Intelligent detection and treatment device for peculiar smell of refrigerator

Abstract

The invention provides an intelligent refrigerator peculiar smell detecting and processing device, and belongs to the technical field of refrigerator production.The intelligent refrigerator peculiar smell detecting and processing device is characterized in that a multi-dimensional data acquisition system comprising a gas sensor array and a temperature and humidity sensor is constructed, and low-frequency trend components of sensor signals are extracted through wavelet transformation; a sparse coding-based dictionary learning and deep expansion network combined temperature drift compensation model is used to realize self-adaptive correction of a sensor baseline, and feature matrix construction and cosine similarity calculation are combined to complete peculiar smell type identification and concentration level quantification. And the power of the ultraviolet light emitting diode and the rotating speed of the fan are dynamically adjusted through a thermodynamic entropy increase minimization purification control algorithm to realize efficient purification, so that the technical problem of insufficient detection accuracy caused by serious baseline drift of the refrigerator odor sensor in a temperature fluctuation environment is solved.

Description

Technical Field

[0001] This invention belongs to the field of refrigerator manufacturing technology, and specifically relates to an intelligent odor detection and treatment device for refrigerators. Background Technology

[0002] Refrigerator odor detection devices typically employ metal-oxide-semiconductor (MOS) gas sensors to identify odor gases. These sensors respond to changes in resistance value by monitoring the concentration of the target gas. In practical applications, the internal temperature of a refrigerator fluctuates frequently due to door opening and closing, refrigeration cycles, and food storage and retrieval, typically ranging from -5°C to 25°C. Traditional temperature compensation methods mainly include polynomial fitting and lookup table correction. Polynomial fitting relies on manually defined function forms, making it difficult to accurately describe the complex nonlinear coupling relationship between sensor resistance and temperature and humidity. Lookup table correction requires pre-calibrating a large number of discrete points across the entire temperature range, demanding significant storage space and failing to address individual sensor differences and changes in characteristics caused by long-term aging. Because existing temperature compensation technologies fail to effectively capture the temperature-dependent characteristics of the sensor response, the sensor baseline drifts significantly with temperature fluctuations under actual refrigerator operating conditions. This makes it impossible to accurately distinguish between resistance fluctuations caused by ambient temperature changes and the true response signal caused by odor gases, resulting in high false alarm and false negative rates. In other words, existing technologies suffer from the technical problem of severe baseline drift in refrigerator odor sensors under fluctuating temperature environments, leading to insufficient detection accuracy. Summary of the Invention

[0003] In view of this, the present invention provides a refrigerator odor intelligent detection and processing device, which can solve the technical problem in the prior art where the baseline drift of the refrigerator odor sensor is serious under temperature fluctuation environment, resulting in insufficient detection accuracy.

[0004] The present invention is implemented as follows: The present invention provides a refrigerator odor intelligent detection and treatment device. The gas sensor array includes six tin dioxide-based semiconductor gas-sensitive elements uniformly distributed in a ring at the inlet of the refrigerator's internal circulating air duct. A temperature and humidity sensor is fixedly installed at the center of the gas sensor array to monitor the ambient temperature and relative humidity inside the refrigerator in real time. A photocatalytic purifier is set in the middle section of the refrigerator's internal air duct and consists of a titanium dioxide coated substrate and an ultraviolet light-emitting diode. A fan assembly is fixedly installed at the end of the circulating air duct and its speed is controlled by a frequency conversion drive circuit. The output of the power management unit is electrically connected to the gas sensor array, temperature and humidity sensor, photocatalytic purifier, and fan assembly to provide a 5V DC working voltage. The control chip is equipped with an adaptive baseline correction and entropy minimization control module to achieve coordinated control of each component.

[0005] The first input port group of the control chip is electrically connected to the six output signal lines of the gas sensor array, and the second input port of the control chip is electrically connected to the data output terminal of the temperature and humidity sensor.

[0006] The first output port of the control chip is electrically connected to the ultraviolet light-emitting diode through a pulse width modulation drive circuit, and the second output port of the control chip is electrically connected to the motor winding of the wind turbine assembly through a frequency conversion drive circuit.

[0007] The adaptive baseline correction and entropy minimization control module synchronously collects the resistance value of the gas-sensitive element and the ambient temperature and humidity values ​​to construct a multi-dimensional time-series data stream. It performs wavelet transform decomposition on the resistance value data to extract low-frequency trend components. The resistance value corresponding to the low-frequency trend components is combined with the ambient temperature and humidity values ​​to form an input feature vector, which is then input into the temperature drift compensation model to obtain the compensated standardized resistance value. The temperature drift compensation model adopts a dictionary learning and deep unfolding network architecture based on sparse coding. It realizes the sparse representation of temperature dependence through dictionary matrix transformation unit, nonlinear shrinkage function unit and residual connection unit. It calculates the rate of change of the compensated standardized resistance value to determine the odor detection status, extracts time-series features to construct a feature matrix and performs similarity calculation with the standard odor feature library to identify the odor type and pollutant concentration level. It calls the thermodynamic entropy minimization purification control algorithm to calculate the optimal power of the ultraviolet light-emitting diode and the optimal speed of the fan component to start the purification process.

[0008] The synchronous acquisition sampling frequency is 10 Hz, and wavelet transform decomposition filters out high-frequency noise components above 5 Hz, retaining low-frequency trend components for subsequent analysis.

[0009] Among them, the temperature drift compensation model expands the traditional iterative shrinkage threshold algorithm into a five-layer learnable network. Each learnable network layer contains a dictionary matrix transformation unit, a nonlinear shrinkage function unit, and a residual connection unit. The dictionary matrix dimension of the dictionary matrix transformation unit is 128 by 64.

[0010] The nonlinear shrinkage function unit uses a soft threshold function to set the components whose absolute value is less than the shrinkage threshold parameter to zero, and subtracts the sign value of the shrinkage threshold parameter from the components whose absolute value is greater than the shrinkage threshold parameter, thereby achieving feature sparsity and noise suppression.

[0011] The training dataset for the temperature drift compensation model was established in a standard temperature and humidity test chamber. The temperature was adjusted in 2-degree increments from -5 degrees Celsius to 25 degrees Celsius, and the relative humidity was adjusted in 10-degree increments from 30% to 90%. 300 sets of baseline data were collected at each operating point, for a total of 42,000 sets of baseline data.

[0012] The training dataset also includes 168,000 sets of labeled data. The labeled data are introduced with four odorous gases, namely ethanol, trimethylamine, hydrogen sulfide and ammonia, under the same temperature and humidity combination working condition matrix. The concentration gradient is set to 5ppm to 50ppm, and 200 sets of response data are collected at each concentration point.

[0013] The temperature drift compensation model training uses mean squared error as the loss function and an adaptive moment estimation optimization algorithm for parameter updates. The initial learning rate is 0.001, which decays to 0.9 times the original value every 20 rounds. The batch size is 64, and the total number of training rounds is 200.

[0014] Furthermore, when calculating the rate of change of the standardized resistance value after compensation, if the absolute value of the rate of change of the standardized resistance value after compensation of any gas-sensitive element exceeds the baseline threshold of 0.15, it is determined that there is interference from odorous gas and the odor detection state is triggered.

[0015] Among them, the time-series features of the compensated standardized resistance values ​​of six gas-sensitive elements within 5 consecutive seconds under odor detection conditions include mean, variance, peak time and decay rate, and a feature matrix is ​​constructed.

[0016] The similarity calculation between the feature matrix and the standard odor feature library uses cosine similarity to identify odor types and quantify pollutant concentration levels into three levels: mild, moderate, and severe.

[0017] Among them, the thermodynamic entropy increase minimization purification control algorithm abstracts the odor purification process into an irreversible thermodynamic system. With minimizing the entropy generation rate per unit time as the optimization objective, the continuous time is discretized into a decision sequence using dynamic programming method. The time step is 5 seconds, the optimal power of the ultraviolet light-emitting diode is divided into 5 levels, and the optimal speed of the fan component is divided into 3 levels.

[0018] During the purification process, the rate of change of the standardized resistance value after compensation of the six gas-sensitive elements is continuously monitored. When the absolute value of the rate of change of the standardized resistance value after compensation of all gas-sensitive elements is lower than the recovery threshold of 0.05 and the duration exceeds 10 seconds, the purification is deemed complete.

[0019] After purification, the optimal power of the ultraviolet light-emitting diode is reduced to a maintenance power of 3 watts, and the optimal speed of the fan assembly is adjusted to a low-speed circulation speed of 30% of the rated speed.

[0020] This invention employs a combination of dictionary learning based on sparse coding and deep unfolded networks to construct a temperature drift compensation model. Dictionary learning automatically discovers sparse basis functions representing the changes in resistance values ​​of gas-sensitive elements with ambient temperature and relative humidity. Compared to manually designed polynomial models, this captures more complex nonlinear coupling relationships. The deep unfolded network embeds iterative optimization algorithms into the neural network architecture, allowing parameters for each layer to be obtained through backpropagation training, avoiding the need for manual adjustment of compensation parameters in traditional methods. The sparse coding mechanism forces the model to use a small number of key features to express temperature dependence, reducing the number of parameters and computational complexity, enabling the control chip to complete real-time compensation in milliseconds. Simultaneously, the dictionary matrix adaptively learns individual sensor differences and long-term drift characteristics through end-to-end training, enhancing model robustness and accurately reproducing the sensor's true gas response signal under different temperature and humidity conditions. This invention combines physics-driven sparse priors with data-driven end-to-end learning, achieving practical performance requirements while ensuring theoretical interpretability. By eliminating the interference of temperature fluctuations on the sensor baseline, it achieves accurate detection of odor gases, solving the technical problem of insufficient detection accuracy in existing refrigerator odor sensors due to severe baseline drift under fluctuating temperature environments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the device of the present invention.

[0022] Figure 2 The graph shows the time-series variation of the standardized resistance values ​​of the six gas-sensitive elements after compensation.

[0023] Figure 3 This is a graph showing the convergence curve of the loss function during the training of the temperature drift compensation model.

[0024] Figure 4 This is a curve showing the dynamic adjustment of ultraviolet light-emitting diode power and fan speed during the odor purification process.

[0025] Figure 5 This is a graph showing the exponential decay of system entropy over purification time. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a schematic representation of the components of an intelligent refrigerator odor detection and processing device provided by the present invention, including: a gas sensor array, a temperature and humidity sensor, a photocatalytic purifier, a fan assembly, a power management unit, and a control chip. The gas sensor array comprises six... A semiconductor gas-sensitive element is evenly distributed in a ring at the inlet of the refrigerator's internal circulating air duct. A temperature and humidity sensor is fixedly installed at the center of the gas sensor array to monitor the refrigerator's internal ambient temperature and relative humidity in real time. A photocatalytic purifier is located in the middle section of the refrigerator's internal air duct and consists of a titanium dioxide-coated substrate and ultraviolet light-emitting diodes (UV LEDs), with the UV LEDs embedded evenly on the surface of the titanium dioxide-coated substrate. A fan assembly is fixedly installed at the end of the circulating air duct, and its speed is controlled by a frequency converter drive circuit. A power management unit is located in the refrigerator's control compartment, and its output is electrically connected to the gas sensor array, the temperature and humidity sensor, the photocatalytic purifier, and the fan assembly, providing 5V power. The control chip operates at a DC voltage of V. It is mounted on the main control circuit board within the control compartment. The first input port group of the control chip is electrically connected to the six output signal lines of the gas sensor array. The second input port of the control chip is electrically connected to the data output terminal of the temperature and humidity sensor. The first output port of the control chip is electrically connected to the ultraviolet light-emitting diode via a pulse width modulation drive circuit. The second output port of the control chip is electrically connected to the motor windings of the fan assembly via the frequency converter drive circuit, thus achieving coordinated control of all components. The control chip incorporates an adaptive baseline correction and entropy minimization control module to eliminate the influence of temperature fluctuations on sensor response and optimize the energy conversion efficiency of the odor purification process.

[0028] The adaptive baseline correction and entropy minimization control module is used to perform the following steps: S01. Simultaneously acquire the resistance values ​​of the six gas-sensitive elements and the ambient temperature and relative humidity values ​​output by the temperature and humidity sensor at a sampling frequency of 10Hz to construct a multi-dimensional time-series data stream; S02. Perform wavelet transform decomposition on the resistance value data in the multidimensional time series data stream, extract low-frequency trend components and high-frequency fluctuation components, retain the low-frequency trend components for subsequent analysis, and filter out high-frequency noise components higher than 5Hz. S03. Combine the resistance value corresponding to the low-frequency trend component with the ambient temperature and relative humidity values ​​at the corresponding time to form an input feature vector, and input it into the pre-trained temperature drift compensation model to obtain the compensated standardized resistance value. S04. Calculate the rate of change of the standardized resistance value after compensation of the six gas-sensitive elements. When the absolute value of the rate of change of the standardized resistance value after compensation of any gas-sensitive element exceeds the baseline threshold of 0.15, it is determined that there is odor gas interference and the odor detection state is triggered. S05. Under the odor detection state, extract the time-series characteristics of the compensated standardized resistance values ​​of the six gas-sensitive elements within 5 consecutive seconds, including the mean, variance, peak time and decay rate, and construct a feature matrix. S06. Perform cosine similarity calculation between the feature matrix and the pre-stored standard odor feature library to identify the odor type and quantify the pollutant concentration level, which is divided into three levels: mild, moderate and severe. S07. Based on the identified pollutant concentration level and the current system entropy value, invoke the thermodynamic entropy increase minimization purification control algorithm to calculate the optimal power of the ultraviolet light-emitting diode and the optimal speed of the fan assembly. S08. The optimal power is converted into a pulse width modulation duty cycle signal and output to the pulse width modulation drive circuit; the optimal speed is converted into a frequency conversion control signal and output to the frequency conversion drive circuit; and the purification process is started. S09. During the purification process, continuously monitor the rate of change of the standardized resistance value after compensation of the six gas-sensitive elements. When the absolute value of the rate of change of the standardized resistance value after compensation of all gas-sensitive elements is lower than the recovery threshold of 0.05 and the duration exceeds 10 seconds, determine that the purification is completed, reduce the optimal power of the ultraviolet light-emitting diode to the maintenance power, and adjust the optimal speed of the fan assembly to the low-speed circulation speed.

[0029] The specific structure of the temperature drift compensation model is as follows: The temperature drift compensation model adopts a combined architecture of dictionary learning based on sparse coding and deep unfolded network. The traditional iterative shrinking threshold algorithm is unfolded into five learnable network layers. Each learnable network layer contains a dictionary matrix transformation unit, a nonlinear shrinking function unit, and a residual connection unit. The dictionary matrix transformation unit has a dictionary matrix dimension of 128×64 and is used to map the input feature vector to the sparse representation space. The nonlinear shrinking function unit uses a soft thresholding function to achieve feature selection. The residual connection unit retains the original input feature vector information to prevent gradient vanishing. The final output layer of the temperature drift compensation model is a fully connected layer, which maps the sparse coding in the sparse representation space to the compensated standardized resistance value. The total number of parameters of the temperature drift compensation model is approximately 87,000. The steps for establishing the training dataset of the temperature drift compensation model specifically include: exposing the gas sensor array to a clean air environment in a standard temperature and humidity test chamber, adjusting the temperature from -5℃ to 25℃ in 2℃ increments, maintaining each temperature point for 30 minutes and then recording the resistance values ​​of the six gas-sensitive elements, while simultaneously adjusting the relative humidity from 30% to 90% in 10% increments to form a temperature and humidity combination working condition matrix, collecting 300 sets of baseline data at each working condition point, for a total of 42,000 sets of baseline data; subsequently, under the same temperature and humidity combination working condition matrix, introducing four typical odor gases—ethanol, trimethylamine, hydrogen sulfide, and ammonia—with concentration gradients set from 5ppm to 50ppm, collecting 200 sets of response data at each concentration point, for a total of 168,000 sets of labeled data, and dividing the 42,000 sets of baseline data and the 168,000 sets of labeled data into a training set and a validation set in a 7:3 ratio; The specific steps for training the temperature drift compensation model include: initializing the dictionary matrix of the dictionary matrix transformation unit to a Gaussian random distribution with a standard deviation of 0.01, initializing the shrinkage threshold parameter of the nonlinear shrinkage function unit to 0.1, using mean squared error as the loss function, and optimizing to minimize the error between the predicted resistance value and the actual compensation value; updating parameters using an adaptive moment estimation optimization algorithm with an initial learning rate of 0.001, decaying to 0.9 times the original value every 20 rounds, a batch size of 64, and a total of 200 training rounds; introducing a dictionary sparsity regularization term with a weight coefficient of 0.001 during training to constrain the dictionary matrix of the dictionary matrix transformation unit to maintain its sparsity characteristics and improve the generalization ability of the temperature drift compensation model; evaluating the performance of the temperature drift compensation model on the validation set every 10 rounds, and terminating training early when the validation error no longer decreases after 30 consecutive rounds, saving the parameters of the temperature drift compensation model with the minimum validation error.

[0030] The principle of combining dictionary learning based on sparse coding with deep unfolded networks lies in mapping the iterative process of traditional optimization algorithms to the forward propagation process of neural networks. Each learnable network layer corresponds to one iteration update of the iterative algorithm. Through end-to-end training, the dictionary matrix of the dictionary matrix transformation unit and the shrinkage threshold parameter of the nonlinear shrinkage function unit are optimized simultaneously, enabling the temperature drift compensation model to possess the expressive power of deep learning while maintaining mathematical interpretability. In specific implementation, the input feature vector is multiplied by the dictionary matrix of the dictionary matrix transformation unit to obtain the initial sparse code. The soft threshold function of the nonlinear shrinkage function unit achieves feature selection and denoising. Then, the signal is reconstructed by transposing the dictionary matrix of the dictionary matrix transformation unit. The residual connection unit ensures information flow, and multiple learnable network layers are cascaded to form a deep network. The dictionary learning enables the temperature drift compensation model to adaptively extract temperature-related intrinsic patterns, and the deep unfolding ensures the convergence and stability of the optimization process.

[0031] The combination of sparse coding-based dictionary learning and deep unfolded network brings three main technical benefits to temperature drift compensation: First, the dictionary learning automatically discovers the sparse representation basis function of the resistance value of the gas-sensitive element as a function of the ambient temperature and relative humidity. Compared with manually designed polynomial or exponential models, this can capture more complex nonlinear relationships and significantly improve the compensation accuracy under different temperature and humidity conditions. Second, the deep unfolded network embeds an iterative optimization algorithm into the neural network architecture. The parameters of each learnable network layer are obtained through backpropagation training, avoiding the need for manual step adjustment required by traditional iterative algorithms. The method addresses the issues of length and threshold, enhancing the robustness of the temperature drift compensation model to individual sensor differences and long-term drift. Third, the sparse coding mechanism forces the temperature drift compensation model to use a small number of key features to express complex temperature dependencies, reducing the number of parameters and computational complexity of the model. This enables the control chip to complete real-time compensation in milliseconds, meeting the requirements for rapid response. Overall, the method combines physics-driven sparse priors with data-driven end-to-end learning, ensuring both theoretical interpretability and practical performance requirements, providing an effective way to solve the sensor baseline drift problem caused by temperature fluctuations.

[0032] The principle of the thermodynamic entropy increase minimization purification control algorithm is as follows: The odor purification process is abstracted as an irreversible thermodynamic system. The initial state entropy value of this irreversible thermodynamic system is determined by the pollutant concentration level and the degree of disorder in its distribution. The purification target state is an ordered state where the pollutant concentration level drops below a threshold. The odor purification process is accompanied by ultraviolet photocatalytic reaction and forced air convection, inevitably resulting in entropy increase. The thermodynamic entropy increase minimization purification control algorithm optimizes the entropy generation rate per unit time. Under the constraints of purification time and final state concentration, it solves for the optimal time series of the optimal power of the ultraviolet light-emitting diode and the optimal speed of the fan assembly. In specific implementation, firstly, based on the current pollutant concentration level and the response characteristics of the feature matrix, the initial state entropy value of the irreversible thermodynamic system is estimated. Then, a state transition equation is established to describe the... The evolution of the initial entropy of the irreversible thermodynamic system under the same control parameters over time is studied. The state transition equation considers three factors: photocatalytic reaction rate, airflow mixing efficiency, and energy dissipation. The continuous-time problem is discretized into a decision sequence using dynamic programming, with a time step of 5 seconds. The optimal power of the ultraviolet light-emitting diode (UVLED) is divided into 5 levels, and the optimal speed of the fan assembly is divided into 3 levels, constructing 15 combinations of control actions. The optimal control action is calculated step by step from the purification target state using the Bellman optimality principle, so that the cumulative entropy from the current initial entropy of the irreversible thermodynamic system to the purification target state reaches the minimum value. At the same time, the optimal power of the UVLED is constrained not to exceed the rated power of 15W, the power of the fan assembly is constrained not to exceed 8W, and the purification time is constrained not to exceed 300 seconds.

[0033] The thermodynamic entropy increase minimization purification control algorithm brings a technical advantage to the device by establishing an essential relationship between energy conversion and pollutant degradation. Traditional control strategies typically operate at a fixed power until the concentration reaches the target, ignoring the physical law that the reaction rate and mass transfer efficiency decrease with the decrease in pollutant concentration during the odor purification process, resulting in low energy utilization efficiency in the later stages. The thermodynamic entropy increase minimization purification control algorithm, based on the thermodynamic framework of irreversible processes, separates the effective work and dissipated work in the odor purification process. By adjusting the optimal power of the ultraviolet light-emitting diode and the optimal speed of the fan assembly to dynamically match the current pollutant concentration level, the photocatalytic reaction rate is always kept in the high-efficiency range, avoiding excessive excitation. The resulting loss of ineffective photons and prolonged purification time due to weak excitation are addressed by the dynamic programming method, which ensures that the combination of control actions satisfies global optimality. Compared with greedy strategies or heuristic methods, it can anticipate future state changes and adjust control parameters in advance, avoiding getting trapped in local optima. Furthermore, the thermodynamic entropy increase minimization purification control algorithm explicitly incorporates power and time constraints into the optimization problem, automatically seeking the Pareto optimal solution for energy consumption and time while ensuring purification effect, adapting to the needs and preferences of different application scenarios. From a system perspective, the thermodynamic entropy increase minimization purification control algorithm reveals the thermodynamic essence of the odor purification process, providing theoretical guidance and computational framework for the control chip, and improving the intelligence level and practical value of the device.

[0034] The soft thresholding function achieves feature sparsity and noise suppression by setting the components of the input features whose absolute value is less than the shrinkage threshold parameter to zero, and subtracting the sign value of the shrinkage threshold parameter from the components whose absolute value is greater than the shrinkage threshold parameter.

[0035] The wavelet transform decomposition projects the resistance value data in the multidimensional time-series data stream onto wavelet basis functions of different scales, separating the low-frequency trend and high-frequency details in the resistance value data. The low-frequency trend component reflects the main changing trend of the sensor response, while the high-frequency fluctuation component includes environmental noise and random disturbances.

[0036] The cosine similarity calculation is performed by taking the inner product of the normalized feature matrix and the vectors in the standard odor feature library to measure the similarity of the vectors in the direction. The value ranges from -1 to 1, and the closer the value is to 1, the higher the similarity.

[0037] The baseline threshold is determined based on the statistical analysis of the resistance stability of the gas-sensitive element in clean air, and is set to three standard deviations of the normal fluctuation range to distinguish between environmental interference and real odor signals.

[0038] The recovery threshold is set to one-third of the baseline threshold to determine whether the purification treatment has achieved the expected effect. The smaller recovery threshold ensures that the device will exit the odor detection state only after completely eliminating odor residue.

[0039] The maintenance power is set to 20% of the rated power of the ultraviolet light-emitting diode, i.e., 3W, to maintain the basic activity of the photocatalytic purifier after the purification is completed and prevent the rapid accumulation of odorous gases.

[0040] The low-speed circulation speed is set to 30% of the rated speed of the fan assembly, which is used to maintain a slow airflow inside the refrigerator after the purification is completed, and to promote uniform temperature distribution.

[0041] The current system entropy value is calculated by multiplying the concentration value corresponding to the pollutant concentration level by the Boltzmann constant, and then multiplying the concentration value by the natural logarithm base to the power of the concentration value. The Boltzmann constant is set to a value of... .

[0042] The number of rows (128) in the dictionary matrix dimension of the dictionary matrix transformation unit is determined based on the number of channels in the gas sensor array and the sampling frequency. The input feature vector contains eight parameters: the resistance values ​​of the six gas-sensitive elements, the ambient temperature value, and the relative humidity value. To fully extract temporal correlation, data from 16 consecutive sampling times are concatenated to form a 128-dimensional input feature vector, corresponding to the number of rows in the dictionary matrix of the dictionary matrix transformation unit. The number of columns (64) in the dictionary matrix dimension of the dictionary matrix transformation unit represents the dimension of the sparse representation space. The number of columns (64) has been experimentally verified to achieve a balance between the complexity and expressive power of the temperature drift compensation model. The initial value of the shrinkage threshold parameter of the nonlinear shrinkage function unit is adaptively set according to the amplitude range of the input feature vector in the training dataset. Specifically, it is the mean of the standard deviation of each component of the input feature vector multiplied by 0.1. During training, the shrinkage threshold parameter decreases with the depth of the learnable network layer, allowing shallow layers to retain more original input feature vector information and deeper layers to achieve stronger feature selection.

[0043] The specific implementation methods of the above steps are described in detail below.

[0044] The specific implementation of step S01 is as follows: The control chip synchronously reads the resistance values ​​of the six gas-sensitive elements at a sampling frequency of 10Hz through the first input port group. The resistance value of each gas-sensitive element is converted into an analog voltage signal through a voltage divider circuit, and then converted into a 12-bit digital signal by an analog-to-digital converter and stored in the data buffer. At the same time, the control chip reads the ambient temperature and relative humidity values ​​output by the temperature and humidity sensor through the second input port. The temperature and humidity sensor directly outputs digital quantities using a digital communication protocol. The sampling period is synchronized with the gas-sensitive elements. The six resistance values, one ambient temperature value, and one relative humidity value at each sampling moment are combined into an eight-dimensional data vector and arranged in chronological order to form a multi-dimensional time-series data stream. The data buffer uses a circular queue structure to store the data of the most recent 160 sampling moments, corresponding to a 16-second historical window, for subsequent time-series feature extraction. The synchronous acquisition mechanism is based on hardware timer interrupt triggering to ensure that the time alignment error of each channel data is less than 1 millisecond.

[0045] The specific implementation of step S02 is as follows: Read the resistance value data of the most recent 80 sampling times (i.e., 8 seconds) from the data buffer, perform wavelet transform decomposition on the resistance value sequence of each gas-sensitive element, and use Daubechies wavelet basis function for three-level decomposition. The first level of decomposition divides the signal into a low-frequency component of 0 to 5 Hz and a high-frequency component of 5 to 10 Hz. The low-frequency component is retained as a low-frequency trend component, which reflects the true response characteristics of the sensor to odor gas. The high-frequency component is discarded as a high-frequency noise component, which mainly comes from circuit thermal noise and electromagnetic interference. Wavelet transform realizes the joint characterization of the signal in the time and frequency domains through multi-scale analysis. Compared with Fourier transform, it can better preserve the local time domain characteristics of the signal. The number of sampling points of the decomposed low-frequency trend component is halved to 40 points, reducing the subsequent calculation burden. The low-frequency trend components of the six gas-sensitive elements are reorganized into six signal sequences according to channels and stored in the temporary storage area.

[0046] The specific implementation of step S03 is as follows: Read the resistance values ​​corresponding to the six low-frequency trend components of the most recent 16 sampling times from the temporary storage area, totaling 96 values. Simultaneously, read 16 ambient temperature values ​​and 16 relative humidity values ​​for the corresponding times. Concatenate the 96 resistance values, 16 ambient temperature values, and 16 relative humidity values ​​in chronological order into a 128-dimensional input feature vector. This input feature vector is input into the first learnable network layer of the temperature drift compensation model. After a 128×64-dimensional dictionary matrix linear transformation by the dictionary matrix transformation unit, a 64-dimensional intermediate representation is obtained. This intermediate representation is then processed by the nonlinear... The soft threshold function processing of the shrinkage function unit sets the components with absolute values ​​less than the current shrinkage threshold parameter to zero to achieve sparsity. Then, the linear mapping of the original input feature vector and the sparsification result are added through the residual connection unit and passed to the next learnable network layer. After five layers of cascaded processing, the fully connected layer of the final output layer of the network outputs six compensated standardized resistance values, which correspond to the standardized response values ​​of the six gas-sensitive elements to eliminate the influence of temperature drift under the current temperature and humidity conditions. The temperature drift compensation model uses sparse coding to automatically learn the nonlinear mapping relationship related to temperature, avoiding the overfitting problem of the traditional polynomial fitting model.

[0047] The specific implementation of step S04 is as follows: Calculate the rate of change of the six compensated standardized resistance values ​​relative to the value at the previous sampling time. The rate of change is equal to the current value minus the previous value, then divided by the previous value, to obtain six rate of change values. Determine whether the absolute value of each rate of change value exceeds the baseline threshold of 0.15. The baseline threshold is determined based on the statistical characteristics of the resistance fluctuation of the gas-sensitive element in a clean air environment, taking the value corresponding to three times the standard deviation. When the absolute value of the rate of change of any gas-sensitive element exceeds the baseline threshold, it is determined that there is odor gas interference in the refrigerator. Set the system status flag to odor detection state, and record the trigger time as the start time of the odor event. The rate of change calculation method is not sensitive to slow temperature drift and can effectively identify rapid resistance changes caused by odor gas. If the absolute values ​​of all rates of change do not exceed the baseline threshold, the system remains in normal monitoring state, and the data acquisition in step S01 continues.

[0048] The specific implementation of step S05 is as follows: Under the odor detection state, compensated standardized resistance value data are continuously collected for 5 seconds from the start time of the odor event, corresponding to 50 sampling points. Four types of time-series features, namely mean, variance, peak time, and decay rate, are calculated for the data sequences of the six gas-sensitive elements. The mean reflects the overall response intensity, the variance reflects the response stability, the peak time reflects the time difference of gas diffusion to the sensor, and the decay rate reflects the rate at which the gas concentration decreases with ventilation. The four types of features of the six gas-sensitive elements, totaling 24 values, are organized into a 6×4-dimensional feature matrix. The rows of the feature matrix correspond to different gas-sensitive elements, and the columns correspond to different feature types. The time-series feature extraction is based on statistical analysis methods, which can compress the original time-series signal into a low-dimensional feature representation, facilitating subsequent pattern recognition. The 5-second window length has been verified to fully capture the dynamic response process of the odor gas.

[0049] The specific implementation of step S06 is as follows: Standard feature matrices corresponding to four typical odor gases—ethanol, trimethylamine, hydrogen sulfide, and ammonia—are read from a pre-stored standard odor feature library. The standard feature matrix for each gas has a dimension of 6×4. Cosine similarity calculations are performed on the feature matrix and each of the four standard feature matrices. First, the feature matrix and each standard feature matrix are expanded into 24-dimensional vectors. The inner product of the two vectors is calculated and divided by the product of their magnitudes to obtain four similarity values. The gas type corresponding to the maximum similarity value is taken as the identification result. The pollutant concentration level is quantified based on the similarity value: a similarity greater than 0.9 indicates heavy pollution; a similarity between 0.7 and 0.9 indicates moderate pollution; and a similarity between 0.5 and 0.7 indicates light pollution. The cosine similarity calculation is based on the vector angle, which is invariant to feature amplitude scaling and is suitable for handling response pattern matching problems at different concentrations.

[0050] The specific implementation of step S07 is as follows: The current system entropy value is determined based on the pollutant concentration level identified in step S06. This current system entropy value is obtained by multiplying the concentration value corresponding to the pollutant concentration level by the Boltzmann constant. The concentration value is then multiplied by the natural logarithm base to a power of the concentration value, where 10 ppm corresponds to light pollution, 30 ppm to moderate pollution, and 50 ppm to heavy pollution. A thermodynamic entropy minimization purification control algorithm is used, with the current system entropy value as the initial state. The target entropy value corresponding to the purification target state is set to the entropy value corresponding to a concentration of 5 ppm. A state transition equation is established to describe the evolution of system entropy over time. This state transition equation includes photocatalytic reaction terms, airflow mixing terms, and energy dissipation terms. A dynamic programming method is used to discretize the 300-second purification time into 60 time steps. Each time step has a duration of 5 seconds. At each time step, five optimal power values ​​(3W, 6W, 9W, 12W, 15W) for the ultraviolet light-emitting diode and three optimal speed values ​​(low, medium, high) for the wind turbine component are enumerated, resulting in 15 possible control action combinations. Using the Bellman optimality principle, the optimal control action and the minimum cumulative entropy are calculated from the target state to the initial state. The time series of the optimal power of the ultraviolet light-emitting diode and the optimal speed of the wind turbine component are obtained. The dynamic programming algorithm guarantees global optimality and avoids the local optima problem of greedy algorithms.

[0051] The specific implementation of step S08 is as follows: The optimal power value of the ultraviolet light-emitting diode (UVLED) obtained in step S07 is read, and the optimal power value is converted into a pulse width modulation (PWM) duty cycle. The duty cycle is equal to the optimal power value divided by the rated power of 15W. The PWM duty cycle signal is output to the PWM drive circuit through the first output port of the control chip. The PWM drive circuit adjusts the on-time ratio of the UVLED according to the duty cycle to achieve continuously adjustable power. Simultaneously, the optimal speed setting of the fan assembly is read, and the setting signal is converted into a frequency conversion control signal. Low speed corresponds to a frequency of 15Hz, medium speed to 30Hz, and high speed to 50Hz. The frequency conversion control signal is output to the frequency conversion drive circuit through the second output port of the control chip. The frequency conversion drive circuit adjusts the motor power supply frequency to control the speed, starting the photocatalytic purifier and the fan assembly to work together. The PWM technology and frequency conversion control technology achieve precise energy regulation, meeting the optimization requirements of the thermodynamic entropy minimization purification control algorithm.

[0052] The specific implementation of step S09 is as follows: After executing step S08 to start the purification process, the monitoring process from step S01 to step S04 is continuously executed. The rate of change of the standardized resistance value after compensation of the six gas-sensitive elements is calculated in real time. It is determined whether the absolute value of the rate of change of all gas-sensitive elements is lower than the recovery threshold of 0.05. The recovery threshold is set to one-third of the baseline threshold of 0.15 to ensure the strictness of the purification effect judgment. When the absolute value of all rates of change is lower than the recovery threshold for more than 10 seconds, it is determined that the odor gas has been fully purified. The optimal power of the ultraviolet light-emitting diode is reduced to the maintenance power of 3W. The maintenance power is 20% of the rated power of 15W to maintain the surface activity of the photocatalyst. The optimal speed of the fan assembly is adjusted to the low-speed circulation speed, which is 30% of the rated speed, to maintain slow air circulation. The system status flag is cleared to exit the odor detection state and return to the normal monitoring mode to continue executing step S01. The duration determination avoids misjudgment caused by instantaneous fluctuations. The maintenance power and low-speed circulation speed ensure the persistence of the purification effect.

[0053] It should be noted that the key technical ideas of this invention include a temperature drift compensation method based on a combination of dictionary learning and deep unfolded networks using sparse coding, and a purification control algorithm based on the second law of thermodynamics that minimizes entropy increase. The temperature drift compensation method unfolds the traditional iterative optimization algorithm into a learnable neural network layer. Through end-to-end training, it simultaneously optimizes the dictionary matrix and shrinkage threshold, enabling the model to adaptively learn the sparse representation basis function of the gas-sensitive element's resistance as a function of temperature and humidity. Compared to traditional polynomial or lookup table compensation methods, it can capture more complex nonlinear temperature dependencies. Simultaneously, the sparse coding mechanism forces the model to use a small number of key features to express the temperature effect, reducing the number of parameters and computational complexity. This allows the embedded control chip to complete the compensation calculation in real time, solving the technical problem of severe sensor baseline drift under frequent temperature fluctuations in refrigerators. The entropy minimization purification control algorithm, based on the thermodynamics of irreversible processes, views the odor purification process as a transformation from a disordered, high-entropy state to an ordered, low-entropy state. It optimizes by minimizing the entropy generation rate per unit time, using dynamic programming to solve for the globally optimal time series of photocatalytic power and fan speed. Compared to traditional fixed-power control strategies, this algorithm dynamically matches the current pollutant concentration to adjust the purification intensity, ensuring the photocatalytic reaction always operates within its high-efficiency range. This avoids energy waste from excessive excitation and prolonged time from insufficient excitation, significantly improving energy utilization efficiency. The synergistic effect of these two approaches lies in the temperature drift compensation method, which provides accurate sensor response signals, eliminating temperature interference and providing reliable input for pollutant concentration identification. The entropy minimization purification control algorithm formulates optimized control strategies based on the accurately identified concentration levels. Together, they form a complete closed loop from perception to decision-making, ensuring both detection accuracy and high purification efficiency. Overall, this provides a theoretically rigorous and engineering-feasible solution for intelligent odor control in refrigerators.

[0054] It should be noted that the specific hardware implementation of the refrigerator odor intelligent detection and processing device is as follows: the gas sensor array includes six... The refrigerator employs a semiconductor gas-sensitive element, with each element utilizing a micro-heater structure. The heater power is 0.5W, and the operating temperature is stable at 300℃. The six gas-sensitive elements are arranged in a circular array with a center-to-center spacing of 50mm, fixed to a PTFE mounting bracket. This bracket is bolted to a metal frame at the inlet of the refrigerator's internal air circulation duct. Each gas-sensitive element's two electrode pins are connected to a voltage divider circuit, which consists of a precision resistor and an operational amplifier, converting resistance changes into a 0-5V voltage signal output. The temperature and humidity sensor uses a digital sensor chip with a 3mm×3mm package, fixedly mounted at the geometric center of the mounting bracket. It connects to the second input port of the control chip via two data lines, transmitting data using a serial communication protocol. The measurement range is temperature -40℃ to 80℃ and humidity 0-100%, with a measurement accuracy of temperature ±0.5℃ and humidity ±3%. The titanium dioxide coated substrate of the photocatalytic purifier is made of aluminum alloy, measuring 200mm×150mm with a thickness of 2mm, and coated with a nano-scale coating. The photocatalyst has a coating thickness of 10μm. Twelve ultraviolet (UV) light-emitting diodes (LEDs), each with a rated power of 1.25W and a peak emission wavelength of 365nm, are uniformly embedded on the surface of the titanium dioxide-coated substrate. The spacing between adjacent LEDs is 40mm. These twelve UV LEDs are connected in parallel and then connected to the pulse width modulation (PWM) drive circuit via wires. The fan assembly is driven by a brushless DC motor with a rated voltage of 12V, a rated power of 8W, and a rated speed of 2000rpm. The fan impeller has a diameter of 80mm and seven blades. The three-phase leads of the motor windings are connected to the output of the frequency converter drive circuit via connectors. The power management unit includes an AC-to-DC conversion module and a multi-channel regulated output module. The input is connected to a 220V AC power supply from the refrigerator. It outputs 5V DC to power the control chip, the gas sensor array, and the temperature and humidity sensor, and 12V DC to power the fan assembly and the UV LEDs. Each output port is equipped with an overcurrent protection circuit. The control chip is a 32-bit microcontroller with a main frequency of 168MHz. It integrates a 12-bit analog-to-digital converter, a pulse width modulation output module, and a communication interface on the chip. It is packaged on the main control circuit board, which is 100mm×80mm in size and uses a four-layer printed circuit board process to lay out signal and power traces. The main control circuit board is fixedly installed in the mounting slot in the control compartment.

[0055] It should be noted that this invention also solves the following technical problem: Existing odor purification systems typically employ a control strategy of continuous operation at a fixed power, failing to consider the physical law that the photocatalytic reaction rate and mass transfer efficiency decrease with decreasing pollutant concentration during the purification process. This leads to low energy utilization efficiency and high overall energy consumption in the later stages of purification. This invention abstracts the odor purification process as an irreversible thermodynamic system, optimizing by minimizing the entropy generation rate per unit time. It uses dynamic programming to solve for the optimal time series of ultraviolet light-emitting diode power and fan speed, ensuring the photocatalytic reaction rate remains within the high-efficiency range. By dynamically matching control parameters with the current pollutant concentration level, it avoids ineffective photon loss due to excessive excitation and prolonged purification time due to insufficient excitation. It establishes a correlation between energy conversion and pollutant degradation at the thermodynamic level, achieving a globally optimal balance between purification effect and energy consumption.

[0056] Specifically, the principle of this invention is as follows: the essence of sensor baseline drift caused by temperature fluctuations is that the oxygen adsorption and desorption processes on the sensor surface change nonlinearly with temperature, leading to a nonlinear change in resistance. This change differs significantly from the gas response signal in its time-domain characteristics; temperature drift exhibits a slow, low-frequency trend, while the gas response shows a rapid, high-frequency abrupt change. This invention first decomposes the sensor signal into low-frequency trend components and high-frequency fluctuation components using wavelet transform, retaining the low-frequency components for temperature compensation analysis. Then, dictionary learning is used to automatically discover the inherent sparse representation patterns related to temperature in a high-dimensional feature space. The deep unfolded network maps each iteration of the traditional iterative threshold shrinkage algorithm to a layer of neural network, achieving end-to-end transformation from the original sensor signal to the compensated, standardized resistance value through multi-layer cascading. The dictionary matrix learns the basis functions corresponding to different temperature and humidity combinations during training. The nonlinear shrinkage function achieves feature selection and noise suppression, and residual connections ensure information flow in the deep network. Because dictionary learning can express complex temperature dependencies with a sparse combination of a small number of basis functions, the model parameters are controlled within a reasonable range, enabling the embedded control chip to perform real-time calculations. At the same time, the mathematical structure of sparse representation ensures the model's adaptability to individual sensor differences and long-term drift, thereby achieving accurate separation of temperature interference and gas response signals in temperature fluctuation environments, thus improving the accuracy of odor detection.

[0057] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0058] The specific implementation of step S01 is as follows: the control chip synchronously acquires six data points at a sampling frequency of 10Hz. The resistance value of the basic semiconductor gas-sensitive element and the ambient temperature and relative humidity values ​​output by the temperature and humidity sensor are used to construct a multi-dimensional time-series data stream. The resistance value of the gas-sensitive element is denoted as... The ambient temperature value is recorded as The relative humidity value is recorded as ,in This indicates the number of the six gas-sensitive elements. Indicates the sampling time, in seconds. A multidimensional time-series data stream is represented as... In the formula, for An eight-dimensional data vector at any given time; For the first Each gas-sensitive element The resistance value at time t, in units of This is the ambient temperature value, in units of... The value represents relative humidity, expressed as a percentage. Step S02 is implemented by performing wavelet transform decomposition on the resistance data in the multidimensional time-series data stream, extracting low-frequency trend components and high-frequency fluctuation components, retaining the low-frequency trend components for subsequent analysis, and filtering out high-frequency noise components above 5Hz. The wavelet transform decomposition uses discrete wavelet transform to extract the resistance data. Projecting onto wavelet basis functions of different scales separates the low-frequency trend and high-frequency details in the resistance data. The low-frequency trend component reflects the main changing trend of the sensor response, while the high-frequency fluctuation component contains environmental noise and random disturbances. The low-frequency trend component obtained after wavelet decomposition is denoted as... The unit is The specific implementation of step S03 is as follows: The resistance value corresponding to the low-frequency trend component is combined with the corresponding ambient temperature and relative humidity values ​​to form an input feature vector. This vector is then input into a pre-trained temperature drift compensation model to obtain the compensated, standardized resistance value. The input feature vector is represented as follows: In the formula, The input feature vector is denoted as . The temperature drift compensation model outputs the compensated, standardized resistance value, denoted as . ,in This value is the dimensionless value after normalization.

[0059] The specific implementation of step S04 is as follows: Calculate the rate of change of the standardized resistance values ​​after compensation for the six gas-sensitive elements. When the absolute value of the rate of change of the standardized resistance value after compensation for any gas-sensitive element exceeds the baseline threshold of 0.15, it is determined that there is interference from odorous gas, triggering the odor detection state. The formula for calculating the rate of change is as follows: In the formula, For the first Each gas-sensitive element The rate of change of the standardized resistance value after compensation at time t is a dimensionless relative rate of change. for The standardized resistance value after compensation at any given time; for The standardized resistance value after compensation at any given time; The time interval is 0.1 seconds. The judgment condition is... Odor detection state is triggered at certain times, where The absolute value of the rate of change is represented by the baseline threshold of 0.15, which is determined based on the statistical analysis of the resistance stability of the gas-sensitive element in clean air. It is set to three times the standard deviation of the normal fluctuation range and is used to distinguish between environmental interference and real odor signals.

[0060] The specific implementation of step S05 is as follows: Under odor detection conditions, extract the time-series features of the compensated standardized resistance values ​​of six gas-sensitive elements within 5 consecutive seconds, including the mean, variance, peak time, and decay rate, and construct a feature matrix. The time-series feature extraction formula is expressed as follows: mean ;variance Peak time decay rate In the formula, For the first The average standardized resistance value of each gas-sensitive element after compensation within a 5-second time window, dimensionless; The variance is dimensionless. The number of sampling points within 5 seconds is equal to 50. For the first Each sampling time point is measured in seconds. Sampling sequence number ; For the first The compensated standardized resistance value at each sampling time; Peak time, in seconds; Indicates to make The moment when the maximum value is reached Peak time, in seconds; Indicates to make The moment when the maximum value is reached The decay rate is expressed in seconds. The end time of the 5-second time window, in seconds; The standardized resistance value after compensation at the peak moment; This represents the compensated, standardized resistance value at the end time. The characteristic matrix is ​​represented as follows: In the formula, It is a 24-dimensional feature matrix.

[0061] The specific implementation of step S06 is as follows: The feature matrix is ​​compared with a pre-stored standard odor feature library using cosine similarity calculation to identify the odor type and quantify the pollutant concentration level. The pollutant concentration level is divided into three levels: mild, moderate, and severe. The cosine similarity calculation formula is expressed as follows: In the formula, Characteristic matrix Compared with the first in the standard odor feature library Odor feature vector The cosine similarity between them is dimensionless. This is the current feature matrix; The first in the standard odor feature library The feature vector of the odor, where The odor type is numbered, and the standard odor feature library contains feature vectors for four typical odors: ethanol, trimethylamine, hydrogen sulfide, and ammonia. Representing vectors and The inner product; Representing vectors The Euclidean norm; Representing vectors The Euclidean norm. Cosine similarity measures the similarity of vectors in a direction by taking the inner product of the normalized feature matrix and vectors in the standard odor feature library. The value ranges from -1 to 1, with values ​​closer to 1 indicating higher similarity. The pollutant concentration level is determined based on the concentration information labeled by the standard feature corresponding to the maximum cosine similarity.

[0062] The specific implementation of step S07 is as follows: Based on the identified pollutant concentration level and the current system entropy value, the thermodynamic entropy increase minimization purification control algorithm is invoked to calculate the optimal power of the ultraviolet light-emitting diode and the optimal speed of the fan assembly. The formula for calculating the current system entropy value is expressed as follows: In the formula, The normalized system entropy value is dimensionless. Here is the Boltzmann constant, which takes the value of Joules per Kelvin; The concentration values ​​corresponding to the pollutant concentration levels are expressed in units of 1000g. For reference concentration, the value is [value to be filled in]. The normalization factor takes a value of Joules per Kelvin; It is the natural logarithm function; The base of the natural logarithm is approximately 2.718. The thermodynamic entropy increase minimization purification control algorithm abstracts the odor purification process as an irreversible thermodynamic system. The initial state entropy value is determined by the pollutant concentration level and the degree of disorder in its distribution. The purification target state is an ordered state where the pollutant concentration level drops below a threshold. Dynamic programming is used to solve for the optimal power of the ultraviolet light-emitting diode. Optimal speed of the fan assembly ,in The unit is watt. The unit is revolutions per minute.

[0063] The specific implementation of step S08 is as follows: The optimal power is converted into a pulse width modulation (PWM) duty cycle signal and output to the PWM drive circuit; the optimal speed is converted into a frequency converter control signal and output to the frequency converter drive circuit; and the purification process is initiated. The formula for calculating the PWM duty cycle is as follows: In the formula, Pulse width modulation duty cycle, dimensionless; The optimal power for an ultraviolet light-emitting diode is expressed in watts. The rated power of the ultraviolet light-emitting diode is 15 watts. The formula for calculating the frequency of the frequency converter control signal is as follows: In the formula, The frequency of the frequency converter control signal is expressed in Hertz (Hz). The optimal speed of the fan assembly, expressed in revolutions per minute; The rated speed of the fan assembly, in revolutions per minute; The rated power supply frequency is 50 Hz.

[0064] The specific implementation method of step S09 is the same as described above, and will not be repeated in detail here.

[0065] The specific implementation of the temperature drift compensation model is as follows: The temperature drift compensation model adopts a combined architecture of dictionary learning based on sparse coding and deep unfolded networks. The traditional iterative shrinkage thresholding algorithm is unfolded into five learnable network layers. Each learnable network layer contains a dictionary matrix transformation unit, a nonlinear shrinkage function unit, and a residual connection unit. The dictionary matrix of the dictionary matrix transformation unit has a dimension of 128 x 64. The dictionary matrix is ​​represented as follows: In the formula, It is a dictionary matrix; This represents a 128-row, 64-column real matrix space. The calculation process of the input feature vector through dictionary matrix transformation is as follows: In the formula, It is a sparse coding vector with a dimension of 64 and is dimensionless; This is the transpose of the dictionary matrix; The input feature vector is a 128-dimensional vector formed by concatenating 16 consecutive sampling times. A dictionary matrix is ​​used to map the input feature vector to a sparse representation space. The nonlinear shrinkage function unit uses a soft thresholding function for feature selection. The residual connection unit preserves information from the original input feature vector to prevent gradient vanishing. The final output layer of the temperature drift compensation model is a fully connected layer, mapping the sparse encoding in the sparse representation space to the compensated, standardized resistance value. The expression for the soft thresholding function is as follows: In the formula, Output of the soft threshold function; For input feature components; This is the shrinkage threshold parameter; For a sign function, when The output is 1 when... The output is negative 1 when... The output is 0. This is a function to find the maximum value. express The absolute value of the contraction threshold parameter. The soft thresholding function achieves feature sparsity and noise suppression by setting the components of the input features whose absolute value is less than the contraction threshold parameter to zero, and subtracting the sign value of the contraction threshold parameter from the components whose absolute value is greater than the contraction threshold parameter. The output is 0 when... The output is This enables the sparsity selection of features.

[0066] The specific implementation steps for establishing the training dataset for the temperature drift compensation model are the same as described above, and will not be repeated in detail here. The specific implementation steps for training the temperature drift compensation model are also the same as described above, and will not be repeated in detail here.

[0067] The recovery threshold is set to one-third of the baseline threshold. This threshold is used to determine whether the purification treatment has achieved the expected results. A smaller recovery threshold ensures that the device completely eliminates any residual odor before exiting the odor detection state. The formula for calculating the recovery threshold is as follows: In the formula, To recover the threshold, dimensionless; The baseline threshold is dimensionless and has a value of 0.15.

[0068] The maintenance power is set to 20% of the rated power of the ultraviolet light-emitting diode (UV LED), which is 3 watts. This is used to maintain the basic activity of the photocatalytic purifier after purification is complete, preventing the rapid accumulation of odor gases. The formula for calculating the maintenance power is as follows: In the formula, To maintain power, the unit is watt; The rated power of the ultraviolet light-emitting diode is 15 watts.

[0069] The low-speed circulation speed is set to 30% of the rated speed of the fan assembly. This is used to maintain slow airflow inside the refrigerator after purification, promoting uniform temperature distribution. The formula for calculating the low-speed circulation speed is as follows: , ; where This refers to the low-speed cycle rotation speed, measured in revolutions per minute. This refers to the rated speed of the fan assembly, measured in revolutions per minute (rpm).

[0070] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To verify the effectiveness of this invention in practical applications, technicians constructed a test environment for the intelligent refrigerator odor detection and treatment device. By simulating the odor generation and purification process inside a household refrigerator, the detection accuracy and purification efficiency of the device were evaluated. The test environment used a 218L three-door refrigerator, with the refrigerator compartment temperature set at 5℃ and relative humidity maintained at 65%, and the freezer compartment temperature set at -18℃. Technicians placed a fish sample stored for more than 72 hours in the refrigerator compartment. Under refrigeration conditions, the sample underwent protein degradation, producing typical odor gases such as trimethylamine, hydrogen sulfide, and ammonia, simulating the odor pollution scenario caused by food spoilage in daily use. The device has six... Semiconductor gas-sensitive elements are evenly distributed at the inlet of the circulating air duct, temperature and humidity sensors are fixed at the center of the sensor array, and photocatalytic purifiers are located in the middle section of the air duct. The coated substrate has 12 ultraviolet light-emitting diodes with a wavelength of 365nm embedded on its surface. The rated speed of the fan assembly is 1800r / min. The control chip adopts the ARM Cortex-M4 architecture and has a main frequency of 168MHz.

[0071] Before the test, the device was run in a clean air environment for 30 minutes to complete baseline calibration, and the initial resistance values ​​of the six gas-sensitive elements stabilized at 82kΩ, 79kΩ, 85kΩ, 81kΩ, 83kΩ, and 80kΩ, respectively. After the fish sample was placed in the refrigerator, the control chip synchronously acquired sensor data at a sampling frequency of 10Hz. Figure 2 As shown, the compensated standardized resistance values ​​of the six gas-sensitive elements began to show a significant decreasing trend starting 15 seconds after sample placement. The resistance values ​​of elements 1 to 6 decreased from the initial 1.00 to 0.73, 0.68, 0.75, 0.71, 0.74, and 0.69, respectively. The absolute values ​​of the resistance change rates all exceeded the baseline threshold of 0.15, prompting the control chip to determine the presence of odorous gas interference and trigger odor detection. The device extracted temporal features over a continuous 5-second period to construct a feature matrix. This feature matrix contained 24 feature parameters, including the mean, variance, peak time, and decay rate of the six channels. By performing cosine similarity calculations with a pre-stored standard odor feature library, the main pollutant was identified as a mixture of trimethylamine and hydrogen sulfide, with cosine similarities reaching 0.87 and 0.82, respectively, classifying the pollutant concentration level as moderate pollution.

[0072] like Figure 3As shown, the temperature drift compensation model exhibits good convergence performance during training. The training loss decreased from the initial 0.342 to 0.018 at the 120th epoch, and the validation loss decreased from 0.358 to 0.021. Both values ​​stabilized after the 150th epoch, indicating that the model successfully learned the intrinsic law of the change in the resistance value of the gas-sensitive element with temperature and humidity. The model employs a deep unfolded architecture with five learnable network layers, a dictionary matrix dimension of 128×64, and approximately 87,000 parameters. The single inference time on the control chip is only 3.2ms, meeting the requirements for real-time compensation. The model extracts key temperature-related features through a sparse coding mechanism. Compared with the traditional third-order polynomial fitting method, the compensation error is reduced by 58% in the temperature range of -5℃ to 25℃, significantly improving the detection stability under different temperature and humidity conditions.

[0073] Based on the identified pollutant concentration levels and the current system entropy value, the control chip uses a thermodynamic entropy-minimizing purification control algorithm to calculate the optimal control parameters. The initial system entropy value is... The target state entropy value for purification is set to The algorithm uses dynamic programming to discretize the 300-second purification cycle into 60 time steps, each lasting 5 seconds. The UV LED power is divided into five levels: 3W, 6W, 9W, 12W, and 15W. The fan speed is divided into three levels: 540 r / min, 1080 r / min, and 1620 r / min. A total of 15 control action combinations are constructed. For example... Figure 4 As shown, the power of the ultraviolet light-emitting diode and the fan speed exhibit dynamic adjustment characteristics during the purification process. For the first 60 seconds, the power is maintained at its maximum value of 15W, and the fan speed is kept at 1620 r / min to rapidly reduce the pollutant concentration. From the 60th to the 180th second, the power gradually decreases to 9W, and the fan speed is adjusted to 1080 r / min to match the lower pollutant concentration at this time, avoiding energy waste caused by excessive excitation. After the 180th second, the power is further reduced to 6W, and the fan speed is maintained at 1080 r / min to ensure the full degradation of residual pollutants. Throughout the purification process, the compensated standardized resistance values ​​of the six gas-sensitive elements continuously increase. By the 240th second, the absolute value of the resistance change rate of all elements is below the recovery threshold of 0.05 and the duration exceeds 10 seconds. The control chip determines that purification is complete, reduces the power of the ultraviolet light-emitting diode to a maintenance power of 3W, and adjusts the fan speed to a low-speed circulation speed of 540 r / min.

[0074] like Figure 5 As shown, the system entropy decreases exponentially with the purification process, from the initial value... It dropped to at 60 seconds It dropped to 120 seconds. It dropped to 180 seconds later. The purification target state was reached at 240 seconds. The cumulative entropy generation is Compared to the cumulative entropy generation of a fixed-power continuous operation strategy The efficiency was reduced by 38.6%, verifying the effectiveness of the purification control algorithm that minimizes thermodynamic entropy increase. Technicians recorded key parameters during the purification process, as shown in Table 1. The data shows that the device automatically adjusted the control strategy at different stages, achieving dynamic optimization of energy conversion efficiency.

[0075] Table 1 Key Parameter Recording Table for Purification Process

[0076] The characteristic matrix response patterns under different pollutant concentration levels showed significant differences. Under heavy pollution, the mean resistance change rate of the six gas-sensitive elements reached 0.32, with a variance of 0.018. The peak value was concentrated between 8 and 12 seconds after the start of detection, and the decay rate was slow. Under moderate pollution, the mean resistance change rate was 0.21, with a variance of 0.012. The peak value was distributed between 15 and 20 seconds, and the decay rate was moderate. Under light pollution, the mean resistance change rate was only 0.09, with a variance of 0.006. The peak value was delayed to 25 to 30 seconds, and the decay rate was relatively fast. These response pattern characteristics provide a reliable basis for the control chip to accurately identify the pollutant type and concentration level.

[0077] The technological advancements brought about by this invention compared to traditional methods are mainly reflected in three aspects. First, the temperature drift compensation model based on sparse coding dictionary learning and deep unfolded networks overcomes the limitations of traditional polynomial fitting methods. By adaptively learning the sparse representation basis function of the gas-sensitive element's resistance value changing with temperature and humidity, it can capture more complex nonlinear relationships, significantly improving the detection accuracy and stability under different temperature and humidity conditions, and avoiding the problem of repeatedly adjusting parameters required by manually designing compensation models. Second, the thermodynamic entropy increase minimization purification control algorithm abstracts the odor purification process into an irreversible thermodynamic system. With minimizing the entropy generation rate per unit time as the optimization objective, it dynamically matches the power of the ultraviolet light-emitting diode and the fan speed with the current pollutant concentration, ensuring that the photocatalytic reaction rate is always in the high-efficiency range. This overcomes the problem of low energy utilization efficiency caused by the traditional fixed-power control strategy ignoring the physical law that the reaction rate and mass transfer efficiency decrease with decreasing concentration. Third, the dynamic programming method ensures global optimality in solving the control action sequence. Compared with greedy strategies or heuristic methods, it can anticipate future state changes and adjust control parameters in advance, avoiding getting trapped in local optima. At the same time, it explicitly incorporates power constraints and time constraints into the optimization problem, automatically finding the Pareto optimal solution for energy consumption and time while ensuring purification effect, thus improving the intelligence level and practical value of the device.

[0078] It should be noted that the variables involved in this invention are explained in detail in Table 2.

[0079] Table 2 Variable Explanation Table

[0080]

[0081] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A refrigerator odor intelligent detection and processing device, comprising a gas sensor array, a temperature and humidity sensor, a photocatalytic purifier, a fan assembly, a power management unit, and a control chip, characterized in that, The gas sensor array comprises six tin dioxide-based semiconductor gas-sensitive elements evenly distributed in a ring at the inlet of the refrigerator's internal circulating air duct. A temperature and humidity sensor is fixedly installed at the center of the gas sensor array to monitor the refrigerator's internal ambient temperature and relative humidity in real time. The photocatalytic purifier, consisting of a titanium dioxide-coated substrate and ultraviolet light-emitting diodes, is located in the middle section of the refrigerator's internal air duct. The fan assembly is fixedly installed at the end of the circulating air duct, and its speed is controlled by a frequency converter drive circuit. The power management unit output is electrically connected to the gas sensor array, temperature and humidity sensor, photocatalytic purifier, and fan assembly to provide a 5V DC operating voltage. The control chip incorporates an adaptive baseline correction and entropy minimization control module to achieve coordinated control of each component.

2. The refrigerator odor intelligent detection and processing device according to claim 1, characterized in that, The first input port group of the control chip is electrically connected to the six output signal lines of the gas sensor array, and the second input port of the control chip is electrically connected to the data output terminal of the temperature and humidity sensor.

3. The intelligent refrigerator odor detection and processing device according to claim 2, characterized in that, The first output port of the control chip is electrically connected to the ultraviolet light-emitting diode through a pulse width modulation drive circuit, and the second output port of the control chip is electrically connected to the motor winding of the fan assembly through a frequency conversion drive circuit.

4. The refrigerator odor intelligent detection and processing device according to claim 3, characterized in that, The adaptive baseline correction and entropy minimization control module synchronously collects the resistance values ​​of the gas-sensitive element and the ambient temperature and humidity values ​​to construct a multi-dimensional time-series data stream. Wavelet transform is used to decompose the resistance data to extract low-frequency trend components. The resistance values ​​corresponding to these low-frequency trend components are combined with the ambient temperature and humidity values ​​to form an input feature vector, which is then input into the temperature drift compensation model to obtain the compensated standardized resistance value. The temperature drift compensation model adopts an architecture combining dictionary learning and deep expansion networks based on sparse coding. It achieves a sparse representation of temperature dependence through dictionary matrix transformation units, nonlinear shrinkage function units, and residual connection units. The rate of change of the compensated standardized resistance value is calculated to determine the odor detection status. Time-series features are extracted to construct a feature matrix, and similarity calculations are performed with a standard odor feature library to identify the odor type and pollutant concentration level. The thermodynamic entropy minimization purification control algorithm is then invoked to calculate the optimal power of the ultraviolet light-emitting diode and the optimal speed of the fan assembly to initiate purification treatment.

5. The intelligent refrigerator odor detection and processing device according to claim 4, characterized in that, The sampling frequency of synchronous acquisition is 10 Hz. Wavelet transform decomposition filters out high-frequency noise components above 5 Hz and retains low-frequency trend components for subsequent analysis.

6. The intelligent refrigerator odor detection and processing device according to claim 5, characterized in that, The temperature drift compensation model expands the traditional iterative shrinkage threshold algorithm into a five-layer learnable network. Each learnable network layer contains a dictionary matrix transformation unit, a nonlinear shrinkage function unit, and a residual connection unit. The dictionary matrix of the dictionary matrix transformation unit has a dimension of 128 x 64.

7. The intelligent refrigerator odor detection and processing device according to claim 6, characterized in that, The nonlinear shrinkage function unit uses a soft threshold function to set the components of the input features whose absolute value is less than the shrinkage threshold parameter to zero, and subtracts the sign value of the shrinkage threshold parameter from the components whose absolute value is greater than the shrinkage threshold parameter, thereby achieving feature sparsity and noise suppression.

8. The intelligent refrigerator odor detection and processing device according to claim 7, characterized in that, The training dataset for the temperature drift compensation model was established in a standard temperature and humidity test chamber. The temperature was adjusted in 2-degree increments from -5 degrees Celsius to 25 degrees Celsius, and the relative humidity was adjusted in 10-degree increments from 30% to 90%. 300 sets of baseline data were collected at each operating point, for a total of 42,000 sets of baseline data.

9. The intelligent refrigerator odor detection and processing device according to claim 8, characterized in that, The training dataset also includes 168,000 sets of labeled data. The labeled data are introduced with four odorous gases, namely ethanol, trimethylamine, hydrogen sulfide and ammonia, under the same temperature and humidity combination working condition matrix. The concentration gradient is set to 5ppm to 50ppm, and 200 sets of response data are collected at each concentration point.

10. The intelligent refrigerator odor detection and processing device according to claim 9, characterized in that, The temperature drift compensation model training uses mean squared error as the loss function and an adaptive moment estimation optimization algorithm for parameter updates. The initial learning rate is 0.001, which decays to 0.9 times the original value every 20 rounds. The batch size is 64, and the total number of training rounds is 200.

Images