Multi-modal fusion reinforcement learning control method for agrochemical operation of humanoid robot

Abstract

The invention discloses a multi-modal fusion reinforcement learning control method for humanoid robot agrochemical operation, and relates to the technical field of agricultural robot control. The problems that in the prior art, multi-modal data fusion is poor, crop demand prediction lags behind, and dynamic environment adaptability is weak can be at least partially solved. The method comprises the steps of collecting soil, crop and environment multi-source data and converting the data into uniform feature vectors; based on the dynamic modal weight, attention weighted fusion is carried out on various types of data; carrying out noise filtering by adopting Kalman filtering and abnormal value detection; predicting future crop demands through the improved time sequence prediction model and performing error correction; determining a job execution sequence based on the demand priority function; dynamically optimizing operation parameters according to a real-time environment by adopting a reinforcement learning algorithm; and the operation precision, efficiency and environmental protection are balanced through a multi-target reward function. According to the invention, multi-dimensional perception and full-process intelligent control are realized, and the method has dynamic environment adaptation capability and operation strategy self-optimization capability.

Description

Technical Field

[0001] This invention relates to the field of agricultural robot control technology, specifically to a multimodal fusion reinforcement learning control method for humanoid robots performing agricultural operations. Background Technology

[0002] Current algorithms for humanoid robots in agricultural operations mostly rely on single sensor feedback and fixed rule control, which presents significant technical bottlenecks in complex agricultural scenarios and makes it difficult to meet the needs of precise and intelligent agricultural operations.

[0003] In terms of multimodal data fusion, existing algorithms mostly process visual, soil, and meteorological data separately without establishing correlation mappings between the data. Judging pests and diseases solely based on leaf images while ignoring the impact of soil moisture on disease transmission leads to a misjudgment rate of over 15% for crop needs, resulting in insufficient targeting of agrochemical solutions.

[0004] In terms of crop demand forecasting, existing algorithms rely on real-time data to trigger agrochemical operations, and cannot predict future demand based on crop growth sequence characteristics. Potassium deficiency in wheat during the grain-filling stage needs to be supplemented 3 to 5 days in advance, but existing algorithms only respond after deficiency symptoms appear, missing the optimal intervention time and resulting in crop yield reduction rates exceeding 8%.

[0005] Regarding dynamic environment adaptation, existing algorithms use fixed operating parameters and cannot dynamically adjust according to real-time environmental changes. When the wind speed is greater than 3m / s, the pesticide drift rate exceeds 30%, and when operating in hilly areas, the crop under-spray rate exceeds 12%.

[0006] In terms of optimizing operational strategies, existing algorithms rely on manually preset rules or traditional optimization algorithms, which have slow iteration and convergence speeds, requiring 30 to 60 seconds for a single optimization, and are prone to getting trapped in local optima. In multi-crop mixed planting scenarios, the optimization error of agrochemical dosage exceeds 20%.

[0007] Therefore, how to establish a humanoid robot agricultural operation control method that integrates multimodal dynamic fusion, temporal prediction, and dynamic optimization to achieve high-precision operation and dynamic environment adaptation is a core problem that has not yet been solved in the field of agricultural intelligent equipment. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a multimodal fusion reinforcement learning control method for humanoid robot agricultural operations.

[0009] To achieve the above objectives, the present invention provides a multimodal fusion reinforcement learning control method for humanoid robots engaged in agricultural operations, comprising: Collect soil data, crop data, and environmental data, and convert the soil data, crop data, and environmental data into feature vectors respectively; The modal weights are determined based on soil type, crop growth stage, and environmental stability. Based on these modal weights, various feature vectors are weighted and fused to generate fused feature values. Kalman filtering is used to smooth the fused feature values, and outlier detection is used to remove data that exceeds a preset range; The filtered fused feature values ​​are input into the time series prediction model, and the crop demand prediction value is output. The sliding window error correction mechanism is used for correction. Priority scores are calculated based on demand gap, crop sensitivity, and environmental impact, and the order of operations is determined according to these priority scores. A state space containing the fused features, demand predictions, and operational parameters is constructed, and an action space containing spraying pressure, walking speed, and nozzle angle is defined. A reinforcement learning algorithm is used to optimize the operational parameters based on the reward function. Based on the operation parameters, the humanoid robot performs agricultural operations and calculates reward values, updates the action value function, and determines convergence when the change in the action value function is less than a preset threshold.

[0010] Furthermore, the dimension allocation of the feature vector is as follows: Soil data comprises 32 dimensions, crop data comprises 64 dimensions, and environmental data comprises 32 dimensions, totaling 128 dimensions; The soil data includes nitrogen content, phosphorus content, potassium content, pH value, and soil moisture. The crop data includes leaf spectral data, disease and pest image data, plant height, and new shoot length; The environmental data includes wind speed, temperature, light intensity, and rainfall.

[0011] Furthermore, the weighted fusion formula is characterized by: ; in, The fusion feature value, , , These are the soil modal weights, the crop modal weights, and the environmental modal weights, respectively. , , These are soil feature vectors, crop feature vectors, and environmental feature vectors, respectively. , , These are the soil mapping matrix, crop mapping matrix, and environment mapping matrix, respectively. The soil modal weights range from 0.3 to 0.5, the crop modal weights range from 0.4 to 0.6, and the environmental modal weights range from 0.1 to 0.2.

[0012] Furthermore, the formula for the Kalman filter is: ; in, This is the filtered data. For predicted values, For Kalman gain, These are the actual collected values; The Kalman gain ranges from 0.1 to 0.3, and the predicted value is calculated based on the average of the first 5 collected data. The outlier detection method is the 3σ criterion. When a certain collected value exceeds the range of the average plus or minus 3 times the standard deviation, it is determined to be an outlier and removed. The average of the previous 3 collected data is used to replace the outlier.

[0013] Furthermore, the time-series prediction model is an improved long short-term memory network model, comprising an input layer, a hidden layer, a fully connected layer, and an output layer; The input layer receives a 128-dimensional fused feature vector; The hidden layer consists of three layers, each containing 256 neurons, with each neuron incorporating attention gating; The fully connected layer contains 128 neurons; The output layer outputs crop demand forecasts for the next 7 days, including nitrogen, phosphorus, and potassium demand gaps, pest and disease risk probabilities, and crop sensitivity.

[0014] Furthermore, the formula for correcting the sliding window error is as follows: ; in, This is the corrected predicted value. These are the initial predicted values. This represents the average error between the initial predicted value and the actual value within the sliding window. The size of the sliding window is set to 7 days.

[0015] Furthermore, the formula for calculating the priority score is as follows: ; in, Score the priority level. Score the size of the demand gap. Score the crop's sensitivity. Score for environmental impact; The score for the size of the demand gap is calculated based on the ratio of the gap to normal demand: 1 point for a ratio of 0% to 10%, 2 points for a ratio of 10% to 20%, 3 points for a ratio of 20% to 30%, and 4 points for a ratio of more than 30%. The crop sensitivity score ranges from 1 to 5. The environmental impact score is calculated based on the degree of environmental impact on the effectiveness of agrochemicals, and the value ranges from 1 to 4 points.

[0016] Furthermore, the action space is defined as follows: The spray pressure adjustment step is 0.05 MPa, and the value range is 0.2 to 0.8 MPa; The walking speed adjustment step size is 0.1 m / s, with a range of 0.3 to 1.2 m / s; The nozzle angle can be adjusted in 5-degree increments, with a range of 0 to 90 degrees.

[0017] Furthermore, the formula for the reward function is: ; in, As a reward value, Bonus for work accuracy. Rewards for work efficiency Punishment for wasting medicine; The operational accuracy bonus is calculated based on the matching degree between agrochemical usage and demand gap. A matching degree of 90% or higher is 10 points, 80% to 90% is 8 points, 70% to 80% is 6 points, and less than 70% is 0 points. The work efficiency bonus is calculated based on the ratio of actual work time to theoretical optimal time; The penalty for drug waste is calculated based on the sum of the drift rate and the missed spray rate.

[0018] Furthermore, the update formula for the action value function is: ; in, For the value of new actions, For the value of old moves, For learning rate, This is the current reward value. As a discount factor, The value of the action in the next state; The learning rate is set to 0.1, and the discount factor is set to 0.9. The strategy is considered convergent when the change in the action value function over 100 consecutive iterations is less than or equal to 0.01.

[0019] The beneficial effects of this invention are as follows: This invention reduces the crop demand misjudgment rate from 15% to below 3% by constructing a multimodal data association and fusion model of soil, crops, and environment, increases the accuracy of pest and disease identification to 96%, and increases the fertilizer application matching degree from 70% to 92%.

[0020] This invention uses an improved time-series prediction model to predict crop demand 3 to 7 days in advance, reducing crop yield reduction from 8% to less than 2%, ensuring timely and effective agrochemical intervention during key growth periods, and reducing the spread rate of pests and diseases from 30% to less than 5%.

[0021] This invention uses a reinforcement learning algorithm to adjust operational parameters in real time. When the wind speed is 3 to 5 m / s, the pesticide drift rate is reduced from 30% to 12%, the missed spraying rate in hilly areas is reduced from 12% to 3%, and the adaptability to complex scenarios is improved by 70%.

[0022] After the strategy optimization algorithm of this invention converges iteratively, the single optimization time is shortened from 30 to 60 seconds to less than 100 milliseconds, the average daily operating area is increased from 2 hectares to 3.5 hectares, and the operating efficiency is improved by 75%. Attached Figure Description

[0023] Figure 1 This is a flowchart of the multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to the present invention; Figure 2 This is a schematic diagram of the module composition of the control system of the present invention; Figure 3 This is a trend chart showing the variation of multimodal data fusion error in this invention. Figure 4 This is a convergence curve of the reinforcement learning strategy of this invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0025] The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations of the present invention is based on a control system.

[0026] See Figure 2 The control system comprises a multimodal perception layer, a temporal prediction layer, a dynamic optimization layer, and an execution control layer. The multimodal perception layer collects multi-source data from soil sensors, hyperspectral cameras, and micro-weather stations, generating a unified state vector through an attention-weighted fusion mechanism. The temporal prediction layer employs an improved long short-term memory network model, predicting future crop requirements based on historical growth data and real-time fused features. The dynamic optimization layer uses reinforcement learning algorithms to dynamically optimize operational parameters according to real-time environmental changes. The execution control layer converts the optimized operational parameters into robot control commands, driving the robotic arm and chassis to perform agricultural operations.

[0027] Example 1 This embodiment uses the application scenario of pest and disease control and foliar fertilizer application before bagging apples in a hilly orchard. The humanoid robot model is JZ-worker01, equipped with a 6-DOF robotic arm with a 5kg end-effector load. The robot's end-effector arm is equipped with a TDR-300 soil sensor and an HSI-400 hyperspectral camera, and a WS-100 miniature weather station is fixed to the top of the robot's body.

[0028] See Figure 1 The control method in this embodiment includes the following steps: Step S1: Multi-source data acquisition and feature transformation.

[0029] The data acquisition module was activated, and the sampling frequencies were set as follows: soil sensor 1 time per second, hyperspectral camera 0.5 times per second, and mini weather station 0.1 times per second. The soil sensor was inserted 8 cm into the soil to collect soil data from the apple tree root zone, yielding nitrogen content of 18 mg / kg, phosphorus content of 12 mg / kg, potassium content of 15 mg / kg, pH value of 6.5, and soil moisture of 25%. The hyperspectral camera, at a distance of 1.5 m, captured vertical images of the apple leaves, identifying a light aphid density of 3 aphids per leaf. The laser rangefinder measured the tree height at 220 cm and the new shoot length at 15 cm. The mini weather station collected data on wind speed (2.5 m / s), temperature (28℃), light intensity (8000 lux), and rainfall (0 mm / h). All three types of data were converted into a 128-dimensional feature vector, with soil data comprising 32 dimensions, crop data 64 dimensions, and environmental data 32 dimensions.

[0030] Step S2: Multimodal dynamic fusion.

[0031] Based on the current scenario, the modal weights are set as follows: soil modality weight 0.4, crop modality weight 0.5, and environment modality weight 0.1. The fusion feature values ​​are then calculated using the attention-weighted fusion formula. .

[0032] in It is a soil feature vector containing nitrogen content 18, phosphorus content 12, potassium content 15, pH value 6.5, and soil moisture 25. It is a crop feature vector containing aphid density of 3, plant height of 220, and new shoot length of 15. It is an environmental feature vector containing wind speed of 2.5, temperature of 28, and light intensity of 8000. , , This is a 128×128 mapping matrix obtained from pre-training. The calculated fused feature values ​​are 128-dimensional vectors, ranging from 0 to 1.

[0033] Step S3: Noise filtering.

[0034] Soil moisture data were smoothed using Kalman filtering, and the Kalman gain was adjusted. Set to 0.2, predicted value Calculated based on the average of the first 5 data collections. Soil moisture was 25% before filtering and 24.8% after filtering. The 3σ criterion was used to detect outliers. The average and standard deviation of the last 10 data collections were calculated. A light intensity of 8000 lux was considered normal and retained if it fell within the range of the average of 7800 lux plus or minus three times the standard deviation of 150. For outliers outside this range, the average of the first 3 data collections was used instead, achieving an outlier removal rate of over 95%.

[0035] Step S4: Crop demand time series forecasting.

[0036] A pre-trained time-series prediction model for apple crops was loaded. This model was trained based on data from apple orchards in Yantai, Shandong Province, from 2019 to 2023, with a validation set accuracy of 92%. Input was fused feature values ​​and historical growth data from the past 7 days, showing plant height increasing from 210cm to 220cm and new shoot length increasing from 12cm to 15cm. The model output initial predictions for the next 5 days: nitrogen demand deficit of 6mg / kg, phosphorus demand deficit of 5mg / kg, potassium demand deficit of 4mg / kg, aphid risk probability of 85%, powdery mildew risk probability of 10%, and crop sensitivity score of 4. The average error over the 7-day sliding window was calculated. The prediction errors for nitrogen demand over the past 7 days were 0.2, 0.3, 0.1, 0.2, 0.3, 0.2, and 0.2 mg / kg, with an average of 0.21 mg / kg. The corrected nitrogen demand deficit was 6 minus 0.21, equaling 5.79 mg / kg, and was set to 5.8 mg / kg.

[0037] Step S5: Prioritize requirements.

[0038] Prioritize aphid control measures based on the following criteria: demand gap size (4 points), aphid density exceeding 30%; crop sensitivity (4 points); environmental impact (3 points), and wind speed of 2.5 m / s (suitable for operation).

[0039] Priority Points are awarded, with 4 points being the highest priority.

[0040] Calculate the priority of foliar fertilizer supplementation: Demand gap size scores 2 points, with a nitrogen gap of 5.8 mg / kg accounting for 23.2% of the normal demand of 25 mg / kg; Crop sensitivity scores 3 points; Environmental impact scores 3 points.

[0041] Priority Points are awarded, with 2 points being the second highest priority.

[0042] Output execution order: Aphid control first, foliar fertilizer application second.

[0043] Step S6: Optimize dynamic job strategy.

[0044] Initial operational parameters were set based on aphid control priorities: spraying pressure 0.5 MPa, walking speed 0.6 m / s, and nozzle angle 45 degrees aimed at the underside of leaves. Initial reinforcement learning parameters were set: learning rate 0.1, discount factor 0.9, and exploration rate 0.2. The state space was constructed as a joint vector of fused feature vectors, nitrogen demand gap of 5.8 mg / kg, aphid risk probability of 85%, spraying pressure of 0.5 MPa, walking speed of 0.6 m / s, and nozzle angle of 45 degrees. The action space was defined as follows: 13 selectable values ​​for spraying pressure from 0.2 to 0.8 MPa, 10 selectable values ​​for walking speed from 0.3 to 1.2 m / s, and 19 selectable values ​​for nozzle angle from 0 to 90 degrees, resulting in a total of 2470 action combinations.

[0045] Step S7: Policy iteration and convergence.

[0046] The robot walks to a plot of land with a 25-degree slope, and the environmental data is updated to a wind speed of 3.2 m / s. Based on an exploration rate of 0.2, it selects the action with the highest current value with a 90% probability: spraying pressure of 0.4 MPa, walking speed of 0.5 m / s, and nozzle angle of 60 degrees. After executing the action, the operation effect is measured: the agrochemical usage matches the demand gap by 92%, resulting in a 10-point bonus for operational accuracy; the actual time is 18 minutes, the theoretical optimal time is 16 minutes, a ratio of 1.12, resulting in a 4-point bonus for operational efficiency; the drift rate of 10% plus the missed spray rate of 2% equals a 12% waste rate, resulting in a 0-point penalty for pesticide waste.

[0047] Calculate reward value Points. Update the action value function: After 500 iterations, the change in the action value function is less than or equal to 0.01 for 100 consecutive iterations, indicating that the strategy has converged. The operational parameters after convergence are: spraying pressure 0.35 MPa, walking speed 0.45 m / s, nozzle angle 50 degrees, and parameter adjustment response time 80 ms.

[0048] Example 2 This embodiment uses precision fertilization and pest control in a greenhouse strawberry cultivation setting as an application scenario. The greenhouse, located in the Shouguang Modern Agricultural Park in Shandong Province, covers an area of ​​2000 square meters and adopts a solar greenhouse structure, equipped with an automatic rolling shutter system and integrated water and fertilizer irrigation facilities. The greenhouse cultivates the Hongyan strawberry variety, with a plant spacing of 25cm and a row spacing of 30cm, totaling 8000 plants. The strawberries are in their flowering and fruiting stage and are highly sensitive to temperature, humidity, and nutrient supply.

[0049] The humanoid robot, model JZ-worker01-G, is specifically designed for greenhouse environments. Its height is adjusted to 1.2m to accommodate the space constraints of greenhouses. It features a lightweight 4-DOF robotic arm with a 3kg end-effector load. The arm's end effector is equipped with a TDR-200S miniature soil sensor and an NIR-350 near-infrared spectroscopy camera, while a temperature and humidity sensor array is fixed to the front of the body. The robot's chassis uses omnidirectional wheels, with a minimum turning radius of 0.3m, allowing for flexible movement between strawberry rows.

[0050] The control method in this embodiment is the same as the steps in Embodiment 1, the difference being the different requirements of parameter configuration and application scenarios.

[0051] In step S1, the soil sensor sampling frequency was configured to twice per second to meet the requirements of precise monitoring of soil moisture changes in the greenhouse environment. The sensor insertion depth was adjusted to 5 cm, corresponding to the shallow root distribution characteristics of strawberries. The collected soil data included: nitrogen content 22 mg / kg, phosphorus content 18 mg / kg, potassium content 20 mg / kg, pH value 6.2, soil moisture 65%, and soil conductivity 1.8 mS / cm. The near-infrared spectral camera shooting distance was adjusted to 1.0 m to suit the relatively short strawberry plants. The camera resolution was set to 1920×1080 pixels, and the frame rate was adjusted to 15 frames per second to capture subtle changes in the leaves. The camera identified mild gray mold symptoms on the strawberry leaves, with an infected area of ​​8%, and also detected a red spider mite density of 2 mites per leaf. The laser rangefinder measured the plant height as 18 cm, the crown diameter as 25 cm, and the fruit maturity coefficient as 0.6. An array of temperature and humidity sensors collected environmental data inside the greenhouse: air temperature 26℃, air humidity 78%, carbon dioxide concentration 450ppm, and photosynthetically active radiation 320μmol / m² / s. Due to the enclosed greenhouse environment, the wind speed was constant at 0.2m / s, and rainfall was zero. The three types of data were uniformly converted into a 128-dimensional feature vector, with soil electrical conductivity and carbon dioxide concentration incorporated as extended features into the corresponding modes.

[0052] In step S2, since the greenhouse environment is relatively stable and controllable, the environmental modal weight is set to a lower limit of 0.1. The soil modal weight is set to 0.45 to strengthen the focus on the water and fertilizer status of the strawberry root zone; the crop modal weight is set to 0.45 to balance the needs of pest and disease identification and nutrient diagnosis. The modal weight allocation formula is adjusted as follows: Mapping matrix , , A pre-trained model specifically for strawberries was used. This model was trained based on historical data from strawberry greenhouses in Shouguang, Shandong Province, from 2020 to 2024, containing 120,000 sets of sample data covering the three main cultivated varieties: Hongyan, Zhangji, and Tianbao. The validation set accuracy reached 94%. After the fused feature values ​​were calculated, the system performed L2 norm normalization on the feature vectors to ensure the comparability of features in each dimension.

[0053] In step S3, the Kalman filter adjusts its parameters to suit the characteristics of the greenhouse environment. Kalman gain. The standard deviation was set to 0.15, lower than the 0.2 used in open-field operations, to accommodate the smaller fluctuations in greenhouse environment data. Soil moisture data was 65% before filtering and 64.8% after filtering. Air humidity data was 78% before filtering and 77.6% after filtering. When using the 3σ criterion for outlier detection, the standard deviation calculation window was expanded to 20 samples to obtain more stable statistical characteristics. When soil conductivity showed a transient jump due to irrigation, the system automatically identified and removed the outlier, replacing it with the average of the first three collected data points, 1.75 mS / cm. The outlier removal rate reached over 97%.

[0054] In step S4, a pre-trained strawberry crop time-series prediction model is loaded. This model is trained based on strawberry greenhouse data from Shouguang, Shandong Province, from 2020 to 2024. The training data covers five growth stages: seedling stage, budding stage, flowering stage, fruit enlargement stage, and harvest stage. The validation set accuracy reaches 93%. Input fused feature values ​​and historical growth data from the past 7 days: plant height increased from 15cm to 18cm, crown diameter increased from 22cm to 25cm, and fruit maturity coefficient increased from 0.4 to 0.6.

[0055] The model outputs initial forecasts for the next 7 days: nitrogen demand deficit 3.2 mg / kg, phosphorus demand deficit 2.8 mg / kg, potassium demand deficit 4.5 mg / kg, gray mold spread risk probability 72%, spider mite reproduction risk probability 45%, powdery mildew risk probability 15%, and crop sensitivity score 5. The average error over the 7-day sliding window is calculated. The potassium demand forecast errors for the past 7 days are 0.3, 0.2, 0.4, 0.3, 0.2, 0.3, and 0.3 mg / kg, with an average of 0.29 mg / kg. The corrected potassium demand deficit is 4.5 mg / kg minus 0.29, which equals 4.21 mg / kg, and is set to 4.2 mg / kg. The gray mold spread risk probability is corrected to 70% using the same method.

[0056] In step S5, the priority for gray mold control is calculated: Demand gap size scores 3 points, with 8% of the gray mold infection area exceeding the 5% warning threshold; Crop sensitivity scores 5 points, as strawberries are extremely sensitive to gray mold during the flowering and fruiting period; Environmental impact scores 4 points, as pesticides can be applied in a windless and humid environment within the greenhouse. Priority Points are awarded, with 4 points being the highest priority. Potassium fertilizer supplementation priority is calculated as follows: Demand gap size: 2 points (potassium gap of 4.2 mg / kg, representing 23.3% of the normal demand of 18 mg / kg); Crop sensitivity: 4 points (high potassium requirement during fruit enlargement); Environmental impact: 4 points. Priority Points are awarded, with 3 points being the second highest priority.

[0057] Calculating the priority of spider mite control: Demand gap size scores 2 points; density of 2 spider mites per leaf is below the control threshold of 5 spider mites per leaf; crop sensitivity scores 3 points; environmental impact scores 4 points. Priority Points are awarded, with 2 points being the third priority. Output execution order: gray mold control first, potassium fertilizer supplementation second, and red spider mite control last.

[0058] In step S6, the initial operating parameters are adjusted to suit the delicate characteristics of strawberry leaves: the spraying pressure is 0.3 MPa, lower than the 0.5 MPa in the apple orchard scenario, to avoid damage to tender leaves and flowers by high-pressure water flow; the walking speed is 0.4 m / s to ensure uniform coverage of the pesticide solution; the nozzle angle is 30 degrees, spraying from the side and below to cover the back of the leaves.

[0059] The action space has been redefined for the greenhouse scenario: spraying pressure has 7 selectable values ​​from 0.2 to 0.5 MPa, with an adjustment step of 0.05 MPa; walking speed has 7 selectable values ​​from 0.3 to 0.9 m / s, with an adjustment step of 0.1 m / s; and nozzle angle has 13 selectable values ​​from 0 to 60 degrees, with an adjustment step of 5 degrees. The total number of action combinations has been adjusted to 637, a significant reduction compared to the 2470 combinations in the open field scenario, to adapt to the constraints of the greenhouse environment. The reinforcement learning parameters remain unchanged: learning rate 0.1, discount factor 0.9, and exploration rate 0.2.

[0060] In step S7, the strategy iteration process is smoother due to the stable environment inside the greenhouse. The robot walks along the strawberry rows, and the environmental data remains relatively constant. Based on an exploration rate of 0.2, the system selects the following actions: spraying pressure of 0.28 MPa, walking speed of 0.35 m / s, and nozzle angle of 25 degrees. After executing the actions, the operational effect is measured: the dosage of pyrimethanil, a gray mold control agent, matches the demand by 95%, resulting in a 10-point bonus for operational accuracy; the actual time is 22 minutes, the theoretical optimal time is 20 minutes, the ratio is 1.1, resulting in a 5-point bonus for operational efficiency; due to the lack of wind inside the greenhouse, the drift rate is 0%, the missed spray rate is 1%, the waste rate is 1%, and the penalty for pesticide waste is 0 points.

[0061] Calculate reward value Because the greenhouse environment is stable, the strategy converged after 350 iterations, which is 30% faster than in the open field scenario.

[0062] Post-convergence operating parameters: spraying pressure 0.25 MPa, walking speed 0.32 m / s, nozzle angle 28 degrees, parameter adjustment response time 65 ms. After a single greenhouse operation, the gray mold control rate reached 96%, potassium fertilizer utilization rate reached 91%, and red spider mite control rate reached 88%.

[0063] Example 3 This embodiment illustrates the performance and comparative verification of the method of the present invention under complex environments. The verification experiments were conducted simultaneously in three typical scenarios: an apple orchard in a hilly area of ​​Yantai, Shandong; a strawberry greenhouse in Shouguang, Shandong; and a wheat field in Zhoukou, Henan. The experimental period lasted for 6 months, covering the key growth stages of the crops.

[0064] See Figure 3 In terms of verifying the accuracy of multimodal fusion, the experiment adopted a double-blind comparison method to compare the multimodal fusion results of the present invention with the results of human expert diagnosis.

[0065] A total of 12,000 sets of valid data were collected in the hilly apple orchard scenario, covering different slope, light, and wind speed conditions. Initially, the fusion error was 15%, comparable to traditional single-sensor methods. With increasing iterations, the fusion error gradually decreased: after 20 iterations, the error dropped to 12%; after 50 iterations, it dropped to 8%; after 80 iterations, it dropped to 5%; and after 100 iterations, the error stabilized below 3%. The error of traditional single-sensor methods fluctuated around 15% throughout the experimental period, without a significant downward trend. Statistical analysis showed that the average absolute error between the crop demand prediction and actual demand using the method of this invention was 0.85 mg / kg, the relative error was 2.8%, and the maximum relative error was 5.2%. The fusion accuracy was even higher in the strawberry greenhouse scenario, with an average relative error of only 2.1%, thanks to the stability of the greenhouse environment. In the field wheat scenario, due to the large area and drastic environmental changes, the average relative error was 3.5%, but this was still significantly better than the 15% error rate of traditional methods.

[0066] In terms of verifying the accuracy of time-series predictions, experiments validated the predictive performance of the improved Long Short-Term Memory (LSTM) network model. Taking the prediction of aphid outbreaks in apple orchards as an example, the model predicted an 85% probability of aphid outbreak risk 5 days in advance, with the actual outbreak time differing from the predicted time by 1 day, resulting in a prediction lead time of 4 days. During the 6-month experimental period, a total of 35 pest and disease events were predicted, of which 32 were accurate, achieving a prediction accuracy rate of 91.4% and an average prediction lead time of 4.2 days. Compared with the traditional real-time triggering method, the average intervention lead time increased from 0 days to 4.2 days, and the timely intervention rate of agrochemicals during key growth periods improved from 65% to 96%. Regarding fertilizer demand prediction, the model achieved prediction accuracies of 93%, 91%, and 89% for nitrogen, phosphorus, and potassium, respectively. The slightly lower accuracy for potassium was attributed to the greater influence of environmental factors on crop potassium demand. The sliding window error correction mechanism was highly effective; the average prediction error before correction was 8%, which was reduced to below 3% after correction, a reduction of 62.5%.

[0067] In terms of dynamic environment adaptability verification, various extreme working conditions were set up for pressure testing. In the wind speed test, when the wind speed gradually increased from 0 m / s to 5 m / s, the agent drift rate of the traditional fixed parameter method rapidly increased from 5% to 35%; when using the method of this invention, the system automatically adjusted the spraying pressure from 0.5 MPa to 0.3 MPa, the nozzle angle from 45 degrees to 60 degrees, and the walking speed from 0.6 m / s to 0.4 m / s, controlling the drift rate within 12%, a reduction of 65.7%. In the slope test, when the slope of the plot increased from 0 degrees to 35 degrees, the missed spray rate of the traditional method increased from 3% to 18%; when using the method of this invention, the system automatically increased the robotic arm compensation angle and adjusted the spraying trajectory, controlling the missed spray rate within 4%, a reduction of 77.8%. In temperature testing, when the greenhouse temperature rose from 20℃ to 38℃, the system automatically adjusted the pesticide concentration coefficient to prevent pesticide damage caused by high temperatures, reducing the incidence of crop pesticide damage from 8% using traditional methods to below 1%. A comprehensive evaluation of its adaptability to complex scenarios shows that the method of this invention is 70% more efficient than traditional methods.

[0068] See Figure 4In verifying the convergence of reinforcement learning strategies, the experiment recorded the trend of reward value changes during strategy optimization. In the hilly apple orchard scenario, the initial reward value was 2 points. After 100 iterations, the reward value increased to 4 points, after 200 iterations it increased to 5.2 points, and after 300 iterations it reached a convergence point of 6 points. From 300 to 500 iterations, the reward value fluctuated slightly between 5.8 and 6.2 points, eventually stabilizing at 6.2 points. The convergence criterion was that the change in the action value function over 100 consecutive iterations was less than or equal to 0.01, which was achieved at 420 iterations. In the strawberry greenhouse scenario, due to the stable environment, convergence was faster, achieving the convergence criterion at 280 iterations, with the final reward value stabilizing at 6.5 points. In the wheat field scenario, due to the large operating area and drastic environmental changes, convergence was slower, achieving the convergence criterion at 550 iterations, with the final reward value stabilizing at 5.8 points.

[0069] In terms of operational efficiency verification, experiments compared the operational efficiency metrics of the method of this invention with those of traditional methods. Regarding parameter optimization time, traditional genetic algorithms require 35 seconds for a single optimization, and traditional particle swarm optimization algorithms require 28 seconds. The reinforcement learning method of this invention requires only 80 to 100 milliseconds for parameter adjustment after policy convergence, representing an efficiency improvement of over 350 times. In terms of average daily operating area, traditional methods, due to frequent downtime for parameter tuning, result in an average daily operating area of ​​2 hectares. The method of this invention, achieving dynamic parameter adjustment, increases the average daily operating area to 3.5 hectares, an improvement of 75%. In multi-crop mixed planting scenarios, traditional methods require manual switching of crop modes, with parameter adjustment time reaching up to 2 hours. The method of this invention, by automatically identifying crop types and loading corresponding models, shortens the parameter switching time to 1 minute, achieving an efficiency improvement of 120 times.

[0070] In terms of human-machine collaborative decision-making verification, experiments simulated unknown disease identification scenarios. When the hyperspectral camera detected abnormal leaf symptoms not present in the training dataset, the system automatically triggered the human-machine collaborative decision-making process. A total of 12 collaborative decisions were triggered in the experiment, 10 of which involved real unknown diseases and 2 were sensor misjudgments. In real unknown disease scenarios, the system completed data upload within an average of 8 seconds, human experts completed diagnosis and issued treatment instructions within an average of 3 minutes, and the robot completed strategy updates and resumed operations within an average of 15 seconds. Compared with the traditional manual inspection and discovery mode, the disease discovery time was shortened from an average of 2 days to 5 minutes, and the disease spread rate decreased from 15% of the traditional method to 2% of the method of this invention, improving decision-making efficiency by more than 500 times. In safety incident scenarios such as pesticide leaks, the robot can detect anomalies and trigger emergency shutdown within 3 seconds, and complete leak point location and reporting within 10 seconds, improving operational safety by 90% compared to traditional methods.

[0071] Regarding the verification of economic benefits, taking a 10-hectare apple orchard as the calculation object, the annual benefit analysis after adopting the method of this invention is as follows: In terms of pesticide savings, due to precision application reducing waste, pesticide usage decreased from 8 kg per hectare using the traditional method to 5.2 kg per hectare, a reduction of 35%, resulting in annual pesticide cost savings of 14,000 yuan. In terms of fertilizer savings, due to precision fertilization reducing losses, fertilizer utilization rate increased from 35% using the traditional method to 58%, resulting in annual fertilizer cost savings of 9,000 yuan. In terms of loss reduction, due to advance prediction and timely intervention, the pest and disease loss rate decreased from 12% using the traditional method to 3%, resulting in annual yield loss reduction of 28,000 yuan. In terms of efficiency improvement, due to a 75% increase in operational efficiency, annual labor costs are saved by 12,000 yuan. The comprehensive annual net benefit is 63,000 yuan. Based on an equipment investment of 150,000 yuan, the investment payback period is 2.4 years. The investment payback period for a strawberry greenhouse is even shorter, at 1.8 years.

[0072] In summary, the embodiments disclosed herein have at least the following technical effects: This invention significantly improves the accuracy of crop demand assessment by constructing a multimodal data association and fusion model of soil, crops, and environment, reducing the demand misjudgment rate from 15% to below 3%.

[0073] This invention enables advance prediction of crop demand through an improved time-series forecasting model, solving the problem of delayed intervention in traditional algorithms and extending the advance time for pest and disease intervention from 0 days to 4.2 days.

[0074] This invention utilizes a reinforcement learning algorithm to dynamically optimize operational parameters, enhancing adaptability to complex environments. At wind speeds of 3 to 5 m / s, the drift rate is reduced from 30% to 12%. The strategy optimization algorithm of this invention significantly improves convergence speed, reducing parameter adjustment response time from 30 seconds to 80 milliseconds, resulting in a 75% increase in operational efficiency.

[0075] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A multimodal fusion reinforcement learning control method for humanoid robot agricultural operations, characterized in that, include: Collect soil data, crop data, and environmental data, and convert the soil data, crop data, and environmental data into feature vectors respectively; The modal weights are determined based on soil type, crop growth stage, and environmental stability. Based on these modal weights, various feature vectors are weighted and fused to generate fused feature values. Kalman filtering is used to smooth the fused feature values, and outlier detection is used to remove data that exceeds a preset range; The filtered fused feature values ​​are input into the time series prediction model, and the crop demand prediction value is output. The sliding window error correction mechanism is used for correction. Priority scores are calculated based on demand gap, crop sensitivity, and environmental impact, and the order of operations is determined according to these priority scores. A state space containing the fused features, demand predictions, and operational parameters is constructed, and an action space containing spraying pressure, walking speed, and nozzle angle is defined. A reinforcement learning algorithm is used to optimize the operational parameters based on the reward function. Based on the operation parameters, the humanoid robot performs agricultural operations and calculates reward values, updates the action value function, and determines convergence when the change in the action value function is less than a preset threshold.

2. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The dimension allocation of the feature vector is as follows: Soil data comprises 32 dimensions, crop data comprises 64 dimensions, and environmental data comprises 32 dimensions, totaling 128 dimensions; The soil data includes nitrogen content, phosphorus content, potassium content, pH value, and soil moisture. The crop data includes leaf spectral data, disease and pest image data, plant height, and new shoot length; The environmental data includes wind speed, temperature, light intensity, and rainfall.

3. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The formula for the weighted fusion is: ； in, The fusion feature value, , , These are the soil modal weights, the crop modal weights, and the environmental modal weights, respectively. , , These are soil feature vectors, crop feature vectors, and environmental feature vectors, respectively. , , These are the soil mapping matrix, crop mapping matrix, and environment mapping matrix, respectively. The soil modal weights range from 0.3 to 0.5, the crop modal weights range from 0.4 to 0.6, and the environmental modal weights range from 0.1 to 0.

2.

4. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The formula for the Kalman filter is: ； in, This is the filtered data. For predicted values, For Kalman gain, These are the actual collected values; The Kalman gain ranges from 0.1 to 0.3, and the predicted value is calculated based on the average of the first 5 collected data. The outlier detection method is the 3σ criterion. When a certain collected value exceeds the range of the average plus or minus 3 times the standard deviation, it is determined to be an outlier and removed. The average of the previous 3 collected data is used to replace the outlier.

5. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The time-series prediction model is an improved long short-term memory network model, including an input layer, a hidden layer, a fully connected layer, and an output layer; The input layer receives a 128-dimensional fused feature vector; The hidden layer consists of three layers, each containing 256 neurons, with each neuron incorporating attention gating; The fully connected layer contains 128 neurons; The output layer outputs crop demand forecasts for the next 7 days, including nitrogen, phosphorus, and potassium demand gaps, pest and disease risk probabilities, and crop sensitivity.

6. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The formula for correcting the sliding window error is as follows: ； in, This is the corrected predicted value. These are the initial predicted values. This represents the average error between the initial predicted value and the actual value within the sliding window. The size of the sliding window is set to 7 days.

7. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The formula for calculating the priority score is as follows: ； in, Score the priority level. Score the size of the demand gap. Score the crop's sensitivity. Score for environmental impact; The score for the size of the demand gap is calculated based on the ratio of the gap to normal demand: 1 point for a ratio of 0% to 10%, 2 points for a ratio of 10% to 20%, 3 points for a ratio of 20% to 30%, and 4 points for a ratio of more than 30%. The crop sensitivity score ranges from 1 to 5. The environmental impact score is calculated based on the degree of environmental impact on the effectiveness of agrochemicals, and the value ranges from 1 to 4 points.

8. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The action space is defined as follows: The spray pressure adjustment step is 0.05 MPa, and the value range is 0.2 to 0.8 MPa; The walking speed adjustment step size is 0.1 m / s, with a range of 0.3 to 1.2 m / s; The nozzle angle can be adjusted in 5-degree increments, with a range of 0 to 90 degrees.

9. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The formula for the reward function is: ； in, As a reward value, Bonus for work accuracy Rewards for work efficiency Punishment for wasting medicine; The operational accuracy bonus is calculated based on the matching degree between agrochemical usage and demand gap. A matching degree of 90% or higher is 10 points, 80% to 90% is 8 points, 70% to 80% is 6 points, and less than 70% is 0 points. The work efficiency bonus is calculated based on the ratio of actual work time to theoretical optimal time; The penalty for drug waste is calculated based on the sum of the drift rate and the missed spray rate.

10. The multimodal fusion reinforcement learning control method for humanoid robot agricultural operations according to claim 1, characterized in that, The update formula for the action value function is: ； in, For the value of new actions, For the value of old moves, For learning rate, This is the current reward value. As a discount factor, The value of the action in the next state; The learning rate is set to 0.1, and the discount factor is set to 0.

9. The strategy is considered convergent when the change in the action value function over 100 consecutive iterations is less than or equal to 0.01.

Images