Method and system for intelligent regulation and control of pericarpium citri ricti storage environment based on multi-source data fusion

Abstract

The application discloses a pericarpium citri reticulatae storage environment intelligent regulation and control method and system based on multi-source data fusion, and relates to the technical field of intelligent regulation and control; the method comprises the following steps: deploying an environment sensing network in a selected storage area, accessing a meteorological data interface, and integrating storage business data; the technical key points are as follows: the comprehensive environment index obtained by the scheme is not only used as the input of a mixed temperature and humidity regulation and control model to generate accurate temperature and humidity regulation and control targets, but also can be used for dynamically correcting the regulation and control priority score of the storage area, realizes effective reuse of multi-source information in the prediction and decision-making double links, balances long-term value orientation and short-term risk response, solves the technical problems that the regulation and control strategy of the traditional storage system is rigid and cannot respond to sudden environmental disturbances in a coordinated manner, and achieves the effects of pericarpium citri reticulatae storage guarantee, active defense in high-risk scenarios and multi-storage resource collaborative optimization, and the stability, safety and intelligent level of the pericarpium citri reticulatae storage environment are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to an intelligent control method and system for the storage environment of dried tangerine peel based on multi-source data fusion. Background Technology

[0002] Current intelligent temperature and humidity control technologies for warehousing mainly rely on the Internet of Things, cloud computing, and artificial intelligence. They collect environmental data in real time through high-precision sensors and upload it to a cloud platform via wireless methods such as LoRa, ZigBee, or Wi-Fi. The platform analyzes the data based on preset rules to achieve anomaly warnings and automatic control. At the same time, the execution end can link with equipment such as air conditioners, dehumidifiers, humidifiers, and fans, supporting both manual and automatic modes. Some systems also integrate technologies such as digital twins, blockchain evidence storage, and controlled atmosphere sealing, making them suitable for environmentally sensitive warehousing scenarios such as pharmaceuticals, grains, and tobacco.

[0003] In traditional tangerine peel storage management, temperature and humidity control generally relies on fixed thresholds for start-stop control, lacking the ability to respond to differences in the age and type of tangerine peel, or even external weather changes. This results in using the same standards for old and new tangerine peels, which may lead to mold risks due to insufficient control, or energy waste and cracking due to excessive dehumidification. At the same time, the control strategies of each storage area are isolated from each other. When adjacent storage areas trigger equipment commands simultaneously due to local temperature and humidity fluctuations, dehumidifiers and humidifiers often operate synchronously, which may cause mutual interference due to improper distance or isolation. This not only reduces system efficiency but also accelerates equipment aging. In addition, when faced with sudden disturbances such as rainy season or air conditioning failures, the system can often only passively alarm and cannot actively adjust resource priorities. For example, storage areas for tangerine peels older than ten years and ordinary raw material storage areas should have different response levels under the same humidity, but the existing system cannot dynamically identify and assign differentiated treatment weights, causing high-value stored products to be exposed to the risk of irreversible quality loss. All of these problems will seriously restrict the development or improvement of the stability and safety of tangerine peel storage. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for intelligent control of the storage environment of dried tangerine peel based on multi-source data fusion, the method including:

[0006] An environmental sensor network is deployed within the selected warehousing area, connected to a meteorological data interface, and integrated with warehousing business data to construct a multi-source heterogeneous dataset.

[0007] Under the condition of dealing with different target small warehouses in the selected storage area, three types of feature data are extracted based on multi-source heterogeneous datasets to establish a feature vector set, and the configured data fusion strategy is run to comprehensively process the feature vector set in order to obtain the fused comprehensive environmental index.

[0008] A hybrid temperature and humidity control model is constructed by combining a Long Short-Term Memory (LSTM) network with a Random Forest Regressor (RF). After obtaining the LSTM predictions and RF correction terms, the model jointly outputs the recommended target temperature (TL) and the recommended target humidity (HL), and triggers the model running and online update mechanism.

[0009] The control level rule engine runs on the target small warehouse, and the control sequence is determined according to the control level corresponding to the target small warehouse. The multi-warehouse collaborative control algorithm is implemented simultaneously, and multiple early warning mechanisms are triggered. The operation of the control level rule engine is based on: constructing an initial control priority score, calling comprehensive environmental indicators to complete the weighted correction action to obtain the final control priority score, and comparing the final control priority score with the configured standard threshold range to obtain the control level corresponding to the target small warehouse.

[0010] Furthermore, the deployed environmental sensor network is a distributed temperature and humidity sensor array; the meteorological data access interface means accessing third-party weather services via API to obtain real-time meteorological data for the selected storage area, including at least: outdoor temperature, relative humidity, rainfall probability, and wind speed; the storage business data includes at least: warehousing information, storage period, transaction voucher number, and pesticide residue test report status; among which, the warehousing information includes the variety, year, origin, weight of dried tangerine peel, and whether drying and care services are required; the constructed multi-source heterogeneous dataset is: the temperature and humidity data, meteorological data, and storage business data obtained by the environmental sensor network are used as raw data, and after data cleaning, they are stored.

[0011] Furthermore, the extracted three types of feature data are: local environment feature vector, neighboring warehouse environment feature vector, and environmental business feature vector. Among them, the local environment feature vector F1 includes at least: temperature T_local and humidity H_local, the average temperature T_avg and the average humidity H_avg of the past 1 hour, the temperature change rate dT / dt, and the humidity change rate dH / dt; the neighboring warehouse environment feature vector F2 includes at least: the temperature difference ΔT_adj and the humidity difference ΔH_adj between adjacent warehouses, the weighted average temperature T_wei of other warehouses in the same target warehouse, and the weighted average humidity H_wei of other warehouses in the same target warehouse; the environmental business feature vector F3 includes at least: outdoor temperature T_ext, rainfall probability P_rain, tangerine peel variety onehot, year Year, and care mark Cfg.

[0012] Furthermore, the process of running the data fusion strategy based on the feature vector set is as follows: the local environment feature vector, the neighboring warehouse environment feature vector, and the environmental business feature vector are normalized to [0, 1] respectively, and then fused into a single scalar score according to the weights to obtain the fused comprehensive environmental index F_final.

[0013] Furthermore, the operation of the Long Short-Term Memory (LSTM) network is as follows: it takes the 24-hour temperature and humidity time series as input and outputs the LSTM prediction values, namely the temperature prediction value T_pred and the humidity prediction value H_pred for the next hour; the operation of the Random Forest Regressor (RF) is as follows: it takes the comprehensive environmental index F_final as input and outputs the RF correction terms, namely the temperature correction term ΔT_corr and the humidity correction term ΔH_corr.

[0014] Furthermore, the basis for the joint output based on the LSTM prediction and the RF correction term is as follows:

[0015] The predicted temperature value T_pred is summed with the temperature correction term ΔT_corr to generate the recommended target temperature TL.

[0016] The predicted humidity value H_pred is summed with the humidity correction term ΔH_corr to generate the recommended target humidity HL.

[0017] The basis for triggering the model execution and online update mechanism is as follows:

[0018] Supervised learning training was conducted using historical datasets, with labels set as pre-defined ideal temperature and humidity ranges. Cross-validation was used to divide the training and testing sets, and an online learning mechanism was introduced. Whenever a new batch of dried tangerine peel entered or left the warehouse, the actual operating results were collected and fed back to the hybrid temperature and humidity control model for incremental training.

[0019] Furthermore, the initial score S0 for regulatory priority is constructed based on the tangerine peel value score and environmental risk score contained in F_final. The tangerine peel value score V_pel is set according to the following rules: when the year is ≥5, V_pel=1; when 5 > year ≥3, V_pel=0.8; when the year is <3, V_pel=0.6. The environmental risk score R_env is calculated by weighting the temperature change rate dT / dt, the temperature difference ΔT_adj between adjacent small warehouses, and the rainfall probability P_rain obtained from historical statistics, and normalized to [0,1]. The tangerine peel value score V_pel and the environmental risk score R_env are weighted and summed to generate the initial score S0 for regulatory priority. The process of calling F_final to complete the weighted correction is as follows:

[0020] A correction factor σ(F_final) is introduced: the temperature change rate dT / dt, the temperature difference ΔT_adj between adjacent small warehouses, and the rainfall probability P_rain extracted from the comprehensive environmental index F_final are normalized and then weighted to obtain the correction factor σ(F_final); a risk amplification coefficient is introduced: the correction factor σ(F_final) is multiplied by the risk amplification coefficient, and the first result is accumulated with 1 to obtain the risk correction term; the final control priority score S is defined: the initial control priority score S0 is multiplied by the risk correction term to generate the final control priority score S.

[0021] Furthermore, the final control priority score S is compared with the configured standard threshold range [Q1, Q2] as follows:

[0022] When S≥Q2, the control level corresponding to the target small warehouse is marked as high priority;

[0023] When Q2>S≥Q1, the control level corresponding to the target small warehouse is marked as medium priority;

[0024] When S < Q1, the control level corresponding to the target small warehouse is marked as low priority;

[0025] Where Q1 < Q2, and the values ​​of Q1 and Q2 are both in the range of 0 to 1.

[0026] Furthermore, the process of implementing the multi-warehouse collaborative control algorithm is as follows: obtain the recommended target temperature TL and recommended target humidity HL for each target warehouse; check the control levels of adjacent warehouses; if the control levels of adjacent warehouses are the same and there is a conflict, call the game theory model to solve the Nash equilibrium solution, output the final instruction and send it to the execution unit.

[0027] The process of triggering the multiple early warning mechanism is as follows: the final actual temperature and humidity of each target compartment are obtained. When the actual temperature deviates from the corresponding recommended target temperature TL±1℃ or the actual humidity deviates from the corresponding recommended target humidity HL±5%, a level one early warning is triggered. When the actual temperature deviates from the corresponding recommended target temperature TL±1℃ and the actual humidity deviates from the corresponding recommended target humidity HL±5% for at least 3 consecutive monitoring cycles, a level two early warning is triggered. When a mold image is detected by a pre-configured camera, a level three early warning is triggered.

[0028] A smart control system for the storage environment of dried tangerine peel based on multi-source data fusion, the system includes:

[0029] Heterogeneous data construction module: Deploy an environmental sensor network within the selected warehousing area, connect to the meteorological data interface, and integrate warehousing business data to construct a multi-source heterogeneous dataset;

[0030] Environmental indicator detection module: Under the condition of dealing with different target small warehouses in the selected storage area, it extracts three types of feature data based on multi-source heterogeneous datasets, establishes feature vector sets, and runs the configured data fusion strategy to comprehensively process the feature vector sets to obtain the fused comprehensive environmental indicators.

[0031] The regulation model building module combines a Long Short-Term Memory (LSTM) network with a Random Forest Regressor (RF) to construct a hybrid temperature and humidity regulation model. After obtaining the LSTM predictions and RF correction terms, it jointly outputs the recommended target temperature (TL) and recommended target humidity (HL) and triggers the model running and online update mechanism.

[0032] Collaborative Control and Early Warning Module: The control level rule engine runs on the target small warehouse, formulates the control sequence according to the control level corresponding to the target small warehouse, implements multi-warehouse collaborative control algorithm simultaneously, and triggers multiple early warning mechanisms. The operation of the control level rule engine is based on: constructing an initial control priority score, calling comprehensive environmental indicators to complete the weighted correction action to obtain the final control priority score, and comparing the final control priority score with the configured standard threshold range to obtain the control level corresponding to the target small warehouse.

[0033] This invention provides a method and system for intelligent control of the storage environment of dried tangerine peel based on multi-source data fusion, which has the following beneficial effects:

[0034] (1) The comprehensive environmental indicators obtained by this scheme are not only used as inputs to the hybrid temperature and humidity control model to generate accurate temperature and humidity control targets, but also used to dynamically correct the control priority score of the warehouse area. This realizes the effective reuse of multi-source information in both prediction and decision-making, balances long-term value orientation and short-term risk response, and to a certain extent solves the technical problems of rigid control strategies and inability to coordinate responses to sudden environmental disturbances in traditional warehousing systems. It achieves the effects of ensuring the storage of dried tangerine peel, completing active defense in high-risk scenarios, and optimizing the collaborative resources of multiple warehouses, significantly improving the stability, safety and intelligence level of the dried tangerine peel storage environment.

[0035] (2) This solution achieves dynamic and precise control of the storage environment of dried tangerine peel. It integrates three types of characteristic data to construct a hybrid temperature and humidity control model, and generates personalized temperature and humidity targets that change with time, category and weather, i.e. recommended target temperature and humidity, replacing the traditional fixed threshold control. At the same time, it introduces a corresponding mechanism for control priority scoring, so that high-year and high-risk warehouses can obtain more sensitive responses, solving the problems of energy waste or quality control failure that are easy to occur in traditional warehouse control, and achieving the effect of on-demand adjustment, quality preservation and energy saving. The control level is dynamically divided by the control level rule engine, and the game theory is called to solve the Nash equilibrium when the instructions of adjacent warehouses conflict, avoiding equipment confrontation operation. Combined with three-level early warning to complete different degrees of response, the overall robustness of the operating system is improved, ensuring the effective development of regional intelligence. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall process of the intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion in this invention. Detailed Implementation

[0037] 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. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1: Please refer to Figure 1 This embodiment provides an intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion. The core of this method lies in its ability to combine IoT sensing, machine learning modeling, and dynamic feedback control to achieve a complete management system with adaptive adjustment capabilities. The recorded scheme not only improves the stability of the storage environment but also enhances the ability to ensure the quality of dried tangerine peel under complex scenarios. Therefore, the specific scheme of this method is as follows:

[0039] S1. Deploy an environmental sensor network within the selected storage area, connect to the meteorological data interface, and integrate storage business data to construct a multi-source heterogeneous dataset; wherein, the selected storage area is the area selected in actual work, and the corresponding intelligent control system for the storage environment of dried tangerine peel based on multi-source data fusion is deployed in this area;

[0040] The deployed environmental sensing network is a sensor array. In this embodiment, a distributed temperature and humidity sensor array is used. It is specifically deployed in the target warehouse or the smaller warehouses contained within the target warehouse in the selected storage area. The target warehouse is, for example, Warehouse A, Warehouse 1 or Warehouse B, Warehouse 2, etc. The distributed temperature and humidity sensor array can be selected with industrial-grade sensors, specifically the SHT31-D model. Each sensor in the distributed temperature and humidity sensor array can upload data to the central database every 5 minutes during actual operation, and mark the timestamp and spatial coordinates (X, Y, Z). For example, Warehouse A, Warehouse 1 has a total of 8 temperature and humidity sensors installed, located at the four corners and the center, forming a three-dimensional distributed array monitoring network.

[0041] Accessing the meteorological data interface means connecting to the meteorological bureau's open platform or third-party weather services via API to obtain real-time meteorological data for the selected storage area. This data includes at least outdoor temperature, relative humidity, probability of rainfall, and wind speed. The meteorological data can predict indoor environmental trends over the next 24 hours, allowing for advance adjustments to control strategies. For example, if the forecast indicates heavy rainfall in the next 6 hours, the dehumidification power of all target warehouses needs to be increased to prevent moisture penetration and mold growth. The specific increase in dehumidification power can be set based on historical data or experience.

[0042] The integrated warehousing business data includes at least: inbound information, storage period, transaction voucher number, and pesticide residue test report status. The warehousing business data is structured and stored in a relational database, with each record corresponding to a unique ID and bound to the physical storage location of the target warehouse. The inbound information includes the variety of dried tangerine peel, such as: red peel / green peel, year, place of origin, weight, and whether drying and care services are required.

[0043] The basis for constructing the multi-source heterogeneous dataset is as follows: establish a distributed data lake with Hadoop+Spar as the core, and use temperature and humidity data obtained from environmental sensor networks, meteorological data obtained from meteorological data interfaces, and warehousing business data as raw data, and store them uniformly after data cleaning; the established distributed data lake is easy to call later, and supports fast retrieval by warehouse area, time window, category and other dimensions.

[0044] S2. Under the condition of dealing with different target small warehouses in the selected storage area, extract three types of feature data based on multi-source heterogeneous datasets, establish feature vector sets, and run the configured data fusion strategy to comprehensively process the feature vector sets in order to obtain the fused comprehensive environmental index F_final.

[0045] The extracted feature data consists of three types: local environment feature vector, neighboring warehouse environment feature vector, and environmental business feature vector. Specifically, the local environment feature vector includes: temperature T_local and humidity H_local, the average temperature T_avg and average humidity H_avg over the past hour, the temperature change rate dT / dt, and the humidity change rate dH / dt. Therefore, the local environment feature vector F1 is represented as: F1 = [T_local, H_local, T_avg, H_avg, dT / dt, dH / dt]. It should be noted that F1... This reflects the environmental state of the target warehouse itself; the temperature T_local and humidity H_local are also calculated by averaging the temperature and humidity obtained by the distributed temperature and humidity sensor array; the average temperature T_avg and the average humidity H_avg over the past 1 hour can be selected for more than just 1 hour, and other required times can be selected according to actual needs. In this embodiment, 1 hour is used as an example, and the average of 12 equally spaced time points is taken; the temperature change rate dT / dt and the humidity change rate dH / dt need to be calculated using the central difference method, which will not be elaborated on in detail.

[0046] The neighboring warehouse environmental feature vector includes: the temperature difference ΔT_adj and humidity difference ΔH_adj between adjacent warehouses, the weighted average temperature T_wei of other warehouses within the same target warehouse, and the weighted average humidity H_wei of other warehouses within the same target warehouse. It should be noted that there are many target warehouses, such as Warehouse A, Warehouse B, etc., and within the same target warehouse, such as Warehouse A, there are also several smaller warehouses, such as Warehouse A1 and Warehouse A2, etc. Therefore, the neighboring warehouse environmental feature vector F2 is expressed as: F2 = [ΔT_adj, ΔH ... [j, T_wei, H_wei]; It should be noted that F2 considers physical proximity and ventilation coupling effect; ΔT_adj is obtained by subtracting the temperature value T_nea of ​​the nearest neighboring small warehouse from T_local, and ΔH_adj is the same, so it will not be explained in detail; T_wei is the weighted average of the temperature values ​​of other small warehouses, where the corresponding weight is inversely proportional to the distance, which refers to the distance between other small warehouses and the target small warehouse; the same applies to H_wei, so it will not be explained in detail here.

[0047] The environmental business feature vector includes: outdoor temperature T_ext, rainfall probability P_rain, and tangerine peel variety, which uses One-Hot encoding and is denoted as onehot; year; and care indicator Cfg. Therefore, the environmental business feature vector F3 is represented as: F3=[T_ext, P_rain, onehot, Year, Cfg]. The outdoor temperature T_ext, rainfall probability P_rain, and year are all obtained directly from the warehousing business data. For the tangerine peel variety, One-Hot encoding is used, resulting in: red tangerine peel as [1, 0] and green tangerine peel as [0, 1]. The year is set according to the number of years of storage, for example, if it has been stored for 3 years, then Year=3. The care indicator Cfg corresponds to whether drying and care services are required. If so, the care indicator Cfg=1, otherwise it is 0.

[0048] Therefore, the established feature vector set is a set formed by the local environment feature vector, the neighboring warehouse environment feature vector, and the environmental business feature vector;

[0049] The process of implementing a data fusion strategy based on the feature vector set is as follows:

[0050] A weighted fusion method is used to comprehensively process the feature vector set. The local environment feature vector, the neighboring warehouse environment feature vector, and the environmental business feature vector are normalized to [0, 1], and then fused into a single scalar score according to weights to obtain the fused comprehensive environmental index F_final. Regarding the weight setting in this strategy, if the current warehouse is in a closed state, the weights corresponding to the local environment feature vector, the neighboring warehouse environment feature vector, and the environmental business feature vector decrease sequentially to emphasize local stability. If an outbound operation is being performed, the weight corresponding to the neighboring warehouse environment feature vector increases to enhance the sensitivity to disturbances from neighboring warehouses. If a rainstorm is forecast, the weight corresponding to the environmental business feature vector increases. The value range of each weight is always maintained within the range of 0 to 1.

[0051] The following are examples illustrating real-world scenarios:

[0052] Scenario 1: When the warehouse is in a closed state, i.e., at night or during non-operational periods, the warehouse door is closed and no personnel or vehicles enter or exit. At this time, the primary concern is the safety and equipment operation status within the warehouse, minimizing external interference and macro-business impact. Therefore, the weights can be decreased sequentially, set as follows: the weights w1, w2, and w3 corresponding to the local environment feature vector, the neighboring warehouse environment feature vector, and the environmental business feature vector are 0.7, 0.2, and 0.1, respectively. Calculation example: Assuming the normalized local environment score is 0.9 (good status), the neighboring warehouse environment score is 0.6 (slight anomaly), and the environmental business score is 0.5, then F_final = 0.7 × 0.9 + 0.2 × 0.6 + 0.1 × 0.5 = 0.63 + 0.12 + 0.05 = 0.8. The high-weighted local features dominate the final score, effectively shielding the slight anomalies of the neighboring warehouse, ensuring that the overall solution will not generate false alarms due to irrelevant external fluctuations, and maintaining the stability of local monitoring.

[0053] Scenario 2: Outbound operations are underway, requiring enhanced sensitivity to disturbances from neighboring warehouses. During peak daytime hours, multiple forklifts traverse the warehouse area, and this warehouse shares logistics channels with neighboring warehouses. The weighting is set as follows: Outbound operations have extremely high requirements for coordination with the surrounding environment; sudden events in neighboring warehouses, such as channel congestion or equipment failure, can easily affect operations in this warehouse. Therefore, the weights of the neighboring warehouse environmental feature vectors are significantly increased. Combining this with the weights corresponding to Scenario 1, the weights are set as follows: w1, w2, and w3 are set to 0.4, 0.5, and 0.1 respectively. Assuming... The local environment score after standardization is 0.8, the neighboring warehouse environment score is 0.4 (congestion warning issued), and the environmental business score is 0.6; therefore, F_final = 0.4 × 0.8 + 0.5 × 0.4 + 0.1 × 0.6 = 0.32 + 0.20 + 0.06 = 0.58; by assigning the neighboring warehouse characteristics the highest weight, the comprehensive environmental index F_final is significantly lowered by the low scores of the neighboring warehouses, which prompts the early warning to be issued and the dispatchers to adjust the outbound rhythm, reflecting the extremely high sensitivity to nearby disturbances.

[0054] S3. The Long Short-Term Memory (LSTM) network is combined with the Random Forest Regressor (RF) to construct a hybrid temperature and humidity control model. After obtaining the LSTM predictions and RF correction terms, the model is jointly output as recommended target temperature (TL) and recommended target humidity (HL), and the model running and online update mechanism is triggered.

[0055] The Long Short-Term Memory (LSTM) network operates as follows: A 24-hour temperature and humidity time series is used as input, with 288 time steps (each 5 minutes). The input for each time step is a two-dimensional vector, which passes through an LSTM unit. Internally, the LSTM dynamically regulates the update of the cell state and hidden state through forget gates, input gates, and output gates: the forget gate determines how much historical information is retained, the input gate filters new information and writes it into the cell state, and the output gate controls the output of the current hidden state. After 288 recursive steps, the LSTM outputs the predicted temperature (T_pred) and humidity (H_pred) for the 12th time step (1 hour later). The LSTM predictions effectively capture long-term dependent patterns such as diurnal temperature variations and equipment start-up / shutdown, achieving temporal modeling of environmental trends. The Random Forest Regressor (RF) then operates... The process involves taking the comprehensive environmental index F_final as input, which includes the local environment, the impact of neighboring warehouses, and global business attributes, typically with 10 to 20 dimensions. The RF (Radio Frequency) model is composed of multiple decision trees, each independently trained based on a subset of features and a subset of samples. During prediction, F_final is fed into all decision trees, and each tree, according to feature splitting rules (e.g., if the year > 5 and P_rain > 0.8, requires additional dehumidification), outputs a correction term. The final value of all decision tree outputs is taken to obtain the temperature correction term ΔT_corr and the humidity correction term ΔH_corr. The RF correction terms are: temperature correction term ΔT_corr and humidity correction term ΔH_corr. This process, through nonlinear piecewise constant fitting, effectively compensates for the static business rules and external abrupt changes ignored by LSTM, thereby improving control accuracy. Therefore, it is necessary to combine the two to construct the required hybrid temperature and humidity control model.

[0056] The basis for determining the joint output based on the LSTM prediction and the RF correction term is as follows:

[0057] The predicted temperature value T_pred is summed with the temperature correction term ΔT_corr to generate the recommended target temperature TL.

[0058] The predicted humidity value H_pred is summed with the humidity correction term ΔH_corr to generate the recommended target humidity HL.

[0059] The basis for triggering the model execution and online update mechanism is as follows:

[0060] Supervised learning training was conducted using historical datasets, labeled with pre-defined ideal temperature and humidity ranges, such as: 18℃ to 25℃, 40%RH to 60%RH for Da Hong Pi (a type of dried tangerine peel). Cross-validation was used to divide the training and testing sets to ensure the model's generalization ability. The historical dataset could be selected from the data types of the hybrid temperature and humidity control model that have been input over the past 3 years. An online learning mechanism was also introduced, collecting actual operating results whenever a new batch of dried tangerine peel entered or left the warehouse, such as whether mold or odor changes occurred, and feeding them back to the hybrid temperature and humidity control model for incremental training to gradually optimize prediction accuracy. Details are omitted here.

[0061] S4. Run the control level rule engine on the target small warehouse, formulate the control sequence according to the control level corresponding to the target small warehouse, implement the multi-warehouse collaborative control algorithm simultaneously, and trigger multiple early warning mechanisms; the operation of the control level rule engine is based on: constructing an initial control priority score S0, calling the comprehensive environmental index F_final to complete the weighted correction action to obtain the final control priority score S, comparing S with the configured standard threshold range to obtain the control level corresponding to the target small warehouse;

[0062] The initial score S0 for regulatory priorities is constructed based on the tangerine peel value score and environmental risk score contained in F_final. First, the tangerine peel value score V_pel is set according to the following rules: if the year is ≥5, then V_pel=1; if 5 > the year is ≥3, then V_pel=0.8; if the year is <3, then V_pel=0.6. Second, the environmental risk score R_env is calculated by weighting the temperature change rate dT / dt, the temperature difference ΔT_adj between adjacent warehouses, and the rainfall probability P_rain obtained from historical statistics, and normalized to [0, 1]. Then, the tangerine peel value score V_pel and the environmental risk score R_env are weighted and summed to generate the initial score S0 for regulatory priorities. This considers both the tangerine peel value attribute and environmental risk, namely, local volatility + neighboring disturbances + external threats, thus deriving the final required initial score S0 for regulatory priorities. However, to balance the accuracy and scientific validity of this score, subsequent calls to the comprehensive environmental index F_final are needed to complete the weighted correction.

[0063] The process of calling the comprehensive environmental metric F_final to complete the weighted adjustment is as follows:

[0064] First, a correction factor σ(F_final) is introduced: representing the comprehensive environmental anomaly degree extracted from F_final. The derivation process is as follows: after normalizing the temperature change rate dT / dt, the temperature difference ΔT_adj between adjacent small cells, and the rainfall probability P_rain extracted from the comprehensive environmental index F_final, a weighted calculation is performed to obtain the correction factor σ(F_final).

[0065] It should be noted that although the correction factor σ(F_final) and the environmental risk score R_env both reflect environmental risk, their positioning and purpose are different. The environmental risk score R_env is used as a basic static risk score when calculating the initial score S0 of the control priority. It has a low update frequency and focuses on long-term stability assessment. On the other hand, the correction factor σ(F_final) is a dynamic anomaly index. The signals extracted from F_final are all instantaneous and highly sensitive signals. It is used to correct the initial score S0 of the control priority and emphasizes the risks of suddenness and coupling. Simply put, the environmental risk score R_env answers whether the warehouse is stable at normal times, while the correction factor σ(F_final) answers whether it is dangerous at this moment. The two are complementary. The former sets the baseline, and the latter adjusts the response.

[0066] Secondly, a risk amplification factor is introduced: the correction factor σ(F_final) is multiplied by the risk amplification factor, and the first result is then added to 1 to obtain the risk correction term; the risk amplification factor has a value range of [0, 1], which is used to control the correction strength, and the specific value can be set according to actual needs;

[0067] The core of introducing a risk amplification coefficient lies in giving the overall solution greater sensitivity to dynamic adjustments in response to sudden risks during operation. In actual operation, comprehensive environmental indicators reflect the risk baseline under normal conditions, but in the face of extreme weather or sudden failures, a single indicator often fails to accurately measure the destructive power of the risk. This coefficient acts as a valve for the risk amplifier. When a high-risk state is determined, increasing this coefficient can amplify the effect of the correction factor many times over, causing the risk correction item to rise rapidly and triggering an emergency circuit breaker or a high-level warning. Conversely, decreasing this coefficient during a stable period can suppress minor fluctuations and avoid frequent false alarms from the system. This design transforms the originally linear risk assessment into a flexible nonlinear response mechanism, which not only makes the calculation of the risk correction item smoother and more controllable, but also ensures that the system can both keenly detect fatal hidden dangers and remain robust in daily operation, achieving the best balance between security defense and business continuity.

[0068] Finally, the final control priority score S is defined: the initial control priority score S0 is multiplied by the risk correction term to generate the final control priority score S; the instability of the current overall environment is quantified by the correction factor σ(F_final), and the initial control priority score S0 is nonlinearly enhanced. When the system is in a high disturbance state, even if the value of dried tangerine peel is moderate, the control level should be temporarily increased to prevent cascading failures. This weighted correction action enables the control process to not only maintain value in normal times, but also to ensure safety in emergencies, significantly improving the robustness of the overall scheme operation.

[0069] By adopting the above technical solutions, the resulting comprehensive environmental indicators are not only used as inputs to the hybrid temperature and humidity control model to generate precise temperature and humidity control targets, but also used to dynamically correct the control priority scores of the warehouse area. This enables the effective reuse of multi-source information in both prediction and decision-making processes, balancing long-term value orientation and short-term risk response. For example, it balances factors such as the age of the tangerine peel, its category and temperature change rate, disturbances from neighboring warehouses, and rainstorm warnings. To a certain extent, it solves the technical problems of rigid control strategies and inability to coordinate responses to sudden environmental disturbances in traditional warehousing systems. This achieves the effects of ensuring tangerine peel storage, proactive defense in high-risk scenarios, and collaborative optimization of multi-warehouse resources, significantly improving the stability, safety, and intelligence level of the tangerine peel storage environment.

[0070] The process of comparing S with the configured standard threshold range [Q1, Q2] is as follows:

[0071] When S≥Q2, the control level corresponding to the target small warehouse is marked as high priority;

[0072] When Q2>S≥Q1, the control level corresponding to the target small warehouse is marked as medium priority;

[0073] When S < Q1, the control level corresponding to the target small warehouse is marked as low priority;

[0074] Where Q1 < Q2, and the values ​​of Q1 and Q2 are both in the range of 0 to 1. In this embodiment, the value of Q1 is usually 0.65, and the value of Q2 is 0.85. The higher the priority, the earlier the control order, so that the control resources are tilted towards high-value and high-risk warehouses, thereby improving the overall warehousing efficiency and security.

[0075] The process of implementing the multi-warehouse collaborative control algorithm is as follows:

[0076] Obtain the recommended target temperature (TL) and recommended target humidity (HL) for each target warehouse; check the control levels of adjacent warehouses. If adjacent warehouses have the same control level and there is a conflict, such as warehouse A1 needing to cool down and warehouse A2 needing to heat up, then call the game theory model to solve for the Nash equilibrium solution. The game theory model operates based on the following: the two conflicting warehouses are the participants, and their respective strategy sets are optional temperature control commands, including: strong cooling, weak cooling, and maintaining the current temperature. The payoff function comprehensively considers the value of the dried tangerine peel, the penalty cost of deviating from the target temperature and humidity, and energy consumption. By solving the Nash equilibrium of this game, a set of strategy combinations is found such that unilaterally changing the strategy of any warehouse cannot reduce its own total cost. For example, the equilibrium solution may be that warehouse A1 starts local cooling to prioritize the protection of high-aged dried tangerine peel, and warehouse A2 delays the heating up and activates temporary heat insulation curtains. Under the premise of ensuring overall storage efficiency, resources are rationally allocated to avoid equipment conflict. The final command is output and sent to the execution unit, and the execution unit includes at least: air conditioner, dehumidifier, ventilation fan, etc.

[0077] The process of triggering the multiple early warning mechanism is as follows:

[0078] The system acquires the final actual temperature and humidity of each target compartment. A Level 1 alert is triggered when the actual temperature deviates from the corresponding recommended target temperature TL±1℃ or the actual humidity deviates from the corresponding recommended target humidity HL±5%. The typical strategy is to notify the administrator via SMS or telephone voice. A Level 2 alert is triggered when the actual temperature deviates from the corresponding recommended target temperature TL±1℃ and the actual humidity deviates from the corresponding recommended target humidity HL±5% for at least three consecutive monitoring periods. The corresponding strategy is to activate backup power and emergency dehumidification to eliminate momentary interference, such as opening doors for ventilation, ensuring the alarm reflects the true trend of deviation. A Level 3 alert is triggered when mold images are detected by a pre-configured camera, requiring immediate sealing of the target compartment and alarm activation.

[0079] This solution achieves dynamic and precise control of the tangerine peel storage environment. It integrates three types of characteristic data to construct a hybrid temperature and humidity control model, generating personalized temperature and humidity targets that change with time, product category, and weather—i.e., recommended target temperature and humidity—replacing traditional fixed threshold control. Simultaneously, it introduces a corresponding mechanism for control priority scoring, enabling higher-year and higher-risk warehouses to receive more sensitive responses. This solves the energy waste or quality control problems that easily occur in traditional warehousing control, achieving on-demand adjustment and quality preservation while saving energy. The control level rule engine dynamically divides control levels and uses game theory to solve Nash equilibrium when adjacent warehouses conflict, avoiding equipment-on-device conflict. Combined with a three-level early warning system, it completes different levels of response, overcoming the shortcomings of previous intelligent warehouses operating independently and unable to perceive each other. This improves the overall robustness of the operating system and ensures the effective development of regional intelligence.

[0080] Example 2: Based on Example 1, this example also provides an intelligent control system for the storage environment of dried tangerine peel based on multi-source data fusion. The system includes: a heterogeneous data construction module: deploying an environmental sensor network in a selected storage area, accessing a meteorological data interface, and integrating storage business data to construct a multi-source heterogeneous dataset.

[0081] Environmental indicator detection module: Under the condition of dealing with different target small warehouses in the selected storage area, it extracts three types of feature data based on multi-source heterogeneous datasets, establishes feature vector sets, and runs the configured data fusion strategy to comprehensively process the feature vector sets to obtain the fused comprehensive environmental indicators.

[0082] The regulation model building module combines a Long Short-Term Memory (LSTM) network with a Random Forest Regressor (RF) to construct a hybrid temperature and humidity regulation model. After obtaining the LSTM predictions and RF correction terms, it jointly outputs the recommended target temperature (TL) and recommended target humidity (HL) and triggers the model running and online update mechanism.

[0083] Collaborative Control and Early Warning Module: The control level rule engine runs on the target small warehouse, formulates the control sequence according to the control level corresponding to the target small warehouse, implements multi-warehouse collaborative control algorithm simultaneously, and triggers multiple early warning mechanisms. The operation of the control level rule engine is based on: constructing an initial control priority score, calling comprehensive environmental indicators to complete the weighted correction action to obtain the final control priority score, and comparing the final control priority score with the configured standard threshold range to obtain the control level corresponding to the target small warehouse.

[0084] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0085] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent control of the storage environment of dried tangerine peel based on multi-source data fusion, characterized in that, The method includes: An environmental sensor network is deployed within the selected warehousing area, connected to a meteorological data interface, and integrated with warehousing business data to construct a multi-source heterogeneous dataset. Under the condition of dealing with different target small warehouses in the selected storage area, three types of feature data are extracted based on multi-source heterogeneous datasets to establish a feature vector set, and the configured data fusion strategy is run to comprehensively process the feature vector set in order to obtain the fused comprehensive environmental index. A hybrid temperature and humidity control model is constructed by combining a Long Short-Term Memory (LSTM) network with a Random Forest Regressor (RF). After obtaining the LSTM predictions and RF correction terms, the model jointly outputs the recommended target temperature (TL) and the recommended target humidity (HL), and triggers the model running and online update mechanism. The control level rule engine runs on the target small warehouse, and the control sequence is determined according to the control level corresponding to the target small warehouse. The multi-warehouse collaborative control algorithm is implemented simultaneously, and multiple early warning mechanisms are triggered. The operation of the control level rule engine is based on: constructing an initial control priority score, calling comprehensive environmental indicators to complete the weighted correction action to obtain the final control priority score, and comparing the final control priority score with the configured standard threshold range to obtain the control level corresponding to the target small warehouse.

2. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 1, characterized in that, The deployed environmental sensing network is a distributed array of temperature and humidity sensors. Accessing the meteorological data interface means connecting to a third-party weather service via API to obtain real-time meteorological data for the selected storage area, including at least: outdoor temperature, relative humidity, probability of rainfall, and wind speed; warehousing business data includes at least: inbound information, storage period, transaction voucher number, and pesticide residue test report status; among which, the inbound information includes the variety, year, origin, weight, and whether drying and care services are required for dried tangerine peel; the constructed multi-source heterogeneous dataset is: using temperature and humidity data, meteorological data, and warehousing business data obtained from the environmental sensor network as raw data, and storing them after data cleaning.

3. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 1, characterized in that, The three types of extracted feature data are: local environment feature vector, neighboring warehouse environment feature vector, and environmental business feature vector. Among them, the local environment feature vector F1 includes at least: temperature T_local and humidity H_local, the average temperature T_avg and the average humidity H_avg of the past 1 hour, the temperature change rate dT / dt, and the humidity change rate dH / dt. The neighboring warehouse environment feature vector F2 includes at least: the temperature difference ΔT_adj and the humidity difference ΔH_adj between adjacent warehouses, the weighted average temperature T_wei of other warehouses in the same target warehouse, and the weighted average humidity H_wei of other warehouses in the same target warehouse. The environmental business feature vector F3 includes at least: outdoor temperature T_ext, rainfall probability P_rain, tangerine peel variety onehott, year, and care mark Cfg.

4. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 3, characterized in that, The process of running the data fusion strategy based on the feature vector set is as follows: the local environment feature vector, the neighboring warehouse environment feature vector, and the environmental business feature vector are normalized to [0, 1] respectively, and then fused into a single scalar score according to the weights to obtain the fused comprehensive environmental index F_final.

5. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 1, characterized in that, The operation of the Long Short-Term Memory (LSTM) network is as follows: it takes a 24-hour temperature and humidity time series as input and outputs LSTM prediction values, namely the temperature prediction value T_pred and the humidity prediction value H_pred for the next hour. The operation of the Random Forest Regressor (RF) is as follows: it takes the comprehensive environmental index F_final as input and outputs RF correction terms, namely the temperature correction term ΔT_corr and the humidity correction term ΔH_corr.

6. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 5, characterized in that, The basis for determining the joint output based on the LSTM prediction and the RF correction term is as follows: The predicted temperature value T_pred is summed with the temperature correction term ΔT_corr to generate the recommended target temperature TL. The predicted humidity value H_pred is summed with the humidity correction term ΔH_corr to generate the recommended target humidity HL. The basis for triggering the model execution and online update mechanism is as follows: Supervised learning training was conducted using historical datasets, with labels set as pre-defined ideal temperature and humidity ranges. Cross-validation was used to divide the training and testing sets, and an online learning mechanism was introduced. Whenever a new batch of dried tangerine peel entered or left the warehouse, the actual operating results were collected and fed back to the hybrid temperature and humidity control model for incremental training.

7. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 6, characterized in that, The initial score S0 for regulatory priority is constructed based on the tangerine peel value score and environmental risk score contained in F_final. The tangerine peel value score V_pel is set according to the following rules: when the year is ≥5, V_pel=1; when 5>year ≥3, V_pel=0.8; when the year is <3, V_pel=0.

6. The environmental risk score R_env is calculated by weighting the temperature change rate dT / dt, the temperature difference ΔT_adj between adjacent warehouses, and the rainfall probability P_rain obtained from historical statistics, and normalized to [0,1]. The tangerine peel value score V_pel and the environmental risk score R_env are weighted and summed to generate the initial score S0 for regulatory priority. The process of calling F_final to complete the weighted correction is as follows: The correction factor σ(F_final) is introduced: the temperature change rate dT / dt, the temperature difference ΔT_adj between adjacent small cells, and the rainfall probability P_rain extracted from the comprehensive environmental index F_final are normalized and then weighted to obtain the correction factor σ(F_final); the risk amplification coefficient is introduced: the correction factor σ(F_final) is multiplied by the risk amplification coefficient, and the first result is then added to 1 to obtain the risk correction term; Define the final control priority score S: multiply the initial control priority score S0 by the risk correction term to generate the final control priority score S.

8. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 7, characterized in that, The process of comparing the final control priority score S with the configured standard threshold range [Q1, Q2] is as follows: When S≥Q2, the control level corresponding to the target small warehouse is marked as high priority; When Q2>S≥Q1, the control level corresponding to the target small warehouse is marked as medium priority; When S < Q1, the control level corresponding to the target small warehouse is marked as low priority; Where Q1 < Q2, and the values ​​of Q1 and Q2 are both in the range of 0 to 1.

9. The intelligent control method for the storage environment of dried tangerine peel based on multi-source data fusion according to claim 6, characterized in that, The process of implementing the multi-warehouse collaborative control algorithm is as follows: obtain the recommended target temperature TL and recommended target humidity HL for each target warehouse; check the control levels of adjacent warehouses; if the control levels of adjacent warehouses are the same and there is a conflict, call the game theory model to solve the Nash equilibrium solution, output the final instruction and send it to the execution unit. The process of triggering the multiple early warning mechanism is as follows: obtain the final actual temperature and actual humidity of each target compartment. When the actual temperature deviates from the corresponding recommended target temperature TL±1℃ or the actual humidity deviates from the corresponding recommended target humidity HL±5%, a first-level early warning is triggered. A Level II warning is triggered when the actual temperature deviates from the corresponding recommended target temperature TL±1℃ and the actual humidity deviates from the corresponding recommended target humidity HL±5% for at least three consecutive monitoring cycles; a Level III warning is triggered when a mold image is detected by a pre-configured camera.

10. An intelligent control system for the storage environment of dried tangerine peel based on multi-source data fusion, characterized in that, The system includes: Heterogeneous data construction module: Deploy an environmental sensor network within the selected warehousing area, connect to the meteorological data interface, and integrate warehousing business data to construct a multi-source heterogeneous dataset; Environmental indicator detection module: Under the condition of dealing with different target small warehouses in the selected storage area, it extracts three types of feature data based on multi-source heterogeneous datasets, establishes feature vector sets, and runs the configured data fusion strategy to comprehensively process the feature vector sets to obtain the fused comprehensive environmental indicators. The regulation model building module combines a Long Short-Term Memory (LSTM) network with a Random Forest Regressor (RF) to construct a hybrid temperature and humidity regulation model. After obtaining the LSTM predictions and RF correction terms, it jointly outputs the recommended target temperature (TL) and recommended target humidity (HL) and triggers the model running and online update mechanism. Collaborative Control and Early Warning Module: The control level rule engine runs on the target small warehouse, formulates the control sequence according to the control level corresponding to the target small warehouse, implements multi-warehouse collaborative control algorithm simultaneously, and triggers multiple early warning mechanisms. The operation of the control level rule engine is based on: constructing an initial control priority score, calling comprehensive environmental indicators to complete the weighted correction action to obtain the final control priority score, and comparing the final control priority score with the configured standard threshold range to obtain the control level corresponding to the target small warehouse.

Images