Temperature control method and device for UV LED module

Abstract

The present invention relates to a temperature control method and device for a UV LED module. The method includes: collecting temperature data of a UV LED module to obtain a historical temperature data set, analyzing the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function; calculating a temperature influence factor according to the temperature-output relationship function; constructing a deep neural network UV evaluation model based on the temperature influence factor and the historical temperature data set; performing real-time temperature monitoring on multiple target partitions in the UV LED module to obtain a real-time temperature data set for each target partition; predicting UV output control parameters of the real-time temperature data set through a deep neural network UV evaluation model to obtain multiple UV output control parameters, and creating a corresponding intelligent temperature control instruction sequence according to the multiple UV output control parameters. The implementation of the present invention can accurately sense temperature changes, predict UV output characteristics, and realize adaptive control.

Description

Technical Field

[0001] The present invention relates to the technical field of UV LED module control, and in particular to a temperature control method and device for a UV LED module. Background Art

[0002] As a new type of light source, UV LED modules are widely used in printing, coating, medical and other fields. However, UVLED generates a lot of heat during operation, causing the temperature to rise, which in turn affects its output performance and service life. Traditional temperature control methods often use simple switch control or PID control, which is difficult to adapt to the temperature change characteristics of UV LED modules under complex working conditions, and are prone to problems such as low control accuracy and slow response speed.

[0003] As UV LED modules develop towards high power and high density, their heat dissipation problem becomes more prominent. Uneven temperature distribution and local overheating frequently occur, seriously affecting the stability and reliability of UV LED modules. In addition, there is a complex nonlinear relationship between the output characteristics of UVLED modules and temperature. Traditional temperature control methods are difficult to accurately capture this relationship and cannot achieve precise output control. Summary of the invention

[0004] The main purpose of the present invention is to provide a temperature control method and device for a UV LED module, so as to accurately sense temperature changes, predict UV output characteristics, and achieve adaptive control.

[0005] To achieve the above object, the present invention provides a temperature control method for a UV LED module, comprising the following steps:

[0006] Collecting temperature data of the UV LED module to obtain a historical temperature data set, and analyzing the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function;

[0007] Calculating the temperature influence factor according to the temperature-output relationship function;

[0008] Constructing a deep neural network UV evaluation model based on the temperature influencing factor and the historical temperature data set;

[0009] Perform real-time temperature monitoring on multiple target partitions in the UV LED module to obtain a real-time temperature data set for each target partition;

[0010] The deep neural network UV evaluation model is used to predict UV output control parameters of the real-time temperature data set to obtain multiple UV output control parameters, and a corresponding intelligent temperature control instruction sequence is created according to the multiple UV output control parameters.

[0011] The present invention also provides a temperature control device for a UV LED module, comprising:

[0012] The acquisition module is used to collect temperature data of the UV LED module to obtain a historical temperature data set, and analyze the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function;

[0013] A calculation module, used for calculating a temperature influence factor according to the temperature-output relationship function;

[0014] A construction module, used to construct a deep neural network UV evaluation model based on the temperature influencing factor and the historical temperature data set;

[0015] A monitoring module, used to perform real-time temperature monitoring on multiple target partitions in the UV LED module, and obtain a real-time temperature data set for each target partition;

[0016] A prediction module is used to predict UV output control parameters of the real-time temperature data set through the deep neural network UV evaluation model to obtain multiple UV output control parameters, and create corresponding intelligent temperature control instruction sequences according to the multiple UV output control parameters.

[0017] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0019] In summary, the technical solution provided by the present invention establishes a relationship model between temperature, current input and UV output intensity by performing polynomial regression analysis on historical temperature data sets. This method can accurately capture the nonlinear output characteristics of the UV LED module, and by calculating the temperature change rate and acceleration, constructing a temperature impact assessment matrix, and performing singular value decomposition, the temperature impact characteristics after dimensionality reduction are obtained. This method can fully consider the dynamic characteristics of temperature changes and improve the accuracy of temperature impact assessment. A complex deep learning model is constructed by adopting a multimodal input structure, a multi-branch convolutional neural network, a multi-head self-attention mechanism and a bidirectional long short-term memory network. This model can make full use of the information of temperature data and temperature influencing factors to achieve accurate prediction of UV output control parameters. By constructing a temperature control state machine including normal state, overheating state and cooling state, and introducing dynamic threshold adjustment and online learning mechanism, intelligent and adaptive control of UVLED module temperature is achieved. Real-time temperature monitoring of multiple target partitions in the UV LED module can timely detect local overheating problems and improve the refinement of temperature control. Based on the predicted UV output control parameters, PWM modulation signals and cooling system control instructions are generated to form an intelligent temperature control instruction sequence. This method can achieve precise control of UV LED modules and improve the response speed and control accuracy of the system. The cross entropy loss function and adaptive moment estimation optimizer are adopted, and residual connection and dropout technologies are introduced to improve the training effect and generalization ability of the deep neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 1 is a schematic diagram of the steps of a temperature control method for a UV LED module in one embodiment of the present invention;

[0021] Figure 2 is a structural block diagram of a temperature control device for a UV LED module in one embodiment of the present invention;

[0022] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0023] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] Reference Figure 1, this embodiment provides a temperature control method for a UV LED module, comprising the following steps:

[0026] S1, collecting temperature data of the UV LED module to obtain a historical temperature data set, and analyzing the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function;

[0027] Specifically, a temperature sensor array is installed in the UV LED module, and the temperature data of the UV LED module in different time periods are collected through the temperature sensor array to form a historical temperature data set, which reflects the temperature change of the UV LED module under different working conditions. The historical temperature data set is normalized to eliminate the data inconsistency caused by the difference in the measurement range between different sensors; denoising is performed to remove abnormal data caused by sensor noise or other external interference; interpolation is performed to complete the missing data that may be caused by sensor failure or other reasons during the acquisition process, and the preprocessed temperature data set is obtained. Based on the preprocessed temperature data set, the corresponding UV LED module input current data and UV output intensity data are collected to construct a multidimensional output data set including temperature, current input and UV output intensity. Polynomial regression analysis is performed on the constructed multidimensional output data set to establish a relationship model between temperature, current input and UV output intensity, and obtain the initial temperature-output relationship function. This function can describe the influence of temperature change on the output intensity of the UV LED module. In the optimization process, the least squares method is used to solve the polynomial regression coefficient, and a more accurate optimization relationship function is obtained by minimizing the error between the predicted value and the actual value. The optimization relationship function is model-verified. The prediction accuracy of the model is evaluated by calculating the coefficient of determination and root mean square error. If the coefficient of determination is greater than or equal to the preset threshold and the root mean square error is less than or equal to the preset threshold, the model is considered to have been verified and the obtained optimization relationship function is valid. At this time, it can be used as the final temperature-output relationship function for subsequent temperature control of the UV LED module to achieve precise control of UV output. If the model verification fails to meet expectations, it is necessary to further adjust the model parameters or select a more complex regression method to re-model until the requirements are met.

[0028] Feature selection is performed on the multidimensional output data set to identify the features that have the greatest impact on the UV output intensity, and redundant or less relevant features are removed to obtain the feature data set after dimensionality reduction. Based on the feature data set after dimensionality reduction, a polynomial feature matrix is ​​constructed, and the original features are expanded into a higher-dimensional feature matrix by performing polynomial expansion on the temperature and current features. The polynomial expansion of the features can capture the nonlinear relationship between the features, thereby improving the fitting ability of the model. The expanded feature matrix is ​​regularized. In order to suppress overfitting, the ridge regression method is used to introduce the L2 regularization term in the regression model. By adding a penalty term proportional to the sum of squares of the regression coefficients, the complexity of the model is constrained to obtain a regularized feature matrix. Based on the regularized feature matrix, polynomial regression analysis is performed, and the regression coefficients are iteratively optimized using the gradient descent algorithm. The gradient descent algorithm continuously adjusts the regression coefficients so that the error between the model prediction value and the true value is gradually reduced to obtain the initial polynomial regression model. The initial polynomial regression model is cross-validated. The K-fold cross-validation method is used to divide the data set into K subsets, and each subset is trained and validated separately. The performance of the model on different training sets and validation sets is evaluated, and the cross-validation score is calculated. This process can reveal the stability and generalization ability of the model and prevent model bias caused by improper data set division. According to the cross-validation score, the initial polynomial regression model is hyperparameter tuned, and the grid search algorithm is used to optimize the order and regularization coefficient of the polynomial. The best parameter combination is searched in a set of preset parameter grids to maximize the prediction accuracy and generalization ability of the model. Based on the optimized parameter combination, the model is retrained and the initial relationship function is obtained.

[0029] S2, calculate the temperature influence factor according to the temperature-output relationship function;

[0030] Specifically, the temperature-output relationship function is numerically differentiated to obtain the sensitivity function of temperature to UV output, which reflects the degree of influence of temperature change on UV output intensity. Based on the sensitivity function, the temperature change rate is calculated, and the temperature data is processed by sliding window to obtain a dynamic temperature change rate sequence. Sliding window processing can capture the local trend of temperature change over time, making the calculation of temperature change rate more accurate and real-time. According to the dynamic temperature change rate sequence, the acceleration of temperature change is calculated, and the temperature change acceleration sequence is obtained by performing second-order difference processing on the temperature change rate. Based on the temperature change acceleration sequence, a temperature impact assessment matrix is ​​constructed, and the temperature, temperature change rate and temperature change acceleration are combined into a multidimensional feature vector. The temperature impact assessment matrix is ​​subjected to singular value decomposition, and the most important features are extracted to obtain the temperature impact features after dimensionality reduction. Singular value decomposition can effectively remove noise and redundant information, simplify the complexity of the model and improve computational efficiency. Based on the temperature impact features after dimensionality reduction, the cumulative effect of temperature impact is calculated, and the exponentially weighted moving average method is used to obtain the cumulative temperature impact factor. The exponentially weighted moving average can give greater weight to recent data and better capture the trend and cumulative effect of temperature changes. The cumulative temperature impact factor is a comprehensive assessment of the long-term impact of temperature on UV output and is an important parameter for the entire temperature control system. Based on the cumulative temperature impact factor, a temperature impact prediction model is constructed to calculate the short-term temperature impact prediction value, and then a weighted fusion is performed based on the short-term temperature impact prediction value and the cumulative temperature impact factor to obtain the final temperature impact factor.

[0031] S3, builds a deep neural network UV evaluation model based on temperature influencing factors and historical temperature datasets;

[0032] Specifically, a deep neural network UV evaluation model is constructed. The input layer of the model adopts a multimodal input structure, including a temperature data input channel and a temperature influence factor input channel. Among them, the temperature data input dimension is n×m, n represents the time step, and m represents the number of temperature sensors; the temperature influence factor input dimension is k×1, where k is the number of influence factors. Through the multimodal input structure, the model can comprehensively process historical temperature data and temperature influence factors. The feature extraction module of the deep neural network UV evaluation model is constructed. The module contains three parallel convolutional neural network branches, each of which consists of three one-dimensional convolutional layers. Among them, the convolution kernel sizes of the first branch are 3, 5, and 7 respectively, the convolution kernel sizes of the second branch are 5, 7, and 9 respectively, and the convolution kernel sizes of the third branch are 7, 9, and 11 respectively, and each layer is configured with 64 convolution kernels. Each convolution layer is followed by a batch normalization layer and a rectified linear unit activation function. Through multi-scale convolution operations, the model can capture features at different scales, and improve the nonlinear expression ability of the model through batch normalization and activation functions. In the feature extraction module model, a multi-head self-attention mechanism is introduced to construct an 8-head self-attention layer. The dimension of each attention head is 64. The attention weight is obtained by scaling the dot product attention calculation. Then the outputs of different heads are spliced ​​and linearly transformed, so that the model can adaptively focus on the important parts of the input features, effectively improving the model's ability to capture key information. The time series modeling module of the deep neural network UV evaluation model is constructed. The module adopts a bidirectional long short-term memory network (BiLSTM) structure, which contains 2 layers of bidirectional long short-term memory units, each with 128 hidden nodes. Each long short-term memory unit contains a forget gate, an input gate, an output gate, and a unit state. Through the bidirectional LSTM structure, the model can simultaneously consider the front-end and back-end dependencies of the time series and better capture the dynamic characteristics of temperature data over time. The model also contains a global average pooling layer and a global maximum pooling layer, and the pooling window size is equal to the time step. By splicing the output of the pooling layer on the feature dimension, the model extracts global information, reduces the feature dimension, and retains important time series information. Global features can effectively summarize the overall trend of time series data and enhance the stability of the model. Three fully connected layers are introduced in the feature extraction and time series modeling model. The first layer uses 256 neurons and adopts the rectified linear unit activation function; the second layer uses 128 neurons and adopts the same activation function; the third layer uses 64 neurons and adopts the hyperbolic tangent activation function. Each fully connected layer is followed by a dropout layer, and the dropout rate is set to 0.5 to reduce the risk of overfitting. At the same time, in order to retain the original input information, the model adds a residual connection after the fully connected layer. Through the element-by-element addition operation, the original features of the input layer are fused with the output of the fully connected layer to achieve short-circuit transmission of information and ensure that key information is not lost in the deep network.Based on the output of the residual connection, the model constructs the output layer and uses the softmax activation function to generate the probability distribution of the UV output control parameters. The output dimension is q, where q represents the number of possible UV output control parameter categories. In order to ensure the prediction accuracy of the model, the temperature influencing factors and the historical temperature data set are preprocessed to obtain the preprocessed training data set. On this basis, the cross entropy loss function and the adaptive moment estimation optimizer (Adam) are used to train the model. The model parameters are updated through multiple rounds of iterations to finally obtain the trained deep neural network UV evaluation model. The model can accurately evaluate the temperature impact of UV LED modules.

[0033] The temperature influencing factors and historical temperature data sets are normalized, and the data of different dimensions are adjusted to the same range to eliminate the influence of the difference between the eigenvalues ​​on the model training, and the normalized data sets are obtained. Based on the normalized data sets, the sliding window method is used to divide the time series. The historical temperature data set is divided into multiple segments according to the time step to obtain the preprocessed training data set. The preprocessed training data set is input into the input layer of the deep neural network UV evaluation model to obtain the temperature data input matrix and the temperature influencing factor input vector. In order to fully extract the characteristic information in temperature and temperature influencing factors, the model designs three parallel one-dimensional convolution branches, each branch independently performs convolution operations on the input data to extract features at different scales. The output of each branch is a feature matrix, and then the feature matrices of the three branches are spliced ​​on the feature dimension to form a spliced ​​feature matrix. The multi-branch convolution structure helps to capture the feature information at different dimensions and scales. Multi-head self-attention calculation is performed on the spliced ​​feature matrix. Including multiplying the spliced ​​feature matrix by three learnable weight matrices to obtain the query matrix, key matrix and value matrix. By calculating the dot product of the query matrix and the key matrix, followed by scaling and softmax operations, the attention weights are obtained to reflect the importance of each position in the feature matrix. The attention weights are then multiplied by the value matrix to obtain the weighted feature representation. Based on the weighted feature representation, the model performs time series modeling through a bidirectional long short-term memory network to obtain a forward hidden state and a backward hidden state. These two hidden states capture the forward and reverse dependencies of the time series respectively. By concatenating the forward hidden state and the backward hidden state, a time series feature matrix is ​​obtained, which fully reflects the dynamic characteristics of the input data in the time dimension. Global average pooling and global maximum pooling operations are performed on the time series feature matrix to obtain two feature vectors. The global pooling operation can compress the feature matrix into a vector of fixed length, thereby extracting global features and reducing the complexity of the model. After concatenating the two feature vectors, they are input into the fully connected layer for further processing to obtain a reduced feature vector. The reduced feature vector is more compact while retaining key feature information. Based on the reduced feature vector, the probability distribution of the UV output control parameters is calculated by the softmax function. The feature vector is multiplied by the weight matrix of the output layer and the bias vector is added. After the softmax operation, the probability distribution of each UV output control parameter category is obtained. The cross entropy loss between the probability distribution and the true label is calculated to obtain the loss function value. Based on the loss function value, the adaptive moment estimation optimizer (Adam) is used to update the parameters. The Adam optimizer can adaptively adjust the learning rate according to the first-order and second-order momentum, thereby accelerating convergence, and finally obtaining a trained deep neural network UV evaluation model.

[0034] Based on the loss function value, the gradient of each parameter of the deep neural network UV evaluation model is calculated to obtain a gradient matrix, which reflects the rate of change of the loss function relative to each model parameter. Based on the gradient matrix, the first-order momentum estimate is updated to obtain a first-order momentum vector. The first-order momentum vector accumulates historical gradient information by exponentially weighted moving average, so that the parameter update not only considers the current gradient, but also integrates the trends of the previous updates, thereby enhancing the stability of the update. The gradient matrix is ​​squared to obtain a gradient square matrix, and the second-order momentum estimate is updated based on the gradient square matrix to obtain a second-order momentum vector. The second-order momentum vector is used to capture the amplitude of the gradient change and reflects the variance information of the gradient by accumulating the square value of the gradient. The first-order and second-order momentum vectors are bias corrected respectively to eliminate the bias effect of the momentum vector in the initial stage. Especially in the early stage of the optimization process, the direct use of the uncorrected momentum vector may lead to inaccurate updates because the cumulative effect of the momentum has not yet fully taken effect. Through bias correction, more accurate corrected first-order momentum vectors and corrected second-order momentum vectors are obtained. Based on the corrected momentum vector, the adaptive learning rate is calculated to obtain the adaptive learning rate vector of each parameter. The adaptive learning rate dynamically adjusts the learning rate of each parameter by combining the information of the first-order and second-order momentum, so that the model can be updated at an appropriate pace at different training stages and different parameter dimensions. When the gradient of the parameter is large, the learning rate will decrease accordingly to avoid excessive update steps; conversely, when the gradient is small, the learning rate will increase to accelerate the convergence speed. According to the adaptive learning rate vector and the corrected first-order momentum vector, the update amount of each parameter is calculated to obtain the parameter update vector. The parameter update vector determines how the parameters of each layer should be adjusted in the current iteration to minimize the prediction error of the model. Based on the parameter update vector, the parameters of each layer of the deep neural network UV evaluation model are updated. The entire optimization process continues until the preset number of training rounds is reached or the early stopping condition is met. The setting of the early stopping condition is usually based on the performance of the validation set. When the loss of the validation set no longer decreases or begins to increase, the optimization process will terminate early to prevent the model from overfitting. After the above steps, the trained deep neural network UV evaluation model is finally obtained.

[0035] S4, performing real-time temperature monitoring on multiple target partitions in the UV LED module to obtain a real-time temperature data set for each target partition;

[0036] Specifically, temperature sensors are reasonably arranged in each target partition of the UV LED module. Each target partition should be equipped with at least one high-precision temperature sensor, which can detect the temperature changes of the partition in real time and convert them into electrical signals. The sensors are managed uniformly by a central control unit. The central control unit adopts a multi-channel data acquisition module, which can simultaneously receive temperature sensor signals from different partitions. The central control unit usually adopts a high-frequency sampling rate to ensure that the dynamic changes of temperature can be captured in time. In each sampling cycle, the temperature data of all partitions will be synchronously collected and transmitted to the data processing module. In the data processing module, the received temperature signal is preprocessed, including signal filtering, amplification and normalization processing, to eliminate the influence of sensor noise and environmental interference on the data. After preprocessing, the temperature data of each partition will be organized into a structured data set, usually identified in the form of timestamps and partition labels, to form a real-time temperature data set for multiple target partitions.

[0037] S5, predicting UV output control parameters of the real-time temperature data set through a deep neural network UV evaluation model, obtaining multiple UV output control parameters, and creating corresponding intelligent temperature control instruction sequences according to the multiple UV output control parameters.

[0038] Specifically, the real-time temperature data set is preprocessed, including data normalization and time window division, to obtain the preprocessed real-time temperature data set. Normalization is to standardize the temperature data collected by different sensors into a unified range to eliminate the calculation error caused by dimensional differences; while time window division is to divide the continuous temperature data into time segments of fixed length so that the deep neural network can effectively process the time series data and capture the dynamic characteristics of temperature changes. The preprocessed real-time temperature data set is input into the deep neural network UV evaluation model. The model extracts the key features in the temperature data through the feature extraction layer, and then calculates the weights at different time steps through the self-attention mechanism to enhance the focus on important temperature changes. The model models the time series data through a time series modeling module (such as a bidirectional long short-term memory network) to capture the complex pattern of temperature changes over time. Through the fully connected layer processing, the model generates the probability distribution of the UV output control parameters, which reflects the probability of each control parameter under the current temperature conditions. Based on the probability distribution of the UV output control parameters, the maximum a posteriori probability method is used to select the optimal UV output control parameters. By calculating the posterior probability of each possible control parameter, the parameter with the highest posterior probability is selected as the optimal control parameter, and the UV output control parameters of multiple target partitions are obtained. In order to avoid the impact of sudden changes in control parameters on the system, the UV output control parameters of multiple target partitions are smoothed. By using the moving average or exponentially weighted moving average method, the change of the control parameter sequence is ensured to be more stable and continuous, and the fluctuation of the system during execution is reduced, and the smoothed UV output control parameter sequence is obtained. Based on the smoothed UV output control parameter sequence, a temperature control state machine is constructed. The state machine includes normal state, overheating state and cooling state, which correspond to the working modes of the UV LED module under different temperature conditions. According to the current temperature and UV output control parameters, the state transition conditions are determined. For example, when the temperature exceeds the set threshold, it switches from the normal state to the overheating state, and triggers cooling measures. According to the temperature control state machine and the smoothed UV output control parameter sequence, the corresponding PWM modulation signal and cooling system control instructions are generated to form a complete intelligent temperature control instruction sequence. These instructions achieve intelligent control of the module temperature by adjusting the input current of the UV LED module and the working status of the cooling system, ensuring that the UV LED module operates efficiently and stably within the optimal temperature range.

[0039] Based on the smoothed UV output control parameter sequence, the temperature thresholds of the normal state, overheating state and cooling state are set to obtain the state transition temperature threshold set. The state transition temperature threshold set is dynamically adjusted. By combining the historical temperature data and the statistical characteristics of the UV output control parameters, the trend of temperature change and the fluctuation mode of the control parameters are identified, and an adaptive state transition temperature threshold that is more suitable for the current working environment is obtained. Through dynamic adjustment, the system can adapt to changes in ambient temperature or aging of the UV LED module, ensuring that the temperature control state machine can maintain good responsiveness and reliability under different working conditions. Based on the adaptive state transition temperature threshold, a state transition rule matrix is ​​constructed. The matrix clarifies the transition conditions and transition probabilities between the normal state, overheating state and cooling state. By defining the transition paths and possibilities between these states, when the temperature and control parameters change, the state transition is realized according to the preset rules, and a state transition model that can accurately reflect the actual working conditions is constructed. The state transition model is initialized. The initial state is determined based on the current temperature and UV output control parameters to ensure that the temperature control state machine can start running from the most reasonable state at startup. The initialized temperature control state machine ensures that the system can quickly adapt to the temperature conditions at startup and implement the correct temperature control strategy from the beginning. Based on the initialized temperature control state machine, the state holding time and state jump penalty mechanism are designed to avoid frequent state switching caused by slight fluctuations in temperature or control parameters. Frequent state switching will not only reduce the stability of the system, but also may cause unnecessary losses to the UV LED module. By setting the state holding time, the system requires that the state must be maintained for a certain length of time before the state transition; and the state jump penalty mechanism introduces a penalty term to limit the frequent jumps between states, thereby obtaining a more stable temperature control state machine. Online learning and optimization are performed on the stable temperature control state machine. Online learning means that the system dynamically adjusts the state transition conditions and control strategies based on real-time temperature data and feedback on control effects to ensure that the state machine can always adapt to changes in the actual working environment. Through adaptive optimization, the temperature control state machine can continuously improve its control effect, so that the UV LED module can maintain stable and efficient operation under various temperature conditions, and finally obtain an adaptive optimization temperature control state machine that can be intelligently adjusted.

[0040] In one example, temperature data of a UV LED module is collected to obtain a historical temperature data set, and the relationship between temperature and output characteristics of the historical temperature data set is analyzed to obtain a temperature-output relationship function, including:

[0041] A temperature sensor array is installed in the UV LED module, and a historical temperature data set of the UV LED module is collected through the temperature sensor array;

[0042] Perform normalization, denoising and interpolation processing on the historical temperature data set to obtain a preprocessed temperature data set;

[0043] Based on the preprocessed temperature data set, the corresponding UV LED module input current data and UV output intensity data are collected to construct a multi-dimensional output data set;

[0044] Perform polynomial regression analysis on the multidimensional output data set, establish the relationship model between temperature, current input and UV output intensity, and obtain the initial relationship function;

[0045] The parameters of the initial relationship function are estimated, and the least square method is used to solve the polynomial regression coefficients to obtain the optimized relationship function;

[0046] The optimized relationship function is model verified, and the determination coefficient and root mean square error are calculated. If the determination coefficient is greater than or equal to the preset threshold and the root mean square error is less than or equal to the preset threshold, the model is considered valid and the temperature-output relationship function is output.

[0047] In this example, a temperature sensor array is installed inside the UV LED module to collect historical temperature data of the UV LED module. This data is usually obtained through continuous monitoring over a period of time, recording the temperature changes of the module under various operating environments. The historical temperature data set is normalized, denoised, and interpolated. The purpose of normalization is to adjust the temperature data of different sensors to a unified range to eliminate the impact of dimensional differences on the model. Denoising is performed. Common denoising methods include using a low-pass filter or moving average method to smooth the temperature data and remove high-frequency noise. Data interpolation is used to complete missing data caused by sensor failure or other reasons during the acquisition process. Common interpolation methods include linear interpolation and polynomial interpolation. When selecting a specific method, the continuity and change trend of the data should be considered. After processing, the obtained preprocessed temperature data set is cleaner and more complete, and can accurately reflect the temperature changes of the UV LED module. Based on the preprocessed temperature data set, the corresponding UV LED module input current data and UV output intensity data are collected. By combining the data of temperature, input current, and UV output intensity, a multidimensional output data set is constructed. Each row of the data set contains information about the temperature, input current, and output intensity of the UV LED module at a certain time or under a certain operating condition. Polynomial regression analysis is used to establish a relationship model between temperature, current input, and UV output intensity. Polynomial regression is a model that extends linear regression and can capture nonlinear relationships between variables. Assume that the temperature is , the input current is , UV output intensity is , then the relationship between them can be expressed by a polynomial regression model in the following form:

[0048] ;

[0049] in, is the regression coefficient, which is a parameter that needs to be determined by data fitting. The model can capture the complex interaction between temperature and input current by expanding the polynomials of them and more accurately predict the UV output intensity. After obtaining the initial relationship function, the model parameters are estimated and the least squares method is used to solve these polynomial regression coefficients. The goal of the least squares method is to determine the optimal regression coefficient by minimizing the square error between the predicted value and the actual value. Specifically, it can be expressed as:

[0050] ;

[0051] in, is the actual UV output intensity, is the UV output intensity predicted by the model, is the number of samples. By solving the optimization problem, the optimal regression coefficient is obtained and the optimized relationship function is constructed. The model of the optimized relationship function is verified. The key indicators of verification include the determination coefficient and root mean square error (RMSE). The coefficient of determination is used to measure the goodness of fit of the model, indicating the proportion of data variance explained by the model, and the formula is:

[0052] ;

[0053] in, is the mean UV output intensity. The closer the value is to 1, the better the model fit is. The root mean square error is the standard deviation between the predicted value and the actual value, reflecting the prediction accuracy of the model. If the value is greater than or equal to the preset threshold and the RMSE is less than or equal to the preset threshold, the model is considered valid and the final temperature-output relationship function can be output.

[0054] In one example, a polynomial regression analysis is performed on a multidimensional output data set to establish a relationship model between temperature, current input and UV output intensity, and an initial relationship function is obtained, including:

[0055] Perform feature selection on the multi-dimensional output data set to obtain a feature data set after dimensionality reduction;

[0056] Based on the feature data set after dimensionality reduction, a polynomial feature matrix is ​​constructed, and the temperature and current features are polynomially expanded to obtain an extended feature matrix;

[0057] Regularize the extended feature matrix, introduce the L2 regularization term using the ridge regression method, and obtain the regularized feature matrix;

[0058] Based on the regularized feature matrix, polynomial regression analysis is performed, and the regression coefficients are iteratively optimized using the gradient descent algorithm to obtain the initial polynomial regression model;

[0059] The initial polynomial regression model was cross-validated, and the K-fold cross-validation method was used to evaluate the generalization ability of the model and obtain the cross-validation score;

[0060] According to the cross-validation score, the hyperparameters of the initial polynomial regression model are tuned, and the grid search algorithm is used to optimize the polynomial order and regularization coefficient to obtain the initial relationship function.

[0061] In this example, feature selection is performed on the multidimensional output data set to reduce the dimension of the data and eliminate redundant information. Common feature selection methods include correlation analysis, principal component analysis, recursive feature elimination, etc. Assuming that the data set contains multiple features such as temperature, input current, UV output intensity, and ambient humidity, after feature selection, it may be found that temperature and input current have the greatest impact on UV output intensity, while other features have less impact. Therefore, the data set is reduced to retain only the two key features of temperature and input current to obtain a reduced feature data set. Based on the reduced feature data set, a polynomial feature matrix is ​​constructed. Through polynomial expansion, the original features are nonlinearly transformed to generate a richer feature space in order to capture more complex relationships between features. Set the temperature feature to And the current characteristics are , construct the following polynomial feature matrix:

[0062] ;

[0063] in, represents the polynomial characteristic matrix, and are the polynomial orders of temperature and current characteristics, respectively. By expanding, the originally linearly independent features can be combined into more complex nonlinear features, making the model have stronger fitting ability. Polynomial expansion is suitable for processing data with nonlinear relationships, such as the relationship between temperature, current and output intensity in UV LED modules. Regularization is performed on the expanded feature matrix. Ridge regression is a commonly used regularization method that constrains the complexity of the model by introducing the L2 regularization term in the regression model. The objective function of ridge regression can be expressed as:

[0064] ;

[0065] in, is the regression coefficient, is the actual output value, is the model prediction value, is the regularization parameter, is the number of features. L2 regularization term Used to penalize the sum of squares of all regression coefficients, The larger the value of , the stronger the penalty, which can effectively prevent the model from overfitting the training data and obtain a regularized feature matrix. Based on the regularized feature matrix, perform polynomial regression analysis. Polynomial regression finds the optimal regression coefficient by minimizing the prediction error. The gradient descent algorithm is used to iteratively optimize the regression coefficient. The basic principle of the gradient descent algorithm is to gradually update the regression coefficient according to the gradient information of the objective function until the minimum error is reached or the preset convergence condition is met. Assume that the objective function is:

[0066] ;

[0067] in, is the mean square error, is the L2 regularization term. The gradient descent algorithm continuously adjusts value, reduce The value of is used to obtain the initial polynomial regression model. The initial polynomial regression model is cross-validated to evaluate the generalization ability of the model. K-fold cross validation is a commonly used technique that divides the data set into K subsets, uses K-1 subsets for training each time, and uses the remaining 1 subset for validation, and calculates the average validation result. This method can effectively avoid overfitting, and obtain the performance of the model on different data sets, and obtain a more reliable cross-validation score. Based on the cross-validation score, the initial polynomial regression model is hyperparameter tuned. By adjusting the order of the polynomial and the regularization parameter , find the best configuration of the model. The commonly used method is the grid search algorithm, which finds the optimal parameter combination by exhaustively searching in the preset parameter grid. Grid search can effectively evaluate the impact of different parameters on model performance, thereby optimizing the polynomial order and regularization coefficient to obtain the initial relationship function.

[0068] In one example, the temperature impact factor is calculated according to the temperature-output relationship function, including:

[0069] Numerically differentiate the temperature-output relationship function to obtain the sensitivity function of temperature to UV output;

[0070] Based on the sensitivity function, the temperature change rate is calculated, and the temperature data is processed by sliding window to obtain a dynamic temperature change rate sequence;

[0071] According to the dynamic temperature change rate sequence, the acceleration of temperature change is calculated, and the temperature change rate is subjected to second-order difference to obtain the temperature change acceleration sequence;

[0072] Based on the temperature change acceleration sequence, a temperature impact assessment matrix is ​​constructed, and the temperature, temperature change rate and temperature change acceleration are combined into a multi-dimensional feature vector;

[0073] Perform singular value decomposition on the temperature impact assessment matrix to obtain the temperature impact characteristics after dimensionality reduction;

[0074] Based on the temperature influence characteristics after dimensionality reduction, the cumulative effect of temperature influence is calculated, and the cumulative temperature influence factor is obtained by exponentially weighted moving average;

[0075] According to the cumulative temperature impact factor, a temperature impact prediction model is constructed and the short-term temperature impact prediction value is calculated. The short-term temperature impact prediction value and the cumulative temperature impact factor are weightedly fused to obtain the temperature impact factor.

[0076] In this example, the temperature-output relationship function is numerically differentiated to obtain the sensitivity function of temperature to UV output. The temperature-output relationship function usually indicates the effect of temperature on the output intensity of the UV LED, while the sensitivity function reveals how a small change in temperature affects the UV output intensity. Assuming the temperature-output relationship function is ,in is the temperature, is the UV output intensity, the sensitivity function can be obtained by Perform first-order numerical differentiation to obtain:

[0077] ;

[0078] in, is the temperature sensitivity function, which describes the immediate impact of temperature change on UV output intensity. The temperature change rate is calculated based on the sensitivity function. The temperature change rate is an important indicator to describe how fast the temperature changes over time. The temperature data is processed by sliding window to obtain a dynamic temperature change rate sequence. The sliding window process divides the temperature data into fixed time windows and calculates the temperature change rate in each window. Set the temperature sequence to , then the temperature change rate in the sliding window can be expressed as:

[0079] ;

[0080] in, Indicates at time The temperature change rate at time, is the length of the time window. Through sliding window processing, a continuous dynamic temperature change rate sequence is obtained, which can reflect the gradual change of temperature over time. According to the dynamic temperature change rate sequence, the acceleration of temperature change is calculated. The calculation of acceleration can be achieved by taking the second-order difference of the temperature change rate. Suppose the temperature change rate sequence is , then the temperature change acceleration can be expressed as:

[0081] ;

[0082] in, Indicates the acceleration of temperature change, indicating the rate of temperature change over time The increase and decrease trend at each moment. Through calculation, the temperature change acceleration sequence is obtained. Based on the temperature change acceleration sequence, the temperature impact assessment matrix is ​​constructed. The temperature, temperature change rate and temperature change acceleration are combined into a multi-dimensional feature vector to form an assessment matrix that can fully reflect the temperature change characteristics. Assuming the temperature is , the temperature change rate is , the temperature change acceleration is , then the evaluation matrix can be expressed as:

[0083] ;

[0084] in, is the temperature impact assessment matrix, each row represents a multidimensional vector of temperature and its change characteristics at a certain moment. The temperature impact assessment matrix is ​​reduced in dimension. The singular value decomposition technique is used to extract the most important features by decomposing the matrix to remove noise and redundant information, and obtain the temperature impact characteristics after dimensionality reduction. Assume that the matrix It can be broken down into:

[0085] ;

[0086] in, and are orthogonal matrices, It is a diagonal matrix containing the singular values ​​of the matrix. By selecting the first few largest singular values ​​and their corresponding eigenvectors, a low-dimensional matrix is ​​reconstructed to retain the most important temperature influence features. Based on the temperature influence features after dimensionality reduction, the cumulative effect of temperature influence is calculated. The cumulative effect of temperature influence refers to the long-term comprehensive effect of temperature on UV output. The exponentially weighted moving average (EWMA) method is used to calculate the cumulative temperature influence factor. The calculation formula of EWMA is:

[0087] ;

[0088] in, For in time The cumulative temperature influence factor is is a smoothing factor, usually between 0 and 1. Through recursive calculation, EWMA can give greater weight to recent temperature impacts while retaining the cumulative effect of historical data to obtain a cumulative temperature impact factor that reflects long-term trends. Based on the cumulative temperature impact factor, a temperature impact prediction model is constructed, and the short-term temperature impact prediction value is calculated. The short-term temperature impact prediction value can be predicted by the current temperature impact characteristics and cumulative factors. The predicted value is weighted and fused with the cumulative temperature impact factor to obtain a more accurate temperature impact factor. Assuming the short-term prediction value is , the accumulation factor is , then the weighted fusion temperature influence factor can be expressed as:

[0089] ;

[0090] in, and is a weighting coefficient that determines the contribution of short-term and long-term predictions to the final temperature impact factor. The temperature impact factor is used to adjust the operating parameters of the UV LED module in real time to ensure that it maintains optimal performance under different temperature conditions.

[0091] In one example, a deep neural network UV assessment model is constructed based on temperature influencing factors and historical temperature datasets, including:

[0092] The input layer of the deep neural network UV assessment model is constructed, which adopts a multimodal input structure, including temperature data input channel and temperature influence factor input channel. The temperature data input dimension is n×m, n is the time step, m is the number of temperature sensors, and the temperature influence factor input dimension is k×1, k is the number of influence factors.

[0093] The feature extraction module of the deep neural network UV assessment model is constructed, including three parallel convolutional neural network branches: the first branch contains 3 one-dimensional convolutional layers, with convolution kernel sizes of 3, 5, and 7, respectively, and 64 convolution kernels in each layer; the second branch contains 3 one-dimensional convolutional layers, with convolution kernel sizes of 5, 7, and 9, respectively, and 64 convolution kernels in each layer; the third branch contains 3 one-dimensional convolutional layers, with convolution kernel sizes of 7, 9, and 11, respectively, and 64 convolution kernels in each layer; each convolution layer is followed by a batch normalization layer and a rectified linear unit activation function;

[0094] The multi-head self-attention mechanism is adopted to construct an 8-head self-attention layer of the deep neural network UV evaluation model. The dimension of each attention head is 64. The attention weight is obtained by scaling the dot product attention calculation, and the outputs of different heads are spliced ​​and linearly transformed.

[0095] Construct the time series modeling module of the deep neural network UV evaluation model, using a bidirectional long short-term memory network structure, including 2 layers of bidirectional long short-term memory units, 128 hidden nodes in each layer, and each long short-term memory unit contains a forget gate, input gate, output gate and unit state;

[0096] Construct the global average pooling layer and global maximum pooling layer of the deep neural network UV evaluation model, with the pooling window size equal to the time step, and concatenate the outputs of the two in the feature dimension;

[0097] The fully connected layer of the deep neural network UV evaluation model is constructed, which includes 3 fully connected layers: the first layer uses 256 neurons and adopts the rectified linear unit activation function; the second layer uses 128 neurons and adopts the rectified linear unit activation function; the third layer uses 64 neurons and adopts the hyperbolic tangent activation function; each layer is followed by a dropout layer with a dropout rate of 0.5;

[0098] Add a residual connection after the fully connected layer to perform element-by-element addition operation on the original features of the input layer and the output of the fully connected layer to achieve short-circuit transmission of information;

[0099] The output layer is constructed based on the output of the residual connection, and the softmax activation function is used to generate the probability distribution of the UV output control parameters. The output dimension is q, where q is the number of possible UV output control parameter categories.

[0100] The temperature influencing factors and the historical temperature data set are preprocessed to obtain a preprocessed training data set, and based on the preprocessed training data set, the constructed deep neural network UV evaluation model is trained using a cross entropy loss function and an adaptive moment estimation optimizer to obtain a trained deep neural network UV evaluation model.

[0101] In this example, the input layer is designed to adopt a multimodal input structure to make full use of different types of data sources. The input layer contains two main input channels: temperature data input channel and temperature influencing factor input channel. The dimension of the temperature data input channel is ,in represents the time step, Represents the number of temperature sensors; the dimension of the temperature influence factor input channel is ,in is the number of influencing factors. This design allows the model to process temperature data and other key influencing factor data from multiple sensors simultaneously, thereby improving the accuracy of UV output characteristic prediction. The model enters the feature extraction module, which consists of three parallel one-dimensional convolutional neural network (CNN) branches, each of which consists of three one-dimensional convolutional layers. The convolution kernel sizes of the first branch are 3, 5, and 7, the convolution kernel sizes of the second branch are 5, 7, and 9, and the convolution kernel sizes of the third branch are 7, 9, and 11. Each convolution layer is configured with 64 convolution kernels, and the convolution operation is followed by a batch normalization layer and a rectified linear unit (ReLU) activation function. Batch normalization helps to accelerate training and improve the stability of the model, while the ReLU activation function introduces nonlinearity, enabling the model to capture the complex nonlinear relationship between temperature data and UV output. After feature extraction is completed, the model uses a multi-head self-attention mechanism to strengthen the processing of input features. The model constructs an 8-head self-attention layer with a dimension of 64 for each attention head. The self-attention mechanism calculates the scaled dot product between the query matrix (Query) and the key matrix (Key) and performs a softmax operation to obtain the attention weight. The formula for this operation is:

[0102] ;

[0103] in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key matrix. The calculated attention weights are used to weight the value matrix to obtain the output results of each head, and these outputs are concatenated and linearly transformed. The multi-head self-attention mechanism allows the model to focus on different parts of the input features, capture important information patterns, and improve prediction accuracy. The time series modeling module of the deep neural network UV evaluation model is constructed. This module adopts a bidirectional long short-term memory network (BilSTM) structure, which contains 2 layers of bidirectional LSTM units, each with 128 hidden nodes. Each LSTM unit consists of a forget gate, an input gate, an output gate, and a unit state, which can effectively capture the front-to-back dependencies in the time series data. Through the bidirectional LSTM, the model can not only consider the past information of the temperature data, but also use future information to improve the prediction ability of the UV output characteristics. The model extracts global features through a global average pooling layer and a global maximum pooling layer. The window size of the two pooling layers is equal to the time step. By calculating the average and maximum values ​​of the time series data respectively, the model can extract a more compact and representative feature vector. The outputs of the global average pooling and the maximum pooling are concatenated in the feature dimension and input into the fully connected layer. The fully connected layer consists of three layers: the first layer uses 256 neurons and adopts the ReLU activation function; the second layer uses 128 neurons and also uses the ReLU activation function; the third layer uses 64 neurons and uses the hyperbolic tangent (Tanh) activation function. Each fully connected layer is followed by a dropout layer, and the dropout rate is set to 0.5 to reduce the risk of overfitting. In order to retain the information of the original features of the input layer, a residual connection is added after the fully connected layer. Through element-by-element addition operations, the original features of the input layer are fused with the output of the fully connected layer to ensure that the information is not lost during the transmission process in the network. The formula for the residual connection is:

[0104] ;

[0105] in, is the output of the fully connected layer, is the feature of the input layer. Through the residual connection, the model realizes the short-circuit transmission of information, avoiding the gradient vanishing problem that may occur in the deep network. Based on the output of the residual connection, the output layer of the model is constructed. The output layer uses the softmax activation function to generate the probability distribution of the UV output control parameters. The softmax function can convert the output of the model into a probability distribution, and its formula is:

[0106] ;

[0107] in, For the The probability of class output parameters, is the linear combination output of the fully connected layer, is the number of possible UV output control parameter categories. Through the softtmax operation, the model can normalize the output parameters of each category so that the sum of the probabilities of all categories is 1, which is convenient for classification decisions. The temperature influencing factors and historical temperature data sets are preprocessed. The preprocessing steps include data normalization, denoising and time series partitioning to obtain the preprocessed training data set. The purpose of preprocessing is to improve the quality of the data and ensure that the model can fully learn the effective information of the data during the training process. The preprocessed training data set is input into the deep neural network UV evaluation model, and the cross entropy loss function is used as the loss metric, combined with the adaptive moment estimation optimizer (Adam) to train the model. The Adam optimizer is an adaptive learning rate optimization algorithm based on first-order and second-order momentum, and its update rule is:

[0108] ;

[0109] in, are model parameters, is the learning rate, and are the estimates of the first-order and second-order momentum, is a small constant used to prevent division by zero errors. In this way, the Adam optimizer can dynamically adjust the learning rate to speed up the convergence of the model. After multiple rounds of training, the trained deep neural network UV evaluation model is finally obtained. The model can accurately predict the output control parameters of the UV LED module based on the input temperature data and temperature influencing factors, thereby achieving precise temperature control and ensuring the efficient operation of the UV LED module under different working conditions.

[0110] In one example, the temperature influencing factor and the historical temperature data set are preprocessed to obtain a preprocessed training data set, and based on the preprocessed training data set, a cross entropy loss function and an adaptive moment estimation optimizer are used to train the constructed deep neural network UV evaluation model to obtain a trained deep neural network UV evaluation model, including:

[0111] The temperature influencing factors and the historical temperature data set are normalized to obtain a normalized data set, and based on the normalized data set, a sliding window method is used to perform time series division to obtain a preprocessed training data set;

[0112] Input the preprocessed training data set into the input layer of the deep neural network UV evaluation model to obtain the temperature data input matrix and the temperature influence factor input vector;

[0113] The temperature data input matrix and the temperature influencing factor input vector are respectively subjected to feature extraction through three parallel one-dimensional convolution branches. The output of each branch is a feature matrix. The feature matrices of the three branches are then concatenated in the feature dimension to obtain a concatenated feature matrix.

[0114] Perform multi-head self-attention calculation on the concatenated feature matrix. The specific steps include: multiplying the concatenated feature matrix by three learnable weight matrices to obtain the query matrix, key matrix and value matrix; calculating the dot product of the query matrix and the key matrix, and performing scaling and softmax operations to obtain the attention weights; multiplying the attention weights by the value matrix to obtain the weighted feature representation;

[0115] Based on the weighted feature representation, the time series modeling is performed through the bidirectional long short-term memory network to obtain the forward hidden state and the backward hidden state, and the forward hidden state and the backward hidden state are concatenated to obtain the time series feature matrix;

[0116] Perform global average pooling and global maximum pooling operations on the time series feature matrix to obtain two feature vectors. After concatenating the two feature vectors, process them through a fully connected layer to obtain a feature vector after dimensionality reduction.

[0117] Based on the reduced-dimensional feature vector, the probability distribution of the UV output control parameters is calculated by the softmax function. The specific steps include: multiplying the feature vector by the output layer weight matrix and adding the bias vector; performing a softmax operation on the result to obtain the probability of each UV output control parameter category;

[0118] Calculate the cross entropy loss between the probability distribution and the true label to get the loss function value;

[0119] Based on the loss function value, the adaptive moment estimation optimizer is used to update the parameters to obtain the trained deep neural network UV evaluation model.

[0120] In this example, the temperature influencing factors and historical temperature data sets are normalized to eliminate the dimensional differences between different features, so that all feature data are scaled to the same range to avoid certain features from having too much impact on model training, and obtain the normalized data set. The sliding window method is used to divide the normalized data set into time series. The sliding window method divides the time series data into multiple fixed-length time segments, each of which can be regarded as an independent sample for model training. Set the window length to , the time series data is , the sliding window method can be expressed as:

[0121] ;

[0122] The preprocessed training dataset obtained by sliding window processing can capture local trends in time series data and provide the model with richer samples for learning the dynamic impact of temperature on UV output. After the preprocessed training dataset is input into the input layer of the deep neural network UV evaluation model, the data is divided into two main parts: the temperature data input matrix and the temperature impact factor input vector. The temperature data input matrix contains multiple time steps and sensor data, while the temperature impact factor input vector summarizes the impact factor data at different time points. The feature extraction module of the model consists of three parallel one-dimensional convolutional neural network (CNN) branches, each of which independently performs convolution operations on the temperature data input matrix and the temperature impact factor input vector. The output of each convolution layer is a feature matrix that captures data features at different scales and modes. By splicing the feature matrices of these three branches on the feature dimension, a richer spliced ​​feature matrix is ​​formed. Multi-head self-attention calculation is performed on the spliced ​​feature matrix. The self-attention mechanism identifies the important parts of the input features by calculating the interaction between the query matrix (Query), the key matrix (Key), and the value matrix (Value). The concatenated feature matrix is ​​multiplied by the three learnable weight matrices to obtain the query matrix, key matrix, and value matrix. The dot product of the query matrix and the key matrix is ​​calculated and scaled. The formula is:

[0123] ;

[0124] in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key matrix. Through the softmax operation, the attention weights obtained are used to weight the value matrix and generate a weighted feature representation. The time series modeling module uses a bidirectional long short-term memory network (BiLSTM) to simultaneously process the input sequence through forward and backward LSTM units to capture the bidirectional dependencies in the time series data. The output of BiLSTM includes a forward hidden state and a backward hidden state, which are spliced ​​together to form a time series feature matrix. Bidirectional time series modeling can better understand the complex pattern of temperature data changing over time and improve the prediction ability of UV output. Global average pooling and global maximum pooling operations are performed on the time series feature matrix. The window size of the global pooling operation is equal to the time step. By calculating the average and maximum values ​​of the sequence, two feature vectors are obtained respectively. These two feature vectors capture the global trend of the time series data. The pooled results are spliced ​​on the feature dimension to obtain the spliced ​​feature vector. Based on the spliced ​​feature vector, the model further reduces the dimension through the fully connected layer. The fully connected layer performs a linear transformation on the input features and uses an activation function for nonlinear processing. Assume that the output of the fully connected layer is , the weight matrix is , the bias vector is , then the linear transformation can be expressed as:

[0125] ;

[0126] in, is the input feature vector. After being processed by the activation function, these output feature vectors enter the final softmax layer, and the probability distribution of the UV output control parameters is calculated by the softmax function. The function of the softmax function is to convert the output value of the fully connected layer into a probability distribution, and its formula is:

[0127] ;

[0128] in, Indicates that the UV output control parameters belong to The probability of the class, is the number of categories of output control parameters. In this way, the model can generate the probability of various output control parameters based on the input data and achieve accurate prediction. Calculate the cross entropy loss between the probability distribution and the true label. The cross entropy loss function measures the difference between the model prediction value and the true value, and its formula is:

[0129] ;

[0130] in, is the binary value of the true label, is the prediction probability of the model. By minimizing the cross entropy loss, the model can gradually optimize its parameters and improve the prediction accuracy. Based on the cross entropy loss value, the adaptive moment estimation optimizer (Adam) is used to update the parameters. The Adam optimizer combines the first-order and second-order momentum methods, which can dynamically adjust the learning rate, accelerate the convergence process of the model, and avoid falling into the local optimal solution. The update rule of the Adam optimizer is:

[0131] ;

[0132] in, are model parameters, is the learning rate, and are the estimates of the first-order and second-order momentum, is a small constant used to prevent division by zero errors. Through repeated iterative optimization, a trained deep neural network UV evaluation model is obtained, which can accurately predict the output control parameters of the UV LED module and realize intelligent temperature control and optimization.

[0133] In one example, based on the loss function value, an adaptive moment estimation optimizer is used to update parameters to obtain a trained deep neural network UV evaluation model, including:

[0134] Calculate the gradient of the loss function value with respect to each parameter of the deep neural network UV evaluation model to obtain the gradient matrix, and based on the gradient matrix, update the first-order momentum estimate to obtain the first-order momentum vector;

[0135] The gradient matrix is ​​squared to obtain a square gradient matrix, and based on the square gradient matrix, the second-order momentum estimate is updated to obtain a second-order momentum vector;

[0136] Performing deviation correction on the first-order momentum vector to obtain a corrected first-order momentum vector, and performing deviation correction on the second-order momentum vector to obtain a corrected second-order momentum vector;

[0137] Based on the corrected first-order momentum vector and the corrected second-order momentum vector, the adaptive learning rate is calculated to obtain the adaptive learning rate vector of each parameter;

[0138] According to the adaptive learning rate vector and the modified first-order momentum vector, the parameter update amount is calculated to obtain the parameter update vector;

[0139] Based on the parameter update vector, the parameters of each layer of the deep neural network UV evaluation model are updated until the preset number of training rounds is reached or the early stopping condition is met, and a trained deep neural network UV evaluation model is obtained.

[0140] In this example, the gradient of the loss function value with respect to each parameter is calculated for the error between the model's output and the target value. This process can be implemented by the back propagation algorithm, and the loss function is usually a cross entropy loss or mean square error (MSE). Assume that the loss function is ,in is the parameter set of the model, then for each parameter The gradient of can be expressed as:

[0141] ;

[0142] in, For the time step The gradient matrix of Indicates the parameters The gradient operation. The gradient matrix contains the derivative information of the loss function with respect to each parameter, reflecting the sensitivity of the current parameter configuration to the loss. Based on the gradient matrix, the first-order momentum estimate is updated to obtain the first-order momentum vector. Momentum is an important concept in the optimization process, which is used to accelerate the convergence in the direction of gradient descent and reduce oscillation in other directions. The update formula of the first-order momentum estimate is:

[0143] ;

[0144] in, is the current time step The first-order momentum vector of is the exponential decay rate of the momentum term, which is generally set at around 0.9. is the current gradient matrix. This formula takes a weighted average of the current gradient and the previous momentum, allowing the model to smoothly follow the gradient direction when updating parameters, avoiding excessive parameter update steps. After obtaining the first-order momentum vector, the gradient matrix is ​​squared to capture the magnitude of the gradient change and obtain the gradient square matrix. The squared gradient is used to update the second-order momentum estimate, which represents the variance information of the gradient. The update formula for the second-order momentum estimate is:

[0145] ;

[0146] in, is the current time step The second-order momentum vector of is the exponential decay rate of the second-order momentum, usually set at around 0.999, Represents the square operation of the gradient matrix. The role of the second-order momentum vector is to capture the fluctuation of the gradient, making the model more robust when updating parameters and preventing drastic changes in parameters caused by excessive gradients. After updating the momentum estimate, in order to improve the accuracy of the algorithm in the initial stage, the first-order and second-order momentum vectors are corrected for deviations. Since the momentum may cause deviations in the initial stage due to the initial value being set to zero, the first-order momentum vector can be expressed as follows after adjustment through the deviation correction formula:

[0147] ;

[0148] The corrected second-order momentum vector is expressed as:

[0149] ;

[0150] in, and are the modified first-order and second-order momentum vectors, respectively, and is the momentum decay rate at The correction operation makes the momentum vector more accurate in the initial stage of training, avoiding the slow convergence problem caused by insufficient initial estimation. Based on the corrected first-order momentum vector and second-order momentum vector, the adaptive learning rate is calculated to obtain the adaptive learning rate vector of each parameter. The adaptive learning rate is a method of adjusting the update step of each parameter so that the parameter update can flexibly adapt to different gradient changes in each dimension. The calculation formula of the adaptive learning rate is:

[0151] ;

[0152] in, is the time step The adaptive learning rate vector, is the global learning rate, is a small constant used to prevent division by zero errors. In this way, the learning rate of each parameter is automatically adjusted according to the fluctuation of its gradient, avoiding unstable updates caused by excessive gradients, and also speeding up convergence. The parameter update amount is calculated based on the adaptive learning rate vector and the corrected first-order momentum vector. The parameter update amount is the parameter value that the model needs to adjust in each step, which determines the change of the model parameters in this time step. The calculation formula for the parameter update amount is:

[0153] ;

[0154] in, is the current time step The parameter update amount, is the adaptive learning rate vector, is the corrected first-order momentum vector. The update amount represents the movement amplitude of the model parameters in the current gradient direction, so that the model can be closer to the global optimal solution in the next time step. Based on the parameter update vector, the parameters of each layer of the deep neural network UV evaluation model are updated. This process continues until the preset number of training rounds is reached or the early stopping condition is met. The early stopping condition is usually that the loss of the validation set no longer decreases or starts to increase. At this point, the update process of the model parameters stops, and the trained deep neural network UV evaluation model achieves a good fitting effect on the given data set.

[0155] In one example, a deep neural network UV evaluation model is used to predict UV output control parameters for a real-time temperature data set to obtain multiple UV output control parameters, and a corresponding intelligent temperature control instruction sequence is created according to the multiple UV output control parameters, including:

[0156] Preprocessing the real-time temperature data set, including normalization processing and time window division, to obtain a preprocessed real-time temperature data set;

[0157] The preprocessed real-time temperature data set is input into the deep neural network UV evaluation model. After feature extraction, self-attention calculation, time series modeling and full connection layer processing, the probability distribution of UV output control parameters is obtained.

[0158] Based on the probability distribution of UV output control parameters, the maximum a posteriori probability method is used to select the optimal UV output control parameters, and the UV output control parameters of multiple target partitions are obtained;

[0159] Smoothing the UV output control parameters of multiple target partitions to obtain a smoothed UV output control parameter sequence;

[0160] Based on the smoothed UV output control parameter sequence, a temperature control state machine is constructed, including normal state, overheating state and cooling state, and the state transition condition is determined according to the current temperature and UV output control parameters;

[0161] According to the temperature control state machine and the smoothed UV output control parameter sequence, the corresponding PWM modulation signal and cooling system control instruction are generated to form an intelligent temperature control instruction sequence.

[0162] In this example, the real-time temperature data set is normalized to ensure that temperature data from different sources are processed on the same scale to avoid model errors caused by dimensional differences. The real-time temperature data set is divided into time windows to divide the continuous time series data into multiple segments so that the model can better capture the short-term dynamic characteristics of temperature changes. The length of the time window is set to , real-time temperature data is , it can be expressed by the sliding window method:

[0163] ;

[0164] The preprocessed real-time temperature data set is input into the deep neural network UV evaluation model, and the model extracts features from the data. A one-dimensional convolutional neural network (CNN) is used to extract multi-scale features from the temperature data through convolution kernels of different sizes. Each convolution kernel can capture the temperature change characteristics at different time steps and convert them into higher-level feature representations. These features then enter the self-attention mechanism to enhance the model's attention to important features. After time window division, the preprocessed real-time temperature data set obtained can provide rich feature information and capture the changing trend of temperature at different time points. The self-attention mechanism identifies the temperature features that have the greatest impact on UV output by calculating the relationship between the query matrix (Query), the key matrix (Key), and the value matrix (Value). The core formula for self-attention calculation is:

[0165] ;

[0166] in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key matrix. By calculating the dot product of the query and the key and performing scaling and softmax processing, the attention weights obtained are used to weight the value matrix to generate a weighted feature representation. The model enters the time series modeling stage and uses a bidirectional long short-term memory network (BiLSTM) to model the time series characteristics of the temperature data. The bidirectional LSTM can simultaneously consider the forward and backward dependencies of the time series data. The forward LSTM captures past temperature change information, and the backward LSTM captures future temperature change information. The outputs of the two are spliced ​​together to form a time series feature matrix. After obtaining the time series feature matrix, the model reduces the dimensionality of the high-dimensional feature vector through the fully connected layer, and calculates the probability distribution of the UV output control parameters through the softmax function. The softmax function converts the output of the fully connected layer into a probability distribution, and its formula is:

[0167] ;

[0168] in, Indicates The probability of UV-like output control parameters, is the linear combination output of the fully connected layer, is the number of possible UV output control parameter categories. Through the softmax operation, the model can classify each possible output parameter and generate a probability distribution for each category. Based on the probability distribution of the UV output control parameters, the maximum a posteriori probability (MAP) method is used to select the optimal UV output control parameters. The maximum a posteriori probability method calculates the posterior probability of each category and selects the parameter with the highest posterior probability as the final output control parameter. The formula of the MAP method is:

[0169] ;

[0170] in, For the optimal UV output control parameters, For Category The posterior probability of the model. This method can ensure that the model selects the most likely control parameters under uncertainty and improve the reliability of decision-making. The UV output control parameters of multiple target partitions are smoothed to eliminate fluctuations between parameters and improve the stability of the system. The smoothing process usually adopts the moving average method or the exponentially weighted moving average method. Through the smoothing operation, a smoother UV output control parameter sequence can be obtained, which reduces the frequent adjustment of the control system and improves the stability and response speed of the overall system. Based on the smoothed UV output control parameter sequence, a temperature control state machine is constructed. The temperature control state machine is used to describe the working state of the UV LED module under different temperature conditions, mainly including normal state, overheating state and cooling state. According to the current temperature and UV output control parameters, the state machine can determine the transition conditions between each state. For example, when the temperature exceeds the set threshold, the state machine will switch from the normal state to the overheating state and trigger the cooling mechanism; when the temperature falls back to a safe range, the state machine will switch back to the normal state. According to the temperature control state machine and the smoothed UV output control parameter sequence, the corresponding PWM modulation signal and cooling system control instructions are generated. The PWM modulation signal is used to adjust the input current of the UV LED module to control its output intensity; the cooling system control command is used to control the working state of the cooling device to keep the UV LED module running within a suitable temperature range. The generation of the PWM signal can be achieved by the following formula:

[0171] ;

[0172] Among them, PWM represents the duty cycle, is the current input voltage, and The minimum and maximum input voltages are set respectively. By adjusting the duty cycle of the PWM signal, the output of the UV LED module can be precisely controlled. The intelligent temperature control instruction sequence formed can effectively adjust the temperature and output intensity of the UV LED module to ensure that it can maintain optimal performance under different working conditions.

[0173] In one example, based on the smoothed UV output control parameter sequence, a temperature control state machine is constructed, including a normal state, an overheating state, and a cooling state, and a state transition condition is determined according to the current temperature and the UV output control parameter, including:

[0174] Based on the smoothed UV output control parameter sequence, the temperature thresholds of the normal state, the overheating state and the cooling state are set to obtain the state transition temperature threshold set;

[0175] Dynamically adjust the state transition temperature threshold set, and obtain the adaptive state transition temperature threshold according to the historical temperature data and the statistical characteristics of the UV output control parameters;

[0176] Based on the adaptive state transition temperature threshold, a state transition rule matrix is ​​constructed to define the transition conditions and transition probabilities between states, and a state transition model is obtained.

[0177] Initialize the state transition model, determine the initial state according to the current temperature and UV output control parameters, and obtain the initialized temperature control state machine;

[0178] Based on the initialized temperature control state machine, the state retention time and state jump penalty mechanism are designed to avoid frequent state switching and obtain a stable temperature control state machine;

[0179] The stable temperature control state machine is learned and optimized online. According to the real-time temperature data and control effect feedback, the state transition conditions and control strategies are dynamically adjusted to obtain an adaptively optimized temperature control state machine.

[0180] In this example, the temperature thresholds for normal state, overheating state, and cooling state are set. The temperature threshold is the basis for the system to judge the operating state of the UV LED module. By analyzing the smoothed UV output control parameter sequence, it is determined within what temperature range the UV LED module operates normally, and under what circumstances overheating protection or cooling measures need to be initiated. The process of setting the temperature threshold can be implemented in the following way: Assume that the temperature threshold set is ,in is the upper limit temperature of normal state, The lower limit temperature of overheating is the trigger temperature of the cooling state. The initial state transition temperature threshold set is dynamically adjusted to adapt to changes under different working conditions. By analyzing the statistical characteristics of historical temperature data and UV output control parameters, a more reasonable adaptive state transition temperature threshold is obtained. These characteristics include the mean, variance, peak value, etc. of the temperature, as well as the stability and change trend of the UV output control parameters. Dynamic adjustment infers future temperature change trends based on historical data and adjusts the threshold accordingly, making the system more flexible and accurate in actual operation. The adjusted adaptive state transition temperature threshold can be expressed as , where each threshold is optimized based on the current operating state and historical data. Based on the adaptive state transition temperature threshold, a state transition rule matrix is ​​constructed. The state transition rule matrix is ​​used to define the transition conditions and transition probabilities between states to ensure that the system can automatically adjust the operating state under different temperature conditions. The state transition rule matrix can be expressed as:

[0181] ;

[0182] in, Represents the probability of transitioning from state A to state B. By defining the transition conditions and probabilities of each state, a complete state transition model is constructed. The model not only considers the instantaneous impact of temperature changes, but also integrates the transition probabilities between states to ensure that the system has a certain robustness when performing state transitions. Initialize the state transition model. Determine the initial state of the system based on the current temperature and UV output control parameters. Assume that the current temperature is , UV output control parameters are , then the initial state can be determined according to the following rules:

[0183] ;

[0184] in, The initial state is determined based on the current temperature and UV output control parameters, so that it can immediately enter the appropriate working mode at startup. The state holding time and state jump penalty mechanism are designed in the state transition model. The state holding time stipulates that after each state transition, the system must remain in this state for a certain period of time to avoid frequent state switching, which will lead to system instability or increase the hardware load. The state jump penalty mechanism introduces a penalty term to limit the possibility of frequent jumps in a short period of time and ensure the smoothness and consistency of state transitions. The setting of state holding time and jump penalty enables the system to remain stable in a dynamic environment and avoid frequent state switching due to short-term temperature fluctuations. Online learning and optimization are performed on the stable temperature control state machine. Online learning means that the system dynamically adjusts the state transition conditions and control strategies according to real-time temperature data and control effect feedback during operation. For example, the current temperature deviation and UV output deviation are calculated to determine whether the state transition rules need to be updated or the temperature threshold needs to be adjusted. Online optimization uses a feedback loop to compare the actual operation results with the model prediction results, and continuously adjusts the state transition conditions and temperature thresholds through incremental learning. After going through multiple feedback cycles, it can gradually converge to an optimal control strategy, allowing the temperature control state machine to adaptively respond to different working environments and load conditions.

[0185] Reference Figure 2 , this embodiment provides a temperature control device for a UV LED module, comprising:

[0186] The acquisition module 1 is used to collect temperature data of the UV LED module to obtain a historical temperature data set, and analyze the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function;

[0187] Calculation module 2, used for calculating the temperature influence factor according to the temperature-output relationship function;

[0188] Building module 3, for building a deep neural network UV evaluation model based on temperature influencing factors and historical temperature datasets;

[0189] Monitoring module 4, used for real-time temperature monitoring of multiple target partitions in the UV LED module, and obtaining a real-time temperature data set of each target partition;

[0190] The prediction module 5 is used to predict the UV output control parameters of the real-time temperature data set through the deep neural network UV evaluation model, obtain multiple UV output control parameters, and create a corresponding intelligent temperature control instruction sequence according to the multiple UV output control parameters.

[0191] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0192] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0193] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0194] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0195] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0196] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0197] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A temperature control method for a UV LED module, characterized in that: The following steps are involved: Collecting temperature data of the UV LED module to obtain a historical temperature data set, and analyzing the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function; Calculating the temperature influence factor according to the temperature-output relationship function; Constructing a deep neural network UV evaluation model based on the temperature influencing factor and the historical temperature data set; Perform real-time temperature monitoring on multiple target partitions in the UV LED module to obtain a real-time temperature data set for each target partition; The deep neural network UV evaluation model is used to predict UV output control parameters of the real-time temperature data set to obtain multiple UV output control parameters, and a corresponding intelligent temperature control instruction sequence is created according to the multiple UV output control parameters.

2. The temperature control method of the UV LED module according to claim 1, characterized in that: The temperature data of the UV LED module is collected to obtain a historical temperature data set, and the relationship between temperature and output characteristics of the historical temperature data set is analyzed to obtain a temperature-output relationship function, including: Installing a temperature sensor array in the UV LED module, and collecting a historical temperature data set of the UV LED module through the temperature sensor array; Performing normalization, denoising and interpolation processing on the historical temperature data set to obtain a preprocessed temperature data set; Based on the preprocessed temperature data set, corresponding UV LED module input current data and UV output intensity data are collected to construct a multi-dimensional output data set; Performing polynomial regression analysis on the multidimensional output data set, establishing a relationship model between temperature, input current and UV output intensity, and obtaining an initial relationship function; Performing parameter estimation on the initial relationship function, solving the polynomial regression coefficients using the least square method, and obtaining an optimized relationship function; The optimization relationship function is model verified, and the determination coefficient and the root mean square error are calculated. If the determination coefficient is greater than or equal to the preset threshold and the root mean square error is less than or equal to the preset threshold, the model is considered valid and the temperature-output relationship function is output.

3. The temperature control method of the UV LED module according to claim 2, characterized in that: The polynomial regression analysis is performed on the multidimensional output data set to establish a relationship model between temperature, input current and UV output intensity to obtain an initial relationship function, including: Performing feature selection on the multi-dimensional output data set to obtain a feature data set after dimensionality reduction; Based on the feature data set after dimensionality reduction, a polynomial feature matrix is ​​constructed, and polynomial expansion is performed on the temperature and current features to obtain an extended feature matrix; Regularizing the extended feature matrix, introducing an L2 regularization term using a ridge regression method, and obtaining a regularized feature matrix; Based on the regularized feature matrix, a polynomial regression analysis is performed, and a gradient descent algorithm is used to iteratively optimize the regression coefficients to obtain an initial polynomial regression model; Cross-validating the initial polynomial regression model, using a K-fold cross-validation method to evaluate the generalization ability of the model, and obtaining a cross-validation score; According to the cross-validation score, the initial polynomial regression model is subjected to hyperparameter tuning, and a grid search algorithm is used to optimize the polynomial order and the regularization coefficient to obtain an initial relationship function.

4. The temperature control method of the UV LED module according to claim 1, characterized in that: The calculating the temperature influence factor according to the temperature-output relationship function comprises: Numerically differentiating the temperature-output relationship function to obtain a sensitivity function of temperature to UV output; Based on the sensitivity function, the temperature change rate is calculated, and the temperature data is subjected to sliding window processing to obtain a dynamic temperature change rate sequence; Calculating the acceleration of temperature change according to the dynamic temperature change rate sequence, and performing second-order difference on the temperature change rate to obtain a temperature change acceleration sequence; Based on the temperature change acceleration sequence, a temperature impact assessment matrix is ​​constructed, and the temperature, the temperature change rate and the temperature change acceleration are combined into a multi-dimensional feature vector; Performing singular value decomposition on the temperature impact assessment matrix to obtain temperature impact features after dimensionality reduction; Based on the temperature influence characteristics after dimensionality reduction, the cumulative effect of temperature influence is calculated, and the cumulative temperature influence factor is obtained by using exponentially weighted moving average; According to the cumulative temperature impact factor, a temperature impact prediction model is constructed and a short-term temperature impact prediction value is calculated, and a weighted fusion is performed based on the short-term temperature impact prediction value and the cumulative temperature impact factor to obtain a temperature impact factor.

5. The temperature control method of the UV LED module according to claim 1, characterized in that: The constructing of a deep neural network UV evaluation model based on the temperature influencing factor and the historical temperature data set includes: The input layer of the deep neural network UV assessment model is constructed, which adopts a multimodal input structure, including temperature data input channel and temperature influence factor input channel. The temperature data input dimension is n×m, n is the time step, m is the number of temperature sensors, and the temperature influence factor input dimension is k×1, k is the number of influence factors. The feature extraction module of the deep neural network UV assessment model is constructed, including three parallel convolutional neural network branches: the first branch contains 3 one-dimensional convolutional layers, with convolution kernel sizes of 3, 5, and 7, respectively, and 64 convolution kernels in each layer; the second branch contains 3 one-dimensional convolutional layers, with convolution kernel sizes of 5, 7, and 9, respectively, and 64 convolution kernels in each layer; the third branch contains 3 one-dimensional convolutional layers, with convolution kernel sizes of 7, 9, and 11, respectively, and 64 convolution kernels in each layer; each convolution layer is followed by a batch normalization layer and a rectified linear unit activation function; The multi-head self-attention mechanism is adopted to construct an 8-head self-attention layer of the deep neural network UV evaluation model. The dimension of each attention head is 64. The attention weight is obtained by scaling the dot product attention calculation, and the outputs of different heads are spliced ​​and linearly transformed. Construct the time series modeling module of the deep neural network UV evaluation model, using a bidirectional long short-term memory network structure, including 2 layers of bidirectional long short-term memory units, 128 hidden nodes in each layer, and each long short-term memory unit contains a forget gate, input gate, output gate and unit state; Construct the global average pooling layer and global maximum pooling layer of the deep neural network UV evaluation model, with the pooling window size equal to the time step, and concatenate the outputs of the two in the feature dimension; The fully connected layer of the deep neural network UV evaluation model is constructed, which includes 3 fully connected layers: the first layer uses 256 neurons and adopts the rectified linear unit activation function; the second layer uses 128 neurons and adopts the rectified linear unit activation function; the third layer uses 64 neurons and adopts the hyperbolic tangent activation function; each layer is followed by a dropout layer with a dropout rate of 0.5; Adding a residual connection after the fully connected layer to perform element-by-element addition operation on the original features of the input layer and the output of the fully connected layer to achieve short-circuit transmission of information; An output layer is constructed based on the output of the residual connection, and a softmax activation function is used to generate a probability distribution of UV output control parameters, where the output dimension is q, and q is the number of possible UV output control parameter categories; The temperature influencing factor and the historical temperature data set are preprocessed to obtain a preprocessed training data set, and based on the preprocessed training data set, a cross entropy loss function and an adaptive moment estimation optimizer are used to train the constructed deep neural network UV evaluation model to obtain a trained deep neural network UV evaluation model.

6. The temperature control method of the UV LED module according to claim 5, characterized in that: The temperature influencing factor and the historical temperature data set are preprocessed to obtain a preprocessed training data set, and based on the preprocessed training data set, a cross entropy loss function and an adaptive moment estimation optimizer are used to train the constructed deep neural network UV evaluation model to obtain a trained deep neural network UV evaluation model, including: Normalizing the temperature influencing factors and the historical temperature data set to obtain a normalized data set, and based on the normalized data set, using a sliding window method to perform time series division to obtain a preprocessed training data set; Inputting the preprocessed training data set into the input layer of the deep neural network UV evaluation model to obtain a temperature data input matrix and a temperature influence factor input vector; The temperature data input matrix and the temperature influencing factor input vector are respectively subjected to feature extraction through three parallel one-dimensional convolution branches, the output of each branch is a feature matrix, and then the feature matrices of the three branches are spliced ​​in the feature dimension to obtain a spliced ​​feature matrix; Performing multi-head self-attention calculation on the concatenated feature matrix, the specific steps include: multiplying the concatenated feature matrix by three learnable weight matrices respectively to obtain a query matrix, a key matrix and a value matrix; calculating the dot product of the query matrix and the key matrix, and performing scaling and softmax operations to obtain attention weights; multiplying the attention weights by the value matrix to obtain a weighted feature representation; Based on the weighted feature representation, a bidirectional long short-term memory network is used to perform time series modeling to obtain a forward hidden state and a backward hidden state, and the forward hidden state and the backward hidden state are concatenated to obtain a time series feature matrix; Performing global average pooling and global maximum pooling operations on the time series feature matrix to obtain two feature vectors, concatenating the two feature vectors and processing them through a fully connected layer to obtain a feature vector after dimensionality reduction; Based on the reduced-dimensional feature vector, the probability distribution of the UV output control parameter is calculated by a softmax function, and the specific steps include: multiplying the feature vector by the output layer weight matrix and adding the bias vector; performing a softmax operation on the result to obtain the probability of each UV output control parameter category; Calculate the cross entropy loss between the probability distribution and the true label to obtain a loss function value; Based on the loss function value, an adaptive moment estimation optimizer is used to update parameters to obtain a trained deep neural network UV evaluation model.

7. The temperature control method of the UV LED module according to claim 6, characterized in that: Based on the loss function value, an adaptive moment estimation optimizer is used to update parameters to obtain a trained deep neural network UV evaluation model, including: Calculate the gradient of the loss function value with respect to each parameter of the deep neural network UV evaluation model to obtain a gradient matrix, and update the first-order momentum estimation based on the gradient matrix to obtain a first-order momentum vector; Squaring the gradient matrix to obtain a gradient square matrix, and updating the second-order momentum estimation based on the gradient square matrix to obtain a second-order momentum vector; Performing deviation correction on the first-order momentum vector to obtain a corrected first-order momentum vector, and performing deviation correction on the second-order momentum vector to obtain a corrected second-order momentum vector; Based on the modified first-order momentum vector and the modified second-order momentum vector, calculating an adaptive learning rate to obtain an adaptive learning rate vector of each parameter; Calculating a parameter update amount according to the adaptive learning rate vector and the modified first-order momentum vector to obtain a parameter update vector; Based on the parameter update vector, the parameters of each layer of the deep neural network UV evaluation model are updated until a preset number of training rounds is reached or an early stopping condition is met, thereby obtaining a trained deep neural network UV evaluation model.

8. The temperature control method of the UV LED module according to claim 1, characterized in that: The method of predicting UV output control parameters of the real-time temperature data set by using the deep neural network UV evaluation model to obtain a plurality of UV output control parameters, and creating a corresponding intelligent temperature control instruction sequence according to the plurality of UV output control parameters, comprises: Preprocessing the real-time temperature data set, including normalization processing and time window division, to obtain a preprocessed real-time temperature data set; The preprocessed real-time temperature data set is input into a deep neural network UV evaluation model, and after feature extraction, self-attention calculation, time series modeling and full connection layer processing, a probability distribution of UV output control parameters is obtained; Based on the probability distribution of the UV output control parameters, the maximum a posteriori probability method is used to select the optimal UV output control parameters to obtain UV output control parameters of multiple target partitions; Smoothing the UV output control parameters of the multiple target partitions to obtain a smoothed UV output control parameter sequence; Based on the smoothed UV output control parameter sequence, a temperature control state machine is constructed, including a normal state, an overheating state, and a cooling state, and a state transition condition is determined according to the current temperature and the UV output control parameter; According to the temperature control state machine and the smoothed UV output control parameter sequence, corresponding PWM modulation signals and cooling system control instructions are generated to form an intelligent temperature control instruction sequence.

9. The temperature control method of the UV LED module according to claim 8, characterized in that: The temperature control state machine is constructed based on the smoothed UV output control parameter sequence, including a normal state, an overheating state and a cooling state, and a state transition condition is determined according to the current temperature and the UV output control parameter, including: Based on the smoothed UV output control parameter sequence, setting temperature thresholds of a normal state, an overheating state, and a cooling state to obtain a state transition temperature threshold set; Dynamically adjusting the state transition temperature threshold set, and obtaining an adaptive state transition temperature threshold according to the historical temperature data and the statistical characteristics of the UV output control parameters; Based on the adaptive state transition temperature threshold, a state transition rule matrix is ​​constructed to define transition conditions and transition probabilities between states, thereby obtaining a state transition model; Initializing the state transition model, determining the initial state according to the current temperature and UV output control parameters, and obtaining an initialized temperature control state machine; Based on the initialized temperature control state machine, a state retention time and a state jump penalty mechanism are designed to avoid frequent state switching and obtain a stable temperature control state machine; The stable temperature control state machine is subjected to online learning and optimization, and the state transition conditions and control strategies are dynamically adjusted according to real-time temperature data and control effect feedback to obtain an adaptively optimized temperature control state machine.

10. A temperature control device for a UV LED module, characterized in that: The device is used to perform the temperature control method of the UV LED module according to any one of claims 1 to 9, comprising: The acquisition module is used to collect temperature data of the UV LED module to obtain a historical temperature data set, and analyze the relationship between temperature and output characteristics of the historical temperature data set to obtain a temperature-output relationship function; A calculation module, used for calculating a temperature influence factor according to the temperature-output relationship function; A construction module, used to construct a deep neural network UV evaluation model based on the temperature influencing factor and the historical temperature data set; A monitoring module, used to perform real-time temperature monitoring on multiple target partitions in the UV LED module, and obtain a real-time temperature data set for each target partition; A prediction module is used to predict UV output control parameters of the real-time temperature data set through the deep neural network UV evaluation model to obtain multiple UV output control parameters, and create corresponding intelligent temperature control instruction sequences according to the multiple UV output control parameters.

Images