Warehousing silkworm feed microbial pollution risk early warning analysis method

Abstract

The invention relates to a microbial pollution early warning technology based on multi-mode sensing and an intelligent model in a storage environment. In order to solve the problems of delayed microbial pollution detection and low early warning accuracy in existing silkworm group production, accurate alignment and purification of multi-modal monitoring data are realized through highly collaborative time synchronization and noise filtering, and silkworm body biological behavior and physiological anomaly characteristics are further extracted from the multi-modal monitoring data. Based on a graph attention network and an LSTM-GCN hybrid model, the microecological stress response and microbial growth situation of the silkworm group are deduced, and a dynamic threshold and adaptive weight mechanism is introduced to realize real-time evaluation and early warning of pollution risks. According to the method, the sensitivity and accuracy of pollution early warning in a storage scene are effectively improved, the method has self-closed-loop adjustment and real-time response capabilities, and the silkworm body health guarantee level is improved.

Description

Technical Field

[0001] This invention relates to the field of early warning technology for microbial contamination risks in silkworm feed storage, and in particular to an early warning analysis method for microbial contamination risks in stored silkworm feed. Background Technology

[0002] Currently, the field of microbial contamination risk early warning for silkworm feed storage mainly relies on static threshold methods driven by environmental parameters for contamination risk assessment. Traditional risk early warning systems often employ the following technical model: during storage monitoring, environmental parameters such as temperature, humidity, carbon dioxide, and ammonia, as well as physicochemical indicators such as feed water activity and pH, are collected through IoT sensors. Fixed thresholds or ranges are set based on statistical experience as contamination risk assessment standards. When parameters are detected to exceed the predetermined range, a potential microbial contamination risk is identified, and an early warning is issued. This type of method has the advantages of being simple to implement and having a fast reasoning speed, and is therefore widely used in the storage safety monitoring of bulk commodities such as grains, forage, and mushrooms.

[0003] With the development of technologies such as big data analytics and artificial intelligence, the industry is exploring new approaches such as multi-source parameter weighted fusion, multi-model integrated prediction, knowledge graph rule reasoning, and adaptive thresholding through transfer learning to improve the early warning system's ability to perceive abnormal fine-grained fluctuations. For example, by constructing multi-factor weighted models, a generalized assessment of pollution risk under different storage environments, feed types, and seasonal cycles can be achieved; or, adaptive sliding adjustment of thresholds using statistical mean / standard deviation can be adopted to address false alarms or missed alarms caused by external disturbances to some extent. Although these optimization methods have certain practical effects in specific fields, they generally still follow the static parameter experience or data-driven unidirectional modeling approach, neglecting the crucial "feed-microorganism-silkworm body" ternary biological feedback link in the stored feed system.

[0004] Practical applications show that the dynamic evolution of microbial contamination is influenced by the nonlinear coupling of multiple factors. Using a single environmental or physicochemical parameter as a warning threshold is insufficient for achieving high-precision, low-latency capture of risk events in complex scenarios. Specifically, the following prominent problems and technical bottlenecks exist: Fixed threshold models are difficult to adapt to dynamic changes such as different storage environments, seasonal changes, and batch heterogeneity of feed. When external climate, warehouse loading layout, batch raw material status, etc. change, the original static threshold often leads to a significant increase in risk misjudgment rate, manifested as an increase in false alarm rate and an aggravation of false alarm rate, which seriously restricts the practicality and generalization ability of the early warning system.

[0005] Existing major technological approaches have failed to effectively utilize the unique advantage of silkworms as highly sensitive "biological indicators." Traditional methods focus only on environmental parameters or feed physicochemical characteristics, neglecting the sensitive physiological responses of silkworm populations to microbial contamination, such as decreased feeding activity, group disorder, and surface stress temperature changes. These feedback signals have a direct temporal causal relationship with contamination risk.

[0006] While sliding adjustments based on statistical mean, standard deviation, quantiles, or multi-source data weighting optimization can dynamically correct thresholds to some extent, they are still essentially driven by environmental data and lack criteria that truly integrate biological feedback. When encountering sudden or hidden pollution events, the response is lagging and the reliability of the judgment is poor.

[0007] Most systems lack a closed-loop feedback mechanism encompassing risk assessment, biofeedback, and model calibration, hindering the adaptive online evolution of risk thresholds. The threshold parameters remain unchanged over extended periods, failing to adjust to actual scenario conditions, leading to a significant decline in system stability and long-term accuracy, particularly in high-risk, low-morbidity scenarios where rapid self-optimization is difficult.

[0008] Therefore, the field of early warning for microbial contamination risks in stored silkworm feed urgently needs a technical approach that can integrate biological feedback information from silkworms and achieve adaptive generation of dynamic risk thresholds. By using biological phenotypic signals such as silkworm feeding rhythm, body surface temperature changes, and population activity as core criteria, and combining environmental and physicochemical information, a closed-loop dynamic threshold adjustment method integrating physical, biological, and environmental domains can be established. This will greatly improve the sensitivity and accuracy of identifying contamination risk events, and effectively suppress false alarms and missed alarms under changing scenarios. It will fundamentally solve the technical limitations of existing methods, such as their inability to achieve dynamic environmental adaptability and the lack of biological criteria and feedback self-evolution mechanisms. This need lays the foundation for the innovative approach and core protection content of this invention. Summary of the Invention

[0009] This application provides a method for early warning analysis of microbial contamination risks in stored silkworm feed, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0010] This application provides a method for early warning and analysis of microbial contamination risks in stored silkworm feed, specifically including: S1: Acquire millimeter-wave radar signals and environmental parameters at multiple different locations within the storage unit, and uniformly define the collected results as raw monitoring data; S2: Perform cross-modal time-series alignment and noise filtering on the original monitoring data to eliminate signal interference and obtain preprocessed monitoring data; S3: Based on the preprocessed monitoring data, extract the feeding rhythm decay rate and the low-frequency amplitude of the body surface microtemperature, and define the extraction results as silkworm biological characteristic variables; S4: Based on the biological characteristic variables of the silkworm, input the lightweight graph attention network to perform mapping calculations and obtain the stress response intensity of the silkworm; S5: Based on the timestamp of the stress response intensity of the silkworm, synchronize the environmental parameters, use the LSTM-GCN hybrid model to infer the microbial growth status, and generate the microbial growth potential index; S6: Perform a weighted summation operation based on the microbial growth potential index and the stress response intensity of the silkworm, and generate a dynamic threshold by combining the rate of change compensation factor; S7: Compare the real-time risk value with the dynamic threshold and if the value exceeds the threshold, generate a pollution risk warning signal. S8: Based on the consistency verification between the pollution risk warning signal and the actual detection results, the weight coefficients in the weighted summation operation are corrected online to generate updated weight coefficients.

[0011] The method for early warning and analysis of microbial contamination risk in stored silkworm feed provided in this application has the following beneficial effects: (1) By introducing a dynamic threshold generation mechanism based on the live behavior signals of silkworm colonies, this scheme effectively overcomes the fundamental defects of traditional warehouse microbial risk monitoring, such as poor adaptability and delayed response caused by relying on static environmental thresholds or statistical mean correction. Existing technologies usually rely on fixed temperature and humidity thresholds or multi-source data weighted fusion for risk judgment, which makes it difficult to capture the true temporal characteristics of biological cascade reactions and is prone to false alarms or missed alarms under complex working conditions. However, this invention uses a non-contact sensing array to continuously collect three types of biological response signals: feeding action density, surface micro-temperature fluctuation spectrum, and population aggregation entropy value. Combined with Granger causality test, the average response delay of "colony growth → behavioral inhibition → physiological imbalance" is calibrated to be 4.2 ± 0.7 hours, realizing cross-modal temporal alignment and causal chain locking from pollution inducing factors to biological stress. This design enables the system to sense changes in silkworm rhythms hundreds of milliseconds in advance before microbial contamination reaches the physical detection limit, significantly improving the early warning sensitivity and predictability, and solving the problem of blind spots in risk identification caused by the lack of biological feedback dimensions in traditional methods.

[0012] (2) An online optimization mechanism is constructed by integrating a dual-channel risk mapping function and interpretability weights, enabling the scheme to have strong robustness and self-evolutionary capability while ensuring prediction accuracy. The forward channel adopts an LSTM-GCN hybrid model, which outputs the microbial growth potential index P by combining environmental and feed physicochemical parameters. t The backward channel reflects external pollution trends; the backward channel is based on the real-time stress intensity S of the silkworm colony. t and its rate of change |dS t / dt| Construct a biofeedback response surface to form a two-way, mutually verified risk assessment architecture. Especially when S tWhen the mutation exceeds the 95th percentile of history and its derivative turns from negative to positive, the dynamic threshold T dynamic Automatically compress to P t The system achieves a 72% success rate, triggering a high-sensitivity early warning mode and effectively capturing acute stress events. More importantly, the system embeds an online reinforcement learning framework to continuously fine-tune the weight coefficients α, β, and γ, and performs credibility verification every 24 hours: if three consecutive early warnings are not accompanied by the detection of the target pathogen, a feedback compensation mechanism is activated, increasing the weight of γ and injecting "false positive" samples for incremental training, thereby achieving closed-loop evolution of the model in actual operation. This mechanism completely eliminates the reliance on external priors such as Bayesian optimization, transfer learning, or knowledge graph rules, significantly reducing the need for manual intervention and parameter tuning costs, and improving the system's autonomous adaptability and deployment convenience.

[0013] (3) The proposed edge-end collaborative architecture fully meets the stringent requirements of real-time performance and computational efficiency in the silkworm feed storage scenario, while also achieving the feasibility of high-precision risk assessment in engineering implementation. The entire link signal processing and inference process is completed on the edge side, with the edge-side inference latency controlled within 80ms, supporting a high-frequency monitoring rhythm of once every 10 minutes, ensuring near real-time response to sudden pollution events. Compared with methods that rely on centralized knowledge graph inference or batch fusion of multi-source heterogeneous data, this solution significantly reduces the computational load by using a lightweight graph attention network and sliding window normalization strategy, while ensuring an F1-score of 0.93 (37.6% higher than the fixed threshold method), avoiding complex parameter tuning and high computational costs. In addition, the entire system uses live biological feedback as the core criterion, which not only enhances the biological rationality and interpretability of risk warning, but also provides a new "biological-digital" fusion perception paradigm for the agricultural Internet of Things field, with good scalability and cross-scenario migration potential.

[0014] In summary, this solution, by constructing a complete closed loop of "perception-inference-decision-verification-evolution," integrates the dynamic behavior of living indicator organisms into the risk threshold generation process for the first time. It overcomes the inherent limitations of static threshold mechanisms in terms of dynamic environmental adaptability, and achieves synergistic enhancement of early warning sensitivity, prediction accuracy, and system adaptability. This provides a new intelligent monitoring path for silkworm feed storage safety that is highly reliable, low-latency, parameter-tuning-free, and capable of sustainable evolution. Attached Figure Description

[0015] Figure 1 This is the main flowchart of a method for early warning analysis of microbial contamination risks in stored silkworm feed; Figure 2 This is a sub-flowchart of a method for early warning analysis of microbial contamination risks in stored silkworm feed; Figure 3This is another sub-flowchart of a method for early warning analysis of microbial contamination risks in stored silkworm feed. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0018] like Figure 1 As shown, this application provides a method for early warning analysis of microbial contamination risks in stored silkworm feed, specifically including: S1: Acquire millimeter-wave radar signals and environmental parameters at multiple different locations within the storage unit, and uniformly define the collected results as raw monitoring data; S2: Perform cross-modal time-series alignment and noise filtering on the original monitoring data to eliminate signal interference and obtain preprocessed monitoring data; S3: Based on the preprocessed monitoring data, extract the feeding rhythm decay rate and the low-frequency amplitude of the body surface microtemperature, and define the extraction results as silkworm biological characteristic variables; S4: Based on the biological characteristic variables of the silkworm, input the lightweight graph attention network to perform mapping calculations and obtain the stress response intensity of the silkworm; S5: Based on the timestamp of the stress response intensity of the silkworm, synchronize the environmental parameters, use the LSTM-GCN hybrid model to infer the microbial growth status, and generate the microbial growth potential index; S6: Perform a weighted summation operation based on the microbial growth potential index and the stress response intensity of the silkworm, and generate a dynamic threshold by combining the rate of change compensation factor; S7: Compare the real-time risk value with the dynamic threshold and if the value exceeds the threshold, generate a pollution risk warning signal. S8: Based on the consistency verification between the pollution risk warning signal and the actual detection results, the weight coefficients in the weighted summation operation are corrected online to generate updated weight coefficients.

[0019] Step S1: Acquire millimeter-wave radar signals and environmental parameters from multiple locations within the storage unit, and uniformly define the collected results as raw monitoring data. Specifically, this includes: S1.1: Deploy millimeter-wave radar sensor arrays and environmental parameter sensing nodes at multiple preset spatial coordinate points within the storage unit. Generate radar transmission control commands and environmental sampling trigger signals with unified timestamps based on the pulse repetition frequency synchronization mechanism to establish a spatiotemporal reference framework for multi-source data acquisition.

[0020] For the deployment process of millimeter-wave radar sensor arrays and environmental parameter sensing nodes at preset spatial coordinate points within a storage unit, the input conditions are known overall spatial layout data of the storage unit and physical specifications of each sensor, including the operating frequency band of the millimeter-wave radar signal, the directivity index of the antenna array, and the detection range and accuracy level of the environmental sampling module.

[0021] A coordinate mapping algorithm based on the spatial layout matrix of storage units is adopted (parameters: the origin of the coordinate system is the entrance of the storage unit, and the mapping resolution is 0.1m) to achieve precise positioning of sensor nodes in three-dimensional space, and a sensor deployment index table is generated on the layout matrix. Furthermore, a timing coordination algorithm based on the pulse repetition frequency (PRF) synchronization mechanism (parameter: PRF set value is 10Hz, time reference comes from a unified GPS clock) is adopted to achieve time synchronization between radar transmission control commands and environmental sampling trigger signals, and generate a unified timestamp index sequence.

[0022] Furthermore, through the multi-source trigger signal generation module, the synchronous trigger conditions for radar transmission and environmental sampling signals are constructed using logical AND operations, and the trigger signals are coupled to the control interfaces of each node to generate trigger packets that can be used for downstream data acquisition.

[0023] By combining pulse synchronization and coordinate mapping, the deployment location and unified trigger time are bound as spatiotemporal reference data, achieving spatial and temporal consistency of multi-source acquisition.

[0024] For example, in a silkworm feed storage unit measuring 20 meters long, 10 meters wide, and 4 meters high, the spatial layout matrix is ​​divided into 200×100×40 grid cells. The sensor coverage radius is set to 1.5 meters, the millimeter-wave radar operates at 77 GHz, and the temperature accuracy of the environmental sensing nodes is ±0.1℃, and the humidity accuracy is ±1%. A coordinate mapping algorithm is used to deploy the radar nodes at a height of 2 meters above the ground, with 16 radar nodes arranged according to a minimum distance threshold. Twelve environmental sensing nodes are deployed at the four corners and the center of the storage unit. A PRF synchronization mechanism is used to unify the radar transmission control commands and environmental sampling trigger signals within a 0.1-second accuracy range, generating a unified timestamp index sequence to ensure cross-source spatiotemporal consistency of data collected by all nodes. In actual testing, this deployment scheme maintains a significantly improved signal coverage across the entire space under full-load silkworm feed conditions, and the trigger delay remains stable within 5 milliseconds, providing a highly reliable spatiotemporal reference for the subsequent synchronous execution of millimeter-wave radar transmission and environmental sampling.

[0025] S1.2: Based on the radar transmission control command, drive the millimeter-wave radar sensor array to transmit frequency-modulated continuous wave signals to the silkworm activity area, and receive the millimeter-wave radar echo signals formed by reflection from the surface of the silkworm body, so as to obtain the raw radar data stream containing the micro-motion Doppler frequency shift information of the silkworm group.

[0026] Driven by radar transmission control commands, a millimeter-wave radar sensor array is used to achieve non-contact micro-motion imaging scanning of the silkworm colony's activity area by employing a frequency-modulated continuous wave transmission method (parameters: frequency modulation range 76GHz-81GHz, modulation period 100ms).

[0027] Furthermore, by using a coherent demodulation method (parameters: the local oscillation signal frequency is synchronized with the transmitted signal frequency, and the sampling rate is at least 10kHz), the phase and amplitude information of the radar echo signal is acquired, and a baseband signal sequence containing the micro-motion reflection characteristics of the silkworm body surface is obtained.

[0028] Furthermore, using Fast Fourier Transform (FFT) spectral analysis (parameters: Hanning window type, window length 1024 points), the frequency domain distribution of the baseband signal sequence is calculated, and a spectral vector containing Doppler frequency shift information is generated.

[0029] Furthermore, the instantaneous velocity of the silkworm colony's micro-movements was calculated using the Doppler frequency shift extraction formula: Among them, f d For the Doppler frequency shift of the radar echo, f c denoted as the center frequency of the transmitted signal, and c as the speed of light.

[0030] Furthermore, a bandwidth-limited filter (parameter: cutoff frequency of ±0.1Hz) is used to remove low-frequency background drift not originating from silkworms, outputting the raw radar data stream containing the Doppler frequency shift characteristics of the silkworm colony.

[0031] By transmitting frequency-modulated continuous waves, coherent demodulation, FFT spectrum analysis, and Doppler frequency calculation, the radar transmission control commands generated in the previous step are transformed into high-precision raw radar data that can be used for subsequent biomarker extraction, thereby achieving precise capture of the micro-movement state of silkworm colonies.

[0032] For example, within a storage unit, a millimeter-wave radar array is configured with four sensors, a center frequency of 77.5 GHz, a modulation bandwidth of 4 GHz, a modulation period of 80 ms, and a sampling rate of 12.5 kHz. The local oscillation signal and the transmitted signal are synchronized using phase-locked loop (PLL). The FFT window length is set to 2048 points. After correction using the Hanning window function coefficients, a Doppler frequency shift of 0.05 Hz corresponding to the spectral peak is obtained. Applying the formula, the instantaneous velocity of the silkworm colony is calculated to be 0.0097 m / s. After removing background drift using a bandpass filter with a cutoff of ±0.1 Hz, the output raw radar data stream shows a significantly improved signal-to-noise ratio in subsequent feeding behavior analysis. The micro-motion frequency shift curve remains stable during continuous monitoring, meeting the timing accuracy requirements of the dynamic threshold generation module.

[0033] S1.3: Based on the environmental sampling trigger signal, drive the environmental parameter sensing node to perform multi-point parallel sampling operation, and capture temperature field distribution data, humidity field distribution data, carbon dioxide concentration data and ammonia concentration data in the storage space in real time to generate a multi-dimensional environmental parameter dataset.

[0034] Based on the environmental sampling trigger signal, the environmental parameter sensing nodes are driven, and a multi-point parallel sampling scheduling algorithm is adopted (parameters: sampling period of 2 seconds, number of parallel threads = number of sensing nodes) to simultaneously collect temperature, humidity, carbon dioxide concentration and ammonia concentration at preset coordinate points in the storage space.

[0035] Furthermore, by using the spatial distribution matrix reconstruction method (parameters: node coordinate set, sensor identification code), the real-time measurement results of each sampling point are mapped onto the three-dimensional grid model of the storage space, and the temperature field distribution data matrix and humidity field distribution data matrix are obtained.

[0036] By employing multi-point parallel sampling scheduling, spatial matrix reconstruction, data drift compensation, and multi-dimensional encapsulation processing, the real-time measurement results of each sensing node are transformed into a unified format of multi-dimensional environmental parameter dataset, thereby achieving the synchronization and standardization of original multi-source environmental information.

[0037] For example, 10 temperature and humidity sensors, and 5 carbon dioxide and 5 ammonia sensors are deployed in the silkworm feed storage unit. The nodes are evenly distributed according to the storage plane grid. The sampling period is set to 2 seconds, and the number of parallel threads is 20, the total number of sensor nodes, to ensure synchronous data acquisition from all nodes. The temperature sensor outputs in degrees Celsius, which is reconstructed into a 10×10 resolution temperature field matrix after spatial distribution matrix reconstruction. The humidity sensor outputs in percentage, which is reconstructed into a 10×10 humidity field matrix after spatial distribution matrix reconstruction. The carbon dioxide and ammonia sensors output in ppm, which are converted using concentration formulas. Where C represents the original concentration value from the sensor, S is the standardization coefficient, and K is the calibration constant, converted to standard units conforming to GB / T18204.2, and a 1.5 ppm drift error is eliminated through time-division sampling compensation. The four types of distributed data are encapsulated into a multi-dimensional environmental parameter dataset under a unified timestamp. Its data structure is time index + temperature matrix + humidity matrix + carbon dioxide matrix + ammonia matrix. Verification shows that this dataset maintains high integrity and consistency in subsequent time-series alignment and noise filtering, significantly improving the environmental input accuracy for dynamic threshold prediction.

[0038] S1.4: Perform frame-level encapsulation processing based on a unified timestamp on the original radar data stream and the multidimensional environmental parameter dataset, and use data packaging to integrate heterogeneous signal streams into data packets with a fixed frame structure to form a standardized original monitoring data sequence.

[0039] S1.5: Based on the standardized raw monitoring data sequence, perform integrity verification and outlier marking processing, and use sliding window statistical verification to remove invalid data fragments caused by instantaneous sensor failures. Finally, output the confirmed usable raw monitoring data for subsequent preprocessing modules to call.

[0040] Step S2: Perform cross-modal time-series alignment and noise filtering on the raw monitoring data to eliminate signal interference and obtain preprocessed monitoring data. Specifically, this includes: S2.1: Perform linear interpolation resampling on the millimeter-wave radar point cloud sequence and environmental sensor timestamps in the original monitoring data to unify the sampling frequency of multi-source heterogeneous data and generate a synchronized monitoring data stream with a unified time reference.

[0041] Linear interpolation resampling processing was performed on the millimeter-wave radar point cloud sequence and the environmental sensor timestamps in the original monitoring data (parameter: millimeter-wave radar sampling frequency f). r Environmental parameter sampling frequency f e Unified target sampling frequency f t This enables waveform reconstruction of signals acquired from different sources under a unified time reference.

[0042] Furthermore, by using a multi-segment piecewise linear interpolation algorithm (segment width set according to the original timestamp interval Δt), the intensity values ​​of millimeter-wave radar point clouds and environmental parameter values ​​are continuously estimated within the corresponding time period, and an interpolation data matrix containing the timestamps of each interpolation node is obtained.

[0043] Furthermore, time base normalization is employed (to unify the target sampling frequency f). t Based on the benchmark, an integer multiple mapping transformation is performed on all timestamps to achieve index alignment of interpolated data from different sources on a unified sampling period and generate a multi-source synchronized time series.

[0044] Furthermore, a sampling density smoothing algorithm is used (with the window length ω set to f). t (integer multiples of the time series) to smooth interpolation noise in the synchronized time series, ensuring signal continuity of data under a unified time base, and generating a synchronized monitoring data stream with a unified time base.

[0045] Through the above multi-segment linear interpolation and time base normalization processing, the original monitoring data results from the previous step are transformed into data objects with a unified sampling frequency and consistent time index, thereby achieving complete synchronization of multi-source heterogeneous signals in the time dimension.

[0046] For example, in the monitoring of a silkworm feed storage unit, the sampling frequency f of the millimeter-wave radar sensor is... r The sampling frequency f of the environmental parameter sensing node is 10Hz. e The target sampling frequency is 1Hz, and the sampling frequency f is uniform. t The frequency was set to 5Hz. The millimeter-wave radar point cloud sequence (one data point every 100ms) and environmental parameter timestamps (one data point every 1s) were input into a piecewise linear interpolation algorithm. The parameter Δt was divided into 1s intervals between the millimeter-wave radar data and the environmental parameter data. Interpolation calculations generated equally spaced data nodes at 0.2s intervals. Time base normalization was applied, mapping all data timestamps to indices that are multiples of 0.2s. A sliding smoothing algorithm with a window length ω=1s was used to perform mean smoothing on the interpolation sequence, eliminating sharp fluctuations during the interpolation process and achieving smooth signal synchronization. After processing, the generated synchronized monitoring data stream showed no time offset between the millimeter-wave radar and environmental parameters, meeting the accuracy requirements for cross-modal time sequence alignment and laying the data foundation for subsequent cross-correlation peak calculations.

[0047] S2.2: Calculate the peak position of the cross-correlation function based on the synchronized monitoring data stream to quantify the transmission delay difference between the millimeter-wave radar signal and the environmental parameter signal, and generate a time-series offset correction vector containing precise time delay compensation.

[0048] A cross-correlation analysis method (parameter settings: maximum lag range ±120 seconds, interpolation step size 0.1 seconds) was used to quantitatively determine the correlation between millimeter-wave radar signal sequences and environmental parameter signal sequences in the synchronized monitoring data stream.

[0049] Furthermore, by using the cross-correlation function calculation method (the formula follows the principle of discrete time series correlation calculation), the delay difference is numerically expressed, and the cross-correlation function curve is obtained.

[0050] For example, within a storage unit, the sampling frequency of the millimeter-wave radar signal is 10Hz, and the sampling frequency of the environmental parameter signal is 1Hz. After linear interpolation processing in S2.1, these frequencies are unified to 4Hz. In the same batch of monitoring, the maximum lag range for cross-correlation analysis is set to ±60 seconds, with an interpolation step size of 0.05 seconds. The calculation results show that the peak of the cross-correlation between the radar signal and the environmental humidity signal occurs at a delay of 14.3 seconds, with a peak amplitude of 0.92. After processing according to the normalized interval [-1,1], the accurate time delay compensation is 0.238. This compensation is encapsulated as a double-precision array with a length equal to the number of sensors, where each element corresponds to the time delay correction value of different sensor nodes. In subsequent time axis translation operations, this time offset correction vector is directly applied to synchronize the monitoring data stream, significantly reducing the analysis bias caused by cross-modal phase asynchrony. This shrinks the time error range of the subsequent S2.3 cross-modal alignment dataset to ±0.3 seconds, effectively improving the accuracy of silkworm biofeature extraction.

[0051] S2.3: Use the time offset correction vector to perform a sliding window translation operation on the synchronized monitoring data stream to eliminate the phase asynchrony between multimodal sensors and generate a cross-modal aligned dataset with strict time axis alignment.

[0052] S2.4: Apply an adaptive wavelet threshold denoising algorithm to the cross-modal aligned dataset to separate and filter out high-frequency electromagnetic interference and mechanical vibration noise in the storage environment, generating a purified monitoring signal matrix with a significantly improved signal-to-noise ratio.

[0053] S2.5: Based on the purified monitoring signal matrix, perform outlier removal and linear data filling to repair data breaks caused by sensor momentary failures, and finally output complete and continuous preprocessed monitoring data that meets the requirements of analytical accuracy.

[0054] like Figure 2 As shown, step S3 involves extracting the feeding rhythm decay rate and the low-frequency amplitude of the body surface microtemperature based on the preprocessed monitoring data, and uniformly defining the extraction results as biological characteristic variables of the silkworm. Specifically, this includes: S3.1: Perform short-time Fourier transform processing on the millimeter-wave radar echo signal sequence in the preprocessed monitoring data to obtain a time-frequency energy distribution map containing Doppler frequency shift information of the silkworm colony micro-movement, which serves as the basic data source for identifying the frequency changes of feeding actions.

[0055] The millimeter-wave radar echo signal sequence in the preprocessed monitoring data is decomposed in the time-frequency domain using a short-time Fourier transform algorithm (parameter settings: window function type is Hanning window, window length is 256 sampling points, frame shift is 128 sampling points, and FFT number is 512).

[0056] Furthermore, by introducing a framing mechanism with an overlap ratio of 50% in the short-time Fourier transform operation, high-resolution capture of the micro-motion cycle of the silkworm flock's feeding action is achieved, and dual-channel time-frequency matrix data results containing amplitude spectrum and phase spectrum are obtained.

[0057] Furthermore, based on the dual-channel time-frequency matrix data results, normalization processing is performed (amplitude is normalized to the [0,1] interval, and phase is normalized to the [-π,π] interval) to achieve scale uniformity of signal intensity at different radar acquisition locations and generate time-frequency spectrum data that can be compared across locations.

[0058] Furthermore, using the Doppler frequency shift extraction algorithm, the energy centroid along the frequency axis in the normalized amplitude spectrum is calculated, and the energy centroid frequency is calculated using the following formula: Where K is the frequency index, P is the energy value at the corresponding frequency, and f is the physical frequency corresponding to the frequency index.

[0059] Furthermore, by utilizing the time-series variation curve of the energy centroid frequency, Doppler frequency shift features related to the feeding action cycle of the silkworm colony are extracted, and a time-frequency energy distribution map containing the distribution of micro-motion frequencies within each time window is generated.

[0060] By using short-time Fourier transform and Doppler frequency shift extraction, the preprocessed radar data from the previous step is transformed into a time-frequency energy distribution map that can intuitively reflect the micro-movement patterns and frequency changes of feeding actions of silkworm colonies, thus achieving the basic data extraction effect of dynamic changes in feeding behavior.

[0061] For example, a millimeter-wave radar sensor array was deployed in a batch of silkworm feed storage unit, with a sampling frequency of 100Hz. The preprocessed radar echo sequence collected had a length of 60,000 points. A Hanning window with a window length of 256 points was used, and a short-time Fourier transform was performed with a frame shift of 128 points, resulting in an FFT of 512 points and a time-frequency matrix of 468 frames. The amplitude spectrum was normalized, with energy values ​​uniformly set to [0,1], and the phase spectrum was normalized to [-π, π]. The energy centroid frequency of each frame amplitude spectrum was calculated, where the energy centroid frequency was concentrated in the range of 0.35Hz to 0.45Hz during feeding activities and decreased to the range of 0.08Hz to 0.12Hz during non-feeding activities. The generated time-frequency energy distribution map showed a significant bandwidth expansion characteristic during the peak feeding activity period, verifying that the feature extraction effect of this step could be used for subsequent peak detection and calculation of the feeding rhythm decay rate.

[0062] S3.2: Based on the time-frequency energy distribution map, perform peak detection and trajectory association to extract the effective feeding action count sequence within a unit time window, and calculate the slope change value of the count sequence over time through linear regression fitting, thereby generating the feeding rhythm decay rate that characterizes the declining trend of the silkworm population's feeding activity.

[0063] S3.3: Perform bandpass filtering on the infrared thermal imaging temperature field data stream in the preprocessed monitoring data, and use a Butterworth filter with a cutoff frequency of 0.01 Hz to 0.1 Hz to separate high-frequency environmental noise and low-frequency physiological fluctuations to obtain a pure low-frequency fluctuation signal sequence of the silkworm body surface micro-temperature.

[0064] The infrared thermal imaging temperature field data stream in the preprocessed monitoring data is subjected to a bandpass filtering method (cutoff frequency parameters: 0.01Hz and 0.1Hz) to achieve frequency domain separation of the original temperature signal.

[0065] Furthermore, by using a second-order Butterworth filter design method (normalized frequency: 0.01 to 0.1, sampling frequency automatically calculated based on imaging acquisition frequency), the high-frequency environmental noise components and extremely low-frequency trend drift components are weakened, and the surface temperature fluctuation data of silkworms are obtained within the stress response frequency band.

[0066] Furthermore, a filter amplitude-frequency response correction method is adopted (parameters: filter order = 2, maximum passband attenuation). (dB) to compensate for the amplitude distortion of the output signal and generate a low-frequency temperature fluctuation sequence with minimized phase distortion.

[0067] Furthermore, a buffer smoothing process is performed using a window function smoothing algorithm (Hanning window, window length = sampling frequency × 10 seconds) to suppress the discontinuities at the edges of the low-frequency fluctuation sequence and generate a pure low-frequency fluctuation signal of the silkworm body surface micro-temperature that meets the requirements of temporal continuity.

[0068] By using the above-mentioned bandpass filtering and smoothing methods, the temperature field data from the previous step is transformed into low-frequency temperature fluctuation time series data that can be directly used for stress response analysis, thereby achieving precise isolation of the physiological fluctuation signals of silkworms.

[0069] For example, under the condition that the infrared thermal imaging module of the silkworm feed storage unit acquires data at a frequency of 50Hz, the preprocessed temperature field data is input into a second-order Butterworth bandpass filter, and the normalized cutoff frequency is calculated as follows: to (where the sampling frequency divided by 2 is the Nyquist frequency of 25Hz), thus obtaining the filter coefficient set.

[0070] The temperature data sequence was convolved onto a filter coefficient set to eliminate high-frequency interference faster than 100 seconds and extremely low-frequency drift slower than 10 seconds. A Hanning window smoothing process was applied with a window length of 500 sampling points, and amplitude transition control was performed on the edge waveforms to generate a time-continuous silkworm body surface temperature fluctuation signal containing only a 0.01-0.1Hz component. In subsequent Hilbert-Huang transform decomposition, this signal showed a significantly improved amplitude fluctuation in the stress response frequency band, while the background noise amplitude was significantly reduced, meeting the accuracy requirements for biometric variable extraction.

[0071] S3.4: Based on the low-frequency fluctuation signal sequence of the silkworm body surface micro-temperature, perform Hilbert-Huang transform decomposition processing to extract the instantaneous amplitude envelope of the intrinsic mode function components corresponding to the stress response frequency band, and calculate the root mean square statistic of the envelope within the sliding window, thereby generating the low-frequency amplitude of the silkworm body surface micro-temperature to characterize the degree of metabolic disorder.

[0072] Based on the low-frequency fluctuation signal sequence of the silkworm's body surface micro-temperature, the Hilbert-Huang transform algorithm (parameters: empirical mode decomposition resolution 0.0005℃, stress response frequency band defined as 0.01-0.1Hz) is used to achieve adaptive decomposition of non-stationary low-frequency physiological signals.

[0073] Furthermore, the Empirical Mode Decomposition (EMD) method (parameters: stopping criterion threshold of 0.05, maximum number of iterations of 20) is used to decompose the input signal into a set of intrinsic mode function components (IMFs). k And obtain the time series representation of each component.

[0074] Furthermore, by utilizing the Hilbert transform (parameter: Fast Fourier Transform (FFT) window size of 1024 points), the analysis of each IMF can be performed. k Analytical signal generation is performed to obtain the instantaneous amplitude envelope sequence A. k (t).

[0075] Furthermore, by using a bandpass filter, only the instantaneous amplitude envelope A of the stress response frequency band is retained. s (t), and use it as the core feature input of the stress response signal.

[0076] Furthermore, the root mean square value (RMS) of the instantaneous amplitude envelope within the time window is calculated using the sliding window method (window length = number of equivalent sampling points equal to the average stress response delay τ = 4.2 hours). A This allows for the creation of a quantitative indicator to characterize the degree of metabolic disorder in silkworms. The formula is as follows: Among them, A s Let t be the instantaneous amplitude envelope of the stress frequency band. i Where N is the sampling time, and N is the total number of samples within the window.

[0077] Through the aforementioned algorithm or processing method, the low-frequency fluctuation signal from the previous step is transformed into a quantifiable biological characteristic—the low-frequency amplitude of surface micro-temperature—enabling a precise characterization of the degree of metabolic disorder in silkworms.

[0078] For example, in a silkworm feed storage management scenario, bandpass filtering is performed on the infrared thermal imaging temperature field data stream, with cutoff frequency parameters set to 0.01Hz and 0.1Hz, and a sampling frequency of 1Hz. EMD decomposition yields six IMF components, with IMF3 concentrated in the stress response frequency band. A Hilbert transform is performed on IMF3 to obtain the instantaneous amplitude envelope, with a peak value of approximately 0.12℃. The root mean square (RMS) is calculated using a sliding window length of 15120 points (corresponding to a 4.2-hour delay), yielding a value of 0.083℃. This RMS value is above the 95th percentile of historical data, indicating that the silkworm population is under significant stress. The output low-frequency amplitude index of surface microtemperature serves as a biofeedback correction factor in subsequent dynamic threshold generation, significantly improving the system's sensitivity and stability to microbial contamination risks.

[0079] S3.5: Perform spatiotemporal coordinate binding and format encapsulation processing on the feeding rhythm decay rate and the low-frequency amplitude of the body surface micro-temperature. Use a unified timestamp to integrate the two types of heterogeneous scalar and vector data into a structured data object. Finally, output a standard dataset defined as the biological characteristic variables of the silkworm for subsequent graph attention network calls.

[0080] like Figure 3As shown, step S4 involves inputting the silkworm's biological characteristic variables into a lightweight graph attention network for mapping calculation to obtain the silkworm's stress response intensity. Specifically, this includes: S4.1: Based on the feeding rhythm decay rate and body surface micro-temperature low-frequency amplitude in the biological characteristic variables of the silkworm, construct heterogeneous graph data containing spatial location information and time series characteristics, define each millimeter-wave radar sensor node in the storage unit as a graph node, define the spatial adjacency relationship and signal correlation between nodes as graph edges, and generate an initial topological graph structure to characterize the distribution of micro-ecological disturbances in the silkworm colony.

[0081] Based on the encapsulated dataset of silkworm biological characteristic variables, a spatial mapping construction algorithm (parameters: spatial coordinates of millimeter-wave radar nodes, physical layout matrix within the storage unit) is used to extract and standardize the spatial location information of each sensor node.

[0082] Furthermore, a time series feature fusion method (parameters: continuous value sequence of feeding rhythm decay rate, low frequency amplitude spectrum matrix of body surface microtemperature) is used to bind the time series features corresponding to each node and associate them with spatial coordinate information to form node description data containing time series attributes.

[0083] Furthermore, an initial graph structure generation method (parameters: number of nodes, edge weight mode) is constructed using spatial adjacency relationships to achieve full definition of graph nodes and graph edges. Each millimeter-wave radar sensor node in the storage unit is explicitly mapped as a graph node, and graph edges are generated based on the joint conditions of correlation and adjacency, thus completing the generation of the initial topology graph for the characterization of micro-ecological disturbance distribution.

[0084] By using the above spatial and temporal information fusion processing method, the biological characteristic variables of the silkworm in the previous step are transformed into heterogeneous graph data containing node spatial location, time series characteristics and weighted edge relationships, so as to realize a topological description of the distribution of microecological disturbances in the silkworm population.

[0085] For example, in a silkworm feed storage unit with an area of ​​120 square meters and a height of 3.5 meters, 12 millimeter-wave radar sensor nodes were deployed. The spatial coordinate accuracy of each node was ±0.05 meters, and the physical layout matrix was generated by measuring the actual distance between the nodes. The feeding rhythm decay rate data was recorded in units of 0.002 actions / second, and the low-frequency amplitude spectrum of body surface micro-temperature was obtained with a bandwidth of 0.05 Hz. The Pearson correlation coefficient was used to calculate the correlation of node signals, with a threshold set at 0.65. Combined with the measured delay compensation vector, cross-node data synchronization errors were eliminated. The edge weights between nodes adopted the product mode of the correlation value and the spatial adjacency coefficient. The generated initial topology graph contained 12 nodes and 28 weighted edges, with edge weights ranging from 0.42 to 0.89. This topology graph effectively preserved the spatial propagation pattern of silkworm colony stress response in the subsequent graph attention network, and it was verified that it can significantly improve the model's capture ability under weak pollution signal conditions.

[0086] S4.2: Perform node feature embedding processing on the initial topology graph structure. Use a linear transformation matrix to map the scalar value of the feeding rhythm decay rate and the low-frequency amplitude spectrum vector of the body surface microtemperature to a high-dimensional latent space to generate node feature embedding vectors carrying multi-dimensional biological behavioral semantic information, so as to eliminate the interference of different physical dimensions on the subsequent attention mechanism calculation.

[0087] S4.3: Based on the node feature embedding vector, perform multi-head self-attention mechanism calculation, quantify the association weight of biological behavioral features between adjacent nodes by performing dot product operation of query vector, key vector and value vector, and generate dynamic attention coefficient matrix reflecting the abnormal propagation path of local silkworm population, so as to identify the group stress response pattern caused by microbial contamination.

[0088] Based on the node feature embedding vector input multi-head self-attention mechanism calculation module, by setting the number of attention heads h and the query vector dimension d, the multi-head self-attention mechanism calculation module is used. q , key vector dimension d k The dimension d of the sum vector v The parameters enable parallel quantization of the correlation between features of adjacent nodes.

[0089] Furthermore, through the linear projection matrix W Q W K W V Matrix multiplication is performed on the node feature embedding vectors to map the original high-dimensional semantic features into three types of vectors: query, key, and value, resulting in grouped feature representations for different attention heads.

[0090] Furthermore, the similarity matrix is ​​subjected to exponential normalization by the softmax function (parameter: temperature factor τ) to realize the probabilistic expression of the association weights and obtain the neighborhood attention distribution matrix for each node.

[0091] Furthermore, by combining the spatial topological constraints within the storage unit, the normalized attention coefficients are limited to the set of physically adjacent nodes to generate a dynamic attention coefficient matrix that reflects the abnormal propagation path of the local silkworm colony.

[0092] By employing a multi-head attention merging strategy, the dynamic attention coefficient matrices output by each attention head are spliced ​​and fused along the feature dimension to obtain a comprehensive dynamic attention coefficient matrix that fully covers multiple biological behavioral patterns, thereby enabling the identification of mass stress response patterns caused by microbial contamination.

[0093] For example, the millimeter-wave radar sensor data nodes in the silkworm feed storage unit are 16, the node feature embedding vector dimension is 64, the number of multiple heads h is set to 8, and the dimensions of the query, key, and value vectors are all 8. This is achieved through the projection matrix W. Q W K W V Each 64-dimensional embedding vector is mapped to an 8-dimensional query, key, and value vector at the head, and grouped by head. A scaling factor is applied. Dot product similarity calculations were performed to generate a neighborhood similarity matrix of size (16×16), which was then normalized using softmax (temperature factor τ=0.9) to obtain the probability weight matrix. Spatial topological masking was performed on the probability weights, retaining only the coefficients of physically adjacent nodes to generate a dynamic attention coefficient matrix. The coefficient matrices corresponding to the eight heads were concatenated and fused dimensionally to obtain a comprehensive coefficient matrix of size (16×128), which was applied to the weight allocation before neighborhood feature aggregation. Ultimately, in a pollution event simulation, four stress response paths triggered by the spread of microbial contamination were identified, demonstrating a significantly improved ability to capture weak pollution signals.

[0094] S4.4: The dynamic attention coefficient matrix is ​​used to perform a weighted aggregation operation on the feature embedding vectors of neighboring nodes. Combined with the residual connection mechanism, the original features of the current node and the neighborhood context information are fused to generate an updated node hidden state representation containing global spatial dependencies, so as to enhance the model's ability to capture weak pollution signals.

[0095] S4.5: Based on the hidden state representation of the updated node, perform fully connected layer mapping and normalization processing to compress the high-dimensional hidden state into a single-dimensional scalar output, and generate the final evaluation value characterizing the stress response intensity of the overall silkworm body of the storage unit at the current moment, which serves as the biological feedback correction factor in the subsequent dynamic threshold adjustment mechanism.

[0096] Based on the hidden state representation of the updated node, the input fully connected mapping layer (parameters: hidden state dimension H=256, output dimension O=1) is used to compress multi-dimensional spatially dependent features into a single risk quantification index.

[0097] Furthermore, by combining weight matrix multiplication with bias vector operation (parameter: initialization method is Xavier, initial bias value is zero), a linear mapping from the hidden state to the scalar space is achieved, and the original value of the unnormalized silkworm stress response intensity is obtained.

[0098] Furthermore, batch normalization (parameters: batch size B=32, momentum coefficient μ=0.9) is used to standardize the mean and variance of the original values, generating normalized stress response intensity data with numerical stability and comparability.

[0099] For example, within a silkworm feed storage unit, the hidden state of an updated node is represented as a 256-dimensional vector. Each node is composed of the feeding rhythm decay rate and the low-frequency amplitude of body surface micro-temperature, collected by a millimeter-wave radar sensor and an infrared thermal imaging module. The fully connected mapping layer is initialized with Xavier initialization for the weight matrix, a bias vector of zero, a batch size of 32, and a momentum coefficient of 0.9. Batch normalization adjusts the mean of the original stress response intensity value to 0 and the standard deviation to 1. Subsequently, the Sigmoid function maps this value to the 0-1 interval. For example, a hidden state with a batch normalized value of 2.35 is represented by the formula... The normalized stress response intensity value was calculated to be approximately 0.913. This value is used as a biofeedback correction factor input in the subsequent dynamic threshold generation process. When this value continues to rise above the historical 95th percentile, the system triggers a high-sensitivity early warning mode, effectively achieving early identification and stable prediction of microbial contamination risks.

[0100] Step S5: Based on the timestamp of the silkworm stress response intensity, synchronize environmental parameters, and use the LSTM-GCN hybrid model to infer the microbial growth status and generate a microbial growth potential index. Specifically, this includes: S5.1: Perform sliding window cross-correlation matching processing based on the timestamp of the stress response intensity of the silkworm and the collection timestamp of the environmental parameters to eliminate clock drift errors between multiple source sensors and obtain a time-synchronized fusion dataset.

[0101] S5.2: Perform spatial topology mapping construction processing on the environmental parameters in the time synchronization fusion dataset, define the node connection relationship using the physical layout of the storage unit, and obtain the environmental state diagram structure data.

[0102] Furthermore, by using the Euclidean distance calculation method (parameter: three-dimensional coordinate difference), the spatial distance matrix between any two environmental parameter collection points is generated, and the data results used for adjacency determination are obtained.

[0103] Furthermore, a weighted adjacency matrix is ​​generated based on the initial adjacency matrix by performing a weight calculation method (parameters: temperature and humidity difference, CO2 concentration difference, and ammonia concentration difference), thereby reflecting the quantitative relationship of environmental state differences between different nodes.

[0104] Furthermore, through graph structure construction processing, node indexes, node features (current environmental parameter values), and weighted edge information are encapsulated into environmental state graph structure data, realizing a digital representation of spatial topology mapping.

[0105] By using a spatial topology mapping algorithm, the time-synchronized fusion data from the previous step is transformed into an environmental state graph structure, achieving the expected technical effect of graph convolutional network modeling with spatially distributed feature inputs.

[0106] For example, in a silkworm feed storage unit, 24 temperature and humidity sensors, 24 CO2 sensors, and 24 NH3 sensors are deployed. The overall coordinate range is X: 0-12 meters, Y: 0-8 meters, and Z: 0-3 meters. The sensor coordinate matrix accuracy reaches 0.01 meters. The internal partition wall thickness of the storage unit is 0.2 meters, and the airflow channel width is not less than 0.5 meters. A distance threshold d is set. thresh It is 2.0 meters, according to the formula Calculate the spatial distance between each pair of nodes and remove edges exceeding a threshold. For the retained edges, calculate the absolute values ​​of temperature difference, humidity difference, CO2 concentration difference, and NH3 concentration difference, take a weighted average as the edge weight, and generate a weighted adjacency matrix. Encapsulate the node indices (0-71), node features (four types of environmental parameter values), and weighted edge information into an environmental state graph data structure conforming to the GCN input format. In actual operation, this graph contains 72 nodes with an average node degree of 5.3, effectively reflecting the physical and environmental coupling relationships within the storage space. Subsequent input into GCN can significantly improve the ability to capture spatial features of microbial diffusion trends.

[0107] S5.3: Based on the environmental state graph structure data input graph convolutional network layer, perform spatial feature aggregation operation to extract the microbial diffusion coupling features between adjacent storage areas and obtain a spatially dependent feature vector sequence.

[0108] The input environmental state graph structure data includes node and edge relationship information obtained by mapping the physical layout of the storage units, along with the environmental parameter feature vectors of each node.

[0109] A graph convolutional network neighbor aggregation algorithm (parameters: 2 convolutional layers, 1 aggregation radius) is adopted to achieve weighted summation and fusion of environmental parameter features of each node and its neighboring nodes.

[0110] Furthermore, by weight normalization (parameter: normalization method is symmetric normalization), the influence of neighbors is evenly distributed across different node inputs, and a preliminary spatial feature representation is obtained.

[0111] Furthermore, the ReLU nonlinear activation function (parameter: threshold 0) is used to perform nonlinear transformation on the fused spatial features, thereby enhancing the model's ability to adapt to the complex structure of microbial diffusion patterns.

[0112] Furthermore, by utilizing a multi-layer stacked graph convolutional structure (number of layers: 3), iterative accumulation from local neighbor features to global environmental features is achieved, generating a comprehensive spatial feature matrix containing cross-regional information.

[0113] By using max pooling (with the pooling window covering all nodes), the node spatial feature matrix from the previous step is transformed into a sequence of spatially dependent feature vectors, enabling the global extraction of microbial diffusion coupling features between adjacent storage areas.

[0114] S5.4: Input the spatially dependent feature vector sequence and the stress response intensity of the silkworm body into the long short-term memory network layer to perform temporal evolution inference processing, so as to capture the nonlinear dynamic law of microbial growth changing over time and obtain the temporal hidden state feature representation.

[0115] Based on the spatially dependent feature vector sequence and the stress response intensity of the silkworm, a temporal input matrix is ​​constructed. A Long Short-Term Memory (LSTM) unit stacking structure (3 layers, 128 hidden units) is adopted to realize the joint sequence modeling of multi-time spatial features and biological feedback signals.

[0116] Furthermore, the state vectors of the input gate, forget gate, and output gate are calculated through a gating mechanism. The input gate is used to control the degree of writing of the current time-in features, the forget gate is used to adjust the retention ratio of historical hidden states, and the output gate is used to generate the observable representation of the current hidden state.

[0117] Furthermore, the spatially dependent feature vector sequence and the stress response intensity of the silkworm are normalized in the time dimension and then synchronously input into the LSTM unit. The weight parameter matrix is ​​updated using the sequence backpropagation algorithm (BPTT) to obtain the hidden state sequence that is sensitive to the nonlinear changes in microbial growth.

[0118] Through the aforementioned network structure and state update mechanism, the spatial features and biological feedback signals from the previous step are transformed into temporal hidden state features that can reflect the evolution of microbial growth over time, thereby achieving accurate capture of nonlinear dynamic laws.

[0119] For example, in a silkworm feed storage unit, the spatially dependent feature vector sequence is configured with a length of 96 time steps, with each vector having a dimension of 64. After regularization, it is concatenated with the silkworm stress response intensity at the corresponding time step (scalar normalized to the 0.0-1.0 range) to form a 65-dimensional temporal input matrix. This matrix is ​​input into a three-layer stacked LSTM network with 128 hidden units, using a hyperbolic tangent activation function, and the gated state vector is initialized to a Gaussian distribution with a mean of 0 and a variance of 0.01. After 300 iterations of training using the BPTT algorithm, the hidden state sequence is output when the loss function converges to a stable value. In this scenario, the hidden state features represent the inflection point time that can stably capture the microbial growth potential under different seasons and storage conditions, significantly advancing the predicted trigger time and significantly improving the stability of risk warning.

[0120] S5.5: Based on the temporal hidden state feature representation, numerical mapping transformation is performed through a fully connected regression head to quantify the microbial reproduction trend within a preset time period and obtain the microbial growth potential index.

[0121] Based on the temporal hidden state feature representation, the fully connected regression head is input to perform an affine mapping operation (parameters: weight matrix dimension matching temporal hidden state feature vector length and target prediction step size), thereby compressing the high-dimensional nonlinear dynamic feature space into a single-dimensional potential microbial growth potential prediction value.

[0122] Furthermore, the model weights are numerically optimized using the least squares regression training method (parameter: mean squared error (MSE) as the loss function), and the regression residual distribution characteristics are obtained to evaluate the prediction stability.

[0123] Furthermore, based on the regularized regression method (parameter: L2 regularization coefficient λ is set to 0.01), norm constraints on the weight matrix are implemented to reduce the risk of overfitting and generate a convergent and stable set of regression coefficients.

[0124] For example, in an environment where the temperature and humidity fluctuate between 22.5℃ and 26.3℃, humidity is 55%-62%, CO2 concentration is capped at 850ppm, and NH3 concentration does not exceed 30ppm in a silkworm feed storage unit, the input time-series hidden state feature vector has a length of 64, predicting the microbial growth trend over the next 8 hours. The fully connected regression head weight matrix is ​​configured as 64×1, and the training is performed using MSE loss for 500 iterations, with a convergence threshold of 1e. -5 The L2 regularization coefficient λ is set to 0.01, and the moving average window length is set to 3 time steps. In the calculation formula, T is taken as 8 hours, and the measured y... pred Within a continuous sampling period, y meanThe summation of the difference sequences, divided by T, yields a gpi of 2.46. This value, after standardization by the system, is mapped to a high growth potential level. In this scenario, the growth potential index serves as the core input for the subsequent dynamic threshold adjustment mechanism, significantly improving the accuracy of predicting microbial outbreak trends.

[0125] Step S6: Perform a weighted summation operation based on the microbial growth potential index and the silkworm stress response intensity, and generate a dynamic threshold by combining the rate of change compensation factor. Specifically, this includes: S6.1: Perform sliding window difference operation on the real-time collected time series data of silkworm stress response intensity to extract the stress response change rate that characterizes the trend of biological feedback mutation, as the basic input variable for subsequent compensation calculation.

[0126] S6.2: Based on the stress response change rate and the preset historical quantile benchmark, perform deviation measurement processing to generate a change rate compensation factor for quantifying the degree of abnormality of the current biological state, thereby realizing nonlinear correction of environmental disturbances.

[0127] S6.3: Using the interpretable weight coefficients output by the online reinforcement learning framework, a weighted summation operation is performed on the microbial growth potential index and the stress response intensity of the silkworm to construct a basic risk benchmark value that reflects the dual driving forces of environment and biology.

[0128] Based on the synchronized numerical vectors of the microbial growth potential index and the silkworm stress response intensity, the policy gradient algorithm in the online reinforcement learning framework (parameters: learning rate η=0.01, discount factor γ) is used. rl =0.95), enabling dynamic estimation and updating of the weight coefficients.

[0129] Furthermore, using the feature contribution analysis method in the interpretability module (parameter: SHAP sampling number = 1000), the relative importance of environmental and biological driving variables in risk assessment is quantified, and a temporary weight coefficient matrix W is obtained. temp .

[0130] Furthermore, by using a normalized constraint optimization method (parameter: L2 norm constraint threshold = 1), the W is optimized. temp Each weight coefficient is scaled to ensure that the weight ratio is within the range of numerical stability, and a normalized weight coefficient vector W is generated. norm .

[0131] Furthermore, by using the feedback loop of online reinforcement learning to input recent warning triggering and verification states as reward signals into the policy gradient update rule, continuous iterative optimization of the weight coefficient sequence is achieved, generating a stable and convergent interpretable set of weight coefficient parameters W. final .

[0132] By using the aforementioned online reinforcement learning framework and interpretable weight coefficient calculation method, the microbial growth potential index and silkworm stress response intensity from the previous step are transformed into basic risk benchmark values ​​that reflect the dual drivers of environment and biology, thereby achieving the synergistic integration of environmental and biological information in dynamic threshold generation.

[0133] For example, in the silkworm feed storage environment, the microbial growth potential index P t The value is 0.68, representing the stress response intensity of the silkworm. The value is set to 0.52, the online reinforcement learning framework sets the learning rate η=0.01, and the discount factor γ. rl =0.95, and the average reward signal obtained in the verification state collected in the most recent 24 hours was 0.87. Interpretability analysis showed that the contribution of environmental driving variables was 0.62, and the contribution of biological driving variables was 0.38. After L2 constraint normalization, α=0.62 and β=0.38 were obtained. Using the weighted summation formula: The calculated result was 0.618. Finally, through online reinforcement learning iterative optimization, the matching degree between the baseline risk value and the actual pollution event triggering ratio on the validation set was significantly improved, ensuring the stability and generalization ability of the dynamic threshold adjustment process.

[0134] S6.4: Perform linear superposition and fusion processing on the basic risk benchmark value and the rate of change compensation factor to generate an initial dynamic threshold containing a dynamic adjustment mechanism, thus completing the paradigm shift from static criteria to biological response-driven criteria.

[0135] Based on the structured numerical input of the basic risk benchmark value and the rate of change compensation factor, a linear superposition fusion algorithm (parameters: the fusion coefficient is set according to the optimization results of the weight matrix α, β and the rate of change compensation factor γ) is used to achieve the orderly merging of environmental driving indicators and biological feedback corrections.

[0136] Furthermore, the initial dynamic threshold is calculated using a weighted linear combination formula, as follows: in, The initial dynamic threshold is generated. Based on the baseline risk value, This is the rate of change compensation factor.

[0137] Furthermore, by utilizing the adaptive adjustment mechanism of the fusion coefficient in the linear superposition process (parameter: real-time calculated rate of change compensation factor and historical statistical quantile deviation), the benchmark value is dynamically compressed or expanded, so that the judgment interval corresponding to the threshold achieves a balance between high-sensitivity early warning and stability.

[0138] Furthermore, by binding structured data, the initial dynamic threshold is synchronously associated with the subsequent risk value calculation module according to a unified timestamp, so as to ensure the temporal consistency of the threshold and risk value in the real-time comparison process.

[0139] By using a linear superposition fusion algorithm, the basic risk benchmark value and rate of change compensation factor from the previous step are transformed into an initial dynamic threshold containing a dynamic adjustment mechanism, thereby realizing a paradigm shift from static criteria to biological response-driven criteria.

[0140] For example, in the environmental monitoring scenario of silkworm feed storage units, the basic risk benchmark value R is determined by the microbial growth potential index (P). t =4.35) and the stress response intensity of silkworms (S t =2.18) is obtained by weighted summation with weights α=0.65 and β=0.35, i.e. The calculation result is R=3.56, and the rate of change compensation factor C is determined by the stress response rate of change ΔS. t =0.42 and the deviation measure of the historical 95th percentile of 0.15 were calculated using a nonlinear correction model (coefficient γ=0.28), that is... The value is 0.16. The initial dynamic threshold is obtained by linearly superimposing R and C. =3.72, which serves as a dynamic criterion for subsequent risk value comparison. Under different seasonal conditions, if a sudden change in temperature and humidity causes C to increase to 0.31, T will adjust accordingly to 3.87, achieving real-time adaptation and sensitive response of the threshold, and significantly improving the accuracy of risk assessment in volatile environments.

[0141] S6.5: Perform boundary constraint verification on the initial dynamic threshold to ensure that the generated final dynamic threshold is within a physically achievable reasonable range, and output a standard dynamic threshold for real-time risk comparison.

[0142] Logical verification is performed on the initial dynamic threshold input boundary constraint rule set using an interval verification algorithm (parameter: lower threshold L). min upper limit of threshold L max This enables basic compliance checks on physically realizable regions.

[0143] Furthermore, a multivariate statistical analysis method based on historical operational data of the storage unit (parameter: time window W) is used. h Environmental parameter gain coefficient κ env This allows for dynamic adjustment of boundary values, resulting in a feasible threshold range corrected by both temporal and spatial constraints.

[0144] Furthermore, through a numerical normalization algorithm (parameter: normalization reference μ) ref Standard deviation σref The initial dynamic threshold is mapped to a dimensionless form, and a comparison operation is performed in combination with the boundary value range to generate a boundary over-limit flag.

[0145] For example, in a batch of silkworm feed storage units, the initial dynamic threshold T init The measured value is 32.5, and the lower limit L is set at the boundary. min The value is 28.0, and the upper limit is L. max The value is 36.0, and the time window W is... h Set to 24 hours, environmental gain coefficient κ env The normalized baseline μ is 0.85. ref The value is 30, and the standard deviation σ is 30. ref The value is 2.1. During boundary verification, T is first determined using an interval verification algorithm. init It does not exceed the static boundary range. Using multivariate statistical analysis, the upper limit is dynamically adjusted to 35.5, and the lower limit is dynamically adjusted to 28.3. After normalization, the comparison results have no limit-crossing flags. If T init If a limit is exceeded during future data collection, the condition reset algorithm will reset the condition by Δ. reset =0.5 and λ damp The parameter = 0.9 is continuously adjusted within the correction range. Finally, T is calculated using the formula. final =32.5, serving as the standard dynamic threshold for real-time risk comparison. During subsequent comparisons, this threshold remains stable without sudden breaches, ensuring that the early warning module can significantly improve judgment accuracy and execution reliability under various environmental fluctuations.

[0146] Step S7: Compare the real-time risk value with the dynamic threshold; if the value exceeds the threshold, generate a pollution risk warning signal. Specifically, this includes: S7.1: Perform a weighted fusion calculation on the microbial growth potential index and the stress response intensity of the silkworm, and combine it with the rate of change compensation factor to generate a real-time risk value, which serves as the core input variable for subsequent comparison and judgment.

[0147] S7.2: Based on the timestamp index of the real-time risk value, the dynamic threshold at the current moment is indexed, and the threshold fluctuation sequence within the most recent response delay period is extracted using a sliding window mechanism to construct an adaptive comparison benchmark.

[0148] S7.3: Perform difference calculation and sign discrimination processing on the real-time risk value and the adaptive comparison benchmark to identify the instantaneous out-of-bounds state where the real-time risk value exceeds the dynamic threshold and generate a primary out-of-bounds flag.

[0149] The difference calculation method (parameters: risk value sequence R(t), comparison benchmark sequence B(t)) is used to calculate the difference between the real-time risk value and the adaptive comparison benchmark to quantify the degree of deviation of the current risk value from the dynamic threshold.

[0150] Furthermore, a threshold-based boundary judgment rule (parameter: D(t)>0 is judged as boundary violation, D(t)≤0 is judged as not boundary violation) is adopted to realize instantaneous boundary violation state recognition and generate a primary boundary violation flag sequence.

[0151] For example, in a silkworm feed storage unit environment, the real-time risk value R(t) is calculated by fusing the microbial growth potential index and the silkworm's stress response intensity, with a value ranging from 0 to 1. The benchmark B(t) is output by the dynamic threshold generation module, with a value ranging from 0.3 to 0.8. When performing the difference calculation, the formula is used: in This is the risk value. As a benchmark, For time indexing. When determining symbols, when... If the value is greater than 0, an out-of-bounds flag of 1 is generated. If the value is less than or equal to 0, a flag bit of 0 is generated. To improve detection capability when approaching the threshold, sensitivity enhancement processing is employed to address the condition. Differences less than 0.05 are multiplied by a multiplication factor of η=3, i.e.: In actual operation, at a certain moment, when R(t) = 0.75, B(t) = 0.72, and the difference D(t) = 0.03, after sensitivity amplification, D'(t) = 0.09, and the sign discrimination result is an out-of-bounds flag of 1. Verification results show that this process can still achieve highly stable out-of-bounds determination even when environmental temperature and humidity fluctuate significantly, effectively improving the sensitivity of microbial contamination risk detection.

[0152] S7.4: Perform time filtering based on the duration of the primary boundary crossing flag to remove transient interference pulses shorter than the minimum biological response time, generate a confirmed boundary crossing event label, and ensure the stability of the early warning trigger.

[0153] S7.5: Based on the confirmed boundary crossing event label, invoke the preset early warning protocol to generate a pollution risk early warning signal containing pollution level, location information and expected outbreak time, completing the closed loop from state judgment to decision output.

[0154] Based on the confirmed boundary crossing event label, the event classification module in the preset early warning protocol is invoked, and a multi-threshold segmentation judgment method is adopted (parameters: the dynamic threshold quantile division interval is [0, 0.6], [0.6, 0.85], [0.85, 1]) to realize the corresponding mapping between boundary crossing state and pollution level.

[0155] Furthermore, by analyzing location information (parameters: millimeter-wave radar node coordinate matrix, environmental sensor node topology), the spatial coordinates of the storage unit containing the boundary crossing event source are calculated in reverse, and the location vector data of the pollution source is obtained.

[0156] Furthermore, by using the time-series regression method in the outbreak time prediction module (parameters: historical microbial growth potential index sequence, silkworm stress response intensity sequence, response delay coefficient τ = 4.2), the predicted outbreak time after the pollution index exceeds the high sensitivity threshold is calculated, and a pollution risk time-series prediction index is generated.

[0157] Furthermore, a data fusion processing method is adopted to encapsulate the pollution level results, location information vector data, and expected outbreak time indicators into structured fields, generating a pollution risk warning signal data packet that conforms to the preset warning protocol.

[0158] By using an event broadcasting mechanism with a pre-set early warning protocol, the early warning signal data packets from the previous step are transformed into unified risk information that can be received simultaneously by multiple terminals (warehouse management platform, mobile monitoring APP, and on-site audible and visual alarms), thereby achieving real-time alarm and closed-loop decision output for pollution risks across the entire domain.

[0159] For example, in a silkworm feed storage unit, the millimeter-wave radar node deployment coordinate matrix is ​​(x = 2.0m, y = 1.0m), (x = 5.0m, y = 1.0m), (x = 2.0m, y = 4.0m), and (x = 5.0m, y = 4.0m), forming a four-node ring topology for environmental sensing nodes. The dynamic threshold quantile division intervals are set to [0, 0.6], [0.6, 0.85], and [0.85, 1]. A risk value of 0.92 for boundary crossing events corresponds to a high-sensitivity warning level. The location information parsing input is the signal delay difference matrix of the millimeter-wave radar nodes affected by the boundary crossing event, and the pollution source location vector is calculated as (x = 5.0m, y = 4.0m). The outbreak time prediction module inputs the microbial growth potential index sequence and the silkworm stress response intensity sequence of the most recent 48 hours, and combines them with the response delay coefficient τ = 4.2, calculating the expected outbreak time as 6.5 hours from the current moment using a time-series regression method. The data fusion processing method encapsulates the risk level ("high sensitivity"), location vector (5.0m, 4.0m), and outbreak time (6.5 hours) into a structured data packet, and pushes it in real time to the warehouse management platform, mobile monitoring APP, and on-site audible and visual alarms via an event broadcast mechanism, achieving real-time alarm across the entire area. The application effect is that, under the triggering condition of a high-sensitivity warning, the warehouse management takes preventative measures such as feed isolation and environmental disinfection, effectively reducing the spread of highly pathogenic bacteria within the storage unit.

[0160] Step S8: Based on the consistency verification between the pollution risk warning signal and the actual detection results, the weight coefficients in the weighted summation operation are corrected online to generate updated weight coefficients. Specifically, this includes: S8.1: Obtain the pollution risk warning signals triggered three times consecutively and the actual microbial detection results within the corresponding time window. Perform a logical consistency comparison between the pollution risk warning signals and the actual microbial detection results to generate a verification status sequence containing false positive event markers.

[0161] S8.2: Based on the false positive event markers in the verification state sequence, extract the corresponding historical silkworm stress response intensity change rate data, and use the sliding window statistical algorithm to perform distribution characteristic analysis on the historical silkworm stress response intensity change rate data to generate a statistical quantile index characterizing the current threshold sensitivity.

[0162] S8.3: Based on the degree of deviation between the statistical quantile index and the preset confidence interval, construct a weight compensation factor calculation model, and perform gradient descent iterative operation on the weight compensation factor calculation model to generate an initial correction increment value for the rate of change compensation factor.

[0163] S8.4: The original weight coefficient matrix in the weighted summation operation is updated linearly by using the initial correction increment value, and an incremental training set is constructed by combining the manually labeled false positive event sample data to generate updated weight coefficients with adaptive evolution capabilities.

[0164] S8.5: Based on the updated weight coefficients replacing the original parameters in the dynamic threshold generation formula, re-execute the fusion calculation of the microbial growth potential index and the stress response intensity of the silkworm to generate a calibrated new dynamic threshold and complete the closed-loop feedback adjustment.

[0165] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0166] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0167] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for early warning analysis of microbial contamination risks in stored silkworm feed, specifically including: S1: Acquire millimeter-wave radar signals and environmental parameters at multiple different locations within the storage unit, and uniformly define the collected results as raw monitoring data; S2: Perform cross-modal time-series alignment and noise filtering on the original monitoring data to eliminate signal interference and obtain preprocessed monitoring data; S3: Based on the preprocessed monitoring data, extract the feeding rhythm decay rate and the low-frequency amplitude of the body surface microtemperature, and define the extraction results as silkworm biological characteristic variables; S4: Based on the biological characteristic variables of the silkworm, input the lightweight graph attention network to perform mapping calculations and obtain the stress response intensity of the silkworm; S5: Based on the timestamp of the stress response intensity of the silkworm, synchronize the environmental parameters, use the LSTM-GCN hybrid model to infer the microbial growth status, and generate the microbial growth potential index; S6: Perform a weighted summation operation based on the microbial growth potential index and the stress response intensity of the silkworm, and generate a dynamic threshold by combining the rate of change compensation factor; S7: Compare the real-time risk value with the dynamic threshold and if the value exceeds the threshold, generate a pollution risk warning signal. S8: Based on the consistency verification between the pollution risk warning signal and the actual detection results, the weight coefficients in the weighted summation operation are corrected online to generate updated weight coefficients.

2. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 1, characterized in that, Step S1 specifically includes: S1.1: Deploy millimeter-wave radar sensor arrays and environmental parameter sensing nodes at multiple preset spatial coordinate points within the storage unit. Generate radar transmission control commands and environmental sampling trigger signals with unified timestamps based on the pulse repetition frequency synchronization mechanism to establish a spatiotemporal reference framework for multi-source data acquisition. S1.2: Based on the radar transmission control command, drive the millimeter-wave radar sensor array to transmit frequency-modulated continuous wave signals to the silkworm activity area, and receive the millimeter-wave radar echo signals formed by reflection from the surface of the silkworm body, so as to obtain the original radar data stream containing the micro-motion Doppler frequency shift information of the silkworm group. S1.3: Based on the environmental sampling trigger signal, drive the environmental parameter sensing node to perform multi-point parallel sampling operation, and capture temperature field distribution data, humidity field distribution data, carbon dioxide concentration data and ammonia concentration data in the storage space in real time to generate a multi-dimensional environmental parameter dataset. S1.4: Perform frame-level encapsulation processing based on a unified timestamp on the original radar data stream and the multidimensional environmental parameter dataset, and use data packaging to integrate heterogeneous signal streams into data packets with a fixed frame structure to form a standardized original monitoring data sequence. S1.5: Based on the standardized raw monitoring data sequence, perform integrity verification and outlier marking processing, and use sliding window statistical verification to remove invalid data fragments caused by instantaneous sensor failures. Finally, output the confirmed usable raw monitoring data for subsequent preprocessing modules to call.

3. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 1, characterized in that, Step S2 specifically includes: S2.1: Perform linear interpolation resampling on the millimeter-wave radar point cloud sequence and environmental sensor timestamps in the original monitoring data to unify the sampling frequency of multi-source heterogeneous data and generate a synchronized monitoring data stream with a unified time reference. S2.2: Calculate the peak position of the cross-correlation function based on the synchronized monitoring data stream to quantify the transmission delay difference between the millimeter-wave radar signal and the environmental parameter signal, and generate a time-series offset correction vector containing precise time delay compensation. S2.3: Use the time offset correction vector to perform a sliding window translation operation on the synchronized monitoring data stream to eliminate the phase asynchrony between multimodal sensors and generate a cross-modal aligned dataset with strict time axis alignment. S2.4: Apply an adaptive wavelet threshold denoising algorithm to the cross-modal aligned dataset to separate and filter out high-frequency electromagnetic interference and mechanical vibration noise in the storage environment, and generate a purified monitoring signal matrix with significantly improved signal-to-noise ratio. S2.5: Based on the purified monitoring signal matrix, perform outlier removal and linear data filling to repair data breaks caused by sensor momentary failures, and finally output complete and continuous preprocessed monitoring data that meets the requirements of analytical accuracy.

4. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 1, characterized in that, Step S3 specifically includes: S3.1: Perform short-time Fourier transform processing on the millimeter-wave radar echo signal sequence in the preprocessed monitoring data to obtain a time-frequency energy distribution map containing Doppler frequency shift information of silkworm colony micro-movements, as the basic data source for identifying the frequency changes of feeding actions; S3.2: Based on the time-frequency energy distribution map, execute the peak detection and trajectory association algorithm to extract the effective feeding action count sequence within a unit time window, and calculate the slope change value of the count sequence over time through linear regression fitting, thereby generating the feeding rhythm decay rate that characterizes the declining trend of the silkworm population's feeding activity; S3.3: Perform bandpass filtering on the infrared thermal imaging temperature field data stream in the preprocessed monitoring data, and use a Butterworth filter with a cutoff frequency of 0.01 Hz to 0.1 Hz to separate high-frequency environmental noise and low-frequency physiological fluctuations to obtain a pure low-frequency fluctuation signal sequence of the silkworm body surface micro-temperature. S3.4: Based on the low-frequency fluctuation signal sequence of the silkworm body surface micro-temperature, perform Hilbert-Huang transform decomposition processing to extract the instantaneous amplitude envelope of the intrinsic mode function component corresponding to the stress response frequency band, and calculate the root mean square statistic of the envelope within the sliding window, thereby generating the low-frequency amplitude of the silkworm body surface micro-temperature to characterize the degree of metabolic disorder. S3.5: Perform spatiotemporal coordinate binding and format encapsulation processing on the feeding rhythm decay rate and the low-frequency amplitude of the body surface micro-temperature. Use a unified timestamp to integrate the two types of heterogeneous scalar and vector data into a structured data object. Finally, output a standard dataset defined as the biological characteristic variables of the silkworm for subsequent graph attention network calls.

5. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 1, characterized in that, Step S4 specifically includes: S4.1: Based on the feeding rhythm decay rate and body surface micro-temperature low-frequency amplitude in the biological characteristic variables of the silkworm, construct heterogeneous graph data containing spatial location information and time series characteristics, define each millimeter-wave radar sensor node in the storage unit as a graph node, define the spatial adjacency relationship and signal correlation between nodes as graph edges, and generate an initial topological graph structure to characterize the distribution of micro-ecological disturbances in the silkworm colony. S4.2: Perform node feature embedding processing on the initial topology graph structure, and use a linear transformation matrix to map the scalar value of the feeding rhythm decay rate and the low-frequency amplitude spectrum vector of the body surface microtemperature to a high-dimensional latent space to generate node feature embedding vectors carrying multi-dimensional biological behavioral semantic information, so as to eliminate the interference of different physical dimensions on the subsequent attention mechanism calculation. S4.3: Based on the node feature embedding vector, perform multi-head self-attention mechanism calculation, quantify the association weight of biological behavior features between adjacent nodes by performing dot product operation of query vector, key vector and value vector, and generate dynamic attention coefficient matrix reflecting the abnormal propagation path of local silkworm population, so as to identify the group stress response pattern caused by microbial contamination. S4.4: The dynamic attention coefficient matrix is ​​used to perform a weighted aggregation operation on the feature embedding vectors of neighboring nodes. Combined with the residual connection mechanism, the original features of the current node and the neighborhood context information are fused to generate an updated node hidden state representation containing global spatial dependencies, so as to enhance the model's ability to capture weak contamination signals. S4.5: Based on the hidden state representation of the updated node, perform fully connected layer mapping and normalization processing to compress the high-dimensional hidden state into a single-dimensional scalar output, and generate the final evaluation value characterizing the stress response intensity of the overall silkworm body of the storage unit at the current moment, which serves as the biological feedback correction factor in the subsequent dynamic threshold adjustment mechanism.

6. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 5, characterized in that, A coordinate mapping algorithm is used to accurately locate the sensor deployment in three-dimensional space. A pulse repetition frequency synchronization mechanism is used with a PRF frequency of 10Hz. The millimeter-wave radar signal operates in the 76GHz-81GHz frequency band with a modulation period of 100ms. Fast Fourier transform and Doppler frequency shift extraction are used to achieve high-precision analysis of the micro-movements of the silkworm. The environmental parameter sampling period is 2 seconds, and the number of parallel threads is equal to the number of sensor nodes.

7. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 3, characterized in that, Cross-correlation function analysis was performed with a maximum lag range of ±120 seconds and a cross-correlation peak normalization interval of [-1,1]. Sensing signal delay compensation was performed based on the peak position. Short-time Fourier transform was used with a Hanning window type, a window length of 256 points, a frame shift of 128 points, and 512 FFT points.

8. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 6, characterized in that, The millimeter-wave radar nodes are bound to biometric variables through spatial mapping and time series fusion. The Pearson correlation coefficient threshold of 0.65 is used to determine the weighted edges for the correlation of node signals. The environmental state graph structure is established by establishing node adjacency with a Euclidean distance threshold of 2 meters. The edge weights are taken as the weighted average of temperature and humidity, carbon dioxide and ammonia concentration differences, and a multi-head attention mechanism is adopted.

9. The method for early warning analysis of microbial contamination risk in stored silkworm feed according to claim 1, characterized in that, The baseline risk value is calculated by weighting the current microbial growth potential index and the stress response intensity of the silkworm body using online reinforcement learning weights α and β. The interpretability analysis uses SHAP with a sample size of 1000 and L2 norm normalization. The upper and lower limits of the dynamic threshold boundary constraint interval are dynamically adjusted by multivariate statistical analysis of historical storage data according to time window and environmental gain coefficient, and the final dynamic threshold is output.

Images