Multi-dimensional canal flow monitoring method for farmland irrigation

Abstract

The invention provides a canal flow multi-dimensional monitoring method for farmland irrigation, and relates to the field of data prediction, and the method comprises the specific steps: firstly, converting original multi-modal monitoring data into an enhanced state integrated tensor through a modal asymmetric potential energy response coding module, and solving the modeling problem of asynchronous and asymmetric coupling of multi-modal response; then, a modal phase disturbance guide fusion module is used for carrying out direction perception fusion on coding features, and self-adaptive weighting and conflict resolution of response direction differences between modals are achieved; then, converting a fusion result into a structure code with disturbance perception capability by means of a conditional guidance structure coding module, and establishing an explicit mapping relationship between a dynamic disturbance structure and a prediction condition; and finally, self-adaptive prediction of trend dominance and historical balance is realized by using a monitoring module, and high-precision monitoring of the flow of the farmland irrigation canal is completed.

Description

Technical Field

[0001] This invention belongs to the field of data prediction, specifically relating to a multidimensional monitoring method for irrigation canal flow in farmland. Background Technology

[0002] With the continuous improvement of the intensification and precision of modern agriculture, farmland irrigation systems, as the core infrastructure for ensuring food production and water resource allocation, have become a key link in achieving efficient water resource utilization through real-time monitoring of their operation status and accurate flow prediction. The dynamic changes in canal flow directly affect irrigation uniformity and crop growth status. Deep integration and intelligent analysis of multi-source monitoring data is an important technical foundation for improving the regulation and control capabilities and risk warning level of irrigation systems.

[0003] In recent years, the widespread application of IoT technology and intelligent sensing devices has provided a wealth of data sources for canal monitoring. Some studies have attempted to use water level-flow relationship models for single-parameter prediction and traditional time series analysis methods to fit historical flow data. Some scholars have further introduced machine learning models to integrate meteorological factors for flow regression prediction. Other studies have explored deep learning structures to capture the temporal dependence of flow changes. These methods have achieved initial success in improving the automation of prediction and mining the nonlinear characteristics of data.

[0004] However, existing methods for monitoring canal flow still have significant limitations: First, the monitoring data comes from multiple types of sensors, and the data from each modality show significant differences in temporal evolution and disturbance response. Traditional methods fail to fully consider the inherent inconsistencies in direction and asymmetric coupling relationships when fusing data. Second, most models are insufficient in capturing the dynamic and non-stationary characteristics of the canal system under external disturbances, making it difficult to effectively cope with the sudden changes in flow and complex operating conditions commonly encountered during irrigation. Therefore, there is an urgent need for a multi-dimensional flow monitoring method that can accurately model the differences in multi-modal responses, possesses dynamic adaptability and strong generalization ability, in order to achieve accurate perception and forward-looking prediction of the operating status of farmland irrigation canals, and provide core technical support for the construction of smart irrigation systems. Summary of the Invention

[0005] This invention provides a multidimensional monitoring method for irrigation canal flow in farmland. It proposes a multidimensional monitoring model for complex irrigation canal flow data in farmland, which consists of a modal asymmetric potential energy response encoding module, a modal phase perturbation guided fusion module, a condition-guided structure encoding module, and a monitoring module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps:

[0007] S1. Collect the flow rate of irrigation canals in farmland and construct the original dataset;

[0008] S2. Based on the original dataset, construct a perturbation difference tensor to quantize the perturbation intensity, design an energy sensing function based on the perturbation difference tensor, construct inter-modal asymmetric coupling weights using an asymmetric coupling function, construct modal reflection coding through weighted fusion, and introduce a nonlinear reflection enhancement term to obtain the first dataset.

[0009] S3. The first dataset is processed by the unit angle mapping function to obtain the unit angle. Based on the unit angle, a nonlinear perturbation weighting function is constructed to generate a fusion weighting factor. A modal perturbation enhancement expression mechanism is designed to fuse modal feature representations to obtain the second dataset.

[0010] S4. Based on the second dataset, calculate the perturbation deviation index, construct the anti-perturbation response distribution function, use the guided response rate function to obtain the perturbation guided rate, propose a guided structure encoding mechanism to integrate the perturbation fusion representation and the perturbation guided rate to obtain the third dataset, which is divided into a training set and a prediction set;

[0011] S5. Calculate the trend-guided adjustment coefficient based on the training set, construct a weighted fusion prediction mechanism by combining the structure encoding results and the weighted average of the structure encoding, and design a loss function for training optimization.

[0012] S6. The prediction set is input into the trained multi-dimensional monitoring model of irrigation canal flow in farmland, and the final output is the predicted flow value.

[0013] Preferably, in step S2, based on the original dataset, a perturbation differential tensor is constructed to quantize the perturbation intensity, an energy sensing function is designed based on the perturbation differential tensor, an asymmetric coupling function is used to construct inter-modal asymmetric coupling weights, a modal reflection code is constructed through weighted fusion, and a nonlinear reflection enhancement term is introduced to process and obtain the first dataset.

[0014] For any mode At time step The original input vector The perturbation difference tensor is obtained by calculating the difference between the current time and past time. The perturbation difference tensor is input into the self-sensing function, and the result is calculated. The square term of the L2 norm is used, and a minimal positive number is introduced as a stability compensation. The reciprocal of the sum of the two terms is used to obtain the self-energy value. The product of the absolute value of the self-energy difference between modes and the Euclidean distance between the perturbation difference tensors is calculated, and a stability compensation term composed of the self-energy product is introduced to construct the numerator of the asymmetric coupling function. By normalizing the numerator with the sum of the difference measures of all mode pairs relative to the current mode, the asymmetric coupling weight is finally obtained. Using the asymmetric coupling weights and perturbation difference tensors as inputs, the modal reflection coding is constructed through weighted aggregation. The modal reflection code is multiplied by the self-energy value and then input into the hyperbolic tangent activation function. This is then multiplied element-wise by the original modal reflection code and multiplied by the reflection nonlinear enhancement coefficient. Generate the nonlinear reflection enhancement term, and then combine the nonlinear reflection enhancement term with... Linear superposition yields the conditional information encoding vector Finally, the enhanced state integration tensor is obtained. This serves as the first dataset.

[0015] Furthermore, addressing the issues of asynchronous modal responses, inconsistent disturbance propagation directions, and asymmetric coupling characteristic of multi-source monitoring in farmland irrigation systems, this invention proposes a modal asymmetric potential energy response encoding module. This module aims to achieve differentiated modeling and state enhancement of multimodal dynamic responses. First, it quantifies the local disturbance intensity of each mode on a short timescale by constructing a disturbance difference tensor. Then, it designs a self-energy sensing function to transform the disturbance difference amplitude into a self-energy value that measures stability. Next, it constructs an asymmetric coupling function based on potential energy differences and disturbance differences, fusing the self-energy value difference with the disturbance... The dynamic differential tensor difference constructs direction-specific coupling weights; finally, modal reflection codes are generated through weighted aggregation, and a nonlinear reflection enhancement term based on the hyperbolic tangent function is introduced to fuse the modal energy state and coupling information, forming a conditional information encoding vector with both linear robustness and nonlinear sensitivity. The final output is an enhanced state integration tensor as the first dataset. The module can transform the implicit stability differences and asymmetric coupling relationships in the original multimodal monitoring data into explicit and computable feature representations, providing a highly discriminative coding foundation for subsequent modal fusion and flow prediction.

[0016] Preferably, in step S3, the first dataset is processed by a unit angle mapping function to obtain a unit angle, a nonlinear perturbation weighting function is constructed based on the unit angle to generate a fusion weighting factor, and a modal perturbation enhancement expression mechanism is designed to fuse modal feature representations to obtain a second dataset;

[0017] Based on vector geometric relationships, a mapping function for the unit angle between modal pairs is constructed. The inner product of the conditional information encoding vectors of the two modes is calculated as the numerator. The product of a minimum positive number and the L2 norm of the two vectors is added as the denominator. The ratio of the numerator to the denominator is input into an inverse cosine function to obtain the unit angle. The unit included angle The nonlinear perturbation weighting function is constructed by inputting an exponential function, and an angle attenuation control coefficient is introduced. Generate the fusion weight factor Based on the conditional information encoding vector of the current modality, the directional difference vectors of all other modalities and the current modality are superimposed to construct the modal perturbation enhancement representation mechanism. This mechanism is then adjusted by the fusion weight factor to form a perturbation fusion representation. Finally, the fusion tensor is obtained. This serves as the second dataset.

[0018] Furthermore, to address the issues of inconsistent response directions and asynchronous disturbance propagation in multimodal monitoring data within farmland irrigation systems, this invention proposes a modal phase perturbation-guided fusion module. This module aims to achieve quantitative perception and adaptive fusion of directional differences between modalities. First, it calculates the unit angle between modal conditional information encoding vectors using a unit angle mapping function, constructing a quantitative index of directional deviation. Then, based on a nonlinear perturbation weighting function, it transforms the unit angle into a fusion weight factor, enhancing directionally consistent modes and suppressing directionally conflicting modes. Finally, through a modal perturbation enhancement expression mechanism, it fuses directional-weighted cross-modal difference information using the current modal's conditional information encoding vector as a benchmark, generating a perturbation fusion representation with directional discriminative power. The final output is a fusion tensor used as a second dataset. This module can transform directional differences in multimodal responses into computable fusion weights, significantly improving the ability to identify inconsistent flow trends and the robustness of multi-source information fusion.

[0019] Preferably, in step S4, a disturbance deviation index is calculated based on the second dataset, an anti-disturbance response distribution function is constructed, a disturbance guidance rate is obtained using the guidance response rate function, and a guidance structure encoding mechanism is proposed to integrate the disturbance fusion representation and the disturbance guidance rate to obtain a third dataset, which is divided into a training set and a prediction set;

[0020] The sum of squared Euclidean distances between the current modality's perturbation fusion representation and the perturbation fusion representations of all other modalities is calculated as the numerator. The sum of squared L2 norms of the perturbation fusion representations of all modalities is calculated, and a minimal positive number is introduced as the denominator to ensure numerical stability. The perturbation resistance response distribution function is constructed using the ratio of the numerator to the denominator, thus obtaining the perturbation deviation index. The disturbance deviation index is compared with the curvature adjustment factor. The ratio of the original perturbation deviation index (scaled and then increased by one) is used to construct the guidance response rate function, and the perturbation guidance rate is calculated. The perturbation guidance rate is applied to the original perturbation fusion representation to construct the guidance structure coding mechanism, thereby obtaining the structure coding result. Finally, the fusion structure encoding is obtained. The third dataset is divided into a training set and a prediction set in a 7:3 ratio.

[0021] Furthermore, to address the issues of insufficient response adaptability and limited condition-driven capability of modal perturbation fusion results in downstream prediction, this invention proposes a condition-guided structure encoding module. This module aims to achieve stable representation and adaptive condition injection of dynamic perturbation structures. First, the ratio of the sum of squared Euclidean distances between the modal perturbation fusion representations to the sum of squared global energy norms is calculated using the anti-perturbation response distribution function, constructing a perturbation deviation index to quantify structural differences between modes. Then, a saturated nonlinear transformation is applied to the deviation index using a guided response rate function to generate a perturbation guidance rate with boundary convergence characteristics. Finally, the perturbation guidance rate is multiplied by the original perturbation fusion representation through a guided structure encoding mechanism, forming a structure encoding result with adaptive adjustment capabilities. The final output is the fusion structure encoding as a third dataset. This module can transform structural differences in multimodal perturbation fusion results into stable condition-guided signals, significantly improving the monitoring model's ability to perceive the dominant perturbation mode and the accuracy of dynamic extrapolation.

[0022] Preferably, in step S5, a trend-guided adjustment coefficient is calculated based on the training set, a weighted fusion prediction mechanism is constructed by combining the structure encoding results with the weighted average of the structure encoding, and a loss function is designed for training optimization;

[0023] Based on the structural coding results Introducing stability tuning hyperparameters Calculate the trend-guided adjustment coefficient The structural encoding result at the current moment is obtained through the trend-guided adjustment coefficient. Weighted average of structural coding Weighted fusion is performed, specifically, the trend-guided adjustment coefficient is adjusted. and The product of, same The complement and the weighted average of the structure encoding The products of these factors are added together to construct the weighted fusion prediction mechanism, yielding the model's predicted values. The loss function is constructed based on the weighted mean square error to measure and optimize the difference between the model's predicted values ​​and the actual observed values.

[0024] Furthermore, to achieve accurate prediction of dynamic changes in canal flow, this invention constructs a multi-dimensional canal flow monitoring model and establishes a complete training and optimization process. First, based on the structural encoding results, a trend-guided adjustment mechanism is introduced. A trend-guided adjustment coefficient is constructed through stability adjustment hyperparameters to quantify the degree of trend dominance at the current moment. Then, a weighted fusion prediction mechanism is proposed, using the trend-guided adjustment coefficient as a dynamic weight to adaptively weight and fuse the current structural encoding results with the historical state mean, achieving a balance between short-term disturbance response and long-term steady-state prediction. Finally, a loss function based on weighted mean square error is constructed, and the model parameters are iteratively optimized through a gradient descent algorithm, ultimately obtaining a trained multi-dimensional canal flow monitoring model. Through the collaborative prediction mechanism of trend perception and historical equilibrium, the module significantly improves the model's adaptability and prediction accuracy under non-stationary flow signals.

[0025] Preferably, in step S6, the prediction set is input into the trained multi-dimensional monitoring model of irrigation canal flow in farmland, and the final output is the predicted flow value.

[0026] In summary, this invention proposes a multidimensional monitoring method for irrigation canal flow in farmland. The method comprises a modal asymmetric potential energy response encoding module, a modal phase perturbation-guided fusion module, a condition-guided structure encoding module, and a monitoring module. First, the modal asymmetric potential energy response encoding module transforms the original multimodal monitoring data into an enhanced state integration tensor, completing the encoding construction from raw data to dynamic response features and solving the modeling challenge of asynchronous and asymmetric coupling of multimodal responses. Next, the modal phase perturbation-guided fusion module performs direction-aware fusion of the encoded features, realizing the direction-aware fusion of intermodal responses. The invention employs an adaptive weighting and conflict resolution mechanism to address differences. Then, a condition-guided structural coding module transforms the fusion results into a structured code capable of perturbation sensing, establishing an explicit mapping between dynamic perturbation structures and prediction conditions. Finally, a monitoring module enables adaptive prediction based on trend dominance and historical equilibrium, ultimately achieving high-precision, multi-dimensional prediction of canal flow. Through the collaborative processing of these modules, the invention effectively overcomes the limitations of traditional methods in adapting to directional conflicts and expressing dynamic perturbations in the fusion of multi-source heterogeneous monitoring data, providing a precise and reliable flow monitoring solution for farmland irrigation systems. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the steps of a multidimensional monitoring method for irrigation canal flow in farmland.

[0028] Figure 2 Structure diagram of a multidimensional monitoring model for irrigation canal flow in farmland.

[0029] Figure 3 This is a structural diagram of the modal asymmetric potential energy response encoding module.

[0030] Figure 4 This is a structural diagram of the modal phase perturbation-guided fusion module.

[0031] Figure 5 This is a structural diagram of the conditional guidance structure coding module.

[0032] Figure 6 This is a diagram of the model training process.

[0033] Figure 7 The figure shows the fitting effect of the model on the prediction of irrigation canal flow in farmland. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Please see Figures 1-7 This invention provides a technical solution: a method for multi-dimensional monitoring of irrigation canal flow in farmland, comprising a modal asymmetric potential energy response encoding module, a modal phase perturbation-guided fusion module, a condition-guided structure encoding module, and a monitoring module. First, the modal asymmetric potential energy response encoding module performs dynamic feature extraction and enhanced encoding on multi-source monitoring data. Next, the modal phase perturbation-guided fusion module is used to achieve direction-aware multi-modal adaptive fusion. Then, the condition-guided structure encoding module constructs perturbation-sensitive structure encoding. Finally, the monitoring module is used to achieve high-precision prediction of irrigation canal flow in farmland. Specific steps are as follows: Figure 1 As shown.

[0036] Construct a multi-dimensional monitoring model structure for irrigation canal flow in farmland, as shown in the diagram. Figure 2 As shown.

[0037] S1. Collect the water flow of irrigation canals in farmland to construct the original dataset.

[0038] Furthermore, the dataset of this invention contains 2000 irrigation canal flow data for farmland, including five types of farmland irrigation monitoring modal data. Data was collected using an IoT sensor network, meteorological monitoring stations, and a hydrological monitoring platform, including canal flow velocity sequences, water level sequences, soil moisture sequences, meteorological rainfall sequences, and ambient light intensity sequences. The canal flow velocity sequences include instantaneous velocity of water flow within the canal, cross-sectional average velocity, and velocity change rate. The instantaneous velocity of water flow within the canal is continuously collected using an ultrasonic current meter at a sampling interval of 5 minutes, with a measurement accuracy of ±0.01 m / s. The cross-sectional average velocity is collected three times per hour at a standard cross-section using a Doppler profile current meter. The flow rate was calculated using an integral algorithm with an accuracy of ±0.02 m / s; the velocity change rate was calculated based on the instantaneous velocity difference over 10 consecutive sampling periods, with a resolution of 0.001 m / s²; the water level height sequence included real-time water level height, reference water level deviation, and water level change trend slope. Real-time water level height was collected once per minute by a pressure-type water level sensor, with a measurement error of less than ±2 mm; the reference water level deviation was calculated based on the difference between the real-time water level and the design water level elevation, with an accuracy of ±1 mm; the water level change trend slope was obtained based on the linear fitting slope of water level data over the past hour, with a resolution of 0.1 mm / h; the soil moisture sequence included soil volumetric water content and stratified moisture gradient. The study included measurements of soil moisture content and saturation water deficit rate. Soil volumetric water content was synchronously collected once per hour at depths of 10cm, 30cm, and 50cm using a frequency domain reflectance sensor, with a measurement accuracy of ±2%. Stratified humidity gradients were calculated based on the humidity difference between adjacent depth layers, with a resolution of 0.1% / cm. Soil saturation water deficit rate was calculated based on the difference between the soil's maximum water holding capacity and the measured volumetric water content, with an accuracy of ±1%. Meteorological precipitation sequences included regional cumulative rainfall, rainfall intensity, and short-term rainfall forecast data. Regional cumulative rainfall was collected per rainfall event using a tipping bucket rain gauge, with a resolution of 0.1mm. Rainfall intensity was calculated based on minute-by-minute rainfall data, with an accuracy of ±0.5mm / h. Short-term rainfall forecast data is obtained from the meteorological department's application programming interface (API) to provide quantitative precipitation forecasts for the next 6 hours, with a spatial resolution of 1 km × 1 km. The ambient light intensity sequence includes total solar irradiance, photosynthetically active radiation (RALED) intensity, and cumulative sunshine duration. Total solar irradiance is collected once per minute by a total radiation sensor, with a measurement range of 0-2000 W / m² and an accuracy of ±5%. RLED intensity is collected three times per hour at crop canopy height using a quantum sensor, with an accuracy of ±2 μmol / m² / s. The cumulative sunshine duration is calculated based on the duration of daily total radiation exceeding 120 W / m², with a resolution of 1 minute. This data forms the original dataset of farmland irrigation canal flow.

[0039] S2. Based on the original dataset, construct a perturbation difference tensor to quantize the perturbation intensity, design an energy sensing function based on the perturbation difference tensor, construct inter-modal asymmetric coupling weights using an asymmetric coupling function, construct modal reflection coding through weighted fusion, and introduce a nonlinear reflection enhancement term to obtain the first dataset.

[0040] Furthermore, in step S2, a modal asymmetric potential energy response encoding module is constructed for the dynamic responses of different modes. After acquiring multidimensional modal input data, the response differences between modes under perturbation are identified, and modal conditional information encoding that can be used for subsequent anomaly detection and comparative learning is generated. The structure and flow are as follows: Figure 3 As shown, the modalities refer to monitoring data from different sources, including canal flow velocity sequences, water level sequences, soil moisture sequences, meteorological rainfall sequences, and ambient light intensity sequences. Each modality is defined at a specific time step. The input is represented as a vector. The specific implementation steps are as follows:

[0041] To obtain the changes of each mode between adjacent time intervals, a perturbation difference tensor for each mode is constructed. For any mode... At time step The original input vector The perturbation difference tensor is obtained by calculating the difference between the current time and past time. The mathematical model is:

[0042] ;

[0043] In the formula, For modality At time step The original input vector, The disturbance time interval is taken in this embodiment. , representing the difference between adjacent time frames, quantifies the perturbation intensity of modes in a short time scale by characterizing the perturbation amplitude of modal features during the temporal evolution process, and provides a foundation for subsequent autonomous energy construction.

[0044] Furthermore, the perturbation difference tensor is input into the self-sensing function, and the result is calculated... The square term of the L2 norm is used, and a minimal positive number is introduced as a stability compensation. The reciprocal of the sum of the two terms is used to obtain the self-energy value. It is used to measure the trend and non-stationarity of the mode itself. The mathematical model is as follows:

[0045] ;

[0046] In the formula, For modality At time step The self-energy value is used to reflect the relative stability of modal perturbations. In this embodiment, since all monitored quantities are real scalars, the L2 norm is equivalent to the square of the absolute value. Therefore, the calculation process is directly the reciprocal of the sum of the square of the perturbation difference tensor and the smallest positive number. To ensure the denominator is a very small positive number and prevent it from being zero, this embodiment uses... If the perturbation difference is small, it indicates that the physical quantity changes slowly, and the square of the second norm will approach 0, and the final self-energy value will approach 1, indicating that the mode is currently in a stable state. If the perturbation difference is large, the self-energy value will decrease significantly, indicating that the mode is currently in an unstable state. Using a reciprocal structure can amplify the influence of small perturbation modes and suppress the influence of violent perturbation modes, making it easier for the subsequent coupling process to capture the mode states that are truly anomalous.

[0047] Furthermore, to describe the mutual influence between different modes within the same time step, this embodiment constructs an asymmetric coupling function based on potential energy difference and perturbation difference. It calculates the product of the absolute value of the self-energy difference between modes and the Euclidean distance between the perturbation difference tensors, and introduces a stability compensation term composed of the self-energy product to construct the numerator of the asymmetric coupling function. By normalizing the numerator with the sum of the difference measures of all mode pairs relative to the current mode, the asymmetric coupling weight is finally obtained. The mathematical model is:

[0048] ;

[0049] In the formula, For time step Time modality For modes Asymmetric coupling weights, For modality At time step The self-energy value, For modality At time step The self-energy value, For modality At time step The perturbation difference tensor, For modality At time step The perturbation difference tensor, For modality At time step The perturbation difference tensor, The total number of modes is 5 in this embodiment, corresponding to the canal flow velocity sequence, water level sequence, soil moisture sequence, meteorological rainfall sequence, and ambient light intensity sequence.

[0050] Furthermore, in this embodiment, the asymmetric coupling weights and perturbation difference tensors are used as input terms to construct the modal reflection code through weighted aggregation. The mathematical model is:

[0051] ;

[0052] Subsequently, the modal reflection code is multiplied by the self-energy value and then input into the hyperbolic tangent activation function. This is then multiplied element-wise with the original modal reflection code and multiplied by the reflection nonlinear enhancement coefficient. Generate the nonlinear reflection enhancement term, and then combine the nonlinear reflection enhancement term with... Linear superposition yields the conditional information encoding vector The mathematical model is:

[0053] ;

[0054] In the formula, For modality The conditional information encoding vector, The reflection nonlinearity enhancement coefficient, with a value of 0.2, is used to adjust the relative influence between the linear and nonlinear components, enhance the stability of the encoding, and prevent the nonlinear term from amplifying noise. This makes it suitable for the practical characteristics of canal flow monitoring, where flow is stable most of the time but fluctuates drastically a few times. This represents element-wise multiplication. The hyperbolic tangent function is represented by a linear part that reflects the coupling and synthesis information between modes, and a nonlinear part that reflects the state sensitivity of the modes themselves, ultimately yielding the enhanced state integration tensor. As the first dataset.

[0055] S3. The first dataset is processed by the unit angle mapping function to obtain the unit angle. Based on the unit angle, a nonlinear perturbation weighting function is constructed to generate the fusion weight factor. A modal perturbation enhancement expression mechanism is designed to fuse modal feature representations to obtain the second dataset.

[0056] Furthermore, to enhance the modeling ability of directional differences between different modes and improve the discriminative and dynamic adaptability of the fused representation during multimodal collaborative analysis, this example proposes a modal phase perturbation-guided fusion module, the structure of which is as follows: Figure 4 As shown, the specific steps are as follows:

[0057] First, to achieve dynamic collaborative modeling of multidimensional canal flow signals, it is necessary to address the coordination difficulties caused by the directional differences among various sensing modes during temporal evolution. Therefore, to accurately capture the differences in response direction among modes at the same time step and improve the early perception capability of abnormal hydrological behavior, this example constructs a unit angle mapping function between mode pairs based on vector geometric relationships. The inner product of the conditional information encoding vectors of the two modes is calculated as the numerator, and the product of the minimum positive number and the L2 norm of the two vectors is introduced as the denominator. The ratio of the numerator to the denominator is input into the inverse cosine function to obtain the unit angle. The mathematical model is as follows:

[0058] ;

[0059] In the formula, For modality and At time step The unit angle between them For modality The conditional information encoding vector, It should be a very small positive number to prevent the denominator from being zero.

[0060] Furthermore, in this embodiment, Representing modes With mode The dot product of vectors at the current time step measures the similarity of the directions of two modal responses. If the two modal vectors are oriented in the same direction, the dot product value is close to the modal norm product; if the directions are not aligned, the dot product value tends to be negative, reflecting directional conflict. Using the inverse cosine function, the similarity of the normalized vectors is mapped to the angle value, thus obtaining a clear geometric quantification index for judging the degree of response deviation between modes. The denominator... The product of the Euclidean norms of the modal vectors is used to normalize the inner product value. In this embodiment, it is used. .

[0061] Furthermore, to enhance the fusion process's sensitivity to differences in intermodal disturbance directions, enabling it to effectively suppress and strengthen multimodal information even when the propagation directions of canal flow disturbances are inconsistent, this embodiment uses the unit included angle... The nonlinear perturbation weighting function is constructed by inputting an exponential function, and an angle attenuation control coefficient is introduced. Generate the fusion weight factor The mathematical model is:

[0062] ;

[0063] In the formula, Representing modes For modes At time step The fusion weighting factor is used to dynamically adjust the mode. In relation to mode The degree of contribution in the information fusion process. The angle attenuation control coefficient controls the intensity of the impact of angle changes on the fusion weights; a larger value indicates higher angle sensitivity.

[0064] Furthermore, in this embodiment, the angle attenuation control coefficient Setting it to 0.8 improves the ability to identify directional divergences while ensuring convergence stability. When it approaches 0, it represents a mode. With mode The response directions are highly consistent, meaning that the two react synchronously to the disturbance trend. Then it approaches 1, the fusion weights approach the maximum value, and the contribution of modal information to the final decision is enhanced; when Approaching At this time, the modal response directions are approximately opposite, exhibiting significant directional deviation and relative conflict. Approaching 0, it automatically suppresses interference from modal information, preventing the introduction of meaningless or even misleading fusion results. The nonlinear perturbation weighting function has good directional sensitivity and suppression, and can be effectively used in farmland irrigation flow monitoring for modal combinations with inconsistent response directions due to differences in sensor types, thereby improving the accuracy and robustness of modeling modal coordination capabilities under sudden hydrological events.

[0065] Furthermore, to enhance the modal fusion process's responsiveness to differences in disturbance direction among different canal monitoring modes and improve the dynamic identification of inconsistent flow trends, the modal disturbance enhancement expression mechanism is constructed by superimposing the directional difference vectors of all other modes with the current mode, using the conditional information encoding vector of the current mode as a benchmark. This mechanism is then adjusted by the fusion weight factor to form a disturbance fusion representation. The mathematical model is:

[0066] ;

[0067] In the formula, Representing modes At time step The perturbation fusion representation, as an enhanced feature used for trend recognition after the fusion of multi-source sensor information, Modal The conditional information encoding vector, Representing modes With mode The directional differences in the response space reflect the inconsistency in the perturbation trends between modes. Through the above operations, a stable and direction-discriminative modal fusion representation can be constructed to address the inconsistency in the directional responses of multimodal sensors at spatial nodes. Finally, the perturbation representations of all modes are fused to construct a fusion tensor. As the second dataset.

[0068] S4. Based on the second dataset, calculate the perturbation deviation index, construct the anti-perturbation response distribution function, use the guided response rate function to obtain the perturbation guidance rate, propose a guided structure encoding mechanism to integrate the perturbation fusion representation and the perturbation guidance rate to obtain the third dataset, which is divided into a training set and a prediction set.

[0069] Furthermore, to enhance the response adaptability and condition-driven capability of modal perturbation fusion results in downstream prediction, this embodiment proposes a condition-guided structure encoding module to replace the explicit condition injection operation performed by splicing and attention in traditional deep structures. This addresses the problems of insufficient representation of dynamic perturbation structures and weak generalization ability in existing methods. The specific steps are as follows:

[0070] To accurately identify the degree of perturbation difference between modes, the sum of squared Euclidean distances between the perturbation fusion representation of the current mode and the perturbation fusion representations of all other modes is calculated as the numerator. The sum of squared L2 norms of the perturbation fusion representations of all modes is calculated, and a minimal positive number is introduced as the denominator to ensure numerical stability. The perturbation resistance response distribution function is constructed using the ratio of the numerator to the denominator, thus obtaining the perturbation deviation index. The mathematical model is:

[0071] ;

[0072] In the formula, For modality In time The perturbation deviation index indicates that the larger the value, the more significant the deviation of the perturbation structure from other modes. Representing modes At time step The perturbation fusion representation on the above, It should be a very small positive number to prevent the denominator from being zero.

[0073] Furthermore, in this embodiment Molecular part This represents the Euclidean difference in the disturbance vector between the current mode and other modes. The denominator is used to normalize the disturbance scale to avoid bias caused by differences in mode energy strength. Through the above design, the response heterogeneity of each mode in the canal flow monitoring at the same time step can be effectively identified, a disturbance mapping with directional distribution information can be constructed, and the dynamic adaptability under the background of multi-source mode redundancy and interference can be improved.

[0074] Furthermore, to further enhance the nonlinear response characteristics of the disturbance deviation index and prevent extreme disturbances from causing instability in the guiding structure, the disturbance deviation index is combined with a curvature adjustment factor. The ratio of the original perturbation deviation index (scaled and then increased by one) is used to construct the guidance response rate function, and the perturbation guidance rate is calculated. The mathematical model is:

[0075] ;

[0076] In the formula, For modality In time The perturbation guidance rate. This is a curvature adjustment factor used to control the saturation range and boundary convergence of the guiding rate. In this embodiment, With the value set to 1.2, the above structural design enables the disturbance response to achieve both discriminability and stability across different modes, making it particularly suitable for the structural heterogeneous modeling requirements between multi-source irrigation canal monitoring nodes in farmland irrigation systems.

[0077] Furthermore, the perturbation guidance rate is applied to the original perturbation fusion representation to construct the guidance structure encoding mechanism, thereby obtaining the structure encoding result. The mathematical model is:

[0078] ;

[0079] In the formula, For modality The structure encoding result guided by the perturbation guidance rate adaptively adjusts the conditional expression ability of different modes, ultimately achieving modeling of the directional sensitivity and driving ability of the modal perturbation fusion result. This enables subsequent diffusion prediction structures to be extrapolated based on the dominant perturbation mode, resulting in the fused structure encoding. As the third dataset, the third dataset is divided into training and test sets in a 7:3 ratio.

[0080] S5. Calculate the trend-guided adjustment coefficient based on the training set, construct a weighted fusion prediction mechanism by combining the structure encoding results and the weighted average of the structure encoding, and design a loss function for training optimization.

[0081] Furthermore, this invention proposes a monitoring module that inputs the training set into a multi-dimensional monitoring model for irrigation canal flow in farmland. The model employs the PyTorch deep learning framework and is accelerated using an NVIDIA V100 32GB GPU. During training, the batch size is set to 128. To achieve dynamic prediction of multi-dimensional flow signals in irrigation canals during farmland irrigation and further improve the response capability to abnormal fluctuations and changes in water source regulation demand, a multi-dimensional monitoring model for canal flow is proposed. First, based on the aforementioned structural encoding results... Introducing stability tuning hyperparameters Calculate the trend-guided adjustment coefficient The mathematical model for determining the dominance of quantitative trends in current forecasts is as follows:

[0082] ;

[0083] In the formula, The stability adjustment hyperparameter is used to control the sensitivity of the trend adjustment factor when... When the value approaches 0, the model is more sensitive to trend-driven changes. When the value approaches 1, the trend is suppressed, and the prediction is more biased towards the historical mean. The initial value is 0.3. During the training phase, the value is optimized through cross-validation to adapt to the sensitivity requirements of different datasets for trend response. The above design realizes the adaptive fusion of trend enhancement prediction and historical mean prediction, providing dynamic adjustment capability for subsequent training modules.

[0084] Furthermore, to enhance the monitoring model's ability to dynamically balance historical mean and trend disturbances, and to improve its adaptability under non-stationary disturbance flow signals, a weighted fusion prediction mechanism based on a trend adjustment factor is proposed. This mechanism uses the trend-guided adjustment coefficient to encode the structural result at the current moment. Weighted average of structural coding Weighted fusion is performed, specifically, the trend-guided adjustment coefficient is adjusted. and The product of, same The complement and the weighted average of the structure encoding The products of these factors are added together to construct the weighted fusion prediction mechanism, yielding the model's predicted values. The mathematical model is:

[0085] ;

[0086] In the formula, For modality exist Predicted output at time step For modality The structure-encoded weighted average, through the above design, not only retains the ability to respond quickly to short-term disturbances, but also enhances the robust predictive performance in the long-term historical context.

[0087] Furthermore, to minimize the deviation between the model's predicted values ​​and the actual observed labels, and to improve the accuracy and stability of the prediction results, this embodiment constructs the loss function based on the weighted mean square error to measure and optimize the difference between the model's predicted values ​​and the actual observed values. The mathematical model is as follows:

[0088] ;

[0089] In the formula, These are actual observations. This represents the average squared error of all channels, serving as the loss benchmark for model parameter updates. During training, the loss function is calculated using backpropagation gradients in batches. The optimizer uses Adam for gradient descent, with a learning rate set to 0.001. The training process is as follows: Figure 6 As shown in the figure, the number of training rounds is 200. It can be seen from the figure that as the number of training rounds increases, the model converges and stabilizes during the training process. Finally, a well-trained multi-dimensional monitoring model of irrigation canal flow in farmland is obtained.

[0090] S6. The prediction set is input into the trained multi-dimensional monitoring model of irrigation canal flow in farmland, and the final output is the predicted flow value.

[0091] Furthermore, the multi-dimensional monitoring model for irrigation canal flow prediction achieves the following fitting effect: (see figure). Figure 7 As shown in the figure, the horizontal axis represents time, the vertical axis represents flow rate, the solid dotted line represents the actual observed flow rate, and the dashed cross line represents the model's predicted flow rate. It can be seen from the figure that the trends of the predicted values ​​and the actual observed flow rates are roughly similar. The experimental results show that the multi-dimensional monitoring model for irrigation canal flow can effectively capture the trend of canal flow and can predict the canal flow for irrigation in farmland relatively well.

Claims

1. A method for multidimensional monitoring of irrigation canal flow in farmland, characterized in that: Collect the flow rate of irrigation canals in farmland to construct the original dataset; Based on the original dataset, a perturbation difference tensor is constructed to quantize the perturbation intensity. An energy sensing function is designed based on the perturbation difference tensor. An asymmetric coupling function is used to construct intermodal asymmetric coupling weights. A modal reflection code is constructed through weighted fusion. A nonlinear reflection enhancement term is introduced to process and obtain the first dataset. The first dataset is processed by a unit angle mapping function to obtain a unit angle. A nonlinear perturbation weighting function is constructed based on the unit angle to generate a fusion weighting factor. A modal perturbation enhancement expression mechanism is designed to fuse modal feature representations to obtain a second dataset. Based on the second dataset, the perturbation deviation index is calculated, the anti-perturbation response distribution function is constructed, the perturbation guidance rate is obtained using the guidance response rate function, and a guidance structure encoding mechanism is proposed to integrate the perturbation fusion representation and the perturbation guidance rate to obtain the third dataset, which is divided into a training set and a prediction set. Based on the training set, the trend-guided adjustment coefficient is calculated, and a weighted fusion prediction mechanism is constructed by combining the structure encoding results and the weighted average of the structure encoding. A loss function is designed for training optimization. The prediction set is input into the trained multidimensional monitoring model of irrigation canal flow in farmland, and the final output is the predicted flow value.

2. The method for multidimensional monitoring of irrigation canal flow in farmland according to claim 1, characterized in that, The original dataset includes a canal flow velocity sequence, a water level height sequence, a soil moisture sequence, a meteorological precipitation sequence, and an ambient light intensity sequence. The canal flow velocity sequence includes the instantaneous velocity of water flow within the canal, the average velocity across the cross-section, and the rate of change of velocity. The water level height sequence includes real-time water level height, the deviation from the baseline water level, and the slope of the water level change trend. The soil moisture sequence includes soil volumetric water content, stratified moisture gradient, and soil saturation water deficit rate. The meteorological precipitation sequence includes regional cumulative rainfall, rainfall intensity, and short-term rainfall forecast data. The ambient light intensity sequence includes total solar irradiance, photosynthetically active radiation intensity, and cumulative sunshine duration. The original dataset is obtained using these sequence parameters.

3. The method for multidimensional monitoring of irrigation canal flow in farmland according to claim 1, characterized in that, For any mode At time step The original input vector The perturbation difference tensor is obtained by calculating the difference between the current time and past time. The perturbation difference tensor is input into the self-sensing function, and the result is calculated. The square term of the L2 norm is used, and a minimal positive number is introduced as a stability compensation. The reciprocal of the sum of the two terms is used to obtain the self-energy value. The product of the absolute value of the self-energy difference between modes and the Euclidean distance between the perturbation difference tensors is calculated, and a stability compensation term composed of the self-energy product is introduced to construct the numerator of the asymmetric coupling function. By normalizing the numerator with the sum of the difference measures of all mode pairs relative to the current mode, the asymmetric coupling weight is finally obtained. .

4. The method for multidimensional monitoring of irrigation canal flow in farmland according to claim 3, characterized in that, Using the asymmetric coupling weights and perturbation difference tensors as inputs, the modal reflection code is constructed through weighted aggregation. The modal reflection code is multiplied by the self-energy value and then input into the hyperbolic tangent activation function. This is then multiplied element-wise by the original modal reflection code and multiplied by the reflection nonlinear enhancement coefficient. Generate the nonlinear reflection enhancement term, and then combine the nonlinear reflection enhancement term with... Linear superposition yields the conditional information encoding vector Finally, the enhanced state integration tensor is obtained. This serves as the first dataset.

5. A method for multidimensional monitoring of irrigation canal flow in farmland according to claim 1, characterized in that, Based on vector geometric relationships, a unit angle mapping function is constructed between modal pairs. The inner product of the two modal conditional information encoding vectors is calculated as the numerator. The product of a minimum positive number and the L2 norm of the two vectors is added as the denominator. The ratio of the numerator to the denominator is input into an inverse cosine function to obtain the unit angle. .

6. A method for multidimensional monitoring of irrigation canal flow in farmland according to claim 5, characterized in that, The unit included angle The nonlinear perturbation weighting function is constructed by inputting an exponential function, and an angle attenuation control coefficient is introduced. Generate the fusion weight factor Based on the conditional information encoding vector of the current modality, the directional difference vectors of all other modalities and the current modality are superimposed to construct the modal perturbation enhancement representation mechanism. This mechanism is then adjusted by the fusion weight factor to form a perturbation fusion representation. Finally, the fusion tensor is obtained. This serves as the second dataset.

7. A method for multidimensional monitoring of irrigation canal flow in farmland according to claim 1, characterized in that, The sum of squared Euclidean distances between the current modality's perturbation fusion representation and the perturbation fusion representations of all other modalities is calculated as the numerator. The sum of squared L2 norms of the perturbation fusion representations of all modalities is calculated, and a minimal positive number is introduced as the denominator to ensure numerical stability. The perturbation resistance response distribution function is constructed using the ratio of the numerator to the denominator, thus obtaining the perturbation deviation index. .

8. A method for multidimensional monitoring of irrigation canal flow in farmland according to claim 7, characterized in that, The disturbance deviation index and the curvature adjustment factor are combined. The ratio of the original perturbation deviation index (scaled and then increased by one) is used to construct the guidance response rate function, and the perturbation guidance rate is calculated. The perturbation guidance rate is applied to the original perturbation fusion representation to construct the guidance structure coding mechanism, thereby obtaining the structure coding result. Finally, the fusion structure encoding is obtained. The third dataset is divided into a training set and a prediction set in a 7:3 ratio.

9. A method for multidimensional monitoring of irrigation canal flow in farmland according to claim 1, characterized in that, Based on the structural coding results Introducing stability tuning hyperparameters Calculate the trend-guided adjustment coefficient The structural encoding result at the current moment is obtained through the trend-guided adjustment coefficient. Weighted average of structural coding Weighted fusion is performed, specifically, the trend-guided adjustment coefficient is adjusted. and The product of, same The complement and the weighted average of the structure encoding The products of these factors are added together to construct the weighted fusion prediction mechanism, yielding the model's predicted values. .

10. A method for multidimensional monitoring of irrigation canal flow in farmland according to claim 9, characterized in that, The loss function is constructed based on the weighted mean square error to measure and optimize the difference between the model prediction and the actual observation, thereby obtaining the trained multi-dimensional monitoring model of irrigation canal flow in farmland, and finally outputting the flow prediction value.

Images