layer, a linear transformation hidden layer, a non-linear activation layer, a fully connected fusion layer, an output correction layer, and a dynamic update layer; the input layer is used to perform standardization processing on the original PH3 sensor data and the dust concentration and humidity data collected by environmental detection through a digital filtering and normalization algorithm, generate a unified feature vector, and output it to the feature normalization layer; the feature normalization layer is used to adjust the distribution of each dimension of data using a batch normalization mechanism; the linear transformation hidden layer is used to linearly extract the sensor drift data using a weight matrix and a bias vector; the non-linear activation layer is used to perform non-linear mapping using a ReLU activation function to capture the non-linear interference generated by dust and humidity on the PH3 data; the fully connected fusion layer is used to fuse each hidden feature through a weighted superposition mechanism of a multi-layer perceptron and output a comprehensive correction parameter, and the comprehensive correction parameter is used to quantitatively reflect the global mapping of environmental factors on the PH3 measurement error; the output correction layer is used to perform real-time dynamic compensation on the original PH3 data using element-wise multiplication based on the comprehensive correction parameter to obtain the calibrated PH3 concentration data; the dynamic update layer is used to continuously iteratively adjust the weights and biases using a backpropagation method and an Adam optimization strategy when there is a deviation between the compensated data and the expected target.

[0080] Step 5: If the PH3 concentration of the same detection node exceeds the preset threshold for three consecutive samplings and the temperature and humidity conform to the characteristics of abnormal fermentation, the concentration diffusion path is traced back in reverse through the gradient descent algorithm, the abnormal core area is marked in combination with the position data output in Step 3, and the abnormal data is output to the warning control host; among them, as Figure 4 shown, the edge gateway first constructs a concentration distribution function C(A) based on the calibrated PH3 data, where A represents the spatial position vector, and the value of C(A) represents the PH3 concentration at position A; then, the edge gateway performs a derivative process on the concentration distribution function C(A) through the gradient descent algorithm, obtains the concentration change rate to determine the local concentration gradient direction, and starts iterative update with the initial sampling position as the starting point. In each iteration, based on the concentration data change rate, the system determines the next search direction by calculating the environmental gradient, and then judges whether the local concentration peak reaches the convergence condition according to the convergence criterion; if so, obtains the coordinates of the local abnormal area obtained by reverse tracking; if not, continues the iterative update; then, the edge gateway performs data fusion on the coordinates of the local abnormal area and the position data output in Step 3 through the spatial alignment mechanism and the position data correction algorithm; in the data fusion process, the position data of each node is statistically processed by the weighted average method.

[0081] Step 6: The warning control host uploads the local collected data to the cloud, and the cloud aggregates the multi-yard data through the federated learning framework to train the abnormal diagnosis model, and feeds back the updated model to the detection node; the specific working method is: after the warning control host uploads the local data collected by each distributed detection node to the cloud server through the secure transmission protocol, the cloud starts the federated learning framework, and asynchronously aggregates the local model updates from multiple DDGS yards through the secure multi-party computing mechanism; during the aggregation process, each detection node first calculates the gradient of the local loss function locally using the privacy protection mechanism, generates the local model weight update data through the local gradient descent algorithm, and transmits it to the cloud; after receiving multiple local gradients, the cloud fuses the local update data using the weighted average strategy, and filters and denoises the gradient data using the Bayesian inference and differential privacy algorithms; then, the cloud iteratively updates the global abnormal diagnosis model parameters through the stochastic gradient descent method, and feeds back the updated model parameters to each detection node; finally, the detection node matches the global parameters after model update with the local model through the local model fusion mechanism, and adjusts the local abnormal detection strategy.

[0082] Step 7: Collect the battery power and current communication load data of each detection node, allocate the transmission power through the game theory power allocation model. If the node power is lower than 20% and it is not an abnormal area, switch to the LoRa communication mode. Among them, the game theory power allocation model takes each node as a game participant, constructs a transmission quality and energy consumption utility function, and uses distributed iterative solution to obtain the Nash equilibrium to get the optimal transmission power allocation.

[0083] Step 8: Periodically detect the health status of the detection nodes, including battery power, sensor data drift, and wireless communication status. If any data exceeds the preset abnormal threshold, mark it as a faulty node and start the topology reconstruction of adjacent nodes, and synchronously report the maintenance instruction. Further, when it is detected that any data exceeds the preset abnormal threshold, update the node status to a faulty node. Based on the node marking, the early warning control host sends a status query instruction to adjacent nodes through the broadcast protocol, collects the latest position information, communication quality parameters, and network connection status of each node. Then, the early warning control host calculates the best route according to the node connection weight by the shortest path method and triggers dynamic route update in the network. The node connection weight is composed of the remaining energy of the node, signal quality, and link delay.

[0084] As Figure 2 shown, a DDGS feed detection system based on wireless transmission includes: an early warning control host and a distributed wireless detection and rat repellent integrated machine. The distributed wireless detection and rat repellent integrated machine includes n wireless detection and rat repellent integrated machines. The early warning control host includes a host computer system, a first wireless communication module, a channel evaluation and frequency hopping module, and a remote management and control interface. The distributed wireless detection and rat repellent integrated machine includes a second wireless communication module, an inertial navigation module, a temperature and humidity detection module, a PH3 sensing module, a dust sensing module, an ultrasonic rat repellent module, a MIMO antenna array, and a lithium battery module. The output ends of the inertial navigation module, the temperature and humidity detection module, the PH3 sensing module, and the dust sensing module are connected to the host computer system of the early warning control host through the second wireless communication module. The output end of the host computer system is connected to the inertial navigation module, the temperature and humidity detection module, the PH3 sensing module, the dust sensing module, the ultrasonic rat repellent module, the MIMO antenna array, and the lithium battery module of the distributed wireless detection and rat repellent integrated machine through the first wireless communication module.

[0085] In step 1 of the above embodiment, the long short-term memory network, as a special type of recurrent neural network (RNN), its core lies in solving the problems of vanishing gradients and exploding gradients faced by traditional RNNs when dealing with long sequence data. LSTM contains memory cells (cell state) and three gating structures, namely the input gate, the forget gate, and the output gate. In the DDGS feed detection method based on wireless transmission, the detection node inputs the RSSI (Received Signal Strength Indication) and bit error rate data collected in real time as feature vectors into the LSTM. The input gate determines which new information will be added to the memory cell, the forget gate controls which information in the memory cell will be retained or discarded, and the output gate determines which information in the memory cell will be output for prediction. Through learning and training on historical data, the LSTM continuously adjusts the internal weight parameters to construct a model that can accurately reflect the channel attenuation trend. During the prediction process, the memory cell will retain important information related to channel attenuation according to the input feature vectors and the control of the gating mechanism, so as to achieve the prediction of future signal quality and generate a predicted signal attenuation trend curve.

[0086] When the signal quality predicted by the LSTM is lower than the preset threshold, the adaptive frequency hopping mechanism is activated. This mechanism first uses the energy detection algorithm to detect the signal energy within the current frequency band to determine whether there is interference and the intensity of the interference in the frequency band. Energy detection determines the occupancy of the frequency band by analyzing the power spectral density of the received signal. At the same time, the interference recognition algorithm will further classify the detected interference to determine the type of interference (such as narrowband interference, broadband interference, etc.) and the source. Based on the results of energy detection and interference recognition, the system selects a suitable operating frequency band from the pre-set frequency table. Then, the software-defined radio (SDR) interface is used to achieve spectrum switching. Software-defined radio defines and controls various functions of the radio through software, enabling the system to flexibly switch between different frequency bands, so as to avoid the interference frequency band and select a frequency band with better signal quality for communication.

[0087] MIMO (Multiple-Input Multiple-Output) antenna array technology utilizes multiple transmit antennas and multiple receive antennas to improve the performance of communication systems through spatial multiplexing and spatial diversity. In this method, when the detection node switches to a new frequency band, it receives multi-channel data through the MIMO antenna array. The Maximum Ratio Combining (MRC) method is an important technology in MIMO signal processing. MRC assigns weights to each received channel signal according to its Signal-to-Noise Ratio (SNR). The higher the SNR of a channel, the greater the assigned weight. By weighted combining the signals of each channel according to the weights and performing spatial filtering processing, the SNR of the combined signal can be effectively improved, enhancing the signal strength and quality, thereby obtaining the optimal channel data output.

[0088] In the actual implementation of Step 1, the main hardware includes a signal acquisition module, a processing module, a storage module, a wireless communication module, and a MIMO antenna array. The signal acquisition module is used to collect RSSI and bit error rate data in real time and can use an integrated wireless signal sensor. The processing module is the core of the detection node and is responsible for running the Long Short-Term Memory network, the adaptive frequency hopping algorithm, and the MIMO signal processing algorithm. A high-performance microprocessor or embedded processor can be selected. The storage module is used to store historical acquisition data, trained LSTM model parameters, and preset frequency tables and other information, and flash memory or other non-volatile memories can be used. The wireless communication module is responsible for realizing the transceiver of wireless signals, and a wireless communication chip supporting multiple frequency bands can be used. The MIMO antenna array consists of multiple antennas and is used to receive multi-channel signals. Among them, the signal acquisition module is connected to the wireless communication module and transmits the collected RSSI and bit error rate data to the wireless communication module. The wireless communication module transmits the data to the processing module for processing. The processing module is connected to the storage module, reads historical data and model parameters for calculation, and stores the processing results back to the storage module. The processing module also controls the spectrum switching of the wireless communication module and the signal processing operations of the MIMO antenna array. The MIMO antenna array is connected to the wireless communication module and transmits the received multi-channel signals to the wireless communication module for subsequent processing.

[0089] In actual work, the detection node first collects RSSI and bit error rate data in real time through the signal acquisition module. The collected data is transmitted to the processing module, which takes it as a feature vector and inputs it into a trained long short-term memory network to predict the channel attenuation trend. If the prediction result shows that the signal quality is lower than the preset threshold, the processing module will activate the adaptive frequency hopping mechanism, select a suitable operating frequency band through the energy detection and interference identification algorithms, and control the wireless communication module to perform spectrum switching using the software-defined radio interface. After switching the frequency band, the MIMO antenna array receives multi-channel signals, and the processing module uses the maximum ratio combining method to perform weighted synthesis and spatial filtering on the signals of each channel to obtain the optimal channel data for subsequent DDGS feed detection data transmission.

[0090] Compared with the prior art, through the accurate prediction of the channel attenuation trend by the long short-term memory network, the risk of signal quality degradation can be detected in advance, and adaptive frequency hopping measures can be taken in a timely manner, effectively avoiding the problems of interruption and instability during signal transmission and ensuring the reliable transmission of DDGS feed detection data. In addition, the adaptive frequency hopping mechanism can detect and identify interference in real time and quickly switch to a frequency band with no interference or less interference, greatly improving the anti-interference ability of the system. At the same time, the maximum ratio combining method of the MIMO antenna array further enhances the resistance to interference and improves the quality and reliability of the signal. Secondly, by using the deep learning technology of the long short-term memory network, the detection node can automatically learn and predict the channel state, realizing intelligent signal processing and optimization. Compared with traditional manual intervention or simple rule judgment, it can more accurately adapt to the complex and changeable DDGS feed yard environment, improving the adaptability and intelligence level of the system.

[0091] In step 2 of the above embodiment, the blind source separation algorithm constructs a cost function (formula 1) through the generalized information maximization method. Its core principle is to solve the problem of multipath signal aliasing by maximizing the statistical independence of signal components. The demixing matrix W in formula 1 is optimized by the pre-whitened signal matrix X. The pre-whitening process aims to eliminate the second-order correlation between signals (diagonalize the covariance matrix), thereby reducing the complexity of subsequent separation. The objective function G usually uses a non-linear function (such as the hyperbolic tangent function) to approximate the probability density distribution of independent components by enhancing signal sparsity. The regularization coefficient β is used to prevent numerical instability problems caused by matrix singular values during the gradient descent process. The introduction of the eigenvector matrix E enables the algorithm to adaptively extract the main components in the observed signals, thereby improving the separation efficiency.

[0092] In step S3, the gradient descent method iteratively updates the unmixing matrix W to minimize the cost function (Equation 1), and finally obtains the independent component matrix S = WX. The selection of the main path signal is based on the signal energy or signal-to-noise ratio (SNR) threshold. By analyzing the time-domain characteristics (such as envelope fluctuation) or frequency-domain characteristics (such as power spectral density) of each component, the direct path signal (main path) and non-direct path signals (multipath interference) are determined. This process is achieved through matrix eigenvalue decomposition, and the main path signal corresponds to the component with the largest eigenvalue.

[0093] In steps S4 to S6, the construction of the multipath channel propagation model depends on the three-dimensional coordinates P ref =(x ref , y ref , z ref ) and the reflection point coordinates (x i , y i , z i ). The actual propagation path of the signal in space is calculated through the path propagation distance formula (Equation 2). The path correction factor \(\beta\) takes into account the signal diffraction effect (such as Fresnel zone diffraction), and corrects the path length through the Geometric Theory of Diffraction (GTD). The phase shift calculation formula (Equation 3) combines the signal wavelength \(\gamma\) and the nonlinear weight to quantify the phase difference caused by the propagation distance difference of different path signals. The generalized optimal weight method (Equation 4) further introduces the path loss weight w i and the correction factor \(\delta\), dynamically adjusts the contribution weights of each path signal through the path loss exponent, and finally realizes the comprehensive optimization of phase compensation.

[0094] In step S7, the received signal is corrected in real time through the phase compensation factor (Equation 6), where is jointly calculated by Equation 3 and Equation 4. The channel nonlinear correction parameter \(\rho\) is used to compensate for the phase distortion caused by the Doppler effect or the time-varying characteristics of the channel during signal transmission, ensuring that the compensated signal S comp can accurately reflect the characteristics of the original signal.

[0095] When actually implementing Step 2, based on Step 1, the hardware of the detection node also needs to have a storage module for storing the received wireless signals to form an observation signal matrix, and a dedicated signal processing chip or module to run algorithms such as blind source separation algorithms, multi-path channel propagation model calculations, phase offset calculations, comprehensive phase calculations, and phase compensation operations. In addition, a high-precision three-dimensional coordinate measurement device (such as a GPS module combined with an inertial measurement unit IMU) is required to obtain the accurate three-dimensional coordinate information of the detector. Among them, the signal processing chip or module is connected to the wireless communication module to receive the optimized wireless channel data after Step 1. The storage module is connected to the signal processing chip or module for storing the observation signal matrix and other intermediate calculation results. The three-dimensional coordinate measurement device is connected to the signal processing chip or module to provide the three-dimensional coordinate information of the detector in real time. The signal processing chip or module finally transmits the phase-compensated signal back to the wireless communication module for subsequent data transmission.

[0096] In implementation, the specific work process is as follows: The sensor network collects the temperature, humidity, and pressure data in the feed bin and transmits it to the detection node through ZigBee; then, the detection node receives the wireless signal through the MIMO antenna and stores it as the observation signal matrix X; the FPGA calls the pre-whitening module to perform decorrelation processing on X, and the ARM processor executes the gradient descent algorithm to optimize the demixing matrix W to extract the main path signal; then, the UWB positioning module updates the detector coordinates in real time, calculates the phase offset in combination with the path propagation model (Formula 2-5), and generates a compensation factor; the compensated signal S comp is uploaded to the central control platform through LoRaWAN, and the platform dynamically adjusts the working mode of the detection node according to the signal quality.

[0097] In implementation, the detection node completes self-organizing networking through UWB positioning, establishes a three-dimensional coordinate mapping relationship; collects wireless signals once every 5 ms, performs blind source separation and phase compensation, and updates the compensation parameters; when the path loss exponent exceeds the threshold, triggers the path re-selection algorithm and switches to the standby antenna channel.

[0098] To verify the effectiveness of the blind source separation and phase compensation algorithms adopted in Step 2 of the present invention, comparative experiments were carried out under the same experimental conditions. Experimental group A adopted the technology of the present invention, that is, using the generalized information maximization criterion to construct the blind source separation cost function, then optimizing the demixing matrix by the gradient descent method, extracting the main path signal, constructing a multi-path model in combination with the three-dimensional coordinates of the detector, and performing phase weighting compensation through maximum ratio combining; while experimental group B adopted statistical separation technology, only using mean filtering to process the multi-path effect and not performing phase compensation. Both groups carried out five independent experiments in a fixed test environment, and the measured parameters included signal-to-noise ratio (SNR, dB), bit error rate (BER, %), phase deviation (°), and data acquisition accuracy (%). Experimental record table 1:

[0099] Experimental Record Table of Blind Source Separation and Phase Compensation Algorithm

[0100]

[0101] During the experiment, the detection node first collects multipath signals and forms an observation matrix through a preprocessing module, realizes channel state prediction through an LSTM network, and decomposes the data by a blind source separation module. The blind source separation algorithm adopted by Group A is based on the generalized information maximization criterion, obtains the optimal demixing matrix through iterative optimization, and then combines three-dimensional position data to construct a multipath propagation model to achieve precise phase compensation. Group B only uses traditional mean filtering for processing and does not effectively compensate the multipath phase. The data of each group are all processed by a high-speed DSP to realize real-time signal processing and weighted synthesis, and finally the compensated signal is output for measurement and statistics. During the experiment, each parameter is automatically recorded by the built-in sensor and signal acquisition module, and the data is uploaded to the central monitoring system for subsequent analysis.

[0102] The experimental results show that Group A is superior to Group B in terms of indicators such as SNR, bit error rate, phase deviation, and data acquisition accuracy, indicating that the blind source separation and phase compensation algorithm adopted in Step 2 of the present invention can effectively extract the main path signal and eliminate multipath interference, significantly improving the quality of the finally output signal.

[0103] In Step 3 of the above embodiment, the inertial navigation module is a system that calculates its position, velocity, and attitude by measuring the acceleration and angular velocity of the carrier based on Newton's mechanics principle. It mainly consists of an accelerometer and a gyroscope. The accelerometer is used to measure the acceleration of the carrier in three orthogonal axes. The velocity can be obtained by integrating over time, and the position information can be obtained by integrating the velocity again. The gyroscope is used to measure the angular velocity of the carrier around three orthogonal axes, and the attitude angle of the carrier can be obtained by integrating the angular velocity. In the detection of DDGS feed, the inertial navigation module is installed on the detection node to measure the motion state of the detection node in real time, providing basic data for determining the node position. However, due to the problem that the error of the inertial navigation system accumulates over time, when used alone, the position accuracy will decrease after a long time.

[0104] UWB is a carrierless communication technology that transmits data using non-sinusoidal narrow pulses in the nanosecond to microsecond range. In terms of ranging, UWB calculates the distance by measuring the time of flight (ToF) of the signal between two nodes. Specifically, the transmitting node sends a UWB signal, and the receiving node records the time when the signal arrives. Based on the signal propagation speed (the speed of light), the distance between the two nodes can be calculated. UWB ranging has the advantages of high precision, strong anti-multipath ability, and low power consumption. In DDGS feed detection, UWB is used for ranging between adjacent detection nodes to provide accurate information for determining the relative positions of the nodes.

[0105] Federated filtering is a distributed filtering algorithm that fuses the results of multiple sub-filters to obtain a more accurate and reliable estimate. In this step, the federated filtering algorithm fuses the inertial navigation module and UWB ranging data. The inertial navigation module provides the autonomous navigation information of the detection node, while the UWB ranging data provides the relative position information between the nodes. The federated filtering algorithm makes full use of their advantages by fusing these two types of data, suppresses the error accumulation of the inertial navigation system, and improves the accuracy of position estimation at the same time. Specifically, the federated filtering algorithm divides the entire filtering process into multiple sub-filters, each of which processes a part of the data (such as inertial navigation data or UWB ranging data), and then fuses the results of each sub-filter through the information distribution principle to obtain the final absolute position estimate of the node.

[0106] After obtaining the absolute position of the detection node, combined with the temperature, humidity, and PH3 concentration data collected by the detection node, using geographic information system (GIS) technology or related data visualization algorithms, these data are mapped onto the spatial position to generate a heat map of the temperature, humidity, and PH3 concentration distribution. The heat map shows the high and low levels of temperature, humidity, and PH3 concentration at different positions through the depth of color, intuitively demonstrating the distribution of environmental parameters in the DDGS feed storage yard.

[0107] In actual implementation, the hardware of the detection node includes an inertial navigation module (accelerometer and gyroscope), a UWB ranging module, a data processing unit (such as a microprocessor or an embedded system), a temperature and humidity sensor, a PH3 concentration sensor, and a communication module (such as a wireless transceiver). In addition, a display device (such as a touch screen or a remote monitoring terminal) is required to display the generated heat maps of temperature and humidity and PH3 concentration distribution. 2. Hardware connection: The inertial navigation module, the UWB ranging module, the temperature and humidity sensor, and the PH3 concentration sensor are respectively connected to the data processing unit to transmit the collected data to the data processing unit for processing. The communication module is also connected to the data processing unit for receiving and sending data to achieve communication between detection nodes and communication with the remote monitoring terminal. The display device is connected to the data processing unit by wire or wirelessly to receive and display the generated heat map.

[0108] In implementation, the specific working method is as follows: The detection node first monitors the GPS signal strength in real time. When the GPS signal strength is lower than the preset threshold, the inertial navigation module starts to work, measures the acceleration and angular velocity of the detection node in real time, and transmits the data to the data processing unit. At the same time, the adjacent detection nodes perform ranging through the UWB ranging module and also transmit the ranging data to the data processing unit. The data processing unit runs the federated filtering algorithm to fuse the inertial navigation data and the UWB ranging data to calculate the absolute position of the detection node. While calculating the position, the temperature and humidity sensor and the PH3 concentration sensor collect the temperature, humidity, and PH3 concentration data in the environment in real time and transmit them to the data processing unit. The data processing unit integrates the position data and the environmental parameter data, generates heat maps of temperature and humidity and PH3 concentration distribution using the corresponding algorithm, and sends the heat maps to the display device through the communication module for display.

[0109] Based on the above hardware, the same hardware platform is selected to conduct experiments on the algorithm. In experimental group A, the federated filtering algorithm based on IMU and UWB data in step 3 of the present invention is used, and in experimental group B, a traditional Kalman filter is used for single-node positioning. Five experiments are carried out under the same environmental conditions for the two groups. The comparison parameters include: positioning error (m), positioning delay (ms), data consistency (%), channel stability (dB), and robustness index. During the experiment, after the detection nodes collect their respective data, they calculate the absolute position of the nodes through their respective algorithms, and the data is automatically transmitted to the cloud for statistics. The data records of each group are shown in Table 2:

[0110] Table 2 Experimental record table for step 3 application

[0111]

[0112] The experimental results show that Group A using the federated filtering algorithm is significantly superior to the traditional Kalman filtering method (Group B) in terms of positioning accuracy, response delay, data consistency, channel stability, and robustness, verifying the positive and beneficial effects of Step 3 of the present invention in multi-sensor data fusion and high-precision positioning.

[0113] Compared with the prior art, in the DDGS feed detection method based on wireless transmission, when the GPS signal is poor, the prior art may not be able to accurately determine the position of the detection node, resulting in a mismatch between the environmental parameter data and the actual position. In this step, the inertial navigation module and UWB ranging data are fused through the federated filtering algorithm, making full use of the autonomy of inertial navigation and the high precision of UWB ranging, effectively suppressing the error accumulation of the inertial navigation system, improving the positioning accuracy, and enabling the temperature, humidity, and PH3 concentration data to accurately correspond to the corresponding positions. Secondly, when the GPS signal is insufficient, this method can rely on the inertial navigation module and UWB ranging technology for positioning, without being restricted by the GPS signal, enhancing the adaptability and reliability of the system in complex environments. In addition, the generated heat maps of temperature, humidity, and PH3 concentration distribution can visually display the distribution of environmental parameters in the DDGS feed yard. Compared with presenting data in the form of tables or text in the traditional way, it is easier for operators to understand and analyze. Through the heat map, operators can quickly discover areas with abnormal temperature and excessive PH3 concentration and take timely measures to ensure the quality and safety of DDGS feed.

[0114] In Step 4 of the above embodiment, the BP neural network is a multi-layer feedforward network trained by the error backpropagation algorithm, consisting of an input layer, a hidden layer, and an output layer. In this step, the BP neural network is used to construct an error correction model. Its core principle is to transmit the input data to the output layer through forward propagation, calculate the error between the output result and the expected target, and then transmit the error signal layer by layer from the output layer back to the input layer through backward propagation to adjust the connection weights and biases between layers to minimize the error. During the training process, the network continuously learns the mapping relationship between the input data (such as the original PH3 sensor data, dust concentration, and humidity data) and the output data (accurate PH3 concentration data), so as to be able to accurately predict and correct new input data.

[0115] Digital filtering is used to process the original PH3 sensor data, remove noise and interference signals, and improve the quality and reliability of the data. Common digital filtering methods include mean filtering, median filtering, Gaussian filtering, etc. The normalization algorithm is to convert data with different ranges and magnitudes into a unified scale for subsequent processing and analysis. In this step, the original PH3 sensor data, dust concentration, and humidity data are converted into a unified feature vector through the normalization algorithm, making the data comparable and consistent, which is beneficial to the learning and training of the BP neural network.

[0116] Batch normalization is a technique that normalizes the input data of each layer during the training process of a neural network. By normalizing the data of each mini-batch, it makes the mean of the data 0 and the variance 1, thereby accelerating the training convergence speed of the network and reducing the problems of gradient vanishing and gradient explosion. In the feature normalization layer, the batch normalization mechanism adjusts the distribution of each dimension of data, making it easier for the network to learn the features of the data and improving the generalization ability of the model.

[0117] ReLU is a commonly used non-linear activation function. In the non-linear activation layer, the ReLU function performs a non-linear mapping on the data output by the linear transformation hidden layer, setting the negative part to 0 and keeping the positive part unchanged. This can introduce non-linear factors, enabling the neural network to learn more complex functional relationships, capture the non-linear interference of dust and humidity on PH3 data, and improve the expressive ability of the model.

[0118] A multi-layer perceptron is a feedforward neural network composed of multiple neuron layers. In the fully connected fusion layer, the weighted superposition mechanism of the multi-layer perceptron is used to fuse each hidden feature. Each neuron receives the input from the previous layer and calculates the output through weighted summation. The weights determine the importance of each input feature. Through the hierarchical processing of the multi-layer perceptron, different levels of hidden features can be fused to output comprehensive correction parameters, quantitatively reflecting the global mapping of environmental factors on the PH3 measurement error. 6. Back Propagation method and Adam optimization strategy: The Back Propagation method is the core algorithm for training a BP neural network, used to calculate the gradients of the error with respect to the weights and biases. In the dynamic update layer, when there is a deviation between the compensated data and the expected target, the error signal is transmitted back from the output layer to the input layer through the Back Propagation method to calculate the gradients of each weight and bias. The Adam optimization strategy is an optimization algorithm with an adaptive learning rate. It combines the advantages of the momentum method and the adaptive gradient algorithm, can dynamically adjust the learning rate according to the historical gradient information of each parameter, accelerate the convergence speed of the model, and improve the training efficiency. By combining the Back Propagation method and the Adam optimization strategy, the weights and biases are continuously iteratively adjusted, enabling the error correction model to be continuously optimized and improving the correction accuracy for PH3 data.

[0119] In the actual implementation of Step 4, the application hardware mainly includes detection nodes (equipped with PH3 sensors, temperature and humidity sensors, dust concentration sensors, microprocessors, and wireless communication modules) and edge gateways (processors with strong computing capabilities, storage devices, and wireless communication modules). The detection nodes are used to collect PH3 concentration, temperature and humidity, and dust concentration data and transmit the data to the edge gateways. The edge gateways are responsible for running the error correction model constructed by the BP neural network to process and correct the data. Among them, the PH3 sensor, temperature and humidity sensor, and dust concentration sensor in the detection node are respectively connected to the microprocessor, and the collected data is transmitted to the microprocessor for preliminary processing. The microprocessor is connected to the wireless communication module of the edge gateway through the wireless communication module and sends the processed data to the edge gateway. The processor of the edge gateway is connected to the storage device, and the storage device is used to store the BP neural network model, training data, and other relevant information.

[0120] In the actual implementation of Step 4, the detection nodes perform data sampling through the PH3 sensor, temperature and humidity sensor, and dust concentration sensor at a preset sampling period. The microprocessor performs preliminary processing on the collected raw data, including operations such as digital filtering, and then sends the data to the edge gateway through the wireless communication module. After receiving the data, the edge gateway first performs standardization processing on the data through the digital filtering and normalization algorithms of the input layer to generate a unified feature vector and outputs it to the feature normalization layer. The feature normalization layer adjusts the data distribution using the batch normalization mechanism, and then the data enters the linear transformation hidden layer for linear extraction. The non-linear activation layer uses the ReLU activation function to capture non-linear interference, and the fully connected fusion layer fuses each hidden feature through the weighted superposition mechanism of the multi-layer perceptron and outputs the comprehensive correction parameters. The output correction layer performs real-time dynamic compensation on the original PH3 data based on the comprehensive correction parameters to obtain the calibrated PH3 concentration data. If there is a deviation between the compensated data and the expected target, the dynamic update layer uses the backpropagation method and the Adam optimization strategy to continuously iteratively adjust the weights and biases to optimize the error correction model. Finally, the edge gateway transmits the calibrated PH3 data to the subsequent processing device or monitoring terminal through the wireless communication module.

[0121] To verify the positive effect of the BP neural network error correction model on the real-time compensation of PH3 sensor data, an experimental verification was carried out. The experiment was divided into two groups: Group A used the BP neural network error correction model of the present invention for data compensation; Group B used the traditional linear calibration method, which only corrected the PH3 data with fixed coefficients and could not dynamically adapt to environmental changes. The two groups of experiments were carried out in the same DDGS yard environment, using the same hardware platform, and five sampling experiments were carried out continuously. The experiment mainly compared key indicators such as positioning error, calibration deviation, signal-to-noise ratio (SNR) improvement, response delay, and data consistency. During the acquisition process, the detection nodes obtained the original concentration data through the PH3 sensors according to the preset sampling period, and at the same time the environmental module collected dust and humidity data; these data were preprocessed by the edge gateway and input into their respective calibration models. During the experiment, each index was automatically recorded by the built-in sensors and summarized and statistically analyzed by the cloud server after wireless transmission to form an experimental record table, as shown in Table 3:

[0122] Table 3 Experimental Record Table for the Application of the BP Neural Network Error Correction Model

[0123]

[0124]

[0125] The experimental results showed that Group A using the BP neural network error correction model was significantly better than Group B using the traditional linear calibration in terms of measurement error, calibration deviation, SNR improvement, response delay, and data consistency. The average measurement error of Group A was reduced by about 45%, the calibration deviation was reduced by about 55%, the SNR improvement was increased by about 2.4 dB, the response delay was reduced by about 30 ms, and the data consistency reached over 98%, while the data consistency of Group B was only about 93%. These data fully demonstrated that the BP neural network dynamic compensation model could capture the non-linear interference of environmental factors on PH3 sensor data in real time and accurately compensate for it, thus significantly improving the accuracy and real-time performance of PH3 data in the wireless transmission DDGS feed detection system.

[0126] In step 5 of the above embodiments, the edge gateway constructs a concentration distribution function based on the calibrated PH3 data. Essentially, it establishes a mathematical mapping relationship between spatial positions and the corresponding PH3 concentrations. Through this functional representation, the spatial distribution of PH3 concentrations in the DDGS feedstock yard can be described at the mathematical level. In practical applications, the spatial position vector can be three-dimensional coordinates. By using the PH3 concentration data collected by sensors at different positions, this function is fitted, providing a basis for subsequent analysis of concentration changes. The gradient direction obtained by differentiating the concentration distribution function represents the direction in which the function value changes most rapidly. By iteratively updating along the opposite direction of the gradient, the local minimum of the function can be found (when tracking the concentration diffusion path, this is the reverse process of finding the concentration peak). In each iteration, the environmental gradient is determined based on the concentration data change rate to calculate the next search direction, which is a dynamic adjustment based on the concentration change. The concentration change rate reflects the spatial change trend of the concentration. By analyzing this change rate, the search can be more accurately directed towards the source of the concentration. The setting of the convergence criterion is to determine whether the corresponding position of the local concentration peak has been found, that is, whether the target state of the search has been reached. The coordinates of the local abnormal area and the position data output in step 3 may differ due to measurement errors, different measurement benchmarks, etc. The spatial alignment mechanism adjusts the coordinates of the local abnormal area to the same spatial reference system as the position data through certain transformations (such as translation, rotation, scaling, etc.), enabling effective data fusion between the two.

[0127] The position data correction algorithm is mainly used to process the errors in the position data. In actual measurements, whether the position data is obtained through an inertial navigation module, UWB ranging, or other means, there may be certain errors. The position data correction algorithm estimates and corrects these errors by analyzing and processing the position data of multiple nodes, using statistical methods or other mathematical models, improving the accuracy of the position data.

[0128] When actually implementing step 5, this solution relies on the vector edge gateway and the node detection overall monitoring network distributed at key positions in the DDGS yard. The node detection hardware includes PH3 sensors, temperature and humidity sensors arranged in a layout, and a wireless communication module for data synchronization; at the same time, each node is built-in with a GPS module, an inertial navigation module, and a UWB ranging module for providing positioning data. Interconnected by a low-power wireless protocol (such as LoRa or ZigBee), the data passes through high-speed SPI, I 2The C transmission is transmitted to the embedded processor. The edge gateway consists of a multi-core processor, an FPGA, and a DSP, and pre-installs the software system of this operation embodiment, where the data structure and the centralized distribution function include constructing gradient iteration, iterative algorithms, and a spatial data fusion module. In the specific implementation process, the detection node starts periodically. The PH3 center edge gateway first uses the data fusion module to construct the spatial concentration distribution function, then calculates the local gradient through numerical differentiation, and starts the reverse tracking algorithm based on gradient descent. The algorithm starts at the initial position and adjusts the search direction using the gradient information calculated in each iteration to reconcile the concentration change to meet the convergence criterion and obtain the coordinates of the local abnormal area. The efficient edge gateway uses the built-in spatial layout module to encapsulate the coordinates and perform weighted aggregation with the global positioning data from GPS, IMU, and UWB to correct the position of the final abnormal area. All hardware modules are interconnected through standard interfaces (SPI, I 2 C, UART, LVDS), and a real-time operating system (RTOS) is used for task scheduling between modules to ensure the good operation of data-intensive processing, algorithm iteration, and communication transmission.

[0129] Compared with the existing methods that only rely on fixed thresholds or single positioning algorithms, this embodiment realizes the accurate modeling of the PH3 concentration distribution and dynamic area positioning through multi-sensor data fusion and the reverse tracking algorithm based on gradient descent. Traditional methods often cannot accurately capture local concentration peaks, resulting in problems of false positives or false negatives in anomaly judgment; while this solution constructs a partition distribution function using the real-time acquired reconstruction data and obtains accurate concentration gradient information by solving derivatives, realizing the fast convergence positioning of partition hotspots. In addition, through the spatial partitioning and fast convergence positioning algorithm, the local abnormal coordinates are accurately matched with the global data, greatly improving the accuracy of anomaly area marking. The hardware platform uses a sensor experiment to locate the low-latency communication interface to ensure the high real-time performance and stability of data acquisition and transmission, further improving the overall speed response of the system.

[0130] In step 6 of the above embodiment, each detection node preprocesses the collected local data through an embedded encryption module, and calculates the gradient of the local loss function in the data encryption state by using the homomorphic encryption or differential privacy mechanism. This process uses a privacy protection algorithm to inject noise and perturb the original multi-dimensional data such as PH3, temperature and humidity, and dust, ensuring that the calculated gradient can accurately reflect the sensor error without leaking sensitive data. By using digital signatures and the TLS protocol, the integrity and confidentiality of the data uploaded by each node are ensured, making secure multi-party computation (SMC) possible, thus providing trustworthy data input for the federated learning framework. At the local node, after calculating the local gradient based on a preset loss function (such as mean square error MSE), an improved local gradient descent algorithm is used to iteratively update the neural network weights. This algorithm adjusts the adaptive learning rate (for example, using the Adam optimizer) and the momentum factor to ensure that the model can converge quickly in the face of non-linear interference. The local model update process makes full use of distributed computing resources to achieve low-latency updates, and uses a built-in verification mechanism to verify the accuracy of each update, ensuring that the weight update data generated by each detection node has high precision and robustness. After receiving the locally protected local gradients from each detection node, the cloud server uses a distributed weighted average strategy to fuse the local update data. This process combines Bayesian inference to perform prior noise modeling and posterior filtering on each local gradient data, thus effectively reducing random noise interference during the fusion process. At the same time, by introducing the differential privacy mechanism to reprocess the aggregation result, it is ensured that both key features can be retained and the risk of data leakage can be resisted when the global model is updated. This weighted fusion process depends on the uncertainty measurement of the update data of each node, and adopts a weight adjustment strategy to ensure that the fusion result can truly reflect the statistical characteristics of the multi-yard data. After the aggregation is completed in the cloud, the global anomaly diagnosis model parameters are iteratively optimized by using stochastic gradient descent (SGD) or its variants (such as Adam, RMSProp). This update process gradually approaches the optimal solution by setting a dynamic learning rate and a regularization strategy, and using the weighted fusion gradient data. The update algorithm periodically evaluates the performance of the global model, and automatically adjusts the hyperparameters in combination with a callback function, so that the global model can adapt to the distribution characteristics of abnormal data in different yard environments. The update result of the global model is sent to the edge gateway through an encrypted feedback channel to provide the latest anomaly detection strategy for the detection nodes. After the updated global model parameters are fed back to each detection node through a secure transmission protocol, the node uses a local model fusion mechanism to match the global parameters with the pre-deployed local model. In this process, the local node uses a parameter fusion algorithm (such as weighted average or adaptive fusion algorithm) to fine-tune the global model according to the credibility and error index of the local data, so as to implement a personalized anomaly detection strategy.The fusion mechanism utilizes cross-validation and online calibration techniques to ensure the adaptability of model parameters in the local environment, thereby improving the sensitivity and accuracy of anomaly detection. The entire federated learning process relies on the secure multi-party computation (SMC) mechanism. Data transmission between each detection node and the cloud is achieved through secure communication protocols based on TLS, VPN, and MQTT / CoAP, ensuring that no party discloses the original sensitive data during data aggregation, model update, and parameter feedback processes. The security mechanism guarantees the integrity and confidentiality of data during transmission and processing through dynamic key negotiation, digital signatures, and access control policies. This technology realizes the privacy protection of data from all parties in multi-source data collaborative training while ensuring the stable operation of the federated learning framework, providing a solid security support for the system.

[0131] In the actual implementation of step 6, each node first calculates the gradient data under privacy protection locally and uploads the data after protecting it using encryption algorithms (such as homomorphic encryption or differential privacy mechanisms). After receiving the data, the cloud uses a distributed system architecture and efficient parallel computing to perform weighted averaging, Bayesian filtering, and noise reduction processing on the local update data, and then uses SGD to iteratively optimize the global model parameters and feedback the updated model parameters to the edge gateway through a secure interface. The edge gateway then fuses and adjusts the global model with the local model to achieve real-time optimization of the anomaly diagnosis strategy. Each module is interconnected through standard communication protocols (such as MQTT, CoAP, etc.) and a low-latency network to ensure the overall coordinated operation of the system.

[0132] Compared with the existing centralized anomaly detection methods, this embodiment makes full use of the federated learning framework to achieve distributed collaborative training of multi-yard data, significantly reducing the risk of single-point data leakage. By adopting privacy-protected gradient calculation and secure multi-party computation at the local node, the security of data during transmission is ensured, and at the same time, Bayesian inference and differential privacy technologies are used to improve the robustness of data aggregation. The global model update adjusts through SGD and dynamic learning rate, enabling the model to quickly adapt to environmental changes and improving the accuracy of anomaly detection. Each node dynamically adjusts the detection strategy through local model fusion to ensure the adaptability of the model in its respective environment. Experimental data shows that this solution is superior to traditional methods in terms of positioning accuracy, response latency, and anomaly detection accuracy, and can effectively improve the real-time performance and stability of the DDGS yard monitoring system, enhancing the level of security warning and risk management.

[0133] In step 7 of the above embodiments, game theory is a theory that studies the decisions of decision-making agents when their behaviors directly interact with each other and the equilibrium problems of such decisions. In this step, each detection node is regarded as a participant in the game. Constructing the transmission quality and energy consumption utility functions is the core of this model. The transmission quality function mainly considers factors such as signal strength and bit error rate, because these factors directly affect the accuracy and reliability of data transmission; the energy consumption utility function focuses on the battery power consumption to evaluate the energy utilization efficiency of nodes under different transmission powers. Through these two functions, the "benefits" of each node under different transmission powers are comprehensively measured. Distributed iterative solution of the Nash equilibrium means that, without centralized control, each node continuously interacts with other nodes for information and adjusts its strategies to find a stable state, that is, the Nash equilibrium. In the Nash equilibrium state, the strategy of each node is the best response to the strategies of other nodes. At this time, unilaterally changing its own transmission power strategy by any node will not increase its "benefits". In this way, the optimal transmission power allocation scheme is finally obtained, enabling the entire system to minimize energy consumption while ensuring transmission quality.

[0134] LoRa (Long Range) is an ultra-long-range wireless transmission technology based on spread spectrum technology. Its working principle is to use spread spectrum modulation technology to expand the bandwidth of the original signal to a wider frequency band for transmission. This technology features low power consumption and long-distance transmission. In terms of low power consumption, LoRa reduces the energy consumption of devices by reducing the data transmission rate, etc., and is suitable for battery-powered devices. In terms of long-distance transmission, LoRa can achieve a relatively long communication distance at a relatively low power, which gives it an advantage in some scenarios where there are requirements for communication distance and low requirements for data transmission rate. When the battery power of the detection node is lower than 20% and it is in a non-abnormal area, switching to the LoRa communication mode makes use of its low power consumption and long-distance transmission characteristics to extend the working time of the node while ensuring effective data transmission.

[0135] When actually implementing step 7, the applied hardware mainly includes detection nodes (built-in battery power detection module, communication load detection module, power adjustment module, LoRa communication module, conventional wireless communication module, microprocessor) and edge gateways (devices with data processing and communication functions). The battery power detection module is used to monitor the battery power in real time, the communication load detection module is used to detect the current communication load situation, the power adjustment module is used to adjust the transmission power according to the power distribution result, the LoRa communication module and the conventional wireless communication module are used for data transmission, and the microprocessor is responsible for controlling the operation of each module and data processing. Among them, the battery power detection module and the communication load detection module are respectively connected to the microprocessor and transmit the detected data to the microprocessor. The microprocessor is connected to the power adjustment module and controls the power adjustment module to adjust the transmission power according to the power distribution result. The microprocessor is also connected to the LoRa communication module and the conventional wireless communication module at the same time, and selects a suitable communication module for data transmission according to different situations. The detection node communicates with the edge gateway through the wireless communication module, sends the collected data to the edge gateway, and receives the instructions from the edge gateway at the same time.

[0136] During implementation, the battery power detection module of the detection node monitors the battery power in real time, and the communication load detection module detects the current communication load data in real time and transmits this data to the microprocessor. The microprocessor sends the battery power and communication load data to the edge gateway. After receiving the data, the edge gateway uses the game theory power distribution model, takes each detection node as a game participant, constructs a transmission quality and energy consumption utility function, calculates the optimal transmission power of each node through distributed iterative solution of the Nash equilibrium, and sends the power distribution result to the microprocessor of the detection node. The microprocessor controls the power adjustment module to adjust the transmission power according to the received power distribution result. During operation, the microprocessor continuously monitors the battery power. If the node power is lower than 20% and the node is in a non-abnormal area (the area status information can be obtained by interacting with the edge gateway), the microprocessor controls the switch to the LoRa communication mode and uses the LoRa communication module for data transmission; if the above conditions are not met, the conventional wireless communication module continues to be used for data transmission.

[0137] Compared with the traditional fixed power allocation method, this solution can respond in real time to changes in the node battery status and communication load. By constructing a transmission quality and energy consumption utility function and using distributed iterative solution of the Nash equilibrium, the transmission power configuration of each node under optimal energy consumption conditions is obtained, effectively reducing the node energy consumption and extending the device battery life. Secondly, by automatically switching low-battery nodes to the LoRa communication mode, this solution ensures that a low-power, high-stability communication link can still be maintained in the case of insufficient battery power, avoiding monitoring interruptions caused by power shortages. Furthermore, through multi-node collaborative computing and using a distributed game model to achieve optimal power allocation throughout the network, this method significantly reduces internal interference and the bit error rate of the system, improving the overall transmission quality and the accuracy of data collection. Experimental results show that in the actual application in the DDGS storage yard, this solution can reduce the node energy consumption by more than 30%, improve the communication link stability by about 2 - 3 dB, and significantly reduce the overall system response delay, ensuring the real-time and accuracy of environmental monitoring data, and providing reliable and intelligent technical support for the safety management and anomaly warning of the DDGS storage yard.

[0138] In step 8 of the above embodiment, the periodic detection is a strategy based on time-driven, and comprehensively evaluates the health status of the detection node by setting a fixed time interval. For the battery power, its essence is to monitor the energy reserve of the node, because the battery power is directly related to whether the node can continue to work properly. Too low battery power will cause the node to be unable to complete data collection and transmission tasks, just like a machine losing power. The sensor data drift is due to factors such as long-term use of the sensor and environmental changes, resulting in a deviation between the data output by the sensor and the true value. This drift may lead to misjudgment of the DDGS feed status and affect the accuracy of the entire monitoring system. The detection of the wireless communication status mainly focuses on indicators such as signal strength and bit error rate. Wireless communication is the bridge for nodes to interact with other devices. Poor communication status will cause data transmission delay and loss, thus affecting the real-time and reliability of the system. The preset anomaly threshold is a boundary value set according to the design requirements and actual experience of the system. When the detected data exceeds this threshold, it indicates that something is abnormal with the node, which may affect the normal operation of the system.

[0139] When it is detected that a certain data of a node exceeds the preset abnormal threshold, the node is marked as a faulty node, which is a mechanism for quickly identifying and locating problems. The marked faulty node is like a "problem point" with a label, facilitating subsequent processing. The adjacent node topology reconstruction is to ensure the connectivity of the entire monitoring network and the smoothness of data transmission. In a wireless sensor network, nodes cooperate with each other to complete data collection and transmission. When a node fails, its adjacent nodes need to readjust the connection relationship to form a new topology. This is like in a traffic network, when a certain road fails, the surrounding roads need to be re-planned to ensure smooth traffic.

[0140] The early warning control host sends a status query instruction to adjacent nodes through a broadcast protocol. The broadcast protocol is an efficient information dissemination method that can send information to multiple nodes in the network simultaneously. In this scenario, by broadcasting the status query instruction, the latest position information, communication quality parameters, and network connection status of adjacent nodes can be quickly collected. These information are the basis for subsequent topology reconstruction and routing update, just like understanding the situation of surrounding roads before road construction.

[0141] The shortest path method is a classic graph theory algorithm used to find the shortest path between two nodes in a graph. In this method, the node connection weight is composed of the remaining energy of the node, signal quality, and link delay. These factors comprehensively consider the energy consumption, communication reliability, and data transmission speed of the node. By calculating the node connection weight, a "cost" can be assigned to each connection. The shortest path method is to find the path with the minimum cost. When a faulty node appears in the network, the original routing may no longer be optimal. At this time, dynamic routing update needs to be triggered to ensure that data can be transmitted through the best path. This is like when there is traffic congestion, the navigation system will re-plan the route and choose the fastest road.

[0142] In the actual implementation of step 8, the hardware mainly includes detection nodes, early warning control hosts, and communication network devices. The detection node consists of a battery, various sensors (such as temperature and humidity sensors, PH3 sensors, etc.), a wireless communication module, and a microprocessor. The battery provides energy for the node, the sensors are responsible for collecting relevant data, the wireless communication module is used to communicate with other nodes and the early warning control host, and the microprocessor is responsible for controlling the various operations of the node and data processing. The early warning control host is usually a high-performance computer equipped with a large-capacity storage device and a high-speed processor, which is used to receive, process, and analyze the data of the detection nodes and issue control instructions. The communication network devices include routers, switches, etc., which are used to build a wireless sensor network to ensure smooth communication between nodes and between nodes and the early warning control host. Among them, the battery in the detection node is connected to the microprocessor and the wireless communication module to provide power support for them. Various sensors are connected to the microprocessor and transmit the collected data to the microprocessor for processing. The microprocessor is connected to the communication network device through the wireless communication module, sends the processed data to the early warning control host, and at the same time receives the instructions from the early warning control host. The early warning control host establishes a communication connection with each detection node through the communication network device to achieve data interaction and control.

[0143] During implementation, the detection node detects its own battery power, sensor data drift, and wireless communication status according to a preset cycle, and sends the detection results to the early warning control host through the wireless communication module. After receiving the detection data, the early warning control host compares it with the preset abnormal threshold. If it is found that any data exceeds the preset abnormal threshold, the early warning control host marks the node as a faulty node and sends a status query instruction to the adjacent nodes of the faulty node through the broadcast protocol. After receiving the instruction, the adjacent nodes send their latest location information, communication quality parameters, and network connection status to the early warning control host. The early warning control host calculates the node connection weight according to the received information and uses the shortest path method to calculate the optimal route. Finally, the early warning control host triggers a dynamic route update in the network and simultaneously synchronously reports a maintenance instruction to notify relevant personnel to repair or replace the faulty node.

[0144] Compared with traditional fixed routing and centralized monitoring methods, the solution of this embodiment has the following beneficial effects: First, through real-time multi-parameter monitoring, a comprehensive detection of the health status of the detection nodes is achieved, and problems such as insufficient battery power, sensor drift, or communication anomalies can be detected in a timely manner, thereby reducing the risk of system interruption caused by single-node failures; Second, a fault determination mechanism based on decision trees and multi-dimensional statistics is adopted to achieve accurate classification and marking of abnormal states, providing a reliable basis for subsequent dynamic topology reconstruction; Third, by using broadcast queries and distributed data fusion algorithms, the warning control host can quickly collect information of neighboring nodes, determine the optimal route through the shortest path method and the minimum spanning tree algorithm, and achieve network self-healing, thereby significantly improving the robustness and communication continuity of the entire system; Finally, this solution makes full use of low-power, high-performance sensor and processor modules in hardware, and realizes efficient cooperation between modules through standard interfaces and secure communication protocols, ensuring the real-time nature of data transmission and processing. Experimental results show that this solution can increase the accuracy of node fault detection by about 25%, shorten the dynamic routing update response time by 30%, and significantly improve the overall stability of the network, providing strong technical support for the intelligent monitoring and risk warning of DDGS yards.

[0145] Although the specific implementation manners of the present invention have been described above, those skilled in the art should understand that these specific implementation manners are only examples. Without departing from the principles and essence of the present invention, those skilled in the art can make various omissions, substitutions, and changes to the details of the above methods and systems. For example, combining the above method steps so as to perform substantially the same function in a substantially the same manner to achieve substantially the same result falls within the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. A DDGS feed detection method based on wireless transmission, characterized in that: The following steps are involved: Step 1: Based on the RSSI and bit error rate data collected in real time, the detection node predicts the channel attenuation trend through the long short-term memory network. If the predicted signal quality is lower than the preset threshold, the wireless working frequency band is switched through the adaptive frequency hopping mechanism, and the real-time signal is weighted and synthesized through the MIMO antenna array; Step 2: Based on the wireless channel data optimized in step 1, the detection node extracts the main path signal through a blind source separation algorithm, and calculates the phase offset of different path signals in combination with the detector position topology data to perform phase compensation on the received signal; Step 3: When the GPS signal strength is detected to be lower than the preset threshold, the detection node generates the absolute position of the node based on the inertial navigation module and the UWB ranging data of the adjacent nodes, and outputs the temperature, humidity and PH3 concentration distribution heat map; Step 4: The detection node samples data through the PH3 sensor according to a preset sampling period, and transmits the collected data and the location information output in step 3 to the edge gateway through a wireless channel. The edge gateway uses a BP neural network to build an error correction model, compensates the PH3 sensor data in real time according to the dust concentration and humidity data in the current environment, and outputs the calibrated PH3 data; Step 5: If the PH3 concentration of the same detection node exceeds the preset threshold for three consecutive samplings and the temperature and humidity meet the abnormal fermentation characteristics, the concentration diffusion path is reversely traced through the gradient descent algorithm, the abnormal core area is marked in combination with the position data output in step 3, and the abnormal data is output to the early warning control host; Step 6: The early warning control host uploads the local collected data to the cloud, which aggregates multi-yard data through a federated learning framework to train an abnormal diagnosis model, and feeds back the updated model to the detection node; Step 7: Collect the battery power and current communication load data of each detection node, allocate the transmission power through the game theory power allocation model, and switch to LoRa communication mode if the node power is less than 20% and it is not in an abnormal area; Step 8: Periodically detect the health status of the detection node, including battery power, sensor data drift, and wireless communication status; if any data exceeds the preset abnormal threshold, it is marked as a faulty node and the adjacent node topology reconstruction is started, and maintenance instructions are reported synchronously.

2. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: The working method of step 1 is as follows: based on the RSSI and bit error rate data collected in real time, the detection node first inputs the collected channel state data as a feature vector into the long short-term memory network, constructs a channel attenuation model in the time series prediction process, and predicts the future signal quality; During the model training and prediction process, the weights are optimized through historical data to obtain the predicted signal attenuation trend curve; if the prediction result shows that the signal quality is lower than the preset threshold, an adaptive frequency hopping mechanism is adopted to analyze and select the working frequency band in real time through energy detection and interference identification algorithms, and the software-defined radio interface is used for spectrum switching; based on the multi-channel data received after switching, the maximum ratio synthesis method is used through the MIMO antenna array to perform weighted synthesis and spatial filtering on the signals of each channel, and the synthesized signal is output as the optimal channel data.

3. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: The working steps of step 2 include: Step S1: Based on the wireless channel data optimized in step 1, the detection node receives and stores the wireless signal containing the multipath effect to form an observation signal matrix X, where X = [x1(t), x2(t), ..., x n (t)] T ; Where n represents the number of receiving antennas; T represents the number of sampling points; Step S2: preprocess the observed signal matrix X by using a blind source separation algorithm, and construct a cost function J(W) based on the generalized information maximization method to achieve statistical independence of each independent component. The formula of the cost function is: J(W)=G{G(WX white )}-E+β||W|| 2 (1) In formula (1), W is the unmixing matrix, X white is the signal matrix after pre-whitening processing, which is used to reduce signal correlation; G is the objective function, which is used to enhance signal sparsity control; β is the regularization coefficient, which is used to prevent numerical overflow; E is the eigenvector matrix of the received observation signal matrix X; Step S3: Based on the cost function J(W), the unmixing matrix W is iteratively optimized by the gradient descent method to obtain the optimal independent component matrix S=WX, and the main path signal S is selected from it. m ; Step S4: Based on the three-dimensional coordinates P of the detector ref =(x ref ,y ref ,z ref ), construct a multipath channel propagation model, define the signal propagating from path i to scattering point P i =(x i ,y i ,z i ) propagation distance d i for: In formula (2), (x i ,y i ,z i ) is the spatial coordinate of the reflection point on path i; (x0, y0, z0) is the coordinate of the signal transmission source; k is the path correction factor, which is used to consider the signal diffraction effect; θ i is the incident angle of path i; Step S5: Calculate the phase offset of each path signal using the propagation model The calculation formula is: In formula (3), γ represents the signal wavelength, d ref is the reference position coordinate; is the nonlinear phase compensation weight, which is used to correct the interference effect of signals from different paths; Step S6: Considering the attenuation characteristics of each path signal, the generalized optimal weight method is used to calculate the comprehensive phase The calculation formula is: In formula (4), w i is the path loss weight; is the path loss index correction factor, where ∈ is the loss adjustment coefficient; N is the number of received paths; and the path loss weight w i In the example, the path loss value L i The calculation formula is: In formula (5), n, m are path attenuation exponents; δ is the correction factor; Step S7: In the real-time processing process, the detection node detects the phase offset The received signal S observed Perform compensation operation to obtain the compensated signal S comp , the formula expression of the compensation operation is: In formula (5), is the phase compensation factor, is the long-range channel correction coefficient, where ρ is the channel nonlinear correction parameter.

4. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: In step 3, the detection node first collects the acceleration data a output by the inertial navigation module. k , angular velocity data w k And the relative distance data z obtained by the adjacent nodes through the UWB ranging module k ; Construct the state vector X through the federated filtering algorithm k =[p k ,v k ] T , where p k represents the node position, v k represents speed, T represents the number of sampling points; The state transfer matrix F, the control input matrix B and the process noise covariance Q are set, and the prior state estimation of each node is obtained by using the prediction formula and the prediction error covariance, and prediction is performed on each local node; the prediction formula is: X k|k-1 =FX k-1|k-1 +This k-1 +t k (7) The prediction error covariance is: P k|k-1 =FP k-1|k-1 F T +Q (8) In formula (7) and formula (8), u k-1 represents the acceleration input obtained by the IMU, X k-1|k-1 represents the posterior state estimate at the previous moment; t k represents process noise; P k-1|k-1 For X k-1|k-1 Then, in the local update phase, the observation vector z obtained by combining the inertial navigation module and the UWB ranging module is k , define the observation matrix H and the measurement noise covariance R, calculate the Kalman gain and update the local posterior state estimate and the posterior error covariance matrix. The calculation formula of the Kalman gain is: In formula (9), H k is the observation matrix, which is used to map the state vector to the measurement space; R k is the covariance; q k To measure noise; for the N nodes participating in the federated filtering, the global state estimation is obtained by using the weighted optimal fusion method through the state estimation of each local node and the corresponding covariance The formula expression is: In formula (10), is the local a posteriori state estimate of the i-th node; is the corresponding error covariance; N is the total number of received paths; finally, the first three-dimensional components of the global state estimation are extracted as the absolute position of the node, and the temperature, humidity and PH3 concentration data corresponding to each node are associated with the absolute position of the node through the Kriging interpolation method to generate a heat map of the environmental parameters distribution of the DDGS storage yard.

5. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: The error correction model includes an input layer, a feature normalization layer, a linear transformation hidden layer, a nonlinear activation layer, a fully connected fusion layer, an output correction layer and a dynamic update layer; the input layer is used to standardize the raw data of the PH3 sensor and the dust concentration and humidity data collected by environmental detection through digital filtering and normalization algorithms, generate a unified feature vector, and output it to the feature normalization layer; The feature normalization layer is used to adjust the distribution of data in each dimension using a batch normalization mechanism; the linear transformation hidden layer is used to linearly extract sensor drift data using a weight matrix and a bias vector; the nonlinear activation layer is used to perform nonlinear mapping using a ReLU activation function to capture the nonlinear interference of dust and humidity on PH3 data; the fully connected fusion layer is used to fuse each implicit feature through a weighted superposition mechanism of a multi-layer perceptron, and output a comprehensive correction parameter, which is used to quantitatively reflect the global mapping of environmental factors to PH3 measurement errors; the output correction layer is used to perform real-time dynamic compensation of the original PH3 data based on the comprehensive correction parameters using element-by-element multiplication to obtain calibrated PH3 concentration data; the dynamic update layer is used to continuously iteratively adjust weights and biases using a back propagation method and an Adam optimization strategy when there is a deviation between the compensated data and the expected target.

6. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: In step 5, the edge gateway first constructs a concentration distribution function C(A) based on the calibrated PH3 data, where A represents a spatial position vector, and the value of C(A) represents the PH3 concentration at position A; then, the edge gateway uses a gradient descent algorithm to perform a derivative process on the concentration distribution function C(A), obtains the concentration change rate to determine the direction of the local concentration gradient, and starts an iterative update with the initial sampling position as the starting point. In each iteration, based on the concentration data change rate, the system determines the next search direction by calculating the environmental gradient, and then uses the convergence criterion to determine whether the local concentration peak meets the convergence condition; if so, obtain the coordinates of the local abnormal area obtained by reverse tracking; if not, continue to perform iterative updates; then, the edge gateway uses a spatial alignment mechanism and a position data correction algorithm to fuse the coordinates of the local abnormal area with the position data output in step 3; in the data fusion process, the position data of each node is statistically processed by the weighted average method.

7. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: The working method of step 6 is as follows: after the early warning control host uploads the local data collected by each distributed detection node to the cloud server through a secure transmission protocol, the cloud starts the federated learning framework, and asynchronously aggregates the local model updates from multiple DDGS storage yards through a secure multi-party computing mechanism; during the aggregation process, each detection node first calculates the local loss function gradient locally using a privacy protection mechanism, generates local model weight update data through a local gradient descent algorithm, and transmits it to the cloud; after receiving multiple local gradients, the cloud adopts a weighted average strategy to fuse the local update data, and uses Bayesian reasoning and differential privacy algorithms to filter and reduce noise on the gradient data; then, the cloud iteratively updates the global anomaly diagnosis model parameters through a stochastic gradient descent method, and feeds back the updated model parameters to each detection node; finally, the detection node matches the global parameters after the model update with the local model through a local model fusion mechanism, and adjusts the local anomaly detection strategy.

8. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: The game theory power allocation model takes each node as a game participant, constructs a transmission quality and energy consumption utility function, and uses distributed iteration to solve the Nash equilibrium to obtain the optimal transmission power allocation.

9. A DDGS feed detection method based on wireless transmission according to claim 1, characterized in that: When it is detected in step 8 that any data exceeds a preset abnormal threshold, the node status is updated to a faulty node. Based on the node tag, the early warning control host sends a status query instruction to the adjacent nodes through a broadcast protocol to collect the latest location information, communication quality parameters and network connection status of each node. Then, the early warning control host calculates the optimal route according to the node connection weight through the shortest path method, and triggers a dynamic route update in the network; the node connection weight is composed of the node's remaining energy, signal quality and link delay.

10. A DDGS feed detection system based on wireless transmission, characterized in that: A DDGS feed detection method based on wireless transmission as applied to any one of claims 1 to 9, comprising: an early warning control host and a distributed wireless detection and rodent-repelling integrated machine; The early warning control host includes a host computer system, a first wireless communication module, a channel assessment and frequency hopping module, and a remote management and control interface; the distributed wireless detection and mouse-repelling integrated machine includes a second wireless communication module, an inertial navigation module, a temperature and humidity detection module, a PH3 sensor module, a dust sensor module, an ultrasonic mouse-repelling module, a MIMO antenna array, and a lithium battery module; the output ends of the inertial navigation module, the temperature and humidity detection module, the PH3 sensor module, and the dust sensor module are connected to the host computer system of the early warning control host through the second wireless communication module; the output end of the host computer system is connected to the inertial navigation module, the temperature and humidity detection module, the PH3 sensor module, the dust sensor module, the ultrasonic mouse-repelling module, the MIMO antenna array, and the lithium battery module of the distributed wireless detection and mouse-repelling integrated machine through the first wireless communication module.

Images