Method and system for synergistically regulating multiple factors of respiration intensity of fruits and vegetables

Abstract

The application discloses a multi-factor synergistic regulation method and system for fruit and vegetable respiration intensity, and belongs to the technical field of automatic control, which comprises the following steps: collecting multi-modal data and performing pretreatment, performing feature analysis, generating a fusion feature vector, and identifying a respiration state; converting the fusion feature vector into a pulse sequence, setting a fuzzy decision method, constructing a three-layer pulse neural network, generating a hidden pulse sequence in combination with a dynamic threshold mechanism, constructing a causal regulation pulse through a fuzzy cognitive map, generating a fusion regulation pulse in combination with an activation variance, and thermodynamically optimizing the fusion regulation pulse in combination with an entropy generation rate change rate; constructing a hypergraph model, setting a synergistic regulation method, optimizing a regulation strategy, and using PCO to generate a synergistic pulse sequence with the frequency of a main regulation factor as a benchmark and according to a phase difference determined by causal intensity; collecting multi-dimensional feedback data and performing pretreatment, setting a constraint learning method, generating candidate optimization parameters and updating them in real time, and realizing autonomous optimization of the regulation strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to a method and system for multi-factor synergistic regulation of fruit and vegetable respiration intensity, belonging to the field of automatic control technology. BACKGROUND

[0002] In the field of fruit and vegetable storage and logistics preservation, dynamic regulation of respiration intensity is the core technology to maintain the freshness of fruits. As the basis of postharvest physiological activity, the intensity of respiration metabolism directly affects the maturation and aging process of fruits. Precise regulation of respiration intensity requires simultaneous matching of the dynamic changes of fruit and vegetable physiological characteristics and environmental factors.

[0003] However, the existing technology does not consider the dynamic differences of fruit and vegetable material characteristics, especially the nonlinear coupling effect of the respiratory jump peak and ethylene release of jump-type fruits, as well as the synergistic interaction of temperature, humidity, gas concentration and other environmental factors. Such dynamic differences make the respiration intensity show nonlinear and non-steady-state fluctuation characteristics, making it difficult for the prediction model of the adaptive regulation system to capture the true law of physiological changes, and thus unable to achieve precise and effective adjustment of parameters. On the one hand, the exponential growth of respiration rate caused by ethylene burst has randomness, and the traditional system relies on fixed thresholds, resulting in frequent misjudgment of the regulation timing, causing the actuator to fall into a vicious cycle of "ineffective adjustment-response lag". On the other hand, when multiple factors interfere synergistically and break the linear law of respiration response, the sensor data and actual metabolic demand are disconnected, making the key nodes that really need dynamic adjustment be ignored, such as the early stage of the jump peak. Not only does this exacerbate the volatility of system energy consumption, but also accelerates the wear and tear of hardware devices due to frequent start-stop. Thus, the actuator action cannot be converted into positive effects of energy consumption optimization and life extension, but rather reduces the matching degree of regulation precision and fruit and vegetable metabolic demand, making it difficult to build a dynamic coupling precise control logic. SUMMARY

[0004] In view of the deficiencies of the existing technology, the purpose of the present application is to provide a method and system for multi-factor synergistic regulation of fruit and vegetable respiration intensity, which synchronously collects data through three-dimensional sensing and pre-processes, uses pulse neural networks and fuzzy cognitive maps to process time series and causal reasoning, dynamically topological multi-agent synergistic regulation, and combines with thermodynamic constraint evolutionary learning.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] The method for multi-factor synergistic regulation of fruit and vegetable respiration intensity comprises:

[0007] Collecting multi-modal data and pre-processing, performing feature analysis, generating a fusion feature vector, and identifying the respiration state;

[0008] convert the fusion feature vector into a pulse sequence, set a fuzzy decision method, construct a three-layer pulse neural network, generate a hidden pulse sequence combined with a dynamic threshold mechanism, construct a causal regulation pulse through a fuzzy cognitive map, dynamically fuse the hidden pulse sequence and the causal regulation pulse combined with the node activation degree variance to generate a fusion regulation pulse, and perform thermodynamic optimization on the fusion regulation pulse combined with the entropy generation rate change rate;

[0009] construct a hypergraph model, set a synergistic regulation method, optimize the regulation strategy, and generate a synergistic pulse sequence using PCO as the benchmark of the main regulation factor frequency and determining the phase difference according to the causal strength;

[0010] collect multi-dimensional feedback data and preprocess, set a constraint learning method, generate candidate optimization parameters and update in real time.

[0011] Specifically, the fuzzy decision method comprises:

[0012] rate encoding is performed on the fusion feature vector, and the feature amplitude is converted into a pulse frequency;

[0013] The key features are screened out, the key features are linearly transformed through the encoding coefficient, and the tangent hyperbolic value is obtained to obtain a first pulse time;

[0014] The pulse frequency and the first pulse time are converted into a binary pulse sequence to generate a multi-dimensional pulse sequence;

[0015] A three-layer pulse neural network is constructed, wherein a dynamic threshold mechanism is used for the hidden layer neurons to generate a dynamic threshold;

[0016] The weight matrix is initialized, and the weight matrix is updated through the STDP rule;

[0017] The trigger condition of the pulse is set, the membrane potential is calculated based on exponential decay, and the pulse is fired once the membrane potential exceeds the dynamic threshold, otherwise not fired, to generate a hidden pulse sequence.

[0018] Specifically, the fuzzy decision method further comprises:

[0019] The fusion feature vector is defined as a fuzzy set to generate a node activation degree vector;

[0020] The joint distribution of each node in the historical data is statistically analyzed, the mutual information between nodes is calculated, the sensitivity of each feature in the fusion feature vector to the entropy generation rate is calculated based on a thermodynamic model, the mutual information and the sensitivity are combined according to the normalized weight, the causal weight between nodes is calculated and updated to obtain a causal strength matrix;

[0021] A three-level causal hierarchy is set, node activation is calculated in multi-scale fusion, and is mapped to a regulatory pulse parameter to generate a causal regulatory pulse in combination with a regulatory priority;

[0022] The variance of the node activation is calculated to determine the activation weight, the hidden pulse sequence is fused with the causal regulatory pulse, and the final regulatory pulse is obtained through thermodynamic optimization based on the change rate of entropy production rate.

[0023] Specifically, the specific steps of feature analysis include:

[0024] Collect multi-modal data, unify timestamps and denoise, perform three-layer preprocessing, and generate standard time series;

[0025] Perform Hilbert-Huang transform on the standard time series to obtain a time-frequency feature set;

[0026] Establish a gas concentration differential equation and use Kalman filter algorithm to calculate the respiration rate;

[0027] Calculate the entropy production rate, consider the ethylene diffusion entropy flow and metabolic entropy flow, and generate a thermodynamic feature set;

[0028] Calculate the mutual information matrix and perform cross-domain correlation analysis on the time-frequency feature set and the thermodynamic feature set through canonical correlation analysis;

[0029] Use principal component analysis to reduce the dimensionality of the time-frequency feature set and the thermodynamic feature set to generate a fusion feature vector.

[0030] Specifically, the specific steps of feature analysis further include:

[0031] Input the fusion feature vector into a Bayesian classifier, model the conditional probability through kernel density estimation, and dynamically adjust the prior probability to identify the respiratory state;

[0032] Combine the thermodynamic-time-frequency coupling factor, time-frequency energy entropy change rate, and entropy production rate change rate to calculate the warning parameter;

[0033] Set a basic warning threshold, dynamically adjust the basic warning threshold based on the integral of the entropy production rate and ethylene concentration, and trigger the jump peak warning when the real-time calculated warning parameter exceeds the warning threshold;

[0034] Organize the respiratory state recognition results, jump peak warning signals, and key features to generate a real-time analysis report, regularly calibrate the sensor, and update the Onsager coefficient and Bayesian classifier parameters.

[0035] Specifically, the specific steps of the dynamic threshold mechanism include:

[0036] Set the initial threshold value with reference to the resting potential of biological neurons;

[0037] The connection weight of the previous time step input layer to the hidden layer is multiplied by the corresponding pulse signal, and the historical threshold is exponentially attenuated by combining the forgetting factor to generate the basic threshold;

[0038] The ratio of the current entropy generation rate to the historical maximum entropy generation rate and the entropy generation rate change rate are calculated, and a thermodynamic conditioning factor is obtained by weighting and summing according to a preset weight coefficient;

[0039] The instantaneous frequency variance is calculated and compared with the reference frequency, and a time-frequency conditioning factor is obtained by combining and weighting the time-frequency energy entropy;

[0040] The thermodynamic-time-frequency coupling factor is compared with the preset coupling threshold, combined with the fusion slope coefficient, converted into a conditioning fusion coefficient by a Sigmoid function, and smoothed by a first-order low-pass filter;

[0041] The smoothed conditioning fusion coefficient is used as the weight of the thermodynamic conditioning factor, and the remaining weight is allocated to the time-frequency conditioning factor. The influence of the thermodynamic conditioning factor and the time-frequency conditioning factor on the basic threshold is calculated and summed with the basic threshold to obtain a conditioning threshold;

[0042] The identification result of the respiratory state is obtained in real time. If it is in the jump period, the deviation of the current entropy generation rate from the reference entropy generation value is calculated, and the conditioning threshold is fine-tuned online according to the deviation ratio to obtain a final dynamic threshold, otherwise the conditioning threshold is used as the final dynamic threshold.

[0043] Specifically, the specific steps of thermodynamic optimization of the fusion control pulse include:

[0044] A basic coefficient is set, and the entropy generation rate change rate at the current time and at a preset time point in the past is calculated;

[0045] The entropy generation rate change rate is compared with the reference entropy generation rate to obtain a relative ratio. The metabolic adjustment coefficient is generated by performing a hyperbolic tangent function conversion on the relative ratio, and the metabolic optimization coefficient is obtained by combining the basic coefficient;

[0046] Based on the metabolic optimization coefficient, the optimized pulse width and frequency are calculated, and the theoretical entropy generation rate change value is calculated by combining the control influence coefficient;

[0047] If the theoretical entropy generation rate change value is less than zero, the metabolic optimization coefficient is reset to zero, and the pulse width and frequency are recalculated, otherwise the metabolic optimization coefficient is applied.

[0048] Specifically, the synergistic control method includes:

[0049] A hypergraph model is constructed, an agent node is set, an interaction relationship between factors is represented by a hyperedge, and a hyperedge weight is calculated;

[0050] When receiving a new pulse sequence, recalculate the factor pulse time synchronization, and update the super-edge weight in a weighted update manner;

[0051] Define the agent state and design the regulation reward function;

[0052] Each agent uses a distributed reinforcement learning algorithm, takes the initial regulation strategy as the initial regulation strategy, and updates the initial regulation strategy according to the current state and reward;

[0053] Set the pulse coupled oscillator, define the main regulation factor, and generate the collaborative pulse sequence;

[0054] Construct a global optimization target, periodically synchronize the super-edge weight and regulation strategy parameters;

[0055] Solve the global optimization problem using a consensus algorithm and an alternating direction multiplier method, perform thermodynamic feasibility checking on the optimized regulation strategy, adjust the illegal regulation strategy, and obtain the optimized regulation strategy.

[0056] Specifically, the constraint learning method comprises:

[0057] Collect multi-dimensional feedback data and align them with regulation actions according to timestamps, and calculate the mean and variance through a sliding window;

[0058] Design a feedback reward function based on thermodynamic feedback characteristics, which includes entropy change convergence reward, energy efficiency reward, thermodynamic constraint term and state recognition reward;

[0059] Encode the parameters in the fuzzy decision method and the collaborative regulation method into chromosomes, simulate plant vascular system design evolution operators to generate candidate optimization parameters;

[0060] Thermodynamic test on the candidate optimization parameters, including correcting the Onsager coefficient to meet the symmetry, adjusting the thermodynamic force parameter to ensure the non-negativity of entropy production rate, and verifying the effectiveness of the candidate optimization parameters;

[0061] Update the candidate optimization parameters through a thermodynamic adaptive step size, which is related to the entropy production rate and the reward gradient;

[0062] Every time a regulation cycle is completed, Trigger evolution iteration and apply the optimal parameters.

[0063] The multi-factor collaborative regulation system of fruit and vegetable respiration intensity comprises a feature analysis module, a pulse decision module, a regulation module and a constraint learning module.

[0064] The feature analysis module is used for collecting multi-modal data through three-dimensional grid deployment of sensors, and after time stamp synchronization, hardware filtering and preprocessing, time-frequency features are generated through Hilbert-Huang transformation, thermodynamic features are generated by combining gas concentration differential equations and Kalman filtering, and after fusion, the respiratory state is identified by inputting the Bayesian classifier;

[0065] The pulse decision module is used for converting the fusion feature vector into a pulse sequence through rate coding and time intermediate Encoding, constructing a pulse neural network and a fuzzy cognitive map, generating a hidden pulse sequence and a causal regulation pulse, and outputting a fusion regulation pulse after thermodynamic optimization;

[0066] The regulation module is used for constructing a hypergraph model, updating hyperedge weights according to causal strength and pulse time synchronization, generating a collaborative pulse sequence through a pulse coupled oscillator based on a reinforcement learning algorithm of each intelligent agent, and outputting a regulation strategy after optimization by a consistency algorithm.

[0067] The constraint learning module is used for collecting feedback data, designing a feedback reward function, encoding parameters into chromosomes, generating candidate optimization parameters through biological heuristic operators, and iteratively updating the optimization parameters after thermodynamic verification.

[0068] The beneficial effects of the present application are:

[0069] Through three-dimensional grid sensors and multi-layer data preprocessing, the comprehensiveness and accuracy of the collected data are ensured, high-quality input is provided for feature analysis, Hilbert-Huang transformation and thermodynamic calculation are combined to realize cross-domain correlation of time-frequency features and metabolic intensity, the respiratory state recognition accuracy is improved, and the foundation for jump peak early warning is laid, biological neuron mechanism and fuzzy cognitive map are combined to generate regulation pulses with time sequence sensitivity and causal reasoning ability, a dynamic threshold mechanism is used to adaptively adjust the regulation time through thermodynamic and time-frequency adjustment factors, the regulation time accuracy is improved, a hypergraph model and reinforcement learning are used to realize dynamic representation and collaborative optimization of multi-factor interaction, a pulse coupled oscillator is used to ensure the synchronization of the regulation pulse, reduce factor interference, an embedded reward function and biological heuristic operators are used to promote self-optimization of system parameters, and finally the comprehensive effects of improving jump peak early warning accuracy, reducing regulation energy consumption and enhancing long-term adaptability of the system are realized. BRIEF DESCRIPTION OF DRAWINGS

[0070] Fig. 1 It is a multi-factor collaborative regulation method for fruit and vegetable respiration intensity;

[0071] Fig. 2 It is a flowchart of feature analysis of the present application;

[0072] Fig. 3 It is a flowchart of the fuzzy decision method of the present application;

[0073] Fig. 4 Flow chart for dynamic threshold mechanism of the present application;

[0074] Fig. 5 Flow chart for thermodynamic optimization of the present application;

[0075] Fig. 6 Structural diagram of the multi-factor synergistic regulation system for fruit and vegetable respiration intensity. DETAILED DESCRIPTION

[0076] The technical solutions of the present application will be described in detail below with reference to the drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solutions of the present application, rather than limitations of the technical solutions of the present application. In the case of no conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0077] Embodiment 1:

[0078] Reference Figs. 1 to 5 As shown in the figure, the present embodiment introduces a multi-factor synergistic regulation method for fruit and vegetable respiration intensity, including the following steps:

[0079] A three-dimensional grid sensor is deployed in the fruit and vegetable storage environment, and multi-modal data is synchronously collected at a sampling interval. The multi-source time series is generated by unifying the time stamp through a GPS clock and hardware low-pass filtering. The standard time series is generated through three-layer preprocessing, and the standard time series is analyzed for features. The time-frequency features are extracted through Hilbert-Huang transform. The respiration rate is estimated by combining the gas concentration differential equation and Kalman filtering. The thermodynamic characteristics are calculated. After cross-domain correlation analysis and fusion dimensionality reduction of the two types of features, the fusion feature vector is generated. The respiration state is identified by inputting the fusion feature vector into the Bayesian classifier. The warning parameters are calculated by comprehensively considering multiple factors. The threshold is dynamically adjusted to realize the warning of the jump peak.

[0080] The fusion feature vector is converted into a pulse sequence through rate coding and time coding to solve the time correlation problem of traditional digital signals. A fuzzy decision method is set. A three-layer pulse neural network is constructed. The weight matrix is updated using the STDP rule. The hidden pulse sequence is generated by combining the dynamic threshold mechanism. At the same time, a fuzzy cognitive map is constructed. The node activation and causal weight are calculated. The multi-scale fusion activation is obtained through three-level causal layering. The causal regulation pulse is generated. The weight is calculated according to the variance of the node activation. The hidden pulse sequence and the causal regulation pulse are fused to generate a fusion regulation pulse. The fusion regulation pulse is thermodynamically optimized by combining the entropy generation rate change rate to ensure compliance with the laws of thermodynamics. The dynamic threshold mechanism is self-adaptively adjusted through the calibration of thermodynamics, time-frequency regulation factors, and respiration states. The thermodynamic optimization adjusts the pulse parameters by calculating the metabolic optimization coefficient and verifies the feasibility.

[0081] An ultra-graph model is constructed with multi-factors as agent nodes, a collaborative regulation method is set, the ultra-edge weight is dynamically updated based on the causal strength and pulse time synchronicity, the agent optimizes the regulation strategy based on the local state through the reward function and PPO algorithm, and the PCO is used to generate a collaborative pulse sequence based on the main regulation factor frequency and the phase difference determined by the causal strength, to construct a global optimization target, to realize the global optimization of topology and strategy through the consistency algorithm and the thermodynamic test, to solve the factor interaction representation and collaborative regulation problem;

[0082] Multi-dimensional feedback data is collected and preprocessed, a constraint learning method is set, features are extracted and a feedback reward function is set, the laws of thermodynamics are embedded, parameters are encoded as chromosomes, evolution operators are designed to generate candidate optimization parameters, thermodynamic tests and multi-scale simulations are performed on the candidate optimization parameters, the candidate optimization parameters are updated by adaptive step size, and the optimal parameters obtained by evolution are applied to realize autonomous optimization of the regulation strategy after completing N regulation cycles.

[0083] Specifically, the specific steps of feature analysis include:

[0084] A variety of sensors are deployed in the fruit and vegetable storage environment, the sensors are distributed in a three-dimensional grid to ensure comprehensive data collection, the sampling interval is set to Multi-modal data is collected synchronously according to the sampling interval, the multi-modal data is time-stamped uniformly by a GPS clock to ensure consistency in time, high-frequency noise interference is removed by a hardware low-pass filter, and multi-source time series are generated;

[0085] The multi-source time series are preprocessed in three layers to generate standard time series, the first layer is to use a sliding window to perform median filtering on each time series to eliminate sudden noise, the second layer is local neighborhood-based outlier detection, and the abnormal values in outlier detection are repaired using linear interpolation, and the third layer is to normalize the data to convert it to a standard normal distribution, i.e. each data point is subtracted from the global mean and then divided by the global standard deviation, so that different types of data are comparable;

[0086] The standard time series are subjected to Hilbert-Huang transform to generate a time-frequency feature set, including iterative decomposition of each time series to generate intrinsic mode functions (IMFs), and for each IMF component, Hilbert transform is performed to generate an analytical signal, and then the instantaneous frequency, instantaneous amplitude, energy proportion, and time-frequency energy entropy are calculated to extract time-frequency features and obtain the time-frequency feature set;

[0087] Considering the gas exchange factor in the storage environment, the rate of change of carbon dioxide concentration is calculated based on the difference between the carbon dioxide generation rate of fruits and vegetables due to respiration and the concentration change rate caused by ventilation, thereby establishing a differential equation about the gas concentration, since it is difficult to directly measure the respiration rate, the Kalman filtering algorithm is used to estimate the respiration rate, including inverse calculation of the differential equation, and through the Kalman gain, the historical value and the current measured value are fused to suppress noise interference, and a more accurate respiration rate is obtained;

[0088] The chemical potential difference of oxygen, carbon dioxide and ethylene is calculated according to the chemical potential difference formula, and then the thermodynamic force is calculated from the chemical potential difference, and the entropy generation rate is calculated by using the Onsager reciprocal relationship to reflect the degree of irreversibility in the respiratory metabolism process, in the process of calculating the entropy generation rate, the Onsager coefficient will be modified according to the Arrhenius formula with the change of temperature, considering the diffusion entropy flow caused by the ethylene concentration gradient and the metabolic entropy flow caused by the ethylene metabolic heat, the entropy flow density caused by ethylene diffusion is calculated, and the thermodynamic feature set is generated combined with the respiration rate;

[0089] Cross-domain feature correlation analysis is performed on the time-frequency features and the thermodynamic features, including calculating the mutual information matrix to quantify the correlation between the time-frequency features and the thermodynamic features, and finding the optimal linear combination of the two groups of features through canonical correlation analysis to strengthen the coupling relationship;

[0090] The time-frequency feature set and the thermodynamic feature set are fused, since the feature dimension is high, principal component analysis is used for dimension reduction, thereby generating a fusion feature vector, including calculating the covariance matrix of the feature vector, performing eigenvalue decomposition, selecting principal components with cumulative contribution rate exceeding 95%, and generating a low-dimensional feature vector after fusion, retaining the key information that best reflects the respiratory state;

[0091] The respiratory state is divided into five stages: immature, pre-jump, jump, post-jump and aging, the fusion feature vector is input into the Bayesian classifier, the posterior probability is calculated, and the respiratory state is determined, wherein when calculating the posterior probability, the kernel density estimation is used to model the class-conditional probability of the fusion feature vector under each state, and the prior probability is dynamically adjusted according to the storage time of fruits and vegetables;

[0092] To realize the jump peak early warning, the thermodynamic-time frequency coupling factor, the time frequency energy entropy change rate and the entropy production rate change rate are comprehensively considered, the early warning parameters are obtained through weighted calculation, and the potential trend of respiratory jump is quantified; wherein the weight coefficient in the weighted calculation is determined by historical data training, reflecting the contribution of each element to the jump peak; wherein the thermodynamic-time frequency coupling factor quantifies the interaction intensity of thermodynamic entropy change and respiratory signal time frequency characteristics, for each IMF component, the partial derivative of entropy production rate to IMF component instantaneous frequency is calculated to represent the influence degree of frequency change on metabolic entropy, the Pearson correlation coefficient between entropy production rate and IMF component instantaneous frequency is calculated, combined with the partial derivative and the correlation coefficient, the thermodynamic-time frequency coupling factor is obtained by weighted summation according to the energy proportion, and the time frequency energy entropy change rate and the entropy production rate change rate are obtained by derivation on the time frequency energy entropy and the entropy production rate;

[0093] The basic early warning threshold is set to determine the reference value of whether the early warning parameter triggers an alarm, and the integral of the entropy production rate and the ethylene concentration is combined to dynamically adjust the basic early warning threshold, when the real-time calculated early warning parameter exceeds the early warning threshold, the jump peak early warning is triggered to predict the occurrence of respiratory jump in advance;

[0094] The respiratory state recognition result, the jump peak early warning signal and the key thermodynamic and time frequency characteristics are sorted into real-time analysis reports, which are provided to the subsequent regulation system as decision basis, at the same time, to ensure the accuracy and adaptability of the system, the sensor is calibrated by standard gas every fixed period, and the historical data is used to update the Onsager coefficient and the Bayesian classifier parameters, so as to continuously optimize the model performance.

[0095] Specifically, the fuzzy decision method includes:

[0096] The fusion feature vector is converted into a pulse sequence that can be processed by biological neurons, the time sequence characteristics of non-steady fluctuation are retained through time coding, and the problem of losing time correlation in traditional digital signal processing is solved; for each feature parameter in the fusion feature vector, rate coding is performed, the feature amplitude is converted into the corresponding pulse frequency through linear mapping, and it is ensured that the pulse frequency is within the reasonable working range of biological neurons, and the larger the feature amplitude is, the higher the pulse frequency is;

[0097] Based on the mutual information matrix, the features with mutual information higher than a preset value with the entropy production rate are screened out and defined as key features, the key features are linearly transformed through coding coefficients and then the hyperbolic tangent values are taken, so as to obtain the first pulse time combined with the reference time, so as to map the key features to the time dimension;

[0098] The pulse frequency or the first pulse time of each feature is converted into a binary pulse sequence, including periodically emitting pulses according to the frequency for rate-coded features, and generating pulses according to the first pulse time and a fixed interval for key features, and finally aligning the pulse sequences of all features by timestamps to form a multi-dimensional pulse sequence;

[0099] A three-layer pulse neural network is constructed, including an input layer, a hidden layer and an output layer, the number of input layer neurons is consistent with the feature dimension of the fusion feature vector, and the input layer neurons directly receive the pulse sequence, the hidden layer neurons adopt a dynamic threshold mechanism to generate a dynamic threshold, so that each neuron has an adaptively adjusted threshold;

[0100] In the three-layer pulse neural network, the layers are connected through a weight matrix, the weight matrix is randomly initialized and updated through the STDP rule, once the hidden layer neurons receive the pulse sequence of the input layer neurons, the connection weight is adjusted according to the time difference between the preceding and subsequent pulses, if the input pulse precedes the output pulse, indicating that the preceding pulse triggers the subsequent pulse, the weight is increased; if the output pulse precedes the input pulse or arrives at the same time as the input pulse, the weight is decreased, to strengthen the pulse connection with a causal relationship and weaken irrelevant connections;

[0101] At the same time, the trigger condition of the pulse is set, the membrane potential of the hidden neuron is calculated based on the weighted sum of the current time step input pulse and the exponential decay of the previous time step membrane potential, once the membrane potential of the hidden neuron exceeds the dynamic threshold, the pulse is emitted, otherwise it is not emitted, and the membrane potential is reset or updated according to the decay rule, thereby generating a hidden pulse sequence;

[0102] For the features in the fusion feature vector, a unified fuzzy set is defined, the membership of the feature values to each set is calculated using a Gaussian membership function, and a node set containing environmental factors and respiratory factors is established, including but not limited to temperature, ethylene concentration, entropy generation rate and respiratory state, the state value of each node is the activation of the feature after fuzzy, realizing the continuous representation of the factor activation, thereby generating a node activation vector;

[0103] The joint distribution of each node in the historical data is statistically analyzed, the mutual information between nodes is calculated to quantify the statistical correlation between nodes, the greater the mutual information, the stronger the synergistic change trend, and the sensitivity of each feature in the fusion feature vector to the entropy generation rate is calculated based on a thermodynamic model, the mutual information and the sensitivity are combined according to the normalized weight to calculate the causal weight between nodes, thereby generating an initial causal strength matrix;

[0104] The causal weight is updated based on the Hebbian learning rule of biological neurons, when the activation of two nodes exceeds a preset activation threshold, the two nodes are synergistically activated at this time, the causal weight between the nodes is enhanced, the update amplitude is limited by the current weight saturation, and the updated causal strength matrix is output.

[0105] A three-level causal hierarchy is set, and the node activation vector and the causal strength matrix are combined to calculate the multi-scale fused node activation. The three-level causal hierarchy includes direct causality, indirect causality, and high-order causality. The direct causality is calculated by collecting directly connected nodes, calculating the activation weight sum, and generating the direct causal activation through the Sigmoid function. Nodes with a causal weight absolute value exceeding a preset weight value are defined as directly connected nodes. The indirect causality is based on the results of the direct causality, and the above process is repeated to calculate the indirect causal activation. The high-order causality captures the long-chain influence of twice indirect connection, and the high-order causal activation is generated in the same way. The three-level causal results are weighted and summed according to the hierarchical weight to obtain the multi-scale fused node activation, balancing the contributions of different levels of causality.

[0106] The multi-scale fused node activation is mapped to the regulation pulse parameters, and the causal regulation pulse is generated in combination with the regulation priority. This includes setting the change time step, calculating the activation gradient by calculating the absolute value of the change of the node activation within the change time step, multiplying the sum of the absolute values of all causal weights in the corresponding node to obtain the regulation priority, and generating the regulation pulse for the node with high priority. At the same time, the regulation direction is determined according to the size of the node activation, such as generating a cooling pulse for a temperature node with high activation. The pulse key parameters are calculated according to the activation value, and the pulse width is taken as an example. The reference width is associated with the node activation to obtain a width value that is positively correlated with the node activation. The frequency parameter is calculated using the same method, and the emission frequency is adjusted according to the node activation. Finally, the regulation pulse is generated by integrating various parameters to ensure that the pulse parameters can reflect the influence degree of the environmental factors represented by the node activation, realizing the conversion from causal reasoning to specific regulation action.

[0107] For all multi-scale fused node activations, the variance is calculated through a sliding window, compared with the variance threshold to obtain the variance difference, and the activation weight is calculated based on the Sigmoid function. The hidden pulse sequence and the causal regulation pulse are fused based on the activation weight to generate the fused regulation pulse.

[0108] Based on the change rate of entropy generation rate, the fused regulation pulse is thermodynamically optimized to obtain the final regulation pulse. When the change rate of entropy generation rate increases, the pulse intensity is enhanced to ensure that the regulation pulse meets the thermodynamic requirements of respiratory metabolism and improves the physical feasibility.

[0109] Specifically, the specific steps of the dynamic threshold mechanism include:

[0110] The initial threshold is set with the resting potential of biological neurons as a reference, as the baseline level of neuron activation, to avoid non-specific activation. Meanwhile, the product of the accumulated connection weight from the previous time step input layer to the hidden layer and the corresponding pulse signal is calculated to represent the influence strength of the historical input on the current threshold. The historical threshold is exponentially decayed by a forgetting factor to generate a basic threshold with short-term memory characteristics, simulating the adaptive decay of biological neurons to repeated stimulation and avoiding the infinite shift of the threshold due to continuous input.

[0111] The ratio of the current entropy production rate to the historical maximum entropy production rate is calculated to reflect the metabolic intensity. The entropy production rate change rate is introduced to reflect the trend of metabolic acceleration or deceleration. The two indicators are weighted and summed according to the preset weight coefficient to obtain the thermodynamic conditioning factor, which represents the inhibitory regulation of the metabolic intensity in the body on the neural threshold. The preset weight coefficient is determined by historical data training.

[0112] The instantaneous frequency variance is calculated to reflect the dynamic change amplitude of the signal frequency. The reference frequency is compared to quantify the degree of fluctuation abnormality. The time-frequency energy entropy is combined for weighted summation to obtain the time-frequency conditioning factor, which represents the sensitivity enhancement regulation of the neural conditioning to the fluctuating signal.

[0113] To balance the adjustment weight of the thermodynamic and time-frequency conditioning factors, the thermodynamic-time-frequency coupling factor is compared with the preset coupling threshold to calculate the coupling relative size. The coupling relative size is converted into a conditioning fusion coefficient by a Sigmoid function through a fusion slope coefficient to determine the weight proportion of the thermodynamic conditioning factor and the time-frequency conditioning factor in the modulation process. To prevent the fluctuation of the thermodynamic-time-frequency coupling factor from causing a sharp jump in the conditioning fusion coefficient, a first-order low-pass filter is used for smoothing to ensure a smooth transition of the conditioning fusion coefficient and prevent threshold modulation from oscillating.

[0114] The smoothed conditioning fusion coefficient is used as the weight of the thermodynamic conditioning factor, and the remaining weight is allocated to the time-frequency conditioning factor. The influence of the two on the basic threshold is calculated and summed, and then added to the basic threshold to obtain the conditioning threshold, which realizes the dynamic adjustment of the threshold to adapt to different metabolic states and signal fluctuation conditions. When the conditioning fusion coefficient is 1, it represents a strong coupling state, such as the early stage of respiratory transition, and the thermodynamic conditioning factor dominates the regulation, with the threshold increasing with the increase of metabolic intensity to inhibit excessive activation. When the conditioning fusion coefficient is zero, it represents a weak coupling state, such as immaturity, and the time-frequency conditioning factor dominates the regulation, with the threshold decreasing with the increase of signal fluctuation to enhance the response sensitivity. Through the dynamic modulation of the conditioning factor, the threshold realizes the adaptive response to the multi-scale characteristics of respiratory metabolism.

[0115] The identification result of the current breathing state is obtained in real time, and it is judged whether it is in the transition period; if it is in the transition period, the threshold calibration mechanism is activated, the deviation of the current entropy production rate and the reference entropy production value is calculated, the greater the deviation, the farther the metabolism deviates from the normal state, at this time the online fine tuning of the conditioning threshold is carried out in proportion to the deviation, if the deviation is positive, the conditioning threshold is adjusted upward, the response is inhibited, if the deviation is negative, the conditioning threshold is adjusted downward, the adjustment amplitude is nonlinearly enhanced with the deviation degree, so as to calculate the final dynamic threshold; if it is in the non-transition period, the conditioning threshold is maintained unchanged, and no additional calibration is performed, at this time the dynamic threshold is the conditioning threshold.

[0116] Specifically, the specific steps of thermodynamic optimization of the fusion regulation pulse include:

[0117] The basic coefficient is set as the default regulation intensity when the metabolic activity is in a stable state, which provides a reference value for subsequent dynamic adjustment, and ensures that the regulation intensity is moderate when there is no obvious metabolic change;

[0118] The change rate of the entropy production rate at the current time and the past preset time point is calculated in real time, so as to evaluate the change direction and rate of the metabolic activity; if the change rate is positive, it means that the metabolism is accelerated, such as the respiratory transition period; if the change rate is negative, it means that the metabolism is decelerated, such as the aging stage; the greater the absolute value of the difference, the more intense the metabolic change;

[0119] By comparing the change rate of the entropy production rate with the reference entropy production rate, the relative ratio is calculated to measure the degree of deviation of the current metabolic intensity from the reference level, the metabolic adjustment coefficient is generated by nonlinear conversion of the relative ratio by the hyperbolic tangent function, and the final metabolic optimization coefficient is generated by combining the basic coefficient; wherein the metabolic adjustment coefficient is the product of the hyperbolic tangent value and the basic coefficient;

[0120] Based on the metabolic optimization coefficient, the optimized pulse parameters including width and frequency are calculated, and the theoretical entropy production rate change value is calculated by combining the regulation influence coefficient obtained by historical fitting, so as to represent the theoretical influence of the regulation pulse on metabolism;

[0121] If the theoretical entropy production rate change value is less than zero, at this time the regulation leads to the decrease of the entropy production rate, which violates the second law of thermodynamics, the metabolic optimization coefficient is reset to zero immediately, and the pulse width and frequency are recalculated to ensure that the theoretical entropy change is non-negative, so as to ensure that the regulation strategy conforms to the physiological metabolism law; otherwise, the metabolic optimization coefficient is applied.

[0122] Specifically, the synergistic regulation method includes:

[0123] The temperature, humidity, oxygen, carbon dioxide, ethylene and breathing state, and the regulation pulse parameters are set as the agent nodes, an hypergraph model is constructed, the interaction relationship between factors is represented by hyperedges, the interaction intensity is reflected by the weight of hyperedge, the weight of hyperedge is calculated combined with the causal strength matrix and the time synchronization of regulation pulse, and the historical weight is introduced in the calculation process to decay and avoid weight shock;

[0124] Each time a new pulse sequence is received, the time synchronization of factor pulse is recalculated to update the weight of hyperedge, and the weighted update method is used to adjust the topology structure in real time with the change of factor interaction, so as to ensure that the coupling relationship of factors in the critical period such as jump period can be accurately characterized;

[0125] The state of each agent includes the value of its own factor, the connection weight with other factors and the state of neighboring agents, forming an information set required for local decision making. The action of the agent includes the regulation direction, amplitude and pulse parameters. At the same time, in order to enable the agent to learn the optimal regulation strategy autonomously, a regulation reward function is designed, which includes multiple objectives, such as the change of entropy production rate, the change of energy consumption and the smoothness of action. When the respiratory entropy production rate approaches the reference value, a positive reward is obtained, and when the energy consumption decreases, a positive reward is obtained, while frequent start-stop is inhibited to reduce equipment wear and tear.

[0126] Each agent uses a distributed reinforcement learning algorithm such as PPO algorithm, takes the regulation pulse as the initial regulation strategy to avoid random exploration, and constantly updates the regulation strategy according to the current state and reward to optimize its own regulation action. Each agent can make decisions quickly based on local information, solving the problem of response lag in traditional centralized control.

[0127] To realize the coordinated regulation of multiple factors, each agent is set as a pulse coupled oscillator (PCO), and a main regulation factor is defined. The pulse frequency of the main regulation factor is set as the reference frequency of the PCO network, and the initial phase of other factors is determined according to the causal strength. The higher the causal strength, the smaller the phase difference with the main regulation factor. At the same time, the inherent frequency and the topology weight of adjacent agents are combined to update the phase of each agent in real time, and the time difference of regulation action is adjusted. Finally, the phase difference is mapped to the pulse time difference, and the topology weight is mapped to the pulse width, to generate a regulation pulse sequence for multiple factor coordination.

[0128] A global optimization objective is constructed by comprehensively considering the entropy generation rate deviation, energy consumption entropy increase ratio, and consistency of the strategy and control pulse. The consistency term forces the control strategy and control pulse parameters to match, ensuring thermodynamic constraints. The agents periodically synchronize the topological weights and strategy parameters, use a consistency algorithm to make the parameters in the neighborhood consistent, and solve the global optimization problem using an alternating direction multiplier method (ADMM). The optimized control strategy is checked for thermodynamic feasibility, the change in entropy generation rate after the execution of the control strategy is simulated, and if the thermodynamic law is violated, the action direction is adjusted. The control strategy that violates the rules is penalized using a causal strength matrix, and the optimized control strategy that matches the dynamic topology and meets the global thermodynamic objective is finally obtained.

[0129] Specifically, the constraint learning method includes:

[0130] Multi-dimensional feedback data is collected and aligned with the corresponding control actions by timestamp. The mean and variance are calculated through a sliding window to eliminate random fluctuations.

[0131] A feedback reward function is designed based on thermodynamic feedback characteristics, including entropy change convergence reward, energy efficiency reward, thermodynamic constraint term, and state recognition reward. This avoids the problem of the optimization direction being disconnected from the physical law in traditional methods, ensuring that the evolution process always complies with the principles of non-equilibrium thermodynamics. The entropy change convergence reward punishes the entropy generation rate deviation and its rate of change, promoting the convergence of respiratory metabolism to the target state. The energy efficiency reward rewards the improvement of the energy consumption entropy increase ratio while suppressing its fluctuations, ensuring energy-saving and stable control. The thermodynamic constraint term enforces the non-negativity of the entropy generation rate, imposes a penalty when violated, and penalizes the accumulation of device dissipation entropy and frequent changes in control actions. The state recognition reward combines the accuracy of respiratory state recognition to strengthen the accurate judgment of the metabolic stage.

[0132] The various parameters in the above methods are encoded as chromosomes, and evolution operators are designed to simulate the material transport and signal transduction mechanisms of plant vascular systems to generate candidate optimization parameters, including: selecting parents based on topological weight similarity, performing vascular bundle crossover at the super-edge weight layer, adjusting the mutation amplitude based on the entropy generation rate deviation for key parameters affecting entropy change, the larger the deviation, the higher the mutation strength, directional search for optimal parameters, and screening historical optimal individuals that comply with the laws of thermodynamics to ensure the physical feasibility of the evolution direction.

[0133] The candidate optimization parameters are strictly tested for thermodynamics, including: correcting the Onsager coefficient to satisfy symmetry, ensuring that the thermodynamic flow coupling relationship complies with physical laws, adjusting thermodynamic force parameters to ensure non-negative entropy generation rate, imposing a penalty when violated and re-optimizing, and verifying parameter effectiveness on short-term and long-term scales to balance short-term control accuracy and long-term energy optimization.

[0134] Meanwhile, the candidate optimization parameters are updated by a thermodynamic adaptive step size, and the step size is related to the entropy generation rate and the reward gradient, ensuring that each iteration responds to real-time feedback and meets the thermodynamic constraints;

[0135] Every time a control cycle is completed, an evolutionary iteration is triggered, and the optimal parameters obtained by evolution are applied.

[0136] Embodiment 2:

[0137] Please refer to Fig. 6 , the present application provides another embodiment: a multi-factor synergistic control system for fruit and vegetable respiration intensity, comprising: a feature analysis module, a pulse decision module, a control module and a constraint learning module;

[0138] The feature analysis module is used to collect and store multi-modal environmental data and generate a fusion feature vector. Temperature, humidity and gas concentration sensors are deployed on a three-dimensional grid. The sensors are synchronized and sampled at fixed intervals. The time series are generated by GPS clock synchronization and hardware filtering. After preprocessing by median filtering, outlier repair and standardization, the data is subjected to Hilbert-Huang transform to generate intrinsic mode functions and calculate time-frequency features such as instantaneous frequency and energy entropy. The respiration rate is calculated by combining the gas concentration differential equation and Kalman filter inversion. The chemical potential difference and entropy generation rate are calculated to generate thermodynamic features. Finally, the time-frequency and thermodynamic features are fused by mutual information analysis and principal component dimensionality reduction, and input into a Bayesian classifier to identify the respiration state.

[0139] The pulse decision module is used to convert the fusion feature vector into a control pulse sequence. The features are mapped into pulse frequency and timing by rate coding and time coding. A three-layer pulse neural network is constructed. The weight matrix is dynamically updated using the STDP rule. The neuron threshold is adjusted by combining thermodynamic and time-frequency adjustment factors. The hidden layer pulse sequence is generated. A fuzzy cognitive map is also constructed. The node activation and causal weight are calculated. The causal control pulse is generated by three-level causal layering. Finally, the control pulse sequence is dynamically fused according to the node activation variance. The precise control pulse is output after thermodynamic optimization.

[0140] The control module is used to implement a multi-factor synergistic control strategy. A hypergraph model containing environmental factors is constructed. The hyperedge weight is updated according to the causal strength and pulse time synchronicity. The key factor interaction is focused. Each agent optimizes the control strategy based on reinforcement learning. The entropy change convergence and energy efficiency are used as objective functions. The action parameters are iteratively updated using the PPO algorithm. The synergistic pulse sequence is generated by the pulse coupled oscillator network according to the topological weight and phase difference. Finally, the global strategy is optimized by a consensus algorithm to ensure that the control meets the thermodynamic constraints and the respiration state requirements.

[0141] ​The constraint learning module is used for autonomous evolutionary optimization of system parameters, collects feedback data, designs a feedback reward function containing thermodynamic constraints, encodes parameters into chromosomes, generates candidate optimization parameters through biological heuristic operators, verifies symmetry through Onsager coefficients and non-negativity of entropy production rate, iteratively updates candidate optimization parameters using adaptive step length, triggers evolutionary iteration every N control cycles, and improves long-term adaptability of the system.

[0142] In summary, the present application solves the problems of inaccurate respiratory state recognition and delayed jump peak warning by collecting multi-modal data, generating multi-source time series through GPS clock unified timestamp and hardware low-pass filtering, generating standard time series through median filtering, isolated point repair and standardization processing, performing Hilbert-Huang transform and thermodynamic calculation on the standard time series, estimating respiratory rate combined with Kalman filtering, generating fusion feature vectors through mutual information analysis and principal component dimensionality reduction, inputting the Bayesian classifier and dynamically adjusting the warning threshold; the fusion feature vectors are converted into pulse sequences, a pulse neural network is constructed and a dynamic threshold mechanism is introduced, the weight matrix is updated through the STDP rule, and the problem of losing time correlation in traditional signal processing is solved; a fuzzy cognitive map is constructed, causal control pulses are generated through three-level causal layering, thermodynamic optimization is performed combined with the change rate of entropy production rate, and the problem of insufficient multi-factor nonlinear coupling reasoning is solved; a dynamic hypergraph model is constructed with multi-factors as agents, the regulation strategy is optimized based on reinforcement learning, and collaborative pulse sequences are generated using pulse coupled oscillators, through consistency algorithm and thermodynamic verification, the problems of multi-factor collaborative regulation asynchronization and local optimum are solved; feedback data are collected and a feedback reward function with thermodynamic constraints is designed, parameters are encoded and evolved through biological heuristic operators, and parameters are updated through thermodynamic verification and multi-scale simulation, solving the problems of regulation strategy solidification and poor long-term adaptability. This process solves the core technical problems of signal processing, causal reasoning, collaborative optimization and strategy evolution in respiratory regulation through multi-module collaboration from data acquisition to strategy evolution.

[0143] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the scope of the present application should be considered within the protection scope of the present application. It should be noted that for ordinary technical personnel in the technical field, some improvements and decorations without departing from the principles of the present application should also be considered within the protection scope of the present application.

Claims

1. A method for multi-factor synergistic regulation of respiration intensity of fruits and vegetables, characterized in that, The method comprises the following steps: Collecting multi-modal data and preprocessing, feature analysis, generating fusion feature vector, and identifying respiratory state; Convert the fusion feature vector to pulse sequence, set fuzzy decision method, construct three-layer pulse neural network, combine dynamic threshold mechanism to generate hidden pulse sequence, define fuzzy set for the fusion feature vector, generate node activation vector, build causal regulation pulse through fuzzy cognitive map, dynamically fuse the hidden pulse sequence and the causal regulation pulse based on node activation variance to generate fusion regulation pulse, and thermodynamically optimize the fusion regulation pulse based on the change rate of entropy production rate; wherein the three-layer pulse neural network comprises an input layer, a hidden layer and an output layer, and the hidden layer neurons adopt a dynamic threshold mechanism to generate a dynamic threshold, so that each neuron has an adaptively adjusted threshold; Build a hypergraph model with multiple factors as agent nodes, set a collaborative regulation method, dynamically update the hyperedge weight based on causal strength and pulse time synchronicity, and optimize the regulation strategy through reward function and PPO algorithm based on the local state of the agent, and use the pulse coupled oscillator to generate a collaborative pulse sequence with the main regulation factor frequency as the reference and the phase difference determined by the causal strength; wherein temperature, humidity, oxygen, carbon dioxide, ethylene, respiratory state and regulation pulse parameters are set as agent nodes to construct a hypergraph model, the interaction between factors is represented by hyperedges, and the interaction strength is reflected by the weight of hyperedges; the weight of hyperedges is calculated based on the causal strength matrix and the time synchronicity of the regulation pulse, and historical weights are introduced during the calculation process; Collect multi-dimensional feedback data and preprocess, set constraint learning method, generate candidate optimization parameters and update in real time; wherein the parameters in the fuzzy decision method and the collaborative regulation method are encoded as chromosomes, and evolution operators are designed to simulate plant vascular systems to generate candidate optimization parameters.

2. The multi-factor collaborative regulation method for fruit and vegetable respiration intensity according to claim 1, wherein: The fuzzy decision method comprises: Rate coding of the fusion feature vector converts feature amplitude to pulse frequency; Filtering out key features, taking the tangent hyperbolic value after linear transformation of the key features by coding coefficients to obtain the first pulse time; Convert the pulse frequency and the first pulse time to a binary pulse sequence to generate a multi-dimensional pulse sequence; Constructing a three-layer pulse neural network, wherein the hidden layer neurons adopt a dynamic threshold mechanism to generate a dynamic threshold; Initializing the weight matrix and updating the weight matrix through the STDP rule; Setting the trigger condition of the pulse, calculating the membrane potential based on exponential decay, firing a pulse once the membrane potential exceeds the dynamic threshold, otherwise not firing, and generating a hidden pulse sequence.

3. The multi-factor collaborative regulation method for fruit and vegetable respiration intensity according to claim 2, wherein: The fuzzy decision method further comprises: Defining a fuzzy set for the fusion feature vector to generate a node activation vector; Statistical analysis is performed on the joint distribution of each node in the historical data, mutual information between nodes is calculated, the sensitivity of each feature in the fusion feature vector to the entropy production rate is calculated based on a thermodynamic model, the mutual information and the sensitivity are combined according to a normalized weight, the causal weight between nodes is calculated and updated, and a causal strength matrix is obtained; A three-level causal hierarchy is set, node activation is calculated for multi-scale fusion, and causal control pulses are generated by mapping the activation to control pulse parameters and combining control priorities; wherein the three-level causal hierarchy includes direct causality, indirect causality and high-order causality, the direct causality is calculated by collecting directly connected nodes and calculating the weighted sum of activation, and the direct causality activation is generated by a Sigmoid function, while the nodes with absolute causal weight exceeding a preset weight value are defined as directly connected nodes, the indirect causality is based on the results of the direct causality, and the indirect causality activation is calculated by repeating the above process, and the high-order causality is generated by capturing the long-chain influence of twice indirect connection. The variance of the node activation is calculated to determine the activation weight, the hidden pulse sequence and the causal control pulse are fused, and the final control pulse is obtained based on the rate of change of the entropy production rate.

4. The multi-factor synergistic regulation method for fruit and vegetable respiration intensity according to claim 3, characterized in that: The specific steps of feature analysis include: Collecting multi-modal data, unifying time stamps and denoising, performing three-layer preprocessing, and generating standard time series; Performing Hilbert-Huang transform on the standard time series to obtain a time-frequency feature set; Establishing a gas concentration differential equation and using Kalman filter algorithm to inversely calculate the respiration rate; Calculating the entropy production rate, considering the ethylene diffusion entropy flow and metabolic entropy flow, and generating a thermodynamic feature set; Calculating the mutual information matrix and performing cross-domain correlation analysis on the time-frequency feature set and the thermodynamic feature set through canonical correlation analysis; Using principal component analysis to reduce the dimensionality of the time-frequency feature set and the thermodynamic feature set to generate a fusion feature vector.

5. The multi-factor synergistic regulation method for fruit and vegetable respiration intensity according to claim 4, characterized in that: The specific steps of feature analysis further include: Inputting the fusion feature vector into a Bayesian classifier, modeling the conditional probability through kernel density estimation and dynamically adjusting the prior probability to identify the respiration state; Integrating the thermodynamic-time-frequency coupling factor, the time-frequency energy entropy change rate and the entropy production rate change rate to calculate the warning parameter by weighting; Setting a basic warning threshold, dynamically adjusting the basic warning threshold based on the integral of the entropy production rate and the ethylene concentration, and triggering the jump peak warning when the real-time calculated warning parameter exceeds the warning threshold; Organizing the respiration state identification results, jump peak warning signals and key features to generate a real-time analysis report, and regularly calibrating the sensors and updating the Onsager coefficient and the Bayesian classifier parameters.

6. The multi-factor synergistic regulation method for fruit and vegetable respiration intensity according to claim 5, characterized in that: The specific steps of the dynamic threshold mechanism include: Setting an initial threshold value with reference to the resting potential of biological neurons; The connection weight of the previous time step input layer to the hidden layer is multiplied by the corresponding pulse signal, and the historical threshold is exponentially attenuated by combining the forgetting factor to generate the basic threshold; The ratio of the current entropy generation rate to the historical maximum entropy generation rate and the entropy generation rate change rate are calculated, and the weighted sum is obtained by weighting according to the preset weight coefficient to obtain the thermodynamic conditioning factor; The instantaneous frequency variance is calculated and compared with the reference frequency, and the time-frequency energy entropy is weighted and summed to obtain the time-frequency conditioning factor; The thermodynamic-time-frequency coupling factor is compared with the preset coupling threshold, combined with the fusion slope coefficient, converted into the conditioning fusion coefficient through the Sigmoid function, and smoothed by a first-order low-pass filter; The smoothed conditioning fusion coefficient is used as the weight of the thermodynamic conditioning factor, and the remaining weight is allocated to the time-frequency conditioning factor. The influence of the thermodynamic conditioning factor and the time-frequency conditioning factor on the basic threshold is calculated and summed with the basic threshold to obtain the conditioning threshold; The identification result of the respiratory state is obtained in real time. If it is in the transition period, the deviation of the current entropy generation rate from the reference entropy generation value is calculated, and the conditioning threshold is fine-tuned online according to the deviation ratio to obtain the final dynamic threshold, otherwise the conditioning threshold is used as the final dynamic threshold.

7. The multi-factor synergistic regulation method of fruit and vegetable respiration intensity according to claim 6, characterized in that: The specific steps of thermodynamic optimization of the fusion regulation pulse include: setting a basic coefficient, calculating the entropy generation rate change rate at the current time and a preset time point in the past; comparing the entropy generation rate change rate with the reference entropy generation rate to obtain a relative ratio, converting the relative ratio through a hyperbolic tangent function to generate a metabolic adjustment coefficient, and combining the basic coefficient to obtain a metabolic optimization coefficient; based on the metabolic optimization coefficient, calculating the optimized pulse width and frequency, and combining the regulation influence coefficient to calculate the theoretical entropy generation rate change value; if the theoretical entropy generation rate change value is less than zero, reset the metabolic optimization coefficient to zero and recalculate the pulse width and frequency, otherwise apply the metabolic optimization coefficient.

8. The multi-factor synergistic regulation method of fruit and vegetable respiration intensity according to claim 7, characterized in that: the synergistic regulation method includes: constructing a hypergraph model, setting an agent node, representing the interaction relationship between factors with a hyperedge, and calculating the hyperedge weight; when a new pulse sequence is received, recalculate the factor pulse time synchronization, and update the hyperedge weight in a weighted update manner; define the state of the agent, design the regulation reward function; each agent uses a distributed reinforcement learning algorithm, takes the regulation pulse as the initial regulation strategy, and updates the initial regulation strategy according to the current state and reward; set up a pulse coupled oscillator, define a main regulation factor, and generate a synergistic pulse sequence; construct a global optimization goal, periodically synchronize the hyperedge weight and the regulation strategy parameter; use the consensus algorithm and the alternating direction multiplier method to solve the global optimization problem, and perform thermodynamic feasibility check on the optimized regulation strategy to adjust the illegal regulation strategy, and obtain the optimized regulation strategy.

9. The multi-factor synergistic regulation method of fruit and vegetable respiration intensity according to claim 8, characterized in that: the constraint learning method includes: Collect multi-dimensional feedback data and align with the time stamp of the regulation action, calculate the mean and variance through the sliding window; Design a feedback reward function based on the thermodynamic feedback characteristics, which includes entropy change convergence reward, energy efficiency reward, thermodynamic constraint term and state recognition reward; Encode the parameters in the fuzzy decision method and the collaborative regulation method as chromosomes, simulate plant vascular system design evolution operators to generate candidate optimization parameters; Thermodynamic test on the candidate optimization parameters, including correcting the Onsager coefficient to meet the symmetry, adjusting the thermodynamic force parameter to ensure the non-negativity of entropy production rate, and verifying the effectiveness of the candidate optimization parameters; Update the candidate optimization parameters through the thermodynamic adaptive step, which is related to the entropy production rate and the reward gradient; Every time the secondary control loop is completed, the evolutionary iteration is triggered and the optimal parameters are applied.

10. A multi-factor synergistic regulation system for respiration intensity of fruits and vegetables, which is used to realize the multi-factor synergistic regulation method for respiration intensity of fruits and vegetables as claimed in any one of claims 1-9, characterized in that, Comprise: Feature analysis module, pulse decision module, regulation module and constraint learning module; The feature analysis module is used for deploying sensors on a three-dimensional grid to collect multi-modal data, after time stamp synchronization, hardware filtering and preprocessing, generating time-frequency features through Hilbert-Huang transform, combining gas concentration differential equation and Kalman filter to generate thermodynamic features, and inputting into Bayesian classifier to identify respiratory state after fusion; The pulse decision module is configured to encode the fusion feature vector into a pulse sequence through rate coding, construct a pulse neural network and a fuzzy cognitive map, generate a hidden pulse sequence and a causal regulation pulse, and output a fusion regulation pulse after thermodynamic optimization. The regulation module is used for constructing a hypergraph model, updating the hyperedge weight according to the causal strength and the time synchronization of pulse, and generating a collaborative pulse sequence through a pulse coupled oscillator based on the reinforcement learning algorithm of each intelligent agent, and outputting the regulation strategy after optimization by the consistency algorithm; wherein, temperature, humidity, oxygen, carbon dioxide, ethylene and respiratory state, regulation pulse parameters are set as intelligent agent nodes, a hypergraph model is constructed, the interaction relationship between factors is represented by hyperedge, the interaction strength is reflected by hyperedge weight, the hyperedge weight is calculated combined with the causal strength matrix and the time synchronization of regulation pulse, and the historical weight is introduced in the calculation process; The pulse decision module is configured to encode the fusion feature vector into a pulse sequence through rate coding, construct a pulse neural network and a fuzzy cognitive map, generate a hidden pulse sequence and a causal regulation pulse, and output a fusion regulation pulse after thermodynamic optimization. The constraint learning module is used for collecting feedback data, designing a feedback reward function, encoding parameters as chromosomes, generating candidate optimization parameters through biological heuristic operators, and iteratively updating after thermodynamic test to optimize system parameters. ​

Images