Multi-sensor fusion grape tree abnormal growth early warning method

Abstract

The application discloses a multi-sensor fusion grapevine growth anomaly early warning method, and relates to the field of intelligent agriculture and agricultural information technology.The method comprises the following steps: collecting original plant time series signals and environment time series signals through a first type of sensor and a second type of sensor arranged in a target environment; processing the original plant time series signals based on a feature engineering method, fusing the processing result with the environment time series signals, and constructing a multi-dimensional feature matrix; obtaining historical normal growth cycle data, corresponding to constructing and training a dynamic baseline model based on a long short-term memory neural network, and determining a graded early warning threshold; inputting the multi-dimensional feature matrix into the dynamic baseline model, and combining the graded early warning threshold to perform growth anomaly early warning.The application effectively improves the accuracy and real-time performance of grapevine growth anomaly early warning.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and agricultural information technology, specifically to a method for early warning of abnormal growth in grapevines using multi-sensor fusion. Background Technology

[0002] With the rapid development of precision agriculture and agricultural information technology, utilizing multiple sensors to collect environmental and crop data, and then using data fusion and machine learning models to assess growth status and provide early warnings of anomalies, has become an important direction for improving the intelligence of orchard management. Existing technologies mainly include two categories: morphological monitoring based on vision / remote sensing and stress inference based on environmental sensing. These methods aim to indirectly determine the plant's growth status through image feature extraction or environmental factor analysis.

[0003] However, existing technologies still face many limitations in practical applications. Visual or remote sensing-based methods struggle to effectively isolate the inherent physiological state of grapevines from the background interference of the trellis structure, resulting in weak model generalization ability and significantly limited warning accuracy due to scene constraints. Environmental sensing-based stress inference techniques cannot directly perceive plant physiological responses, relying on indirect environmental indicators, leading to high false alarm and false negative rates. Furthermore, existing methods often employ static or simple time-series models, failing to fully consider the dynamic processes and lag effects of environmental stress on plant status, thus limiting the accuracy and real-time performance of warnings. Summary of the Invention

[0004] This invention provides a method for early warning of abnormal growth in grapevines using multi-sensor fusion, aiming to solve the technical problems of insufficient accuracy and real-time performance in early warning of abnormal growth in grapevines in the prior art.

[0005] In view of the above problems, the present invention provides a method for early warning of abnormal growth in grapevines using multi-sensor fusion, comprising: Original plant time-series signals and environmental time-series signals were collected by first-type and second-type sensors deployed in the target environment. The original plant time-series signal is processed using feature engineering methods, and the processing results are fused with the environmental time-series signal to construct a multidimensional feature matrix. Acquire historical normal growth cycle data based on big data, construct and train a dynamic baseline model based on long short-term memory neural network, and determine the graded early warning threshold; The multidimensional feature matrix is ​​input into the dynamic baseline model, and growth anomaly warning is given in combination with the hierarchical warning threshold.

[0006] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a multi-sensor fusion method for early warning of abnormal grapevine growth. It accurately acquires intrinsic physiological temporal signals and environmental signals from multiple sources through synchronous data acquisition, laying a high-quality data foundation for subsequent analysis. Through deep feature engineering extraction and optimization, irrelevant interference is effectively removed, and the representation of the plant's core physiological laws is strengthened. A dynamic baseline model constructed based on a long short-term memory neural network accurately captures the temporal correlation and lag effect between environmental and physiological responses, forming a dynamic standard that conforms to normal growth patterns. Combined with tiered early warning thresholds, real-time anomaly identification and tiered push notifications are achieved, ultimately effectively improving the accuracy, real-time performance, and cross-scenario adaptability of early warning for abnormal grapevine growth, providing efficient and reliable technical support for refined orchard management. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart illustrating the multi-sensor fusion method for early warning of abnormal grapevine growth provided in an embodiment of the present invention. Detailed Implementation

[0009] This invention provides a multi-sensor fusion method for early warning of abnormal growth in grapevines, which addresses the technical problems of insufficient accuracy and real-time performance in early warning of abnormal growth in grapevines in existing technologies.

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0011] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0012] Examples, such as Figure 1 As shown, this invention provides a method for early warning of abnormal grapevine growth based on multi-sensor fusion, the method comprising: S100: Collects original plant time-series signals and environmental time-series signals through first-class and second-class sensors deployed in the target environment.

[0013] In this embodiment of the invention, first and second type sensors deployed in the target environment are used to collect raw plant time-series signals and environmental time-series signals. The plant's growth state is dynamically influenced by the coupling of its own physiological rhythms and environmental factors, and the environment's response to physiological changes exhibits a lag effect. Furthermore, the raw sensor data is susceptible to interference from high-frequency noise, spatiotemporal asynchrony, and data gaps. Directly using the raw data for analysis can lead to distorted feature extraction and biased model judgments, failing to accurately reflect the intrinsic physiological state of the grapevine. Therefore, a comprehensive processing flow involving classification and acquisition, synchronous control, noise reduction, temporal alignment, and data normalization is required to ensure the integrity, synchronization, and effectiveness of the data, laying a high-quality data foundation for subsequent feature engineering and model training.

[0014] Step S100 in the method provided in this embodiment of the invention includes: According to the preset spatiotemporal window, the first type of sensor and the second type of sensor are controlled to synchronously collect data to obtain the first time-series signal and the second time-series signal; The first time-series signal acquired is subjected to filtering and noise reduction processing; Based on statistical analysis methods, the physiological response lag time window is determined, and the second time-series signal is time-aligned with the first time-series signal after filtering and noise reduction. Missing values ​​are imputed based on the time alignment results, and the data is uniformly resampled to a preset time resolution. The output consists of the original plant time series signal and the environmental time series signal.

[0015] The first type of sensor is deployed on the grapevine itself, and the second type of sensor is deployed in the grapevine's growing environment. The first type of sensor includes at least an inertial sensor and a diameter growth sensor.

[0016] First, according to a preset spatiotemporal window, the first type of sensor and the second type of sensor are controlled to synchronously collect data, acquiring a first time-series signal and a second time-series signal. The first type of sensor is deployed on the grapevine itself, while the second type of sensor is deployed in the grapevine's growing environment. The first type of sensor includes at least an inertial sensor and a diameter growth sensor. The preset spatiotemporal window refers to the data collection time and spatial range pre-set according to agronomic knowledge. Temporally, it typically covers a complete growing season or key phenological periods, such as budding or fruit enlargement; spatially, it refers to multiple monitoring points within the vineyard selected using statistical methods that represent the overall condition of the vineyard. According to the preset spatiotemporal window, through a Network Time Protocol (NTP) or hardware clock synchronization mechanism, the first type of sensor and the second type of sensor are controlled to simultaneously start data acquisition, respectively acquiring a first time-series signal reflecting the vine's own state and a second time-series signal reflecting the growing environment. The first type of sensor is placed on the grapevine itself, such as at the base of the main vine, while the second type of sensor is placed in the growing environment, such as the soil and canopy. The first and second types of sensors begin recording data at the same time stamp.

[0017] For example, at five monitoring points in a Cabernet Sauvignon vineyard, on the main vines of three target plants at each monitoring point, IMU sensors and stem diameter growth sensors simultaneously collected triaxial acceleration, angular velocity, and stem diameter micro-change data to obtain the first time-series signal. At the same time, soil water potential / temperature sensors at a depth of 40cm / 60cm and a canopy weather station simultaneously collected soil water potential, air temperature and humidity, and photosynthetically active radiation data to obtain the second time-series signal. All data were received synchronously through a Zigbee gateway, and the timestamps were uniformly calibrated to Beijing time.

[0018] Secondly, the acquired first time-series signal is filtered and denoised. The first time-series signal specifically refers to the raw signal acquired from the plant's own sensors, which typically contains low-frequency useful signals reflecting physiological activities and high-frequency noise from mechanical vibrations, electromagnetic interference, etc. To address the high-frequency noise in the first time-series signal, a Butterworth low-pass filter is selected, with a reasonable cutoff frequency set such as 5Hz, to filter the raw signal, remove irrelevant interference, and retain the effective signal components reflecting the plant's physiological state.

[0019] For example, the raw triaxial acceleration data acquired by the IMU sensor contains 12Hz high-frequency vibration noise caused by wind. By using a Butterworth low-pass filter with a cutoff frequency of 5Hz, the raw data is processed to filter out the 12Hz high-frequency interference and retain the low-frequency signal of 0.1-2Hz. This low-frequency signal corresponds to the physiological activities of the plant stem, such as daily contraction and expansion, and micro-movements during growth.

[0020] Furthermore, based on statistical analysis methods, a physiological response lag window is determined, and the second time-series signal is time-aligned with the filtered and denoised first time-series signal. The physiological response of plants to environmental changes is delayed; for example, after soil water shortage, stem diameter shrinkage takes 2-3 hours to manifest. The specific delay duration obtained through statistical analysis is the physiological response lag window. Time alignment refers to adjusting the timestamp of the second time-series signal according to the physiological response lag window, so that the environmental change signal and the corresponding plant physiological response signal are on the same time dimension, ensuring the accuracy of the causal relationship between the two. Statistical analysis methods are algorithms used to calculate the correlation between environmental factors and physiological responses, such as Pearson correlation analysis, the core of which is to find the delay duration when the correlation between the environmental signal and the physiological signal is highest. Using Pearson correlation analysis or mutual information entropy analysis, the correlation between each environmental factor in the second time-series signal and the physiological indicators in the filtered first time-series signal is calculated to determine the physiological response lag window; the timestamp of the second time-series signal is shifted backward by this duration to achieve precise alignment with the first time-series signal.

[0021] For example, using Pearson correlation analysis, the correlation between the soil water potential at 40cm depth in the second time-series signal and the stem diameter shrinkage in the filtered first time-series signal was calculated: when the soil water potential signal timestamp was shifted forward by 2 hours, the correlation coefficient r = 0.83, indicating the highest correlation coefficient between the two signals. Therefore, the physiological response lag time window was determined to be 2 hours. The timestamps of all second-series signals were uniformly shifted forward by 2 hours, ensuring that the soil water shortage signal at 10:00 and the stem shrinkage response signal at 12:00 were consistent, thus completing the time-series alignment.

[0022] Finally, missing values ​​are imputed based on the time-series alignment results, and the data is uniformly resampled to a preset time resolution. The output consists of the original plant time-series signal and the environmental time-series signal. Missing values ​​refer to a small number of data gaps caused by temporary communication interruptions or unstable power supply of the sensors, typically 1-2 consecutive data points. Linear interpolation is used to calculate the estimated value of the missing position based on the valid data points before and after the missing value, ensuring data continuity. Uniform resampling refers to converting signals from different acquisition frequencies into a uniform format with a preset time resolution, such as 1 hour per sample, to facilitate subsequent feature extraction and model input. For the time-series aligned data, a small number of missing values ​​are imputed using linear interpolation; then, according to the preset time resolution, the first and second time-series signals are uniformly resampled, such as segmented by 1 hour, and the mean or median of each segment is taken to output the normalized original plant time-series signal and the environmental time-series signal.

[0023] For example, after time-series alignment, it was found that the soil water potential data at 40cm depth at a certain monitoring point at 14:00 was missing. The valid data before and after it were 25kPa at 13:00 and 23kPa at 15:00. The missing value at 14:00 was calculated to be 24kPa through linear interpolation. Subsequently, uniform resampling was performed: the data after IMU filtering was averaged in 1-hour segments, and the environmental sensor data was directly integrated in 1-hour segments. Finally, the original plant time-series signal and environmental time-series signal with continuous timestamps and uniform resolution were output.

[0024] In this embodiment of the invention, by classifying and collecting multi-source signals from the plant itself and the environment, and combining synchronous control, low-pass filtering, temporal alignment, and data normalization, the problems of noise interference, spatiotemporal asynchrony, lag mismatch, and data loss in the original data are effectively solved. The output data not only retains the core effective information of plant physiological activities, but also establishes a precise causal relationship between environmental factors and physiological responses, and has the characteristics of temporal continuity and dimensional uniformity. This provides high-quality and highly reliable data support for subsequent extraction of intrinsic physiological characteristics without trellis interference and construction of a precise dynamic baseline model.

[0025] S200: The original plant time-series signal is processed based on the feature engineering method, and the processing result is fused with the environmental time-series signal to construct a multi-dimensional feature matrix.

[0026] In this embodiment of the invention, the original plant time-series signal is processed using feature engineering methods, and the processed result is fused with the environmental time-series signal to construct a multi-dimensional feature matrix. The original plant time-series signal is only raw data collected by sensors, such as acceleration and stem diameter values, which do not directly represent the core physiological state of the grapevine and contain redundant information; at the same time, a single physiological signal or environmental signal cannot reflect the coupling relationship between the environment and physiology. If directly input into the model, it will lead to low model training efficiency, weak feature representation ability, and difficulty in accurately depicting normal growth rhythms. Therefore, it is necessary to first screen key physiological rhythm features through historical data, and then extract real-time physiological features based on this list and fuse them with environmental signals to construct a multi-dimensional feature matrix that has both physiological significance and coupling correlation, providing efficient and effective input data for the dynamic baseline model.

[0027] Step S200 in the method provided in this embodiment of the invention includes: Specifically, the original plant time-series signal is processed using feature engineering methods, and the processed result is fused with the environmental time-series signal to construct a multi-dimensional feature matrix. Prior to this, the process includes: During the normal growth cycle of the grapevines, sample time-series signals obtained by the first type of sensor and the second type of sensor were collected simultaneously. Based on the sample time-series signal, multiple candidate rhythm features representing different physiological dimensions are calculated and obtained, forming a candidate rhythm feature set; Principal component analysis is performed on the candidate rhythm feature set, and feature filtering is performed on the candidate rhythm feature set in combination with a preset cumulative contribution rate threshold. Multiple candidate rhythm features that exceed the cumulative contribution rate threshold are extracted to generate a list of physiological rhythm features.

[0028] First, within the historical normal growth cycle of the grapevines, sample time-series signals were simultaneously collected by the first and second types of sensors. The historical normal growth cycle refers to a complete growing season confirmed by agronomic experts as free from severe stress and in a healthy growth state; for example, the previous growing season for Cabernet Sauvignon was approximately 180 days. The sample time-series signals refer to the historical vine time-series signals and historical environmental time-series signals, simultaneously collected by the first and second types of sensors and preprocessed by the S100 system within the historical normal cycle. Sensor data collected within the historical normal growth cycle of the target vineyard was retrieved, processed by the S100 system, and standardized historical sample time-series signals, including historical vine time-series signals and historical environmental time-series signals, were obtained as the basis for feature selection.

[0029] For example, historical data from the previous growing season (March to October) of a Cabernet Sauvignon vineyard were retrieved and processed by S100 to obtain a 180-day historical sample time-series signal. This includes IMU-filtered data from 24 sample points per day, stem diameter variation data (i.e., historical plant time-series signal), and corresponding soil water potential, air temperature and humidity, and photosynthetically active radiation data (i.e., historical environmental time-series signal). All data have a time resolution of 1 hour per sample.

[0030] Secondly, based on the sample time-series signals, multiple candidate rhythmic features representing different physiological dimensions are calculated and obtained, forming a candidate rhythmic feature set. Candidate rhythmic features refer to characteristic parameters calculated based on historical plant time-series signals that can represent different physiological dimensions and have clear physiological significance. The candidate rhythmic feature set is a collection of all candidate rhythmic features, covering different physiological dimensions, providing sufficient feature sources for subsequent screening. For historical plant time-series signals, starting from physiological dimensions such as water transport and growth vigor, multiple candidate rhythmic features are extracted through statistical calculations, waveform analysis, and other methods, and then summarized to form the candidate rhythmic feature set.

[0031] For example, based on the IMU vertical acceleration data in the historical plant time-series signal, the daily maximum contraction amplitude, contraction rate, afternoon recovery start time, and nighttime signal spectrum entropy are extracted; based on the stem diameter micro-variation data, the daily maximum contraction amount and daily cumulative growth amount are extracted. A total of 6 candidate rhythm features are obtained, forming a candidate rhythm feature set. For example, the candidate rhythm feature set for a typical growth day of a single healthy plant within a 24-hour time window is: {Daily maximum contraction amplitude 0.35mm, contraction rate 0.09mm / h, afternoon recovery start time 14:00, nighttime signal spectrum entropy 1.3, daily maximum contraction amount 0.28mm, daily cumulative growth amount 0.08mm}.

[0032] Finally, principal component analysis (PCA) is performed on the candidate rhythm feature set. Combined with a preset cumulative contribution rate threshold, the candidate rhythm feature set is screened to extract multiple candidate rhythm features exceeding the cumulative contribution rate threshold, generating a physiological rhythm feature list. Principal component analysis (PCA) is a data dimensionality reduction and feature selection algorithm that converts multiple related features into a few unrelated principal components, retaining the core information of the original data and eliminating redundant features. The cumulative contribution rate threshold refers to the minimum proportion of the original data variance explained by the principal components, used to determine the number of principal components to retain, thereby screening key candidate features. The physiological rhythm feature list refers to the set of core features that, after screening, maximize the representation of the plant's physiological state and have low redundancy, serving as the basis for real-time feature extraction. PCA is performed on the candidate rhythm feature set to calculate the contribution of each candidate feature to the principal components; the features are sorted from high to low according to the cumulative contribution rate of the principal components, and candidate features with a cumulative contribution rate reaching the preset threshold are retained to generate the physiological rhythm feature list.

[0033] For example, principal component analysis was performed on the above six candidate rhythm features to calculate the contribution of each feature: maximum daily contraction amplitude 28%, maximum daily contraction amount 25%, contraction rate 18%, afternoon recovery start time 12%, nighttime signal spectrum entropy 10%, and daily cumulative growth 7%. With a preset cumulative contribution rate threshold of 85%, the cumulative contribution rate of the first four features is 28% + 25% + 18% + 12% = 83%, and the cumulative contribution rate of the first five features is 93%, exceeding 85%. Therefore, the first five features were selected to generate a list of physiological rhythm features: {maximum daily contraction amplitude, maximum daily contraction amount, contraction rate, afternoon recovery start time, and nighttime signal spectrum entropy}.

[0034] Specifically, the original plant time-series signal is processed using feature engineering methods, and the processed result is fused with the environmental time-series signal to construct a multi-dimensional feature matrix, including: Based on a pre-defined list of physiological rhythm features, feature extraction is performed on the original plant time-series signals using feature engineering methods to obtain a set of physiological rhythm time-series features of grape plants. The environmental time-series signal and the physiological rhythm time-series feature set are concatenated to obtain the multidimensional feature matrix.

[0035] First, based on a pre-defined list of physiological rhythm features, feature extraction is performed on the original plant time-series signal using feature engineering methods to obtain a physiological rhythm time-series feature set for the grapevine. The physiological rhythm time-series features include at least one of the following: daily maximum contraction amplitude, contraction rate, afternoon recovery start time, nighttime signal spectral entropy, and daily cumulative growth. The physiological rhythm time-series feature set refers to the set of core physiological features extracted from the real-time preprocessed original plant time-series signal according to the physiological rhythm feature list, which changes over time and dynamically reflects the current physiological state of the plant. Feature engineering methods refer to target feature extraction methods based on the list, including statistical calculation, time node identification, and spectral analysis. Using the pre-defined list of physiological rhythm features, the original plant time-series signal processed by S100 is processed using the same method as historical feature calculation to extract each core physiological feature one by one, arranging them in time sequence to form the physiological rhythm time-series feature set.

[0036] For example, for the real-time raw plant time-series signal at a monitoring point during the current growing season, features are extracted according to the list of physiological rhythm features: from the daily variation curve of IMU vertical acceleration, the maximum daily contraction amplitude is calculated to be 0.32 mm, the contraction rate is 0.08 mm / h, the afternoon recovery start time is 15:10, and the nighttime signal spectrum entropy is 1.2; from the stem diameter data, the maximum daily contraction amount is calculated to be 0.25 mm; the five features are arranged in a time series of 1 hour / sample, for a total of 24 time points, forming a 24×5 dimension physiological rhythm time-series feature set.

[0037] Secondly, the environmental time-series signal and the physiological rhythm time-series feature set are concatenated to obtain the multidimensional feature matrix.

[0038] The method for obtaining the multidimensional feature matrix by concatenating the environmental time-series signal with the physiological rhythm time-series feature set includes: Acquire image data of the target scene and extract real-time shape feature vectors representing the current geometric shape of the shed; Based on the historical normal growth cycle data, a standard shape feature vector representing the standard trellis morphology is extracted. Calculate the residual vector between the real-time shape feature vector and the standard shape feature vector, and input the residual vector into the pre-trained shape influence correction model to obtain the corresponding correction coefficient vector; The physiological rhythm time series feature set is corrected using the correction coefficient vector; The modified physiological rhythm time series features are concatenated with the environmental time series signals to generate the multidimensional feature matrix.

[0039] First, image data of the target scene is acquired, and real-time shape feature vectors representing the current geometric shape of the trellis are extracted. Target scene image data is visual data used to represent the current geometric shape of the trellis; it can be drone point cloud data or high-resolution RGB images, accurately reflecting the spatial structure of the trellis. The real-time shape feature vector refers to the low-dimensional vector representing the current trellis shape extracted from the real-time image data. It is usually the core feature after dimensionality reduction using Principal Component Analysis (PCA), with a fixed dimension and consistent with the standard vector. UAVs equipped with LiDAR are periodically used to collect trellis point cloud data at various monitoring points in the target vineyard. After denoising and registration using point cloud processing software such as CloudCompare, the scores of the first N principal components are extracted using the PCA algorithm to construct the real-time shape feature vector.

[0040] For example, on the 5th of each month, a drone is used to scan the pergola at 5 monitoring points in the Cabernet Sauvignon vineyard to obtain the pergola point cloud data at monitoring point 3. After denoising, the scores of the first 3 principal components are extracted by PCA to generate a 3-dimensional real-time shape feature vector: [2.35, 1.12, 0.87], where each dimension represents the span, height, and density characteristics of the pergola.

[0041] Secondly, based on the historical normal growth cycle data, standard shape feature vectors representing the standard trellis morphology are extracted. The standard trellis morphology refers to the trellis morphology confirmed by agronomic experts during the historical normal growth cycle, characterized by no structural deformation and meeting the plant's growth requirements—that is, an interference-free baseline morphology. The standard shape feature vector is a shape feature vector extracted from trellis image data of the historical normal growth cycle, obtained by averaging multiple samples, and used as a benchmark for judging the deviation of the current trellis morphology. Point cloud data of the trellis at each monitoring point during the historical normal growing season are retrieved, and the shape feature vectors are extracted for each instance using the same method described above; the mean of the corresponding dimensions of all vectors is calculated to generate the standard shape feature vector.

[0042] For example, retrieve 12 point cloud data of the trellis at monitoring point 3 in the previous growing season and extract 12 shape feature vectors, such as [2.28, 1.09, 0.91] and [2.32, 1.11, 0.89]. Calculate the mean of each dimension: first dimension (2.28 + 2.32 + ... + 2.30) / 12 ≈ 2.30, second dimension (1.09 + 1.11 + ... + 1.10) / 12 ≈ 1.10, third dimension (0.91 + 0.89 + ... + 0.90) / 12 ≈ 0.90, and finally obtain the standard shape feature vector: [2.30, 1.10, 0.90].

[0043] Further, the residual vector between the real-time shape feature vector and the standard shape feature vector is calculated, and the residual vector is input into the pre-trained shape influence correction model to obtain the corresponding correction coefficient vector.

[0044] The training steps for the shape influence correction model include: Based on the historical normal growth cycle data, multiple sets of standard shape feature vectors and corresponding historical physiological rhythm features, neighborhood shape feature vectors and historical neighborhood physiological rhythm features are extracted. The model input is the residual vector between the standard shape feature vector and the neighborhood shape feature vector, and the model output is the multidimensional ratio vector formed by the ratio of the corresponding historical neighborhood physiological rhythm feature to the ratio of multiple dimensions of the historical physiological rhythm feature. A training sample set is constructed. By combining the training sample set, a regression model is constructed and trained to obtain the shape influence correction model.

[0045] First, based on the historical normal growth cycle data, multiple sets of standard shape feature vectors and corresponding historical physiological rhythm features, neighborhood shape feature vectors, and historical neighborhood physiological rhythm features are extracted. The standard shape feature vector refers to the trellis morphology feature vector obtained by averaging multiple samples from the target monitoring point during the historical normal growing season; this is the baseline trellis morphology vector. The historical physiological rhythm features refer to the physiological characteristics of healthy plants extracted from the physiological rhythm feature list at the target monitoring point during the historical normal growth cycle; this is the baseline physiological characteristics. The neighborhood shape feature vector refers to the trellis shape feature vector extracted from historical data within a 10-20 meter radius of the target monitoring point, covering different trellis deviation scenarios, in neighboring monitoring areas where trellis morphology exhibits natural differences. The historical neighborhood physiological rhythm features refer to the physiological rhythm features collected from the neighboring monitoring area on the corresponding growing day and extracted by S200; these are physiological characteristics affected by the neighboring trellis morphology. Retrieve point cloud data of the trellis during the historical normal growth cycle and the corresponding plant time-series signals. Extract and group the data according to the following rules: Determine the standard shape feature vector and corresponding historical physiological rhythm features of the target monitoring point; Select multiple neighborhood regions of the target monitoring point and extract the neighborhood shape feature vector and corresponding historical neighborhood physiological rhythm features of each neighborhood; Ensure that each group of data corresponds to the same growth day and the environmental conditions are consistent, eliminate environmental interference, and form multiple sets of matching data of standard shape vector - historical physiological features - neighborhood shape vector - historical neighborhood physiological features.

[0046] For example, taking monitoring point 3 as the target area, its standard shape feature vector is [2.30, 1.10, 0.90], and the historical physiological rhythm characteristics at a single time point on a certain growth day are [0.35mm, 0.28mm, 0.09mm / h, 14.67h, 1.3]. Two neighboring regions 3-1 and 3-2 around monitoring point 3 are selected, and the neighboring shape feature vectors for the same growth day are extracted: Region 3-1 [2.40, 1.15, 0.8]. 5], Region 3-2 [2.20, 1.08, 0.95], corresponding to the historical neighborhood physiological rhythm characteristics: Region 3-1 [0.37mm, 0.29mm, 0.08mm / h, 14.75h, 1.25], Region 3-2 [0.33mm, 0.27mm, 0.095mm / h, 14.60h, 1.35]; finally, two sets of matching data were formed, covering two deviation scenarios: the trellis is too wide and too high, and the trellis is too narrow and too low.

[0047] Secondly, a training sample set is constructed using the residual vector between the standard shape feature vector and the neighborhood shape feature vector as the model input, and the multidimensional ratio vector formed by the ratios of the corresponding historical neighborhood physiological rhythm features to the historical physiological rhythm features in multiple dimensions as the model output. The residual vector is the difference vector between the standard shape feature vector of the target monitoring point and the neighborhood shape feature vector, representing the degree of morphological deviation between the neighborhood scaffold and the standard scaffold. The multidimensional ratio vector is the dimension-wise ratio of the historical neighborhood physiological rhythm features of the neighborhood region to the historical physiological rhythm features of the target region, representing the proportion of physiological feature deviation caused by scaffold deviation, i.e., the true correction proportion. The training sample set is a one-to-one data set composed of the residual vector and the multidimensional ratio vector, used for model training. For each set of matched data, construct samples according to the following steps: Calculate input: residual vector = neighborhood shape feature vector - standard shape feature vector; Calculate output: multidimensional ratio vector = historical neighborhood physiological rhythm feature / historical physiological rhythm feature, divide each dimension; Summarize all residual vector - multidimensional ratio vector pairs to form a training sample set.

[0048] For example, the following sample is calculated for the two sets of matching data mentioned above: Neighborhood 3-1: Residual vector = [2.40-2.30, 1.15-1.10, 0.85-0.90] = [0.10, 0.05, -0.05]; Multidimensional ratio vector = [0.37 / 0.35≈1.06, 0.29 / 0.28≈1.04, 0.08 / 0.09≈0.89, 14.75 / 14.67≈1.01, 1.25 / 1.3≈0.96]. Neighborhood 3-2: Residual vector = [2.20-2.30, 1.08-1.10, 0.95-0.90] = [-0.10, -0.02, 0.05]; Multidimensional ratio vector = [0.33 / 0.35≈0.94, 0.27 / 0.28≈0.96, 0.095 / 0.09≈1.06, 14.60 / 14.67≈0.99, 1.35 / 1.3≈1.04]. After summarizing, a basic set containing 2 samples is formed. Expanding this to 100 sets of historical data yields a training sample set containing 100 samples.

[0049] Furthermore, a regression model is constructed and trained using the training sample set to obtain the shape influence correction model. The regression model uses a multilayer perceptron (MLP), a supervised learning model based on neural networks that can fit the nonlinear mapping relationship between the input residual vector and the output multidimensional ratio vector. The shape influence correction model refers to an MLP model that, after training, can stably output a correction coefficient vector; the input is the real-time residual vector, and the output is the correction coefficient corresponding to the physiological feature.

[0050] For example, the MLP model structure is built as follows: the number of neurons in the input layer equals the dimension of the residual vector, which is 3; there are two hidden layers with 16 and 8 neurons respectively, using ReLU as the activation function; the number of neurons in the output layer equals the dimension of the multidimensional ratio vector, which is 5, using Sigmoid as the activation function, ensuring that the output coefficients are in the range of 0.8-1.2; the training parameters are set as follows: the loss function is mean squared error (MSE), the optimizer is Adam, the learning rate is 0.001, the number of training epochs is 50, and early stopping is used, i.e., training stops if the loss on the validation set does not decrease after 3 epochs to prevent overfitting; the model is trained using the training sample set, and the loss on the training and validation sets is monitored in real time. After training, the model parameters are saved, which is the shape influence correction model. For example, training is performed on a training set of 100 samples: after 35 rounds of training, the loss on the validation set drops to 0.002, triggering early stopping; the predicted output of the trained model for a certain sample in the validation set is [0.97, 0.98, 1.03, 0.99, 1.02], with an error of only 0.01 compared to the true multidimensional ratio vector [0.96, 0.97, 1.04, 1.00, 1.01], showing a good fit; saving this model yields the final shape influence correction model.

[0051] Then, the residual vector between the real-time shape feature vector and the standard shape feature vector is calculated, and the residual vector is input into the pre-trained shape influence correction model to obtain the corresponding correction coefficient vector. The residual vector is the difference vector between the real-time shape feature vector and the standard shape feature vector, representing the degree of deviation between the current scaffold shape and the standard shape. The larger the difference, the more significant the scaffold interference. The correction coefficient vector is a numerical vector whose dimension is consistent with the physiological rhythm feature set output by the model. Each value corresponds to the correction ratio of a physiological feature. The residual vector is calculated by subtracting the standard shape feature vector from the real-time shape feature vector; the residual vector is input into the pre-trained shape influence correction model, and the model outputs the corresponding correction coefficient vector.

[0052] For example, the residual vector = real-time vector [2.35, 1.12, 0.87] - standard vector [2.30, 1.10, 0.90] = [0.05, 0.02, -0.03]. Inputting this residual vector into the pre-trained shape influence correction model, the model outputs a 5-correction coefficient vector: [1.02, 1.01, 0.99, 1.00, 1.03], which correspond to the correction ratios of the 5 physiological rhythm features.

[0053] Subsequently, the physiological rhythm time-series feature set is corrected using the correction coefficient vector. Feature correction refers to performing dimension-by-dimensional operations on the physiological rhythm time-series feature set using the correction coefficient vector to offset the interference of trellis morphology deviations on physiological characteristics and restore the intrinsic physiological state of the plant. The feature value at each time point in the physiological rhythm time-series feature set is multiplied by the coefficient of the corresponding dimension in the correction coefficient vector to obtain the corrected physiological rhythm time-series feature set.

[0054] For example, the original values ​​of the real-time physiological rhythm time series feature set of a certain monitoring point are: [Daily maximum contraction amplitude 0.32mm, daily maximum contraction amount 0.25mm, contraction rate 0.08mm / h, afternoon recovery start time 15:10 (quantized as 15.17h), nighttime signal spectrum entropy 1.2]; each feature value is multiplied by the correction coefficient vector [1.02, 1.01, 0.99, 1.00, 1.03] to obtain the corrected feature values: [0.32×1.02=0.326mm, 0.25×1.01=0.253mm, 0.08×0.99=0.079mm / h, 15.17×1.00=15.17h, 1.2×1.03=1.236]; corrected point by point according to 24 time points to form the corrected 24×5-dimensional physiological rhythm time series feature set.

[0055] Finally, the corrected physiological rhythm time-series features are concatenated with the environmental time-series signal to generate the multidimensional feature matrix. The corrected multidimensional feature matrix refers to the structured data formed by concatenating the corrected physiological rhythm time-series feature set with the environmental time-series signal, eliminating trellis interference and better reflecting the coupling relationship between the plant's true physiological state and the environment. The corrected physiological rhythm time-series feature set is aligned with the environmental time-series signal processed by S100 according to the timestamp, and feature vectors at the same time point are combined using a horizontal concatenation method, arranged in a time series to generate the multidimensional feature matrix.

[0056] For example, the corrected physiological rhythm time series feature set is aligned with the environmental time series signal by timestamp, and the feature vector of each time point is [corrected daily maximum contraction amplitude, corrected daily maximum contraction amount, corrected contraction occurrence rate, corrected recovery start time, corrected spectral entropy, soil water potential, air VPD, cumulative photosynthetically active radiation]; finally forming a corrected multidimensional feature matrix with a dimension of 24×8, which serves as the input to the dynamic baseline model.

[0057] In this embodiment of the invention, by extracting trellis shape features, quantifying deviations, and correcting physiological features, non-physiological interference caused by differences in trellis morphology was successfully eliminated, solving the feature distortion problem caused by the composite features of vines and trellises. The corrected physiological rhythm features better reflect the intrinsic growth state of grapevines, and the multi-dimensional feature matrix constructed by combining environmental signals has higher purity and more accurate correlation. This not only improves the training effect of the dynamic baseline model but also enhances its generalization ability across orchards with different trellis morphologies, providing more reliable feature support for accurate anomaly early warning across different scenarios.

[0058] S300: Acquire historical normal growth cycle data based on big data, construct and train a dynamic baseline model based on a long short-term memory neural network, and determine the graded early warning threshold.

[0059] In this embodiment of the invention, historical normal growth cycle data based on big data is acquired, and a dynamic baseline model based on a long short-term memory neural network is constructed and trained accordingly, and a graded early warning threshold is determined. The normal growth rhythm of grapevines exhibits significant temporal dynamics; for example, the stem contraction during the fruit expansion stage is much greater than that during the budding stage. Static thresholds or ordinary classification models cannot adapt to the physiological characteristic changes at different growth stages. Furthermore, single-dimensional feature errors cannot comprehensively characterize the degree of growth abnormality. Directly using fixed thresholds to judge abnormalities easily leads to misjudgments of growth stages and indiscriminate differentiation of abnormality levels. Therefore, a dynamic baseline model capable of capturing temporal dynamic patterns needs to be constructed based on historical normal growth data. This model learns the temporal characteristic patterns of normal growth through encoding and reconstruction, and then uses the statistical distribution of reconstruction errors to determine graded early warning thresholds, achieving accurate abnormality identification through dynamic baseline comparison and graded early warning.

[0060] Step S300 in the method provided in this embodiment of the invention includes: Acquire historical normal growth cycle data, and perform time-series alignment and feature extraction based on the historical normal growth cycle data to obtain a multi-dimensional sample feature matrix set; Using the multidimensional sample feature matrix set as training data, the dynamic baseline model based on a long short-term memory neural network is constructed and trained. Based on the statistical distribution of the reconstruction error of the training data by the trained dynamic baseline model, and combined with the preset percentile, the graded early warning threshold is determined.

[0061] The dynamic baseline model includes at least one encoder based on a long short-term memory neural network and at least one decoder based on a long short-term memory neural network. The encoder is used to encode the input physiological rhythm temporal features into a low-dimensional space vector of a preset dimension, and the decoder is used to reconstruct an output sequence with the same dimension as the input physiological rhythm temporal features based on the low-dimensional space vector. The number of layers, neurons, and structure of the encoder and decoder are not strictly symmetrical, and the dimension of the low-dimensional space vector is lower than the dimension of the physiological rhythm temporal features.

[0062] First, based on big data methods, historical normal growth cycle data is acquired. Then, time-series alignment and feature extraction are performed on this data to obtain a multi-dimensional sample feature matrix set. Historical normal growth cycle data refers to complete growing season data labeled by agronomic experts, containing rich big data scenarios, free from pest / disease / water / nutrient stress, and showing healthy growth. This data includes plant / environment time-series signals preprocessed by S100 and feature extraction results by S200. This data acquisition path based on big data mining and analysis provides a broad and information-rich data foundation for subsequent training. The multi-dimensional sample feature matrix set refers to the dataset formed by classifying historical normal data according to growth stages and summarizing the multi-dimensional feature matrices within each stage, covering normal feature patterns at different growth stages. Time-series alignment refers to unifying the time resolution (1 hour / sample) and feature dimensions (24×8) of historical data to ensure consistency with real-time data format and eliminate differences in data format across time periods. Historical normal sensor data from the previous growing season of the target vineyard was retrieved and preprocessed using S100. Physiological features and environmental features were extracted from each set of historical data using S20 to generate a single 24×8 multidimensional feature matrix. The data was then classified according to the grape agronomic growth stages: budding period: March-April, new shoot growth period: May, fruit expansion period: June-July, coloring period: August-September, and leaf fall period: October. Extreme values ​​of features caused by sensor malfunctions, such as sudden changes in stem diameter, were removed. The training set and validation set were divided in an 8:2 ratio, and the results were summarized to form a multidimensional sample feature matrix set.

[0063] For example, historical normal data from the previous growing season (March to October) of a Cabernet Sauvignon vineyard were retrieved. After S100 preprocessing and S200 trellis correction, a 24×8 multidimensional feature matrix was generated for each growing day. The matrix was classified by agronomic stage: 45 matrices for budding stage, 30 matrices for new shoot growth stage, 60 matrices for fruit expansion stage, 40 matrices for coloring stage, and 25 matrices for leaf fall stage. After removing the abnormal matrices caused by sensor malfunctions, 197 valid matrices were finally obtained, which constituted a multidimensional sample feature matrix set, including 158 training matrices and 39 validation matrices.

[0064] Secondly, using the multidimensional sample feature matrix set as training data, a dynamic baseline model based on a long short-term memory (LSTM) neural network is constructed and trained. The dynamic baseline model includes at least one encoder and at least one decoder based on an LSTM neural network. The encoder encodes the input physiological rhythm temporal features into a low-dimensional spatial vector of a preset dimension. The decoder reconstructs an output sequence with the same dimension as the input physiological rhythm temporal features based on the low-dimensional spatial vector. The number of layers, neurons, and structures of the encoder and decoder are not strictly symmetrical, and the dimension of the low-dimensional spatial vector is lower than the dimension of the physiological rhythm temporal features. The LSTM neural network is a variant of a recurrent neural network capable of capturing long-term temporal dependencies, adapting to the dynamic patterns of plant temporal features, and solving the gradient vanishing problem of ordinary RNNs. The encoder compresses high-dimensional temporal features into a low-dimensional spatial vector, extracting the core temporal patterns of normal growth, with a dimension lower than the input feature dimension. The decoder reconstructs the low-dimensional spatial vector into a temporal feature sequence with the same dimension as the input, restoring the feature patterns of normal growth. Non-strictly symmetric structures refer to encoders and decoders where the number of LSTM layers and neurons is not necessarily the same, adapting to different task requirements of encoding compression-decoding reconstruction.

[0065] Specifically, for planting scenarios with different varieties or different planting methods, multiple sets of dynamic baseline models based on long short-term memory neural networks (LSTM) encoder-decoder can be trained separately. These dynamic baseline models can then be associated with the corresponding variety or planting method encodings, thereby forming a specialized and scenario-based multi-agent ensemble system. This improves the adaptability to different scenarios and helps each dynamic baseline model converge quickly, reducing the complexity and training difficulty of the encoder-decoder.

[0066] For example, a non-strictly symmetric LSTM encoder-decoder structure is constructed: Encoder: 2-layer LSTM, with 64 neurons in the first layer and 32 neurons in the second layer, activation function tanh, input a 24×8 multidimensional feature matrix, output a 16-dimensional low-dimensional vector; Decoder: 1-layer LSTM, 48 neurons, activation function tanh, input a 16-dimensional low-dimensional vector, output a reconstructed 24×8 temporal feature sequence. Training parameters are set as follows: loss function is mean squared error (MSE), optimizer is Adam (learning rate 0.001), batch size is 16, maximum training epochs are 80; early stopping rule: the validation set MSE decreases for 4 consecutive epochs. Then stop training; train the model using the training set of the multidimensional sample feature matrix set, and monitor and reconstruct the MSE in real time using the validation set; after training, save the model parameters to obtain the dynamic baseline model. For example, training with 158 training set matrices: 1-20 rounds: the validation set MSE drops rapidly from 0.03 to 0.008; 21-50 rounds: the MSE decreases slowly, reaching 0.0035 at round 50; 51-54 rounds: the validation set MSE is 0.0035, 0.00349, 0.00348, and 0.00348 respectively, with a continuous decrease in MSE over 4 rounds. Early termination is triggered; training stops at 54 rounds. At this point, the validation set MSE=0.00348. The maximum daily shrinkage amplitude after reconstruction of a certain matrix during the expansion period is 0.348mm vs. the original value of 0.35mm, with an error of only 0.002mm. The reconstruction effect meets the standard, and the trained dynamic baseline model is obtained.

[0067] Finally, based on the statistical distribution of the reconstruction error of the training data using the trained dynamic baseline model, and combined with preset percentiles, the graded early warning threshold is determined. Reconstruction error refers to the element-wise mean square error (MSE) between the output sequence and the original input sequence after the dynamic baseline model reconstructs normal samples. The reconstruction error of normal samples follows a normal distribution. Statistical distribution refers to statistically analyzing the reconstruction errors of all training set samples to obtain features such as mean, standard deviation, and percentiles, reflecting the error distribution pattern of normal samples. Preset percentiles are the core basis for classifying anomaly levels; the higher the percentile, the larger the corresponding error and the more severe the anomaly. The graded early warning threshold refers to the error threshold determined based on percentiles for different levels of anomaly, distinguishing between mild, moderate, and severe anomalies. Calculate the reconstruction error of all samples in the training set: For each multidimensional sample feature matrix, calculate the element-wise MSE of the reconstructed sequence and the original sequence, and take the average reconstruction error of the entire matrix; perform statistics on all single sample error values ​​to verify that they conform to a normal distribution, and draw an error distribution histogram; preset percentiles: 95%: mild anomaly, 99%: moderate anomaly, 99.9%: severe anomaly; calculate the error value of the corresponding percentile as the graded early warning threshold.

[0068] For example, after calculating the reconstruction error of 158 training set samples, the average reconstruction error per sample shows a normal distribution: mean Standard deviation Percentile thresholds of 95%, 99%, and 99.9% were set, and the following were calculated: 95% mild warning threshold = 0.0036, meaning that the error of 95% of normal samples is ≤0.0036, and exceeding this threshold indicates a mild anomaly; 99% moderate warning threshold = 0.0048, meaning that the error of 99% of normal samples is ≤0.0048, and exceeding this threshold indicates a moderate anomaly; 99.9% severe warning threshold = 0.0065, meaning that the error of 99.9% of normal samples is ≤0.0065, and exceeding this threshold indicates a severe anomaly. The final determined tiered warning thresholds are: 0.0036, 0.0048, and 0.0065.

[0069] In this embodiment of the invention, a non-strictly symmetric LSTM codec dynamic baseline model is constructed to effectively learn the physiological-environment coupling temporal patterns at different growth stages of grapes, solving the problem that static thresholds cannot adapt to dynamic growth. Based on the statistical distribution of reconstruction errors of normal samples, a graded warning threshold is determined, achieving quantitative grading of the degree of anomaly, rather than a simple binary classification of normal / abnormal. The dynamic baseline model balances feature compression efficiency and reconstruction accuracy, while the graded thresholds provide a clear quantitative basis for subsequently pushing different levels of warning instructions, effectively improving the hierarchy, accuracy, and practicality of anomaly warnings.

[0070] S400: Input the multidimensional feature matrix into the dynamic baseline model and combine it with the hierarchical early warning threshold to perform growth abnormality early warning.

[0071] In this embodiment of the invention, the multidimensional feature matrix is ​​input into the dynamic baseline model, and combined with the graded early warning thresholds, growth anomaly early warning is provided. The dynamic baseline model has learned the normal temporal characteristics of different growth stages of grapes, and the graded early warning thresholds clarify the quantitative standards for different degrees of anomalies. The core requirement of orchard management is real-time anomaly detection, accurate level determination, and rapid response. Without a real-time early warning process, even if the model and thresholds are accurate, the technological achievements cannot be transformed into actual management effectiveness, easily leading to delayed anomaly handling, such as the continued expansion of water stress affecting yield. Therefore, a closed-loop process of real-time error calculation - threshold comparison and level determination - message push is needed to transform the model output into intuitive and operable early warning instructions, achieving rapid response in detection, early warning, and handling, and maximizing the practical value of the dynamic baseline model and graded thresholds.

[0072] Step S400 in the method provided in this embodiment of the invention includes: The multidimensional feature matrix constructed in real time is input into the dynamic baseline model to calculate the real-time reconstruction error; The real-time reconstruction error is compared with the graded early warning threshold, and a graded early warning instruction of the corresponding level is triggered based on the comparison result. An early warning message is generated according to the tiered early warning instruction and sent to the management terminal. The early warning message includes at least the early warning level and the early warning location.

[0073] First, the multidimensional feature matrix constructed in real time is input into the dynamic baseline model to calculate the real-time reconstruction error. The real-time multidimensional feature matrix refers to a 24×8 dimension feature matrix constructed in real time based on real-time sensor data from the current growth cycle, after S100 preprocessing, S200 feature extraction, and trellis correction, reflecting the current physiological-environment coupling state of the plant. The real-time reconstruction error refers to the mean square error (MSE) between the reconstructed sequence and the real-time input sequence after the dynamic baseline model encodes and reconstructs the real-time multidimensional feature matrix, quantifying the deviation of the current state from the normal baseline. Real-time sensor data is automatically collected at a preset period, and a 24×8 multidimensional feature matrix is ​​constructed in real time using S100 and S200. This real-time matrix is ​​input into the trained dynamic baseline model, which is compressed into a low-dimensional vector by the encoder and then reconstructed by the decoder to produce a normal prediction sequence with the same dimension as the input. The element-wise MSE of the real-time input sequence and the reconstructed sequence is calculated, and the average MSE of the entire matrix is ​​taken as the real-time reconstruction error.

[0074] For example, at 16:00 on a certain day during the fruit expansion period of a Cabernet Sauvignon vineyard, real-time sensor data from monitoring point 3 was collected. After S100 preprocessing and S200 trellis correction, a 24×8 real-time multidimensional feature matrix was constructed. This matrix was input into the dynamic baseline model, and the dynamic baseline model output a reconstructed sequence. The average MSE between the real-time input sequence and the reconstructed sequence was calculated to be 0.0052, that is, the real-time reconstruction error was 0.0052.

[0075] Secondly, the real-time reconstruction error is compared with the tiered early warning threshold, and based on the comparison result, a corresponding level of tiered early warning instruction is triggered. The tiered early warning instruction refers to the corresponding level of early warning instruction generated based on the comparison result between the real-time reconstruction error and the tiered threshold, clearly defining the severity of the anomaly and providing a basis for subsequent message generation. Threshold comparison rules: the smaller the error, the more normal the situation; the larger the error, the more severe the anomaly. The tiers are strictly divided according to the tiered threshold, with no ambiguity.

[0076] For example, the graded early warning thresholds of 0.0036, 0.0048, and 0.0065 determined by S300 are retrieved; the real-time reconstruction error is compared with the thresholds according to the following rules to trigger the corresponding instruction: If the error < 0.0036: no early warning instruction is given, and the current growth status is determined to be normal; if 0.0036 ≤ error < 0.0048: a level 1 early warning instruction is triggered, requiring close monitoring; if 0.0048 ≤ error < 0.0065: a level 2 early warning instruction is triggered, requiring timely verification; if the error ≥ 0.0065: a level 3 early warning instruction is triggered, requiring emergency handling. For example, if the real-time reconstruction error of monitoring point 3 is 0.0052, after retrieving the graded thresholds and comparing them: 0.0048 ≤ 0.0052 < 0.0065, it meets the level 2 early warning triggering condition; a level 2 early warning instruction is automatically generated, marking the abnormality level as alarm level, and the associated location as monitoring point 3.

[0077] Finally, an early warning message is generated based on the tiered early warning instructions and sent to the management terminal. The early warning message includes at least the early warning level and location. The early warning message is structured information generated based on the tiered early warning instructions, containing the core management decision-making content and avoiding redundancy. The management terminal refers to the receiving device used daily by the orchard administrator, such as a mobile early warning app or a computer management platform, supporting real-time message push and historical query. Based on the triggered tiered early warning instructions, an early warning message is automatically generated, including at least: early warning level, early warning location, early warning time, core characteristic deviation prompts, and handling suggestions. The early warning message is synchronously pushed to preset management terminals via IoT communication protocols such as 4G / 5G, such as mobile app pop-ups + SMS reminders, and computer platform message notifications. After receiving the message, the management terminal records a message log, allowing administrators to mark statuses such as "verified" and "handled," forming a closed-loop management system.

[0078] For example, based on the Level 2 warning command, the following warning message is generated: Level 2 Alert: Grape Growth Abnormality Warning! Time: July 15, 2025, 16:00; Location: Cabernet Sauvignon Vineyard Monitoring Point 3; Note: The reconstruction deviation of the maximum daily shrinkage amplitude and maximum daily shrinkage amount reaches 60%, suspected to be water stress; Recommended Action: Please check the soil moisture and drip irrigation system operation status in this area within 2 hours, and start supplementary irrigation if necessary. The message is simultaneously pushed to the administrator's mobile app and computer management platform. The administrator can view detailed information by opening the app.

[0079] In this embodiment of the invention, a real-time early warning closed loop of data-model-decision-handling is constructed through real-time data processing, precise error calculation, hierarchical threshold comparison, and rapid message push. This leverages the accuracy of the dynamic baseline model in capturing temporal anomalies while clearly defining the severity of anomalies through hierarchical early warning, avoiding a one-size-fits-all approach. The early warning messages include specific location, suspected causes, and handling suggestions, reducing the decision-making cost for administrators and achieving the early warning goals of early detection, accurate grading, and rapid response. This effectively reduces yield and quality losses caused by delays in anomaly handling, providing direct and implementable technical support for refined orchard management.

[0080] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a multi-sensor fusion method for early warning of abnormal grapevine growth. In the data acquisition stage, sensors directly deployed on the grapevines acquire intrinsic physiological time-series signals unaffected by trellis morphology, ensuring data independence and accuracy from the source. In the feature construction stage, time-series features with clear physiological significance are extracted and effectively fused with environmental data, enhancing the model's input representational ability and interpretability. In the model training stage, a dynamic baseline model capable of learning long-term temporal dependencies is employed to accurately characterize the complex lag patterns between environmental stress and physiological responses, achieving a more intelligent and physiologically accurate definition of the normal growth state of grapevines. Finally, in the early warning decision-making stage, a hierarchical threshold mechanism based on reconstruction error enables continuous and quantitative judgment from early deviations to severe anomalies, providing timely, accurate, and clearly directional early warning information, effectively supporting early intervention and precise decision-making in orchard management.

[0081] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0083] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A method for early warning of abnormal growth in grapevines using multi-sensor fusion, characterized in that, include: Original plant time-series signals and environmental time-series signals were collected by first-type and second-type sensors deployed in the target environment. The original plant time-series signal is processed using feature engineering methods, and the processing results are fused with the environmental time-series signal to construct a multidimensional feature matrix. Acquire historical normal growth cycle data based on big data, construct and train a dynamic baseline model based on long short-term memory neural network, and determine the graded early warning threshold; The multidimensional feature matrix is ​​input into the dynamic baseline model, and growth anomaly warning is given in combination with the hierarchical warning threshold.

2. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 1, characterized in that, The first type of sensor is deployed on the grapevine itself, and the second type of sensor is deployed in the grapevine's growing environment. The first type of sensor includes at least an inertial sensor and a diameter growth sensor.

3. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 1, characterized in that, Using first-type and second-type sensors deployed in the target environment, raw plant time-series signals and environmental time-series signals are collected, including: According to the preset spatiotemporal window, the first type of sensor and the second type of sensor are controlled to synchronously collect data to obtain the first time-series signal and the second time-series signal; The first time-series signal acquired is subjected to filtering and noise reduction processing; Based on statistical analysis methods, the physiological response lag time window is determined, and the second time-series signal is time-aligned with the first time-series signal after filtering and noise reduction. Missing values ​​are imputed based on the time alignment results, and the data is uniformly resampled to a preset time resolution. The output consists of the original plant time series signal and the environmental time series signal.

4. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 1, characterized in that, The original plant time-series signal is processed using feature engineering methods, and the processed result is fused with the environmental time-series signal to construct a multidimensional feature matrix. This process includes: During the normal growth cycle of the grapevines, sample time-series signals obtained by the first type of sensor and the second type of sensor were collected simultaneously. Based on the sample time-series signal, multiple candidate rhythm features representing different physiological dimensions are calculated and obtained, forming a candidate rhythm feature set; Principal component analysis is performed on the candidate rhythm feature set, and feature filtering is performed on the candidate rhythm feature set in combination with a preset cumulative contribution rate threshold. Multiple candidate rhythm features that exceed the cumulative contribution rate threshold are extracted to generate a list of physiological rhythm features.

5. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 1, characterized in that, The original plant time-series signal is processed using feature engineering methods, and the processed result is fused with the environmental time-series signal to construct a multi-dimensional feature matrix, including: Based on a pre-defined list of physiological rhythm features, feature extraction is performed on the original plant time-series signals using feature engineering methods to obtain a set of physiological rhythm time-series features of grape plants. The environmental time-series signal and the physiological rhythm time-series feature set are concatenated to obtain the multidimensional feature matrix.

6. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 1, characterized in that, Acquire historical normal growth cycle data based on big data, construct and train a dynamic baseline model based on a long short-term memory neural network, and determine graded early warning thresholds, including: Acquire historical normal growth cycle data, and perform time-series alignment and feature extraction based on the historical normal growth cycle data to obtain a multi-dimensional sample feature matrix set; Using the multidimensional sample feature matrix set as training data, the dynamic baseline model based on a long short-term memory neural network is constructed and trained. Based on the statistical distribution of the reconstruction error of the training data by the trained dynamic baseline model, and combined with the preset percentile, the graded early warning threshold is determined.

7. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 1, characterized in that, The multidimensional feature matrix is ​​input into the dynamic baseline model, and growth anomaly warning is performed in combination with the hierarchical warning threshold, including: The multidimensional feature matrix constructed in real time is input into the dynamic baseline model to calculate the real-time reconstruction error; The real-time reconstruction error is compared with the graded early warning threshold, and a graded early warning instruction of the corresponding level is triggered based on the comparison result. An early warning message is generated according to the tiered early warning instruction and sent to the management terminal. The early warning message includes at least the early warning level and the early warning location.

8. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 5, characterized in that, The method further includes performing feature concatenation between the environmental time-series signal and the physiological rhythm time-series feature set to obtain the multidimensional feature matrix, and also includes: Acquire image data of the target scene and extract real-time shape feature vectors representing the current geometric shape of the shed; Based on the historical normal growth cycle data, a standard shape feature vector representing the standard trellis morphology is extracted. Calculate the residual vector between the real-time shape feature vector and the standard shape feature vector, and input the residual vector into the pre-trained shape influence correction model to obtain the corresponding correction coefficient vector; The physiological rhythm time series feature set is corrected using the correction coefficient vector; The modified physiological rhythm time series features are concatenated with the environmental time series signals to generate the multidimensional feature matrix.

9. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 8, characterized in that, The training steps for the shape influence correction model include: Based on the historical normal growth cycle data, multiple sets of standard shape feature vectors and corresponding historical physiological rhythm features, neighborhood shape feature vectors and historical neighborhood physiological rhythm features are extracted. The model input is the residual vector between the standard shape feature vector and the neighborhood shape feature vector, and the model output is the multidimensional ratio vector formed by the ratio of the corresponding historical neighborhood physiological rhythm feature to the ratio of multiple dimensions of the historical physiological rhythm feature. A training sample set is constructed. By combining the training sample set, a regression model is constructed and trained to obtain the shape influence correction model.

10. The multi-sensor fusion method for early warning of abnormal grapevine growth as described in claim 6, characterized in that, The dynamic baseline model includes at least one encoder based on a long short-term memory neural network and at least one decoder based on a long short-term memory neural network. The encoder is used to encode the input physiological rhythm temporal features into a low-dimensional space vector of a preset dimension, and the decoder is used to reconstruct an output sequence with the same dimension as the input physiological rhythm temporal features based on the low-dimensional space vector. The number of layers, neurons, and structure of the encoder and decoder are not strictly symmetrical, and the dimension of the low-dimensional space vector is lower than the dimension of the physiological rhythm temporal features.

Images