A method and system for controlling the conductivity of a pure water cooling system

Abstract

The present invention discloses a method and system for controlling the conductivity of a pure water cooling system. The method includes constructing a conductivity data set; performing adversarial training processing on the conductivity data set to obtain a virtual conductivity sequence; obtaining a conductivity fusion data set; inputting the processed conductivity fusion data set into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain to obtain a conductivity state path and a non-stationary time transition matrix, and the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model; inputting the conductivity state path and the non-stationary time transition matrix into a cooling system state evaluation model to obtain the state of the cooling medium; determining a conductivity control signal for the target pure water cooling system based on the state of the cooling medium, and executing the conductivity control signal. The method provided by the embodiments of the present invention can control the conductivity situation in the pure water cooling system.

Description

Technical Field

[0001] The present invention relates to the technical field of conductivity control, and in particular to a method and system for controlling the conductivity of a pure water cooling system. Background Art

[0002] During the operation of a fixed ice melting converter, the silicon stack composed of high-power semiconductor devices and a water-cooled radiator, which is its core component, relies on a closed pure water cooling system for heat dissipation. This system ensures the circulating flow of the cooling medium under constant pressure and flow rate through the coordinated work of the internal cooling unit and the external cooling unit. However, the working frequency of the fixed ice melting converter is relatively low and the downtime is relatively long. As a result, during the downtime, the main circulation cooling module in the closed pure water cooling system is in a stopped state. At this time, the ion concentration in the cooling medium will gradually increase, which may potentially threaten the stable operation of the converter.

[0003] Therefore, it has become an urgent technical problem for those skilled in the art to ensure that the conductivity of the cooling medium in the cooling system always remains within a reasonable range during the downtime of the converter. Summary of the Invention

[0004] The present invention provides a method and system for controlling the conductivity of a pure water cooling system, which controls the conductivity of the pure water cooling system to ensure that the conductivity of the cooling medium in the cooling system always remains within a reasonable range during the downtime of the converter.

[0005] To solve the above technical problems, an embodiment of the present invention provides a method for controlling the conductivity of a pure water cooling system, and the method includes:

[0006] Constructing a conductivity data set based on the temperature, downtime, and historical conductivity time series data obtained in the target pure water cooling system in the shutdown state;

[0007] Under the constraint conditions of the diffusion-thermodynamic coupling equation, performing adversarial training on the conductivity data set based on the metal area of the inner wall of the obtained target pure water cooling system to obtain a virtual conductivity sequence with the same dimension as the conductivity data set;

[0008] Performing fusion processing on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set;

[0009] Inputting the processed conductivity fusion data set into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain, and using the Viterbi algorithm to solve the cooling system state acquisition model to obtain a conductivity state path and a non-stationary time transition matrix, wherein the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model.

[0010] Construct a cooling system state evaluation model using a dynamic attention residual network that includes at least a dynamic attention layer and a residual jump connection layer, and input the conductivity state path and the non-steady state time transfer matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system;

[0011] Determine the conductivity control signal of the target pure water cooling system based on the state of the cooling medium, and control the functional components corresponding to the pure water cooling system to execute the conductivity control signal.

[0012] As one of the preferred solutions, the process of performing adversarial training on the conductivity data set includes:

[0013] Based on the obtained metal area of the inner wall of the target pure water cooling system, determine the additional ion data in the target pure water cooling system;

[0014] Use a generator constructed based on the Wasserstein generative adversarial network model to perform adversarial training on the conductivity data set and the additional ion data to obtain the virtual conductivity data set;

[0015] Obtain the minimized Wasserstein distance between the virtual conductivity sequence and the conductivity data set through a diffusion-thermodynamic coupling equation, where the diffusion-thermodynamic coupling equation includes at least thermodynamic constraints, diffusion coefficient constraints, and Fick's law constraints;

[0016] Use the minimized Wasserstein distance and the conductivity data set to extract features from the virtual conductivity sequence to obtain a virtual conductivity sequence with the same dimension as the conductivity data set.

[0017] As one of the preferred solutions, before fusing the conductivity data set and the virtual conductivity sequence, the conductivity control method for the pure water cooling system further includes:

[0018] Use a discriminator constructed based on the Wasserstein generative adversarial network model to process the virtual conductivity sequence and the conductivity data set to obtain the maximized Wasserstein distance between the virtual conductivity sequence and the conductivity data set;

[0019] Construct a loss function based on the minimized Wasserstein distance and the maximized Wasserstein distance, and use the loss function to iteratively process the virtual conductivity sequence to obtain the optimal virtual conductivity sequence.

[0020] As one of the preferred solutions, the process of fusing the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set includes:

[0021] Feature extraction is performed on the conductivity data set and the virtual conductivity sequence respectively to obtain a first conductivity feature and a second conductivity feature;

[0022] Weight evaluation is performed on the first conductivity feature and the second conductivity feature, and the conductivity fusion data set is obtained based on the weight evaluation result.

[0023] As one of the preferred solutions, the processed conductivity fusion data set is input into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain. The processing process of the conductivity fusion data set sequentially includes:

[0024] Verification processing and divergence processing;

[0025] Diffusion-thermodynamic equation verification processing is performed on the conductivity fusion data set to obtain a residual result. If the residual result is greater than a first threshold, data cleaning processing is performed;

[0026] Divergence processing is performed on the cleaned conductivity fusion data set to obtain a distribution similarity result between the conductivity fusion data set and the conductivity data set, and the conductivity fusion data set is optimized based on the distribution similarity result.

[0027] As one of the preferred solutions, the use of the Viterbi algorithm to solve the cooling system state acquisition model includes:

[0028] Online estimation is performed on the cooling system state acquisition model based on the Expectation-Maximization (EM) algorithm to obtain a conductivity time-varying transition matrix;

[0029] Probability calculation is performed on the conductivity fusion data set to obtain a conductivity time-varying transition probability;

[0030] Decoding is performed on the conductivity time-varying transition matrix and the conductivity time-varying transition probability based on the Viterbi algorithm to obtain a conductivity state path and a non-stationary time transition matrix.

[0031] As one of the preferred solutions, the dynamic attention residual network at least includes a serial self-attention module, a deep residual module, a hierarchical residual module, and a depthwise separable convolutional network;

[0032] The self-attention module includes a serial dynamic attention layer, a batch normalization layer, and a global linear unit activation layer; the deep residual module includes a serial batch normalization layer, a residual skip connection layer, and a weighted batch normalization layer.

[0033] As one of the preferred solutions, the input of the conductivity state path and the non-stationary time transition matrix into the cooling system state evaluation model includes:

[0034] Perform an embedding mapping process on the conductivity state path, and perform a sine position encoding process on the result of the embedding mapping process to obtain a state embedding sequence;

[0035] Perform a tensor expansion process on the non-steady-state time transfer matrix and perform tensor reconstruction to obtain a non-steady-state time series with the same dimension as the state embedding sequence;

[0036] Input the state embedding sequence and the non-steady-state time series into the cooling system state evaluation model for cross-attention feature fusion processing to obtain the state of the cooling medium in the cooling system.

[0037] As one preferred solution, the state of the cooling medium includes conductivity prediction data and conductivity transfer probability;

[0038] Determining the conductivity control signal of the target pure water cooling system based on the state of the cooling medium includes:

[0039] Determine the conductivity control signal based on the historical conductivity time series data, the conductivity prediction data, and the conductivity transfer probability, where the conductivity control signal includes at least one of starting a standby circulation device, starting an ion filtration device, and starting an automatic dosing device.

[0040] Another embodiment of the present invention provides a conductivity control system for a pure water cooling system, the system includes:

[0041] A construction module for constructing a conductivity data set based on the temperature, shutdown duration, and historical conductivity time series data obtained in the target pure water cooling system in the shutdown state;

[0042] A training module for performing adversarial training on the conductivity data set based on the metal area of the inner wall of the target pure water cooling system obtained under the constraint conditions of the diffusion-thermodynamic coupling equation to obtain a virtual conductivity sequence with the same dimension as the conductivity data set;

[0043] A fusion module for fusing the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set;

[0044] A solution module for inputting the processed conductivity fusion data set into a cooling system state acquisition model constructed by a non-steady-state hidden Markov chain, and using the Viterbi algorithm to solve the cooling system state acquisition model to obtain a conductivity state path and a non-steady-state time transfer matrix, where the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model;

[0045] An evaluation module for constructing a cooling system state evaluation model by using a dynamic attention residual network including at least a dynamic attention layer and a residual jump connection layer, inputting the conductivity state path and the non-steady state time transition matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system;

[0046] A control module for determining a conductivity control signal of the target pure water cooling system based on the state of the cooling medium, and controlling the functional components corresponding to the pure water cooling system to execute the conductivity control signal.

[0047] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0048] Construct a conductivity data set based on the temperature, shutdown duration, and historical conductivity time series data obtained in the target pure water cooling system in the shutdown state; under the constraint condition of the diffusion-thermodynamic coupling equation, perform adversarial training processing on the conductivity data set based on the metal area of the inner wall of the obtained target pure water cooling system to obtain a virtual conductivity sequence with the same dimension as the conductivity data set; perform fusion processing on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set; input the processed conductivity fusion data set into a cooling system state acquisition model constructed by a non-steady state hidden Markov chain, and use the Viterbi algorithm to solve the cooling system state acquisition model to obtain a conductivity state path and a non-steady state time transition matrix, where the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model; construct a cooling system state evaluation model by using a dynamic attention residual network including at least a dynamic attention layer and a residual jump connection layer, input the conductivity state path and the non-steady state time transition matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system; determine the conductivity control signal of the target pure water cooling system based on the state of the cooling medium, and control the functional components corresponding to the pure water cooling system to execute the conductivity control signal. Compared with the prior art, this method processes the conductivity of the cooling medium to control the conductivity of the pure water cooling system, and the conductivity of the cooling medium is always maintained within a reasonable range, thereby ensuring the stable operation of the converter. Description of the Drawings

[0049] Figure 1 is a schematic flow chart of a method for controlling the conductivity of a pure water cooling system in one embodiment of the present invention;

[0050] Figure 2 is a schematic structural diagram of a conductivity control system of a pure water cooling system in one embodiment of the present invention.

[0051] Reference numerals:

[0052] Among them, 11, building block; 12, training module; 13, fusion module; 14, solving module; 15, evaluation module; 16, control module. Detailed implementation manners

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] In the description of the present application, the terms "first", "second", "third", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined with "first", "second", "third", etc. may explicitly or implicitly include one or more of such features. In the description of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0055] In the description of the present application, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected", "connected" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected, or indirectly connected through an intermediate medium, and it may be the communication inside two elements. The terms "vertical", "horizontal", "left", "right", "up", "down" and similar expressions used herein are only for the purpose of illustration, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation of the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to specific circumstances.

[0056] In the description of the present application, it should be noted that unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used in the description of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to specific circumstances.

[0057] During the operation of a fixed ice - melting converter, its core component - the silicon stack composed of high - power semiconductor devices and a water - cooled radiator - relies on a closed - type pure - water cooling system for heat dissipation. This system ensures the circulating flow of the cooling medium under constant pressure and flow rate through the coordinated operation of the internal cooling unit and the external cooling unit, removing the heat generated by high - power power - electronic devices. However, the operating frequency of the fixed ice - melting converter is relatively low and the downtime is relatively long. As a result, during the downtime, the main circulation cooling module in the closed - type pure - water cooling system is in a stopped state. At this time, the ion concentration in the cooling medium will gradually increase, which may potentially threaten the stable operation of the converter.

[0058] Therefore, an embodiment of the present invention provides a method for controlling the conductivity of a pure - water cooling system. Specifically, please refer to Figure 1 , Figure 1 which shows a schematic flow chart of the method for controlling the conductivity of a pure - water cooling system in one embodiment of the present invention. The method includes:

[0059] S1: Construct a conductivity data set based on the temperature, downtime duration, and historical conductivity time - series data obtained in the target pure - water cooling system in the shutdown state.

[0060] Specifically, according to the preset temperature sensor, obtain the real - time temperature data in the target pure - water cooling system. In the shutdown state, it is impossible to obtain the conductivity data of the cooling medium in real - time, and only the historical conductivity time - series data in the startup state of the pure - water cooling system can be obtained. Construct a conductivity data set based on the real - time temperature data and the historical conductivity time - series data.

[0061] S2: Under the constraint conditions of the diffusion - thermodynamics coupling equation, perform adversarial training processing on the conductivity data set based on the metal area of the inner wall of the obtained target pure - water cooling system to obtain a virtual conductivity sequence with the same dimension as the conductivity data set.

[0062] In the startup state of the pure - water cooling system, the cooling water circulates in the system, exerting a continuous scouring effect on the metal inner wall. This scouring effect helps to reduce the accumulation of deposits and corrosion products on the metal surface, thereby reducing the chance of chemical reactions between the metal surface and ions in the water. Furthermore, the electric ions generated on the metal surface can be ignored.

[0063] However, in the stationary state, the cooling water no longer circulates, increasing the rate of the electrochemical reaction between the metal surface and ions in the water, thereby generating additional ion data. Therefore, in the stationary state, it is necessary to consider the metal area of the inner wall of the pure - water cooling system, determine the additional ion data based on the metal area under the constraint of the corrosion kinetics model, and thus be able to control the conductivity within a certain range.

[0064] It should be noted that, in addition to the corrosion kinetics model constraints, it also includes geometric and time constraints, thermodynamic equilibrium constraints, etc.

[0065] The Wasserstein Generative Adversarial Network (WGAN) is an improved generative adversarial network that can solve problems such as training instability and mode collapse existing in traditional GANs. Therefore, by using the Wasserstein Generative Adversarial Network to train and process the conductivity dataset, a virtual conductivity sequence with high robustness and extremely close to the conductivity dataset can be obtained.

[0066] The process of obtaining the virtual conductivity sequence is as follows:

[0067] Use the generator constructed based on the Wasserstein Generative Adversarial Network model to perform adversarial training on the conductivity dataset and additional ion data to obtain a virtual conductivity dataset; through the diffusion-thermodynamic coupling equation, obtain the minimized Wasserstein distance between the virtual conductivity dataset and the conductivity dataset, where the diffusion-thermodynamic coupling equation at least includes thermodynamic constraints, diffusion coefficient constraints, and Fick's law constraints; use the minimized Wasserstein distance and the conductivity dataset to extract features from the virtual conductivity dataset to obtain a virtual conductivity sequence with the same dimension as the conductivity dataset.

[0068] This process generates a virtual conductivity sequence through adversarial training based on the Wasserstein Generative Adversarial Network, and its core goal is to solve the coupling problem of the scarcity of real data and physical law constraints.

[0069] Specifically, the generator is constructed by deeply integrating the statistical characteristics of the conductivity dataset and the dynamic correlation of additional ion data. The additional ion data is embedded in the generator as a conditional variable to ensure the dynamic relevance of the virtual data to the metal corrosion process.

[0070] The diffusion-thermodynamic coupling equation at least includes thermodynamic constraints, diffusion coefficient constraints, and Fick's law constraints. Among them, the thermodynamic constraint quantifies the influence of temperature on ion migration as an explicit function, and the diffusion coefficient constraint enforces the virtual sequence to satisfy diffusion equilibrium through Fick's second law; the boundary constraint limits the ion concentration gradient on the metal surface to zero to simulate the ion accumulation effect in a static environment.

[0071] Through the diffusion-thermodynamic coupling equation, obtain the minimized Wasserstein distance between the virtual conductivity dataset and the conductivity dataset, and achieve the statistical consistency between the virtual data and the real data through the alignment of the minimized Wasserstein distance and the feature space.

[0072] At the same time, use principal component analysis or dynamic time warping to perform dimension compression and phase alignment on the virtual sequence to ensure its complete isomorphism with the real dataset in the time domain and spatial domain.

[0073] S3: Perform a fusion process on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set.

[0074] It should be noted that before performing the fusion process on the conductivity data set and the virtual conductivity sequence, it is also necessary to optimize the obtained virtual conductivity sequence. During the optimization process, the Wasserstein generative adversarial network model is also used to process the data. The process is as follows:

[0075] Use a discriminator based on the Wasserstein generative adversarial network model to process the virtual conductivity sequence and the conductivity data set to obtain the maximum Wasserstein distance between the virtual conductivity sequence and the conductivity data set. Construct a loss function based on the minimum Wasserstein distance and the maximum Wasserstein distance, and use the loss function to perform iterative processing on the virtual conductivity sequence to obtain an optimal virtual conductivity sequence.

[0076] Specifically, the discriminator adopts a multi-scale time series feature extraction structure, captures local fluctuations and global trends through convolutional kernels, and measures the distribution difference between real and virtual data based on the Wasserstein distance. This adversarial training mechanism not only avoids the mode collapse problem of traditional GANs but also can generate reasonable virtual samples in low-data-density regions, such as extreme temperatures or long downtime.

[0077] The core task of the discriminator is to maximize the Wasserstein distance between the virtual conductivity sequence and the real conductivity data set, thereby quantifying the distribution difference between the two and guiding the generator to optimize.

[0078] This process not only ensures the dimensional consistency of the virtual data but also makes the finally generated virtual conductivity sequence have both physical rationality and statistical authenticity, which can effectively support subsequent non-stationary hidden Markov chain modeling and real-time control strategy optimization.

[0079] Perform a fusion process on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set. The specific process includes: performing feature extraction on the conductivity data set and the virtual conductivity sequence respectively to obtain a first conductivity feature and a second conductivity feature; performing weight evaluation on the first conductivity feature and the second conductivity feature, and obtaining the conductivity fusion data set based on the weight evaluation result.

[0080] Specifically, the first conductivity feature is extracted from real data, which includes three types of key information: statistical features, frequency-domain features, and physical constraint features. Among them, statistical features include mean, variance, skewness, etc., reflecting the macroscopic distribution characteristics of the conductivity time series; frequency-domain features include extracting low-frequency trends and high-frequency noise components through Fourier transform or wavelet transform, representing the fluctuation patterns at different time scales; physical constraint features include diffusion equations and temperature correlations, used to quantify the degree of compliance of the data with the laws of thermodynamics. The second conductivity feature is designed for virtual sequences. In addition to the above basic features, a generation confidence index is additionally introduced, such as discriminator score, Wasserstein distance, and physical constraint deviation such as the theoretical error between metal ion concentration and conductivity, to evaluate the reliability of virtual data.

[0081] An entropy weight method-attention mechanism hybrid strategy is used to dynamically allocate the fusion weights of the two types of features. During the weight allocation process, significant scenario dependence is presented: in the high-density region of real data, such as common temperature ranges, the entropy weight biases towards real data; while in the low-density or extreme working condition regions, such as extremely long downtime, the weight of virtual data increases, and it will only be adopted when the physical constraint deviation is lower than the threshold. Therefore, the fused dataset not only retains the statistical authenticity of real data but also fills the distribution blind spots through virtual data, reducing the generalization error of the conductivity prediction model in extreme scenarios.

[0082] It should also be noted that feature-level fusion can avoid the problem of dimensional explosion caused by splicing of original signals, compress the data scale, and thus improve the subsequent modeling efficiency.

[0083] S4: Input the processed conductivity fused dataset into the cooling system state acquisition model constructed by a non-stationary hidden Markov chain, and use the Viterbi algorithm to solve the cooling system state acquisition model to obtain the conductivity state path and the non-stationary time transition matrix. Among them, the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model.

[0084] Specifically, the processing process of the conductivity fused dataset sequentially includes verification processing and divergence processing;

[0085] Perform diffusion-thermodynamic equation verification processing on the conductivity fused dataset to obtain a residual result. If the residual result is greater than the first threshold, perform data cleaning processing. The verification processing is to verify whether the fused data conforms to the diffusion-thermodynamic coupling equation and eliminate abnormal data that violates physical laws. Different processing strategies are adopted for different abnormal data. For isolated abnormal points, they are directly deleted, and the gaps are filled by linear interpolation or the mean of adjacent data; for continuous abnormal segments, such as the entire segment offset caused by sensor failure, alternative data is regenerated based on the numerical solution of the diffusion equation.

[0086] Perform divergence processing on the cleaned conductivity fusion dataset to obtain the distribution similarity result between the conductivity fusion dataset and the conductivity dataset, and optimize the conductivity fusion dataset based on the distribution similarity result. Divergence processing is to use the hybrid divergence index to evaluate the similarity between the fusion dataset and the original dataset, avoiding the introduction of distribution shift due to cleaning or correction:

[0087] Cleaning and divergence optimization form a closed-loop process. After each round of optimization, the residuals and divergence are recalculated. The finally processed data needs to satisfy both physical constraints and statistical constraints. This processing process combines hard physical verification and flexible statistical optimization to ensure that the fused data has both thermodynamic rationality and distribution authenticity, thereby improving the accuracy of conductivity state path prediction under the shutdown state.

[0088] Considering that the conductivity change under the shutdown state is still affected by time and the state transition probability is not fixed, while the traditional Markov chain assumes that the state transition probability is fixed, the non-stationary hidden Markov chain allows the transition matrix to change with time and can better capture this dynamic change. Therefore, a non-stationary hidden Markov chain is used to construct a cooling system state acquisition model.

[0089] However, at the same time, the parameters of the non-stationary HMM need to satisfy both data-driven statistical characteristics and physical law constraints. Traditional maximum likelihood estimation or EM algorithm only relies on data statistics and will generate parameters that violate physical laws. The physics-informed neural network (PINN) adjusts the parameters of the cooling system state acquisition model, embeds the diffusion-thermodynamic coupling equation as a hard constraint into the loss function, so that the state path prediction conforms to both data statistical laws and thermodynamics and diffusion kinetics.

[0090] The steps of using the Viterbi algorithm to solve the cooling system state acquisition model include:

[0091] Based on the expectation-maximization (EM) algorithm, perform online estimation on the cooling system state acquisition model to obtain the time-varying conductivity transition matrix; perform probability calculation on the conductivity fusion dataset to obtain the time-varying conductivity transition probability; based on the Viterbi algorithm, decode the time-varying conductivity transition matrix and the time-varying conductivity transition probability to obtain the conductivity state path and the non-stationary time transition matrix.

[0092] Specifically, through the online estimation of the expectation-maximization (EM) algorithm, the conductivity state transition matrix is updated in real time to directly capture the influence of dynamic variables such as temperature and metal corrosion rate on the conductivity evolution during the shutdown process; the Viterbi algorithm traces back the optimal state path step by step based on the time-varying transition matrix and the observation probability, enabling this method to quickly respond to the state changes of the cooling system.

[0093] S5: Construct a cooling system state evaluation model using a dynamic attention residual network that at least includes a dynamic attention layer and a residual jump connection layer. Input the conductivity state path and the non-steady state time transfer matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system.

[0094] Construct a cooling system state evaluation model using a dynamic attention residual network. Specifically, the dynamic attention residual network at least includes a self-attention module, a deep residual module, a hierarchical residual module, and a depthwise separable convolutional network in series. It should be emphasized that the self-attention module includes a dynamic attention layer, a batch normalization layer, and a global linear unit activation layer in series, which together implement a pipeline process of "focusing - normalizing - selecting" to improve the feature quality and reduce the risk of overfitting. The deep residual module includes a batch normalization layer, a residual jump connection layer, and a weighted batch normalization layer in series. The jump connection ensures direct gradient passing, and the weighted BN realizes differential enhancement of feature channels. The combination of the two rapidly improves the convergence speed of deep network training. The hierarchical residual module enables the model to take into account both local mutations and long-term evolution, reducing the root mean square error of state evaluation. The depthwise separable convolution improves the computational efficiency of the model while ensuring the feature extraction ability.

[0095] Input the conductivity state path and the non-steady state time transfer matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system, including:

[0096] Perform embedding mapping processing on the conductivity state path, and perform sine position encoding processing on the result of the embedding mapping processing to obtain a state embedding sequence. Perform tensor expansion processing on the non-steady state time transfer matrix and perform tensor reconstruction to obtain a non-steady state time sequence with the same dimension as the state embedding sequence. Input the state embedding sequence and the non-steady state time sequence into the cooling system state evaluation model for cross-attention feature fusion processing to obtain the state of the cooling medium in the cooling system.

[0097] The embedding mapping processing of the conductivity state path aims to transform a discrete and symbolic state sequence (such as "low - medium - high" conductivity labels) into a continuous low-dimensional vector space to capture the semantic associations and dynamic evolution laws between states. Superimposing sine position encoding on the embedding mapping result, its core role is to inject absolute and relative time information into the sequence. Since the evolution of the conductivity state path has significant time dependence, such as the shutdown duration affecting the corrosion rate, the position encoding of the sine function marks the position of each state on the time axis through a combination of positive and cosine waves with different frequencies. This encoding method enables the model to perceive temporal relationships without relying on a recursive structure, which is especially suitable for capturing long-range dependencies and avoiding the gradient vanishing problem of traditional recurrent neural networks.

[0098] Rather than the tensor expansion and reconstruction of the non - steady - state time - transfer matrix, the core objective is to preserve the dynamic characteristics of the transfer probability and convert it into temporal features adapted to the attention mechanism. The state embedding encodes the semantics of the conductivity itself, while the time series reflects the dynamic law of state transitions. The two complement each other to accurately decode the real - time state of the cooling medium.

[0099] The cross - attention feature fusion of the state - embedding sequence and the non - steady - state time series. This fusion strategy breaks through the limitations of single - modality features: the state embedding encodes the semantics of the conductivity itself, and the time series reflects the dynamic law of state transitions. The two complement each other to accurately decode the real - time state of the cooling medium, and further provide a quantitative basis for the conductivity control strategy.

[0100] S6: Determine the conductivity control signal of the target pure - water cooling system based on the state of the cooling medium, and control the functional components corresponding to the pure - water cooling system to execute the conductivity control signal.

[0101] Specifically, determine the conductivity control signal based on the historical conductivity time - series data, the conductivity prediction data, and the conductivity transfer probability. Among them, the conductivity control signal includes at least one of starting the standby circulation device, starting the ion filtration device, and starting the automatic dosing device.

[0102] Another embodiment of the present invention provides a conductivity control system for a pure - water cooling system. Specifically, please refer to Figure 2 , Figure 2 which is shown as the structural schematic diagram of the conductivity control system of the pure - water cooling system in one embodiment of the present invention. The system includes:

[0103] A construction module 11, configured to construct a conductivity data set based on the temperature, shutdown duration, and historical conductivity time - series data obtained in the target pure - water cooling system in the shutdown state;

[0104] A training module 12, configured to perform adversarial training on the conductivity data set based on the metal area on the inner wall of the obtained target pure - water cooling system under the constraint conditions of the diffusion - thermodynamics coupling equation, and obtain a virtual conductivity sequence with the same dimension as the conductivity data set;

[0105] A fusion module 13, configured to perform fusion processing on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set;

[0106] The solution module 14 is configured to input the processed conductivity fusion data set into the cooling system state acquisition model constructed by the non-steady-state hidden Markov chain, and use the Viterbi algorithm to solve the cooling system state acquisition model to obtain the conductivity state path and the non-steady-state time transition matrix. Wherein, the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model;

[0107] The evaluation module 15 is configured to construct a cooling system state evaluation model by using a dynamic attention residual network including at least a dynamic attention layer and a residual jump connection layer, and input the conductivity state path and the non-steady-state time transition matrix into the cooling system state evaluation model to obtain the cooling medium state in the cooling system;

[0108] The control module 16 is configured to determine the conductivity control signal of the target pure water cooling system based on the cooling medium state, and control the functional components corresponding to the pure water cooling system to execute the conductivity control signal.

[0109] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0110] Construct a conductivity data set based on the temperature, shutdown duration, and historical conductivity time series data obtained in the target pure water cooling system in the shutdown state; under the constraint conditions of the diffusion-thermodynamic coupling equation, perform adversarial training processing on the conductivity data set based on the metal area of the inner wall of the obtained target pure water cooling system to obtain a virtual conductivity sequence with the same dimension as the conductivity data set; perform fusion processing on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set; input the processed conductivity fusion data set into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain, and use the Viterbi algorithm to solve the cooling system state acquisition model to obtain a conductivity state path and a non-stationary time transition matrix, wherein the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model; construct a cooling system state evaluation model using a dynamic attention residual network including at least a dynamic attention layer and a residual jump connection layer, input the conductivity state path and the non-stationary time transition matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system; determine the conductivity control signal of the target pure water cooling system based on the state of the cooling medium, and control the functional components corresponding to the pure water cooling system to execute the conductivity control signal. Compared with the prior art, the present method processes the conductivity of the cooling medium to control the conductivity of the pure water cooling system so that the conductivity of the cooling medium is always maintained within a reasonable range, thereby ensuring the stable operation of the converter.

[0111] The above embodiments merely represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the invention patent shall be subject to the appended claims.

Claims

1. A method for controlling the conductivity of a pure water cooling system, characterized in that, Including: Constructing a conductivity dataset based on the temperature, shutdown duration, and historical conductivity time-series data obtained in the target pure water cooling system in a shutdown state; Under the constraint conditions of the diffusion-thermodynamic coupling equation, performing adversarial training on the conductivity dataset based on the metal area of the inner wall of the obtained target pure water cooling system to obtain a virtual conductivity sequence with the same dimension as the conductivity dataset; Performing fusion processing on the conductivity dataset and the virtual conductivity sequence to obtain a conductivity fusion dataset; Inputting the processed conductivity fusion dataset into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain, and using the Viterbi algorithm to solve the cooling system state acquisition model to obtain a conductivity state path and a non-stationary time transition matrix, where the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model; Constructing a cooling system state evaluation model using a dynamic attention residual network including at least a dynamic attention layer and a residual jump connection layer, and inputting the conductivity state path and the non-stationary time transition matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system; Determining a conductivity control signal for the target pure water cooling system based on the state of the cooling medium, and controlling a functional component corresponding to the pure water cooling system to execute the conductivity control signal.

2. The conductivity control method of the pure water cooling system according to claim 1, characterized in that, The process of performing adversarial training on the conductivity dataset includes: Determining additional ion data in the target pure water cooling system based on the area of the metal region of the inner wall of the obtained target pure water cooling system; Performing adversarial training on the conductivity dataset and the additional ion data using a generator constructed based on a Wasserstein generative adversarial network model to obtain a virtual conductivity dataset; Obtaining the minimum Wasserstein distance between the virtual conductivity dataset and the conductivity dataset through the diffusion-thermodynamic coupling equation, where the diffusion-thermodynamic coupling equation includes at least thermodynamic constraints, diffusion coefficient constraints, and Fick's law constraints; Using the minimum Wasserstein distance and the conductivity dataset to extract features from the virtual conductivity dataset to obtain a virtual conductivity sequence with the same dimension as the conductivity dataset.

3. The conductivity control method of the pure water cooling system according to claim 2, characterized in that Before performing fusion processing on the conductivity dataset and the virtual conductivity sequence, the conductivity control method for the pure water cooling system further includes: Processing the virtual conductivity sequence and the conductivity dataset using a discriminator constructed based on a Wasserstein generative adversarial network model to obtain the maximum Wasserstein distance between the virtual conductivity sequence and the conductivity dataset; Constructing a loss function based on the minimum Wasserstein distance and the maximum Wasserstein distance, and using the loss function to iteratively process the virtual conductivity sequence to obtain an optimal virtual conductivity sequence.

4. The conductivity control method of the pure water cooling system according to claim 1, characterized in that The process of performing fusion processing on the conductivity dataset and the virtual conductivity sequence to obtain a conductivity fusion dataset includes: Feature extraction is respectively performed on the conductivity data set and the virtual conductivity sequence to obtain a first conductivity feature and a second conductivity feature; Weight evaluation is performed on the first conductivity feature and the second conductivity feature, and a conductivity fusion data set is obtained based on the weight evaluation result.

5. The conductivity control method of the pure water cooling system according to claim 1, characterized in that, The processed conductivity fusion data set is input into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain. The processing process of the conductivity fusion data set sequentially includes: Verification processing and divergence processing; Diffusion-thermodynamic equation verification processing is performed on the conductivity fusion data set to obtain a residual result. If the residual result is greater than a first threshold, data cleaning processing is performed; Divergence processing is performed on the cleaned conductivity fusion data set to obtain a distribution similarity result between the conductivity fusion data set and the conductivity data set, and the conductivity fusion data set is optimized based on the distribution similarity result.

6. The conductivity control method of the pure water cooling system according to claim 1, characterized in that Solving the cooling system state acquisition model by using the Viterbi algorithm includes: Online estimation is performed on the cooling system state acquisition model based on the expectation maximization (EM) algorithm to obtain a conductivity time-varying transition matrix; Probability calculation is performed on the conductivity fusion data set to obtain a conductivity time-varying transition probability; Decoding is performed on the conductivity time-varying transition matrix and the conductivity time-varying transition probability based on the Viterbi algorithm to obtain a conductivity state path and a non-stationary time transition matrix.

7. The conductivity control method of the pure water cooling system according to claim 1, characterized in that The dynamic attention residual network at least includes a serial self-attention module, a deep residual module, a hierarchical residual module, and a depthwise separable convolutional network; The self-attention module includes a serial dynamic attention layer, a batch normalization layer, and a global linear unit activation layer; the deep residual module includes a serial batch normalization layer, a residual skip connection layer, and a weighted batch normalization layer.

8. The conductivity control method of the pure water cooling system according to claim 1, characterized in that, Inputting the conductivity state path and the non-stationary time transition matrix into the cooling system state evaluation model to obtain the state of the cooling medium in the cooling system includes: Performing embedding mapping processing on the conductivity state path, and performing sine position encoding processing on the result of the embedding mapping processing to obtain a state embedding sequence; Performing tensor unfolding processing on the non-stationary time transition matrix and performing tensor reconstruction to obtain a non-stationary time sequence with the same dimension as the state embedding sequence; Inputting the state embedding sequence and the non-stationary time sequence into the cooling system state evaluation model for cross-attention feature fusion processing to obtain the state of the cooling medium in the cooling system.

9. The conductivity control method of the pure water cooling system according to claim 1, characterized in that, The state of the cooling medium includes conductivity prediction data and conductivity transition probability; Determining the conductivity control signal of the target pure water cooling system based on the state of the cooling medium includes: Determining the conductivity control signal based on the historical conductivity time series data, the conductivity prediction data, and the conductivity transition probability, where the conductivity control signal at least includes one of starting a standby circulation device, starting an ion filtration device, and starting an automatic dosing device.

10. A conductivity control system for a pure water cooling system, characterized in that, Including: A construction module for constructing a conductivity data set based on temperature, shutdown duration, and historical conductivity time series data obtained in a target pure water cooling system in a shutdown state; A training module for performing adversarial training processing on the conductivity data set based on the metal area of the inner wall of the obtained target pure water cooling system under the constraint conditions of the diffusion-thermodynamic coupling equation to obtain a virtual conductivity sequence with the same dimension as the conductivity data set; A fusion module for performing fusion processing on the conductivity data set and the virtual conductivity sequence to obtain a conductivity fusion data set; A solving module for inputting the processed conductivity fusion data set into a cooling system state acquisition model constructed by a non-stationary hidden Markov chain, and using the Viterbi algorithm to solve the cooling system state acquisition model to obtain a conductivity state path and a non-stationary time transition matrix, wherein the construction process includes adjusting the parameters of the cooling system state acquisition model through a physics-informed neural network to optimize the cooling system state acquisition model; An evaluation module for constructing a cooling system state evaluation model using a dynamic attention residual network including at least a dynamic attention layer and a residual jump connection layer, and inputting the conductivity state path and the non-stationary time transition matrix into the cooling system state evaluation model to obtain the cooling medium state in the cooling system; A control module for determining a conductivity control signal of the target pure water cooling system based on the cooling medium state, and controlling the functional components corresponding to the pure water cooling system to execute the conductivity control signal.

Images