Intelligent dynamic electric field fresh-keeping device with self-adaptive regulation and control function

Abstract

The invention discloses a self-adaptive regulation and control intelligent dynamic electric field fresh-keeping device, relates to the related technical field of livestock and poultry meat fresh-keeping, and aims to solve the problems of single monitoring and data lagging of a traditional device. The self-adaptive regulation and control intelligent dynamic electric field fresh-keeping device comprises an intelligent electric field generator arranged on the outer wall of a fresh-keeping cavity, and a high-voltage electrode arranged in the fresh-keeping cavity and configured to generate an alternating electric field; the sensor group comprises a temperature sensor for monitoring the environment temperature, a humidity sensor for monitoring the relative humidity of the environment and an electric field intensity sensor for monitoring the actual intensity of the alternating electric field; the signal input end of the embedded control unit is electrically connected with the driving circuit of the intelligent electric field generator and the sensor group, and the control output end of the embedded control unit is electrically connected with the intelligent electric field generator. Closed-loop control is formed by comparing the actual and target electric field intensities in real time and dynamically adjusting the voltage, it is ensured that the electric field in the cavity stably approaches the target value, the situation that the fresh-keeping effect is affected by electric field fluctuation is avoided, and the quality guarantee period of livestock and poultry meat is remarkably prolonged.

Description

Technical Field

[0001] This invention relates to the technical field of livestock and poultry meat preservation. More specifically, this invention relates to an adaptively controlled intelligent dynamic electric field preservation device. Background Technology

[0002] Livestock and poultry meat is rich in nutrients such as protein and fat. During storage, it is susceptible to microbial growth, enzymatic reactions, and oxidation, leading to problems such as excessive total bacterial count, increased volatile basic nitrogen, color deterioration, and juice loss, which seriously affect its edible quality and commercial value. Traditional methods of preserving livestock and poultry meat have significant limitations: while low-temperature refrigeration can slow down the rate of spoilage, its ability to inhibit microorganisms is limited, and long-term storage can still lead to deterioration in the flavor and sensory characteristics of fresh meat, and it also consumes a lot of energy.

[0003] In recent years, electric field preservation technology has gradually gained attention due to its green and residue-free advantages. However, most existing applications use a fixed electric field strength output mode, failing to adapt to dynamic changes in the storage environment (such as temperature fluctuations and humidity differences) and the different types of livestock and poultry meat and storage durations. Moreover, most devices lack the ability to accurately monitor the actual electric field strength and overall environmental temperature and humidity within the preservation chamber, resulting in uneven electric field distribution. This makes it difficult to match the quality maintenance requirements of livestock and poultry meat at different storage stages, and thus fails to achieve efficient, stable, and precise preservation.

[0004] Therefore, it is necessary to design a technical solution that can overcome the above-mentioned defects. Summary of the Invention

[0005] One objective of this invention is to provide an adaptive and adjustable intelligent dynamic electric field preservation device that can significantly extend the shelf life of livestock and poultry meat and better maintain its color, juice, and nutritional quality.

[0006] To achieve these and other advantages of the present invention, according to one aspect of the present invention, an adaptively controlled intelligent dynamic electric field preservation device is provided, comprising: an intelligent electric field generator disposed on the outer wall of a preservation cavity, wherein its high-voltage electrode is disposed inside the preservation cavity, the intelligent electric field generator being configured to generate an alternating electric field; a sensor group including a temperature sensor for monitoring ambient temperature, a humidity sensor for monitoring ambient relative humidity, and an electric field strength sensor for monitoring the actual strength of the alternating electric field; an embedded control unit having its signal input terminals electrically connected to the drive circuit of the intelligent electric field generator and the sensor group, respectively, and its control output terminal electrically connected to the intelligent electric field generator; the embedded control unit executes... The following control steps are as follows: S1: Periodically collect real-time ambient temperature data, real-time relative humidity data, and real-time electric field strength data of the preservation chamber through the sensor group, and package them into a monitoring data package; S2: Call the prediction model, input the real-time ambient temperature data, real-time relative humidity data, livestock and poultry meat category code, and storage time label as input feature vectors into the prediction model, and output the target electric field strength set value after calculation; S3: Compare the target electric field strength set value with the real-time electric field strength data to generate a voltage control signal; S4: Send the voltage control signal to the intelligent electric field generator to adjust its output voltage so that the real-time electric field strength in the preservation chamber tends to the target electric field strength set value.

[0007] Furthermore, the prediction model is a gradient boosting decision tree model. The gradient boosting decision tree model is trained on a cloud server. The cloud server obtains historical monitoring data packets, each of which contains ambient temperature data, ambient relative humidity data, livestock and poultry meat category codes, and corresponding storage time labels. The cloud server uses ambient temperature, ambient relative humidity, livestock and poultry meat category codes, and storage time labels as joint input features, and takes the optimal electric field strength required to maintain the best overall quality as the training objective to supervise the learning of the gradient boosting decision tree model. After training, the cloud server sends the model parameters to the embedded control unit.

[0008] Furthermore, during the model training phase, the cloud server establishes a basic model based on the livestock and poultry meat category codes in historical monitoring data packets. Each basic model corresponds to a specific category of livestock and poultry meat. For each basic model of livestock and poultry meat category, the training data is divided into several consecutive time stages based on the storage time label, and a dedicated time-weighted prediction sub-model is trained for each time stage. The prediction model combines these sub-models through an attention mechanism, which calculates the attention weight of each sub-model based on the input real-time storage time data. After training, the cloud server sends the structural parameters of the prediction model, the attention weight calculation function, and the corresponding livestock and poultry meat category-time mapping relationship to the embedded control unit. The embedded control unit identifies the current livestock and poultry meat category code and real-time storage time, and then activates the prediction model for the corresponding livestock and poultry meat category. The prediction model generates the weight coefficients of each sub-model based on the real-time storage time through the attention weight calculation function, and then weights and fuses the outputs of the sub-models for each time stage. Real-time ambient temperature data and real-time ambient relative humidity data are input into the weighted prediction model to generate a target electric field strength setting value that matches the current livestock and poultry meat category and storage stage.

[0009] Furthermore, the attention weight calculation function is performed using a Gaussian-based Softmax normalization form; for the i-th time stage sub-model, its attention weight α i(t) The formula for calculating α is: i(t) =exp(-β*(t-μ i ) 2 ) / (Σ j [exp(-β*(t-μ j ) 2 )]); where t is the real-time storage time, μ i Let μ be the center time parameter determined by the training data at the i-th time stage, and β be the bandwidth parameter controlling the width of the weight distribution. During the model training phase, the cloud server determines the optimal center time parameter μ for each time stage through an optimization algorithm. i The bandwidth parameter β is then used as part of the model parameters and sent to the embedded control unit.

[0010] Furthermore, the total bacterial count, volatile basic nitrogen content, juice loss rate, and color parameters of the livestock and poultry meat samples corresponding to each historical monitoring data package are tested. The cloud server pre-sets the weight coefficient and acceptable threshold range of each quality indicator, and calculates the comprehensive quality score of each sample through a weighted algorithm. The optimal comprehensive quality indicator is determined by the highest comprehensive quality score and the fact that all individual indicators do not exceed their acceptable threshold range.

[0011] Furthermore, the embedded control unit pre-assigns a weighting coefficient to each sensor based on its installation location. The magnitude of the weighting coefficient is determined by the proximity of the sensor to the livestock and poultry meat product. In each acquisition cycle, the embedded control unit multiplies the readings of each temperature sensor by its corresponding weighting coefficient and sums them to obtain a weighted temperature sum. Then, the weighted temperature sum is divided by the sum of all weighting coefficients to obtain real-time ambient temperature data that characterizes the overall thermal environment of the chamber. The same weighting operation is performed on the readings of each humidity sensor to obtain real-time ambient relative humidity data.

[0012] Furthermore, the embedded control unit also integrates a wireless communication module for establishing communication connections with the cloud server and the user's mobile terminal, respectively. The user's mobile terminal is configured to: receive and visualize real-time ambient temperature data, real-time ambient relative humidity data, real-time electric field strength data, and target electric field strength setpoint sent by the embedded control unit; receive user-manually set electric field strength adjustment commands based on the visualized information or confirm / modify the target electric field strength setpoint generated by the prediction model, and send the user's commands or confirmed / modified setpoints to the embedded control unit, which then generates the final voltage control signal accordingly; the user's mobile terminal also synchronizes the user's historical operation data to the cloud server, which uses the historical operation data as incremental training samples to continuously optimize the prediction model.

[0013] Furthermore, multiple high-voltage electrodes with a plate-like structure are installed inside the preservation cavity; Multiple temperature sensors, humidity sensors, and electric field strength sensors are arranged inside the preservation chamber to monitor real-time ambient temperature data, real-time ambient relative humidity data, and real-time electric field strength data at different locations. Multiple high-voltage electrodes are grouped into multiple independent electrode groups. During the preservation process, an embedded control unit controls these electrode groups to operate in an alternating mode. The embedded control unit calculates the electric field uniformity index within the preservation cavity based on real-time electric field intensity data from each electric field intensity sensor. When the electric field uniformity index is below a preset threshold, the switching rate of the alternating mode is increased; when the electric field uniformity index reaches or exceeds the preset threshold, the switching rate is decreased. This invention includes at least the following beneficial effects: This invention uses a sensor array to collect temperature, humidity, and electric field strength data in real time, overcoming the limitations of traditional devices that rely on single monitoring methods and data lag. This provides comprehensive and real-time environmental data for precise control, avoiding preservation deviations caused by insufficient environmental perception. The embedded control unit of this invention relies on a predictive model to generate a target electric field strength by combining the type of livestock and poultry meat, storage time, and real-time environmental parameters. This breaks free from the constraints of a fixed electric field output mode and can be specifically adapted to the needs of different types of livestock and poultry meat and different storage stages, making the electric field effect more closely match the meat's preservation characteristics and significantly improving preservation accuracy. This invention forms a closed-loop control by comparing the actual and target electric field strengths in real time and dynamically adjusting the voltage, ensuring that the electric field inside the cavity stably approaches the target value. This avoids electric field fluctuations affecting the preservation effect, effectively inhibiting microbial growth, slowing down enzymatic reactions and oxidation, significantly extending the shelf life of livestock and poultry meat, and better maintaining its color, juice, and nutritional quality.

[0014] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0015] Figure 1 This is a control flowchart of an embedded control unit according to an embodiment of this application. Detailed Implementation

[0016] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0017] It should be understood that terms such as "having," "comprising," and "including" used in the embodiments of this application do not exclude the presence or addition of one or more other elements or combinations thereof. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of components in a specific posture. If the specific posture changes, the directional indication will also change accordingly. When an element is referred to as "fixed to" or "set on" another element, it can be directly on the other element or may have an intervening element present. When an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element through an intervening element. Descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0018] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0019] The embodiments of this application provide an adaptively controlled intelligent dynamic electric field preservation device, comprising: an intelligent electric field generator disposed on the outer wall of the preservation cavity, with its high-voltage electrode disposed inside the preservation cavity, the intelligent electric field generator being configured to generate an alternating electric field; a sensor group including a temperature sensor for monitoring ambient temperature, a humidity sensor for monitoring ambient relative humidity, and an electric field strength sensor for monitoring the actual intensity of the alternating electric field; and an embedded control unit, the signal input terminals of which are electrically connected to the drive circuit of the intelligent electric field generator and the sensor group, respectively, and the control output terminal of which is electrically connected to the intelligent electric field generator; the embedded control unit executes the following control steps: S1: via the sensor The system periodically collects real-time ambient temperature data, real-time relative humidity data, and real-time electric field strength data of the preservation chamber and packages them into a monitoring data package; S2: The system calls the prediction model, inputting the real-time ambient temperature data, real-time relative humidity data, livestock and poultry meat category code, and storage time label as input feature vectors into the prediction model, and outputs the target electric field strength setpoint after calculation; S3: The system compares the target electric field strength setpoint with the real-time electric field strength data to generate a voltage control signal; S4: The system sends the voltage control signal to the intelligent electric field generator to adjust its output voltage so that the real-time electric field strength in the preservation chamber tends to the target electric field strength setpoint.

[0020] For example, the intelligent electric field generator refers to a specialized device capable of generating an adjustable alternating electric field. The outer shell of this device can be made of ABS engineering plastic and is located on the outer wall of the preservation chamber. The preservation chamber is a rectangular metal structure (or a composite insulation structure) used to store livestock and poultry meat in a sealed container. For example, its design dimensions are 5m long × 3m wide × 2.5m high, and the outer wall is the outer surface of the chamber. The intelligent electric field generator can be installed in the middle of the right outer wall of the preservation chamber or near the edge of the top outer wall of the preservation chamber. The high-voltage electrode includes a pair of polycarbonate plates and a wire. The opposing surfaces of the pair of polycarbonate plates are provided with placement grooves that match the wires. The pair of polycarbonate plates are aligned to confine the wires within the placement grooves. The wires are formed by connecting multiple U-shaped units in sequence. The length and width of the polycarbonate plates are 600mm and 500mm, respectively. The intelligent electric field generator is configured to generate an alternating electric field, the frequency of which can be 50Hz or 60Hz. The sensor group is a collection of multiple monitoring sensors. The temperature sensor can be installed in the middle of the left side wall inside the preservation chamber or in the lower part of the rear side wall inside the chamber. The humidity sensor can be installed in the middle of the right side wall inside the chamber or in the upper part of the front side wall inside the chamber. The electric field strength sensor can be installed at the center of the top surface inside the chamber or at the center of the bottom surface inside the chamber. The measurement range of the temperature sensor can be -20℃ to 50℃ or -10℃ to 40℃. The measurement range of the humidity sensor can be 20%RH to 95%RH or 30%RH to 90%RH. The measurement range of the electric field strength sensor can be 0kV / m to 50kV / m or 0kV / m to 80kV / m. An embedded control unit is a miniature control device used to receive signals and output control commands. It can be a microcontroller. The signal input terminal of the microcontroller is electrically connected to the drive circuit of the intelligent electric field generator and the signal output interface of the sensor group through wires. The control output terminal of the embedded control unit is electrically connected to the control interface of the intelligent electric field generator through wires. The embedded control unit can be installed on the lower left outer wall of the preservation cavity or on the middle of the rear outer wall of the cavity.The embedded control unit executes the following control steps: S1 periodically collects data through the sensor group. Periodic collection means collecting data at fixed time intervals, which can be 1 minute or 2 minutes. The process involves the embedded control unit sending a collection command to the sensor group according to the set period. After receiving the command, the temperature sensor measures the temperature inside the cavity and transmits the data to the embedded control unit. The humidity sensor and the electric field strength sensor perform the same operation synchronously. Afterward, the embedded control unit packages the real-time ambient temperature data, real-time ambient relative humidity data, and real-time electric field strength data into a monitoring data package according to JSON or XML format and stores it internally. In the storage module; S2 calls the prediction model, which can be a gradient boosting decision tree model (such as an XGBoost model). The livestock and poultry meat category code is a code that distinguishes different livestock and poultry meats, such as pork code 001 or chicken code 002. The storage time tag is an identifier that records the storage time, such as 12h or 24h. The execution process is that the embedded control unit calls the pre-stored XGBoost model from the internal storage module and combines the real-time ambient temperature data (such as 2℃), the real-time ambient relative humidity data (such as 60%RH), the livestock and poultry meat category code (such as 001), and the storage time tag (such as 12h) into an input feature vector ([2,60,001, 12) Input the model, and the model calculates through internal decision tree nodes (such as judging whether the temperature is less than 3℃, whether the humidity is between 55%-65%RH, etc.), and finally outputs the target electric field strength setting value, which can be 20kV / m or 30kV / m; S3 compares the target electric field strength setting value with the real-time electric field strength data. The comparison process is that the embedded control unit retrieves the target value (such as 20kV / m) and the real-time value (such as 18kV / m), calculates the difference between the two (20kV / m-18kV / m=2kV / m), and adjusts the corresponding relationship according to the preset difference-voltage (such as adjusting the voltage by 0.5V for every 1kV / m difference). The required voltage value (2kV / m × 0.5V / (kV / m) = 1V) is calculated, and a voltage control signal (such as a signal requiring an increase of 1V in the output voltage) is generated. S4 sends the voltage control signal to the intelligent electric field generator. The execution process is that the embedded control unit sends the signal to the control interface of the intelligent electric field generator through the control output terminal wire. After receiving the signal, the intelligent electric field generator adjusts the output voltage according to the signal (such as increasing it from 10V to 11V). The electric field strength output by the high-voltage electrode changes with the voltage adjustment (such as increasing it from 18kV / m to 20kV / m), so that the real-time electric field strength in the cavity tends to the target value.

[0021] In existing technologies, traditional alternating electric field preservation devices mostly adopt a fixed electric field output mode, such as a fixed-parameter electric field generator with a fixed electric field frequency of 50Hz and a fixed intensity of 25kV / m. They lack a complete sensor array and rely solely on manual measurement of the cavity temperature periodically (e.g., every 2 hours) using a common glass thermometer. This method cannot monitor humidity and actual electric field intensity, and lacks an embedded control unit, preventing adjustments to the electric field based on the type of livestock or poultry meat and storage time; it can only continuously output a fixed voltage. Compared to existing technologies, this embodiment achieves real-time and accurate monitoring of temperature, humidity, and electric field intensity through a sensor array, avoiding the lag and errors of manual measurements. Relying on an embedded control unit and predictive model, it can dynamically adjust the electric field intensity based on the type of livestock or poultry meat, storage time, and real-time environmental parameters, rather than providing a fixed output, making the electric field effect more suitable for different preservation needs. Closed-loop control ensures that the electric field intensity stably approaches the target value, effectively improving the problems of electric field fluctuations and poor adaptability in existing technologies, and is more conducive to maintaining the quality of livestock and poultry meat.

[0022] In another embodiment, the prediction model is a gradient boosting decision tree model. The gradient boosting decision tree model is trained on a cloud server. The cloud server obtains historical monitoring data packets, each of which contains ambient temperature data, ambient relative humidity data, livestock and poultry meat category codes, and corresponding storage time labels. The cloud server uses ambient temperature, ambient relative humidity, livestock and poultry meat category codes, and storage time labels as joint input features, and takes the optimal electric field strength required to maintain the best overall quality as the training objective to supervise the learning of the gradient boosting decision tree model. After training, the cloud server sends the model parameters to the embedded control unit.

[0023] For example, the prediction model is a gradient boosting decision tree model, which is an algorithm based on ensemble learning of multiple decision trees. XGBoost or LightGBM models can be used. Its core is to iteratively train multiple weak decision trees and then sum the prediction results of each tree in a weighted manner to obtain the final output. During the computation, the loss function (such as the mean squared error loss function) is minimized using gradient descent. The gradient boosting decision tree model is trained on a cloud server, which is a device used for remotely storing data and running model training programs. This server does not need to be installed on the preservation chamber and establishes a connection with the embedded control unit via wireless communication (such as 4G or Wi-Fi). During model training, the cloud server first acquires historical monitoring data packets. These historical monitoring data packets are datasets containing preservation-related parameters collected in the past. Each data packet contains ambient temperature data (such as 1℃, 3℃), ambient relative humidity data (such as 55%RH, 65%RH), livestock and poultry meat category codes (such as 001, 002), and corresponding storage time labels (such as 6h, 18h). The number of historical data packets can be 1000 or 2000. The cloud server uses ambient temperature, relative humidity, livestock and poultry meat category code, and storage time label as joint input features. Joint input features are model input variables formed by combining multiple independent parameters in a preset order. For example, temperature 2℃, humidity 60%RH, code 001, and time 12h are combined into a vector of [2,60,001,12]. The training objective is to maintain the optimal electric field strength required to achieve the best overall quality. The optimal electric field strength refers to the electric field strength value that keeps indicators such as total bacterial count and volatile basic nitrogen in livestock and poultry meat within an acceptable range and achieves the highest overall score, such as 18kV / m or 25kV / m. The cloud server performs supervised learning on the gradient boosting decision tree model. Supervised learning is a learning method that uses labeled training data (i.e., the optimal electric field strength) to adjust the model parameters. During training, parameters such as the number of iterations (e.g., 100 or 200), learning rate (e.g., 0.1 or 0.2), and tree depth (e.g., 5 or 8 layers) are set. By continuously iterating, the error between the model's predicted value and the actual optimal electric field strength is reduced. After training is completed, the cloud server sends the model parameters to the embedded control unit. The model parameters include the node splitting features of the decision tree, node thresholds, leaf node weights, etc. The sending process is completed through wireless communication. After receiving the parameters, the embedded control unit stores them in the internal Flash storage module.

[0024] In existing technologies, some preservation devices use simple linear regression models for prediction, trained on local microcontrollers. The training data only includes temperature and electric field strength, lacking parameters such as humidity, product category coding, and storage time. Furthermore, the model parameters cannot be updated after training, resulting in low prediction accuracy. In contrast, this embodiment employs a gradient boosting decision tree model, which can handle nonlinear relationships between multiple features and better reflects actual preservation needs than linear models. Training on a cloud server allows access to a large amount of historical data, ensuring more thorough training. Moreover, updated model parameters can be distributed to the embedded control unit, solving the problems of inaccurate predictions caused by simple models, insufficient data, and inability to update parameters in existing technologies, thus making the calculation of the target electric field strength more accurate.

[0025] In another embodiment, during the model training phase, the cloud server establishes a basic model based on the livestock and poultry meat category codes in historical monitoring data packets. Each basic model corresponds to a specific category of livestock and poultry meat. For each basic model of livestock and poultry meat category, the training data is divided into several consecutive time stages based on the storage time label, and a dedicated time-weighted prediction sub-model is trained for each time stage. The prediction model combines these sub-models through an attention mechanism, which calculates the attention weight of each sub-model based on the input real-time storage time data. After training, the cloud server sends the structural parameters of the prediction model, the attention weight calculation function, and the corresponding livestock and poultry meat category-time mapping relationship to the embedded control unit. The embedded control unit identifies the current livestock and poultry meat category code and real-time storage time, and then activates the prediction model for the corresponding livestock and poultry meat category. The prediction model generates the weight coefficients of each sub-model based on the real-time storage time through the attention weight calculation function, and then weights and fuses the outputs of the sub-models for each time stage. Real-time ambient temperature data and real-time ambient relative humidity data are input into the weighted prediction model to generate a target electric field strength setting value that matches the current livestock and poultry meat category and storage stage.

[0026] For example, during the model training phase, the cloud server builds a basic model based on the livestock and poultry meat category codes in historical monitoring data packets. These codes distinguish different types of livestock and poultry meat, such as beef code 003 and mutton code 004. Each basic model corresponds to a specific category of livestock and poultry meat; for example, code 003 corresponds to the beef basic model, and code 004 corresponds to the mutton basic model. The basic model is a gradient boosting decision tree model framework initially constructed for a single livestock and poultry meat category. For each livestock and poultry meat category's basic model, the training data is divided into several consecutive time periods based on storage time labels. Storage time labels are identifiers recording storage duration. The time periods can be divided into 6-hour periods (e.g., 0-6h, 7-12h) or 7-day periods (e.g., 0-7d). Each time period corresponds to a set of training data; for example, the training data for the 0-6h period includes the temperature and humidity, category code, and corresponding optimal electric field strength for that time period. A dedicated time-weighted prediction sub-model is trained for each time stage. This sub-model is a gradient boosting decision tree sub-model built based on the training data for the corresponding time stage. The training process is similar to the base model. By adjusting parameters such as the number of iterations (e.g., 80 or 150) and learning rate (e.g., 0.1 or 0.2), the sub-model can accurately predict the optimal electric field intensity for the corresponding time stage. The prediction model combines these sub-models using an attention mechanism. This attention mechanism is an algorithm that dynamically assigns weights to each sub-model based on the input data. It calculates the attention weights of each sub-model based on the real-time storage time data. For example, when the real-time storage time is 8 hours, higher weights are assigned to the sub-models for the 7-12 hour stage, and lower weights are assigned to the sub-models for the 0-6 hour stage. After training is completed, the cloud server sends the structural parameters of the prediction model (such as the number of sub-models and the basic architecture of each sub-model), the attention weight calculation function (such as the calculation formula based on the Gaussian function), and the corresponding livestock and poultry meat category-time mapping relationship (such as code 003 corresponding to stages such as 0-6h and 7-12h) to the embedded control unit. The sending method is wireless communication, and the embedded control unit stores these data in the internal storage module.The embedded control unit identifies the current livestock and poultry meat category code and real-time storage time. The identification process involves the embedded control unit receiving the category code (e.g., 003) input by the user via a mobile terminal and the start time recorded when the livestock and poultry meat was placed in the cavity. It calculates the real-time storage time (e.g., 8 hours) using the difference between the current time and the start time. Then, it activates the prediction model corresponding to the livestock and poultry meat category, i.e., it calls the prediction model matching code 003. The prediction model generates weight coefficients for each sub-model based on the real-time storage time using an attention weight calculation function. For example, when the real-time time is 8 hours, the weight coefficients for the 0-6 hours sub-model are 0.3 and 7-1, respectively. The weighting coefficient for the 2-hour sub-model is 0.7. Then, the outputs of the sub-models at each time stage are weighted and fused. The fusion process is to multiply the electric field strength value output by each sub-model by its corresponding weighting coefficient and then sum them (e.g., 0-6h sub-model output 18kV / m×0.3 + 7-12h sub-model output 22kV / m×0.7=20.8kV / m). Finally, real-time ambient temperature data (e.g., 2℃) and real-time ambient relative humidity data (e.g., 60%RH) are input into the weighted prediction model to generate a target electric field strength setting value (e.g., 21kV / m) that matches the current livestock and poultry meat category and storage stage.

[0027] In existing technologies, some models train by mixing data from different storage times without dividing the time period or using an attention mechanism, outputting only a single predicted value, which fails to adapt to the needs of different storage stages. In contrast, this embodiment divides the storage time into stages and trains sub-models accordingly. Through an attention mechanism, weights are dynamically allocated, allowing the model to prioritize the output of the corresponding sub-model based on the real-time storage time. This solves the problem of poor adaptability of existing models to different storage stages, making the target electric field strength more aligned with the quality maintenance requirements of livestock and poultry meat at different storage stages, and improving the accuracy of preservation.

[0028] In another embodiment, the attention weight calculation function is calculated using a Gaussian-based Softmax normalized form; for the i-th time stage sub-model, its attention weight α i(t) The formula for calculating α is: i(t) =exp(-β*(t-μ i ) 2 ) / (Σ j [exp(-β*(t-μ j ) 2 )]); where t is the real-time storage time, μ i Let μ be the center time parameter determined by the training data at the i-th time stage, and β be the bandwidth parameter controlling the width of the weight distribution. During the model training phase, the cloud server determines the optimal center time parameter μ for each time stage through an optimization algorithm. i The bandwidth parameter β is then used as part of the model parameters and sent to the embedded control unit.

[0029] For example, the attention weight calculation function is performed using a Gaussian function-based Softmax normalization form. The Gaussian function is a mathematical function with bell-shaped curve characteristics, and Softmax normalization is a method that converts multiple values ​​into a probability distribution. Combining these two methods ensures that the closer the real-time storage time is to the center time of a certain time stage, the greater the weight of the sub-model at that stage. For the i-th time stage sub-model, its attention weight α... i(t) In the calculation formula, t represents the real-time storage time, which can be in days (d), such as t=9d or t=16d; μ i Let μ be the center time parameter determined by the training data for the i-th time stage. The center time parameter is the midpoint of that time stage, such as μ for the 0-7d stage. i μ can be in the 3.5d or 8-14d phase. i It can be 11d; β is the bandwidth parameter that controls the width of the weight distribution. The larger the value of β, the more concentrated the weight distribution. β can be 0.05 or 0.1; exp is the natural exponential function, Σ j This represents the expression for the sub-model at all time stages: exp(-β*(t-μ) j ) 2 The values ​​are summed. For example, when there are four time-phase sub-models, t=9d, μ1=3.5d (0-7d phase), μ2=11d (8-14d phase), μ3=18d, μ4=25d, and β=0.05, the numerator is calculated separately: For the 0-7d phase sub-model, exp(-0.05*(9-3.5)) 2 =exp(-0.05*30.25)=exp(-1.5125)≈0.220; For the sub-model of stage 8-14d, exp(-0.05*(9-11) 2 ) = exp(-0.05*4) = exp(-0.2) ≈ 0.819; then calculate the sum of the denominators: 0.220 + 0.819 = 1.039; finally, we can obtain α. 1(t) =0.220 / 1.039≈0.212, α 2(t) =0.819 / 1.039≈0.788, α 3(t) ≈0.0166, α 4(t) ≈2.6×10 -6 .

[0030] During the model training phase, the cloud server uses optimization algorithms to determine the optimal center time parameter μ for each time stage. i Given the bandwidth parameter β, the optimization algorithm can be either gradient descent or particle swarm optimization. The goal is to minimize the error between the model-predicted target electric field strength and the actual optimal electric field strength, and to continuously adjust μ. iThe value of β, for example, by adjusting β from the initial value of 0.03 to the optimal value of 0.05 through multiple iterations, will affect μ in the 0-6h stage. i The time was adjusted from 2.5h to 3h. After determining the optimal parameters, the cloud server used them as part of the model parameters and sent them to the embedded control unit via wireless communication. The embedded control unit stored these parameters in its internal storage module for subsequent calculation of attention weights.

[0031] In existing technologies, some attention weights are calculated using a simple averaging method, with each sub-model's weight fixed at 1 / n (where n is the number of sub-models). This method cannot be dynamically adjusted based on real-time storage time, resulting in unreasonable weight allocation. Compared to existing technologies, this embodiment uses Softmax normalization based on a Gaussian function to calculate weights. This allows for dynamic weight allocation based on the distance between real-time storage time and the center time of each stage, making the weights more aligned with actual needs. Furthermore, by optimizing the algorithm to determine the optimal parameters, the accuracy of weight calculation is further improved, solving the problems of fixed weights and poor adaptability in existing technologies, resulting in more accurate sub-model fusion results.

[0032] In another embodiment, the total bacterial count, volatile basic nitrogen content, juice loss rate, and color parameters of the livestock and poultry meat samples corresponding to each historical monitoring data package are detected. The cloud server pre-sets the weight coefficient and acceptable threshold range of each quality indicator, and calculates the comprehensive quality score of each sample through a weighted algorithm. The optimal comprehensive quality indicator is determined by the highest comprehensive quality score and the fact that all individual indicators do not exceed their acceptable threshold range.

[0033] For example, multiple quality indicators are tested on the livestock and poultry meat samples corresponding to each historical monitoring data package. The livestock and poultry meat samples refer to the livestock and poultry meat samples that are consistent with the category and storage conditions recorded in the historical monitoring data package, such as the pork sample corresponding to code 001 (pork) and storage time of 12h. The indicators tested include total bacterial count, volatile basic nitrogen content, juice loss rate, and color parameters. Total bacterial count refers to the number of bacteria per gram of meat sample. The detection method can be plate counting, and the instrument can be a mold incubator from Shanghai Jinghong. The unit of measurement is CFU / g. Volatile basic nitrogen content refers to the content of alkaline nitrogenous substances produced by the decomposition of meat during storage. The detection method can be semi-micro nitrogen determination, and the instrument can be an automatic Kjeldahl nitrogen analyzer. The unit of measurement is mg / 100g. Juice loss rate refers to the percentage of juice lost during storage relative to the initial mass of the meat. The detection method can be gravimetric analysis, and the instrument can be a Mettler Toledo PL2002 electronic balance. The result is typically 2% or 4%. Color parameters refer to the color characteristics of meat, including L* value (brightness), a* value (redness), and b* value (yellowness). The detection method can be a colorimeter, and the instrument can be a colorimeter. The cloud server pre-sets the weighting coefficient and acceptable threshold range for each quality indicator. The weighting coefficient measures the degree of influence of each indicator on the overall quality. For example, the weighting coefficient for total bacterial count can be 0.4 or 0.5, the weighting coefficient for volatile basic nitrogen content can be 0.3 or 0.25, the weighting coefficient for juice loss rate can be 0.15 or 0.15, and the weighting coefficient for color parameters can be 0.15 or 0.1. The acceptable threshold range refers to the maximum or minimum range that an indicator can be allowed to exist. For example, the acceptable threshold range for total bacterial count can be ≤1×10⁻⁶. 5 CFU / g or ≤5×10 4 The acceptable threshold ranges for CFU / g, volatile basic nitrogen content, and juice loss rate are ≤5% or ≤4%. The acceptable threshold ranges for color parameters are 45-60 or 48-62 for L* value, 8-15 or 10-16 for a* value, and 5-10 or 6-11 for b* value. The cloud server calculates the overall quality score for each sample using a weighted algorithm. This algorithm standardizes the measured values ​​of each indicator, multiplies them by their corresponding weight coefficients, and then sums the results. Standardization converts the measured values ​​of the indicators into scores from 0 to 100; for example, the total bacterial count is 5 × 10⁻¹⁰. 3 CFU / g, threshold ≤5×10 4 CFU / g, standardized score (5×10) 4 -5×10 3 ) / (5×10 4)×100=90 points; The criterion for judging the optimal comprehensive quality index is that the comprehensive quality score is the highest and all individual indicators do not exceed their acceptable threshold range. For example, if sample A has a comprehensive score of 85 points and all indicators are within the threshold, and sample B has a comprehensive score of 82 points and all indicators are within the threshold, then the electric field strength corresponding to sample A is the optimal electric field strength.

[0034] In existing technologies, some quality assessments focus only on the single indicator of total bacterial count, neglecting other indicators and lacking clear weights and thresholds, resulting in ambiguous judgment criteria. Compared to existing technologies, this embodiment comprehensively considers multiple quality indicators, setting weight coefficients and acceptable thresholds to make the overall quality assessment more comprehensive and objective. The application of a weighted algorithm quantifies the overall quality, clarifies the optimal judgment criteria, and solves the problems of single evaluation and ambiguous standards in existing technologies, providing more accurate target values ​​for model training.

[0035] In another embodiment, the embedded control unit pre-assigns a weighting coefficient to each sensor based on its installation location. The magnitude of the weighting coefficient is determined according to the proximity of the sensor to the livestock and poultry meat product. During each acquisition cycle, the embedded control unit multiplies the readings of each temperature sensor by its corresponding weighting coefficient and sums them to obtain a weighted temperature sum. Then, the weighted temperature sum is divided by the sum of all weighting coefficients to obtain real-time ambient temperature data that characterizes the overall thermal environment of the chamber. The same weighting operation is performed on the readings of each humidity sensor to obtain real-time ambient relative humidity data.

[0036] For example, the embedded control unit pre-assigns a weighting coefficient to each sensor based on its installation location. The sensor installation locations, as described above, include the middle of the left side wall, the lower part of the rear side wall, and the middle of the right side wall inside the cavity. The magnitude of the weighting coefficient is determined based on the proximity of the sensor to the meat product—the closer the proximity, the larger the weighting coefficient. For example, if the meat product is placed on the middle shelf inside the cavity, the temperature sensor installed on the middle side wall inside the cavity is closer to the product, and the weighting coefficient can be 0.6 or 0.7. The temperature sensor installed on the top or bottom of the cavity is farther from the product, and the weighting coefficient can be 0.4 or 0.3. Within each acquisition cycle (which can be 1 minute or 2 minutes), the embedded control unit multiplies the readings of each temperature sensor by their corresponding weighting coefficients and then sums them to obtain a weighted temperature sum. For example, if there are two temperature sensors, sensor 1 reads 2℃ with a weighting coefficient of 0.6, and sensor 2 reads 1.8℃ with a weighting coefficient of 0.4, the weighted temperature sum = 2 × 0.6 + 1.8 × 0.4 = 1.2 + 0.72 = 1.92℃. Then, the weighted temperature sum is divided by the sum of all weighting coefficients (0.6 + 0.4 = 1) to obtain the real-time ambient temperature data used to characterize the overall thermal environment of the chamber, i.e., 1.92℃ / 1 = 1.92℃. The same weighted calculation process is performed on the readings of each humidity sensor. For example, if there are two humidity sensors, sensor A reads 60%RH with a weighting coefficient of 0.7, and sensor B reads 58%RH with a weighting coefficient of 0.3, the weighted humidity sum = 60 × 0.7 + 58 × 0.3 = 42 + 17.4 = 59.4%RH, and the sum of the weighting coefficients = 0.7 + 0.3 = 1. The real-time ambient relative humidity data = 59.4%RH / 1 = 59.4%RH. After the calculation is completed, the embedded control unit stores the real-time ambient temperature data and real-time ambient relative humidity data in the internal storage module for subsequent input to the prediction model and data display.

[0037] In existing technologies, some devices use a simple averaging method to calculate data from multiple sensors, without considering the distance differences between the sensors and the product. This results in data that cannot accurately represent the overall environment of the chamber. Compared with existing technologies, this embodiment assigns weight coefficients based on the proximity of the sensors to the livestock and poultry meat products, and obtains environmental parameters through weighted calculations. This makes the calculation results more consistent with the actual environment of the product, solves the problem of inaccurate data representation in existing technologies, provides more accurate input parameters for the prediction model, and improves the calculation accuracy of the target electric field strength.

[0038] In another embodiment, the embedded control unit also integrates a wireless communication module for establishing communication connections with the cloud server and the user's mobile terminal, respectively. The user's mobile terminal is configured to: receive and visualize real-time ambient temperature data, real-time ambient relative humidity data, real-time electric field strength data, and target electric field strength setpoint sent by the embedded control unit; receive electric field strength adjustment instructions manually set by the user based on the visualized information or confirm / modify the target electric field strength setpoint generated by the prediction model, and send the user's instructions or confirmed / modified setpoint to the embedded control unit, which then generates the final voltage control signal accordingly. The user's mobile terminal also synchronizes the user's historical operation data to the cloud server, which uses the historical operation data as incremental training samples to continuously optimize the prediction model.

[0039] For example, the embedded control unit also integrates a wireless communication module, which is used to realize wireless data transmission. This module is soldered onto the circuit board of the embedded control unit and is used to establish communication connections with the cloud server and the user's mobile terminal, respectively. The connection with the cloud server can be 4G or Wi-Fi, and the connection with the user's mobile terminal can be Wi-Fi or Bluetooth. The user's mobile terminal refers to the user's smartphone or tablet, which has a corresponding preservation device control APP installed. This APP is configured to receive and visualize the data sent by the embedded control unit—the embedded control unit sends real-time ambient temperature data, real-time ambient relative humidity data, real-time electric field strength data, and target electric field strength setpoint to the mobile terminal through the wireless communication module. The APP visualizes this data in the form of numbers, line graphs, or dashboards for easy viewing by the user. The app can also receive manual electric field strength adjustment commands from users based on visual information, or receive confirmation / modification operations from users on the target electric field strength setting value generated by the prediction model. For example, after viewing the data, the user manually sets the electric field strength to 22kV / m, or modifies the 21kV / m generated by the prediction model to 20.5kV / m. The app sends the user's command or the confirmed / modified setting value to the embedded control unit via wireless communication. After receiving the setting, the embedded control unit generates the final voltage control signal accordingly. If the user sets 22kV / m, the control signal is generated by comparing this value with the real-time electric field strength. In addition, the user's mobile terminal will also synchronize the user's historical operation data to the cloud server. The historical operation data includes the electric field strength value manually set by the user, the modification time, and the corresponding environmental parameters. For example, on a certain day in 2025, the user modified the electric field strength from 21kV / m to 20.5kV / m, with a corresponding temperature of 1.92℃ and humidity of 59.4%RH. The cloud server uses the historical operation data as incremental training samples. The incremental training samples are new samples added on the basis of the original training data. They are used to continuously optimize the prediction model. The optimization process is that the cloud server adds incremental samples to the training dataset periodically (e.g., monthly), retrains the model, and adjusts the model parameters to make the model more in line with the user's actual usage needs. After that, the optimized model parameters are sent to the embedded control unit.

[0040] In existing technologies, some devices lack wireless communication capabilities, preventing users from remotely viewing data and performing operations. Furthermore, the models cannot be optimized based on user actions, resulting in poor flexibility. In contrast, this embodiment utilizes a wireless communication module to connect the user's mobile terminal with the embedded control unit and cloud server. Users can remotely view data and adjust parameters, improving ease of use. Simultaneously, historical user operation data is used for model optimization, enabling the model to continuously adapt to actual needs and resolving the issues of poor flexibility and the inability to continuously optimize models inherent in existing technologies.

[0041] In another embodiment, multiple high-voltage electrodes in a plate-like structure are disposed inside the preservation cavity. Multiple temperature sensors, humidity sensors, and multiple electric field strength sensors are arranged within the internal space of the preservation cavity to monitor real-time ambient temperature data, real-time relative humidity data, and real-time electric field strength data at different locations. The multiple high-voltage electrodes are grouped to form multiple independent electrode groups. During the preservation process, an embedded control unit controls the multiple electrode groups to operate in an alternating mode. The embedded control unit calculates the electric field uniformity index within the preservation cavity based on the real-time electric field strength data from each electric field strength sensor. When the electric field uniformity index is lower than a preset threshold, the switching rate of the alternating mode is increased; when the electric field uniformity index reaches or exceeds the preset threshold, the switching rate is decreased. For example, the preservation cavity is a rectangular metal structure, and the material of the metal structure can be 304 stainless steel or galvanized steel plate, for example, with design dimensions of 5m long × 3m wide × 2.5m high, and a base area of ​​15m². 2 The high-voltage electrode itself acts as a grounding electrode—the grounding electrode is used to form an electric field loop with the high-voltage electrode. The preservation cavity is connected to the ground via a wire to ensure a stable and safe electric field. Multiple shelves 3 (e.g., 3 shelves, each 0.8m high) are installed inside the preservation cavity 1. The high-voltage electrode has a plate-like structure. Each high-voltage electrode includes a pair of polycarbonate plates and a wire. The opposing surfaces of the pair of polycarbonate plates have placement grooves that match the wires. The pair of polycarbonate plates are aligned, confining the wires within the placement grooves. The wires are formed by connecting multiple U-shaped units sequentially. The length and width of the polycarbonate plates are 600mm and 500mm respectively (see the applicant's previous patent CN202411116862.0 for details). Sixteen high-voltage electrodes are evenly arranged along the length of the cavity (the cavity is 5m long, with electrodes at 1m, 2m, 3m, and 4m). Four electrodes are installed at each location, with the four electrodes basically on the same plane, totaling 16 electrodes (4 locations x 4 electrodes). The four electrodes at each location are numbered "1, 2, 3, 4" and are installed by bolts to the pre-set mounting holes on the inner side walls of the cavity and the side of the shelf. Temperature and humidity sensors are arranged along the inner walls of the cavity and the side of the shelf (4 temperature sensors and 4 humidity sensors in total). Electric field strength sensors are arranged at the four corners of the top surface, the four corners of the bottom surface, and the center of the cavity (9 in total) to monitor environmental parameters and electric field strength at different locations in real time. The 16 high-voltage electrodes were divided into 4 independent electrode groups (Group 1: 1m-1, 2m-1, 3m-1, 4m-1; Group 2: 1m-2, 2m-2, 3m-2, 4m-2; Group 3: 1m-3, 2m-3, 3m-3, 4m-3; Group 4: 1m-4, 2m-4, 3m-4, 4m-4).

[0042] During the preservation process, the embedded control unit controls four electrode groups to operate in an alternating mode. The initial switching rate of the alternating mode is set to 1 time / minute (i.e., each electrode group works continuously for 1 minute before switching to the next group). The electric field uniformity index is calculated as follows: first, the average value of the real-time data from all electric field intensity sensors is calculated; then, the standard deviation of each sensor's data from the average value is calculated; finally, the standard deviation / average value is used as the electric field uniformity index (the smaller the value, the more uniform the electric field); the preset threshold is set to 0.15. When the calculated electric field uniformity index is 0.18 (above the threshold of 0.15) in a certain acquisition cycle, the embedded control unit automatically increases the switching rate to 2 times / minute; when the index drops to 0.12 (below the threshold) in subsequent cycles, the switching rate is reduced back to 1 time / minute, thus ensuring that the electric field in the cavity remains uniformly distributed. This embodiment solves the problem of uneven electric field distribution caused by fixed electrode operation through the design of multi-electrode grouping, alternating operation and uniformity dynamic control. It is especially suitable for large-space batch preservation scenarios, and can avoid the local electric field being too strong or too weak, which will affect the preservation effect and further improve the stability of livestock and poultry meat quality maintenance.

[0043] The following is a detailed explanation using experiments.

[0044] I. Experimental Group Intelligent electric field generator and preservation chamber: The preservation chamber is a rectangular structure made of 304 stainless steel, with dimensions of 5m long × 3m wide × 2.5m high, and serves as the grounding electrode itself. There are 16 high-voltage electrodes in a plate-like structure. Each high-voltage electrode includes a pair of polycarbonate plates and a wire. The opposing surfaces of the pair of polycarbonate plates have placement grooves that match the wire. The pair of polycarbonate plates are aligned, confining the wire within the placement grooves. The wire is composed of multiple U-shaped units connected sequentially. The length and width of the polycarbonate plates are 600mm and 500mm respectively. The 16 high-voltage electrodes are evenly arranged along the length of the chamber (the chamber is 5m long, and the electrodes are arranged at four positions: 1m, 2m, 3m, and 4m, with four electrodes installed at each position, totaling 16 electrodes (4 positions × 4 electrodes). The four electrodes at each position are numbered "1, 2, 3, and 4"). The alternating electric field frequency is set to 50Hz. The 16 high-voltage electrodes were divided into 4 independent electrode groups according to the principle of "one electrode per location" (Group 1: 1m-1, 2m-1, 3m-1, 4m-1; Group 2: 1m-2, 2m-2, 3m-2, 4m-2; Group 3: 1m-3, 2m-3, 3m-3, 4m-3; Group 4: 1m-4, 2m-4, 3m-4, 4m-4); 4 temperature sensors (arranged in the middle of the inner wall around the cavity), 4 humidity sensors (arranged next to the temperature sensors), and 9 electric field strength sensors (arranged at the four corners of the top surface, the four corners of the bottom surface, and the center of the cavity); the weighting coefficient of the temperature and humidity sensors closer to the shelf was 0.6, and the weighting coefficient of the sensors slightly farther away was 0.4.

[0045] Sensor group: One temperature sensor is installed in the middle of the front side wall of the cavity (next to the "Group 1" electrode), one in the middle of the rear side wall (next to the "Group 2" electrode), one in the middle of the left side wall (next to the "Group 3" electrode), and one in the middle of the right side wall (next to the "Group 4" electrode); one humidity sensor is installed next to the temperature sensor on the front side wall of the cavity, one next to the temperature sensor on the rear side wall, one next to the temperature sensor on the left side wall, and one next to the temperature sensor on the right side wall; nine electric field strength sensors are installed symmetrically at the four corners of the top surface, the four corners of the bottom surface, and the center of the cavity to monitor the electric field strength at different locations within the cavity.

[0046] Embedded control unit and communication module: Installed on the lower part of the rear outer wall of the cavity, integrating a Quectel EC20 4G module. The signal input terminal of the control unit is connected to the drive circuit and sensor group of the intelligent electric field generator, and the control output terminal is connected to the intelligent electric field generator.

[0047] Prediction Model: The prediction model adopts the XGBoost gradient boosting decision tree model. The basic model is established according to the livestock and poultry meat category code (pork code 001). Combined with the long-term storage requirements, the storage time is divided into 7-day stages (0-7d, 8-14d, 15-21d, 22-28d), for a total of 4 consecutive time stages. A dedicated time-weighted prediction sub-model is trained for each stage. The center time parameter μ1=3.5d for the 0-7d (week 1) stage; μ2=10.5d for the 8-14d (week 2) stage; μ3=17.5d for the 15-21d (week 3) stage; and μ4=24d for the 22-28d (week 4) stage. The bandwidth parameter β is maintained at 0.05 (fixed after optimization during model training).

[0048] During model training, the cloud server uses ambient temperature, relative humidity, livestock and poultry meat category code, and storage time label as joint input features, and takes the electric field strength that maintains the best overall quality of pork within 28 days as the training objective. After completing the training of 4 sub-models, the model structure parameters, attention weights, and the mapping relationship between pork code 001 and the 4 time stages are sent to the embedded control unit.

[0049] II. Operational Procedure The embedded control unit collects sensor data at a 1-minute cycle (the collection cycle remains constant to ensure real-time performance).

[0050] Temperature and humidity calculation method: Combining a distributed layout of 4 temperature sensors and 4 humidity sensors (1 each on the front wall, back wall, left wall, and right wall), weights are assigned according to the proximity of the sensors to the pork: Temperature calculation: The weighted temperature sum is calculated as follows: temperature sensor 1 (front wall, near the middle shelf, weight 0.6) reading × 0.6 + temperature sensor 2 (rear wall, near the middle shelf, weight 0.5) reading × 0.5 + temperature sensor 3 (left wall, away from the shelf, weight 0.5) reading × 0.5 + temperature sensor 4 (right wall, away from the shelf, weight 0.4) reading × 0.4. Then, the weighted temperature sum is divided by the sum of the weighting coefficients (0.6 + 0.5 + 0.5 + 0.4 = 2) to obtain the final real-time ambient temperature.

[0051] Humidity calculation: Consistent with temperature logic, the reading of humidity sensor 1 (front wall, weight 0.6) × 0.6 + the reading of humidity sensor 2 (rear wall, weight 0.5) × 0.5 + the reading of humidity sensor 3 (left wall, weight 0.5) × 0.5 + the reading of humidity sensor 4 (right wall, weight 0.4) × 0.4, weighted sum and divided by the total weighted sum of 2, yields the real-time ambient relative humidity.

[0052] Real-time ambient temperature, real-time relative humidity, and real-time electric field intensity data are packaged into a monitoring data package in JSON format and stored in the internal storage module.

[0053] When the embedded control unit calls the XGBoost model, it dynamically inputs feature vectors based on different storage nodes within 28 days and calculates the sub-model weights. Electric field strength closed-loop adjustment: The target electric field strength is compared with the real-time electric field strength, and the difference is calculated. Based on the preset difference-voltage adjustment correspondence (each 1kV / m difference corresponds to a 0.5V voltage adjustment), a voltage control signal is generated and sent to the intelligent electric field generator to ensure that the real-time electric field strength approaches the target value. Electrode group alternating operation and uniformity control: Control the four electrode groups to operate in an alternating mode (the initial switching rate is set to 1 time / minute, that is, each group works continuously for 1 minute before switching to the next group); at the same time, the electric field uniformity index is calculated based on the real-time data of 9 electric field intensity sensors, and the preset threshold is set to 0.15; when the uniformity index is >0.15, the switching rate is automatically increased to 2 times / minute; when the uniformity index is ≤0.15, the switching rate is reduced back to 1 time / minute to ensure the uniform distribution of the electric field in the cavity and ensure the quality of pork at the end of 28 days of storage.

[0054] II. Test Methods All groups used the same breed of fresh pork (hind leg meat), which was processed within 1 hour of purchase. The initial quality indicator was a total bacterial count of 1.2 × 10⁻⁶. 3 CFU / g, volatile basic nitrogen 6.5mg / 100g, juice loss rate 0.8%. Cut pork into uniform samples of 50g each, take 3 replicates for each group, put them into aseptic preservation boxes and place them on the middle shelf of the corresponding group's preservation cavity.

[0055] Apart from the differences in electric field-related design, the initial temperature of the preservation chamber for all groups was set to 4℃ and the initial relative humidity was set to 60%. The placement of the chamber was not changed during the test to avoid interference from external temperature and humidity fluctuations.

[0056] Samples were taken and tested at 0, 7, 14, 21, and 28 days of storage (a total of 5 time points). Total bacterial count was determined using the plate count method, incubated in a Shanghai Jinghong mold incubator before counting. Volatile basic nitrogen was determined using the semi-micro nitrogen determination method, measured using a Shanghai Peiou automatic Kjeldahl nitrogen analyzer. Juice loss rate was determined using the gravimetric method, weighing the samples before and after storage using an electronic balance, and calculating the percentage of the mass difference relative to the initial mass.

[0057] III. Control Group and Blank Group Design Control group 1: Fixed output electric field strength; no temperature, humidity, or electric field strength sensors were installed; the temperature of the cavity was measured manually every 2 hours with a glass thermometer; the electric field strength was not adjusted; there was no alternating operation mode; the preservation cavity, pork sample, and storage conditions were the same as those of the experimental group.

[0058] Control group 2: One temperature sensor and one humidity sensor were installed, but no electric field strength sensor was installed; there was no XGBoost prediction model, and the electric field was adjusted only according to the rule that the electric field strength should be reduced by 5kV / m when the temperature is ≥5℃; the rest of the equipment, samples, and storage conditions were the same as those of the experimental group.

[0059] Control group 3: The sensor configuration was exactly the same as the experimental group; the prediction model was a regular XGBoost model, which was not divided into stages according to storage time, and was directly trained with 28 days of full-cycle data without attention weight allocation; the rest of the devices, samples, and storage conditions were the same as the experimental group.

[0060] Blank group: No electric field effect (simulating conventional cold storage without any electric field preservation measures), using the same specifications of preservation chamber as the experimental group, without installing intelligent electric field generator, sensor group and embedded control unit; only maintaining the refrigeration conditions of chamber temperature 4℃ and relative humidity 60%; pork samples and storage conditions are the same as the experimental group.

[0061] IV. Evaluation Indicators and Results Comparison Three core quality indicators were selected: total bacterial count, volatile basic nitrogen, and juice loss rate. The test results for each group are shown in the table below.

[0062] V. Conclusion Control group 1 used a fixed electric field strength without dynamic adjustment or alternating operation, and the total bacterial count reached 1.5 × 10⁻⁶ after 28 days. 8 CFU / g is 3.5×10⁻⁶ in the experimental group. 6The CFU / g of the experimental group was 43 times that of the experimental group; the volatile basic nitrogen was 24.0 mg / 100g, which was 96.7% higher than the 12.2 mg / 100g of the experimental group. This indicates that the synergistic design of dynamic electric field adjustment and alternating electrode operation in the experimental group can effectively inhibit microbial growth and meet the requirements of long-term storage of 28 days, while traditional fixed electric field devices cannot meet the requirements of long-term uniform preservation.

[0063] Compared to control group 2: Control group 2 relied solely on a single temperature and humidity sensor and simple rule adjustments. Its 28-day juice loss rate was 3.0%, 57.9% higher than the experimental group's 1.9%; volatile basic nitrogen was 21.0 mg / 100g, slightly exceeding the standard. This reflects that a single sensor cannot accurately characterize local environmental differences within a large cavity (such as temperature and humidity fluctuations between shelf shelves), leading to a lag in electric field regulation and making it difficult to adapt to the slow changes in pork quality over 28 days.

[0064] Compared with control group 3: control group 3 had no phased sub-model or attention mechanism, and the total colony count at 28 days was 7.0 × 10⁻⁶. 8 CFU / g is 3.5×10⁻⁶ in the experimental group. 6 The CFU / g concentration was 200 times higher than that of the experimental group; the volatile basic nitrogen concentration was 17.5 mg / 100g, which was 43.4% higher than that of the experimental group. This indicates that the staged attention model designed for the experimental group can achieve precise electric field adaptation for different storage stages within 28 days, which is something that ordinary XGBoost models cannot achieve.

[0065] In summary, the experimental group, by integrating a phased prediction model, multi-dimensional weighted monitoring, and alternating electrode grouping technology, achieved significantly lower total bacterial count, volatile basic nitrogen, and juice loss rate than the control and blank groups within 28 days, thus verifying the superiority of this invention in long-term preservation scenarios.

[0066] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. An adaptively adjustable intelligent dynamic electric field preservation device, characterized in that, include: The intelligent electric field generator is installed on the outer wall of the preservation cavity, and its high-voltage electrode is installed inside the preservation cavity. The intelligent electric field generator is configured to generate an alternating electric field. The sensor array includes a temperature sensor for monitoring ambient temperature, a humidity sensor for monitoring ambient relative humidity, and an electric field strength sensor for monitoring the actual strength of the alternating electric field. An embedded control unit has its signal input terminals electrically connected to the drive circuit and sensor group of the intelligent electric field generator, respectively, and its control output terminal electrically connected to the intelligent electric field generator. The embedded control unit performs the following control steps: S1: Periodically collect real-time ambient temperature data, real-time ambient relative humidity data, and real-time electric field intensity data of the preservation chamber through the sensor group, and package them into a monitoring data package; S2: Call the prediction model, input the real-time ambient temperature data, real-time ambient relative humidity data, livestock and poultry meat category code and storage time label as input feature vectors into the prediction model, and output the target electric field strength set value after the prediction model is calculated; S3: Compare the target electric field strength setpoint with the real-time electric field strength data to generate a voltage control signal; S4: Send the voltage control signal to the intelligent electric field generator to adjust its output voltage so that the real-time electric field strength inside the preservation cavity tends to the target electric field strength set value.

2. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 1, characterized in that, The prediction model is a gradient boosting decision tree model. The gradient boosting decision tree model is trained on a cloud server. The cloud server obtains historical monitoring data packets. Each historical monitoring data packet contains ambient temperature data, ambient relative humidity data, livestock and poultry meat category codes, and corresponding storage time labels. The cloud server uses ambient temperature, ambient relative humidity, livestock and poultry meat category codes, and storage time labels as joint input features, and takes the optimal electric field strength required to maintain the best overall quality as the training objective to supervise the learning of the gradient boosting decision tree model. After training, the cloud server sends the model parameters to the embedded control unit.

3. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 2, characterized in that, During the model training phase, the cloud server establishes a basic model based on the livestock and poultry meat category codes in the historical monitoring data packets. Each basic model corresponds to a specific category of livestock and poultry meat. For each basic model of livestock and poultry meat category, the training data is divided into several consecutive time periods based on the storage time label, and a dedicated time-weighted prediction sub-model is trained for each time period. The prediction model combines these sub-models through an attention mechanism, which calculates the attention weight of each sub-model based on the input real-time storage time data. After training is completed, the cloud server will send the structural parameters of the prediction model, the attention weight calculation function, and the corresponding livestock and poultry meat category-time mapping relationship to the embedded control unit. The embedded control unit identifies the current livestock and poultry meat category code and real-time storage time, and then activates the prediction model for the corresponding livestock and poultry meat category. The prediction model generates the weight coefficients of each sub-model based on the real-time storage time through the attention weight calculation function, and then weights and fuses the outputs of the sub-models at each time stage. Real-time ambient temperature data and real-time ambient relative humidity data are input into the weighted prediction model to generate a target electric field strength setpoint that matches the current livestock and poultry meat category and storage stage.

4. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 3, characterized in that, The attention weight calculation function is performed using a Gaussian-based Softmax normalization form; For the sub-model at the i-th time stage, its attention weight α i(t) The calculation formula is: α i(t) =exp(-β*(t-μ i ) 2 ) / (Σ j [exp(-β*(t-μ j ) 2 )]); where t is the real-time storage time, μ i Let β be the center time parameter determined by the training data for the i-th time stage, and let β be the bandwidth parameter that controls the width of the weight distribution. During the model training phase, the cloud server uses optimization algorithms to determine the optimal center time parameter μ for each time stage. i The bandwidth parameter β is then used as part of the model parameters and sent to the embedded control unit.

5. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 2, characterized in that, For each historical monitoring data package, the total bacterial count, volatile basic nitrogen content, juice loss rate, and color parameters of the corresponding livestock and poultry meat samples were tested. The cloud server pre-sets the weight coefficient and acceptable threshold range of each quality indicator, and calculates the comprehensive quality score of each sample through a weighted algorithm. The optimal comprehensive quality indicator is determined by the highest comprehensive quality score and the fact that none of the individual indicators exceed their acceptable threshold range.

6. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 2, characterized in that, The embedded control unit pre-assigns a weighting coefficient to each sensor based on its installation location. The magnitude of the weighting coefficient is determined by the proximity of the sensor to the livestock and poultry meat product. Within each acquisition cycle, the embedded control unit multiplies the readings of each temperature sensor by their corresponding weighting coefficients and sums them to obtain a weighted temperature sum. Then, the weighted temperature sum is divided by the sum of all weighting coefficients to obtain real-time ambient temperature data that characterizes the overall thermal environment of the chamber. The same weighting operation is performed on the readings of each humidity sensor to obtain real-time ambient relative humidity data.

7. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 2, characterized in that, The embedded control unit also integrates a wireless communication module, which is used to establish communication connections with the cloud server and the user's mobile terminal respectively; The user mobile terminal is configured to: receive and visualize real-time ambient temperature data, real-time ambient relative humidity data, real-time electric field strength data, and target electric field strength setpoint sent by the embedded control unit; receive electric field strength adjustment commands manually set by the user based on the visualized information or confirm / modify the target electric field strength setpoint generated by the prediction model, and send the user commands or confirmed / modified setpoints to the embedded control unit, which then generates the final voltage control signal accordingly. The user's mobile terminal also synchronizes the user's historical operation data to the cloud server. The cloud server uses the historical operation data as incremental training samples to continuously optimize the prediction model.

8. The adaptively controlled intelligent dynamic electric field preservation device as described in claim 2, characterized in that, Multiple high-voltage electrodes in a plate-like structure are installed inside the preservation cavity; Multiple temperature sensors, humidity sensors, and electric field strength sensors are arranged inside the preservation chamber to monitor real-time ambient temperature data, real-time ambient relative humidity data, and real-time electric field strength data at different locations. Multiple high-voltage electrodes are grouped into multiple independent electrode groups. During the preservation process, the embedded control unit controls the multiple electrode groups to operate in an alternating mode. The embedded control unit calculates the electric field uniformity index in the preservation cavity based on the real-time electric field intensity data of each electric field intensity sensor. When the electric field uniformity index is lower than the preset threshold, the switching rate of the alternating mode is increased; when the electric field uniformity index reaches or exceeds the preset threshold, the switching rate is decreased.

Images