Methods, devices, electronic equipment and storage media for early warning of stress before seedling emergence

Abstract

This invention relates to the field of smart agricultural seedling cultivation technology, providing a method, device, electronic device, and storage medium for pre-emergence stress early warning of seedlings. The method includes: acquiring multimodal temporal features of the seedling cultivation environment, including impedance mode features, gas mode features, and temperature change mode features; inputting the multimodal temporal features into a stress discrimination model to obtain a pre-emergence stress early warning result. A single-modal temporal encoding layer is used to independently encode each modal feature; a temporal attention layer is used to temporally weight the encoded modal features to focus on key temporal information; and a modal interaction fusion layer is used to interact and fuse the weighted modal features to generate a pre-emergence stress early warning result. This invention can accurately identify the pre-emergence stress state of seedlings from complex, multi-source weak signals, achieving early and accurate early warning of latent stress before seedling emergence, significantly improving the timeliness and accuracy of pre-emergence stress early warning.

Description

Technical Field

[0001] This invention relates to the field of smart agricultural seedling cultivation technology, and in particular to a method, device, electronic device and storage medium for early warning of stress before seedling emergence. Background Technology

[0002] In modern large-scale agricultural production, the germination rate and vigor of seedlings are key factors that determine the yield and economic benefits of subsequent crops. Therefore, early and accurate identification and warning of stresses that seedlings may suffer during the germination stage, such as seed rot, mold, lack of oxygen, and salt damage, is of great technical demand and application value.

[0003] To meet the above needs, existing technologies mainly employ the following monitoring methods: First, by visual recognition or manual inspection, the health of seedlings is determined by observing the morphology, color, and other appearance characteristics of their leaves after emergence. Second, sensors are deployed to monitor single-dimensional environmental information, such as monitoring only the macroscopic temperature, humidity, or carbon dioxide concentration of the seedling environment, or only the humidity or conductivity of the seedling substrate. The growth status of the seedlings is indirectly inferred through changes in the thresholds of these single parameters. Third, some existing technologies also attempt to integrate signals from multiple sensors, typically by simply splicing or performing a fixed weighted summation of data from different signals to obtain more comprehensive information.

[0004] However, the aforementioned existing technologies still have the following significant shortcomings in practical applications. First, relying on visual recognition or manual inspection after emergence results in a severe delay in early warning. This is because by the time visible stress symptoms appear on the seedling leaves, the seedlings have already suffered irreversible damage during the pre-emergence seed germination stage (usually 2-7 days before emergence), missing the optimal intervention window and leading to poor intervention effects and high costs. Second, schemes relying on single signals for monitoring lack sufficient recognition accuracy and have weak anti-interference capabilities. For example, monitoring carbon dioxide concentration or substrate impedance alone is easily affected by external factors such as greenhouse ventilation, substrate humidity fluctuations, and changes in ambient temperature, and cannot accurately distinguish between different stress types (such as hypoxia and seed rot, both of which can cause abnormal carbon dioxide concentrations), making it difficult to meet the needs of precision seedling cultivation. Finally, some existing multimodal data fusion algorithms are relatively rudimentary. These algorithms typically employ simple feature splicing or fixed-weighted summation, failing to fully consider the temporal correlations within different modal signals and ignoring the dynamic changes in the importance of each modal signal and at different time points under different stress scenarios. Therefore, the models built by these algorithms have weak generalization ability, and their recognition accuracy and stability are difficult to guarantee when faced with complex environments of different crop varieties or different bases. Summary of the Invention

[0005] This invention provides a method, device, electronic device, and storage medium for early warning of stress before seedling emergence, in order to solve the defects in the prior art, such as delayed early warning leading to missed intervention windows, poor anti-interference of single signal monitoring and difficulty in distinguishing stress types, and the failure of multimodal data fusion algorithms to dynamically capture temporal correlation and scenario-based weight changes, resulting in insufficient model generalization ability.

[0006] This invention provides a method for early warning of stress before seedling emergence, comprising the following steps.

[0007] The multimodal temporal characteristics of the seedling cultivation environment are obtained; the multimodal temporal characteristics include impedance mode characteristics, gas mode characteristics, and temperature change mode characteristics. The multimodal temporal features are input into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model; The stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal coding layer is used to independently encode the impedance modal features, the gas modal features, and the temperature-varying modal features to obtain impedance modal coding features, gas modal coding features, and temperature-varying modal coding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal coding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal coding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal coding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

[0008] According to the present invention, a method for early warning of pre-emergence stress in seedlings includes performing intermodal interaction and fusion of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature change mode time-weighted features to generate the pre-emergence stress early warning result, comprising: Construct a modal interaction matrix; the modal interaction matrix is ​​used to characterize the correlation between any two of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature-varying mode time-weighted features; Based on the modal interaction matrix, cross-modal attention weights are determined; the cross-modal attention weights include a first weight, a second weight, and a third weight. Based on the first weight, the second weight, and the third weight, the impedance mode time-series weighted features, the gas mode time-series weighted features, and the temperature change mode time-series weighted features are weighted and fused respectively to generate the pre-emergence stress early warning result.

[0009] According to the present invention, a method for early warning of stress before seedling emergence is provided, wherein the single-modal temporal coding layer includes dilated convolutional units and residual connection units; The dilated convolution unit is used to perform convolution operations with hierarchically increasing dilation coefficients on the impedance mode features, the gas mode features, and the temperature variation mode features, respectively, to obtain deep impedance time series information, deep gas time series information, and deep temperature variation time series information. The residual connection unit is used to perform a residual connection between the impedance deep timing information and the impedance mode feature to obtain the impedance mode coding feature, to perform a residual connection between the gas deep timing information and the gas mode feature to obtain the gas mode coding feature, and to perform a residual connection between the temperature variation deep timing information and the temperature variation mode feature to obtain the temperature variation mode coding feature.

[0010] According to the present invention, a method for early warning of pre-emergence stress in seedlings includes acquiring multimodal temporal features of the seedling cultivation environment, comprising: Acquire raw multimodal signals; the raw multimodal signals include impedance signals, gas signals, and temperature change signals; Wavelet decomposition is performed on the impedance signal and the gas signal to obtain detail coefficients characterizing the high-frequency components; The adaptive dynamic threshold is determined based on the statistical characteristics of the detail coefficients; The corrected detail coefficients are determined based on the detail coefficients and the adaptive dynamic threshold. Wavelet reconstruction is performed based on the corrected detail coefficients to obtain the noise-reduced impedance signal and the noise-reduced gas signal; The temperature change signal is denoised using Kalman filtering to obtain a denoised temperature change signal. Based on the noise-reduced impedance signal, the noise-reduced gas signal, and the noise-reduced temperature change signal, the impedance mode features, the gas mode features, and the temperature change mode features are extracted respectively.

[0011] According to the present invention, a method for early warning of pre-emergence stress in seedlings includes the following step: acquiring raw multimodal signals. The original multimodal signals are collected based on multiple monitoring nodes; the monitoring nodes are deployed in monitoring trays within the seedling cultivation area; at least one mobile verification node is also deployed within the seedling cultivation area. The pre-emergence stress early warning results include the stress level; The step of inputting the multimodal temporal features into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model further includes: The stress level is taken as the known stress level of the monitoring acupoint plate; Determine the spatial distance between the non-monitored seedling trays and each of the monitored seedling trays in the seedling cultivation area, and determine the temporal correlation coefficient used to characterize the data trend correlation between the non-monitored seedling trays and the monitored seedling trays; Based on the spatial distance and the temporal correlation coefficient, the spatiotemporal correlation weights between the non-monitored acupoints and each of the monitored acupoints are determined; Based on the known stress level and the spatiotemporal correlation weight, an interpolation model is constructed to determine the stress level of the non-monitored acupoint.

[0012] According to the present invention, a method for early warning of pre-emergence stress in seedlings is provided, the method further comprising: Based on the stress level of the non-monitored acupoints and the preset review conditions, the target acupoints to be reviewed are determined. The mobile verification node is scheduled to move to the target acupoint, and target verification data is collected based on the mobile verification node; The target verification data is input into the stress discrimination model to obtain the precise stress level output by the stress discrimination model; If the difference between the precise stress level and the stress level of the unmonitored acupoint is greater than a preset correction threshold, the parameters of the interpolation model are corrected.

[0013] According to the present invention, a method for early warning of pre-emergence stress in seedlings includes an impedance modal feature comprising at least one of impedance change rate and impedance variance; a gas modal feature comprising at least one of carbon dioxide concentration gradient, characteristic peak intensity of volatile organic compounds, and correlation coefficient between carbon dioxide and volatile organic compounds; and a temperature change modal feature comprising at least one of temperature change slope, temperature anomaly degree, and temperature change variance.

[0014] The present invention also provides a seedling pre-emergence stress early warning device, comprising the following modules: The acquisition module is used to acquire multimodal temporal characteristics of the seedling cultivation environment; the multimodal temporal characteristics include impedance mode characteristics, gas mode characteristics, and temperature change mode characteristics; The input module is used to input the multimodal temporal features into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model; The stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal coding layer is used to independently encode the impedance modal features, the gas modal features, and the temperature-varying modal features to obtain impedance modal coding features, gas modal coding features, and temperature-varying modal coding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal coding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal coding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal coding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the pre-emergence stress early warning method for seedlings as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the pre-emergence stress early warning method for seedlings as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the pre-emergence stress early warning method for seedlings as described above.

[0018] The present invention provides a method, device, electronic device, and storage medium for pre-emergence stress early warning of seedlings. It acquires multimodal temporal features of the seedling cultivation environment, including impedance mode features, gas mode features, and temperature change mode features. These multimodal temporal features are input into a stress discrimination model to obtain pre-emergence stress early warning results. The stress discrimination model includes a single-modal temporal encoding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal encoding layer independently encodes each modal feature; the temporal attention layer performs temporal weighting on the encoded modal features to focus on key temporal information; and the modal interaction fusion layer interacts and fuses the weighted modal features to generate pre-emergence stress early warning results. This invention, based on acquired impedance, gas, and temperature-varying multimodal temporal features, independently encodes each modal feature through a single-modal temporal coding layer to preserve its unique temporal patterns. Then, a temporal attention layer automatically identifies and focuses on the most critical temporal information for stress assessment from the temporal sequences of each modality. Furthermore, a modal interaction fusion layer deeply mines the synergistic and complementary relationships between different modal features. This allows for accurate identification of pre-emergence stress states from complex, multi-source weak signals, solving the problems of existing technologies that rely on single signals, are susceptible to interference, have low identification accuracy, and use simple fusion algorithms that fail to effectively utilize temporal dynamics and cross-modal correlation information, resulting in delayed warnings and low accuracy. This invention achieves early and accurate warnings of latent stress before seedling emergence, significantly improving the timeliness and accuracy of pre-emergence stress warnings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts of the pre-emergence stress early warning method for seedlings provided by the present invention.

[0021] Figure 2 This is the second flowchart of the pre-emergence stress early warning method for seedlings provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the pre-emergence stress early warning device for seedlings provided by the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and that the objects distinguished by "first," "second," etc., are generally of the same class.

[0026] This invention provides a method for early warning of stress before seedling emergence. This method aims to solve the technical problem in the prior art of making it difficult to identify latent stresses such as seed rot, mold, hypoxia, and salt damage that occur before seedling emergence in an early and accurate manner. Figure 1 This is one of the flowcharts illustrating the pre-emergence stress early warning method for seedlings provided by this invention, such as... Figure 1 As shown, the method includes the following: Step 110: Obtain multimodal temporal characteristics of the seedling cultivation environment; the multimodal temporal characteristics include impedance mode characteristics, gas mode characteristics, and temperature change mode characteristics.

[0027] Specifically, firstly, the multimodal temporal characteristics of the seedling cultivation environment can be obtained. This step refers to deploying one or more sensors in the seedling cultivation environment to periodically collect various types (i.e., different modes) of physical or chemical signals that can reflect the germination state of the seedlings at a preset time frequency, and forming time series data.

[0028] The seedling cultivation environment refers to a place used for industrialized, large-scale seedling cultivation, such as seedbeds or seedling trays in intelligent greenhouses, plant factories, or artificial climate chambers. This embodiment of the invention does not specifically limit this. It should be understood that such an environment typically has controllable macroscopic environmental parameters such as temperature, humidity, and light. The monitoring object of this embodiment of the invention is the microscopic environment within this macroscopic environment, namely the environment inside the seedling substrate and immediately adjacent to the seed germination area.

[0029] Multimodal temporal features are a set of time-series data used to comprehensively and multidimensionally characterize the physiological and metabolic state of seedlings before emergence (seed germination stage) and the dynamic changes of their surrounding microenvironment. By fusing signals from multiple different sources, the limitations of single signals being susceptible to interference and providing incomplete information can be overcome, thus improving the accuracy of stress assessment. In this embodiment, the multimodal temporal features include impedance mode features, gas mode features, and temperature variation mode features.

[0030] Here, impedance modal characteristics are used to reflect comprehensive information about the conductivity and moisture content of the substrate in the seedling cultivation environment, and are closely related to the salt concentration and humidity in the substrate. For example, salt stress can lead to a significant increase in substrate conductivity, thereby reducing impedance; while moisture changes during seed rot can also cause changes in impedance. Impedance modal characteristics can be obtained by deploying a substrate impedance sensor, such as a bipolar stainless steel probe sensor, in the seedling substrate, for example, about 0.5 cm below the seed radicle, at a set frequency, such as every 5 minutes, to collect the substrate impedance values ​​and form an impedance time series.

[0031] Here, gas modal characteristics are used to reflect respiratory metabolic activities during seed germination and characteristic gas changes caused by potential microbial activities (such as mold growth). Healthy seed germination is accompanied by stable respiration, consuming oxygen and releasing carbon dioxide (CO2); while stresses such as mold growth may produce specific volatile organic compounds (VOCs). Therefore, gas modal characteristics can be directly correlated with the physiological state of the seed. Gas modal characteristics can be acquired by deploying miniature gas sensors on or within the substrate of the seedling tray. For example, a composite sensor integrating a non-dispersive infrared (NDIR) CO2 sensor and a photoionization detector (PID) VOCs sensor can be used to collect CO2 concentration C(t) and VOCs concentration V(t) at a frequency synchronized with the impedance sensor, forming a gas concentration time series.

[0032] Here, temperature variation modal characteristics are used to reflect the changes in trace amounts of heat generated by metabolism during seed germination. Seed germination is an energy-consuming process, accompanied by weak heat production, forming a micro-temperature zone distinct from the ambient background temperature. When seeds are subjected to stress, their metabolic activity becomes disordered, leading to abnormal heat production. Temperature variation modal characteristics can be obtained by using a non-contact infrared micro-area temperature sensor, which is aimed directly at the substrate area where the seed is located to measure the temperature, avoiding disturbance to the substrate and interference with the ambient temperature caused by contact measurements, thereby obtaining a time series of the micro-area temperature T(t).

[0033] It's important to note that multimodal temporal features emphasize that the acquired data is a series of data points that change over time, rather than isolated values ​​at a single moment. For example, a sequence of 12 data points, one every 5 minutes over the past 60 minutes, can be collected for subsequent analysis using a stress discrimination model. This temporal sequence is crucial for capturing the dynamic process of stress occurrence.

[0034] Step 120: Input the multimodal temporal features into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model; The stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal coding layer is used to independently encode the impedance modal features, the gas modal features, and the temperature-varying modal features to obtain impedance modal coding features, gas modal coding features, and temperature-varying modal coding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal coding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal coding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal coding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

[0035] Specifically, after obtaining the multimodal time-series features, the multimodal time-series features can be input into the stress discrimination model to obtain the pre-emergence stress warning results of the seedlings output by the stress discrimination model.

[0036] Here, the internal structure of the stress discrimination model is specially designed to process and analyze the aforementioned multimodal time-series data, and ultimately output a judgment on whether the seedlings are under stress and what type of stress they may be under. The stress discrimination model can learn and understand the complex patterns exhibited by different stresses, such as seed rot, mold, hypoxia, and salt damage, in multimodal data.

[0037] The pre-emergence stress warning result is the output of the stress discrimination model. This result can be a classification label, clearly indicating the currently determined stress type, such as normal, seed rot, mold, hypoxia, or salt damage. Furthermore, the pre-emergence stress warning result can also include a confidence score, representing the degree of certainty the stress discrimination model has regarding this judgment; or a stress level, such as 1-5, to quantify the severity of the stress and provide a basis for subsequent precise intervention.

[0038] Here, the stress discrimination model may include a single-mode temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-mode temporal coding layer is used to independently encode the impedance mode features, gas mode features, and temperature-varying mode features to obtain impedance mode coding features, gas mode coding features, and temperature-varying mode coding features.

[0039] Independent encoding refers to a single-modal temporal coding layer with three parallel processing paths, each dedicated to processing the temporal characteristics of a specific modality. During the encoding phase, the information from the three modalities does not interfere with each other. The advantage of this design is that it fully preserves the unique temporal dynamics and inherent patterns of each modality, avoiding mutual influence between the statistical characteristics of different modal signals in the early processing stages. For example, gas signals may exhibit abrupt changes, while temperature-varying signals may exhibit gradual changes; independent encoding can better capture their respective patterns. The single-modal temporal coding layer can be implemented by a neural network module capable of processing temporal data, such as a Temporal Convolutional Network (TCN), a Recurrent Neural Network (RNN), or its variants such as a Long Short-Term Memory (LSTM). This embodiment of the invention does not specifically limit this approach.

[0040] After processing by the single-mode temporal coding layer, the original, relatively shallow three-mode feature sequences are transformed into impedance mode coding features, gas mode coding features, and temperature-varying mode coding features with higher information density and better reflecting temporal dependencies. The impedance mode coding features, gas mode coding features, and temperature-varying mode coding features are respectively deep abstract representations of the impedance mode features, gas mode features, and temperature-varying mode features.

[0041] The temporal attention layer is used to determine the impedance mode temporal weighting features that focus on the first key temporal information based on impedance mode coding features, to determine the gas mode temporal weighting features that focus on the second key temporal information based on gas mode coding features, and to determine the temperature-varying mode temporal weighting features that focus on the third key temporal information based on temperature-varying mode coding features.

[0042] The core function of the temporal attention layer is to capture the dynamic changes of temporal features in each modality, overcoming the shortcomings of traditional models that distribute weights equally across temporal points and fail to focus on key temporal information. By assigning different attention weights to different temporal points, the temporal attention layer focuses on key temporal points before the stress occurs (such as signal abrupt changes), thereby improving the accuracy of stress discrimination models in identifying early stresses. Specifically, the temporal attention weights of the temporal attention layer... Through formula Perform independent calculations for each mode, where For example, the feature vector of the t-th time point after encoding the m-th modality. t=1,2,...,L, m=Z,G,T correspond to three modes respectively. Regarding parameters, The temporal attention weights range from [0,1] and the sum of the weights is 1. The larger the value, the greater the contribution of that temporal point to the coercion identification. It is a learnable weight matrix (e.g., 64×64). This is a scaling factor (e.g., 64) used to prevent the softmax function from saturating; This represents the bias term of the temporal attention layer. Its core innovation lies in the fact that traditional models use "temporal averaging" or "fixed weights" to process temporal features, which cannot distinguish the importance of time points; this invention, on the other hand, calculates attention weights independently for each modality, which can adapt to the unique temporal characteristics of different modalities (such as gas modalities which are mostly abrupt, and temperature-changing modalities which are mostly gradual), automatically enhances the feature contribution of key time points before stress, and weakens the interference of normal time points, thereby significantly improving the recognition accuracy of weak signals such as early hypoxia.

[0043] Here, the first, second, and third key temporal information refer to the information within a specific time point or time period that most significantly predicts the occurrence of stress in the time series of their respective modes. Taking the mold growth process as an example, before mold growth occurs, the concentration of volatile organic compounds may show a slight but continuous inflection point; this moment is the second key temporal information corresponding to the gaseous mode. The temporal attention layer can automatically identify and focus on the temporal location of such key information through a self-learning mechanism.

[0044] Specifically, for each modality's encoded feature sequence, the temporal attention layer calculates a weight score for each time point in the sequence. A higher weight score indicates a greater contribution to the final judgment. Then, these weight scores are used to weight and sum the original encoded feature sequence, generating a new feature representation. The generated impedance mode temporal weighted features, gas mode temporal weighted features, and temperature-varying mode temporal weighted features are attention-weighted features. Compared to the impedance mode encoded features, gas mode encoded features, and temperature-varying mode encoded features, they highlight the most valuable temporal segments for stress discrimination while suppressing the influence of irrelevant or noisy time points.

[0045] Here, the modal interaction fusion layer is used to perform intermodal interaction and fusion of impedance modal time-weighted features, gas modal time-weighted features, and temperature-varying modal time-weighted features to generate pre-emergence stress early warning results. This step is not simply about splicing together the three features—impedance modal time-weighted features, gas modal time-weighted features, and temperature-varying modal time-weighted features—but more importantly, it achieves intermodal interaction. Modal interaction refers to exploring and utilizing the synergistic or complementary relationships between different modal features. For example, when the phenomenon of increased humidity, which may be indicated by a decrease in matrix impedance, occurs simultaneously with an abnormal increase in carbon dioxide concentration, their joint characterization is usually more significant in indicative of seed rot stress than a single signal. The modal interaction fusion layer achieves cross-modal information interaction through specific network structures, such as using a multi-head attention mechanism, or by using operations such as tensor product to systematically capture such cross-modal correlation patterns. After the interaction is completed, the modal interaction fusion layer will further integrate all modal information to form a unified and comprehensive fusion feature. Finally, through one or more fully connected layers, the fusion feature will be mapped to the pre-emergence stress early warning result, completing the end-to-end computation from multimodal data to early warning decision.

[0046] As a specific implementation method, the modal interaction fusion layer can be divided into two steps: the first step is to use the formula Output the probability distribution of the stress type. Wherein, Type identification weight matrix (e.g., 5×64). The probability distribution of stress types corresponds to five states: normal, seed rot, mold, hypoxia, and salinity. Indicates fusion features, This represents the learnable bias term. The second step is to use the formula... Output a coercion level of 1-5. Among them, The weight matrix for determining the level (e.g., 5×64) represents the coercion level. The value range is [1, 5], corresponding to stress levels from mild to severe. Indicates fusion features, This represents the learnable bias term. Ultimately, the stress discrimination model outputs a triple result: "type + level + confidence." For example, when the maximum value in probability P is greater than a preset threshold (e.g., 0.8), it is considered a valid identification; if it is less than the preset threshold (i.e., low confidence), it can trigger AGV (Automated / Automatic Guided Vehicle) movement verification, thereby further reducing the risk of misjudgment and missed judgment, and providing more reliable decision support for subsequent graded early warning and intervention.

[0047] The method provided in this invention acquires multimodal temporal features of the seedling cultivation environment, including impedance modal features, gas modal features, and temperature change modal features; the multimodal temporal features are input into a stress discrimination model to obtain pre-emergence stress warning results. The stress discrimination model includes a single-modal temporal encoding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal encoding layer independently encodes each modal feature; the temporal attention layer performs temporal weighting on the encoded modal features to focus on key temporal information; and the modal interaction fusion layer interacts and fuses the weighted modal features to generate pre-emergence stress warning results. This invention, based on acquired impedance, gas, and temperature-varying multimodal temporal features, independently encodes each modal feature through a single-modal temporal coding layer to preserve its unique temporal patterns. Then, a temporal attention layer automatically identifies and focuses on the most critical temporal information for stress assessment from the temporal sequences of each modality. Furthermore, a modal interaction fusion layer deeply mines the synergistic and complementary relationships between different modal features. This allows for accurate identification of pre-emergence stress states from complex, multi-source weak signals, solving the problems of existing technologies that rely on single signals, are susceptible to interference, have low identification accuracy, and use simple fusion algorithms that fail to effectively utilize temporal dynamics and cross-modal correlation information, resulting in delayed warnings and low accuracy. This invention achieves early and accurate warnings of latent stress before seedling emergence, significantly improving the timeliness and accuracy of pre-emergence stress warnings.

[0048] Based on the above embodiments, the step of performing intermodal interaction and fusion of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature-varying mode time-weighted features to generate the pre-emergence stress early warning result includes: Step 210, construct a modal interaction matrix; the modal interaction matrix is ​​used to characterize the correlation between any two of the impedance mode time-series weighted features, the gas mode time-series weighted features, and the temperature-varying mode time-series weighted features; Step 220: Based on the modal interaction matrix, determine the cross-modal attention weights; the cross-modal attention weights include a first weight, a second weight, and a third weight; Step 230: Based on the first weight, the second weight, and the third weight, the impedance mode time-series weighted features, the gas mode time-series weighted features, and the temperature change mode time-series weighted features are weighted and fused respectively to generate the pre-emergence stress early warning result.

[0049] Specifically, firstly, a modal interaction matrix can be constructed, which is used to characterize the correlation between any two features among the impedance mode time-weighted features, gas mode time-weighted features, and temperature-varying mode time-weighted features.

[0050] Here, the formula for the modal interaction matrix is ​​as follows: ; Where M represents the modal interaction matrix, This represents the impedance mode time-weighted characteristic. The transpose of the impedance mode time-weighted characteristic. Represents the time-weighted characteristics of gas modes. Represents the transpose of the time-weighted characteristics of gas modes. This represents the time-weighted characteristics of temperature-varying modes. This represents the transpose of the time-weighted features of temperature-varying modes.

[0051] Here, the modal interaction matrix M can have a dimension of 192×192. By performing inner and outer product operations on each modal feature, it can comprehensively capture the complex relationships within a modality (diagonal elements) and between modalities (off-diagonal elements).

[0052] In this step, the modal interaction matrix is ​​a numerical matrix used to systematically characterize and quantify the complex correlations within (intra-modal) and between (inter-modal) the three weighted modal features output by the temporal attention layer. The modal interaction matrix can be constructed by performing some form of interaction operation, such as dot product or outer product, between each pair of the impedance modal temporal weighted features, gas modal temporal weighted features, and temperature variation modal temporal weighted features, and then arranging and combining them. By constructing such a comprehensive matrix, the stress discrimination model can capture the cooperative, reinforcing, or suppressive relationships between all modes at once, such as the correlation strength between impedance changes and gas changes, and the internal correlation strength of impedance changes themselves.

[0053] After obtaining the modal interaction matrix, cross-modal attention weights can be determined based on the modal interaction matrix. The cross-modal attention weights include a first weight, a second weight, and a third weight.

[0054] For example, after flattening the modal interaction matrix M, cross-modal attention weights can be calculated using a weight generation network. It includes the first weight Second weight and third weight The core formula for cross-modal attention weights is as follows: Where M represents the modal interaction matrix, For cross-modal attention weights, The sum of the weights is 1, which is used to dynamically adjust and allocate the importance of the three modal features in the final fusion decision. Represents the weight matrix. This represents a learnable bias term.

[0055] For example, when identifying mold stress, the stress discrimination model may learn the weights assigned to the gas modes. It should be greater than 0.5; while when identifying salt stress, the weight assigned to the impedance mode should be greater than 0.5. It may be greater than 0.5, thus enabling adaptive weight allocation for different stress types.

[0056] Here, the first, second, and third weights correspond to the weight values ​​assigned to the impedance mode time-series weighted features, the gas mode time-series weighted features, and the temperature-varying mode time-series weighted features, respectively. For example, when the stress discrimination model determines the current stress mode based on the modal interaction matrix, such as when high salinity has the strongest correlation with the impedance mode, the calculated value of the first weight will be significantly higher than the other two weights. This allows the stress discrimination model to assign a higher contribution to the impedance mode features in subsequent fusion.

[0057] Finally, based on the first, second, and third weights, the impedance mode time-series weighted features, gas mode time-series weighted features, and temperature-varying mode time-series weighted features are weighted and fused respectively to generate pre-emergence stress early warning results. In this step, weighted fusion refers to multiplying the first, second, and third weights determined in the previous step by the corresponding impedance mode time-series weighted features, gas mode time-series weighted features, and temperature-varying mode time-series weighted features, respectively, and then combining these three weighted features to form a final, highly fused feature. This fused feature is then fed into the classifier to generate the final pre-emergence stress early warning result.

[0058] Here, the formula for feature fusion is as follows: ; in, Indicates fusion features, To integrate matrix weights, a high-dimensional interaction matrix can be mapped to a feature space, such as a 64-dimensional one, to keep it consistent with the dimension of the single-modal weighted features, thus achieving effective integration. This represents the impedance mode time-weighted characteristic. Represents the time-weighted characteristics of gas modes. This represents the time-weighted characteristics of temperature-varying modes. This indicates a flattening operation on the modal interaction matrix M.

[0059] The method provided in this invention constructs a modal interaction matrix to comprehensively capture the complex correlations within and between modalities. Then, it calculates cross-modal attention weights based on the modal interaction matrix, reflecting the current importance of each modality. The first, second, and third weights of these cross-modal attention weights are then used to adaptively weight and fuse the features of each modality. This allows the stress discrimination model to dynamically adjust its dependence on different modal information according to real-time data patterns. This solves the problem that traditional fusion algorithms, such as simple concatenation or fixed-weight summation, cannot adaptively adjust modal weights according to different stress types, resulting in weak model generalization and poor adaptability. It achieves a context-aware feature fusion for different stress scenarios, significantly improving the accuracy and robustness of the stress discrimination model in distinguishing different stress types.

[0060] Based on the above embodiments, the single-modal temporal coding layer includes dilated convolution units and residual connection units; The dilated convolution unit is used to perform convolution operations with hierarchically increasing dilation coefficients on the impedance mode features, the gas mode features, and the temperature variation mode features, respectively, to obtain deep impedance time series information, deep gas time series information, and deep temperature variation time series information. The residual connection unit is used to perform a residual connection between the impedance deep timing information and the impedance mode feature to obtain the impedance mode coding feature, to perform a residual connection between the gas deep timing information and the gas mode feature to obtain the gas mode coding feature, and to perform a residual connection between the temperature variation deep timing information and the temperature variation mode feature to obtain the temperature variation mode coding feature.

[0061] Specifically, the single-modal temporal coding layer includes dilated convolutional units and residual connection units. This structure aims to effectively extract long temporal dependencies in temporal features through an innovative design that combines dilated convolution and residual connections, while avoiding the loss of key weak signal information during deep feature extraction.

[0062] The dilated convolution unit performs a special convolution operation called dilated convolution. It performs convolution operations with hierarchically increasing dilation coefficients on impedance mode features, gas mode features, and temperature-varying mode features, respectively, to obtain deep temporal information of impedance, gas, and temperature variations. Compared to standard convolution, dilated convolution increases the receptive field by introducing holes (i.e., dilation coefficients) between the elements of the convolution kernel without increasing computational cost or reducing spatial resolution.

[0063] The hierarchical increasing dilation coefficient means that a dilated convolutional unit may contain multiple dilated convolutional layers, with the dilation coefficient of each subsequent layer being greater than that of the previous layer. For example, the dilation coefficients might be 1, 2, 4, and 8. This design allows the network to simultaneously capture temporal dependencies at different scales, namely short-term, medium-term, and long-term. The impedance deep temporal information, gas deep temporal information, and temperature change deep temporal information obtained after processing by the dilated convolutional unit are feature representations that contain richer contextual information and higher-level abstract patterns, obtained by extracting multi-scale temporal features from the original modal features.

[0064] Here, the residual connection unit is used to residually connect the deep impedance time series information with the impedance mode features to obtain the impedance mode coding features, to residually connect the deep gas time series information with the gas mode features to obtain the gas mode coding features, and to residually connect the deep temperature variation time series information with the temperature variation mode features to obtain the temperature variation mode coding features. The specific implementation of the residual connection is usually to add the output of the dilated convolution unit (i.e., the deep time series information) to the input of the dilated convolution unit (i.e., the original mode features) element-wise.

[0065] As a specific implementation, the single-modal temporal coding layer is implemented through an improved temporal convolutional network. The coding formulas for each modality can be expressed as: Impedance mode coding features are Gas mode coding features are The temperature-varying mode coding features are .

[0066] in, This indicates deep timing information about the impedance. Indicates impedance mode characteristics, Represents deep gas time-series information. Indicates gas modal characteristics, This represents deep temporal information about temperature changes. This indicates the characteristics of temperature-dependent modes.

[0067] The core formula of TCN is: . This represents impedance mode characteristics, gas mode characteristics, or temperature-varying mode characteristics. The TCN is constructed using dilated convolutions with hierarchically increasing dilation coefficients (e.g., d=2,3,4). It captures short, medium, and long temporal dependencies, expanding the receptive field without increasing the number of parameters. Residual connections directly add the input features to the encoded features, effectively solving the gradient vanishing problem in deep networks and ensuring that original features, which may contain weak stress signals, are not filtered out. A single-modal temporal coding layer enhances the original features (e.g., 8-dimensional) to higher dimensions (e.g., 64-dimensional) to mine deeper features; its weight matrix... Through iterative optimization via backpropagation, the bias term... It can then be initialized to a non-zero value (such as 0.1) to avoid gradient vanishing. Through this design that combines hierarchical dilated convolution with residual connections, the single-modal temporal coding layer can capture long temporal dependencies more accurately than traditional CNNs (Convolutional Neural Networks) or ordinary TCNs, while retaining key weak signal features, significantly improving the feature extraction capability for stresses such as early mold growth and mild hypoxia.

[0068] Furthermore, impedance mode time-weighted characteristics The formula is: ,in, This represents the impedance mode characteristics at time t. This represents the impedance mode coding characteristics at time t. L Indicates the sequence length. This represents the attention weight vector.

[0069] Gas mode time-weighted characteristics The formula is: ,in, This represents the gas modal characteristics at time t. This represents the gas mode coding features at time t. L Indicates the sequence length. This represents the attention weight vector.

[0070] Time-weighted characteristics of temperature-varying modes The formula is: ,in, This represents the temperature-varying modal characteristics at time t. This represents the temperature-varying mode coding characteristics at time t. L Indicates the sequence length. This represents the attention weight vector.

[0071] The method provided in this invention expands the receptive field in a single-modal temporal coding layer by using dilated convolutional units with progressively increasing dilation coefficients. This effectively captures long-term temporal dependencies in the signal and extracts deep temporal information. Then, the extracted deep temporal information is directly added to the original input features through residual connection units. This ensures that the original signal, which may contain weak signs of early stress, can be transmitted to deeper layers of the network without loss while learning deep abstract features. This solves the problem that traditional temporal coding networks cannot capture long-term temporal dependencies due to limited receptive fields, or lose key weak signals due to information decay as the network deepens. It achieves efficient deep coding of temporal features and effective preservation of original weak signals, significantly improving the sensitivity and capture ability of the stress discrimination model to early and weak stress signals, thereby further improving the timeliness of pre-emergence stress warning.

[0072] Based on the above embodiments, step 110 includes: Step 110-1: Acquire raw multimodal signals; the raw multimodal signals include impedance signals, gas signals, and temperature change signals; Step 110-2: Perform wavelet decomposition on the impedance signal and the gas signal to obtain detail coefficients characterizing the high-frequency components; Step 110-3: Determine the adaptive dynamic threshold based on the statistical characteristics of the detail coefficients; Step 110-4: Determine the corrected detail coefficients based on the detail coefficients and the adaptive dynamic threshold; Step 110-5: Perform wavelet reconstruction based on the corrected detail coefficients to obtain the noise-reduced impedance signal and the noise-reduced gas signal; Step 110-6: Perform noise reduction processing on the temperature change signal based on Kalman filtering to obtain a noise-reduced temperature change signal; Step 110-7: Based on the noise-reduced impedance signal, the noise-reduced gas signal, and the noise-reduced temperature change signal, the impedance mode features, the gas mode features, and the temperature change mode features are extracted respectively.

[0073] Specifically, the first step is to acquire raw multimodal signals. Raw multimodal signals refer to the unprocessed data streams directly acquired by sensors deployed in the seedling cultivation environment, without any digital signal processing. Raw multimodal signals typically contain useful physiological information as well as various environmental noises and random interferences. Raw multimodal signals include at least impedance signals, gas signals, and temperature variation signals.

[0074] To further illustrate the signal acquisition method, this embodiment employs an optimized grid deployment scheme of "fixed monitoring and mobile verification collaboration" to balance monitoring accuracy and hardware cost. Specifically, a 4×4 standard seedling tray matrix (containing 16 standard seedling trays) is set as a monitoring unit. In the substrate of the center and four corner trays of each monitoring unit, a set of multimodal sensors is deployed as fixed monitoring nodes, for a total of 5 fixed nodes; the sensors are buried at a depth of 2.5±0.5cm, close to the seed radicle, to ensure the authenticity of the acquired signals. Simultaneously, an AGV equipped with a complete set of multimodal sensors is deployed in each greenhouse as a mobile verification node. The fixed nodes periodically collect data at a preset frequency (e.g., every 5 minutes), with each parameter repeatedly collected 3 times and the average value taken to eliminate instantaneous fluctuations. The collection time points are t0, t1, ..., t n (t) n = t0 + 5n minutes), and its average data formula is: ; in, For the valid data at the k-th time point, This represents the data collected for the i-th time. The mobile node, according to system scheduling instructions, performs targeted, high-precision sampling of high-risk acupoints (e.g., stress level ≥ 3) for subsequent verification and model correction.

[0075] Secondly, the impedance signal and the gas signal are processed using improved wavelet denoising. This is because, when responding to stress, the effective signal of the impedance signal and the gas signal may exhibit high-frequency abrupt changes or weak instantaneous variations, which are mixed with high-frequency noise.

[0076] Specifically, wavelet decomposition is performed on the impedance signal and the gas signal to obtain detail coefficients characterizing the high-frequency components. These detail coefficients represent the high-frequency components of the signal at different scales, and may include environmental noise as well as crucial early stress information such as abrupt changes in VOCs signals during the initial stage of mold growth.

[0077] The adaptive dynamic threshold is determined based on the statistical properties of the detail coefficients. Here, the adaptive dynamic threshold is a dynamic threshold used to distinguish between noise and useful signals. Its value is not fixed, but is calculated in real time based on statistical properties such as the median absolute deviation of the detail coefficients of the current signal segment.

[0078] Furthermore, the corrected detail coefficients can be determined based on the detail coefficients and the adaptive dynamic threshold. This step processes the detail coefficients using a modified soft thresholding function. For coefficients larger than the adaptive dynamic threshold, they are shrunk; for coefficients smaller than the adaptive dynamic threshold, they are not directly set to zero, but are retained at a small percentage, such as 10%, to ensure that early, weak stress signals are not completely filtered out.

[0079] Then, wavelet reconstruction is performed based on the corrected detail coefficients to obtain the denoised impedance signal and the denoised gas signal. The denoised impedance signal and the denoised gas signal refer to high-fidelity signals that, after the above wavelet denoising process, have significantly suppressed noise while retaining the core signal features and weak abrupt changes.

[0080] As a detailed explanation of this improved wavelet denoising process, its core formulas include: First, through the formula... ,in Calculate the adaptive dynamic threshold This formula is based on the detail coefficients of wavelet decomposition. Calculate the adaptive dynamic threshold to achieve adaptive noise reduction, where 0.6745 is the conversion factor between the absolute deviation of the median and the standard deviation of the normal distribution, which is applicable to Gaussian noise; The dynamic threshold for the k-th level wavelet decomposition is used to distinguish between noise and valid signals. It is the noise standard deviation of the k-th layer, which is converted from the median absolute deviation of the detail coefficients; represents the detail coefficients obtained from the k-th level wavelet decomposition, which represent the high-frequency components of the signal and mainly contain noise and abrupt change information; N is the total length of the signal (number of sampling points).

[0081] This formula enables adaptive calculation of the dynamic threshold, avoiding the problem of poor handling of noise of varying intensities by a fixed threshold, and providing a basis for subsequent threshold function processing. After determining the threshold, an improved soft threshold function is employed. The detail coefficients are processed.

[0082] Among them, 0.1 is the weak signal retention coefficient, which solves the problem of traditional noise reduction losing weak signals and ensures that early stress signals such as early mold growth and mild hypoxia are not filtered out. The corrected detail coefficients, This is the sign function, used to ensure that the sign of the corrected coefficients is consistent with that of the original coefficients. When When a signal is deemed valid (a sudden change or feature), it is contracted using a traditional soft thresholding method; when When a signal is identified as noise or a weak signal, traditional methods would set it to zero. However, this method retains 10% of its amplitude, ensuring that weak signals generated by early stresses such as early mold growth and mild oxygen deficiency are not filtered out, thus providing the possibility for accurate identification in the future.

[0083] Finally, by reconstructing the formula ,in, This is the final noise-reduced signal. These are the approximation coefficients after level 6 wavelet decomposition, representing the core low-frequency trend of the signal. k = 1 to 6, corresponding to the 6 levels of detail in the level 6 wavelet decomposition. The processed detail coefficients are then compared with the approximation coefficients. The final noise-reduced impedance signal and noise-reduced gas signal are obtained through reconstruction.

[0084] Next, the temperature change signal is denoised using Kalman filtering to obtain a denoised temperature change signal. Since the temperature change signal generated by seedling metabolism is low-frequency and slowly changing, Kalman filtering, through an iterative prediction-update process, can very effectively filter out random fluctuations during the measurement process and extract a smooth, true temperature change trend. The denoised temperature change signal refers to the smooth temperature signal that, after Kalman filtering, accurately reflects the changes in seed metabolic heat.

[0085] As a detailed explanation of the Kalman filtering process, its core formula is as follows: State prediction: , This is the prior state estimate (predicted value) at time k. This is the posterior state estimate at time k-1 (the optimal temperature value obtained from the previous round of filtering). That is, based on the slow-changing characteristics of the temperature signal, the predicted value at the current time is the same as the optimal estimate at the previous time. Covariance prediction: ; As time progresses, the uncertainty of the prediction increases, therefore process noise is superimposed on the covariance of the previous time step. , Let be the prior estimate of the covariance (the uncertainty of the prediction) at time k. The posterior estimate of the covariance at time k-1, The process noise covariance represents the error in the model prediction itself. Since temperature is a slowly varying signal, the model prediction error is very small, so a small value is taken.

[0086] Calculate the Kalman gain: ; This formula is used to calculate a weighting coefficient that determines whether to trust the "predicted value" or the "observation value" more during updates. This represents the Kalman gain, with a value ranging from 0 to 1. To estimate the covariance a priori, The value is 0.01, representing the observation noise covariance, which indicates the error in the sensor's temperature acquisition. Here, the value is greater than... This indicates that the system has more confidence in the model's predictions.

[0087] Status Update: ; This method uses the residuals between observed and predicted values. Multiplying the prediction by the Kalman gain, we correct the prediction to obtain the optimal estimate for the current time. This is the posterior state estimate at time k (the optimal temperature estimate that combines prediction and observation information). This represents the Kalman gain at time k. This represents the prior state estimate at time k; The temperature observation value at time k is Q=0.001 (process noise) and R=0.01 (observation noise). After filtering, the temperature fluctuation is ≤±0.05℃, ensuring that the temperature change signal can truly reflect the changes in seed metabolic heat.

[0088] Finally, based on the denoised impedance signal, denoised gas signal, and denoised temperature change signal, impedance mode features, gas mode features, and temperature change mode features are extracted, respectively. That is, based on the high-quality denoised signal, specific feature engineering calculations such as rate of change, variance, and slope are performed to form the time-series features that are finally input into the stress discrimination model.

[0089] The method provided in this invention employs an improved wavelet denoising technique to process high-frequency impedance and gas signals, and uses Kalman filtering to process slowly varying temperature signals. This effectively filters out various types of noise while preserving key weak signals and true trends related to stress occurrence to the greatest extent possible. It solves the problem that the original sensor signals are easily interfered with by environmental noise, resulting in low quality of input features for the stress discrimination model and thus misjudgment or omission. This method achieves high-fidelity preprocessing of the input data source, significantly improving the accuracy and reliability of subsequent stress identification.

[0090] Based on the above embodiments, step 110-1 includes: The original multimodal signals are collected based on multiple monitoring nodes; the monitoring nodes are deployed in monitoring trays within the seedling cultivation area; at least one mobile verification node is also deployed within the seedling cultivation area. The pre-emergence stress early warning results include the stress level; Step 120, followed by: Step 120-1: The stress level is taken as the known stress level of the monitoring acupoint plate; Step 120-2: Determine the spatial distance between the non-monitoring trays and each of the monitoring trays in the seedling cultivation area, and determine the temporal correlation coefficient used to characterize the data trend correlation between the non-monitoring trays and the monitoring trays; Step 120-3: Based on the spatial distance and the temporal correlation coefficient, determine the spatiotemporal correlation weight between the non-monitored acupoints and each of the monitored acupoints; Step 120-4: Based on the known stress level and the spatiotemporal correlation weight, construct an interpolation model to determine the stress level of the non-monitored acupoints based on the interpolation model.

[0091] Specifically, firstly, the monitoring deployment method is defined. Within the seedling cultivation area, seedling trays equipped with fixed sensors are defined as monitoring trays, and the sensor nodes on the monitoring trays are defined as monitoring nodes. Simultaneously, at least one freely movable verification node, such as an AGV equipped with a full set of sensors, is also deployed within the seedling cultivation area. All seedling trays without fixed sensors are defined as non-monitoring trays.

[0092] Secondly, the pre-emergence stress warning results include the stress level, which is used as the known stress level for the monitoring seedling trays. Then, the stress level is interpolated and inferred for any non-monitoring seedling tray. This process includes: The spatial distance between non-monitored acupoints and each monitored acupoint is determined, as well as the temporal correlation coefficient, which characterizes the correlation of data trends between non-monitored and monitored acupoints. Spatial distance refers to physical geometric distance. The temporal correlation coefficient is an indicator used to quantify the similarity of the changing trends of monitoring data from two acupoints over a past period, reflecting the correlation between non-monitored and monitored acupoints under the influence of the local microenvironment.

[0093] Based on the spatial distance and time series correlation coefficient, a spatiotemporal correlation weight is constructed between non-monitored acupoints and each monitored acupoint. This spatiotemporal correlation weight integrates both spatial proximity and time series similarity information, and its calculation mechanism follows the following principles: in the spatial dimension, the closer the distance, the higher the degree of mutual correlation; in the time dimension, the more similar the historical change trends, the stronger the potential correlation of the current state.

[0094] Based on known stress levels and spatiotemporal correlation weights, an interpolation model is constructed to determine the stress levels of non-monitored acupoints. The interpolation model is a mathematical model used to infer values ​​for unknown points based on data from a few known points.

[0095] As a specific implementation, the interpolation model in this embodiment can be an improved Kriging interpolation algorithm. Its stress level estimate... Through formula The calculation yielded the following result. This represents the stress level of the i-th monitoring acupoint, i.e., the known stress level, n=5 (fixed number of nodes). , These are the temperature gradient compensation term and the humidity gradient compensation term, respectively. The spatiotemporal correlation weights can be expressed by the formula. calculate, For spatial distance, This is the temporal correlation coefficient.

[0096] The core improvement of this interpolation model lies in the introduction of a temperature and humidity gradient compensation term on top of the traditional weighted summation. and It can be expressed by formula and calculate, ; and This is the temperature and humidity gradient compensation value. and Is it a non-monitoring acupoint plate? Real-time temperature and humidity at the location. and These represent the average temperature and average humidity within the seedling area. The temperature and humidity gradient compensation term corrects for interpolation bias caused by macroscopic temperature and humidity gradients within the greenhouse (e.g., lower temperatures near the entrance and higher temperatures in the center). Furthermore, to further correct for systematic biases caused by local environmental differences such as ventilation and lighting, an environmental drift compensation model can be introduced to perform a secondary correction on the interpolation results. The formula is as follows: ,in, Output the estimated value obtained from the improved Kriging interpolation. This is the revised final estimate of the stress level. To quantify the environmental drift coefficient of local environmental differences, This represents the drift correction factor, which corrects for errors caused by differences in greenhouse environments. Through these improvements, the interpolation results can more accurately reflect the stress state of unmonitored seed trays.

[0097] The method provided in this invention constructs an improved interpolation model that integrates the spatiotemporal correlation weights of spatial distance and temporal data correlation. This model can accurately infer the stress level of large-area unmonitored seedling trays, solving the problems of high cost of deploying sensors for full-area coverage in traditional monitoring schemes and large monitoring blind spots due to limited monitoring points. It also overcomes the shortcomings of traditional interpolation algorithms, which suffer from insufficient interpolation accuracy due to failure to consider spatiotemporal dynamic correlation and environmental heterogeneity. This method achieves high-precision, full-area stress state monitoring while controlling costs, significantly improving the practicality and economy of this early warning method in large-scale seedling cultivation scenarios.

[0098] Based on the above embodiments, the method further includes: Step 41: Based on the stress level of the non-monitored acupoints and the preset review conditions, determine the target acupoints to be reviewed; Step 42: Schedule the mobile verification node to move to the target acupoint, and collect target verification data based on the mobile verification node; Step 43: Input the target verification data into the stress discrimination model to obtain the accurate stress level output by the stress discrimination model; Step 440: If the difference between the precise stress level and the stress level of the unmonitored acupoint is greater than a preset correction threshold, the parameters of the interpolation model are corrected.

[0099] Specifically, based on the stress level of the non-monitored acupoints and preset verification conditions, target acupoints to be verified are determined. The verification conditions are a set of conditions used to trigger the mobile verification node to perform on-site verification. For example, they can be set as a stress level calculated by interpolation greater than or equal to level 3, or an uncertainty measure of the interpolation result greater than 0.3. Non-monitored acupoints that meet these verification conditions are identified as target acupoints.

[0100] Then, the mobile verification node is moved to the target acupoint, and target verification data is collected based on the mobile verification node. Finally, the target verification data is input into the stress discrimination model to obtain the precise stress level output by the stress discrimination model. The precise stress level here is directly calculated by the stress discrimination model based on the data collected in real time on site, and can be regarded as the true ground value of the target acupoint.

[0101] If the difference between the precise stress level and the stress level of the unmonitored acupoint (i.e., the interpolation level) exceeds a preset correction threshold, the parameters of the interpolation model are corrected. The preset correction threshold is a limit used to determine whether the interpolation error exceeds an acceptable range; for example, it can be set to one level. When the error exceeds the preset correction threshold, the system automatically triggers a correction of the interpolation model parameters, such as recalculating the spatiotemporal correlation weights. Or fine-tune the temperature and humidity gradient compensation coefficient. , .

[0102] As a supplementary explanation of the overall system operation strategy, this invention also includes a mechanism for dynamic early warning, precise intervention, and data closure. Firstly, the system can, based on seedling variety, age, and environmental parameters, use a dynamic threshold function... Set an adaptive warning threshold, where (Basic threshold) This is the variety correction factor. This is the age correction factor. This is the environmental correction factor.

[0103] Based on the comparison between the stress level and the threshold, a three-level early warning system of mild, moderate and severe is implemented to ensure the accuracy and relevance of the early warning.

[0104] (1) Mild warning (1.5–Th): Mild stress, push notification information, no intervention required, monitor once every 30 minutes; (2) Moderate warning (Th-3.5 level): Moderate stress, push warning information, automatically trigger mild regulation, monitor once every 10 minutes; (3) Severe warning (≥3.5 level): Severe / extremely severe stress, sound and light alarm + SMS notification, automatic emergency control, AGV immediately review, level ≥5 mark requires replay.

[0105] Secondly, based on the type and level of stress detected by the warning, the system can automatically trigger corresponding precise intervention measures. For example, for level 1-2 seed rot stress, ventilation can be strengthened; for level 3-5 mold stress, substrate replacement can be implemented. Finally, data from all stages, including monitoring data, warning information, intervention measures, and subsequent seedling emergence results, are stored in the database, forming a closed-loop data system covering the entire lifecycle. This data can not only be used for quality traceability, but more importantly, it will be screened and labeled for continuous iterative optimization of the stress discrimination model and interpolation model (e.g., fine-tuning every 15 days), thus forming a virtuous cycle of "data-model-optimization-application". In addition, to ensure the generalization ability and high accuracy of the stress discrimination model itself, its initial training is based on a large-scale dataset of over 100,000 multimodal time-series data points from 4 crops, 3 bases, and different temperature and humidity environments. The model is trained using the Adam optimizer and a hybrid loss function, and optimized through strategies such as early stopping, Dropout, and transfer learning, thereby ensuring the robustness of the model in multi-scenario applications.

[0106] The method provided in this invention actively identifies high-risk or high-uncertainty interpolation points by setting verification conditions and scheduling mobile verification nodes to perform targeted on-site precise measurements. Then, by comparing the difference between the interpolation results and the precise measurement results, the parameters of the interpolation model are corrected by feedback when the difference exceeds a preset correction threshold. This establishes a closed-loop control process of "inference-verification-correction," which solves the inference error and uncertainty problems that cannot be completely avoided by any interpolation algorithm. It realizes dynamic verification of interpolation results and continuous self-optimization of the interpolation model, significantly enhancing the long-term stability of the entire stress monitoring system and the reliability of the final output results.

[0107] Based on the above embodiments, the impedance mode characteristics include at least one of impedance change rate and impedance variance; the gas mode characteristics include at least one of carbon dioxide concentration gradient, characteristic peak intensity of volatile organic compounds, and correlation coefficient between carbon dioxide and volatile organic compounds; the temperature change mode characteristics include at least one of temperature change slope, temperature anomaly degree, and temperature change variance.

[0108] Specifically, impedance modal characteristics may include at least one of impedance change rate and impedance variance. The impedance change rate quantifies the relative rate of change of impedance values ​​between adjacent time points, and it sensitively reflects trends caused by rapid changes in salt concentration or matrix water content. The formula for calculating the impedance change rate is: .

[0109] in, In order to be in The rate of change of impedance at time t; In order to be in The noise reduction impedance signal value at any given time; For the previous sampling time The noise reduction impedance signal value.

[0110] Here, impedance variance is used to measure the degree of fluctuation or dispersion of the impedance signal within a time window. Impedance variance can reflect the stability of the matrix environment.

[0111] The formula for calculating impedance variance is as follows: .

[0112] in, For impedance variance; Within the time window The noise reduction impedance signal value at any given time; This is the average value of all noise reduction impedance signal values ​​within this time window.

[0113] Here, the gas modal characteristics include at least one of the following: carbon dioxide concentration gradient, characteristic peak intensity of volatile organic compounds, and correlation coefficient between carbon dioxide and volatile organic compounds.

[0114] Among them, the carbon dioxide concentration gradient is used to reflect the rate of change of carbon dioxide concentration, which is directly related to the strength of seed respiration.

[0115] The formula for calculating the carbon dioxide concentration gradient is: ; in, In order to be in The carbon dioxide concentration gradient at time t; and These are the noise-reduced carbon dioxide concentration signal values ​​at the current and previous moments, respectively; the denominator 5 represents the sampling time interval (e.g., 5 minutes).

[0116] Here, the characteristic peak intensity of volatile organic compounds is used to capture the instantaneous burst of volatile organic compounds caused by microbial activities such as mold growth.

[0117] The formula for calculating the intensity of characteristic peaks of volatile organic compounds is: ; in, In order to be in The intensity of characteristic peaks of volatile organic compounds at any given time; The current value of the denoising VOCs concentration signal; This represents the baseline average concentration over a period of time.

[0118] Here, the correlation coefficient between carbon dioxide and volatile organic compounds (VOCs) is used to quantify the synchronicity of the changing trends in the concentrations of these two gases. It effectively distinguishes between normal respiration (increased CO2, unchanged VOCs) and mold stress (where both CO2 and VOCs may increase simultaneously). The formula for calculating the correlation coefficient between carbon dioxide and VOCs is: .

[0119] in, The correlation coefficient between carbon dioxide and volatile organic compounds; and These are the time series of noise-reduced signals for carbon dioxide and volatile organic compounds, respectively. and These are the average values ​​of the time series of the noise-reduced signals for carbon dioxide and volatile organic compounds, respectively.

[0120] Here, the temperature variation modal characteristics include at least one of temperature variation slope, temperature anomaly degree, and temperature variation variance.

[0121] The temperature change slope reflects the rate of temperature change in a micro-region and is related to the rate of heat production through seed metabolism. The formula for calculating the temperature change slope is: in, In order to be in The slope of the temperature change at any given time; and These are the noise-reduced temperature-varying signal values ​​at the current moment and the previous moment, respectively.

[0122] Here, temperature anomaly is used to measure the degree to which the current temperature deviates from its recent normal fluctuation range, and can effectively identify temperature abrupt changes caused by metabolic disorders. The formula for calculating temperature anomaly is: ; in, Temperature anomaly; This represents the noise-reduced temperature-varying signal value at the current moment; This represents the temperature variance within the window.

[0123] Here, temperature variance is used to measure the fluctuation of temperature signals within a time window, reflecting the stability of seed metabolic activities. The formula for calculating temperature variance is: ; in, Variance due to temperature variation; Within the time window The noise reduction temperature-varying signal value at any given moment; This is the average value of all noise-reduced temperature-varying signal values ​​within this time window.

[0124] After extracting all the features mentioned above, to eliminate the influence of different units and numerical ranges between features and ensure that the model treats all features equally during training, this embodiment also includes a standardization step. Specifically, the Min-Max standardization method can be used, through the formula... Map all extracted feature values ​​f to the interval [0,1]. These are the standardized eigenvalues. and These are the minimum and maximum values ​​of the feature in the training dataset, respectively.

[0125] Based on any of the above embodiments Figure 2 This is the second flowchart of the pre-emergence stress early warning method for seedlings provided by the present invention, as shown below. Figure 2 As shown, this method begins with signal acquisition at the bottom layer. Within the monitoring unit of the seedling tray substrate, sensors collect multimodal raw signals such as impedance, CO2 concentration, VOCs concentration, and micro-area temperature. These multimodal raw signals then enter a signal denoising module. Impedance and gas signals are processed using an improved wavelet denoising algorithm, while temperature variation signals are denoised using Kalman filtering to eliminate noise interference. The denoised signals are then fed into a multidimensional feature extraction and standardization module, from which eight core time-series features are extracted, including impedance change rate, impedance variance, temperature slope, temperature variance, temperature anomaly, CO2 concentration gradient, CO2-VOCs correlation coefficient, and VOCs characteristic peak intensity, and then standardized. These features constitute the input layer of the TAMIFN (Temporal Attention-Modal Interaction Fusion Network) model, whose input matrix is ​​strictly defined as… It is suitable for standard acupuncture point monitoring scenarios. Among them, the impedance mode timing feature matrix... Includes ΔZ, Two features reflecting matrix salinity and water content; gas mode time series feature matrix. Includes dC / dt, and Three features directly correlate with seed respiration and mold state; temperature variation modal temporal feature matrix Includes dT / dt, and Three features reflect the dynamic changes in seed metabolic heat. The input matrix underwent noise reduction preprocessing and standardization (values ​​range [0,1]) to ensure the effectiveness and weight balance of the input features. Within the TAMIFN model, the input features are first encoded using a single-modal temporal encoding layer, extracting features with dimension [0,1]. deep features , , Subsequently, the temporal attention layer weights the encoded features and extracts weighted features focusing on key temporal points. , , Next, the modal interaction fusion layer performs deep interaction and fusion on the weighted features to generate fused features. Finally, the output layer outputs the stress type, level, and confidence score based on the fused features. To ensure the generalization ability and high accuracy of the TAMIFN model, its training is based on a large-scale dataset of over 100,000 multimodal time-series data points from four crops, three bases, and different temperature and humidity environments. The Adam optimizer, a hybrid loss function (cross-entropy loss + mean squared error loss), and optimization strategies including early stopping, Dropout layers, transfer learning, and dynamic iteration with data closure loops are employed. The stress level output by the model for monitored planting pits is used as a known sample and input into an improved Kriging interpolation algorithm + environmental drift compensation model to determine the stress level of non-monitored planting pits. If the interpolation result has an error of greater than or equal to one level compared to the subsequent AGV verification result, the interpolation model will be optimized. Ultimately, based on the identification and inference results, the entire system pushes corresponding early warning messages and intervention measures to users, forming a complete and automated stress early warning process from data collection, noise reduction, feature extraction, intelligent identification, spatial inference to closed-loop review and optimization.

[0126] The pre-emergence stress early warning device for seedlings provided by the present invention is described below. The pre-emergence stress early warning device for seedlings described below can be referred to in correspondence with the pre-emergence stress early warning method for seedlings described above.

[0127] Based on any of the above embodiments, the present invention provides a seedling pre-emergence stress early warning device. Figure 3 This is a schematic diagram of the pre-emergence stress early warning device for seedlings provided by the present invention, as shown below. Figure 3As shown, the device includes: The acquisition module 310 is used to acquire multimodal temporal characteristics of the seedling cultivation environment; the multimodal temporal characteristics include impedance mode characteristics, gas mode characteristics, and temperature change mode characteristics; The input module 320 is used to input the multimodal temporal features into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model; The stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal coding layer is used to independently encode the impedance modal features, the gas modal features, and the temperature-varying modal features to obtain impedance modal coding features, gas modal coding features, and temperature-varying modal coding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal coding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal coding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal coding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

[0128] The apparatus provided in this invention acquires multimodal temporal features of the seedling cultivation environment, including impedance mode features, gas mode features, and temperature change mode features; inputs the multimodal temporal features into a stress discrimination model to obtain pre-emergence stress warning results. The stress discrimination model includes a single-modal temporal encoding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal encoding layer independently encodes each modal feature; the temporal attention layer performs temporal weighting on the encoded modal features to focus on key temporal information; and the modal interaction fusion layer interacts and fuses the weighted modal features to generate pre-emergence stress warning results. This invention, based on acquired impedance, gas, and temperature-varying multimodal temporal features, independently encodes each modal feature through a single-modal temporal coding layer to preserve its unique temporal patterns. Then, a temporal attention layer automatically identifies and focuses on the most critical temporal information for stress assessment from the temporal sequences of each modality. Furthermore, a modal interaction fusion layer deeply mines the synergistic and complementary relationships between different modal features. This allows for accurate identification of pre-emergence stress states from complex, multi-source weak signals, solving the problems of existing technologies that rely on single signals, are susceptible to interference, have low identification accuracy, and use simple fusion algorithms that fail to effectively utilize temporal dynamics and cross-modal correlation information, resulting in delayed warnings and low accuracy. This invention achieves early and accurate warnings of latent stress before seedling emergence, significantly improving the timeliness and accuracy of pre-emergence stress warnings.

[0129] Based on any of the above embodiments, the step of performing intermodal interaction and fusion of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature-varying mode time-weighted features to generate the pre-emergence stress early warning result includes: Construct a modal interaction matrix; the modal interaction matrix is ​​used to characterize the correlation between any two of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature-varying mode time-weighted features; Based on the modal interaction matrix, cross-modal attention weights are determined; the cross-modal attention weights include a first weight, a second weight, and a third weight. Based on the first weight, the second weight, and the third weight, the impedance mode time-series weighted features, the gas mode time-series weighted features, and the temperature change mode time-series weighted features are weighted and fused respectively to generate the pre-emergence stress early warning result.

[0130] Based on any of the above embodiments, the single-modal temporal coding layer includes dilated convolution units and residual connection units; The dilated convolution unit is used to perform convolution operations with hierarchically increasing dilation coefficients on the impedance mode features, the gas mode features, and the temperature variation mode features, respectively, to obtain deep impedance time series information, deep gas time series information, and deep temperature variation time series information. The residual connection unit is used to perform a residual connection between the impedance deep timing information and the impedance mode feature to obtain the impedance mode coding feature, to perform a residual connection between the gas deep timing information and the gas mode feature to obtain the gas mode coding feature, and to perform a residual connection between the temperature variation deep timing information and the temperature variation mode feature to obtain the temperature variation mode coding feature.

[0131] Based on any of the above embodiments, the acquisition module 310 specifically includes: The acquisition module is used to acquire raw multimodal signals; the raw multimodal signals include impedance signals, gas signals, and temperature change signals. The wavelet decomposition module is used to perform wavelet decomposition on the impedance signal and the gas signal to obtain detail coefficients characterizing the high-frequency components. A threshold determination module is used to determine an adaptive dynamic threshold based on the statistical characteristics of the detail coefficients; The correction coefficient module is used to determine the corrected detail coefficients based on the detail coefficients and the adaptive dynamic threshold; The wavelet reconstruction module is used to perform wavelet reconstruction based on the corrected detail coefficients to obtain the noise-reduced impedance signal and the noise-reduced gas signal. The noise reduction processing module is used to perform noise reduction processing on the temperature change signal based on Kalman filtering to obtain a noise-reduced temperature change signal. The extraction module is used to extract the impedance mode features, the gas mode features, and the temperature change mode features based on the noise-reduced impedance signal, the noise-reduced gas signal, and the noise-reduced temperature change signal, respectively.

[0132] Based on any of the above embodiments, the acquisition module is specifically used for: The original multimodal signals are collected based on multiple monitoring nodes; the monitoring nodes are deployed in monitoring trays within the seedling cultivation area; at least one mobile verification node is also deployed within the seedling cultivation area. The pre-emergence stress early warning results include the stress level; It also includes a coercion level determination module, which is specifically used for: The stress level is taken as the known stress level of the monitoring acupoint plate; Determine the spatial distance between the non-monitored seedling trays and each of the monitored seedling trays in the seedling cultivation area, and determine the temporal correlation coefficient used to characterize the data trend correlation between the non-monitored seedling trays and the monitored seedling trays; Based on the spatial distance and the temporal correlation coefficient, the spatiotemporal correlation weights between the non-monitored acupoints and each of the monitored acupoints are determined; Based on the known stress level and the spatiotemporal correlation weight, an interpolation model is constructed to determine the stress level of the non-monitored acupoint.

[0133] Based on any of the above embodiments, a parameter correction module is further included, wherein the parameter correction module is specifically used for: Based on the stress level of the non-monitored acupoints and the preset review conditions, the target acupoints to be reviewed are determined. The mobile verification node is scheduled to move to the target acupoint, and target verification data is collected based on the mobile verification node; The target verification data is input into the stress discrimination model to obtain the precise stress level output by the stress discrimination model; If the difference between the precise stress level and the stress level of the unmonitored acupoint is greater than a preset correction threshold, the parameters of the interpolation model are corrected.

[0134] Based on any of the above embodiments, the impedance mode characteristics include at least one of impedance change rate and impedance variance; the gas mode characteristics include at least one of carbon dioxide concentration gradient, characteristic peak intensity of volatile organic compounds, and correlation coefficient between carbon dioxide and volatile organic compounds; the temperature change mode characteristics include at least one of temperature change slope, temperature anomaly degree, and temperature change variance.

[0135] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a pre-emergence stress early warning method for seedlings. This method includes: acquiring multimodal temporal features of the seedling cultivation environment; the multimodal temporal features include impedance mode features, gas mode features, and temperature change mode features; inputting the multimodal temporal features into a stress discrimination model to obtain a pre-emergence stress early warning result for seedlings output by the stress discrimination model; the stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer; the single-modal temporal coding layer is used to independently encode the impedance mode features, the gas mode features, and the temperature change mode features to obtain... The system comprises impedance mode coding features, gas mode coding features, and temperature-varying mode coding features. The temporal attention layer is used to determine, based on the impedance mode coding features, impedance mode temporal weighting features focusing on the first key temporal information; based on the gas mode coding features, gas mode temporal weighting features focusing on the second key temporal information; and based on the temperature-varying mode coding features, temperature-varying mode temporal weighting features focusing on the third key temporal information. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance mode temporal weighting features, the gas mode temporal weighting features, and the temperature-varying mode temporal weighting features to generate the pre-emergence stress early warning result.

[0136] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0137] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the pre-emergence stress early warning method for seedlings provided by the above methods. The method includes: acquiring multimodal temporal features of the seedling cultivation environment; the multimodal temporal features include impedance mode features, gas mode features, and temperature change mode features; inputting the multimodal temporal features into a stress discrimination model to obtain the pre-emergence stress early warning result of the seedlings output by the stress discrimination model; the stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer; the single-modal temporal coding layer is used to process the impedance... The modal features, gas modal features, and temperature-varying modal features are independently encoded to obtain impedance modal encoding features, gas modal encoding features, and temperature-varying modal encoding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal encoding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal encoding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal encoding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

[0138] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the pre-emergence stress early warning method for seedlings provided by the methods described above. This method includes: acquiring multimodal temporal features of the seedling cultivation environment; the multimodal temporal features include impedance mode features, gas mode features, and temperature change mode features; inputting the multimodal temporal features into a stress discrimination model to obtain a pre-emergence stress early warning result for seedlings output by the stress discrimination model; the stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer; the single-modal temporal coding layer is used to process the impedance mode features, the gas mode features, and the temperature change mode features. The temperature-varying modal features are independently encoded to obtain impedance mode encoding features, gas mode encoding features, and temperature-varying modal encoding features. The temporal attention layer is used to determine the impedance mode temporal weighted features focusing on the first key temporal information based on the impedance mode encoding features, to determine the gas mode temporal weighted features focusing on the second key temporal information based on the gas mode encoding features, and to determine the temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal encoding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance mode temporal weighted features, the gas mode temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

[0139] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0140] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of stress before seedling emergence, characterized in that, include: The multimodal temporal characteristics of the seedling cultivation environment are obtained; the multimodal temporal characteristics include impedance mode characteristics, gas mode characteristics, and temperature change mode characteristics. The multimodal temporal features are input into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model; The stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal coding layer is used to independently encode the impedance modal features, the gas modal features, and the temperature-varying modal features to obtain impedance modal coding features, gas modal coding features, and temperature-varying modal coding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal coding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal coding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal coding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

2. The seedling pre-emergence stress early warning method according to claim 1, characterized in that, The process of performing intermodal interaction and fusion of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature-varying mode time-weighted features to generate the pre-emergence stress early warning result includes: Construct a modal interaction matrix; the modal interaction matrix is ​​used to characterize the correlation between any two of the impedance mode time-weighted features, the gas mode time-weighted features, and the temperature-varying mode time-weighted features; Based on the modal interaction matrix, cross-modal attention weights are determined; the cross-modal attention weights include a first weight, a second weight, and a third weight. Based on the first weight, the second weight, and the third weight, the impedance mode time-series weighted features, the gas mode time-series weighted features, and the temperature change mode time-series weighted features are weighted and fused respectively to generate the pre-emergence stress early warning result.

3. The seedling pre-emergence stress early warning method according to claim 1, characterized in that, The single-modal temporal coding layer includes dilated convolutional units and residual connection units; The dilated convolution unit is used to perform convolution operations with hierarchically increasing dilation coefficients on the impedance mode features, the gas mode features, and the temperature variation mode features, respectively, to obtain deep impedance time series information, deep gas time series information, and deep temperature variation time series information. The residual connection unit is used to perform a residual connection between the impedance deep timing information and the impedance mode feature to obtain the impedance mode coding feature, to perform a residual connection between the gas deep timing information and the gas mode feature to obtain the gas mode coding feature, and to perform a residual connection between the temperature variation deep timing information and the temperature variation mode feature to obtain the temperature variation mode coding feature.

4. The seedling pre-emergence stress early warning method according to any one of claims 1 to 3, characterized in that, The acquisition of multimodal temporal features of the seedling cultivation environment includes: Acquire raw multimodal signals; the raw multimodal signals include impedance signals, gas signals, and temperature change signals; Wavelet decomposition is performed on the impedance signal and the gas signal to obtain detail coefficients characterizing the high-frequency components; The adaptive dynamic threshold is determined based on the statistical characteristics of the detail coefficients; The corrected detail coefficients are determined based on the detail coefficients and the adaptive dynamic threshold. Wavelet reconstruction is performed based on the corrected detail coefficients to obtain the noise-reduced impedance signal and the noise-reduced gas signal; The temperature change signal is denoised using Kalman filtering to obtain a denoised temperature change signal. Based on the noise-reduced impedance signal, the noise-reduced gas signal, and the noise-reduced temperature change signal, the impedance mode features, the gas mode features, and the temperature change mode features are extracted respectively.

5. The seedling pre-emergence stress early warning method according to claim 4, characterized in that, The acquisition of raw multimodal signals includes: The original multimodal signals are collected based on multiple monitoring nodes; the monitoring nodes are deployed in monitoring trays within the seedling cultivation area; at least one mobile verification node is also deployed within the seedling cultivation area. The pre-emergence stress early warning results include the stress level; The step of inputting the multimodal temporal features into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model further includes: The stress level is taken as the known stress level of the monitoring acupoint plate; Determine the spatial distance between the non-monitored seedling trays and each of the monitored seedling trays in the seedling cultivation area, and determine the temporal correlation coefficient used to characterize the data trend correlation between the non-monitored seedling trays and the monitored seedling trays; Based on the spatial distance and the temporal correlation coefficient, the spatiotemporal correlation weights between the non-monitored acupoints and each of the monitored acupoints are determined; Based on the known stress level and the spatiotemporal correlation weight, an interpolation model is constructed to determine the stress level of the non-monitored acupoint.

6. The seedling pre-emergence stress early warning method according to claim 5, characterized in that, The method further includes: Based on the stress level of the non-monitored acupoints and the preset review conditions, the target acupoints to be reviewed are determined. The mobile verification node is scheduled to move to the target acupoint, and target verification data is collected based on the mobile verification node; The target verification data is input into the stress discrimination model to obtain the precise stress level output by the stress discrimination model; If the difference between the precise stress level and the stress level of the unmonitored acupoint is greater than a preset correction threshold, the parameters of the interpolation model are corrected.

7. The seedling pre-emergence stress early warning method according to any one of claims 1 to 3, characterized in that, The impedance mode characteristics include at least one of impedance change rate and impedance variance; the gas mode characteristics include at least one of carbon dioxide concentration gradient, characteristic peak intensity of volatile organic compounds, and correlation coefficient between carbon dioxide and volatile organic compounds; the temperature change mode characteristics include at least one of temperature change slope, temperature anomaly degree, and temperature change variance.

8. A seedling pre-emergence stress early warning device, characterized in that, include: The acquisition module is used to acquire multimodal temporal characteristics of the seedling cultivation environment; the multimodal temporal characteristics include impedance mode characteristics, gas mode characteristics, and temperature change mode characteristics; The input module is used to input the multimodal temporal features into the stress discrimination model to obtain the pre-emergence stress warning result of the seedlings output by the stress discrimination model; The stress discrimination model includes a single-modal temporal coding layer, a temporal attention layer, and a modal interaction fusion layer. The single-modal temporal coding layer is used to independently encode the impedance modal features, the gas modal features, and the temperature-varying modal features to obtain impedance modal coding features, gas modal coding features, and temperature-varying modal coding features. The temporal attention layer is used to determine impedance modal temporal weighted features focusing on the first key temporal information based on the impedance modal coding features, to determine gas modal temporal weighted features focusing on the second key temporal information based on the gas modal coding features, and to determine temperature-varying modal temporal weighted features focusing on the third key temporal information based on the temperature-varying modal coding features. The modal interaction fusion layer is used to perform intermodal interaction and fusion of the impedance modal temporal weighted features, the gas modal temporal weighted features, and the temperature-varying modal temporal weighted features to generate the pre-emergence stress early warning result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the seedling pre-emergence stress early warning method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the seedling pre-emergence stress early warning method as described in any one of claims 1 to 7.

Images