Method and system for early diagnosis of agricultural pests and diseases based on multi-modal bioelectric signals

Abstract

This invention relates to the field of smart agriculture and plant phenotypic monitoring technology, specifically to a method and system for early diagnosis of agricultural pests and diseases based on multimodal bioelectrical signals. A high-sensitivity sensor array deployed at designated parts of the crop collects multimodal bioelectrical signals, including surface potential, transmembrane current, and damage-induced local potential. Microenvironmental parameters are collected using near-ground environmental sensors. A dedicated processing chip extracts time-domain and frequency-domain bioelectrical features from the bioelectrical signals in real time. The bioelectrical features and microenvironmental parameters are input into a multimodal fusion model for joint analysis. This model uses a spatiotemporal attention and adaptive gating fusion mechanism to distinguish between pest and disease stress and environmental interference. Based on the analysis results and a confidence assessment, early warning information is generated. This invention overcomes the limitations of traditional morphological detection, achieving ultra-early and highly reliable early warning by analyzing the crop's own "electrophysiological distress signals" before lesions are visible to the naked eye.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and plant phenotyping technology, specifically to a method and system for early diagnosis of agricultural pests and diseases based on multimodal bioelectric signals. Background Technology

[0002] Early and accurate diagnosis of agricultural pests and diseases is a crucial prerequisite for implementing green pest control, reducing pesticide overuse, and ensuring the safety and yield of agricultural products. Traditional pest and disease diagnosis mainly relies on manual field inspections, judging by observing visible lesions in crop morphology and color (such as lesions, insect holes, wilting, etc.). However, by the time visible symptoms appear, pests and diseases have often entered the middle or late stages, and the window for control has been missed, requiring more investment with limited effectiveness. In recent years, intelligent monitoring methods based on technologies such as remote sensing spectroscopy and machine vision have been developed, but these methods are essentially still indirect observations of crop phenotypes (such as leaf color, texture, and canopy temperature), and remain insensitive to the "latent period" or "physiological pathogenesis period" of early pest and disease infection, failing to issue early warnings when pathogens have established themselves or pests are feeding, before significant morphological changes have occurred.

[0003] Plant electrophysiology research reveals that crops immediately generate a series of electrophysiological responses in the early stages of pest and disease infection, such as fluctuations in cell membrane potential, changes in vascular bundle ion current, and the generation of electrical signals from localized damage. These bioelectrical signals are the most direct and rapid "distress signals" of crops in response to stress, theoretically enabling true ultra-early diagnosis. However, applying bioelectrical signals to real-time field monitoring faces significant challenges: First, plant bioelectrical signals are extremely weak (microvolts to millivolts) and are easily interfered with by environmental electromagnetic noise, the plant's own physiological rhythms, and changes in temperature and humidity, resulting in low signal-to-noise ratios and difficulties in feature extraction. Second, the information from a single electrical signal modality is limited, making it difficult to distinguish between pest and disease types and abiotic stresses. Third, there is a lack of dedicated hardware systems and intelligent analysis methods capable of in-situ, non-destructive, long-term stable acquisition and real-time analysis of multimodal bioelectrical signals and environmental data. Therefore, developing a technology and system that can effectively capture, analyze, and integrate multi-dimensional bioelectrical and environmental information to achieve ultra-early and highly reliable early warning of agricultural pests and diseases has become a critical technological bottleneck that urgently needs to be overcome in the field of smart agriculture.

[0004] Therefore, existing technologies still need further development. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for early diagnosis of agricultural pests and diseases based on multimodal bioelectric signals, so as to solve the problems existing in the prior art.

[0006] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a method for early diagnosis of agricultural pests and diseases based on multimodal bioelectrical signals, comprising: S1. Collect multimodal bioelectrical signals from crops by deploying a high-sensitivity sensor array at designated locations on the crops; S2. Collect microenvironmental parameter data for crop growth through near-ground environmental sensors; S3. Extract time-domain and frequency-domain bioelectrical features related to pest and disease stress from the multimodal bioelectrical signals in real time using a dedicated processing chip; S4. Input the bioelectric characteristics and the microenvironment parameter data into a multimodal fusion model for joint analysis; S5. Based on the results of the joint analysis, generate early warning information for the target pests and diseases.

[0007] Specifically, the multimodal bioelectrical signals include: The sensor array includes: surface potential signals reflecting cell membrane potential and ion channel activity collected by micropotential sensors; transmembrane current signals reflecting ion flow or mass transport within vascular bundles collected by microcurrent sensors; and damage-induced local potential signals related to the activation of damage-related molecular patterns collected by the sensor array at sites of physical damage or chemical stimulation in crops.

[0008] Specifically, the extraction of time-domain and frequency-domain bioelectrical features using a dedicated processing chip includes: Adaptive filtering was applied to the surface potential signal to suppress power frequency and physiological rhythm noise, and its kurtosis, skewness, and power spectral density in a specific low-frequency band were calculated within a time window to characterize cell membrane potential stability and slow wave oscillations. Wavelet packet decomposition was performed on the transmembrane current signal to extract the wavelet coefficient energy entropy corresponding to the cytoplasmic flow velocity fluctuations of phloem sieve molecules, which was used to characterize the degree of interference in assimilate transport. For the damage-induced local potential signal, its rising slope, peak amplitude, and half-life were detected, and the cross-correlation coefficients with typical electrical signal waveforms triggered by known insect feeding or pathogen excitons were calculated.

[0009] Specifically, the bioelectrical characteristics related to pest and disease stress are predefined as specific combinations of characteristic patterns according to different pest and disease types. For piercing-sucking pests, the key features of the characteristic pattern combination include a sharp drop in the energy entropy of the phloem current signal accompanied by an enhancement of potential oscillations in a specific frequency band. For pathogen infection stress, the key features of the characteristic pattern combination include the repetitiveness of damage-induced potentials, low-amplitude excitation, and disorder of the slow wave oscillation rhythm of cell membrane potentials.

[0010] Specifically, the step of inputting bioelectrical characteristics and microenvironmental parameter data into a multimodal fusion model for joint analysis includes: An alignment channel based on time is established to align synchronously acquired bioelectrical characteristic time series with microenvironmental parameter time series. A multimodal fusion model with a spatiotemporal attention mechanism is constructed. This model learns the time series patterns strongly correlated with stress in the bioelectrical characteristic sequence through the first branch network, and analyzes the environmental parameter sequence through the second branch network, and evaluates the coupling effect of temperature, humidity, and light abrupt changes on the bioelectrical background signal. The outputs of the two branches are adaptively weighted and fused through the attention mechanism to dynamically determine whether the observed abnormal bioelectrical patterns are caused by pest and disease stress or by drastic abiotic environmental changes.

[0011] Specifically, the multimodal fusion model incorporates standard stress response maps corresponding to different crop-pest combinations; the joint analysis also includes: The current stress feature vector generated after model fusion is matched with the standard stress response map for similarity. If the matching degree exceeds the first threshold, it is determined that the corresponding early stress of specific pests and diseases has occurred. If the matching degree is lower than the second threshold but the bioelectrical characteristics show significant abnormalities, it is determined to be an unknown or complex stress, and higher sampling frequency monitoring and data uploading are triggered.

[0012] Specifically, the generation of early warning information also includes a confidence assessment of the warning results; the confidence assessment is based on: the number of standard deviations by which the bioelectrical characteristics exceed the baseline of healthy controls, the persistence and repeatability of the characteristic patterns in the continuous monitoring period, and the classification probability value output by the multimodal fusion model; the final warning information is generated and sent only when the confidence assessment result exceeds a preset reliability threshold.

[0013] Specifically, when the method is applied to a specific crop, the deployment location of the sensor array, the combination of bioelectric signal modes collected, and the structural parameters of the multimodal fusion model are all configured individually based on the crop's anatomical structure, vascular system distribution characteristics, and known major pest and disease infection sites.

[0014] Specifically, the method also includes an incremental model learning step: The multimodal data sequences corresponding to subsequent confirmed cases of pest and disease occurrence in the field, as well as the false alarm case data caused by environmental interference, are combined to form an incremental dataset. The parameters of the multimodal fusion model are fine-tuned using the incremental dataset on a dedicated processing chip or in the cloud to optimize the model's adaptability to specific farmland environments and its ability to identify new pest and disease stress patterns.

[0015] According to a second aspect of the present invention, an early diagnosis system for agricultural pests and diseases based on multimodal bioelectric signals is provided, comprising: The sensor array module, including a high-sensitivity micropotential sensor and a microcurrent sensor, is used for in-situ non-destructive acquisition of multimodal bioelectrical signals from crops. An environmental sensor module is used to collect microenvironmental parameter data of the crop root zone or canopy; a dedicated processing chip module with a built-in hardware accelerator is used to extract time-domain and frequency-domain bioelectrical features from the bioelectrical signals in real time. A multimodal fusion model module is used to receive the bioelectrical characteristics and the microenvironmental parameter data, and perform joint analysis to identify early stress patterns of pests and diseases; The early warning module is used to generate and send early warning information based on the output of the multimodal fusion model module.

[0016] Beneficial effects: The method and system for early diagnosis of agricultural pests and diseases based on multimodal bioelectric signals provided by this invention have the following significant advantages compared with existing technologies: First, this invention achieves, for the first time, effective detection and early warning of crop pests and diseases during their "latent period" or early stage of infection. By directly capturing the characteristic bioelectrophysiological fluctuations caused by pests and diseases through a high-sensitivity sensor array, the warning time point is significantly advanced from the traditional morphological lesion stage to the physiological lesion stage, gaining a crucial time window for prevention and control decisions. This is expected to achieve "early treatment and small-scale treatment," significantly reducing prevention and control costs and environmental pressure.

[0017] Secondly, this invention significantly improves the accuracy and reliability of early diagnosis through multimodal information fusion and intelligent analysis. The method not only collects multiple bioelectrical signals but also simultaneously acquires environmental parameters, and utilizes a lightweight neural network with attention mechanisms and adaptive gating for joint analysis. This design enables the system to intelligently distinguish between specific bioelectrical patterns caused by pests and diseases and non-specific electrical signal fluctuations caused by drastic environmental changes (such as strong light or sudden temperature changes), thereby effectively avoiding false alarms caused by the susceptibility of single bioelectrical signals to environmental interference, resulting in more convincing diagnostic results.

[0018] Furthermore, this invention enables embedded real-time processing from complex signals to precise decision-making. A dedicated processing chip with a built-in hardware accelerator can perform real-time noise reduction and feature extraction on multiple weak bioelectrical signals, and deploy a lightweight multimodal fusion model on edge devices. This edge intelligence solution reduces reliance on continuous network bandwidth and cloud computing power, ensures data privacy, and enables real-time, online, and automated diagnosis in farmland, with fast system response and high practicality.

[0019] Finally, this invention possesses excellent scalability and adaptive evolution capabilities. By introducing a matching mechanism based on standard stress response maps and incremental learning functions, the system can not only identify known pests and diseases but also remain alert to unknown or complex stresses. Furthermore, it can continuously learn from localized cases to optimize model performance, enabling it to adapt to the challenges of different crops, different regions, and new pests and diseases. The system's intelligence level and long-term application value are continuously improved. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the method for early diagnosis of agricultural pests and diseases based on multimodal bioelectric signals provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides a method for early diagnosis of agricultural pests and diseases based on multimodal bioelectric signals, comprising: S1. Collect multimodal bioelectrical signals from crops by deploying a high-sensitivity sensor array at designated locations on the crops.

[0024] It should be further noted that the implementation of this method begins with the deployment of the sensor network. Representative crop plants are selected as monitoring points in the target farmland. The sensor array needs to form a stable, low-impedance electrical contact with the plant tissue. For this purpose, specially designed flexible silver / silver chloride (Ag / AgCl) electrodes can be used. The electrode surface is coated with a plant-compatible hydrogel electrolyte and fixed to the midrib, petiole, or young stem epidermis on the underside of the leaf using flexible straps or minimally invasive clamping methods. The sensor array contains multiple sensing units, spatially distributed to cover the most likely primary infection sites. For example, for aphids, the focus is on monitoring the underside of young leaves on new shoots; for soil-borne diseases, additional monitoring points are needed at the base of the stem.

[0025] S2. Collect microenvironmental parameter data for crop growth through near-ground environmental sensors.

[0026] It should be further explained that the near-ground environmental sensor nodes are usually arranged inside the crop canopy, about 20-50 cm away from the plants, to continuously collect air temperature, relative humidity, photosynthetically active radiation (PAR), and soil volumetric water content.

[0027] S3. Using a dedicated processing chip, extract in real time the time-domain and frequency-domain bioelectrical features related to pest and disease stress from the multimodal bioelectrical signals.

[0028] It should be further noted that all sensor data is synchronously transmitted to the local aggregation node via wired (e.g., RS-485) or wireless (e.g., LoRa) methods. This node has a built-in dedicated processing chip. This chip can be a system-on-a-chip (SoC) integrating programmable logic units (e.g., FPGA) and a microprocessor (e.g., ARM Cortex-M series). The real-time feature extraction process is triggered once within each fixed time window (e.g., 300 seconds).

[0029] S4. Input the bioelectric characteristics and the microenvironment parameter data into the multimodal fusion model for joint analysis.

[0030] It should be further explained that the extracted bioelectric feature vector, together with the corresponding time-aligned environmental parameter vector (averaged over a 300-second window), forms a multimodal sample, which is then input into a pre-trained multimodal fusion model deployed in the chip's memory. The model outputs a vector containing the probabilities of various pests and diseases (including the "healthy" category).

[0031] S5. Based on the results of the joint analysis, generate early warning information for the target pests and diseases.

[0032] It should be further explained that the early warning decision module determines whether to generate an early warning based on the probability vector and subsequent confidence assessment rules. Early warning information can be transmitted via local audible and visual alarms, or uploaded to the cloud platform via a 4G / NB-IoT module and pushed to the user terminal.

[0033] Understandably, this method constructs a complete technological closed loop, from physical signal sensing to embedded intelligent analysis and decision services. It enables direct, in-situ, and continuous monitoring of the crop's physiological state, and completes core analysis at the data source through edge computing, reducing reliance on continuous network connections and protecting data privacy. The core of this method's highly reliable, ultra-early warning system lies in fusing and analyzing the difficult-to-forge bioelectrical "stress fingerprint" with easily interfered environmental signals.

[0034] Specifically, the multimodal bioelectric signals include: surface potential signals reflecting cell membrane potential and ion channel activity collected by micropotential sensors in the sensing array; transmembrane current signals reflecting ion flow or mass transport within vascular bundles collected by microcurrent sensors in the sensing array; and damage-induced local potential signals related to the activation of damage-related molecular patterns collected by the sensing array at sites of physical damage or chemical stimulation in crops.

[0035] It should be further explained that the acquisition of the three types of bioelectrical signals has different emphases in terms of hardware circuitry and physiological significance: (1) Surface potential signal: Differential measurement method was used. Two measuring electrodes (E1, E2) were placed along the leaf vein direction with a spacing of 1-3 cm, and a reference electrode (Ref) was placed at the base of the plant or on the stem away from the measuring point. The signal was initially amplified by an instrumentation amplifier (such as AD8421) with an input impedance higher than 1×10¹²Ω and an equivalent input noise voltage lower than 1μVpp (0.1Hz-100Hz). The signal mainly contains ultra-low frequency (<1Hz) variable potential and slow wave oscillation, reflecting the slow fluctuation of membrane potential caused by H+-AT Pase pump activity, potassium ion channel opening and closing, etc., and is a macroscopic electrical manifestation of the overall metabolism and stress response of the crop.

[0036] (2) Transmembrane current signal: A simplified application of the two-electrode voltage clamp principle. A current injection electrode (I_inj) is inserted (or in close contact) at one point in the plant tissue, and a voltage sensing electrode (V_sense) is placed at another point. The distance between the two points is known (e.g., 5 mm). The weak current flowing between them is measured by a high-precision transimpedance amplifier (TIA), with a range of ±1 nA to ±1 μA. This current mainly reflects the current generated by the transport of assimilates within the phloem sieve molecules (assimilate stream) and the current generated by the movement of ions in the xylem transpiration stream. When pests and diseases affect the vascular bundle function, the amplitude and fluctuation pattern of this current will undergo characteristic changes.

[0037] (3) Damage-induced local potential signal: Using a high-gain (×1000 or higher) and high common-mode rejection ratio (>100dB) DC-coupled amplifier, the potential difference between the point of piercing and sucking by the insect's mouthparts or the point of invasion by the pathogen (local electrode L) and the adjacent healthy tissue (reference electrode R_L, approximately 2-5 mm away) is measured. This signal changes rapidly (milliseconds to seconds) and has a small amplitude (microvolts to millivolts), requiring a high sampling rate (at least 500 Hz) to capture. It directly corresponds to the initial events of local cell membrane rupture, ion efflux, and the resulting systemic electrical signaling.

[0038] Understandably, by simultaneously acquiring these three complementary signals at both the spatial scale (from cell membrane to vascular bundle) and temporal dynamics (from millisecond-level damage signals to minute-level slow waves), this method constructs a multidimensional snapshot of the crop's "electrophysiological phenotype." This captures the complete stress chain from local infection points to systemic responses more comprehensively and specifically than a single signal modality, laying a data foundation for the accurate identification of pest and disease types.

[0039] Specifically, the extraction of time-domain and frequency-domain bioelectrical features using a dedicated processing chip includes: adaptively filtering the surface potential signal to suppress power frequency and physiological rhythm noise, and calculating its kurtosis, skewness, and power spectral density in a specific low-frequency band (0.01Hz-1Hz) within a time window, as features characterizing cell membrane potential stability and slow-wave oscillations; performing wavelet packet decomposition on the transmembrane current signal to extract the wavelet coefficient energy entropy corresponding to the frequency band of phloem sieve molecule cytoplasmic flow velocity fluctuations, as features characterizing the degree of interference in assimilate transport; detecting the rise slope, peak amplitude, and half-life of the damage-induced local potential signal, and calculating the cross-correlation coefficient with typical electrical signal waveforms triggered by known insect feeding or pathogen excitons.

[0040] It should be further explained that feature extraction is a core operation on embedded chips, and the specific algorithm implementation and parameter selection are as follows: (1) Surface potential signal processing: First, an adaptive filter based on the Normalized Least Mean Square (NLMS) algorithm is used to eliminate 50Hz power frequency and its harmonic interference. The reference signal is generated by the chip's internal digital phase-locked loop (PLL). After filtering, the kurtosis of the signal data (3000 points) with a sampling rate of 10Hz within a 300-second time window is calculated. and skewness . cliff ,in For signal value, The mean, Points. Skewness. Kurtosis reflects the sharpness of the signal distribution; high kurtosis values ​​may indicate intermittent, large-amplitude membrane potential fluctuations. Skewness reflects the asymmetry of the distribution. Simultaneously, a 1024-point FFT with a Hanning window is performed on the windowed signal to calculate the power spectral density (PSD), and the total power within the 0.01Hz to 1Hz frequency band is integrated. The 0.01Hz-1Hz frequency band was chosen because it encompasses the main physiological rhythms and slow-wave electrical activity of plants, which are closely related to physiological processes such as stomatal movement and photosynthetic product flow. The 300-second window length is a compromise between temporal resolution (diagnosis every 5 minutes) and frequency domain resolution (able to resolve components as low as 0.0033Hz).

[0041] (2) Transmembrane current signal processing: For current signals within the same time window, a 3-level wavelet packet decomposition was performed using the 'db4' wavelet basis to obtain 8 subbands (numbered 0-7) in the 3rd level. According to preliminary experiments, the main fluctuation energy of the phloem current in healthy plants is concentrated in the 0.05Hz-0.5Hz range, corresponding to subbands 2 and 3. The wavelet coefficient energy entropy of these two target subbands was calculated. First, calculate the total energy of the target subband. ,in It is the first The first of the sub-bands Wavelet coefficients, That is, the number of coefficients. Then calculate the total energy of all 8 subbands. Finally, the energy percentage of the target sub-band is calculated. This leads to the energy entropy. When transportation is disrupted, current fluctuations may become more irregular, and the distribution of energy across frequency bands may change, leading to... Change.

[0042] (3) Damage-induced local potential signal processing: First, the moving standard deviation of the signal is calculated. When the absolute value of the signal exceeds three times the standard deviation of the baseline mean, it is identified as a potential event. For each detected event, its starting point, peak point, and decay to half-peak point are precisely determined. Calculation: Rising slope ,in For amplitude, This refers to a specific time point. Peak amplitude. .half life ,in It is the amplitude decay to The time point of the event. Simultaneously, the waveform of this event is compared with several standard waveform templates pre-stored in the chip memory (e.g., "aphid stylet puncture waveform," "fungal elicitor response waveform") to calculate the normalized cross-correlation coefficient. For the templates... and event signals cross-relationship number Take the maximum value As a characteristic. The higher the value, the more similar the current event is to the typical electrical signaling pattern of a specific biological stress.

[0043] It is understandable that the above feature extraction scheme is not a simple application of general signal processing, but rather closely integrated with the intrinsic mechanisms of plant electrophysiology. Kujicic and skewness are effective statistical measures for characterizing membrane potential stability; power in the 0.01-1Hz frequency band is a direct indicator for quantifying the intensity of slow wave oscillations; wavelet packet energy entropy is a sensitive feature for describing the rhythmic disorder of vascular bundle transport function; and waveform cross-correlation is key to identifying the "fingerprint" of specific biological stresses. The combination of these features provides highly discriminative input for subsequent classification models.

[0044] Specifically, the bioelectrical characteristics related to pest and disease stress are predefined as specific combinations of characteristic patterns according to different pest and disease types. For piercing-sucking pests, the key features of the characteristic pattern combination include a sharp drop in the energy entropy of the phloem current signal accompanied by an enhancement of potential oscillations in a specific frequency band. For pathogen infection stress, the key features of the characteristic pattern combination include the repetitiveness of damage-induced potentials, low-amplitude excitation, and disorder of the slow wave oscillation rhythm of cell membrane potentials.

[0045] It should be further explained that during the model training and rule base construction phases, it is necessary to establish multi-dimensional feature space discrimination rules for different pests and diseases based on a large amount of experimental data. This involves not only setting thresholds but also defining the collaborative variation relationships between features.

[0046] (1) For piercing-sucking pests (such as aphids and leafhoppers): When the pest's stylet inserts into the phloem to suck sap, it causes blockage of the sieve plate pores and callose deposition, interfering with the smooth flow of assimilates. Its characteristic pattern is manifested as: the energy entropy of the transmembrane current signal. A significant decrease within a short period (e.g., within 30 minutes), such as a drop of more than 20% from a healthy baseline value (approximately 0.8-0.9), indicates that the current fluctuations have become regular and are no longer uniformly distributed. Simultaneously, the power of the body surface potential signal in the 0.1Hz-0.5Hz frequency band... A relative increase, for example, exceeding twice the standard deviation of the healthy baseline. This is because the insect's feeding stimulates ion flow in local cells, triggering rhythmic oscillations in the membrane potential. The "co-occurrence" of these two characteristics is a key pattern for diagnosing piercing-sucking pests.

[0047] (2) For vivotrophic pathogens (such as powdery mildew and rust): When pathogens infect plants, they continuously release effectors or elicitors, which interact with receptors on the plant cell membrane, triggering multiple, weak defensive responses. The characteristic pattern is the number of times damage-induced local potential signals occur per unit time (e.g., 1 hour). A significant increase, for example, from <5 times / hour in a healthy state to >15 times / hour, but the peak amplitude of individual events... Generally low (e.g., <0.5mV). Furthermore, the slow-wave oscillation rhythm of the body surface potential signal deteriorates, which can be quantified by calculating the approximate entropy of the power spectrum in the 0.01–0.1 Hz frequency band. A decrease indicates a decline in the regularity of the sequence and an increase in complexity. Kujicic It may remain consistently high, reflecting frequent abnormal fluctuations in membrane potential.

[0048] (3) For necrotrophic pathogens or certain bacteria (such as ulcerative bacteria): they often produce toxins or cell wall degrading enzymes, causing extensive cell damage and necrosis. Their characteristic pattern may manifest as: a single amplitude of the damage-induced local potential signal. Larger values ​​(e.g., >2mV), but not high frequency; the baseline of the body surface potential signal drifts slowly; the DC component of the transmembrane current signal may change significantly.

[0049] Understandably, by defining these multidimensional feature pattern combinations closely related to the specific attack mechanisms of pests and diseases, this method can achieve diagnosis with pathophysiological interpretability that goes beyond simple threshold judgment. This pattern recognition approach greatly enhances the model's ability to distinguish between different pests and diseases and reduces misjudgments caused by environmental interference with single features.

[0050] Specifically, the step of inputting bioelectrical characteristics and microenvironmental parameter data into a multimodal fusion model for joint analysis includes: establishing a time-based alignment channel to align the synchronously acquired bioelectrical characteristic time series with the microenvironmental parameter time series; constructing a multimodal fusion model with a spatiotemporal attention mechanism, which learns the time-series patterns strongly correlated with stress in the bioelectrical characteristic sequence through a first branch network, analyzes the environmental parameter sequence through a second branch network, and evaluates the coupling effect of temperature, humidity, and light mutations on the bioelectrical background signal; and adaptively weighting and fusing the outputs of the two branches through the attention mechanism to dynamically determine whether the observed abnormal bioelectrical patterns originate from pest and disease stress or are caused by drastic abiotic environmental changes.

[0051] It should be further noted that the multimodal fusion model is a lightweight neural network that can be deployed on resource-constrained edge devices. Its specific architecture and workflow are as follows: (1) Time alignment and input construction: Bioelectric features are extracted every 300 seconds to form a feature vector. Environmental data (temperature) ,humidity ,illumination Soil moisture The sample is then taken at 1Hz, and the average value is calculated within each corresponding 300-second window to form a vector. The model input is a time series segment, such as data from the last 10 time steps (i.e., the past 50 minutes): The shape is 10× ; It has a shape of 10×4.

[0052] (2) Model structure: a) Bioelectric branch: Input First, a one-dimensional convolutional layer (1D-CNN) is used with 32 convolutional kernels of size 3, a stride of 1, and the ReLU activation function to extract local correlation patterns between features at adjacent time steps. The convolutional output then passes through a max-pooling layer (pooling size 2). The result is then flattened and fed into a bidirectional long short-term memory network (Bi-LSTM) with 16 hidden units. The output of the Bi-LSTM is the concatenation of the forward and backward hidden states at each time step, denoted as . .

[0053] b) Environment Branch: Input By directly inputting a simple fully connected layer and mapping its dimensions to match the hidden states of the bioelectric branch LSTM (e.g., 32-dimensional), and then passing it through a unidirectional LSTM with 8 hidden units, the temporal encoding of environmental features is obtained. .

[0054] c) Spatiotemporal attention mechanism: Applying temporal attention to the bioelectrical branch. Calculating each time step. Attention weights : ,in yes The Middle A vector at each time step. It is the vector of the last time step (representing the current context). , , , These are learnable parameters. The weighted summation yields the weighted bioelectric context vector. The environment branch takes the output of the last time step as the environment context vector. .

[0055] d) Adaptive Gated Fusion: Compute a gated scalar ,in , These are learnable parameters. It's the sigmoid function. The final fused vector. When the environment changes drastically, the model tends to... The smaller the value, the more environmental context information is considered to "explain" the anomalies in bioelectricity.

[0056] e) Classification layer: Fusion vector The system outputs the probability of belonging to each category (healthy, pest A, pest B, etc.) through a fully connected layer and a softmax layer.

[0057] Understandably, this fusion model, through a spatiotemporal attention mechanism, can focus on the time points of bioelectrical characteristics most relevant to stress, thus improving the ability to identify temporal patterns. More importantly, through adaptive gating fusion, the model can dynamically assess the correlation between bioelectrical anomalies and environmental disturbances. For example, when strong sunlight and high temperatures at noon cause stomatal closure, the plant surface potential may also fluctuate. In this case, the output of the environmental branch will indicate the "strong sunlight and high temperature" event, and the gating... The value may decrease, thereby suppressing the overreaction to bioelectrical fluctuations during that period and effectively reducing false positive alarms caused by abiotic stress.

[0058] Specifically, the multimodal fusion model has a built-in standard stress response map corresponding to different crop-pest combinations; the joint analysis also includes: matching the current stress feature vector generated after model fusion with the standard stress response map for similarity; if the matching degree exceeds the first threshold, it is determined that the corresponding early stress of a specific pest has occurred; if the matching degree is lower than the second threshold but the bioelectrical characteristics show significant abnormalities, it is determined to be an unknown or complex stress, and higher sampling frequency monitoring and data uploading are triggered.

[0059] It should be further noted that the standard stress response map is the model's knowledge base, and its construction and usage are as follows: (1) Atlas Construction (Offline Training Phase): A multimodal fusion model is trained using a large number of labeled samples containing various health and pest states. After training, the final softmax classification layer is removed, and the output of the last fully connected layer (i.e., the fusion vector) is used. This is used as a "deep feature". For each specific pest or disease state in the training set (e.g., "tomato-whitefly early stage", "grape-downy mildew late stage"), the mean vector of the deep features of all samples in that class is calculated. This mean vector is the anchor point of the "standard stress response profile" for this category. Anchor points for the health category are also calculated. All anchor points form a graph matrix. .

[0060] (2) Online similarity matching and decision-making (online inference stage): For the data in the current time window, the model is forward propagated to the fusion layer to obtain the current deep feature vector. .calculate With atlas Each anchor point cosine similarity Set two thresholds: a high-confidence matching threshold. Low confidence threshold .

[0061] Furthermore, the method also includes: a) If a certain pest and disease anchor point exists , making ,and If it is clearly closer to a pest or disease than a healthy one, it is determined to be an early stress of a specific pest or disease with a high confidence level.

[0062] b) If Maximum similarity with all known pest and disease anchors However, its similarity to health anchors is... And the original softmax output of the model is the probability of the "healthy" category. If the crop is found to be in an "abnormal" state but cannot be classified as a known pest or disease, the system will trigger "enhanced monitoring mode".

[0063] c) Enhanced monitoring mode: The system temporarily increases the sampling rate of bioelectric signals from 10Hz to 100Hz for 10 minutes to capture richer transient details; at the same time, the data (including the original high-sampling-rate signal segments and extracted fine features) is packaged and uploaded to the cloud server via wireless network for expert systems or more complex cloud models to conduct in-depth analysis, and prompts the user to pay attention.

[0064] Understandably, standard graph matching mechanisms provide an interpretable, distance-based decision-making approach that is more robust than simply relying on neural network output probabilities. Dual threshold strategy ( , The system includes a "confidence zone" and an "uncertain zone." The high threshold ensures high accuracy in early warning of known pests and diseases; the combination of the low threshold and enhanced monitoring mode enables the system to detect new pests or mixed infections, thus achieving system scalability.

[0065] Specifically, the generation of early warning information also includes a confidence assessment of the warning results; the confidence assessment is based on: the number of standard deviations by which the bioelectrical characteristics exceed the baseline of healthy controls, the persistence and repeatability of the characteristic patterns in the continuous monitoring period, and the classification probability value output by the multimodal fusion model; the final warning information is generated and sent only when the confidence assessment result exceeds a preset reliability threshold.

[0066] It should be further explained that confidence assessment is a post-decision processing step that integrates multiple indicators, aiming to ensure the reliability of early warning information. Its specific calculation process is as follows: (1) Calculate the three confidence factors: a) Characteristic Deviation Factor Select the three most critical bioelectrical characteristics for the currently identified pest or disease type (e.g., for piercing-sucking pests, select...). , , For each key feature Calculate its current value Compared to the long-term health baseline of this monitoring point (the mean of health data over the past 7 days) and standard deviation Z-score of ) .Pick The average value and normalized: .so The range is limited to [0,1], with larger values ​​indicating greater deviation from normal. The baseline of the past 7 days is used to accommodate the slow physiological changes at different growth stages of the crop.

[0067] b) Pattern persistence factor : Check the feature pattern of the current determination (e.g.) Decline and Does the increase occur within a past consecutive time window? Let the current time be the [number]th [time window]. A window, check window , , If at least two of the three windows detect the same feature pattern (determined by the trend of feature values), then If only the current window is open If detected, then Otherwise, it is the intermediate value. A high value indicates that the anomalous pattern is persistent and not transient noise.

[0068] c) Model probability factor : That is, the probability output by the softmax layer of the multimodal fusion model corresponding to the currently determined pest or disease category. . .

[0069] (2) Calculation of overall confidence level: final confidence score It is a weighted sum of three factors: Weight , , It can be adjusted according to the identification characteristics of different pests and diseases, and can usually be set to To emphasize the persistence of the pattern, a reliability threshold is set. Only when Only when the specific pest or disease type, location, confidence score, and time of occurrence are confirmed will an early warning message be generated and sent via the communication module. Otherwise, the determination will only be stored as a "suspicious record" in the local log and will not trigger an active early warning.

[0070] Understandably, this multi-factor confidence assessment mechanism cross-validates the early warning results from three dimensions: statistical significance, temporal continuity, and internal model determinism. It effectively filters out single abnormal signals caused by transient strong electromagnetic interference, brief sensor malfunctions, or occasional non-biological stresses, ensuring that the issued warnings are "well-considered." This significantly improves the positive report rate of the early warning system, reduces unnecessary effort for users, and enhances the system's credibility.

[0071] Specifically, when the method is applied to a specific crop, the deployment location of the sensor array, the combination of bioelectric signal modes collected, and the structural parameters of the multimodal fusion model are all configured individually based on the crop's anatomical structure, vascular system distribution characteristics, and known major pest and disease infection sites.

[0072] It should be further noted that the successful application of this method relies on deep customization for the target crop, which involves interdisciplinary knowledge of agronomy, plant physiology, and plant pathology. (1) Personalization of sensor array deployment location: a) For Solanaceae crops such as tomatoes and peppers: The key monitoring area is the midrib on the underside of fully expanded leaves at the top of the plant, especially near the tip third of the leaf, which is the preferred feeding and egg-laying area for whiteflies and aphids. For foliar diseases such as early blight and leaf mold, additional monitoring points should be added on the underside of older leaves in the middle and lower parts of the plant.

[0073] b) For fruit trees such as grapes and apples: In addition to the midrib of the leaves, sensors should also be deployed on the tender green bark of the fruit cluster stalks and new shoots. For grape downy mildew, sensors should be preferentially deployed near the leaf margins where water easily accumulates. For branch and trunk diseases (such as canker), sensors should be deployed on the sun-facing bark of the main trunk or main branches.

[0074] c) For gramineous crops such as wheat and rice: their vascular bundles are scattered, and leaf veins are parallel. Sensors should preferably use strip electrodes, laterally wrapped around the base of the stem or the lower part of the leaf sheath, to measure comprehensive annular current signals. For rice blast, monitoring points should be added at the leaf sheath of the flag leaf.

[0075] (2) Personalized signal mode combination: a) For monitoring of vascular diseases (such as banana wilt and cucumber wilt), transmembrane current signal should be the primary monitoring mode, with surface potential signal as a secondary mode, and the weight of damage-induced signal can be reduced.

[0076] b) For monitoring of piercing-sucking pests (such as aphids and thrips) and viral diseases (transmitted by them), it is necessary to give equal importance to surface electrical potential signals (monitoring systemic electrical signals caused by feeding) and damage-induced signals (monitoring local events of stylet piercing).

[0077] c) In terms of hardware, this can be achieved by configuring an array of sensor probes with different numbers of potential sensors and current sensors.

[0078] (3) Model parameter personalization: The selection and weights of model input features need to be adjusted. For example, for fruit trees, whose physiological response is relatively slow, the time window can be extended from 300 seconds to 600 seconds, and the LSTM network may need more hidden units to capture long-term dependencies. During model training, it is necessary to use a dataset of the target crop in the target growth environment (such as greenhouse or open field) for training and fine-tuning so that the model can learn the crop's unique bioelectrical background pattern and its coupling relationship with environmental factors.

[0079] Understandably, this personalized configuration based on crop type and major pests and diseases allows the general technical framework to be precisely adapted to diverse agricultural production scenarios. It ensures targeted monitoring and accurate diagnosis, avoiding the performance loss that occurs when models trained for broadleaf crops are directly applied to gramineous crops. This is a crucial step in enabling this method to move from the laboratory to practical field applications.

[0080] Specifically, the method further includes an incremental learning step: the multimodal data sequence corresponding to subsequent confirmed cases of pest and disease occurrence in the field, as well as the false alarm case data caused by environmental interference, are combined to form an incremental dataset; the multimodal fusion model is fine-tuned using the incremental dataset on a dedicated processing chip or in the cloud to optimize the model's adaptability to specific farmland environments and its ability to identify new pest and disease stress patterns.

[0081] It should be further explained that incremental learning enables the system to continuously evolve during use, and the specific implementation process is as follows: (1) Incremental data collection and labeling: The system locally stores all early warning events, high-confidence abnormal events, and randomly sampled normal events within a recent period (e.g., 30 days) and their corresponding raw data snapshots (including raw bioelectric signals, environmental data, extracted features, model inference results, and confidence levels for 10 minutes before and after the event). Users or field inspectors verify and provide feedback on system early warnings through a mobile app or cloud platform, labeling the actual situation as: "Confirmed - Pest A", "Confirmed - Pest B", "Confirmed - New Symptoms", "False Alarm - Environmental Interference (e.g., fertilization)", "False Alarm - Mechanical Damage", "No Error - System Misjudgment", etc. These labeled information are associated with the corresponding data snapshots to form an incremental learning sample.

[0082] (2) Incremental dataset construction: Regularly (e.g., weekly), synchronize the collected validated samples from field nodes to the regional edge server or cloud. Construct the incremental dataset. Each sample contains input data. and real labels marked by humans Special attention should be paid to the data on false positive cases, including their labels. It is set to "healthy" or a specific "environmental disturbance" category.

[0083] (3) Model fine-tuning: To avoid catastrophic forgetting (i.e., forgetting old knowledge when learning new knowledge), the Elastic Weight Consolidation (EWC) algorithm is used for fine-tuning. First, the parameters of the original pre-trained model are saved in the cloud. And calculate the Fisher information matrix. The diagonal approximation is used to evaluate each parameter. Importance to the existing task (i.e., existing classification performance). The loss function can be calculated on the original training set. The expectation of the second derivative with respect to the parameters is used for estimation. Then, on the incremental dataset... Next, optimize the new loss function: The first term is the standard cross-entropy loss, which adapts the model to new data; the second term is the EWC regularization term, which penalizes the loss on important old parameters. Significant revisions This is a hyperparameter that balances the importance of new and old knowledge, typically set to 100-1000. It is optimized using gradient descent. The new parameters after fine-tuning are obtained. .

[0084] (4) Model update and deployment: The fine-tuned model Test on the retained old task validation set and incremental dataset to ensure no performance degradation. After passing the tests, distribute the new model parameters to the corresponding field monitoring nodes to complete the model update.

[0085] Understandably, the incremental learning mechanism transforms this system from a static, pre-trained diagnostic tool into an adaptive and evolving agent capable of continuously learning from the local farmland environment. It can not only better adapt to local environmental disturbances by correcting errors (false alarms), but also expand its recognition capabilities by absorbing new pest and disease cases, and even gradually build a unique pest and disease early warning model for that specific plot of land, thus achieving the goal of becoming increasingly accurate with use.

[0086] It should be further explained that, to more clearly illustrate the implementation process of this invention, a specific operational example is used below. This example aims to monitor whether greenhouse tomato plants have suffered early aphid infestation.

[0087] Suppose that within a time window (the t-th window, with a duration of 300 seconds), the system collects and processes the following data: 1. Data Acquisition: The sensor array is fixed to the midrib on the underside of the upper leaves of the tomato plant. An environmental sensor records the average temperature within this window. average humidity Average light Soil moisture After sampling and adaptive filtering, the bioelectrical signals yield a sequence of surface potential signals that can be used for feature extraction. Transmembrane current signal sequence And a damage-induced local potential signal event was detected. .

[0088] 2. Feature extraction calculation process: (1) Characteristics of body surface potential: for Calculate the time-domain statistical characteristics (3000 points, sampling rate 10Hz). The mean of the signal within this window is calculated. Standard deviation . cliff ,in This refers to a single voltage value within a signal sequence. Skewness .right Perform an FFT and calculate the power spectral density, then integrate to obtain the power in the 0.01Hz-1Hz frequency band. .

[0089] (2) Transmembrane current characteristics: for Perform a 3-level 'db4' wavelet packet decomposition. Calculate the total energy of the wavelet coefficients in subbands 2 and 3 (corresponding to ~0.05Hz-0.5Hz). Total energy of all 8 subbands Calculate the energy percentage. Then the energy entropy The plant's baseline health over the past 7 days... mean Standard deviation Therefore, currently The Z-score is .

[0090] (3) Damage-induced potential characteristics: for events The slope of its rising edge was detected. Peak amplitude ,half life Calculate the maximum cross-correlation coefficient between it and the pre-stored "aphid stylet puncture" standard template. .

[0091] 3. Multimodal fusion analysis and decision-making: The above features and environmental parameters are combined to construct the input vector for the current time step. and It, together with the historical data from the previous 9 time steps, forms a sequence. and The pre-trained multimodal fusion model is input. Model forward propagation: the bioelectric branch uses 1D-CNN and Bi-LSTM to extract temporal patterns and calculates weights through a temporal attention mechanism. The environment branch processes environmental data. Assuming the current environment is stable, the gating unit calculates larger fusion weights. This indicates that the model relies more heavily on bioelectrical information. (Fusing vectors) The softmax probability output is obtained through the classification layer: , , , Next, standard map matching is performed. The current fusion features are then calculated. Cosine similarity to the standard anchor point of "aphid" Similarity to the "health" anchor .because and The map matching was successful.

[0092] 4. Confidence Assessment and Early Warning Generation: A confidence assessment was conducted on the "aphid" identification. a) Characteristic Deviation Factor Selecting key features , , . The Z-score has been calculated to be 5.0. Its Z-score is 2.0. The Z-score is 3.0 (calculated based on its healthy baseline). Average Normalization .

[0093] b) Pattern persistence factor The inspection revealed that within the three consecutive windows t-2, t-1, and t, Continued decline and The pattern of continuous increase exists, therefore .

[0094] c) Model probability factor Using weights Calculate the overall confidence level .because If the map match is successful, the system determines that the generated warning is valid.

[0095] Understandably, this case study fully demonstrates the entire automated process from the initial signal to the issuance of an early warning. The calculation process shows that... It plummeted from 0.78 to 0.53 (Z-score=5.0), while rise, The probability of 0.88 perfectly matches the characteristic pattern combination of "piercing-sucking pests." The multimodal fusion model yielded a high-probability (0.82) classification result for aphids, and was highly similar to the standard atlas (0.89). The confidence level (0.946) was extremely high, stemming from significant deviations in features, persistent pattern occurrence, and high model confidence. This case demonstrates that the method of this invention can make a high-confidence early diagnosis by quantitatively analyzing subtle changes in bioelectrical characteristic patterns before pests cause visible symptoms.

[0096] This invention provides another embodiment, which offers an early diagnosis system for agricultural pests and diseases based on multimodal bioelectric signals. The early diagnosis system for agricultural pests and diseases based on multimodal bioelectric signals includes: (1) The sensor array module includes a high-sensitivity micropotential sensor and a microcurrent sensor, which are used for in-situ non-destructive acquisition of multimodal bioelectric signals of crops.

[0097] It should be further noted that this system is an integrated hardware and software device. The sensor array module specifically consists of a high-impedance probe, a front-end amplification and filtering circuit, and an analog multiplexer. The micro-potential sensor probe uses Ag / AgCl electrodes and is connected via shielded wire to an instrumentation amplifier (such as INA116) with an input impedance >1×10¹²Ω and a bias current <1pA. The micro-current sensor uses a transimpedance amplifier (TIA) built based on a low-noise operational amplifier (such as OPA129), with a feedback resistor selectable from 1MΩ to 1GΩ, for measuring currents in the pA to nA range.

[0098] (2) Environmental sensor module, used to collect microenvironmental parameter data of crop root zone or canopy.

[0099] It should be further noted that the environmental sensor module integrates digital temperature and humidity sensors (such as SHT35), light intensity sensors (such as BH1750), and soil moisture sensors (such as FDR type).

[0100] (3) A dedicated processing chip module with a built-in hardware accelerator is used to extract time-domain and frequency-domain bioelectric features from the bioelectric signal in real time.

[0101] It should be further noted that the dedicated processing chip module can be a mixed-signal MCU (such as TI's Sitara series or ST's STM32H7 series) that integrates a high-performance analog front-end (AFE), a floating-point digital signal processor (DSP), and hardware accelerators (such as convolution accelerators). The hardware accelerator exists in the form of a coprocessor or IP core, specifically optimized for performing FFT, wavelet transform, and matrix multiplication-accumulation operations to meet the computing power requirements for real-time processing of multiple bioelectric signals (such as 8 channels).

[0102] (4) Multimodal fusion model module, used to receive the bioelectric characteristics and the microenvironment parameter data, and perform joint analysis to identify early stress patterns of pests and diseases.

[0103] It should be further explained that the multimodal fusion model module is stored in the chip's flash memory in the form of a quantized and compiled neural network model file, which is executed by the chip's main CPU (such as ARM Cortex-M7) calling the inference engine (such as TensorFlow Lite for Micro controllers).

[0104] (5) Early warning module, used to generate and send early warning information based on the output of the multimodal fusion model module.

[0105] It should be further noted that the warning module includes a local alarm unit (RGB LED indicator and buzzer) and a wireless communication unit (such as a LoRa WAN module or 4GCat.1DTU).

[0106] Furthermore, all modules are powered by lithium batteries and can be charged via solar panels. The device housing meets IP65 protection standards to withstand harsh field environments.

[0107] Understandably, this system solidifies the algorithmic flow of the methodology into mass-producible hardware products. Its core innovation lies in integrating complex multimodal signal real-time processing and lightweight artificial intelligence analysis algorithms into low-power, low-cost embedded devices suitable for long-term fieldwork. Feature extraction is achieved through hardware accelerators, resolving the contradiction between the computationally intensive nature of bioelectrical signal processing and the low-power requirements of the devices. This system can operate independently or be deployed in a network to form an early-stage pest and disease sensing IoT covering farmland, providing a reliable hardware foundation and data source for real-time, intelligent decision-making in precision agriculture.

[0108] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the aforementioned method for early diagnosis of agricultural pests and diseases based on multimodal bioelectric signals. This computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0109] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0110] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0111] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0112] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for early diagnosis of agricultural pests and diseases based on multimodal bioelectrical signals, characterized in that, include: S1. Collect multimodal bioelectrical signals from crops by deploying a high-sensitivity sensor array at designated locations on the crops; S2. Collect microenvironmental parameter data for crop growth through near-ground environmental sensors; S3. Extract time-domain and frequency-domain bioelectrical features related to pest and disease stress from the multimodal bioelectrical signals in real time using a dedicated processing chip; S4. Input the bioelectric characteristics and the microenvironment parameter data into a multimodal fusion model for joint analysis; S5. Based on the results of the joint analysis, generate early warning information for the target pests and diseases.

2. The method for early diagnosis of agricultural pests and diseases according to claim 1, characterized in that, The multimodal bioelectric signals include: The sensor array includes: surface potential signals reflecting cell membrane potential and ion channel activity collected by micropotential sensors; transmembrane current signals reflecting ion flow or mass transport within vascular bundles collected by microcurrent sensors; and damage-induced local potential signals related to the activation of damage-related molecular patterns collected by the sensor array at sites of physical damage or chemical stimulation in crops.

3. The method for early diagnosis of agricultural pests and diseases according to claim 2, characterized in that, The extraction of time-domain and frequency-domain bioelectrical features using a dedicated processing chip specifically includes: Adaptive filtering was applied to the surface potential signal to suppress power frequency and physiological rhythm noise, and its kurtosis, skewness, and power spectral density in a specific low-frequency band were calculated within a time window to characterize cell membrane potential stability and slow wave oscillations. Wavelet packet decomposition was performed on the transmembrane current signal to extract the wavelet coefficient energy entropy corresponding to the cytoplasmic flow velocity fluctuations of phloem sieve molecules, which was used to characterize the degree of interference in assimilate transport. For the damage-induced local potential signal, its rising slope, peak amplitude, and half-life were detected, and the cross-correlation coefficients with typical electrical signal waveforms triggered by known insect feeding or pathogen excitons were calculated.

4. The method for early diagnosis of agricultural pests and diseases according to claim 3, characterized in that, The bioelectrical characteristics associated with pest and disease stress are predefined as specific combinations of characteristic patterns according to different pest and disease types; specifically for stress on piercing-sucking insects. Its characteristic pattern combination focuses on the sharp drop in the energy entropy of the phloem current signal, accompanied by an enhancement of potential oscillations in a specific frequency band; targeting pathogen infection stress, Its characteristic pattern combination includes the repetitiveness of damage-induced potential, low-amplitude excitation, and disorder of slow-wave oscillation rhythm of cell membrane potential.

5. The method for early diagnosis of agricultural pests and diseases according to claim 1, characterized in that, The process of inputting bioelectrical characteristics and microenvironmental parameter data into a multimodal fusion model for joint analysis includes: An alignment channel based on time is established to align synchronously acquired bioelectrical characteristic time series with microenvironmental parameter time series. A multimodal fusion model with a spatiotemporal attention mechanism is constructed. This model learns the time series patterns strongly correlated with stress in the bioelectrical characteristic sequence through the first branch network, and analyzes the environmental parameter sequence through the second branch network, and evaluates the coupling effect of temperature, humidity, and light abrupt changes on the bioelectrical background signal. The outputs of the two branches are adaptively weighted and fused through the attention mechanism to dynamically determine whether the observed abnormal bioelectrical patterns are caused by pest and disease stress or by drastic abiotic environmental changes.

6. The method for early diagnosis of agricultural pests and diseases according to claim 5, characterized in that, The multimodal fusion model incorporates standard stress response maps corresponding to different crop-pest combinations; the joint analysis also includes: The current stress feature vector generated after model fusion is matched with the standard stress response map for similarity. If the matching degree exceeds the first threshold, it is determined that the corresponding early stress of specific pests and diseases has occurred. If the matching degree is lower than the second threshold but the bioelectrical characteristics show significant abnormalities, it is determined to be an unknown or complex stress, and higher sampling frequency monitoring and data uploading are triggered.

7. The method for early diagnosis of agricultural pests and diseases according to claim 1, characterized in that, The generation of early warning information also includes a confidence assessment of the warning results; the confidence assessment is based on: the number of standard deviations by which the bioelectrical characteristics exceed the baseline of healthy controls, the persistence and repeatability of the characteristic patterns in the continuous monitoring period, and the classification probability value output by the multimodal fusion model; The final warning message is generated and sent only when the confidence assessment result exceeds the preset reliability threshold.

8. The method for early diagnosis of agricultural pests and diseases according to claim 1, characterized in that, When the method is applied to a specific crop, the deployment location of the sensor array, the combination of bioelectric signal modes collected, and the structural parameters of the multimodal fusion model are all configured individually based on the crop's anatomical structure, vascular system distribution characteristics, and known major pest and disease infection sites.

9. The method for early diagnosis of agricultural pests and diseases according to any one of claims 1 to 8, characterized in that, The method also includes an incremental model learning step: The multimodal data sequences corresponding to subsequent confirmed cases of pest and disease occurrence in the field, as well as the false alarm case data caused by environmental interference, are combined to form an incremental dataset. The parameters of the multimodal fusion model are fine-tuned using the incremental dataset on a dedicated processing chip or in the cloud to optimize the model's adaptability to specific farmland environments and its ability to identify new pest and disease stress patterns.

10. An early diagnosis system for agricultural pests and diseases based on multimodal bioelectric signals, used to implement the method according to any one of claims 1-9, characterized in that, include: The sensor array module, including a high-sensitivity micropotential sensor and a microcurrent sensor, is used for in-situ non-destructive acquisition of multimodal bioelectrical signals from crops. An environmental sensor module is used to collect microenvironmental parameter data of the crop root zone or canopy; a dedicated processing chip module with a built-in hardware accelerator is used to extract time-domain and frequency-domain bioelectrical features from the bioelectrical signals in real time. A multimodal fusion model module is used to receive the bioelectrical characteristics and the microenvironmental parameter data, and perform joint analysis to identify early stress patterns of pests and diseases; The early warning module is used to generate and send early warning information based on the output of the multimodal fusion model module.

Images