Defrosting control method based on near-end strategy optimization and refrigerator

Abstract

The invention provides a defrosting control method based on near-end strategy optimization and a refrigerator. The method comprises the steps that target data of the refrigerator are collected; processing the collected target data, judging whether missing values and abnormal values exist in the data or not, and respectively processing the missing values and the abnormal values; constructing a state space of the multivariable strategy network based on the preprocessed data; meanwhile, energy consumption, freezing chamber temperature fluctuation, compressor starting and stopping times, defrosting interval time and defrosting duration are calculated based on the preprocessed data, and a multi-target reward function is constructed; training the constructed strategy network by using a near-end strategy optimization algorithm; and the trained lightweight strategy network model is embedded into an equipment end control device, and a defrosting control scheme is output. The near-end strategy optimization method of deep reinforcement learning is applied to refrigerator defrosting decision making, the defrosting opportunity is dynamically decided through the deep reinforcement learning strategy network, comprehensive optimization of energy consumption and temperature stability is achieved, energy consumption is reduced, temperature fluctuation is reduced, and the service life of a compressor is prolonged.

Description

Technical Field

[0001] This application relates to the field of refrigerator technology, and in particular to a defrosting control method and refrigerator based on near-end strategy optimization. Background Technology

[0002] As an indispensable refrigeration and preservation device in daily life, the refrigerator's core refrigeration component, the evaporator, condenses and forms a frost layer on its surface due to the temperature difference between the refrigerant and the air. As the frost layer thickens, the evaporator's heat exchange efficiency significantly decreases, leading to reduced cooling capacity, increased energy consumption, and potentially causing temperature fluctuations in the compartments, affecting food preservation. Therefore, timely and effective defrosting control of the evaporator is crucial for ensuring stable refrigerator operation and improving energy efficiency and preservation performance.

[0003] Refrigerator defrosting control technology has undergone several generations of development and can be roughly divided into three stages: early fixed-cycle defrosting technology, which forcibly starts the defrosting program according to a preset fixed time interval. This method does not take into account the actual frost situation and is prone to insufficient or excessive defrosting; the subsequent development of simple threshold-based defrosting technology, which starts defrosting when the monitored value reaches a single threshold such as evaporator temperature and frost thickness, optimizes the timing of defrosting to some extent, but is still difficult to adapt to complex operating conditions; the current mainstream defrosting control technology is fuzzy control defrosting, which builds fuzzy logic rules based on expert experience and combines multiple parameters such as ambient temperature and running time to make defrosting decisions, which has a certain improvement in adaptability compared to the first two types of technology.

[0004] However, existing mainstream fuzzy logic defrosting technology still has significant drawbacks: First, fuzzy logic rules rely on expert experience and are difficult to cover all actual operating conditions, especially in complex scenarios such as high temperature and humidity environments, frequent door opening and closing by users, and drastic changes in the refrigerator's load. It cannot adaptively adjust the defrosting cycle based on real-time dynamic factors, which can easily lead to unnecessary increases in energy consumption and excessive fluctuations in compartment temperature. Second, the compressor start-up and shutdown process is highly coupled with the defrosting process, and frequent start-up and shutdown operations will significantly reduce the compressor's lifespan. Third, there is a lack of a quantitative assessment mechanism for cold loss during the defrosting process, making it difficult to achieve the optimal balance between energy-saving goals and food preservation needs, thus affecting the overall performance of the refrigerator. Summary of the Invention

[0005] This application provides a defrosting control method and refrigerator based on near-end strategy optimization to solve the problem that the existing fuzzy control defrosting technology lacks a quantitative assessment mechanism for cold loss during the defrosting process, making it difficult to achieve the optimal balance between energy-saving goals and food preservation needs, thus affecting the overall performance of the refrigerator.

[0006] In a first aspect, this application provides a defrosting control method based on proximal strategy optimization, the method comprising:

[0007] Collect target data from the refrigerator, including temperature and device component status, and upload it to a big data platform for storage via Internet of Things (IoT) communication protocol.

[0008] The collected target data is processed to determine whether there are missing or outlier values, and then processed accordingly.

[0009] The state space of a multivariate policy network is constructed based on the preprocessed data. The state space includes ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, and heater status. At the same time, energy consumption, freezer temperature fluctuation, compressor start-stop times, defrost interval time, and defrost duration are calculated based on the preprocessed data, and a multi-objective reward function is constructed.

[0010] The constructed policy network is trained using a proximal policy optimization algorithm;

[0011] The trained lightweight policy network model is embedded into the device-side control unit, and a defrosting control scheme is output.

[0012] In some possible implementations, constructing the state space of a multivariate policy network based on preprocessed data includes:

[0013] The preprocessed target data is used to extract 12 feature parameters, including ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, heater status, indoor temperature change, indoor humidity change, freezer set temperature, and door open / close status.

[0014] The feature parameters are standardized according to their minimum and maximum values ​​in historical data, and the parameter values ​​are mapped to the [0,1] interval.

[0015] The 12 standardized feature parameters are arranged in the order of environmental parameters, temperature parameters, and operating status parameters to form a state vector S_t.

[0016] Verify the acquisition timestamps of all feature parameters in the state vector S_t to ensure that the timestamp deviation does not exceed 10 seconds. If there are parameters with timeout deviations, replace them with the latest preprocessed data.

[0017] The time-validated state vector S_t is input into the input layer of the policy network. This input layer has 12 nodes, which correspond one-to-one with the dimensions of the state space.

[0018] In some possible implementations, missing values ​​are handled as follows: if the missing value is a temperature value, it is replaced with the average of the data within the previous 5 minutes; if the missing value is status data, it is replaced with the previous data.

[0019] In some possible implementations, outliers are handled as follows: using the 3x standard deviation method, the mean u and standard deviation σ of the daily sampled data are calculated according to the feature dimension, and data distributed in the intervals [-∞, u-3σ] and [u+3σ, +∞] are removed.

[0020] In some possible implementations, the reward function R = -[w1·ΔE + w2·σ_T1 + w3·frost_time + w4·N_start], where ΔE is the energy consumption increment between two adjacent defrost cycles, σ_T1 is the standard deviation of the freezer compartment temperature between two adjacent defrost cycles, frost_time is the defrost interval duration with a maximum value of 80 hours, and N_start is the number of times the compressor starts and stops.

[0021] In some possible implementations, energy consumption E is obtained using the area enclosed by the compressor frequency and time, specifically calculated...

[0022] The method is as follows:

[0023]

[0024] Where k is the compressor coefficient of performance, and f(t) is the frequency value of the compressor as a function of time.

[0025] In some possible implementations, training the constructed policy network using a proximal policy optimization algorithm includes:

[0026] After each training round, a model generalization validation step is added: three sets of historical operating data from different regions are selected as validation sets. If the model's decision accuracy on the validation set is less than 85%, 500 more training rounds are automatically added and the learning rate is adjusted to 80% of the original initial learning rate. The different regions include high-temperature and high-humidity areas, low-temperature and dry areas, and temperate monsoon areas.

[0027] In some possible implementations, the processing of the collected target data also includes dynamically weighting and correcting the collected environmental humidity data. The correction coefficient is determined based on the deviation between the atmospheric pressure of the day and the historical average pressure. The larger the deviation, the larger the absolute value of the correction coefficient.

[0028] In some possible implementations, the model is trained in the cloud using a GPU cluster, with 2000 rounds of training, each round consisting of 2048 steps. The training is terminated if the cumulative running time of the compressor in each round is greater than the interval between two adjacent defrost cycles. The batch size is set to 128, the initial learning rate is set to 3e-4, and the Adam optimizer is used.

[0029] Secondly, this application provides a refrigerator, which includes the defrosting control method based on proximal strategy optimization described in the first aspect.

[0030] As described above, this application provides a defrosting control method and a refrigerator based on proximal strategy optimization. The method includes: collecting target data from the refrigerator; processing the collected target data to determine if there are missing or outlier values, and processing them accordingly; constructing the state space of a multivariate strategy network based on the preprocessed data; simultaneously calculating energy consumption, freezer temperature fluctuation, compressor start / stop times, defrosting interval, and defrosting duration based on the preprocessed data, and constructing a multi-objective reward function; training the constructed strategy network using a proximal strategy optimization algorithm; embedding the trained lightweight strategy network model into the device-side control unit; and outputting the defrosting control scheme. This application applies a deep reinforcement learning-based proximal strategy optimization method to refrigerator defrosting decisions. By dynamically deciding the defrosting timing through a deep reinforcement learning strategy network, it achieves comprehensive optimization of energy consumption and temperature stability. Compared to traditional fuzzy control defrosting methods, this application can reduce energy consumption, decrease temperature fluctuations, and extend compressor life. Attached Figure Description

[0031] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A schematic diagram of the defrosting control method framework based on proximal strategy optimization provided in this application;

[0033] Figure 2 A schematic diagram of the strategy network structure provided in this application;

[0034] Figure 3 A schematic diagram of the model deployment provided in this application. Detailed Implementation

[0035] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0036] As an indispensable refrigeration and preservation device in daily life, the refrigerator's core refrigeration component, the evaporator, condenses and forms a frost layer on its surface due to the temperature difference between the refrigerant and the air. As the frost layer thickens, the evaporator's heat exchange efficiency significantly decreases, leading to reduced cooling capacity, increased energy consumption, and potentially causing temperature fluctuations in the compartments, affecting food preservation. Therefore, timely and effective defrosting control of the evaporator is crucial for ensuring stable refrigerator operation and improving energy efficiency and preservation performance.

[0037] Refrigerator defrosting control technology has undergone several generations of development and can be roughly divided into three stages: early fixed-cycle defrosting technology, which forcibly starts the defrosting program according to a preset fixed time interval. This method does not take into account the actual frost situation and is prone to insufficient or excessive defrosting; the subsequent development of simple threshold-based defrosting technology, which starts defrosting when the monitored value reaches a single threshold such as evaporator temperature and frost thickness, optimizes the timing of defrosting to some extent, but is still difficult to adapt to complex operating conditions; the current mainstream defrosting control technology is fuzzy control defrosting, which builds fuzzy logic rules based on expert experience and combines multiple parameters such as ambient temperature and running time to make defrosting decisions, which has a certain improvement in adaptability compared to the first two types of technology.

[0038] However, existing mainstream fuzzy logic defrosting technology still has significant drawbacks: First, fuzzy logic rules rely on expert experience and are difficult to cover all actual operating conditions, especially in complex scenarios such as high temperature and humidity environments, frequent door opening and closing by users, and drastic changes in the refrigerator's load. It cannot adaptively adjust the defrosting cycle based on real-time dynamic factors, which can easily lead to unnecessary increases in energy consumption and excessive fluctuations in compartment temperature. Second, the compressor start-up and shutdown process is highly coupled with the defrosting process, and frequent start-up and shutdown operations will significantly reduce the compressor's lifespan. Third, there is a lack of a quantitative assessment mechanism for cold loss during the defrosting process, making it difficult to achieve the optimal balance between energy-saving goals and food preservation needs, thus affecting the overall performance of the refrigerator.

[0039] Based on this, Figure 1 As shown, this application provides a defrosting control method based on proximal strategy optimization, the method comprising:

[0040] Collect target data from the refrigerator, including temperature and device component status, and upload it to a big data platform for storage via Internet of Things (IoT) communication protocol.

[0041] The collected target data is processed to determine whether there are missing or outlier values, and then processed accordingly.

[0042] The state space of a multivariate policy network is constructed based on the preprocessed data. The state space includes ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, and heater status. At the same time, energy consumption, freezer temperature fluctuation, compressor start-stop times, defrost interval time, and defrost duration are calculated based on the preprocessed data, and a multi-objective reward function is constructed.

[0043] The constructed policy network is trained using a proximal policy optimization algorithm;

[0044] The trained lightweight policy network model is embedded into the device-side control unit, and a defrosting control scheme is output.

[0045] By collecting target data such as temperature and equipment component status and uploading it to a big data platform for storage, comprehensive basic data support is provided for model training. Targeted processing of missing and outlier values ​​in the data can effectively prevent low-quality data from interfering with model training, ensuring the accuracy and completeness of the input data and laying a reliable foundation for the subsequent construction and training of the policy network.

[0046] Based on preprocessed data, a state space is constructed including multi-dimensional parameters such as ambient temperature, humidity, and evaporator temperature, comprehensively reflecting the real-time operating conditions of the refrigerator. Simultaneously, a multi-objective reward function is built by combining indicators such as energy consumption and temperature fluctuations. This allows the near-end policy optimization algorithm to comprehensively balance multiple objectives such as energy saving, preservation, and equipment lifespan during training. The resulting policy network can accurately determine the defrosting timing, achieving dynamic and adaptive defrosting control. Using the near-end policy optimization algorithm to train the policy network ensures the stability of the training process due to its "near-end constraint" characteristic, avoiding excessive fluctuations in policy updates and improving model convergence. Embedding the trained lightweight model into the device's control unit enables edge deployment and real-time decision-making, ensuring the defrosting control scheme quickly responds to actual operational needs and reduces latency.

[0047] The defrosting control scheme formed by the above steps can reduce ineffective heating and dynamically adjust the defrosting cycle, thereby reducing refrigerator energy consumption; at the same time, it can reduce compartment temperature fluctuations and improve food preservation; in addition, reasonable defrosting decisions can reduce the probability of frequent compressor start-stop, extend equipment lifespan, and comprehensively improve user experience.

[0048] In some embodiments, constructing the state space of a multivariable policy network based on preprocessed data includes:

[0049] The preprocessed target data is used to extract 12 feature parameters, including ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, heater status, indoor temperature change, indoor humidity change, freezer set temperature, and door open / close status.

[0050] The feature parameters are standardized according to their minimum and maximum values ​​in historical data, and the parameter values ​​are mapped to the [0,1] interval.

[0051] The 12 standardized feature parameters are arranged in the order of environmental parameters, temperature parameters, and operating status parameters to form a state vector S_t.

[0052] Verify the acquisition timestamps of all feature parameters in the state vector S_t to ensure that the timestamp deviation does not exceed 10 seconds. If there are parameters with timeout deviations, replace them with the latest preprocessed data.

[0053] The time-validated state vector S_t is input into the input layer of the policy network. This input layer has 12 nodes, which correspond one-to-one with the dimensions of the state space.

[0054] By extracting 12 feature parameters, covering environmental parameters (ambient temperature, humidity and their changes), temperature parameters (evaporator temperature, freezer temperature and set temperature), and operating status parameters (equipment component status, door opening and closing status, etc.), the system can comprehensively and accurately reflect the real-time operating conditions of the refrigerator, providing rich and crucial input information for the strategy network and avoiding decision-making biases caused by missing features.

[0055] Standardizing the feature parameters and mapping parameters of different dimensions to the [0,1] interval can avoid the impact of large differences in parameter values ​​on the training of the policy network, so that each feature parameter has equal weight reference value in model training, thereby improving the network learning efficiency and the stability of parameter optimization.

[0056] The feature parameters are arranged in the order of environmental parameters, temperature parameters, and operating status parameters to form a state vector. This ensures that the vector structure is consistent with the logical relationship of the refrigerator's operation, making it easier for the policy network to quickly identify the intrinsic relationship between the parameters. At the same time, by verifying the timestamp, it is ensured that the timeliness deviation of all feature parameters is within 10 seconds, which can guarantee that the state vector can truly reflect the refrigerator's operating status at the same moment and avoid misjudgment of the status due to data asynchrony.

[0057] The number of input layer nodes corresponds one-to-one with the dimension of the state space, ensuring that the processed state vector can be completely and accurately input into the policy network. This provides a standardized and reliable input data foundation for the training of the near-end policy optimization algorithm, thereby improving the stability of model training and the accuracy of the final defrosting decision, and helping to achieve the goals of reducing energy consumption and temperature fluctuations.

[0058] In some embodiments, missing values ​​are handled as follows: if the missing value is a temperature value, it is replaced by the average value of the data within the previous 5 minutes; if the missing value is status data, it is replaced by the previous data.

[0059] For missing temperature values, the average value of the previous 5 minutes is used to replace them. This can reasonably fill the gaps based on recent temperature change trends and avoid single-point missing values ​​from disrupting the continuity of temperature time series data. For status data (such as valve status, fan status, etc.), the previous data is used to replace them. This can maintain the stability of the equipment's operating status and prevent data distortion caused by assumptions of sudden changes in status, thereby ensuring that the pre-processed data can fully reflect the refrigerator's operating process.

[0060] Temperature values ​​and equipment status data are the core inputs for constructing the state space of the policy network. Targeted missing value handling can reduce the interference of poor-quality data on model training. The averaging replacement method can balance the volatility of temperature data, while the previous value replacement method can maintain the logic of the state data, making the processed data more consistent with the actual operating scenario and providing a reliable training basis for the policy network.

[0061] Complete and reliable input data is the foundation for accurate decision-making in near-end policy optimization algorithms. By handling missing values ​​in the manner described above, it is ensured that the temperature parameters and equipment state parameters in the state space are true and valid. This allows the policy network to accurately learn the defrosting patterns under different operating conditions during training, thereby improving the accuracy of the final defrosting control decision and helping to achieve the goals of reducing energy consumption and temperature fluctuations.

[0062] In some embodiments, outliers are handled as follows: using the 3x standard deviation method, the mean u and standard deviation σ of the daily sampled data are calculated according to the feature dimension, and data distributed in the intervals [-∞, u-3σ] and [u+3σ, +∞] are removed.

[0063] In some embodiments, the reward function R = -[w1·ΔE + w2·σ_T1 + w3·frost_time + w4·N_start], where ΔE is the energy consumption increment between two adjacent defrost cycles, σ_T1 is the standard deviation of the freezer compartment temperature between two adjacent defrost cycles, frost_time is the defrost interval duration with a maximum value of 80 hours, and N_start is the number of times the compressor starts and stops.

[0064] In some embodiments, energy consumption E is obtained using the area enclosed by the compressor frequency and time, and the specific calculation method is as follows:

[0065]

[0066] Where k is the compressor coefficient of performance, and f(t) is the frequency value of the compressor as a function of time.

[0067] In some embodiments, training the constructed policy network using a proximal policy optimization algorithm includes:

[0068] After each training round, a model generalization validation step is added: three sets of historical operating data from different regions are selected as validation sets. If the model's decision accuracy on the validation set is less than 85%, 500 more training rounds are automatically added and the learning rate is adjusted to 80% of the original initial learning rate. The different regions include high-temperature and high-humidity areas, low-temperature and dry areas, and temperate monsoon areas.

[0069] By selecting historical operating data from three different regions—high temperature and high humidity, low temperature and dryness, and temperate monsoon—as validation sets, the model's decision-making performance under diverse climatic conditions can be effectively tested. This avoids the model being only adapted to specific operating conditions in a single region, enabling it to maintain stable performance in refrigerator operating scenarios across different climatic regions.

[0070] When the model's decision accuracy on the validation set is below 85%, it automatically adds 500 training rounds and adjusts the learning rate to 80% of the original initial learning rate. This dynamic adjustment mechanism enables the model to fully learn the working conditions of different regions during training, reduces the risk of overfitting, and ensures that the model can still make accurate defrosting decisions when facing complex and ever-changing environments in real-world applications.

[0071] By enhancing the model's adaptability and generalization in different regional scenarios, misjudgments of defrosting timing due to regional environmental differences can be reduced, thereby achieving stable reductions in energy consumption and temperature fluctuations in various scenarios, extending compressor lifespan, and improving user experience.

[0072] In some embodiments, the processing of the collected target data also includes dynamically weighting and correcting the collected environmental humidity data. The correction coefficient is determined based on the deviation between the atmospheric pressure of the day and the historical average pressure. The larger the deviation, the larger the absolute value of the correction coefficient.

[0073] Changes in atmospheric pressure affect the air's water-holding capacity, leading to discrepancies between the actual humidity levels reflected by the same humidity value under different pressures. By determining a correction coefficient based on the deviation between the current day's atmospheric pressure and historical average pressure, and dynamically weighting the collected environmental humidity data, the interference of pressure changes on humidity measurements can be offset, making the humidity data more consistent with actual environmental conditions.

[0074] Ambient humidity is one of the key factors affecting the frosting speed of a refrigerator evaporator. Accurate humidity data can provide more reliable input for the policy network. Corrected humidity data can make the state space more realistically reflect the environmental conditions of the refrigerator, helping the near-end policy optimization algorithm to more accurately judge the frosting trend. This allows the defrosting timing decision to be more in line with actual needs, reducing the problem of premature or late defrosting caused by distorted humidity data.

[0075] Atmospheric pressure changes dynamically at different times and under different weather conditions. This correction mechanism ensures that humidity data always matches the current pressure state, ensuring that the defrosting control strategy can maintain good adaptability under complex and changeable weather conditions, thereby stably achieving the effects of reducing energy consumption and reducing temperature fluctuations.

[0076] In some embodiments, the model training is performed in the cloud using a GPU cluster, with 2000 rounds of training, each round consisting of 2048 steps. The training is terminated if the cumulative running time of the compressor in each round is greater than the interval between two adjacent defrost cycles. The batch size is set to 128, the initial learning rate is set to 3e-4, and the Adam optimizer is used.

[0077] Training using cloud-based GPU clusters leverages their powerful parallel computing capabilities to efficiently process large amounts of refrigerator operation data, shortening the model training cycle. Setting a training scale of 2000 rounds with 2048 steps per round ensures that the model fully learns the defrosting patterns under different operating conditions. The mechanism of "terminating if the cumulative compressor running time exceeds the interval between two adjacent defrosting times" in each round avoids invalid training steps and improves training efficiency.

[0078] Setting the batch size to 128 ensures training stability while balancing the accuracy of gradient estimation with computational resource consumption. The initial learning rate of 3e-4, combined with the Adam optimizer, enables the model to iterate and update rapidly in the early stages of training. Furthermore, the adaptive learning rate adjustment feature of the Adam optimizer allows for fine-tuning of parameters in the later stages of training, promoting stable convergence of the model to a better state.

[0079] By setting the above training parameters and combining the near-end constraint characteristics of the near-end policy optimization algorithm, the policy network can be stably updated during training, reducing training fluctuations and enabling the final trained model to have a more reliable defrosting decision-making ability. This effectively reduces energy consumption, temperature fluctuations, and extends compressor life in practical applications.

[0080] In some embodiments, this application also provides a refrigerator, which includes the defrosting control method based on proximal strategy optimization described in the above embodiments.

[0081] This application can accurately determine the frost status of the refrigerator, with good robustness and stability. It optimizes frost control through multi-dimensional sensing data and near-end optimization strategy algorithms, reduces ineffective heating, and dynamically adjusts the defrosting cycle, thereby reducing energy consumption and temperature fluctuations.

[0082] This application discloses a refrigerator defrosting control method based on proximal policy optimization. It is the first to apply the proximal policy optimization method of deep reinforcement learning to the refrigerator defrosting decision. By dynamically deciding the defrosting timing through a deep reinforcement learning policy network, the method achieves comprehensive optimization of energy consumption and temperature stability. Compared with the traditional fuzzy control defrosting method, this invention can reduce energy consumption, reduce temperature fluctuations, extend compressor life and improve user experience.

[0083] Example

[0084] This embodiment presents a defrosting control method based on near-end strategy optimization, as detailed below:

[0085] 1. Data preparation: Collect various data during the refrigerator's operation in real time, including temperature, humidity, compressor speed, fan status, etc., and transmit them to a big data platform for storage via the network;

[0086] 2. Data preprocessing: Determine whether the acquired data contains missing or outlier values ​​and process them accordingly;

[0087] 2.1 Regarding the missing values ​​mentioned in 2 above, if the missing value is a temperature value, this patent uses the average value to replace it, and uses the average value of the data within the previous 5 minutes to replace it; if the missing value is status data, the previous data is used for replacement.

[0088] 2.2 Regarding the outliers mentioned in section 2 above, this patent uses the 3x standard deviation method to remove outliers. The daily sampled data are processed according to their feature dimensions, and the mean u and standard deviation σ are calculated separately. Data distributed in the intervals [-∞, u-3σ] and [u+3σ, +∞] are removed.

[0089] 3. State Space: Construct a state space St from the preprocessed observation data, as shown above. Figure 2 As shown, it includes 12 dimensions such as ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, and heater status;

[0090] 4. Action Space: Construct action space A and design action vector a_t, a_t = {0: no frost, 1: frost}.

[0091] 5. Reward Function: Design a reward function R, R_t=-[w1·ΔE+w2·σ_T1+w3·frost_time+w4·N_start], where ΔE is the energy consumption increment between two adjacent defrost cycles, σ_T1 is the standard deviation of the freezer compartment temperature between two adjacent defrost cycles, frost_time is the defrost interval with a maximum value of 80 hours, N_start is the number of compressor start-stop cycles, and the weights w1~w4 are determined through the experience of refrigeration experts;

[0092] The energy consumption E in point 5 above is obtained using the area enclosed by the compressor frequency and time, and the specific calculation method is shown below:

[0093]

[0094] Where k is the compressor coefficient of performance, and f(t) is the frequency value of the compressor as a function of time;

[0095] 6. Strategy network design, as above Figure 2 The network structure of the policy proposed in this paper is shown. It adopts a 3-layer neural network structure of 12-128-2 and uses ReLU as the activation function.

[0096] 7. Model training: The model is trained on a cloud-based GPU cluster for 2000 rounds, with 2048 steps per round. Training is terminated if the cumulative compressor runtime in each round exceeds the interval between two consecutive defrost cycles. The batch size is set to 128, the initial learning rate is set to 3e-4, and the Adam optimizer is used. Figure 3 As shown, the trained policy network model is converted and quantized, and then ported to the device-side MCU for real-time inference and decision-making.

[0097] As can be seen from the above embodiments, this application provides a defrosting control method and a refrigerator based on proximal strategy optimization. The method includes: collecting target data from the refrigerator; processing the collected target data to determine whether there are missing or outlier values, and processing them accordingly; constructing the state space of a multivariate strategy network based on the preprocessed data; simultaneously calculating energy consumption, freezer temperature fluctuation, compressor start / stop times, defrosting interval, and defrosting duration based on the preprocessed data, and constructing a multi-objective reward function; training the constructed strategy network using a proximal strategy optimization algorithm; embedding the trained lightweight strategy network model into the device-side control unit; and outputting the defrosting control scheme. This application applies a deep reinforcement learning-based proximal strategy optimization method to refrigerator defrosting decisions. By dynamically deciding the defrosting timing through a deep reinforcement learning strategy network, it achieves comprehensive optimization of energy consumption and temperature stability. Compared with traditional fuzzy control defrosting methods, this application can reduce energy consumption, reduce temperature fluctuations, and extend compressor life.

[0098] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. A defrosting control method based on proximal strategy optimization, characterized in that, The method comprises: Collect target data from the refrigerator, including temperature and device component status, and upload it to a big data platform for storage via Internet of Things (IoT) communication protocol. The collected target data is processed to determine whether there are missing or outlier values, and then processed accordingly. The state space of a multivariate policy network is constructed based on the preprocessed data. The state space includes ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, and heater status. At the same time, energy consumption, freezer temperature fluctuation, compressor start-stop times, defrost interval time, and defrost duration are calculated based on the preprocessed data, and a multi-objective reward function is constructed. The constructed policy network is trained using a proximal policy optimization algorithm; The trained lightweight policy network model is embedded into the device-side control unit, and a defrosting control scheme is output.

2. The defrosting control method based on proximal strategy optimization according to claim 1, characterized in that, The state space of a multivariate policy network constructed based on preprocessed data includes: The preprocessed target data is used to extract 12 feature parameters, including ambient temperature, humidity, evaporator temperature, freezer temperature, compressor speed, valve status, fan status, heater status, indoor temperature change, indoor humidity change, freezer set temperature, and door open / close status. The feature parameters are standardized according to their minimum and maximum values ​​in historical data, and the parameter values ​​are mapped to the [0,1] interval. The 12 standardized feature parameters are arranged in the order of environmental parameters, temperature parameters, and operating status parameters to form a state vector S_t. Verify the acquisition timestamps of all feature parameters in the state vector S_t to ensure that the timestamp deviation does not exceed 10 seconds. If there are parameters with timeout deviations, replace them with the latest preprocessed data. The time-validated state vector S_t is input into the input layer of the policy network. This input layer has 12 nodes, which correspond one-to-one with the dimensions of the state space.

3. The defrosting control method based on proximal strategy optimization according to claim 2, characterized in that, The missing values ​​are handled as follows: if the missing value is a temperature value, the average value of the data within the previous 5 minutes is used to replace it; if the missing value is status data, the previous data is used to replace it.

4. The defrosting control method based on proximal strategy optimization according to claim 3, characterized in that, The outlier was handled by using the 3x standard deviation method. The mean u and standard deviation σ of the daily sampled data were calculated according to the feature dimension, and the data distributed in the intervals [-∞, u-3σ] and [u+3σ, +∞] were removed.

5. The defrosting control method based on proximal strategy optimization according to claim 3, characterized in that, The reward function R = -[w1·ΔE + w2·σ_T1 + w3·frost_time + w4·N_start], where ΔE is the energy consumption increment between two adjacent defrost cycles, σ_T1 is the standard deviation of the freezer compartment temperature between two adjacent defrost cycles, frost_time is the defrost interval duration with a maximum value of 80 hours, and N_start is the number of times the compressor starts and stops.

6. The defrosting control method based on proximal strategy optimization according to claim 3, characterized in that, Energy consumption E is obtained by using the area enclosed by the compressor frequency and time, and the specific calculation method is shown below: Where k is the compressor coefficient of performance, and f(t) is the frequency value of the compressor as a function of time.

7. The defrosting control method based on proximal strategy optimization according to claim 6, characterized in that, Training the constructed policy network using the proximal policy optimization algorithm includes: After each training round, a model generalization validation step is added: three sets of historical operating data from different regions are selected as validation sets. If the model's decision accuracy on the validation set is less than 85%, 500 more training rounds are automatically added and the learning rate is adjusted to 80% of the original initial learning rate. The different regions include high-temperature and high-humidity areas, low-temperature and dry areas, and temperate monsoon areas.

8. The defrosting control method based on proximal strategy optimization according to claim 1, characterized in that, The processing of the collected target data also includes dynamic weighted correction of the collected environmental humidity data. The correction coefficient is determined based on the deviation between the atmospheric pressure of the day and the historical average pressure. The larger the deviation, the larger the absolute value of the correction coefficient.

9. The defrosting control method based on proximal strategy optimization according to claim 7, characterized in that, The model was trained in the cloud using a GPU cluster, with 2000 rounds of training, each round consisting of 2048 steps. The training was terminated if the cumulative running time of the compressor in each round exceeded the interval between two adjacent defrost cycles. The batch size was set to 128, the initial learning rate was set to 3e-4, and the Adam optimizer was used.

10. A refrigerator, characterized in that, The refrigerator includes the defrosting control method based on proximal strategy optimization as described in any one of claims 1-9.

Images