Nitrogen adaptive cooling control system and method based on temperature prediction

Abstract

The invention relates to the technical field of equipment control, in particular to a nitrogen self-adaptive cooling control system and method based on temperature prediction, and the system comprises an intelligent sensing module, a dynamic prediction module, a collaborative decision module and a precise execution module. The intelligent sensing module collects temperature, pressure, vibration and motion data of the nitrogen spring, and state feature vectors are generated after fusion processing. And the dynamic prediction module predicts a future temperature interval according to the historical state feature vector sequence. And the collaborative decision-making module generates a self-adaptive cooling control strategy in combination with the safety threshold. And the precise execution module executes a cooling action and collects feedback data. And the feedback data is used for updating the state feature vector and adjusting the prediction model to form closed-loop control. Through temperature prediction and self-adaptive control, the problem of control lag caused by a fixed threshold mechanism is solved, and the cooling efficiency and the equipment reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment control technology, and in particular to a nitrogen adaptive cooling control system and method based on temperature prediction. Background Technology

[0002] The existing delayed nitrogen spring cooling system technology is an integrated solution combining nitrogen spring energy storage, delayed control, and active cooling. Its working principle is based on energy conversion and thermal management logic. The nitrogen spring is filled with high-pressure nitrogen gas, which stores and releases energy during mechanical motion by utilizing the compressibility of the gas, thereby achieving a buffering effect. The delayed function adjusts the gas flow rate through a throttle valve or accumulator, introducing a controllable action delay to smooth the impact response under dynamic loads. The cooling system absorbs and dissipates the heat generated by friction and compression through fluid circulation or phase change materials, maintaining a stable operating temperature and preventing material performance degradation.

[0003] Existing delayed nitrogen spring cooling systems suffer from the following technical challenges: They rely on a fixed temperature threshold triggering mechanism, activating the cooling device only when the real-time temperature reaches a preset limit, lacking the ability to predict dynamic temperature changes during nitrogen spring operation. In high-frequency motion scenarios of automotive molds, nitrogen spring temperatures fluctuate rapidly due to load variations. The fixed threshold mechanism cannot respond to temperature rises in advance, causing cooling to lag behind actual heat accumulation. For example, in continuous stamping operations, the short-cycle reciprocating motion of the mold causes drastic temperature fluctuations in the nitrogen spring. Fixed threshold cooling may activate after the temperature briefly exceeds the limit, but by then the spring is already close to overheating, and seals are easily damaged. Simultaneously, frequent start-stop cycles increase energy consumption, and accelerated wear of mechanical components such as fans or pumps affects system lifespan and stability. This reactive control cannot adapt to load changes, leading to low energy efficiency and reduced equipment reliability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a nitrogen adaptive cooling control system and method based on temperature prediction, which solves the technical problems caused by the inability to predict temperature rise trends in existing nitrogen spring cooling systems due to their fixed threshold triggering mechanism, resulting in lag in cooling control, low energy efficiency, and accelerated equipment wear.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: In a first aspect, the present invention provides a nitrogen adaptive cooling control system based on temperature prediction, comprising: The intelligent sensing module collects temperature data, pressure data, vibration data, and motion data of the nitrogen spring, fuses the collected data, generates a state feature vector, and sends the state feature vector to the dynamic prediction module. The dynamic prediction module receives the state feature vector from the intelligent sensing module, uses the historical state feature vector sequence, calculates and outputs the future temperature prediction range of the nitrogen spring through the prediction model, and sends the temperature prediction range to the collaborative decision-making module. The collaborative decision-making module receives the temperature prediction range from the dynamic prediction module, combines it with a preset safety threshold, generates an adaptive cooling control strategy through a decision-making algorithm, and sends the adaptive cooling control strategy to the precision execution module. The precision execution module receives the adaptive cooling control strategy from the collaborative decision-making module, drives the cooling equipment to perform cooling actions on the nitrogen spring, and collects the temperature data of the cooled nitrogen spring as feedback data. The precise execution module sends feedback data back to the intelligent sensing module; the intelligent sensing module updates the state feature vector using the feedback data; the dynamic prediction module adjusts the parameters of the prediction model using the updated state feature vector to complete closed-loop control.

[0006] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the intelligent sensing module includes a distributed fiber Bragg grating temperature sensor array, a piezoresistive pressure sensor, and a MEMS inertial measurement unit. The sensor captures the surface temperature field, internal pressure pulsation waveform, and three-dimensional vibration acceleration spectrum of the nitrogen spring using a high-frequency synchronous sampling method, generating multi-physics raw data. The acquired multiphysics raw data undergoes sliding window normalization preprocessing to eliminate signal baseline drift; The preprocessed data is fused and estimated using the capacitive Kalman filter algorithm to generate a state feature vector that includes time-domain statistical features, frequency-domain energy entropy, and trajectory morphology descriptors. The generated state feature vectors are written to the shared data pool in the form of timestamp indexes for the dynamic prediction module to read.

[0007] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the dynamic prediction module slides and reads time series window data from a shared data pool. The read time series window data is normalized to eliminate differences in magnitude between features; Normalized data is input into a hybrid neural network structure including a gated recurrent unit and a multi-head self-attention mechanism; the gated recurrent unit learns the short-term local dependency pattern of the temperature sequence and outputs a short-term feature representation; the multi-head self-attention mechanism receives the short-term feature representation, calculates the weight coefficients of the influence of historical state features on future temperature, and generates long-term spatiotemporal features. The output of the hybrid neural network structure generates a probability distribution range by integrating a Monte Carlo dropout prediction model, thus quantifying prediction uncertainty. Once new actual temperature data is stored in the shared data pool, the module fine-tunes the network weights using an elastic weight consolidation algorithm to achieve adaptive tracking of the prediction model.

[0008] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the probability distribution interval output of the dynamic prediction module adopts a quantile regression model to calculate the upper and lower bounds of temperature prediction at different confidence levels. During online learning, the model dynamically adjusts the forget gate weight coefficient of the gated recurrent unit based on the residual sequence between the predicted and actual measured values. After adjusting the weight coefficients, the similarity matrix of query-key pairs in the multi-head self-attention mechanism is recalculated, so that the prediction model learns the feature representation of the current thermodynamic mode first.

[0009] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the collaborative decision-making module encodes the probabilistic temperature range, safety threshold boundary, and real-time load frequency output by the dynamic prediction module into a reinforcement learning state space. The deep reinforcement learning agent uses a proximal policy optimization algorithm to explore the action space composed of continuous fan speed and discrete cooling modes and generate candidate control policies. The digital twin simulation unit is based on the finite volume method and simulates the heat dissipation path under candidate control strategies in real time. The temperature distribution and cooling efficiency indices obtained from the simulation were transformed into a multi-objective reward function; The multi-objective reward function guides the deep reinforcement learning agent to select the Pareto optimal solution that can simultaneously optimize temperature control accuracy and energy consumption through policy gradient updates.

[0010] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the digital twin simulation unit uses an unstructured mesh to divide the geometric model of the nitrogen spring. The partitioned geometric model is used to solve the transient Navier-Stokes equations and simulate the flow and heat transfer process of the cooling medium. A regularization term is introduced when constructing the reward function to verify the consistency between the simulated hotspot temperature and the temperature prediction range of the dynamic prediction module. When the verification finds that the deviation exceeds the tolerance value, the score of the corresponding control strategy is automatically reduced.

[0011] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the precision execution module includes a permanent magnet synchronous variable frequency motor and a proportional and integral servo valve. The variable frequency motor receives the speed command signal issued by the collaborative decision-making module and drives the fan blades to rotate through space vector modulation technology; The proportional and integral servo valve adjusts the valve core opening according to the hydraulic circuit switch command, and controls the flow rate of the cooling medium. During execution, a miniature infrared thermal imager array integrated within the cooling channel acquires the two-dimensional temperature field distribution; The collected two-dimensional temperature field distribution data is used to extract spatial frequency features through fast Fourier transform to quantify cooling uniformity; The processed execution results data are fed back to the shared data pool.

[0012] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the temperature field distribution acquisition of the precision execution module adopts multispectral imaging; Multispectral imaging calculates the true surface temperature by comparing the ratios of infrared radiation intensities at different wavelengths. The energy distribution in the 0.1-10Hz frequency band was extracted using fast Fourier transform processing and used as an evaluation index for cooling uniformity. When a local hot spot is detected, the auxiliary cooling device is automatically triggered; After the auxiliary cooling device is activated, the corrected control parameters are transmitted in reverse to the strategy update loop of the collaborative decision-making module.

[0013] Furthermore, in the nitrogen adaptive cooling control system based on temperature prediction described in this invention, the data fusion strategy of the intelligent sensing module is adaptively adjusted according to the prediction residual provided by the dynamic prediction module. When the prediction residual continues to increase, the sampling frequency is automatically increased and the algorithm is switched to the capacitive Kalman filter. The dynamic prediction module receives guidance from the digital twin simulation results in the collaborative decision-making module through online learning. The physical laws revealed by the digital twin simulation results are used as regularization terms to constrain the direction of model updates; The cooling uniformity data fed back by the precision execution module provides mesh generation accuracy correction parameters for the digital twin model.

[0014] Secondly, the nitrogen adaptive cooling control method based on temperature prediction provided by the present invention is applied to the nitrogen adaptive cooling control system based on temperature prediction as described above, including: Step 1: Collect temperature data, pressure data, vibration data, and motion data of the nitrogen spring; fuse the collected data to generate a state feature vector; and use the state feature vector as the input for Step 2. Step 2: Receive the state feature vector from Step 1, use the historical state feature vector sequence, calculate and output the future temperature prediction range of the nitrogen spring through the prediction model, and use the temperature prediction range as the input of Step 3; Step 3: Receive the temperature prediction range from Step 2, combine it with a preset safety threshold, generate an adaptive cooling control strategy through a decision algorithm, and use the adaptive cooling control strategy as the input for Step 4. Step 4: Receive the adaptive cooling control strategy from Step 3, drive the cooling device to perform cooling action on the nitrogen spring, and collect the temperature data of the cooled nitrogen spring as feedback data. Step 5: Send the feedback data back to Step 1 to update the state feature vector. Use the updated state feature vector to adjust the parameters of the prediction model in Step 2 to complete the closed-loop control.

[0015] Beneficial effects of this invention; This invention collects temperature, pressure, vibration, and motion data of a nitrogen spring through an intelligent sensing module, and fuses them to generate a state feature vector. A dynamic prediction module outputs a future temperature prediction range based on historical state feature vector sequences. A collaborative decision-making module generates an adaptive cooling control strategy by combining preset safety thresholds. A precise execution module drives the cooling equipment to perform cooling actions and collects feedback data. The feedback data is used to update the state feature vector and adjust the prediction model parameters, forming a closed-loop control. This technical solution responds to temperature rise trends in advance through temperature prediction capabilities, overcoming the control lag problem of fixed threshold mechanisms. At the same time, prediction-based decision optimization avoids frequent start-stop of the cooling equipment, reducing energy waste and mechanical wear, thereby improving the overall energy efficiency and equipment reliability of the system and realizing adaptive management of nitrogen spring temperature. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a flowchart of the nitrogen adaptive cooling control method based on temperature prediction according to the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] In a first aspect, the present invention provides a nitrogen adaptive cooling control system based on temperature prediction, comprising: The intelligent sensing module collects temperature data, pressure data, vibration data, and motion data of the nitrogen spring, fuses the collected data, generates a state feature vector, and sends the state feature vector to the dynamic prediction module. The dynamic prediction module receives the state feature vector from the intelligent sensing module, uses the historical state feature vector sequence, calculates and outputs the future temperature prediction range of the nitrogen spring through the prediction model, and sends the temperature prediction range to the collaborative decision-making module. The collaborative decision-making module receives the temperature prediction range from the dynamic prediction module, combines it with a preset safety threshold, generates an adaptive cooling control strategy through a decision-making algorithm, and sends the adaptive cooling control strategy to the precision execution module. The precision execution module receives the adaptive cooling control strategy from the collaborative decision-making module, drives the cooling equipment to perform cooling actions on the nitrogen spring, and collects the temperature data of the cooled nitrogen spring as feedback data. The precise execution module sends feedback data back to the intelligent sensing module; the intelligent sensing module updates the state feature vector using the feedback data; the dynamic prediction module adjusts the parameters of the prediction model using the updated state feature vector to complete closed-loop control.

[0020] The intelligent sensing module captures the surface temperature field, internal pressure pulsation waveform, and three-dimensional vibration acceleration spectrum of the nitrogen spring using a distributed fiber Bragg grating temperature sensor array, a piezoresistive pressure sensor, and a MEMS inertial measurement unit, generating multi-physics raw data through high-frequency synchronous sampling. This raw multi-physics data undergoes sliding window normalization preprocessing to eliminate signal baseline drift and improve data quality. The preprocessed data is then fused and estimated using a capacitive Kalman filter algorithm to generate a state feature vector that includes time-domain statistical features, frequency-domain energy entropy, and trajectory morphology descriptors. This state feature vector is written to a shared data pool with a timestamp index for the dynamic prediction module to access. This process achieves synchronous acquisition, noise reduction, and feature extraction of multi-source data, providing high-precision input for temperature prediction.

[0021] The dynamic prediction module slides through time-series window data from a shared data pool. This data is normalized to eliminate differences in magnitude between features. The normalized data is then input into a hybrid neural network structure including a gated recurrent unit and a multi-head self-attention mechanism. The gated recurrent unit learns short-term local dependency patterns in the temperature sequence and outputs short-term feature representations. The multi-head self-attention mechanism receives these short-term feature representations, calculates the weighting coefficients of historical state features on the impact of future temperatures, and generates long-term spatiotemporal features. The output of the hybrid neural network structure generates a probability distribution interval by integrating a Monte Carlo dropout prediction model, quantifying prediction uncertainty. When new actual temperature data is stored in the shared data pool, the module fine-tunes the network weights using an elastic weight consolidation algorithm, enabling the prediction model to adaptively track the slow drift of nitrogen spring performance. This step utilizes historical data sequences and machine learning models to achieve advanced temperature change prediction and continuous model optimization.

[0022] The collaborative decision-making module encodes the probabilistic temperature range, safety threshold boundary, and real-time load frequency output by the dynamic prediction module into a reinforcement learning state space. The deep reinforcement learning agent employs a proximal policy optimization algorithm to explore the action space composed of continuous fan speed and discrete cooling modes, generating candidate control strategies. The digital twin simulation unit, based on the finite volume method, simulates the heat dissipation path under the candidate control strategies in real time. The simulated temperature distribution and cooling efficiency indices are transformed into a multi-objective reward function. This multi-objective reward function, through policy gradient updates, guides the deep reinforcement learning agent to select a Pareto optimal solution that simultaneously optimizes temperature control accuracy and energy consumption. This process, through simulation verification and a reward mechanism, ensures that the control strategy meets cooling requirements while improving energy efficiency.

[0023] The precision execution module includes a permanent magnet synchronous variable frequency motor and a proportional and integral servo valve. The variable frequency motor receives speed command signals from the collaborative decision-making module and drives the fan blades to rotate using space vector modulation technology. The proportional and integral servo valve adjusts the valve core opening according to hydraulic circuit switching commands to control the cooling medium flow rate. During execution, a miniature infrared thermal imager array integrated in the cooling channel collects a two-dimensional temperature field distribution. The collected two-dimensional temperature field distribution data is processed using fast Fourier transform to extract spatial frequency features and quantify cooling uniformity. The processed execution effect data is fed back to the shared data pool. This step translates the control strategy into physical actions and monitors the cooling effect in real time, providing data support for closed-loop control.

[0024] The precision execution module sends feedback data back to the intelligent sensing module; the intelligent sensing module updates the state feature vector using the feedback data; and the dynamic prediction module adjusts the parameters of the prediction model using the updated state feature vector. This closed-loop control mechanism enables the system to dynamically adjust data acquisition and prediction models based on the actual cooling effect, achieving adaptive management of the nitrogen spring temperature and fundamentally solving the lag problem caused by fixed threshold control. The modules are tightly coupled through a shared data pool and data flow, improving the overall system response speed and reliability.

[0025] The intelligent sensing module of this invention captures the surface temperature field, internal pressure pulsation waveform, and three-dimensional vibration acceleration spectrum of a nitrogen spring using a distributed fiber optic temperature sensor array, a piezoresistive pressure sensor, and a MEMS inertial measurement unit in a high-frequency synchronous sampling manner, generating multi-physics raw data. This raw data undergoes sliding window normalization preprocessing to eliminate signal baseline drift. The preprocessed data is then fused and estimated using a capacitive Kalman filter algorithm to generate a state feature vector that includes time-domain statistical features, frequency-domain energy entropy, and trajectory morphology descriptors. This state feature vector is written to a shared data pool in the form of a timestamp index for the dynamic prediction module to read, thereby achieving synchronous acquisition, noise reduction, and feature extraction of multi-source data, providing high-precision input for temperature prediction.

[0026] The dynamic prediction module of this invention slides to read time series window data from a shared data pool. The read time series window data is normalized to eliminate differences in magnitude between features. The normalized data is input into a hybrid neural network structure including a gated recurrent unit and a multi-head self-attention mechanism. The gated recurrent unit learns the short-term local dependency pattern of the temperature sequence and outputs a short-term feature representation. The multi-head self-attention mechanism receives the short-term feature representation, calculates the weight coefficients of the influence of historical state features on future temperatures, and generates long-term spatiotemporal features. The output of the hybrid neural network structure generates a probability distribution interval by integrating a Monte Carlo dropout prediction model, quantifying prediction uncertainty. When new actual temperature data is stored in the shared data pool, the module fine-tunes the network weights through an elastic weight consolidation algorithm to achieve adaptive tracking of the prediction model, thereby using historical data sequences to achieve advanced prediction of temperature changes and continuous model optimization.

[0027] The probability distribution interval output of the dynamic prediction module of this invention adopts a quantile regression model to calculate the upper and lower bounds of temperature prediction at different confidence levels. During online learning, the model dynamically adjusts the forget gate weight coefficient of the gated recurrent unit based on the residual sequence between the predicted value and the actual measured value. After the weight coefficient is adjusted, the similarity matrix of query-key value pairs in the multi-head self-attention mechanism is recalculated, so that the prediction model learns the feature representation of the current thermodynamic mode first, thereby improving the prediction accuracy and model adaptability.

[0028] The collaborative decision-making module of this invention encodes the probabilistic temperature range, safety threshold boundary, and real-time load frequency output by the dynamic prediction module into a reinforcement learning state space. The deep reinforcement learning agent uses a proximal policy optimization algorithm to explore the action space composed of continuous fan speed and discrete cooling modes, generating candidate control strategies. The digital twin simulation unit, based on the finite volume method, simulates the heat dissipation path under the candidate control strategies in real time. The simulated temperature distribution and cooling efficiency indicators are transformed into a multi-objective reward function. The multi-objective reward function, through policy gradient updates, guides the deep reinforcement learning agent to select the Pareto optimal solution that can simultaneously optimize temperature control accuracy and energy consumption, thereby achieving control strategy optimization through simulation verification and reward mechanisms.

[0029] The digital twin simulation unit of this invention uses an unstructured mesh to divide the geometric model of the nitrogen spring. The divided geometric model is used to solve the transient Navier-Stokes equations to simulate the flow and heat transfer process of the cooling medium. A regularization term is introduced when constructing the reward function to verify the consistency between the hot spot temperature predicted by the simulation and the temperature prediction range of the dynamic prediction module. When the verification finds that the deviation exceeds the tolerance value, the score of the corresponding control strategy is automatically reduced, thereby achieving consistency between the control strategy and the actual thermodynamic behavior.

[0030] The precision execution module of this invention includes a permanent magnet synchronous variable frequency motor and a proportional and integral servo valve. The variable frequency motor receives speed command signals from the collaborative decision-making module and drives the fan blades to rotate through space vector modulation technology. The proportional and integral servo valve adjusts the valve core opening according to the hydraulic circuit switching command to control the cooling medium flow rate. During execution, a miniature infrared thermal imager array integrated in the cooling channel collects two-dimensional temperature field distribution data. The collected two-dimensional temperature field distribution data is processed by fast Fourier transform to extract spatial frequency features and quantify cooling uniformity. The processed execution effect data is fed back to the shared data pool, thereby transforming the control strategy into physical actions and monitoring the cooling effect in real time.

[0031] The temperature field distribution acquisition of the precision execution module of this invention adopts multispectral imaging; multispectral imaging calculates the true surface temperature by the ratio of infrared radiation intensity at different wavelengths; the energy distribution in the 0.1-10Hz frequency band is extracted by fast Fourier transform processing as an evaluation index for cooling uniformity; when a local hot spot is detected, an auxiliary cooling device is automatically triggered; after the auxiliary cooling device is activated, the corrected control parameters are transmitted in reverse to the strategy update loop of the collaborative decision-making module, thereby realizing rapid response to local hot spots and dynamic adjustment of control parameters.

[0032] The data fusion strategy of the intelligent sensing module of this invention adaptively adjusts based on the prediction residuals provided by the dynamic prediction module; when the prediction residuals continue to increase, the sampling frequency is automatically increased and switched to the capacitive Kalman filter algorithm; the online learning of the dynamic prediction module receives guidance from the digital twin simulation results in the collaborative decision-making module; the physical laws revealed by the digital twin simulation results serve as regularization terms to constrain the model update direction; the cooling uniformity data fed back by the precision execution module provides mesh partitioning accuracy correction parameters for the digital twin model, thereby achieving the system's self-evolution and advanced thermal management through nonlinear coupling between modules.

[0033] Secondly, please refer to Figure 1 The nitrogen adaptive cooling control method based on temperature prediction provided by this invention is applied to the nitrogen adaptive cooling control system based on temperature prediction as described above, including: Step 1: Collect temperature data, pressure data, vibration data, and motion data of the nitrogen spring; fuse the collected data to generate a state feature vector; and use the state feature vector as the input for Step 2. Step 2: Receive the state feature vector from Step 1, use the historical state feature vector sequence, calculate and output the future temperature prediction range of the nitrogen spring through the prediction model, and use the temperature prediction range as the input of Step 3; Step 3: Receive the temperature prediction range from Step 2, combine it with a preset safety threshold, generate an adaptive cooling control strategy through a decision algorithm, and use the adaptive cooling control strategy as the input for Step 4. Step 4: Receive the adaptive cooling control strategy from Step 3, drive the cooling device to perform cooling action on the nitrogen spring, and collect the temperature data of the cooled nitrogen spring as feedback data. Step 5: Send the feedback data back to Step 1 to update the state feature vector. Use the updated state feature vector to adjust the parameters of the prediction model in Step 2 to complete the closed-loop control.

[0034] This invention addresses the technical problems of lag in cooling control, low energy efficiency, and accelerated equipment wear caused by the fixed threshold triggering mechanism in existing nitrogen spring cooling systems by constructing an adaptive cooling control system based on temperature prediction. The system employs an intelligent sensing module to collect temperature, pressure, vibration, and motion data from the nitrogen spring, and fuses this multi-source data to generate a state feature vector. This state feature vector is sent to a dynamic prediction module, which uses a historical state feature vector sequence to calculate and output the predicted temperature range for the nitrogen spring through a prediction model. This predictive capability allows the system to anticipate temperature rise trends rather than passively responding to fixed thresholds, thus overcoming the lag in cooling action.

[0035] The temperature prediction range output by the dynamic prediction module is sent to the collaborative decision-making module. The collaborative decision-making module, combined with a preset safety threshold, generates an adaptive cooling control strategy through a decision-making algorithm. During the decision-making process, a deep reinforcement learning agent explores the optimal strategy in the action space, while a digital twin simulation unit simulates the heat dissipation path in real time, transforming the simulation results into a multi-objective reward function to optimize the control strategy. This prediction-based decision-making mechanism avoids frequent start-ups and shutdowns of the cooling equipment, reduces energy waste and mechanical wear, and improves energy efficiency and equipment reliability.

[0036] The adaptive cooling control strategy generated by the collaborative decision-making module is sent to the precision execution module. The precision execution module drives the cooling equipment to perform cooling actions on the nitrogen spring and collects the cooled temperature data as feedback data. This feedback data is sent back to the intelligent sensing module to update the state feature vector; the dynamic prediction module uses the updated state feature vector to adjust the parameters of the prediction model. This closed-loop control process enables the system to continuously optimize data acquisition and prediction models based on the actual cooling effect, achieving adaptive management of the nitrogen spring temperature. Through data flow and feedback loops between modules, the system dynamically adapts to load changes, fundamentally eliminating the inherent defects of fixed threshold control and achieving proactive and precise thermal management.

[0037] A shared data pool, implemented using distributed databases or memory-mapped files, stores state feature vectors and feedback data with timestamp indices, supporting high-concurrency read and write access. A sliding window reading mechanism extracts historical data by setting a fixed-length time-series window and sliding it in steps, ensuring data continuity and real-time performance. In the intelligent sensing module, a capacitive Kalman filter algorithm fuses multi-sensor data using a state-space model to estimate the system state vector and generate a state feature vector containing both temporal and frequency domain features. In the dynamic prediction module, a gated recurrent unit learns short-term dependencies of the time series through input, forget, and output gate structures, while a multi-head self-attention mechanism calculates the importance weights of features at different time steps to capture long-term spatiotemporal correlations. An integrated Monte Carlo dropout technique generates a probability distribution in the prediction model through multiple forward propagations and random neuron dropouts, quantifying prediction uncertainty. An elastic weight consolidation algorithm adjusts the network weight update magnitude based on the importance of new data during online learning, avoiding catastrophic forgetting. The digital twin simulation unit uses an unstructured mesh to generate the geometric model of the nitrogen spring. Based on the discrete transient Navier-Stokes equations using the finite volume method, it simulates the flow and heat transfer process of the cooling medium. A regularization term is used to incorporate the consistency check between the simulation results and the predicted interval into the reward function. In the precision execution module, multispectral imaging eliminates the influence of surface emissivity by calibrating the radiation intensity ratios of different infrared bands, calculating the true temperature field. Fast Fourier transform analysis of spatial frequency characteristics extracts the energy distribution in the 0.1-10Hz frequency band as an indicator of cooling uniformity. The adaptive adjustment mechanism between modules uses a prediction residual monitoring data fusion strategy. When the residual exceeds a threshold, the sampling frequency and filtering algorithm are dynamically switched. Simultaneously, the digital twin simulation results are used as physical constraints to guide model updates, ensuring system self-evolution and thermal management accuracy.

[0038] In high-frequency motion scenarios of automotive molds, nitrogen springs experience rapid temperature rises and falls due to load fluctuations. Existing cooling systems, which use fixed temperature threshold triggering mechanisms, cannot respond to temperature rise trends in advance. This invention presents a nitrogen adaptive cooling control system based on temperature prediction, achieving advanced thermal management through modular design. During system operation, the intelligent sensing module first collects temperature, pressure, vibration, and motion data of the nitrogen springs. The intelligent sensing module uses a distributed fiber optic temperature sensor array to capture the surface temperature field distribution, a piezoresistive pressure sensor to monitor internal pressure pulsation waveforms, and a MEMS inertial measurement unit to record the three-dimensional vibration acceleration spectrum. These sensors generate multi-physics raw data using a high-frequency synchronous sampling method. The raw data undergoes sliding window normalization preprocessing to eliminate signal baseline drift, and then a capacitive Kalman filter algorithm is used for fusion estimation to generate a state feature vector containing time-domain statistical features, frequency-domain energy entropy, and trajectory morphology descriptors. The state feature vector is written to a shared data pool in the form of a timestamp index.

[0039] The dynamic prediction module slides through a shared data pool to read historical state feature vector sequences. Time-series window data is normalized to eliminate differences in the magnitude of features before being input into a hybrid neural network structure. This structure consists of gated recurrent units and a multi-head self-attention mechanism. The gated recurrent units learn short-term local dependency patterns in the temperature sequence and output short-term feature representations. The multi-head self-attention mechanism calculates the weight coefficients of the influence of historical state features on future temperatures to generate long-term spatiotemporal features. The network output generates a probability distribution interval by integrating a Monte Carlo discard prediction model, quantifying prediction uncertainty. When new actual temperature data is stored in the shared data pool, the dynamic prediction module fine-tunes the network weights using an elastic weight consolidation algorithm to achieve adaptive tracking of the prediction model. The probability distribution interval output uses a quantile regression model to calculate the upper and lower bounds of temperature predictions at different confidence levels. During online learning, the model dynamically adjusts the forget gate weight coefficients of the gated recurrent units based on the residual sequence between predicted and actual measurements, and recalculates the similarity matrix of query-key pairs in the multi-head self-attention mechanism, enabling the prediction model to prioritize learning the feature representations of the current thermodynamic pattern.

[0040] The collaborative decision-making module receives the probabilistic temperature range output by the dynamic prediction module and encodes it into a reinforcement learning state space by combining a preset safety threshold and real-time load frequency. The deep reinforcement learning agent employs a proximal policy optimization algorithm to explore the action space composed of continuous fan speed and discrete cooling modes, generating candidate control strategies. The digital twin simulation unit simulates the heat dissipation path under the candidate control strategies in real time using the finite volume method. The digital twin simulation unit uses an unstructured mesh to partition the nitrogen spring geometry model and solves the transient Navier-Stokes equations to simulate the cooling medium flow and heat transfer process. The simulated temperature distribution and cooling efficiency indices are transformed into a multi-objective reward function. A regularization term is introduced during the construction of the reward function to verify the consistency between the simulated hotspot temperature and the temperature prediction range of the dynamic prediction module. When the verification finds a deviation exceeding the tolerance value, the score of the corresponding control strategy is automatically reduced. The multi-objective reward function guides the deep reinforcement learning agent to select the Pareto optimal solution that simultaneously optimizes temperature control accuracy and energy consumption through policy gradient updates.

[0041] The precise execution module receives the adaptive cooling control strategy generated by the collaborative decision-making module. The precise execution module includes a permanent magnet synchronous variable frequency motor and a proportional-integral servo valve. The permanent magnet synchronous variable frequency motor drives the fan blades to rotate using space vector modulation technology, and the proportional-integral servo valve adjusts the valve core opening to control the cooling medium flow rate. During execution, a miniature infrared thermal imager array integrated within the cooling channel acquires a two-dimensional temperature field distribution. This temperature field distribution acquisition uses multispectral imaging technology, calculating the true surface temperature through the ratio of infrared radiation intensities at different wavelengths. The acquired two-dimensional temperature field distribution data is processed using fast Fourier transform to extract spatial frequency features, specifically extracting the energy distribution in the 0.1 Hz to 10 Hz frequency band as an evaluation index for cooling uniformity. When a local hot spot is detected, an auxiliary cooling device is automatically triggered. After the auxiliary cooling device is activated, the corrected control parameters are transmitted back to the strategy update loop of the collaborative decision-making module. The processed execution effect data is fed back to the shared data pool.

[0042] The precise execution module sends the cooled nitrogen spring temperature data back to the intelligent sensing module as feedback data. The intelligent sensing module uses the feedback data to update the state feature vector; the dynamic prediction module uses the updated state feature vector to adjust the parameters of the prediction model, completing closed-loop control. The data fusion strategy of the intelligent sensing module is adaptively adjusted based on the prediction residuals provided by the dynamic prediction module; when the prediction residuals continue to increase, the sampling frequency is automatically increased and the capacitive Kalman filter algorithm is switched. The dynamic prediction module receives guidance from the digital twin simulation results in the collaborative decision-making module for online learning; the physical laws revealed by the digital twin simulation results serve as regularization terms to constrain the model update direction. The cooling uniformity data fed back by the precise execution module provides mesh generation accuracy correction parameters for the digital twin model. Through nonlinear coupling between modules, the system achieves self-evolution and advanced thermal management, fundamentally solving the control lag problem of the fixed threshold mechanism.

[0043] Embodiment 1 of this invention: Application of the system on an automotive mold stamping line. The intelligent sensing module employs a distributed fiber optic temperature sensor array arranged on the surface of a nitrogen spring, a piezoresistive pressure sensor integrated inside the spring, and a MEMS inertial measurement unit mounted on the spring base. It captures the temperature field distribution, pressure pulsation waveform, and three-dimensional vibration spectrum using high-frequency synchronous sampling. The raw multiphysics data undergoes sliding window normalization preprocessing to eliminate baseline drift. Subsequently, a capacitive Kalman filter algorithm is used for data fusion to generate a state feature vector containing time-domain statistical features and frequency-domain energy entropy. The state feature vector is written to a shared data pool in the form of a timestamp index. The dynamic prediction module slides and reads historical state feature vector sequences from the shared data pool. After normalization, the data is input into a hybrid neural network structure. A gated recurrent unit learns short-term temperature dependence patterns, and a multi-head self-attention mechanism calculates historical feature weights to generate long-term spatiotemporal features. The network output generates a probability distribution interval through integrated Monte Carlo dropout technology, quantifying prediction uncertainty. When new temperature data is stored in the shared data pool, the module fine-tunes the network weights using an elastic weight consolidation algorithm to achieve adaptive model tracking. The collaborative decision-making module encodes the probabilistic temperature range, safety threshold, and real-time load frequency into a reinforcement learning state space. A deep reinforcement learning agent employs a proximal policy optimization algorithm to explore candidate policies within the fan speed and cooling mode action space. A digital twin simulation unit solves the transient Navier-Stokes equations using the finite volume method, simulating heat dissipation paths and transforming the temperature distribution into a multi-objective reward function. A regularization term is introduced into the reward function to verify the consistency between the simulated hotspot temperature and the predicted range; when the deviation exceeds the limit, the policy score is automatically reduced. The precise execution module drives the fan rotation via a permanent magnet synchronous variable frequency motor, and proportional and integral servo valves regulate the cooling medium flow rate. During execution, a miniature infrared thermal imager array acquires the two-dimensional temperature field distribution, uses multispectral imaging technology to calculate the true surface temperature, and extracts spatial frequency features through fast Fourier transform to evaluate cooling uniformity. When a local hotspot is detected, an auxiliary cooling device is triggered, and the corrected parameters are fed back to the collaborative decision-making module. The feedback data is used to update the state feature vector and adjust the prediction model parameters, forming a closed-loop control and resolving the control lag problem.

[0044] Embodiment 2 of this invention: Optimized implementation of the system in a high-speed stamping scenario. The intelligent sensing module adaptively adjusts the data fusion strategy based on the prediction residuals provided by the dynamic prediction module; when the residuals continue to increase, the sampling frequency is automatically increased and the system switches to the capacitive Kalman filter algorithm to enhance data acquisition accuracy. The online learning process of the dynamic prediction module receives guidance from the digital twin simulation results in the collaborative decision-making module; the physical laws revealed by the simulation results serve as regularization terms to constrain the model update direction, enabling the prediction model to prioritize learning the current thermodynamic mode. The cooling uniformity data fed back by the precise execution module provides mesh partitioning accuracy correction parameters for the digital twin model, optimizing simulation accuracy. The collaborative decision-making module explores Pareto optimal solutions in the action space through a deep reinforcement learning agent, balancing temperature control accuracy and energy consumption; the digital twin simulation unit uses an unstructured mesh to partition the nitrogen spring geometric model, improving the realism of heat transfer simulation. During execution, multispectral imaging eliminates the influence of emissivity by using infrared radiation intensity ratios, and fast Fourier transform analyzes the energy distribution in the 0.1 Hz to 10 Hz frequency band to quantify cooling uniformity. After local hotspots trigger auxiliary cooling devices, control parameters are transmitted in reverse to the strategy update loop for dynamic adjustment. Through nonlinear coupling between modules, the system achieves self-evolution and advanced thermal management, improving equipment reliability and energy efficiency.

Claims

1. A nitrogen adaptive cooling control system based on temperature prediction, characterized in that, include: The intelligent sensing module collects temperature data, pressure data, vibration data, and motion data of the nitrogen spring, fuses the collected data, generates a state feature vector, and sends the state feature vector to the dynamic prediction module. The dynamic prediction module receives the state feature vector from the intelligent sensing module, uses the historical state feature vector sequence, calculates and outputs the future temperature prediction range of the nitrogen spring through the prediction model, and sends the temperature prediction range to the collaborative decision-making module. The collaborative decision-making module receives the temperature prediction range from the dynamic prediction module, combines it with a preset safety threshold, generates an adaptive cooling control strategy through a decision-making algorithm, and sends the adaptive cooling control strategy to the precision execution module. The precision execution module receives the adaptive cooling control strategy from the collaborative decision-making module, drives the cooling equipment to perform cooling actions on the nitrogen spring, and collects the temperature data of the cooled nitrogen spring as feedback data. The precise execution module sends feedback data back to the intelligent sensing module; the intelligent sensing module updates the state feature vector using the feedback data; the dynamic prediction module adjusts the parameters of the prediction model using the updated state feature vector to complete closed-loop control.

2. The nitrogen adaptive cooling control system based on temperature prediction according to claim 1, characterized in that, The intelligent sensing module includes a distributed fiber optic temperature sensor array, a piezoresistive pressure sensor, and a MEMS inertial measurement unit. The sensor captures the surface temperature field, internal pressure pulsation waveform, and three-dimensional vibration acceleration spectrum of the nitrogen spring using a high-frequency synchronous sampling method, generating multi-physics raw data. The acquired multiphysics raw data undergoes sliding window normalization preprocessing to eliminate signal baseline drift; The preprocessed data is fused and estimated using the capacitive Kalman filter algorithm to generate a state feature vector that includes time-domain statistical features, frequency-domain energy entropy, and trajectory morphology descriptors. The generated state feature vectors are written to the shared data pool in the form of timestamp indexes for the dynamic prediction module to read.

3. The nitrogen adaptive cooling control system based on temperature prediction according to claim 2, characterized in that, The dynamic prediction module reads time series window data from a shared data pool. The read time series window data is normalized to eliminate differences in magnitude between features; Normalized data is input into a hybrid neural network structure including a gated recurrent unit and a multi-head self-attention mechanism; the gated recurrent unit learns the short-term local dependency pattern of the temperature sequence and outputs a short-term feature representation; the multi-head self-attention mechanism receives the short-term feature representation, calculates the weight coefficients of the influence of historical state features on future temperature, and generates long-term spatiotemporal features. The output of the hybrid neural network structure generates a probability distribution range by integrating a Monte Carlo dropout prediction model, thus quantifying prediction uncertainty. Once new actual temperature data is stored in the shared data pool, the module fine-tunes the network weights using an elastic weight consolidation algorithm to achieve adaptive tracking of the prediction model.

4. The nitrogen adaptive cooling control system based on temperature prediction according to claim 3, characterized in that, The probability distribution interval output of the dynamic prediction module uses a quantile regression model to calculate the upper and lower bounds of temperature prediction at different confidence levels. During online learning, the model dynamically adjusts the forget gate weight coefficient of the gated recurrent unit based on the residual sequence between the predicted and actual measured values. After adjusting the weight coefficients, the similarity matrix of query-key pairs in the multi-head self-attention mechanism is recalculated, so that the prediction model learns the feature representation of the current thermodynamic mode first.

5. The nitrogen adaptive cooling control system based on temperature prediction according to claim 1, characterized in that, The collaborative decision-making module encodes the probabilistic temperature range, safety threshold boundary, and real-time load frequency output by the dynamic prediction module into a reinforcement learning state space. The deep reinforcement learning agent uses a proximal policy optimization algorithm to explore the action space composed of continuous fan speed and discrete cooling modes and generate candidate control policies. The digital twin simulation unit is based on the finite volume method and simulates the heat dissipation path under candidate control strategies in real time. The temperature distribution and cooling efficiency indices obtained from the simulation were transformed into a multi-objective reward function; The multi-objective reward function guides the deep reinforcement learning agent to select the Pareto optimal solution that can simultaneously optimize temperature control accuracy and energy consumption through policy gradient updates.

6. The nitrogen adaptive cooling control system based on temperature prediction according to claim 5, characterized in that, The digital twin simulation unit uses an unstructured mesh to divide the geometric model of the nitrogen spring; The partitioned geometric model is used to solve the transient Navier-Stokes equations and simulate the flow and heat transfer process of the cooling medium. A regularization term is introduced when constructing the reward function to verify the consistency between the simulated hotspot temperature and the temperature prediction range of the dynamic prediction module. When the verification finds that the deviation exceeds the tolerance value, the score of the corresponding control strategy is automatically reduced.

7. The nitrogen adaptive cooling control system based on temperature prediction according to claim 1, characterized in that, The precision execution module includes a permanent magnet synchronous variable frequency motor and a proportional and integral servo valve; The variable frequency motor receives the speed command signal issued by the collaborative decision-making module and drives the fan blades to rotate through space vector modulation technology; The proportional and integral servo valve adjusts the valve core opening according to the hydraulic circuit switch command, and controls the flow rate of the cooling medium. During execution, a miniature infrared thermal imager array integrated within the cooling channel acquires the two-dimensional temperature field distribution; The collected two-dimensional temperature field distribution data is used to extract spatial frequency features through fast Fourier transform to quantify cooling uniformity; The processed execution results data are fed back to the shared data pool.

8. The nitrogen adaptive cooling control system based on temperature prediction according to claim 7, characterized in that, The temperature field distribution of the precision execution module is acquired using multispectral imaging. Multispectral imaging calculates the true surface temperature by comparing the ratios of infrared radiation intensities at different wavelengths. The energy distribution in the 0.1-10Hz frequency band was extracted using fast Fourier transform processing and used as an evaluation index for cooling uniformity. When a local hot spot is detected, the auxiliary cooling device is automatically triggered; After the auxiliary cooling device is activated, the corrected control parameters are transmitted in reverse to the strategy update loop of the collaborative decision-making module.

9. The nitrogen adaptive cooling control system based on temperature prediction according to claim 1, characterized in that, The data fusion strategy of the intelligent sensing module is adaptively adjusted based on the prediction residuals provided by the dynamic prediction module. When the prediction residual continues to increase, the sampling frequency is automatically increased and the algorithm is switched to the capacitive Kalman filter. The dynamic prediction module receives guidance from the digital twin simulation results in the collaborative decision-making module through online learning. The physical laws revealed by the digital twin simulation results are used as regularization terms to constrain the direction of model updates; The cooling uniformity data fed back by the precision execution module provides mesh generation accuracy correction parameters for the digital twin model.

10. A nitrogen adaptive cooling control method based on temperature prediction, applied to the nitrogen adaptive cooling control system based on temperature prediction as described in any one of claims 1 to 9, characterized in that, include: Step 1: Collect temperature data, pressure data, vibration data, and motion data of the nitrogen spring; fuse the collected data to generate a state feature vector; and use the state feature vector as the input for Step 2. Step 2: Receive the state feature vector from Step 1, use the historical state feature vector sequence, calculate and output the future temperature prediction range of the nitrogen spring through the prediction model, and use the temperature prediction range as the input of Step 3; Step 3: Receive the temperature prediction range from Step 2, combine it with a preset safety threshold, generate an adaptive cooling control strategy through a decision algorithm, and use the adaptive cooling control strategy as the input for Step 4. Step 4: Receive the adaptive cooling control strategy from Step 3, drive the cooling device to perform cooling action on the nitrogen spring, and collect the temperature data of the cooled nitrogen spring as feedback data. Step 5: Send the feedback data back to Step 1 to update the state feature vector. Use the updated state feature vector to adjust the parameters of the prediction model in Step 2 to complete the closed-loop control.

Images