Equipment temperature adjusting method fusing long-term and short-term memory network

Abstract

The invention discloses an equipment temperature adjusting method fusing a long short-term memory network, and particularly relates to the technical field of temperature control, which comprises the following steps: determining the deployment position of a sensor through thermal simulation, collecting multi-source heterogeneous data, dynamically adjusting the sampling frequency in combination with the temperature and the load current change rate, and adjusting the temperature of the sensor; after data preprocessing, an attention mechanism enhanced LSTM prediction model is constructed, a time sequence sample data set is divided according to equipment thermal response characteristics, training is carried out, and an optimal model is obtained through early stop mechanism optimization; predictive feedback double-closed-loop regulation and control is achieved based on the optimal model, outer loop PI control is combined with an integral separation mechanism to generate a basic control quantity, an inner loop outputs a correction quantity through fuzzification, reasoning and defuzzification, an actuator is driven after superposition, and safety linkage is triggered synchronously; constructing an incremental data buffer pool to screen effective samples, and adaptively updating model parameters by adopting a layered fine tuning strategy; the temperature regulation and control precision and the long-term self-adaptive capability are obviously improved, and the over-temperature risk of equipment is reduced.

Description

Technical Field

[0001] This invention relates to the field of temperature control technology, and more specifically, to a method for regulating device temperature by incorporating a long short-term memory network. Background Technology

[0002] With the advancement of technology, the integration and power density of various electronic equipment continue to increase. Local hotspot temperatures are prone to exceeding safety thresholds, leading not only to chip performance degradation and shortened lifespan but also the risk of system downtime. Temperature control has become a core guarantee for reliable equipment operation. Existing temperature control technologies largely revolve around traditional PID control, experience-based fuzzy rule reasoning, or basic time-series models such as ARIMA and shallow neural networks. These technologies collect data using temperature sensors deployed in fixed locations and trigger adjustment actions based on preset thresholds or static parameters, achieving initial applications in low-to-medium power equipment scenarios. However, with the increasing demand for high-power, long-cycle operation, the industry's requirements for control accuracy, dynamic response speed, and long-term adaptive capabilities have significantly increased, and existing technologies are gradually being adapted.

[0003] However, it still has some drawbacks in practical use, such as: 1. Fixed sampling frequency cannot match dynamic load changes of equipment: Existing methods mostly use preset sampling intervals. When the load changes suddenly or the temperature fluctuates rapidly, the data acquisition lag can easily lead to control delays. At low loads, redundant sampling will increase system overhead. 2. The prediction model is not good at capturing time-series features and has limited generalization ability: Traditional models rely on static features or shallow time-series modeling, which makes it difficult to explore the long-term relationship between temperature and load. When faced with scenarios such as equipment aging and environmental changes, the prediction accuracy drops sharply. 3. The control strategy is too simple and difficult to adapt to nonlinear characteristics: Relying solely on PID control or fuzzy rules cannot balance dynamic response speed and steady-state accuracy, and is prone to overshoot or oscillation in nonlinear scenarios of heat dissipation systems. 4. The model lacks an online update mechanism and has poor long-term adaptability: After the model is trained, the parameters are fixed and it cannot learn new data patterns generated during the operation of the equipment. As components age and dust accumulates, the control performance gradually deteriorates. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides a device temperature regulation method integrating a long short-term memory network, which solves the problems mentioned in the background art through the following solution.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a device temperature regulation method integrating a long short-term memory network, comprising: S1: Multi-source heterogeneous temperature data acquisition and adaptive preprocessing: The sensor deployment location is determined by thermal simulation modeling, and temperature data is collected synchronously to construct a temperature data vector; the temperature change rate and load current change rate are calculated by collecting the load current of the device, and the sampling frequency is dynamically adjusted based on the temperature change rate and load current change rate, and the collected data is cleaned and standardized. S2: Construction and training of LSTM prediction model with attention mechanism enhancement: The input window length and prediction step size are set according to the thermal response characteristics of the device, and a time series sample dataset is constructed based on the constructed temperature data vector, and divided into training set, validation set and test set according to time series; then the network structure is designed by using the attention and LSTM fusion architecture, and the model is trained and optimized, retaining the optimal model; S3: Predictive and feedback dual closed-loop adaptive control: Real-time prediction and deviation calculation are performed based on the optimal model after training and optimization; then, the outer loop PI control takes the dynamic deviation as input, and generates the basic control quantity and limits it by combining the integral separation mechanism; the inner loop fuzzy correction takes the static deviation and its rate of change as input, and outputs the correction quantity after fuzzification, inference and defuzzification, and superimposes the basic control quantity to obtain the final control quantity; based on the final control quantity, control execution and safety interlocking are performed. S4. Model Adaptive Update Based on Online Learning: Based on the data after control execution, an incremental data buffer pool is constructed, effective samples are screened and dynamically updated according to the first-in-first-out rule; a hierarchical fine-tuning strategy is adopted to adaptively update model parameters; and an update triggering and evaluation mechanism is set.

[0006] Preferably, S1 includes: By analyzing the internal thermal field distribution characteristics of the equipment through thermal simulation modeling, the priority deployment positions of the sensors are determined based on the thermal density gradient. The sensors are simultaneously triggered to collect temperature data at their corresponding positions and load current data during equipment operation. A multi-source data vector with timestamps is constructed to ensure the spatial representativeness and temporal consistency of the data collection. The global temperature change rate is calculated based on the collected temperature data, and the load current change rate is calculated based on the load current data. The sampling frequency is set according to the temperature change rate, and the high-frequency sampling mechanism is triggered by the load current change rate to achieve dynamic matching between the sampling frequency and the equipment operating status. Transmission and physical range verification are performed on the collected multi-source data to filter out erroneous data. Time series analysis is used to identify and remove outliers. An appropriate imputation method is selected according to the interval of missing data to ensure the continuity of time series. Statistical features are calculated based on historical valid data. Temperature data is converted into standardized data of a uniform scale, and statistical feature parameters are retained for inverse conversion of predicted values.

[0007] Preferably, the construction of the time-series sample dataset includes: Based on the thermal response characteristics of the equipment, the input window length and prediction step size are set. Based on the preprocessed standardized multi-source time series data, a continuous time series is extracted as the input sequence by a sliding window method. The input sequence includes temperature data and load current data within a preset time period. The device core temperature corresponding to the prediction step size in the input window is used as the sample label to construct sample pairs of input sequence and temperature label. The generated sample pairs are divided into training set, validation set and test set in chronological order. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model performance evaluation, and the test set is used to verify the generalization prediction ability of the model.

[0008] Preferably, the network structure design includes: An encoding, attention, and decoding fusion architecture is adopted, in which the encoding layer consists of a first-layer LSTM network, which is used to extract features from the input sequence in the constructed time-series sample dataset and output a hidden state sequence containing time-series correlation information; the first-layer LSTM network includes: input gate, forget gate, cell state update, output gate and hidden state; An attention mechanism layer is set between the encoding layer and the decoding layer. This attention mechanism layer takes the hidden state sequence output by the encoding layer as input, calculates the weight distribution of the hidden state at each time step, analyzes the key time step features that have a significant impact on the prediction results, and outputs a weighted fusion key temporal feature vector. The decoding layer consists of a second-layer LSTM network, which receives the key temporal feature vectors output by the attention mechanism layer and combines them with the initial hidden state to perform sequence decoding and generate high-dimensional hidden states. The network output layer adopts a fully connected structure, introduces nonlinearity through activation function, extracts key prediction features through dimensionality reduction, and then outputs standardized predicted temperature. Inverse standardization is used to convert the standardized predicted value into the actual temperature. The parameter dimension of the output layer is adapted to the prediction step size.

[0009] Preferably, the model training and optimization includes: Based on the training set in the constructed time-series sample dataset as input, a loss function with regularization is used to quantify the deviation between the predicted value and the sample label. The network parameters are iteratively updated through a gradient descent optimizer, where the regularization term is used to suppress model overfitting. During training, the learning rate is dynamically adjusted, and a step-decay strategy is used to balance the model convergence speed and parameter optimization accuracy. The model prediction performance is monitored in real time based on the validation set. When the validation set performance does not improve for a set of consecutive rounds, an early stop mechanism is triggered to avoid overtraining. The model's generalization ability in different training rounds is evaluated using a test set. The model with the best performance on the validation set and the smallest generalization error on the test set is selected as the final optimal model. At the same time, the model structure and corresponding parameters are saved for subsequent real-time prediction.

[0010] Preferably, the outer loop PI control includes: The dynamic deviation calculated based on the deviation is used as input. The basic control quantity is generated through the collaboration of the proportional term and the integral term. The proportional term is proportional to the current value of the dynamic deviation, and the integral term accumulates the historical dynamic deviation. An integral separation mechanism is introduced to dynamically adjust the integral coefficient according to the absolute value of the dynamic deviation: when the dynamic deviation exceeds the preset threshold, the integral coefficient is reduced to avoid integral saturation; when the dynamic deviation is within the preset threshold, the integral coefficient is restored. Set upper and lower limits for the cumulative amount of the integral term to limit the cumulative range of historical deviations; perform amplitude limiting on the control quantity after the proportional and integral terms are superimposed to keep it within the physical adjustment range of the actuator, and output the amplitude-limited basic control quantity to provide the initial adjustment basis for inner loop fuzzy correction.

[0011] Preferably, the inner ring blur correction includes: Based on the static deviation and its rate of change obtained from the deviation calculation as input, the precise values ​​of the static deviation and the rate of change of deviation are converted into corresponding fuzzy sets through a preset membership function. The fuzzy sets contain linguistic variables that reflect the degree of deviation. Reasoning is performed based on a preset fuzzy rule base, which consists of multiple mapping rules between input fuzzy states and output correction actions. Valid rules are determined by calculating the trigger strength of each rule. A defuzzing algorithm is used to convert the fuzzy output of effective rules into precise correction values. The defuzzing algorithm performs weighted calculations based on the trigger strength of each rule and the corresponding output center value. The precise correction value is superimposed on the basic control value output from the outer loop, and after amplitude limiting, the final control value is obtained, thereby realizing real-time dynamic correction of the basic control value to compensate for the nonlinear characteristics of the equipment and the model prediction error.

[0012] Preferably, the control execution and safety interlock includes: The final control quantity output based on the inner loop fuzzy correction is converted into an adjustment signal adapted to the actuator, which drives the actuator to act according to the intensity corresponding to the control quantity, thereby realizing the dynamic adjustment of the heat dissipation capacity of the equipment. The adjustment signal and the physical adjustment parameters of the actuator are in a preset correlation relationship. Multiple temperature safety thresholds are preset, including warning temperature thresholds and critical temperature thresholds. The warning temperature threshold is higher than the target temperature of the equipment, and the critical temperature threshold is higher than the warning temperature threshold. The system monitors the current core temperature of the equipment and the future temperature prediction value output by the model in real time. When any of the following conditions are met, a safety interlock mechanism is triggered: the current core temperature exceeds the critical temperature threshold, or the temperature prediction value for a preset time in the future exceeds the warning temperature threshold. Once the safety interlock mechanism is activated, the actuator operates at the maximum adjustment intensity and triggers an alarm signal. When the current core temperature of the equipment drops below the warning temperature threshold and the future temperature prediction value stabilizes within the safe range, the safety interlock is released, and the dual closed-loop adaptive control mode is restored.

[0013] Preferably, the incremental data buffer pool includes: A first-in-first-out buffer is constructed to store the valid samples required for model updates. The sample sources are multi-source time-series data collected after control execution and corresponding real temperature values, forming sample pairs of real-time input sequences and real temperature labels. The samples to be stored in the buffer pool are double-verified: first, sensor status verification and data physical rationality verification are performed to filter out data that are faulty or outside the normal operating range of the equipment; then, the model prediction error corresponding to the sample is calculated, and only samples whose prediction error meets the preset threshold are retained to ensure that the sample contains new feature information that the model needs to learn. Set a preset capacity for the buffer pool. When the sample storage reaches the capacity limit, delete the earliest stored sample in chronological order of timestamp and add a new qualified sample. Regularly clean up old samples in the buffer pool that have exceeded the preset storage time to maintain the consistency between the data in the buffer pool and the current operating characteristics of the equipment. The effective sample set output by the buffer pool is used for subsequent hierarchical fine-tuning of the model, providing a high-quality and timely data foundation for adaptive updates of model parameters, and avoiding the impact of noisy or outdated data on the model update effect.

[0014] Preferably, the adaptive update of model parameters includes: A hierarchical parameter adjustment strategy is adopted, which divides the model into frozen layers and fine-tuning layers based on the functional differences of each layer: the layer where the network coding layer learns general temporal features is set as the frozen layer, and its trained parameters are fixed to retain historical knowledge and avoid catastrophic forgetting; the layers where the attention layer, decoding layer and output layer are associated with the current operating characteristics of the device are set as fine-tuning layers, and only the parameters of this type of layer are updated. The effective sample set output from the incremental data buffer pool is used as the update data source. The fine-tuning layer is input in a mini-batch iterative manner, and the loss function with a learning rate and knowledge distillation constraint is configured to be lower than that in the initial training stage. The knowledge distillation constraint term is used to quantify the difference between the output of the new model and the output of the original optimal model, and to balance the new data adaptation ability and the old knowledge retention effect. During the iterative update process, the convergence status of the fine-tuning layer parameters and the sample prediction error are monitored in real time. When the error does not decrease or increases after a series of preset iterations, the current update is terminated. After the update is completed, the generalization error of the new model and the original model is compared on the independent validation sample set. The original model parameters are replaced only if the error of the new model is lower than that of the original model and meets the preset accuracy requirements. Otherwise, the update results are discarded and the original model parameters are kept unchanged.

[0015] The technical effects and advantages of this invention are as follows: 1. Dynamically adjust sampling frequency to match load and temperature changes: Combine the rate of temperature change and the rate of load current change to adjust the sampling frequency in real time. When the load changes abruptly, the sampling density is automatically increased to capture details, and the frequency is reduced to reduce overhead when the load is low, thus balancing data timeliness and system efficiency. 2. Attention-enhanced LSTM deeply captures temporal features and improves generalization ability: Through the encoding, attention, and decoding architecture, it focuses on key time step features, strengthens the long-term and short-term correlation learning between temperature and load, and significantly improves the prediction accuracy for scenarios such as equipment aging and environmental fluctuations. 3. Dual closed-loop control integrates PI and fuzzy correction to adapt to nonlinear characteristics: the outer loop PI control ensures the stability of basic regulation, while the inner loop fuzzy correction dynamically compensates for nonlinear deviations, taking into account both response speed and steady-state accuracy, and effectively suppressing overshoot and oscillation. 4. Online learning enables adaptive model updates and ensures long-term performance: Valid samples are selected through an incremental data buffer pool, and model parameters are fine-tuned in layers. While retaining historical knowledge, new features are learned, and high-precision control can be maintained during long-term operation, reducing the need for manual intervention. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0017] Figure 2 This is a schematic diagram of the S2 process of the present invention.

[0018] Figure 3 This is a schematic diagram of the S3 process of the present invention.

[0019] Figure 4 This is a schematic diagram of the S4 process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] refer to Figures 1-4 The device temperature regulation method shown includes: S1: Multi-source heterogeneous temperature data acquisition and adaptive preprocessing: The sensor deployment location is determined by thermal simulation modeling, and temperature data is collected synchronously to construct a temperature data vector; the temperature change rate and load current change rate are calculated by collecting the load current of the device, and the sampling frequency is dynamically adjusted based on the temperature change rate and load current change rate, and the collected data is cleaned and standardized. S2: Construction and training of LSTM prediction model with attention mechanism enhancement: The input window length and prediction step size are set according to the thermal response characteristics of the device, and a time series sample dataset is constructed based on the constructed temperature data vector, and divided into training set, validation set and test set according to time series; then the network structure is designed by using the attention and LSTM fusion architecture, and the model is trained and optimized, retaining the optimal model; S3: Predictive and feedback dual closed-loop adaptive control: Real-time prediction and deviation calculation are performed based on the optimal model after training and optimization; then, the outer loop PI control takes the dynamic deviation as input, and generates the basic control quantity and limits it by combining the integral separation mechanism; the inner loop fuzzy correction takes the static deviation and its rate of change as input, and outputs the correction quantity after fuzzification, inference and defuzzification, and superimposes the basic control quantity to obtain the final control quantity; based on the final control quantity, control execution and safety interlocking are performed. S4. Model Adaptive Update Based on Online Learning: Based on the data after control execution, an incremental data buffer pool is constructed, effective samples are screened and dynamically updated according to the first-in-first-out rule; a hierarchical fine-tuning strategy is adopted to adaptively update model parameters; and an update triggering and evaluation mechanism is set.

[0022] The specific analysis is as follows: S1: Multi-source heterogeneous temperature data acquisition and adaptive preprocessing S101: Intelligent Sensor Network Deployment and Synchronous Data Acquisition: Thermal simulation software is used to perform thermodynamic modeling of the equipment. Equipment structural parameters, heat source power, and heat dissipation conditions are input. A three-dimensional coordinate system encompassing the entire equipment is set up, and a three-dimensional thermal density cloud map is output after simulation. ,in , , For the internal three-dimensional coordinates of the equipment, This indicates the temperature gradient at that location, and its unit is ℃ / mm. The larger the value, the more drastic the temperature change.

[0023] Next, for the pre-set candidate deployment points inside the device, their priorities are calculated and sorted from high to low. The top D points are selected as sensor deployment locations to ensure coverage of the core area with the most significant temperature changes. All sensors are then connected to the main controller via a bus, and a built-in timer generates a period of: A synchronization pulse signal is used to ensure that all sensors sample at the same time; after each sampling, the sampled temperature is read sequentially according to the sensor number to construct a data vector. ,in Sampling time, The actual sampled temperature is represented by each element in the data vector, which corresponds to the measured temperature of a sensor. The data vector is transmitted to the edge computing node via a bus. The sending time is added to the header of the data frame, and the receiving end records the receiving time to ensure that the transmission delay is less than 5ms. Here, t is the sampling time and k is the sampling period.

[0024] S102: Adaptive Frequency Acquisition: The sampling frequency is dynamically adjusted according to the degree of temperature change to reduce redundant data while ensuring data validity. The specific steps are as follows: The initial parameter settings define three sampling frequencies: a base frequency of 2Hz with a corresponding sampling period of 0.5s, a medium frequency of 5Hz with a corresponding sampling period of 0.2s, and a high frequency of 10Hz with a corresponding sampling period of 0.1s; the base frequency is used by default in the initial state. Constructing a temperature change rate function: Determining dynamic temperature characteristics, calculating the change rate of a single sensor: For the first... Using one sensor, calculate the rate of temperature change between two consecutive samples: ,in This represents the previous sampling period, and the maximum value of the rate of change from all sensors is taken. It serves as an indicator for judging the degree of drastic change in overall temperature.

[0025] Dynamic frequency decision: If This indicates that the temperature is stable, and the sampling period is based on the fundamental frequency. ;like This indicates a moderate temperature change, using a medium-frequency sampling period. ;like This indicates a drastic temperature change, and a high-frequency sampling period is used. ; Load linkage: Real-time acquisition of equipment load current Calculate the rate of change of current. ;like This indicates a rapid increase in load, suggesting a potential surge in temperature, and necessitates a forced switch to a higher frequency sampling period. The timer is maintained for 5 seconds to avoid data loss due to lag caused by temperature changes.

[0026] S103: Data Cleaning and Standardization: Error data is filtered through CRC checksum; temperature values ​​exceeding the physically reasonable range (-40℃-150℃) are removed; outliers and missing values ​​are identified using the sliding window MAD method with a window size of 10, and values ​​with a deviation greater than 3 times the median deviation are removed; short-interval missing values ​​less than or equal to 2s are filled with linear interpolation, and long-interval missing values ​​greater than 2s are filled with historical averages to ensure temporal continuity; finally, Z-score standardization is performed on the temperature data: for each sensor d, the mean is calculated based on historical data. and standard deviation Then, the true temperature is converted into standardized data with a mean of 0 and a standard deviation of 1 using a standardization function to eliminate the influence of differences in temperature ranges from different sensors on the model. Finally, the standardized predicted values ​​output by the model are restored to the actual temperature using an inverse standardization function, which is: The inverse normalization function is: , This is the predicted temperature value.

[0027] S2: Construction and Training of LSTM Prediction Model with Enhanced Attention Mechanism By constructing an end-to-end model that integrates temporal feature encoding, key information focusing, feature decoding, and temperature prediction, high-precision prediction of the future core temperature of the equipment is achieved. The specific analysis is as follows: S201: Construction of Time Series Sample Dataset Window and step size settings: Based on the device's thermal response characteristics, the input window length is set to L=300, corresponding to 30 seconds of historical data, calculated at the highest sampling frequency of 10Hz, containing 300 time steps to ensure coverage of the complete temperature change cycle; the prediction step size is set as follows: Matching the response delay of the temperature control system; Sample generation process: First, input sequences are used to construct the sample from standardized multi-source data. In the middle, the input sequence is truncated using a sliding window method: Where N is the total time step and D is the feature dimension. The end time of the window. Each input sequence contains 30 seconds of historical features; then, labels are defined based on the end time of the window. After that, the The temperature of the core heat-generating component in the step is labeled (the target that the model needs to predict), that is: ,in To obtain the true value of the core temperature, it is necessary to reconstruct it from the standardized data through inverse standardization; Dataset partitioning: The dataset is partitioned chronologically to avoid future data being leaked into historical data: Training set: the first 70% of samples are used for model parameter learning; Validation set: the last 20% of samples are used for hyperparameter tuning; Test set: the last 10% of samples are used for final performance validation. S202: Network Structure Design: An attention and LSTM fusion architecture is adopted, which realizes temporal feature extraction, key information focusing, and future temperature prediction through a three-level structure of encoding, attention, and decoding. The design and computation logic of each layer are as follows: A1: First LSTM temporal feature encoding layer: Converts 30 seconds of multi-source time-series data (L*D) into a high-dimensional temporal feature sequence to capture the long-term dependence of temperature and load; Parameter configuration: The number of hidden units H is 64, the input dimension D is the same as the number of features, and batch input is supported; Gated computation: LSTM dynamically updates the cell state through input gates, forget gates, and output gates, where... The input sequence contains time steps, not global time. The specific analysis is as follows: Input gate: controls the proportion of new information entering the cell state; the larger the value, the more new features are retained. Its specific mathematical function is: ; Forgetting gate: controls the proportion of old cell states retained; the larger the value, the less historical information is forgotten. Its specific mathematical function is: ; Cell state update: combining forgotten old states with newly selected information to store long-term temporal features: , ; Output gate and hidden state: Extract the effective features of the current time step from the cell state and output the hidden state: , ; in Input features for the current time step. This is the previous time step concealment state. For trainable weight matrix, For bias vectors, Here, is the Sigmoid function, tanh is the hyperbolic tangent function, and ⊙ is the element-wise product; Output: The coding layer outputs a sequence of hidden states over 300 time steps. Each hidden state corresponds to a high-dimensional feature at a time step. A2: Attention Layer: Automatically assigns higher weights to time steps that are more important for prediction, improving prediction accuracy. A detailed analysis follows: Parameter configuration: The attention dimension is set to 32 to control the complexity of weight calculation; Energy score calculation: Evaluating the hidden state at each time step Relevance to prediction tasks: ,in The hidden state weight matrix of the attention layer is applied to the hidden state output of the first LSTM layer. This is used to transform the hidden states into a feature space suitable for calculating importance scores; The score vector of the attention layer, and The transformed features are multiplied by an inner product to output the energy score for a single time step. This is the bias vector for the attention layer, used to adjust... Transformed characteristic baseline; Attention weight normalization: Energy scores are converted into weights using the Softmax function. ,in For serial numbers; Context vector generation: Weighted summation of the hidden state sequence according to weights to condense key temporal features: ,in For serial numbers; Output: A single context vector, replacing the last hidden state of a traditional LSTM, retaining only the information most critical to the prediction; A3: Second LSTM feature decoding layer: Further refines the context vector output by the attention layer, mapping key temporal features into core features required for future temperature prediction; Parameter configuration: The number of hidden units is 32, which is lower than the encoding layer to avoid overfitting; the input dimension is 64, which is consistent with the dimension of the context vector. Its computational logic is as follows: It uses the same gating mechanism as the first-layer LSTM, but only takes a context vector as input instead of a time sequence, and outputs the hidden state of the last time step. ,in , These represent the initial hidden state and the cellular state of the decoding layer, respectively. A4: Output Layer Temperature Prediction: This involves converting the high-dimensional hidden states of the decoding layer into the final core temperature prediction value. The specific analysis is as follows: Fully connected hidden layers: Introducing nonlinearity through the ReLU activation function to reduce dimensionality and extract key prediction features. ,in , These are trainable parameters; Linear prediction layer: Outputs standardized predicted temperature (because the input data has been standardized): ,in , These are trainable parameters; Actual temperature restoration: Converting standardized predicted values ​​to actual temperatures (unit: °C) through inverse standardization: ,in , The historical standard deviation and mean of the core temperature; It should be further explained that in the first layer, the subscript 'i' corresponds to the input gate, controlling the proportion of new information entering the cell state; the subscript 'f' corresponds to the forget gate, controlling the proportion of old cell states retained; the subscript 'c' corresponds to the cell state, used to store long-term temporal features; the subscript 'o' corresponds to the output gate, controlling the current hidden state extracted from the cell state; and the subscript 'x' indicates that the input to the weight matrix is ​​the original feature of the current time step. The subscript 'h' indicates that the input to the weight matrix is ​​the hidden state of the previous time step. In the attention layer, 'a' is only used to identify that the parameter is related to the attention mechanism; in the output layer, 'fc' is only used to identify that the parameter is a parameter of the non-linear feature transformation sublayer in the output layer, and 'o' is used to identify that this is a parameter of the final prediction sublayer in the output layer, without any mathematical meaning. S203: Model Training and Optimization: Through loss function definition, optimizer configuration, regularization strategy, and hyperparameter tuning, the model learns patterns on the training set while ensuring generalization ability on the validation and test sets. A detailed analysis follows: Loss function design: The mean squared error (MSE) with L2 regularization is used as the loss function, which measures the deviation between the predicted and actual values ​​while preventing overfitting due to excessively large parameters. Mean squared error: reflects the accuracy of prediction; the smaller the value, the better. ,in The number of samples input for each training session. The standardized predicted value and label of the i-th sample; L2 regularization term, to prevent overfitting: penalizes excessively large model parameters. Regularization strength: ,in This refers to all trainable parameters of the model, including the weight matrix and bias vector. Optimizer configuration: The Adam optimizer is selected, which dynamically adjusts the learning rate and is suitable for training complex time series models. The parameter update steps are as follows: Calculate the gradient of the loss function with respect to the parameters: ; First-order moment estimation: ; Second-order moment estimation: ; Deviation correction: ; Gradient clipping: If the gradient norm Scale the gradient proportionally: Parameter update: The initial learning rate tc is the number of iterations. To prevent smooth terms with a denominator of 0; Regularization strategy: Dropout: A Dropout layer is added after the output of the first LSTM layer. During training, 30% of the hidden state neurons are randomly deactivated, and they are restored during testing. This avoids the model from becoming overly reliant on specific neurons. Its mathematical function is: ,in Let be a Bernoulli distributed random vector with elements of 0 or 1, and probabilities of 0.3 and 0.7 respectively. Early stopping: If the validation set is lost If the parameters do not decrease after 5 consecutive iterations, stop training immediately, save the current optimal parameters, and avoid model overfitting. Learning rate decay: The learning rate decays by 10% every 10 iterations. A larger learning rate is used in the early stages to achieve rapid convergence, and a smaller learning rate is used in the later stages to fine-tune the parameters. Hyperparameter optimization: This involves optimizing key hyperparameters of the model, with the goal of maximizing the coefficient of determination on the validation set. The closer it is to 1, the better the prediction effect. The optimized hyperparameters and search range are as follows: Hyperparameters: Input window length L, search range: [200, 400], controls the length of historical data; Hyperparameters: H is the number of hidden units in the first LSTM layer, and the search range is [32, 128], which controls the ability to extract temporal features; Hyperparameters: learning rate, search range: [1e-4, 1e-2], control parameter update step size; Hyperparameter: Dropout rate, search range: [0.2, 0.5], controls the regularization strength; Bayesian optimization is used to predict the optimal parameter combination based on historical optimization results, and the final hyperparameter configuration is determined after 50 iterations. Model Saving and Evaluation: After training, save the model parameters that minimize the loss on the validation set; evaluate performance on the test set, with key metrics including: Mean Absolute Error (MAE): ; Coefficient of determination R²: This indicates that the model can explain more than 95% of temperature change patterns.

[0028] S3: Adaptive Control with Prediction and Feedback Dual Closed-Loop By employing a dual-layer control structure of prediction and feedback loops, precise regulation of the core temperature of the equipment is achieved. Model prediction is used to anticipate temperature changes, and real-time feedback eliminates instantaneous deviations. Specifically, the system comprises four stages: real-time prediction and deviation calculation, outer-loop PI control, inner-loop fuzzy correction, and control execution and safety interlocking. A detailed analysis follows: S301: Implementation of Prediction and Deviation Calculation: Based on the attention-enhanced LSTM model trained in S2, future temperature predictions are generated in real time, and prediction deviations and real-time deviations are calculated to provide input signals for dual closed-loop control. The specific steps are as follows: Real-time input sequence construction: From the preprocessed standardized data, extract the multi-source features from the most recent 300 time steps to construct the input sequence. ,in For the current time step, D represents the feature dimension, ensuring that the input format is consistent with the S2 training samples; Future temperature forecast: Input the trained model and output the predicted core temperature 15 seconds later: The inverse normalization operation of the model output layer restores the normalized predicted values ​​to the actual temperature, with the unit being °C. Definition of deviation function: Dynamic deviation (prediction loop input): Reflects the degree of deviation between the future temperature trend and the target, used for advance control. ,in The preset target temperature for safe operation of the equipment; Static deviation (feedback loop input): reflects the instantaneous deviation of the current temperature from the target temperature, used for real-time correction. ,in This represents the actual core temperature at the current moment. S302: Outer loop PI control: The outer loop uses dynamic deviation... As input, a basic control quantity is generated through proportional and integral control to achieve early intervention in future temperature trends. At the same time, integral separation is used to avoid control saturation when there is a large deviation. The specific steps are as follows: Construction of PI control function: Proportional term: directly proportional to dynamic deviation; the larger the deviation, the more drastic the control adjustment, quickly suppressing temperature trend deviation. ,in This is the proportionality coefficient; Integral term: Accumulates historical deviations and gradually corrects long-standing small deviations, but introduces an integral separation mechanism to avoid integral saturation when there are large deviations: The integral coefficient The dynamic adjustment rule is: when , ;otherwise , This refers to the time from the start of the control system to the current control cycle. Basic control quantity: the sum of the proportional and integral terms, used as the outer loop output. ; Amplitude limiting: Integral term limit: Limits the range of accumulated integrals to prevent long-term deviations from causing the integral term to become too large. ; Output limiting: Limits the basic control quantity within the physical range of the actuator. ; S303: Inner Ring Blur Correction: The inner ring uses static deviation... Its rate of change is the input, and fuzzy control is used to... Real-time corrections are performed to address the nonlinear characteristics of the equipment's heat dissipation system. The specific analysis is as follows: Input variable fuzzification: converting precise values ​​into fuzzy language, the specific analysis of which is as follows: Static deviation : Range [-5℃, 5℃], covering common temperature fluctuation ranges, defining 5 fuzzy sets: {NB (negative large, temperature is much lower than the target or the temperature drops rapidly), NM (negative medium, temperature is lower than the target and the temperature change rate is stable), ZO (zero, close to the target and the temperature change rate is stable), PM (positive medium, temperature is higher than the target and the temperature change rate is stable), PB (positive large, temperature is much higher than the target or the temperature rises rapidly)}; Membership function: , , , , ; The rate of change of deviation reflects the speed at which the temperature deviates from the target, and is calculated using the following formula: The unit is (℃ / s), where x is The precise value; Construction of a fuzzy rule base: Its fuzzy rule format is as follows: Premise 1: The fuzzy states include: {NB, NM, ZO, PM, PB}; Premise 2: The fuzzy states include: {NB, NM, ZO, PM, PB}; therefore, a total of 25 rules are formed, covering all possible temperature states; Fuzzy Reasoning and Defuzzification: From Fuzzy Decision Making to Precise Correction Quantities; Inference strength calculation: For each rule, calculate the minimum membership degree of the input variable belonging to the corresponding fuzzy set, which serves as the trigger strength of the rule. Where k is the fuzzy rule ordinal number, , respectively , Membership degree; Centroid method for defuzzification: Set output center value for each fuzzy set: {-50, -20, 0, 20, 50}, based on engineering experience, reflecting the magnitude of the correction corresponding to the fuzzy state; The precise correction amount is obtained by summing the outputs of all triggering rules in terms of intensity weighting. ,in The number of rules to be triggered, where k is the sequence number. This is the output center value corresponding to the conclusion of the k-th rule; Final control quantity: The outer loop basic control quantity and the inner loop correction quantity are superimposed and limited to the actuator range. ; S304: Control Execution and Safety Interlock: The final control quantity... This is translated into specific actions by the actuator, while safety monitoring logic prevents temperature from exceeding limits, ensuring equipment safety. The specific operation is as follows: Executor driver function: Fan control: The relationship between fan speed and duty cycle is shown in the model below: , where N is the fan speed; Water-cooled valve control: The relationship model between flow rate and valve opening degree is as follows: Where Q is the cooling water flow rate; Security monitoring logic: Set threshold: Warning temperature: The temperature is too high; critical temperature: The critical condition has been reached; Alarm Trigger: Safety protection will be activated immediately if any of the following conditions are met: B1: Predicted Temperature This indicates that the limit may be exceeded in the future; B2: Current Temperature This indicates that the value is approaching a danger level. After triggering, force This means that the actuator is running at full load, and an audible and visual alarm is activated simultaneously. Alarm cleared: When and When necessary, turn off the alarm; S4: Adaptive Model Update Based on Online Learning During long-term operation of the equipment, the heat dissipation system will gradually deviate from the initial model's training distribution due to factors such as component aging, environmental changes, and load characteristic drift, thereby reducing the accuracy of the prediction model in S2. S4 achieves adaptive updates of the prediction model through incremental data learning and dynamic fine-tuning of model parameters, ensuring that it maintains high-precision prediction capabilities over the long term. Specifically, it consists of three core steps: incremental data buffer pool management, online model fine-tuning strategy, and update triggering and evaluation mechanism.

[0029] S401: Incremental Data Buffer Management: Continuously collects high-quality, representative new data to provide reliable material for model updates and avoids noisy data from polluting the training process.

[0030] C1: Buffer Pool Initialization and Capacity Design Create a first-in-first-out buffer pool Used to store sample pairs of input sequences and real labels. The capacity is set as follows: This ensures that enough samples are included to reflect recent temperature characteristics, while avoiding computational redundancy caused by excessive data volume.

[0031] C2: Sample Selection and Storage Rules After each model outputs a predicted value, you need to wait... To predict the step size, obtain the true core temperature at that moment. And the samples are selected according to the following rules: Validity verification: If the actual temperature exist If the physical conditions are within a reasonable range, then proceed to the next screening step; Information filtering: Calculating prediction error Only retain The sample; if Discard it to save storage space; Sample storage: Samples that meet the criteria Stored in the buffer pool in timestamp order. ,in The 300-step input sequence used for prediction. This is a real label.

[0032] C3: Buffer pool dynamic update When the number of samples in the buffer pool reaches its capacity When this happens, a FIFO update is triggered: the oldest stored sample is deleted, and a new sample is added simultaneously, i.e.: Every day during periods of low device load, old samples older than 7 days in the buffer pool are automatically cleaned up to prevent long-term accumulated aging data from affecting the model's learning of the current state.

[0033] S402: Online model fine-tuning strategy: Based on incremental data in the buffer pool, the model parameters are updated using a small-batch, low-learning-rate fine-tuning method. This retains historical knowledge while learning new patterns, avoiding the loss of the model's predictive ability for early temperature patterns after the update. D1: Fine-tuning data preparation: from the buffer pool Randomly selected from Each sample constitutes a fine-tuning batch, which is then divided into input sequences according to the S201 method. and tags The labels are then Z-score standardized and updated with the latest mean. and standard deviation To ensure consistency with the current data distribution; D2: Parameter Freezing and Training Layer Selection: A hierarchical fine-tuning strategy is adopted: the parameters of the first LSTM layer in the S2 model are frozen, and only the parameters of the attention layer and the second LSTM decoding layer are fine-tuned. These parameters are more focused on device-specific features. The learning cost of the lower-level parameters for general rules is high, and freezing can avoid the loss of historical knowledge. The upper-level parameters are more closely related to the current state of the device, and fine-tuning can quickly adapt to new features.

[0034] D3: Fine-tuning the optimizer and the loss function: Optimizer: The Adam optimizer is used, with the learning rate set to 0.1 times that of the initial training phase to ensure small and stable parameter updates and avoid drastic adjustments that could cause model oscillations. Loss function: The MSE loss with knowledge distillation is used to minimize the prediction error of new data and constrain the difference between the output of the new model and the output of the old model. The formula is as follows: in, The mean squared error of the new data is used to ensure that the model adapts to the new patterns. KL divergence measures the difference between two distributions. For the old model's predictions of the same input, a constraint weight of 0.2 ensures that the new model retains the effective knowledge of the old model; the weight allocations of 0.8 and 0.2 are determined through validation set debugging to balance learning new information with retaining old knowledge. Fine-tuning iteration control: Each fine-tuning iteration is only one. Each epoch involves a complete traversal of 64 samples three times to avoid overfitting to incremental data. During fine-tuning, the training loss is monitored in real time. If the loss increases for two consecutive epochs, the current fine-tuning is stopped immediately to prevent parameter deterioration. S403: Update Triggering and Evaluation Mechanism: Model updates must meet the requirements of necessity and effectiveness: updates are only triggered when model performance drops below a threshold, and the updated model must be superior to the old model before it is allowed to go live, ensuring system stability; Update trigger condition: Calculate the model's mean absolute error over the most recent 100 predicted samples every hour. Fine-tuning is triggered when any of the following conditions are met: Condition 1: The accuracy decreased by half compared to the initial accuracy of 0.5℃, indicating a significant reduction in prediction accuracy. Condition 2: 3 consecutive hours An upward trend indicates a continued deterioration in model performance; Post-update evaluation metrics: After fine-tuning, the performance of the new model is evaluated on an independent validation set (20% of the samples reserved from the buffer pool that were not involved in fine-tuning). Key metrics include: Mean Absolute Error : Must meet The new model has higher accuracy; Maximum error : Must meet There are no extreme prediction errors, avoiding triggering false alarms; Stability Indicators The standard deviation of the prediction error needs to be It has small error fluctuations and more stable control.

[0035] Model replacement and rollback mechanism: If the new model meets all evaluation metrics, then the old model will be replaced and used as the prediction model for S2. If the conditions are not met, discard the new model, retain the old model, record the reason for failure, and adjust the strategy the next time it is triggered. Save the last 3 model snapshots. If an anomaly occurs within 24 hours after the new model is launched, automatically roll back to the most recent stable old model.

[0036] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for regulating device temperature by incorporating a long short-term memory network, characterized in that, include: S1: Multi-source heterogeneous temperature data acquisition and adaptive preprocessing: Determine sensor deployment locations through thermal simulation modeling, and simultaneously acquire temperature data to construct temperature data vectors; By collecting the load current of the device, the rate of temperature change and the rate of load current change are calculated, and the sampling frequency is dynamically adjusted based on the rate of temperature change and the rate of load current change. The collected data is then cleaned and standardized. S2: Construction and training of LSTM prediction model with attention mechanism enhancement: The input window length and prediction step size are set according to the thermal response characteristics of the device, and a time series sample dataset is constructed based on the constructed temperature data vector, and divided into training set, validation set and test set according to time series; then the network structure is designed by using the attention and LSTM fusion architecture, and the model is trained and optimized, retaining the optimal model; S3: Prediction and feedback dual closed-loop adaptive control: Real-time prediction and deviation calculation are performed based on the optimal model after training and optimization; then, the outer loop PI control takes the dynamic deviation as input, and generates the basic control quantity and limits it by combining the integral separation mechanism; the inner loop fuzzy correction takes the static deviation and its rate of change as input, and outputs the correction quantity after fuzzification, inference and defuzzification, and superimposes the basic control quantity to obtain the final control quantity. Control execution and safety interlocking are performed based on the final control quantity; S4. Model adaptive update based on online learning: Based on the data after control execution, an incremental data buffer pool is built, effective samples are screened and dynamically updated according to the first-in-first-out rule; A hierarchical fine-tuning strategy is adopted to adaptively update model parameters; and an update triggering and evaluation mechanism is set up.

2. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: S1 includes: By analyzing the internal thermal field distribution characteristics of the equipment through thermal simulation modeling, the priority deployment positions of the sensors are determined based on the thermal density gradient. The sensors are simultaneously triggered to collect temperature data at their corresponding positions and load current data during equipment operation. A multi-source data vector with timestamps is constructed to ensure the spatial representativeness and temporal consistency of the data collection. The global temperature change rate is calculated based on the collected temperature data, and the load current change rate is calculated based on the load current data. The sampling frequency is set according to the temperature change rate, and the high-frequency sampling mechanism is triggered by the load current change rate to achieve dynamic matching between the sampling frequency and the equipment operating status. Transmission and physical range verification are performed on the collected multi-source data to filter out erroneous data. Time series analysis is used to identify and remove outliers. An appropriate imputation method is selected according to the interval of missing data to ensure the continuity of time series. Statistical features are calculated based on historical valid data. Temperature data is converted into standardized data of a uniform scale, and statistical feature parameters are retained for inverse conversion of predicted values.

3. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The construction of the time-series sample dataset includes: Based on the thermal response characteristics of the equipment, the input window length and prediction step size are set. Based on the preprocessed standardized multi-source time series data, a continuous time series is extracted as the input sequence by a sliding window method. The input sequence includes temperature data and load current data within a preset time period. The device core temperature corresponding to the prediction step size in the input window is used as the sample label to construct sample pairs of input sequence and temperature label. The generated sample pairs are divided into training set, validation set and test set in chronological order. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model performance evaluation, and the test set is used to verify the generalization prediction ability of the model.

4. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The network architecture design includes: An encoding, attention, and decoding fusion architecture is adopted, in which the encoding layer consists of a first-layer LSTM network, which is used to extract features from the input sequence in the constructed time-series sample dataset and output a hidden state sequence containing time-series correlation information; the first-layer LSTM network includes: input gate, forget gate, cell state update, output gate and hidden state; An attention mechanism layer is set between the encoding layer and the decoding layer. This attention mechanism layer takes the hidden state sequence output by the encoding layer as input, calculates the weight distribution of the hidden state at each time step, analyzes the key time step features that have a significant impact on the prediction results, and outputs a weighted fusion key temporal feature vector. The decoding layer consists of a second-layer LSTM network, which receives the key temporal feature vectors output by the attention mechanism layer and combines them with the initial hidden state to perform sequence decoding and generate high-dimensional hidden states. The network output layer adopts a fully connected structure, introduces nonlinearity through activation function, extracts key prediction features through dimensionality reduction, and then outputs standardized predicted temperature. Inverse standardization is used to convert the standardized predicted value into the actual temperature. The parameter dimension of the output layer is adapted to the prediction step size.

5. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The model training and optimization include: Based on the training set in the constructed time-series sample dataset as input, a loss function with regularization is used to quantify the deviation between the predicted value and the sample label. The network parameters are iteratively updated through a gradient descent optimizer, where the regularization term is used to suppress model overfitting. During training, the learning rate is dynamically adjusted, and a step-decay strategy is used to balance the model convergence speed and parameter optimization accuracy. The model prediction performance is monitored in real time based on the validation set. When the validation set performance does not improve for a set of consecutive rounds, an early stop mechanism is triggered to avoid overtraining. The model's generalization ability in different training rounds is evaluated using a test set. The model with the best performance on the validation set and the smallest generalization error on the test set is selected as the final optimal model. At the same time, the model structure and corresponding parameters are saved for subsequent real-time prediction.

6. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The outer loop PI control includes: The dynamic deviation calculated based on the deviation is used as input. The basic control quantity is generated through the collaboration of the proportional term and the integral term. The proportional term is proportional to the current value of the dynamic deviation, and the integral term accumulates the historical dynamic deviation. An integral separation mechanism is introduced to dynamically adjust the integral coefficient according to the absolute value of the dynamic deviation: when the dynamic deviation exceeds the preset threshold, the integral coefficient is reduced to avoid integral saturation; when the dynamic deviation is within the preset threshold, the integral coefficient is restored. Set upper and lower limits for the cumulative amount of the integral term to limit the cumulative range of historical deviations; perform amplitude limiting on the control quantity after the proportional and integral terms are superimposed to keep it within the physical adjustment range of the actuator, and output the amplitude-limited basic control quantity to provide the initial adjustment basis for inner loop fuzzy correction.

7. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The inner ring blur correction includes: Based on the static deviation and its rate of change obtained from the deviation calculation as input, the precise values ​​of the static deviation and the rate of change of deviation are converted into corresponding fuzzy sets through a preset membership function. The fuzzy sets contain linguistic variables that reflect the degree of deviation. Reasoning is performed based on a preset fuzzy rule base, which consists of multiple mapping rules between input fuzzy states and output correction actions. Valid rules are determined by calculating the trigger strength of each rule. A defuzzing algorithm is used to convert the fuzzy output of effective rules into precise correction values. The defuzzing algorithm performs weighted calculations based on the trigger strength of each rule and the corresponding output center value. The precise correction value is superimposed on the basic control value output from the outer loop, and after amplitude limiting, the final control value is obtained, thereby realizing real-time dynamic correction of the basic control value to compensate for the nonlinear characteristics of the equipment and the model prediction error.

8. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The control execution and safety interlock include: The final control quantity output based on the inner loop fuzzy correction is converted into an adjustment signal adapted to the actuator, which drives the actuator to act according to the intensity corresponding to the control quantity, thereby realizing the dynamic adjustment of the heat dissipation capacity of the equipment. The adjustment signal and the physical adjustment parameters of the actuator are in a preset correlation relationship. Multiple temperature safety thresholds are preset, including warning temperature thresholds and critical temperature thresholds. The warning temperature threshold is higher than the target temperature of the equipment, and the critical temperature threshold is higher than the warning temperature threshold. The system monitors the current core temperature of the equipment and the future temperature prediction value output by the model in real time. When any of the following conditions are met, a safety interlock mechanism is triggered: the current core temperature exceeds the critical temperature threshold, or the temperature prediction value for a preset time in the future exceeds the warning temperature threshold. Once the safety interlock mechanism is activated, the actuator operates at the maximum adjustment intensity and triggers an alarm signal. When the current core temperature of the equipment drops below the warning temperature threshold and the future temperature prediction value stabilizes within the safe range, the safety interlock is released, and the dual closed-loop adaptive control mode is restored.

9. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The incremental data buffer pool includes: A first-in-first-out buffer is constructed to store the valid samples required for model updates. The sample sources are multi-source time-series data collected after control execution and corresponding real temperature values, forming sample pairs of real-time input sequences and real temperature labels. The samples to be stored in the buffer pool are double-verified: first, sensor status verification and data physical rationality verification are performed to filter out data that are faulty or outside the normal operating range of the equipment; then, the model prediction error corresponding to the sample is calculated, and only samples whose prediction error meets the preset threshold are retained to ensure that the sample contains new feature information that the model needs to learn. Set a preset capacity for the buffer pool. When the sample storage reaches the capacity limit, delete the earliest stored sample in chronological order of timestamp and add a new qualified sample. Regularly clean up old samples in the buffer pool that have exceeded the preset storage time to maintain the consistency between the data in the buffer pool and the current operating characteristics of the equipment. The effective sample set output by the buffer pool is used for subsequent hierarchical fine-tuning of the model, providing a high-quality and timely data foundation for adaptive updates of model parameters, and avoiding the impact of noisy or outdated data on the model update effect.

10. The device temperature regulation method integrating a long short-term memory network according to claim 1, characterized in that: The adaptive update of model parameters includes: A hierarchical parameter adjustment strategy is adopted, which divides the model into frozen layers and fine-tuning layers based on the functional differences of each layer: the layer where the network coding layer learns general temporal features is set as the frozen layer, and its trained parameters are fixed to retain historical knowledge and avoid catastrophic forgetting; the layers where the attention layer, decoding layer and output layer are associated with the current operating characteristics of the device are set as fine-tuning layers, and only the parameters of this type of layer are updated. The effective sample set output from the incremental data buffer pool is used as the update data source. The fine-tuning layer is input in a mini-batch iterative manner, and the loss function with a learning rate and knowledge distillation constraint is configured to be lower than that in the initial training stage. The knowledge distillation constraint term is used to quantify the difference between the output of the new model and the output of the original optimal model, and to balance the new data adaptation ability and the old knowledge retention effect. During the iterative update process, the convergence status of the fine-tuning layer parameters and the sample prediction error are monitored in real time. When the error does not decrease or increases after a series of preset iterations, the current update is terminated. After the update is completed, the generalization error of the new model and the original model is compared on the independent validation sample set. The original model parameters are replaced only if the error of the new model is lower than that of the original model and meets the preset accuracy requirements. Otherwise, the update results are discarded and the original model parameters are kept unchanged.

Images