A deep neural network-based intelligent evaluation method for crop growth environment in the whole growth cycle

Abstract

The application belongs to the technical field of agricultural intelligent control, and relates to a crop whole-cycle growth environment intelligent evaluation method based on a deep neural network. The method comprises the following steps: dividing a growth stage, configuring an environmental factor threshold range and an evaluation model parameter; constructing a rule scoring model based on agricultural prior knowledge, and calculating a rule comprehensive score; outputting a neural network prediction health score by using a deep feedforward neural network model fused with prior weight constraint; dynamically adjusting a fusion weight, and obtaining a crop growth comprehensive health score by weighted fusion; matching a corresponding regulation strategy according to a current health grade; generating a multi-means echelon regulation sequence; calculating a coupling corrected comprehensive hazard degree, and determining a regulation priority and a multi-factor collaborative regulation execution sequence, and performing dynamic adjustment. The application effectively improves the poor explainability of the black box model and the overfitting problem, improves the regulation effect of a complex abnormal scene, has strong adaptability and expandability, and realizes whole-automatic and accurate management and control of the crop growth environment.

Description

Technical Field

[0001] This invention belongs to the field of agricultural intelligent control technology, specifically relating to an intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks. Background Technology

[0002] Facility agriculture is an important component of modern agriculture. Its core is to create optimal conditions for crop growth by artificially controlling environmental factors. Accurately assessing the suitability of the crop growth environment is a prerequisite for achieving precise control.

[0003] In existing technologies, there are complex nonlinear coupling relationships between environmental factors. For example, the interaction between temperature and light can significantly affect crop photosynthesis. Simple linear weighting cannot accurately fit this relationship, resulting in low evaluation accuracy. The sensitivity of factors varies greatly among different crops and at different growth stages. Hard-coded fixed thresholds and weights have poor adaptability and cannot meet the needs of accurate evaluation throughout the entire growth cycle. The threshold values ​​do not consider the climatic differences across regions and seasons, nor do they allow users to flexibly adjust control standards according to production goals. They fail to balance universality and personalization, making large-scale application in actual production difficult. Particularly near critical thresholds, they cannot promptly detect potential growth stresses. Furthermore, environmental control often employs single-factor independent control strategies, neglecting the coupled hazards of multi-factor anomalies and the side effects of control operations. This easily leads to secondary disasters where "solving one problem triggers another." In addition, existing environmental control systems often use single-point threshold start-stop logic. When environmental parameters are near critical values, this can cause frequent start-stop "oscillations" in the execution equipment, accelerating hardware aging and shortening equipment lifespan, while also causing repeated fluctuations in environmental parameters, which is detrimental to crop growth stability. Simultaneously, existing automatic control systems lack unified safe operation access boundaries and graded verification mechanisms, posing a safety risk of accidental operation exceeding crop physiological limits and inducing diseases in extreme scenarios. In recent years, deep learning technology has been widely used in the agricultural field. However, most existing crop growth assessment methods based on neural networks directly adopt end-to-end black box models, completely ignoring the prior knowledge accumulated by agricultural experts. This results in poor model interpretability and overfitting when there is insufficient data, making it difficult to promote and apply in actual production.

[0004] Therefore, there is an urgent need to develop an intelligent assessment method for crop growth environment that can fully utilize the prior knowledge of agricultural experts, leverage the nonlinear fitting capabilities of neural networks, and possess multi-factor synergistic regulation, multi-scenario adaptation, and target configurability, in order to improve assessment accuracy and practicality. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention is proposed. The present invention provides an intelligent assessment method for the entire crop growth environment based on deep neural networks, comprising: Based on crop type, the entire growth cycle of the target crop is divided into several continuous and standardized growth stages. Corresponding environmental factor threshold ranges and evaluation model parameters are configured for each growth stage. The switching of growth stages is automatically triggered based on the cumulative effective accumulated temperature, and the thresholds and model parameters of the corresponding stage are loaded simultaneously. For each crop type, a rule-based scoring model is constructed based on prior agricultural knowledge. The individual scores are calculated according to the current threshold range of each environmental factor, and the scores are weighted and summed according to the preset prior weights to obtain the comprehensive rule score. For each crop type, a deep feedforward neural network model with prior weight constraints is adopted. Real-time collected environmental factor data is input into the pre-trained deep feedforward neural network model, and the neural network outputs the predicted health score. During the model training process, prior weight regularization constraints are used to ensure that the contribution of each environmental factor conforms to the preset prior weight distribution. The fusion weights are dynamically adjusted based on the current growth stage of the crop and the real-time prediction accuracy of the deep feedforward neural network model. The rule-based comprehensive score and the neural network prediction health score are weighted and fused to obtain the comprehensive health score of crop growth, and the corresponding health level is determined. The system matches the corresponding control strategy based on the current health level; for scenarios where a single environmental factor is abnormal, it generates a multi-method tiered control sequence based on the basic contribution weight, deviation degree, and side effects of the control operation of the abnormal factor; for scenarios where two or more environmental factors are abnormal at the same time, it calculates the comprehensive hazard after coupling correction based on the basic contribution weight, deviation degree, and preset factor coupling relationship matrix of each abnormal factor, and determines the control priority and multi-factor collaborative control execution sequence; it drives the execution terminal to dynamically adjust the regional environmental factors until the target health level, environmental factor stop threshold, and execution safety constraints are met.

[0006] In some optional embodiments, the standardized growth stages include several continuous and non-overlapping sequences of growth stages divided according to the growth and development patterns of the target crop; different numbers and durations of growth stages are set according to the biological characteristics corresponding to the crop type.

[0007] In some optional embodiments, the cumulative effective accumulated temperature is the cumulative value of the daily effective temperature, and the effective temperature is the difference between the daily average air temperature and the crop biological zero degree. Temperatures below the biological zero degree are not included in the accumulated temperature. A corresponding accumulated temperature threshold is set in advance for each growth stage. When the cumulative effective accumulated temperature reaches the upper limit of the accumulated temperature threshold for the current growth stage, the growth stage is automatically triggered to switch, and the accumulated temperature threshold, weight, model and control strategy corresponding to each growth stage are updated synchronously.

[0008] In some optional embodiments, environmental factors include daytime temperature, nighttime temperature, light intensity, soil relative humidity, air relative humidity, and soil pH; daytime temperature is the ambient temperature from 8:00 to 20:00 each day; nighttime temperature is the ambient temperature from 20:00 to 8:00 the next day; and the threshold values ​​for each environmental factor at each growth stage are divided into three levels: optimal threshold range, suboptimal threshold range, and unsuitable threshold range.

[0009] In some alternative embodiments, the structure of a deep feedforward neural network model includes an input layer, a hidden layer, and an output layer; The input layer is used for inputting environmental factor data; The hidden layers include: a first hidden layer, used to perform nonlinear transformation on the input environmental factor data and extract single-factor features of each input environmental factor, whereby the single-factor features characterize the level deviation and suitability status of a single environmental factor; a second hidden layer, used to perform fully connected high-order combination of the single-factor features extracted by the first hidden layer to form interactive features reflecting the nonlinear interactions between environmental factors; and a third hidden layer, used to evaluate factor contribution weights and generate a contribution weight vector consistent with the number of input environmental factors based on the single-factor features and interactive features; and a prior weight L2 regularization term is added to the loss function to constrain the weight ratio output by the third hidden layer to be based on the preset prior weights. The output layer combines interaction features and contribution weight vectors to map the output neural network to predict health scores.

[0010] In some optional embodiments, the pre-training process of the deep feedforward neural network model includes: Collect crop growth sample data under different growth stages and environmental conditions, and label the corresponding health scores; The input environmental factor data is processed by MinMax standardization and mapped to the [0,1] interval to eliminate the influence of different units; The Adam adaptive learning rate optimizer is used to train the deep feedforward neural network model with mean squared error (MSE) as the loss function. The initial learning rate, batch size and number of training rounds are set. During training, a prior weight regularization term is introduced to constrain the feature importance weight distribution of each input environmental factor in the deep feedforward neural network model, so that the relative contribution of the environmental factors learned by the model is consistent with the prior weights given by agricultural experts. The regularization coefficient is set to a range of 0.01 to 0.1, and the relative deviation of the contribution of each environmental factor output by the model is controlled within the set range based on the expert's prior weights.

[0011] In some optional embodiments, the rule-based comprehensive score is weighted and fused with the neural network predicted health score, and the weighting coefficients are dynamically adjusted according to the current growth stage of the crop and the real-time prediction accuracy of the neural network model. set up To arrive at the final comprehensive health score, The overall score is determined by the rules. To predict health scores for neural networks, for t The dynamic weight coefficients of the time-based rule model have a value range of [0.2, 0.8]. Therefore, the weighted fusion formula is: ; The dynamic weight coefficient is calculated as follows: a corresponding basic weight is set for each growth stage in advance. The earlier the growth stage, the higher the basic weight of the rule model. The mean absolute error (MAE) of the neural network prediction scores over the past 7 days is calculated, and the accuracy adjustment coefficient is determined based on the MAE level. The final dynamic weight coefficient is the product of the basic weight and the accuracy adjustment coefficient, and the final dynamic weight coefficient is limited to the range of [0.2, 0.8] by the clamp function.

[0012] In some optional embodiments, the driving execution terminal dynamically adjusts regional environmental factors until the target health level, environmental factor stopping threshold, and execution safety constraints are met. The threshold is then adaptively adjusted based on the target score level. This includes: determining the target score level based on the comprehensive health score of crop growth selected by the user; determining the corresponding threshold leniency coefficient based on the target score level; automatically generating the optimal threshold range, suboptimal threshold range, and unsuitable threshold range corresponding to the target level based on pre-stored benchmark thresholds and adjustment coefficients of each environmental factor; and updating the scoring criteria and control triggering conditions of the rule scoring model based on the generated threshold ranges.

[0013] In some optional embodiments, driving the execution terminal to dynamically adjust regional environmental factors further includes: When two or more abnormal environmental factors are detected simultaneously, dynamic multi-factor combination regulation is executed, including: obtaining the basic contribution weights of each environmental factor output by the deep feedforward neural network model, combining the deviation degree of each abnormal environmental factor and the preset factor coupling relationship matrix, calculating the comprehensive hazard after coupling correction, determining the regulation priority of abnormal environmental factors, retrieving all candidate regulation operations from the safe operation whitelist, filtering regulation operations that aggravate other abnormal environmental factors or break physiological safety constraints based on the regulation operation side effect matrix, calculating the comprehensive net benefit of each candidate regulation operation, and generating a regulation execution sequence based on the comprehensive net benefit. During the execution process, environmental parameter changes are monitored in real time, and the regulation intensity is dynamically adjusted until the stop threshold, the comprehensive health score reaches the target level, and the safety constraint conditions are met. When an abnormal state is detected in a single environmental factor, a multi-pronged coordinated regulation of the single factor is implemented, including: selecting all regulatory operations that can improve the abnormal factor from the current batch's safe operation whitelist to form a candidate set; calculating the comprehensive response weight based on the improvement magnitude of the abnormal factor and the side effects of other normal factors, combined with the prior weights of each environmental factor; generating a tiered regulation sequence according to the comprehensive response weight from high to low, and initiating the regulation operations in sequence; verifying the recovery status of the abnormal factor and the fluctuation of other factors in real time during the regulation process, stopping new operations after the regulation target is reached, and dynamically adjusting the intensity of operations when secondary abnormal risks occur.

[0014] The beneficial effects of this invention are: (1) This invention proposes a dual-model fusion evaluation mechanism that combines prior rules with deep neural networks, which not only retains the professionalism and interpretability of agricultural expert knowledge, but also gives full play to the nonlinear fitting ability of neural networks, thereby improving the evaluation accuracy of crop growth comprehensive health score. (2) This invention constructs a deep neural network model that integrates prior weight constraints. By adding a regularization term during the training process, the model is guided to learn feature representations that conform to agricultural common sense, which effectively solves the problems of poor interpretability and overfitting of pure black box models. (3) The present invention adopts a two-dimensional dynamic weighted fusion algorithm, which dynamically adjusts the fusion weight according to the crop growth stage and the real-time accuracy of the model, and can maintain a stable and reliable evaluation effect in the early growth stage with insufficient data and in the scenario of sudden environmental changes. (4) The present invention supports users to independently select target rating levels, and the system automatically adapts to the corresponding threshold and control strategy, which can flexibly adapt to the planting needs of different production targets. (5) This invention constructs a multi-means coupled response mechanism for single-factor abnormal scenarios. Based on the factor importance and the quantitative calculation of the comprehensive response weight, a step-by-step start-up control strategy is adopted, which not only ensures the recovery efficiency of abnormal factors, but also avoids the secondary environmental problems caused by single control, and significantly improves the scientific nature and comprehensive growth benefits of single-factor control. (6) This invention proposes a dynamic multi-factor combination regulation mechanism, which generates the optimal regulation sequence based on coupling relationship analysis and side effect screening, avoids the secondary disasters of single regulation, and significantly improves the regulation effect under complex abnormal scenarios; (7) This invention can output standardized growth levels and can be directly connected to the greenhouse environment control system to achieve fully automatic and precise control of the crop growth environment; Attached Figure Description Figure 1 This is a flowchart illustrating the principle of an intelligent assessment method for the crop growth environment throughout its entire life cycle based on deep neural networks, as provided in Embodiment 1 of the present invention. Figure 2This is a schematic diagram of a three-tiered database architecture. Figure 3 It is a flowchart that drives the execution terminal to dynamically adjust regional environmental factors. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0016] Example 1 As an example, in order to solve the problems existing in the prior art, this embodiment provides an intelligent assessment method for the crop growth environment throughout its entire life cycle based on deep neural networks.

[0017] The implementation details of the method in this embodiment are described below. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0018] This embodiment presents a method for intelligent assessment of the crop's full-cycle growth environment based on deep neural networks, which can be applied to electronic devices with communication, computing, and data storage capabilities. (See attached image) Figure 1 As shown in the figure, this embodiment provides an intelligent assessment method for the crop growth environment throughout its entire life cycle based on deep neural networks, including steps 110-150.

[0019] Step 110: Divide the entire growth cycle of the target crop into several continuous standardized growth stages according to the crop type, and configure the corresponding environmental factor threshold range and evaluation model parameters for each growth stage; the switching of growth stages is automatically triggered based on the cumulative effective accumulated temperature, and the threshold and model parameters of the corresponding stage are loaded synchronously. The standardized growth stages consist of several continuous and non-overlapping sequences of growth stages divided according to the growth and development patterns of the target crop. Different crops can have different numbers and durations of growth stages set according to their biological characteristics. The system adopts an automatic switching mechanism driven by accumulated temperature: the total accumulated temperature of the crop throughout its entire growth period is predefined and divided into N segments (N≥2). The system accumulates the ambient accumulated temperature in real time, and when the accumulated temperature reaches the accumulated temperature threshold, it automatically switches to the parameter model of the next growth stage.

[0020] As a preferred embodiment, solanaceous crops can be divided into 5 growth stages, corresponding to the first growth stage (germination period), the second growth stage (seedling period), the third growth stage (flowering and fruit setting period), the fourth growth stage (fruit enlargement period), and the fifth growth stage (maturity and harvest period).

[0021] During system initialization, the operating thresholds for the current batch are generated based on a three-level hierarchical database. Specifically, the crop baseline thresholds are loaded from the core expert database, the regional correction coefficients and seasonal correction coefficients from the regional adaptation database are overlaid, and the historical fine-tuning values ​​from the batch self-learning database are combined to generate the optimal, suboptimal, and unsuitable threshold ranges for each environmental factor at the current stage.

[0022] As attached Figure 2 As shown, the system adopts a three-tiered database architecture, including a core expert database, a regional adaptation database, and a batch self-learning database. The core expert database stores nationally applicable basic crop data, including growth stage definitions, baseline thresholds, prior weights, general pre-trained neural networks, regulatory strategy baselines, physiological limit parameters, a general safe operation whitelist, and a baseline environmental factor coupling matrix; it is updated annually and requires review by an expert team. The regional adaptation database stores regional correction data at the provincial, municipal, and county levels, including seasonal correction coefficients for temperature, light, and humidity, regional common problem regulation strategies, a regional supplementary safe operation whitelist, and a regional correction environmental factor coupling matrix; it is updated quarterly, with correction increments not exceeding ±10% of the baseline value. The batch self-learning database stores personalized data for individual planting batches, including threshold fine-tuning values, weight fine-tuning values, neural network incremental parameters, regulation effect records, a batch fine-tuning safe operation whitelist, and a batch fine-tuning environmental factor coupling matrix; it is updated in real-time with growth stage updates, with fine-tuning increments not exceeding ±5% of the regionally corrected value. It adopts a three-level hierarchical database architecture and takes into account the system's universality and regional and batch personalization through a progressive mechanism of general benchmark, regional correction and batch fine-tuning. It also has continuous self-evolution capability, is suitable for a variety of crops and planting scenarios, and has good versatility and scalability.

[0023] During system operation, the final operating parameters are generated according to the formula "core baseline parameter × regional correction coefficient × seasonal correction coefficient + batch fine-tuning value". Data flows and aggregates from bottom to top, enabling the system to continuously self-evolve.

[0024] Control commands are transmitted using either wired or wireless data transmission. When transmitting control commands wirelessly, the execution terminal is equipped with a wireless transceiver to receive the control commands, and the data processor controls the valves or power supply to regulate the operating status via a drive device.

[0025] The cumulative effective accumulated temperature is the cumulative value of the daily effective temperature. The effective temperature is the difference between the daily average air temperature and the crop's biological zero degree. Temperatures below the biological zero degree are not included in the accumulated temperature. A corresponding accumulated temperature threshold is set in advance for each growth stage. When the cumulative effective accumulated temperature reaches the upper limit of the accumulated temperature threshold for the current growth stage, the growth stage is automatically switched, and the accumulated temperature threshold, weight, model and control strategy corresponding to each growth stage are updated synchronously.

[0026] Step 120: For each crop type, construct a rule-based scoring model based on prior agricultural knowledge, calculate the individual score according to the current threshold range of each environmental factor, and sum the scores according to the preset prior weights to obtain the comprehensive rule score.

[0027] Environmental factors include daytime temperature, nighttime temperature, light intensity, soil relative humidity, air relative humidity, and soil pH. Daytime temperature is the ambient temperature from 8:00 to 20:00 each day, and nighttime temperature is the ambient temperature from 20:00 to 8:00 the next day. When scoring environmental factors during the daytime of the day, the nighttime temperature data collected from 20:00 the previous day to 8:00 the current day is used.

[0028] The threshold ranges of each environmental factor at each growth stage are divided into optimal threshold range, suboptimal threshold range, and unsuitable threshold range. The rule-based scoring model adopts a three-level scoring mechanism for each environmental factor: 100 points for being in the optimal threshold range, 60 points for being in the suboptimal threshold range, and 0 points for being in the unsuitable threshold range.

[0029] set up To arrive at the final comprehensive health score, The overall score is determined by the rules. To predict health scores for neural networks, for t The dynamic weight coefficients of the time-based rule model have a value range of [0.2, 0.8]. Therefore, the weighted fusion formula is: The formula for calculating the rule-based composite score is as follows: ; in, For the first Prior weights of environmental factors, For the first Individual scores for each environmental factor.

[0030] The formula for calculating the rule comprehensive score is as follows: Rule comprehensive score = Daytime temperature score × 20% + Nighttime temperature score × 15% + Light intensity score × 25% + Soil moisture score × 20% + Air humidity score × 15% + pH score × 5%.

[0031] Step 130: For each crop type, a deep feedforward neural network model with prior weight constraints is adopted. The real-time collected environmental factor data is input into the pre-trained deep feedforward neural network model, and the neural network predicts the health score. During the model training process, the contribution of each environmental factor is made to conform to the preset prior weight distribution through prior weight regularization constraints.

[0032] In some alternative embodiments, the structure of a deep feedforward neural network model includes an input layer, a hidden layer, an activation layer, and an output layer; The input layer is used for inputting environmental factor data; The hidden layers include: a first hidden layer, used to perform nonlinear transformation on the input environmental factor data, extracting single-factor features of each input environmental factor to characterize the degree of deviation and suitability of each individual environmental factor; a second hidden layer, used to perform fully connected high-order combination of the single-factor features extracted by the first hidden layer to form interaction features reflecting the nonlinear interactions between environmental factors; and a third hidden layer, used to evaluate the factor contribution weights, generating a 6-dimensional weight feature vector consistent with the number of input environmental factors based on the single-factor features and interaction features. A prior weight L2 regularization term is added to the loss function to constrain the weight ratio output by this layer to be based on a preset prior weight, only fine-tuned by interaction features within a limited range, with a relative deviation not exceeding ±10%. The total loss function is: ; in, The main loss is the mean square error. This is the regularization coefficient (ranging from 0.01 to 0.1, with a preferred value of 0.05). The prior weights are L2 regularization terms.

[0033] The output layer combines interaction features and contribution weight vectors to map the output neural network to predict health scores. Specifically, the input layer has 6 neurons, corresponding to six environmental factors: daytime temperature, nighttime temperature, light intensity, soil relative humidity, air relative humidity, and soil pH; the hidden layers have 3 layers with 16, 10, and 6 neurons respectively, all using the ReLU activation function; the output layer has 1 neuron, outputting a health score in the range of 0-100.

[0034] This invention proposes a dual-model fusion evaluation mechanism that combines prior rules with deep neural networks. This mechanism retains the professionalism and interpretability of agricultural expert knowledge while fully leveraging the nonlinear fitting capability of neural networks, thereby improving the accuracy of crop growth health score evaluation. By constructing a deep neural network model that integrates prior weight constraints and adding regularization terms during training, the model is guided to learn feature representations that conform to agricultural common sense, effectively solving the problems of poor interpretability and overfitting of pure black-box models.

[0035] In some optional embodiments, the pre-training process of the deep feedforward neural network model includes: Collect crop growth sample data under different growth stages and environmental conditions, and label the corresponding health scores; The input environmental factor data is processed by MinMax standardization and mapped to the [0,1] interval to eliminate the influence of different units; The Adam adaptive learning rate optimizer is used to train the deep feedforward neural network model with mean squared error (MSE) as the loss function. The initial learning rate, batch size, and number of training epochs are set. For example, the initial learning rate is set to 0.001, the batch size is 32, and the number of training epochs is 100. During training, a priori weight regularization term is added to constrain the weight distribution of the deep feedforward neural network model to be consistent with the prior knowledge of agricultural experts.

[0036] Step 140: Based on the current growth stage of the crop and the real-time prediction accuracy of the deep feedforward neural network model, dynamically adjust the fusion weights, weight and fuse the rule-based comprehensive score with the neural network prediction health score to obtain the comprehensive health score of crop growth, and determine the corresponding health level.

[0037] The rule-based comprehensive score and the neural network-predicted health score are weighted and fused together, and the weight coefficients are dynamically adjusted according to the current growth stage of the crop and the real-time prediction accuracy of the neural network model. set up To arrive at the final comprehensive health score, The overall score is determined by the rules. To predict health scores for neural networks, for t The dynamic weight coefficients of the time-based rule model have a value range of [0.2, 0.8]. Therefore, the weighted fusion formula is: ; in, To arrive at the final comprehensive health score, The overall score is determined by the rules. For the neural network to predict health scores, α is the dynamic weight coefficient of the rule model at time t, with a value range of [0.2, 0.8].

[0038] Dynamic weights are determined by two dimensions: Growth stage dimension: Different base weights are set for different growth stages. The less data there is, the higher the weight of the rule model: First growth stage: ( =0.7) Second growth stage: ( =0.6) Third growth stage: ( =0.4) Fourth growth stage: ( =0.3) Fifth growth stage: ( =0.2).

[0039] Model accuracy dimension: The mean absolute error (MAE) of the neural network prediction is calculated. When the running period is less than 3 days, MAE is not calculated, and the stage base weights are used directly. When the running period is greater than or equal to 3 days but less than 7 days, the full calculation is performed based on the actual running days. When the running period is greater than or equal to 7 days, a fixed 7-day sliding window is used. The calculation formula is as follows: ; N is the total number of valid samples in the past 7 days; Predict the health score for the neural network corresponding to the i-th sample; Let be the baseline true value of the health score corresponding to the i-th sample; Determine the accuracy adjustment coefficient MAE < 3 (Excellent): ( =0.8), 3≤MAE<6 (Good): ( =1.0), 6≤MAE<10 (general): ( =1.2), MAE≥10 (poor): ( =1.5).

[0040] The final dynamic weight calculation formula is: ; The `clamp` function is used to limit the weights to the range [0.2, 0.8] to ensure system stability.

[0041] Step 150: Match the corresponding control strategy according to the current health level; for a single abnormal environmental factor scenario, generate a multi-method tiered control sequence based on the basic contribution weight, deviation degree, and side effect degree of the control operation of the abnormal factor; for a scenario where two or more environmental factors are abnormal at the same time, calculate the comprehensive hazard after coupling correction based on the basic contribution weight, deviation degree, and preset factor coupling relationship matrix of each abnormal factor, and determine the control priority and multi-factor collaborative control execution sequence; drive the execution terminal to dynamically adjust the regional environmental factors until the target health level, environmental factor stop threshold, and execution safety constraints are met.

[0042] Specifically, based on the comprehensive health score, the system determines the crop's overall health level according to the following criteria and outputs corresponding control strategies: 90-100 points: Excellent overall health level, excellent growth; maintain current environmental parameters and implement normal management strategies; 75-89 points: Good overall health level, good growth; make slight adjustments to factors with lower scores based on individual factor scores; 60-74 points: Moderate overall health level, mild stress; issue a system warning to remind growers to pay attention and adjust environmental parameters promptly; 0-59 points: Poor overall health level, severely unsuitable environment; immediately activate the emergency control program to automatically adjust greenhouse environmental parameters to a suitable range.

[0043] Single-factor graded control strategy (taking daytime temperature as an example): Temperature too high (second best, 60 points): Open the top window for ventilation, with a ventilation volume of 30%-50%; Temperature too high (uncomfortable, 0 points): Open the top window and side window for full ventilation, lower the 50% shade net, and turn on the emergency fan if necessary; Temperature too low (second best, 60 points): Close the ventilation window and open the thermal blanket; Temperature too low (uncomfortable, 0 points): Fully enclose the thermal insulation, and turn on the heating system and hot air circulation.

[0044] In some optional embodiments, the driving execution terminal dynamically adjusts regional environmental factors until the target health level, environmental factor stopping threshold, and execution safety constraints are met. The threshold is then adaptively adjusted based on the target score level. This includes: determining the target score level based on the comprehensive health score of crop growth selected by the user; determining the corresponding threshold leniency coefficient based on the target score level; automatically generating the optimal threshold range, suboptimal threshold range, and unsuitable threshold range corresponding to the target level based on pre-stored benchmark thresholds and adjustment coefficients of each environmental factor; and updating the scoring criteria and control triggering conditions of the rule scoring model based on the generated threshold ranges.

[0045] In some alternative embodiments, as shown in the appendix Figure 3 As shown, the driver execution terminal dynamically adjusts regional environmental factors, and also includes: When two or more abnormal environmental factors are detected simultaneously, dynamic multi-factor combination regulation is executed, including: obtaining the basic contribution weights of each environmental factor output by the deep feedforward neural network model, combining the deviation degree of each abnormal environmental factor and the preset factor coupling relationship matrix, calculating the comprehensive hazard after coupling correction, determining the regulation priority of abnormal environmental factors, retrieving all candidate regulation operations from the safe operation whitelist, filtering regulation operations that aggravate other abnormal environmental factors or break physiological safety constraints based on the regulation operation side effect matrix, calculating the comprehensive net benefit of each candidate regulation operation, and generating a regulation execution sequence based on the comprehensive net benefit. During the execution process, environmental parameter changes are monitored in real time, and the regulation intensity is dynamically adjusted until the stop threshold, the comprehensive health score reaches the target level, and the safety constraint conditions are met. When an abnormal state is detected in a single environmental factor, a multi-pronged coordinated regulation of the single factor is implemented, including: selecting all regulatory operations that can improve the abnormal factor from the current batch's safe operation whitelist to form a candidate set; calculating the comprehensive response weight based on the improvement magnitude of the abnormal factor and the side effects of other normal factors, combined with the prior weights of each environmental factor; generating a tiered regulation sequence according to the comprehensive response weight from high to low, and initiating the regulation operations in sequence; verifying the recovery status of the abnormal factor and the fluctuation of other factors in real time during the regulation process, stopping new operations after the regulation target is reached, and dynamically adjusting the intensity of operations when secondary abnormal risks occur.

[0046] The safe operation whitelist serves as a prerequisite safety access boundary for automatic system control. It refers to a set of control operations that, for specific crops and specific growth stages, have been verified and confirmed not to exceed the crop's physiological safety limits, cause irreversible growth damage, or induce common diseases. All automatic control commands must first pass the whitelist verification. Operations not included in the whitelist are prohibited from automatic execution by the system and can only be triggered after manual confirmation.

[0047] The safe operation whitelist is compiled based on crop physiological safety thresholds, national agricultural industry standards, expert prior knowledge, pest and disease risk data, and historical production data. It adopts a three-tiered architecture, managed and updated synchronously with the three-tiered database: The core expert database stores a general basic whitelist, using crop growth stages as the smallest unit. It clearly defines the safe operating range, intensity limits, maximum single operating time, and absolute prohibitions for each control device, serving as a nationwide safety baseline that must not be exceeded in any region or batch. The regional adaptation database stores a supplementary regional whitelist. Based on the core whitelist, it incorporates regional climate characteristics, soil types, and patterns of prevalent pests and diseases to localize operational intensity and prohibitions, with adjustments not exceeding ±10% of the core whitelist parameters. The batch self-learning database stores a batch fine-tuning whitelist. Based on the current batch's crop growth and historical control effect feedback, it makes personalized fine-tuning adjustments to the operational intensity within the whitelist, with adjustments not exceeding ±5% of the regional whitelist parameters. This invention constructs a three-tiered safety operation whitelist mechanism as a prerequisite safety access boundary for automatic regulation. It refines safety rules layer by layer from general baseline and regional adaptation to batch fine-tuning, thereby avoiding the risk of automatic regulation exceeding the physiological limits of crops and inducing diseases from the source. It strengthens the system's safety baseline while ensuring the flexibility of regulation.

[0048] The whitelist has a dynamic self-evolution mechanism: batch running data is aggregated upwards level by level, and the core whitelist is iteratively updated after verification to achieve continuous optimization of security boundaries.

[0049] This system uses basic contribution weight, deviation degree, and comprehensive hazard degree after coupling correction as core quantitative indicators for hierarchical regulation. The definitions and calculation rules for each indicator are unified: the basic contribution weight is taken from the contribution degree of each environmental factor output by the third hidden layer of the deep feedforward neural network model, representing the basic impact of the factor on crop growth and health; the deviation degree is quantified by the difference between the individual environmental factor score and the full score of 100; the larger the difference, the further the factor deviates from the optimal range, and the higher the degree of abnormality; the comprehensive hazard degree after coupling correction is the core judgment indicator for multi-factor abnormal scenarios, obtained by multiplying the sum of the individual hazard degrees of all abnormal factors by the maximum coupling correction coefficient, where: individual factor hazard degree = basic contribution weight × deviation degree. Regulation termination is set with three constraints: first, the target health level constraint, i.e., the comprehensive health score reaches the preset target level lower limit; second, the environmental factor stop threshold constraint, i.e., the abnormal factor parameter reaches the hysteresis stop threshold; and third, the execution safety constraint, i.e., the regulation operation triggers the physiological safety boundary or disease risk threshold, and meeting any one of these conditions triggers the corresponding stop or safety circuit breaker logic.

[0050] Single-factor, multi-method coupled response regulation (as attached) Figure 3(See the path on the left): When a single environmental factor score is detected to be less than or equal to 60, the system is not limited to a single control operation, but adopts a multi-means coordinated coupled control mechanism to quickly eliminate the main anomaly while avoiding secondary environmental problems and maximizing the overall benefits of crop growth.

[0051] The steps for generating the regulatory sequence are as follows: First, identify the main anomalous factor and its anomalous level, and clarify the regulatory target of restoring it to the optimal range. Then, select all regulatory operations from the safe operation whitelist that can directly improve the anomalous factor, forming a candidate operation set. Combining the anomalous factor's basic contribution weight, current deviation degree, and the side effects of the regulatory operation on other normal factors, calculate the comprehensive response weight of each operation and sort them from high to low to generate a multi-method tiered regulatory sequence. The higher the deviation degree, the higher the proportion of improvement benefit of the corresponding operation; the higher the side effects degree, the more the operation weight is deducted. ; In the formula, To measure the improvement magnitude coefficient of the main anomaly factor, The weights of the principal anomaly factors. For operation on the first The side effect amplitude coefficient of each normal factor For the first The prior weights of each environmental factor are considered; the higher the overall response weight, the higher the operational priority. Among these, the prior weights of the environmental factors are... All are dimensionless degree quantification coefficients, with a uniform value range of 0 to 1.

[0052] Control operations are initiated sequentially from high to low based on the comprehensive response weight. The control effect is verified once every preset period. Once the main abnormal factor recovers to the target range, new operations are stopped. If other factors show a trend of fluctuating towards the abnormal range, the current operation intensity is reduced. All environmental factors are monitored synchronously throughout the process. If a side effect causes other factors to enter the suboptimal range, the operation combination is adjusted immediately to avoid secondary abnormal escalation.

[0053] Dynamic multi-factor combination regulation (as attached) Figure 3 (See the path on the right): When two or more individual factor scores are detected to be less than or equal to 60, a combined control mechanism is triggered: The basic contribution weights of each environmental factor output by the neural network model are used as the core benchmark. The single-factor hazard is calculated by combining the deviation of each abnormal factor. Then, a coupling correction coefficient is introduced through the built-in coupling relationship matrix (Table 1 takes the fourth growth stage of tomatoes as an example). The coupling relationship is divided into four levels: strong synergy (correction coefficient 1.4), synergy (correction coefficient 1.2), weak synergy (correction coefficient 1.1), and unrelated (correction coefficient 1.0). The larger the coefficient, the stronger the hazard amplification effect when two factors are abnormal at the same time. The comprehensive hazard is calculated according to the following rules: Single-factor hazard: For each anomalous factor with a score less than or equal to 60, the single-factor hazard is calculated using the following formula: ; In the formula, For the first Individual scores for each environmental factor For the first Prior weights of environmental factors.

[0054] Coupling correction coefficient extraction: Extract the pairwise coupling coefficients of all outliers, and take the maximum value as the overall coupling correction coefficient. The maximum value rule is adopted as a conservative design to prioritize avoiding high-risk collaborative coercion scenarios.

[0055] The overall hazard degree of the multi-factor combination is calculated using the following formula: ; In the formula, This represents the total number of abnormal factors. The higher the overall hazard level, the higher the processing priority for the corresponding abnormal combination.

[0056] Then, the overall hazard level after coupling correction is calculated. The overall hazard level is used as the core basis to determine the control priority of abnormal factors, with higher hazard levels being processed earlier. After prioritization, all candidate control operations are retrieved from the safe operation whitelist and subjected to double filtering based on the control operation side effect matrix: operations that would exacerbate the severity of any abnormal factor are filtered out, as are operations that would trigger physiological safety constraints, retaining only safe and effective candidate operations. The overall net benefit of each candidate control operation is calculated, and a control execution sequence is generated based on the overall net benefit, executed sequentially. The control effect is monitored every 15 minutes, and the control intensity is dynamically adjusted.

[0057] Table 1. Coupling Relationship Matrix of Tomato Fourth Growth Stage

[0058] A three-level hysteresis-based anti-shake stop mechanism covers three dimensions: single factor, multiple operations, and comprehensive score, forming a complete execution feedback loop to avoid frequent equipment start-ups and shutdowns and parameter oscillations. This invention constructs a stop mechanism with hysteresis anti-shake at the control execution end, setting dual-threshold hysteresis for start-stop, combined with multi-operation reverse tiered shutdown and comprehensive score stability judgment rules. This effectively avoids the problem of frequent equipment start-ups and shutdowns near parameter critical points, extends the service life of the execution terminal hardware, ensures smooth transition of environmental parameters, avoids drastic parameter fluctuations causing growth stress to crops, and improves the long-term stability and economy of the system.

[0059] For the regulation of individual environmental factors, separate start and stop thresholds are set. The stop threshold is recessed by a preset hysteresis range within the optimal range, forming a buffer zone. The hysteresis parameters are stored in the core expert database and can be fine-tuned according to a three-level correction rule using the regional adaptation database and the batch self-learning database.

[0060] For scenarios involving multi-method coupled regulation and multi-factor synergistic regulation, a reverse tiered shutdown logic of "last start, first stop" is adopted, which corresponds completely to the tiered start sequence: When the main abnormal factor reaches the stop threshold, the last control device started is shut down first; after one monitoring cycle, if the parameters remain stable, the next level device is shut down; until all control devices are shut down, normal monitoring is resumed.

[0061] The current round of proactive regulation will terminate and transition to routine monitoring when the overall health score meets the following two conditions: the score reaches or exceeds the lower limit of the target level; and the score fluctuation range is less than or equal to 3 points for three consecutive monitoring cycles, indicating a stable state. This rule avoids erroneous termination of regulation due to single data fluctuations, ensuring the sustainability of the regulation effect.

[0062] The system allows users to select their target rating level, including three standard levels: Excellent, Good, and Average, as well as a custom scoring mode. After the user selects the target level, the system automatically calculates the threshold leniency coefficient, generates a corresponding threshold matrix according to the factor differentiation adjustment principle, and simultaneously updates the scoring criteria and control trigger conditions, achieving fully automatic matching of target level, threshold, and control.

[0063] This invention establishes a standardized crop growth cycle division system and multi-factor threshold standard, which is applicable to a variety of crops and has good versatility and scalability.

[0064] Taking the third day of greenhouse cucumber cultivation (first growth stage) as an example, the environmental factor data collected were: daytime temperature 27℃, nighttime temperature 19℃, light intensity 40000 Lux, soil moisture 70%, air humidity 85%, and pH 6.5.

[0065] Rule-based comprehensive score calculation: 100 points for daytime temperature within the optimal range, 100 points for nighttime temperature within the optimal range, 100 points for light intensity within the optimal range, 100 points for soil moisture within the optimal range, 60 points for air humidity within the second-best range, and 100 points for pH within the optimal range. The overall score for the rules is: 100×20%+100×15%+100×25%+100×20%+60×15%+100×5%=94 points.

[0066] Neural network prediction: The standardized environmental factor data was input into the deep feedforward neural network model corresponding to cucumber, and the neural network prediction score was 97.8.

[0067] Dynamic weight calculation: Basic weights for the first growth stage The model's MAE score was 2.8 (excellent, β = 0.8, final weights). ; Integrated score calculation: Overall health score = 94 × 0.56 + 97.8 × 0.44 ≈ 95.67 points.

[0068] Level determination: It belongs to the excellent level, and no environmental factor regulation is required.

[0069] Taking the 65th day of greenhouse tomato cultivation (fourth growth stage) as an example, the environmental data collected were: daytime temperature 32℃, nighttime temperature 16℃, light intensity 50000 Lux, soil moisture 75%, air humidity 65%, and pH 6.3.

[0070] The comprehensive score calculation for tomatoes according to the corresponding rules is as follows: Daytime temperature of 32℃ is within the second-best range (optimal 23-30℃), scoring 60 points; Nighttime temperature of 16℃ is within the best range, scoring 100 points; Light intensity of 50,000 Lux is within the second-best range (optimal 60,000-120,000 Lux), scoring 60 points; Soil moisture is within the best range, scoring 100 points; Air humidity is within the best range, scoring 100 points; pH value is within the best range, scoring 100 points. Therefore: The combined score is 60×20%+100×15%+60×25%+100×20%+100×15%+100×5%=82 points; The standardized environmental data was input into the deep feedforward neural network model corresponding to tomatoes, and the neural network prediction score was 79.2 points. A dynamic weighted fusion algorithm was used, and the basic weights in this stage... =0.3, the model's real-time MAE=4.8 (good level, β=1.0), and the final dynamic weight α(t)=0.3; Integrated score calculation: Overall health score = 82 × 0.3 + 79.2 × 0.7 = 80.04 points; Grade determination: The grade of 80.04 is good, but there are two suboptimal abnormal factors: higher daytime temperature and lower light intensity, which trigger a multi-factor synergistic regulation mechanism. The sum of the hazard levels of each single factor is: (100-60)×0.2+(100-60)×0.25=18; Coupling coefficient: The coupling coefficient between daytime temperature and illumination, obtained from the matrix lookup, is 1.2, i.e. =1.2; Overall Hazard Severity: Based on the factor contribution weights output by the neural network, the hazard severity of each individual factor is calculated by combining the factor deviation: the deviation of daytime temperature is 100-60=40, and the hazard severity of each individual factor is 0.2×40=8; the deviation of illumination is 100-60=40, and the hazard severity of each individual factor is 0.25×40=10; the sum of the hazard severity of each individual factor is 18. Looking up the factor coupling relationship matrix, the coupling correction coefficient for daytime temperature and illumination is 1.2, i.e., the maximum coupling correction coefficient Cmax=1.2. The overall hazard severity after coupling correction is 18×1.2=21.6.

[0071] The combined hazard of this combination is higher than the simple sum of the hazards of individual factors. Therefore, daytime temperature and light intensity were identified as the objects of synchronous regulation with equal priority, generating a multi-factor synergistic regulation execution sequence. Priority was given to activating full-volume ventilation through the top and side windows and the plant supplemental lighting system. Coupled hazard analysis revealed a synergistic negative coupling relationship between high daytime temperature and low light intensity, with a combined hazard level exceeding the sum of the hazards of the two individual factors, necessitating synchronous regulation. Based on the factor contribution weights output by the neural network, combined with the factor deviation and coupling correction coefficient, daytime temperature and light intensity were determined as the objects of synchronous regulation with equal priority. Cooling and supplemental lighting candidate operations were retrieved from the safe operation whitelist, and double filtering was performed based on the side effect matrix of the regulation operations: shading operations would further reduce light intensity and exacerbate insufficient light, so they were filtered out; continuous high-humidity spraying would exceed the safety constraints for disease induction, so its use was restricted. Finally, ventilation and LED supplemental lighting were retained as effective candidate operations. Ventilation only slightly reduces air humidity, and supplemental lighting only slightly increases ambient temperature; both side effects are within a controllable range, making this the highest priority regulation combination.

[0072] After one monitoring cycle, if the daytime temperature remains above 30℃ and has not reached the hysteresis stop threshold, intermittent spraying will be initiated to further enhance the cooling effect. If the relative humidity exceeds 80% during spraying, posing a risk of disease induction due to high humidity, spraying will be stopped immediately, and the current ventilation and supplemental lighting will be maintained to prevent the aggravation of secondary anomalies.

[0073] Verify environmental parameters and control effectiveness after a monitoring cycle: When the daytime temperature drops below 28℃ and the light intensity rises above 60,000 Lux, maintain the current operating status; when the daytime temperature stabilizes below 28℃ (stop threshold, 2℃ lower than the start threshold), the light intensity stabilizes above 65,000 Lux, and the comprehensive health score remains above 90 for three consecutive monitoring cycles, gradually shut down the equipment in the reverse order of "supplementary lighting → ventilation" to avoid a drastic rebound in environmental parameters.

[0074] This invention can output standardized growth levels and can directly interface with greenhouse environmental control systems to achieve fully automatic and precise control of the crop growth environment.

[0075] In some optional embodiments, crop batch information includes batch code, crop type, region number, planting date, and expected yield.

[0076] In some alternative embodiments, the execution terminals include irrigation valves, fertigation machines, ventilators, shade net roll-up machines, and plant supplemental lighting units located in various areas.

[0077] When it is necessary to regulate environmental factor data, taking soil moisture control as an example, the control methods are as follows: (1) If the soil moisture score is 60, the irrigation valve is opened and the irrigation amount is the standard irrigation amount for the current time period multiplied by 0.6; (2) If the soil moisture score is 0, the irrigation valve is opened and the irrigation amount is the standard irrigation amount for this cycle multiplied by 1.0. (3) If the soil moisture score is 100, then keep the irrigation valve closed.

[0078] The comprehensive health score and level of crop growth corresponding to the crop type are updated in real time until the comprehensive health score of crop growth reaches the optimal comprehensive health score level or the target comprehensive health score level.

[0079] In some optional embodiments, the method further includes displaying real-time environmental factor data for each region, real-time comprehensive crop growth health scores, real-time operating status of each execution terminal, and crop batch information.

[0080] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent assessment of the crop's full-cycle growth environment based on deep neural networks, characterized in that, include: Based on crop type, the entire growth cycle of the target crop is divided into several continuous and standardized growth stages. Corresponding environmental factor threshold ranges and evaluation model parameters are configured for each growth stage. The switching of growth stages is automatically triggered based on the cumulative effective accumulated temperature, and the thresholds and model parameters of the corresponding stage are loaded simultaneously. For each crop type, a rule-based scoring model is constructed based on prior agricultural knowledge. The individual scores are calculated according to the current threshold range of each environmental factor, and the scores are weighted and summed according to the preset prior weights to obtain the comprehensive rule score. For each crop type, a deep feedforward neural network model with prior weight constraints is adopted. Real-time collected environmental factor data is input into the pre-trained deep feedforward neural network model, and the neural network outputs the predicted health score. During the model training process, prior weight regularization constraints are used to ensure that the contribution of each environmental factor conforms to the preset prior weight distribution. The fusion weights are dynamically adjusted based on the current growth stage of the crop and the real-time prediction accuracy of the deep feedforward neural network model. The rule-based comprehensive score and the neural network prediction health score are weighted and fused to obtain the comprehensive health score of crop growth, and the corresponding health level is determined. Match the corresponding control strategy according to the current health level; For a single abnormal environmental factor scenario, a multi-method tiered regulation sequence is generated based on the basic contribution weight of the abnormal factor, the degree of deviation, and the degree of side effects of the regulation operation. For scenarios where two or more environmental factors are abnormal simultaneously, the comprehensive hazard after coupling correction is calculated based on the basic contribution weight, deviation degree, and preset factor coupling relationship matrix of each abnormal factor, and the control priority and multi-factor collaborative control execution sequence are determined; the execution terminal is driven to dynamically adjust the regional environmental factors until the target health level, environmental factor stop threshold and execution safety constraints are met.

2. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, Standardized growth stages include several continuous and non-overlapping growth stage sequences divided according to the growth and development patterns of the target crop; different numbers and durations of growth stages are set according to the biological characteristics corresponding to the crop type.

3. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, The cumulative effective accumulated temperature is the cumulative value of the daily effective temperature. The effective temperature is the difference between the daily average air temperature and the crop's biological zero degree. Temperatures below the biological zero degree are not included in the accumulated temperature. A corresponding accumulated temperature threshold is set in advance for each growth stage. When the cumulative effective accumulated temperature reaches the upper limit of the accumulated temperature threshold for the current growth stage, the growth stage is automatically switched, and the accumulated temperature threshold, weight, model and control strategy corresponding to each growth stage are updated synchronously.

4. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, Environmental factors include daytime temperature, nighttime temperature, light intensity, soil relative humidity, air relative humidity, and soil pH. Daytime temperature is the ambient temperature from 8:00 to 20:00 each day; nighttime temperature is the ambient temperature from 20:00 to 8:00 the next day. The threshold values ​​of each environmental factor at each growth stage are divided into three levels: optimal threshold range, suboptimal threshold range, and unsuitable threshold range.

5. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, The structure of a deep feedforward neural network model includes an input layer, hidden layers, and an output layer; The input layer is used for inputting environmental factor data; The hidden layers include: a first hidden layer, used to perform nonlinear transformation on the input environmental factor data and extract single-factor features of each input environmental factor, whereby the single-factor features characterize the level deviation and suitability status of a single environmental factor; a second hidden layer, used to perform fully connected high-order combination of the single-factor features extracted by the first hidden layer to form interactive features reflecting the nonlinear interactions between environmental factors; and a third hidden layer, used to evaluate factor contribution weights and generate a contribution weight vector consistent with the number of input environmental factors based on the single-factor features and interactive features; and a prior weight L2 regularization term is added to the loss function to constrain the weight ratio output by the third hidden layer to be based on the preset prior weights. The output layer combines interaction features and contribution weight vectors to map the output neural network to predict health scores.

6. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, The pre-training process of a deep feedforward neural network model includes: Collect crop growth sample data under different growth stages and environmental conditions, and label the corresponding health scores; The input environmental factor data is processed by MinMax standardization and mapped to the [0,1] interval to eliminate the influence of different units; The Adam adaptive learning rate optimizer is used to train the deep feedforward neural network model with mean squared error (MSE) as the loss function. The initial learning rate, batch size and number of training rounds are set. During training, a prior weight regularization term is introduced to constrain the feature importance weight distribution of each input environmental factor in the deep feedforward neural network model, so that the relative contribution of the environmental factors learned by the model is consistent with the prior weights given by agricultural experts. The regularization coefficient is set to a range of 0.01 to 0.1, and the relative deviation of the contribution of each environmental factor output by the model is controlled within the set range based on the expert's prior weights.

7. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, The rule-based comprehensive score and the neural network-predicted health score are weighted and fused together, and the weight coefficients are dynamically adjusted according to the current growth stage of the crop and the real-time prediction accuracy of the neural network model. set up To arrive at the final comprehensive health score, The overall score is determined by the rules. To predict health scores for neural networks, for t The dynamic weight coefficients of the time-based rule model have a value range of [0.2, 0.8]. Therefore, the weighted fusion formula is: ； The dynamic weight coefficient is calculated as follows: a corresponding basic weight is set for each growth stage in advance. The earlier the growth stage, the higher the basic weight of the rule model. The mean absolute error (MAE) of the neural network prediction scores over the past 7 days is calculated, and the accuracy adjustment coefficient is determined based on the MAE level. The final dynamic weight coefficient is the product of the basic weight and the accuracy adjustment coefficient, and the final dynamic weight coefficient is limited to the range of [0.2, 0.8] by the clamp function.

8. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, The driver execution terminal dynamically adjusts regional environmental factors until the target health level, environmental factor stopping threshold, and execution safety constraints are met. At this point, the threshold is adaptively adjusted based on the target score level. This includes: determining the target score level based on the comprehensive health score of crop growth selected by the user; determining the corresponding threshold leniency coefficient based on the target score level; automatically generating the optimal threshold range, suboptimal threshold range, and unsuitable threshold range corresponding to the target level based on the pre-stored benchmark threshold and adjustment coefficients of each environmental factor; and updating the scoring criteria and control triggering conditions of the rule scoring model based on the generated threshold range.

9. The intelligent assessment method for the crop's full-cycle growth environment based on deep neural networks according to claim 1, characterized in that, The driver for the execution terminal to dynamically adjust regional environmental factors also includes: When two or more abnormal environmental factors are detected simultaneously, dynamic multi-factor combination regulation is executed, including: obtaining the basic contribution weights of each environmental factor output by the deep feedforward neural network model, combining the deviation degree of each abnormal environmental factor and the preset factor coupling relationship matrix, calculating the comprehensive hazard after coupling correction, determining the regulation priority of abnormal environmental factors, retrieving all candidate regulation operations from the safe operation whitelist, filtering regulation operations that aggravate other abnormal environmental factors or break physiological safety constraints based on the regulation operation side effect matrix, calculating the comprehensive net benefit of each candidate regulation operation, and generating a regulation execution sequence based on the comprehensive net benefit. During the execution process, environmental parameter changes are monitored in real time, and the regulation intensity is dynamically adjusted until the stop threshold, the comprehensive health score reaches the target level, and the safety constraint conditions are met. When an abnormal state is detected in a single environmental factor, a multi-pronged coordinated regulation of the single factor is implemented, including: selecting all regulatory operations that can improve the abnormal factor from the current batch's safe operation whitelist to form a candidate set; calculating the comprehensive response weight based on the improvement magnitude of the abnormal factor and the side effects of other normal factors, combined with the prior weights of each environmental factor; generating a tiered regulation sequence according to the comprehensive response weight from high to low, and initiating the regulation operations in sequence; verifying the recovery status of the abnormal factor and the fluctuation of other factors in real time during the regulation process, stopping new operations after the regulation target is reached, and dynamically adjusting the intensity of operations when secondary abnormal risks occur.

Images