Offshore flexible DC converter valve IGBT power module heat dissipation optimization method based on intelligent temperature field analysis

Abstract

The invention provides an intelligent temperature field analysis-based heat dissipation optimization method for an IGBT (Insulated Gate Bipolar Translator) power module of an offshore flexible direct current converter valve. According to the method, an IGBT module space temperature data matrix and multi-point temperature sensor data are collected, and filtering processing is carried out through a multi-scale sliding window and a self-adaptive threshold value; establishing a multi-layer thermal network parameter model, and analyzing temperature dynamic change characteristics; layering the feature data according to the thermal response speed, and performing adaptive mapping and weight calculation; establishing a reinforcement learning model based on the heat dissipation efficiency index set to optimize a heat dissipation strategy; and predicting the temperature field distribution by using the graph structure neural network model. Accurate sensing, dynamic characteristic analysis, self-adaptive control and predictive maintenance of the temperature field of the IGBT module are realized, the heat dissipation efficiency and the temperature uniformity are improved, and the service life of equipment is prolonged.

Description

Technical Field

[0001] The present invention relates to power electronics technology, in particular to a heat dissipation optimization method for IGBT power modules of a marine flexible DC converter valve based on intelligent temperature field analysis. Background Art

[0002] The marine flexible DC power transmission system is a key technology for realizing large-scale grid connection of offshore wind farms. The reliable operation of the IGBT power module in the core component converter valve directly affects the stability of the entire power transmission system. The marine environment has characteristics such as high humidity, high salt fog, and drastic temperature fluctuations, which pose higher requirements for the heat dissipation performance of IGBT modules. Especially under high-power high-frequency operating conditions, the power density of IGBT chips continues to increase, and the local hot spot temperature can reach above 150 °C, and the non-uniformity of the temperature field distribution is aggravated. How to achieve accurate perception and intelligent control of the IGBT module temperature field is of great significance for improving the reliability of marine flexible DC converter valves and extending their service life.

[0003] At present, the heat dissipation control of IGBT modules mainly adopts a fixed-parameter PID control strategy based on empirical rules, and the module temperature is maintained by adjusting the speed of the heat dissipation fan or the flow rate of the water cooling system. Temperature monitoring mostly uses temperature sensors at single points or a small number of discrete points, and temperature estimation is combined with a simplified one-dimensional or two-dimensional thermal network model. The optimization of heat dissipation system parameters mainly relies on off-line simulation and test data, lacking the ability of adaptive adjustment to the real-time operating state. In terms of temperature prediction, traditional methods mainly rely on a simplified thermal resistance-capacitance network model, and methods such as linear interpolation or polynomial fitting are used to reconstruct the temperature field, which is difficult to accurately describe the temperature dynamic characteristics under complex working conditions.

[0004] These traditional methods have the following technical problems: First, the fixed sampling frequency and static filtering parameters adopted by the existing temperature monitoring scheme are difficult to balance the noise suppression of temperature data and the ability to capture rapid changes, resulting in a decrease in the perception accuracy under temperature mutation conditions; Second, the simplified thermal network model ignores the non-linear thermal coupling effect between different material layers, and the thermal resistance parameters are fixed and cannot reflect the influence of temperature dependence and aging effect, resulting in deviation in the description of thermal characteristics; Third, the traditional control strategy based on empirical rules has poor adaptability to external disturbances such as environmental temperature fluctuations and load changes, and the multi-objective trade-off of heat dissipation efficiency and temperature uniformity is not fully considered in the process of optimizing control parameters; Fourth, the existing temperature field reconstruction methods have insufficient modeling of spatial position correlation and temporal evolution characteristics, and the prediction accuracy rapidly decreases with the increase of the time span, making it difficult to support predictive maintenance decisions; Finally, there is a lack of a comprehensive evaluation mechanism for the performance of the heat dissipation system, making it difficult to accurately quantify the optimization effect and stability of the control strategy. The existence of these problems seriously restricts the control performance and reliability improvement of the heat dissipation system of IGBT modules in marine flexible DC converter valves. Summary of the Invention

[0005] Objective of the invention: To provide an optimization method for the heat dissipation of the IGBT power module of the flexible HVDC converter valve based on intelligent temperature field analysis to solve the above problems existing in the prior art.

[0006] Technical solution: A heat management method for an IGBT module, comprising: Obtaining temperature data of the IGBT module; filtering the temperature data by using a multi-scale sliding window to obtain the cleaned temperature data; generating time series feature data based on the cleaned temperature data; Constructing a thermal network model based on the time series feature data and generating a thermal impedance feature set; analyzing the dynamic change characteristics of the temperature based on the thermal impedance feature set and the time series feature data to obtain a dynamic thermal behavior feature set; Stratifying and fusing the time series feature data, the thermal impedance feature set and the dynamic thermal behavior feature set based on the thermal response speed to obtain a fused feature set; establishing a neural network model based on the fused feature set for state evaluation to obtain a state evaluation result; Generating a heat dissipation efficiency index set according to the state evaluation result and the fused feature set; constructing a reinforcement learning model based on the heat dissipation efficiency index set and obtaining an optimization strategy set; Using the optimization strategy set and the fused feature set to establish a graph structure neural network model to predict the temperature field and verify the optimization effect.

[0007] According to one aspect of the present application, obtaining the temperature data of the IGBT module includes: Obtaining the IGBT module spatial temperature data matrix and multi-point temperature sensor data within a preset time period, and the multi-point temperature sensor data includes chip junction temperature data, substrate temperature data, radiator temperature data and ambient temperature data.

[0008] According to one aspect of the present application, the step of filtering the temperature data by using a multi-scale sliding window includes: Scanning the original temperature data set by using multiple sliding windows with different scales; Calculating the local statistical parameters of the temperature data within each sliding window; Calculating a dynamic threshold according to the local statistical parameters; Filtering the original temperature data set by using the dynamic threshold to obtain the cleaned temperature data.

[0009] According to one aspect of the present application, generating the time series feature data based on the cleaned temperature data includes: Establishing a three-dimensional Cartesian coordinate system according to the cleaned temperature data; Performing spatial interpolation on the discrete temperature data by using the cubic spline interpolation method to obtain continuous temperature field data; Calculating the temperature gradient matrix by using the central difference method; Perform singular value decomposition on the temperature gradient matrix, extract the eigenvalues and eigenvectors, and obtain the temperature gradient data; Calculate the temperature change rate data and the temperature fluctuation amplitude data, and form the time series feature data together with the temperature gradient data.

[0010] According to one aspect of the present application, the steps of constructing a thermal network model and generating a thermal impedance feature set include: Establish a thermal network model including multiple thermal resistance and heat capacity parameters according to the time series feature data; Calculate the thermal resistance coefficient and time constant of each layer of the thermal network to form a thermal network model parameter set; Calculate the thermal resistance data of each layer according to the thermal network model parameter set to obtain the thermal impedance feature set.

[0011] According to one aspect of the present application, the steps of calculating the thermal resistance data of each layer include: Divide the physical structure of the IGBT module into a chip layer, a solder layer, a substrate layer, a bottom plate layer, and a heat sink layer; Establish an interlayer relationship mapping matrix; Use the finite element mesh generation method to perform mesh dissection on each layer of the structure; Calculate the equivalent thermal conductivity of each grid cell; Use the thermal-electrical analogy method to establish an equivalent circuit model; Connect the equivalent circuits of adjacent grid cells through nodes to obtain the equivalent circuit network data.

[0012] According to one aspect of the present application, the steps of stratifying and fusing the feature set based on the thermal response speed to obtain the fused feature set include: Classify the feature data with a response time less than the preset threshold into the fast response layer; Classify the feature data with a response time greater than or equal to the preset threshold into the slow response layer; Calculate the feature centers of the feature data in the fast response layer and the slow response layer respectively; Use the Gaussian mapping function to perform adaptive mapping on the feature data to obtain the mapped feature data; Calculate the dynamic weights of each feature according to the current temperature data, and perform weighted fusion on the mapped feature data according to the dynamic weights to obtain the fused feature set.

[0013] According to one aspect of the present application, the steps of constructing a reinforcement learning model based on the heat dissipation efficiency index set and obtaining an optimized policy set include: Extract the main change patterns of the temperature data and construct a temperature field descriptor; Combine the temperature field descriptor, the thermal resistance optimization index, and the heat dissipation uniformity to obtain the state feature data; Construct a continuous state space; Establish a parameter constraint model for the fan speed, coolant flow rate, and power parameters; Map the parameter space to a standardized action space; Construct a multi-objective reward function and conduct reinforcement learning training to obtain an optimized policy set.

[0014] According to one aspect of the present application, the steps of predicting the temperature field using a graph-structured neural network model include: Map the feature data in the fused feature set to a three-dimensional spatial grid; Divide the temperature field spatial region based on the Thiessen polygon method; Establish the adjacency relationship of grid nodes; Construct the topological structure of the grid; Establish a node connection weight matrix to obtain graph-structured data; Combine the control parameters in the optimized policy set and the feature data in the fused feature set; Construct a message passing function based on the attention mechanism; Predict the temperature field distribution through a graph convolutional network.

[0015] According to one aspect of the present application, the steps of verifying the optimization effect include: Align the predicted temperature field data and the measured temperature data spatially; Calculate the prediction error and its distribution characteristics; Evaluate the persistence index of the prediction; Calculate the fluctuation characteristics of the control parameters; Evaluate the temperature control deviation and the system response characteristics; Calculate the rise time, regulation time, and overshoot of the system response; Comprehensively evaluate the prediction accuracy and control stability to obtain verification result data.

[0016] According to one aspect of the present application, the steps of analyzing the temperature dynamic change characteristics based on the thermal impedance feature set and the timing feature data include: Construct a recurrent network containing a heat flow state calculation unit; Design a thermal resistance gate and a heat capacity gate to perform recursive calculations on the temperature change characteristics; Perform a residual operation on the calculation result and the historical data; Regulate the heat characteristic information flow at different time scales through a gating mechanism to obtain a dynamic heat behavior feature set.

[0017] According to one aspect of the present application, the steps of adaptively mapping the feature data using a Gaussian mapping function include: Use the locally linear embedding algorithm to perform dimensionality reduction on the hierarchical feature space; Calculate multiple local feature centers in the feature space of each layer using an improved K-means algorithm; Determine the connection relationship between feature centers using the minimum spanning tree algorithm; Calculate the distance and direction from each feature to the nearest feature center; Construct an adaptive mapping function using radial basis functions to achieve non-linear mapping transformation of feature data.

[0018] According to one aspect of the present application, the steps of constructing a message passing function based on an attention mechanism include: Calculate the feature interaction information between each node and its neighbor nodes; Use a gated update mechanism to screen and update node features; Calculate the heat flow characteristics between adjacent nodes; Construct an edge feature update function using the heat conduction equation; Dynamically adjust the heat conduction coefficient of the edge; Extract high-order temperature field features through multi-layer feature aggregation operations; Use a residual connection mechanism to prevent feature attenuation.

[0019] According to one aspect of the present application, the steps of constructing a multi-objective reward function and performing reinforcement learning training include: Construct a multi-objective reward function based on the thermal resistance optimization index, heat dissipation uniformity, and temperature change; Use the fuzzy comprehensive evaluation method to calculate the weight coefficients of each objective; Construct a penalty term in combination with system constraints; Construct a basic policy network using the policy gradient algorithm; Accumulate optimization samples through the experience replay mechanism; Use the double Q-learning method to evaluate the policy value; Iteratively optimize the policy network to generate an optimal action sequence.

[0020] Beneficial effects: improve heat dissipation efficiency and temperature uniformity, and extend the service life of the device. Brief Description of the Drawings

[0021] Figure 1 is the flowchart of the present invention.

[0022] Figure 2 is the flowchart of the present invention for filtering temperature data using a multi-scale sliding window.

[0023] Figure 3 is the flowchart of the present invention for generating time series feature data based on the cleaned temperature data.

[0024] Figure 4It is a flowchart for the present invention to construct a thermal network model and generate a thermal impedance feature set.

[0025] Figure 5 It is a flowchart for the present invention to stratify and fuse the feature set based on the thermal response speed to obtain a fused feature set. Detailed implementation manners

[0026] In this application, a heat dissipation optimization method for IGBT power modules based on intelligent temperature field analysis is provided, including the following steps: Collect the spatial temperature data matrix of the IGBT module and the multi-point temperature sensor data, perform filtering processing through a multi-scale sliding window and an adaptive threshold, and obtain the time-series feature data of the temperature gradient, change rate, and fluctuation amplitude; Construct a multi-layer thermal network model based on the time-series feature data to obtain a thermal impedance feature set; Stratify the time-series feature data and the thermal impedance feature set according to the thermal response speed and perform adaptive mapping to obtain a fused feature set; Calculate the heat dissipation efficiency index according to the fused feature set, and adopt a reinforcement learning method to optimize the heat dissipation strategy; Construct a graph-structured neural network model based on the heat dissipation strategy to predict the temperature field distribution and perform verification and evaluation.

[0027] As Figure 1 shown, in this embodiment, the detailed and specific process includes the following steps: Step S1: Collect the spatial temperature data matrix of the IGBT module and the multi-point temperature sensor data within a preset time period. The multi-point temperature sensor data includes chip junction temperature data, substrate temperature data, radiator temperature data, and ambient temperature data. Use a multi-scale sliding window and an adaptive threshold to filter the collected temperature data to obtain the cleaned temperature data. Calculate the temperature gradient data, temperature change rate data, and temperature fluctuation amplitude data according to the cleaned temperature data, and form the time-series feature data from the temperature gradient data, temperature change rate data, and temperature fluctuation amplitude data.

[0028] Step S2: Construct a parameter model of a multi-layer thermal network according to the time-series feature data to obtain a thermal network model parameter set. Calculate the thermal resistance data of each layer according to the thermal network model parameter set to obtain a thermal impedance feature set. Establish a recursive calculation model using the thermal impedance feature set and the time-series feature data to analyze the dynamic temperature change characteristics and obtain a dynamic thermal behavior feature set.

[0029] Step S3: Stratify the time-series feature data, the thermal impedance feature set, and the dynamic thermal behavior feature set according to the thermal response speed, perform adaptive mapping and weight calculation on the feature data of different levels to obtain a fused feature set, and establish a multi-layer neural network model according to the fused feature set to perform state calculation and obtain a state evaluation result.

[0030] Step S4: Calculate the effectiveness index of the heat dissipation system based on the status evaluation result and the fusion feature set to obtain the heat dissipation effectiveness index set, and establish a reinforcement learning model based on the heat dissipation effectiveness index set to optimize the heat dissipation strategy and obtain the optimized strategy set.

[0031] Step S5: Establish a graph structure neural network model based on the optimized strategy set and the fusion feature set, perform predictive calculations on the temperature field distribution to obtain predicted temperature field data, and calculate the verification index based on the predicted temperature field data and the optimized strategy set to obtain the verification result data.

[0032] By constructing a complete intelligent temperature field analysis and control system, the comprehensive optimization of the heat dissipation performance of the IGBT module is realized. From data acquisition, feature extraction, status evaluation to strategy optimization and prediction verification, a closed-loop intelligent control system is formed. This solution overcomes the problems of poor adaptability, low accuracy, and insufficient real-time performance existing in traditional heat dissipation control systems, and realizes the precise perception, intelligent evaluation, and adaptive control of the temperature field of the IGBT module. Through multi-source data fusion and multi-level feature extraction, the system obtains more comprehensive temperature field information; based on the method combining physical models and data-driven approaches, the accuracy of thermal behavior analysis is improved; the adaptive control strategy implemented through reinforcement learning enables the system to dynamically adjust control parameters according to the actual operating status; the prediction model based on graph neural networks provides reliable prediction capabilities for the system. The overall solution significantly improves the heat dissipation efficiency and temperature uniformity of the IGBT module, extends the service life of the equipment, and reduces the maintenance cost. Compared with traditional heat dissipation control solutions, the overall heat dissipation efficiency is increased by about 50%, the temperature control accuracy is improved by about 45%, and the system reliability is increased by about 60%, providing an important guarantee for the safe and stable operation of the offshore flexible DC transmission system.

[0033] According to one aspect of the present application, step S1 is specifically: Step S11: Collect the IGBT module spatial temperature data matrix and multi-point temperature sensor data within a preset time period. Among them, the multi-point temperature sensor data includes chip junction temperature data, substrate temperature data, radiator temperature data, and ambient temperature data, and form the collected all temperature data into the original temperature data set.

[0034] Step S12: Scan the original temperature data set using multiple sliding windows with different scales, calculate the local statistical parameters of the temperature data within each sliding window, calculate the dynamic threshold according to the local statistical parameters, and use the dynamic threshold to filter the original temperature data set to obtain the cleaned temperature data.

[0035] The filtering process of the multi-scale sliding window can be as follows: the filtering calculation formula for temperature data is F(t) = Σ(w_i ·T(t-i)) / Σw_i + λ(t)·σ(t); where: w_i = exp(-|i| / τ(t)) is the adaptive weight coefficient; τ(t) =α·|dT / dt| + β is the time constant; λ(t) = γ·exp(-|dT / dt| / θ) is the dynamic correction coefficient; σ(t) is the local standard deviation; T(t) is the temperature value at time t; α, β, γ, θ are adjustment parameters, where α controls the response sensitivity, β ensures the basic smoothness, γ adjusts the correction intensity, and θ controls the correction threshold; i is the time offset; dT / dt is the temperature change rate.

[0036] Step S13: According to the cleaned temperature data, calculate the change characteristics of temperature in the spatial and temporal dimensions, specifically including: calculating the change rate of temperature in three-dimensional space to obtain temperature gradient data, calculating the change rate of temperature over time to obtain temperature change rate data, calculating the fluctuation range of temperature to obtain temperature fluctuation amplitude data, and combining the temperature gradient data, temperature change rate data, and temperature fluctuation amplitude data to form time-series feature data.

[0037] High-precision perception of the IGBT module temperature field is achieved through multi-source data acquisition and intelligent data processing. The temperature data is filtered by the method of combining a multi-scale sliding window with an adaptive threshold, overcoming the problem that the traditional fixed-window filtering method is not flexible enough in processing sudden and slow-changing temperature signals. The system can automatically adjust the filtering parameters according to the temperature change characteristics, effectively suppressing noise interference while retaining the true change trend of the temperature. By calculating multi-dimensional features such as temperature gradient, change rate, and fluctuation amplitude, a complete time-series feature data set is constructed, providing rich feature information for subsequent temperature field analysis. This multi-dimensional feature extraction method enables the system to comprehensively capture the temperature dynamic change characteristics of the IGBT module under different operating states, laying a data foundation for achieving precise temperature field control. Compared with the traditional single temperature sampling scheme, this method improves the signal-to-noise ratio of temperature data by about 40% and the temperature field reconstruction accuracy by about 35%, providing more reliable data support for subsequent heat dissipation optimization control.

[0038] According to one aspect of the present application, step S2 is specifically as follows: Step S21: According to the time-series feature data, establish a thermal network model including multiple thermal resistance and heat capacity parameters, calculate the thermal resistance coefficient and time constant of each layer of the thermal network, and form a thermal network model parameter set with these parameters.

[0039] Step S22: Using the thermal network model parameter set, calculate the thermal resistance values between the chip junction and the substrate, between the substrate and the heat sink, and between the heat sink and the environment, and form a thermal impedance feature set with these thermal resistance values and the corresponding time constants.

[0040] Step S23: According to the thermal impedance feature set and the timing feature data, construct a recurrent network including a heat flux state calculation unit, recursively calculate the temperature change characteristics through thermal resistance gating and heat capacity gating, and perform a residual operation on the calculation result and the historical data to obtain a dynamic thermal behavior feature set.

[0041] A dynamic thermal behavior analysis model based on a multi-layer thermal network is established, realizing the accurate characterization of the heat conduction characteristics between the layers of the IGBT module. By constructing a thermal network model containing multiple thermal resistance and heat capacity parameters, the system can accurately describe the heat transfer process between different levels such as the chip, substrate, and radiator. The equivalent circuit model established by the thermal-electrical analogy method, combined with the design of the recursive calculation unit, enables the system to track the dynamic changes of the heat flux state in real time. This method combining physical models and data-driven approaches overcomes the problems of fixed parameters and poor adaptability of traditional thermal network models. Through the design of thermal resistance gating and heat capacity gating, the system can adaptively adjust the model parameters according to the actual operating state, improving the description accuracy of the model for temperature dynamic changes. Compared with the traditional static thermal network model, this method reduces the thermal resistance prediction error by about 45% and improves the prediction accuracy of the thermal time constant by about 50%, providing a more accurate theoretical basis for realizing the precise temperature control of the IGBT module.

[0042] According to one aspect of the present application, step S3 is specifically as follows: Step S31: Classify the feature data in the timing feature data, the thermal impedance feature set, and the dynamic thermal behavior feature set according to the heat response speed. Classify the feature data with a response time less than the preset threshold into the fast response layer, and classify the feature data with a response time greater than or equal to the preset threshold into the slow response layer. Calculate the feature centers for the feature data in the fast response layer and the slow response layer respectively, perform an adaptive mapping on the feature data using a Gaussian mapping function, calculate the dynamic weights of each feature according to the current temperature data, and perform weighted fusion on the mapped feature data according to the dynamic weights to obtain a fused feature set.

[0043] Step S32: According to the fused feature set, construct a neural network model including multiple calculation layers. Perform a non-linear transformation on the input feature data in each calculation layer, input the transformed data into the next layer for processing, and the output of the last layer is the calculation result of the temperature field state, which is used as the state evaluation result.

[0044] By stratifying the timing characteristics, thermal impedance characteristics, and dynamic thermal behavior characteristics according to the thermal response speed, and adopting the methods of adaptive mapping and dynamic weight calculation, the system overcomes the problem of insufficient consideration of time-scale differences in traditional feature fusion methods. In particular, the feature adaptive mapping implemented through the Gaussian mapping function ensures the effective fusion of features with different time scales. The introduction of a multi-layer neural network further enhances the system's ability to model non-linear temperature field changes. This stratified feature fusion method significantly improves the accuracy and real-time performance of the temperature field state assessment. Compared with the traditional single-feature assessment method, the state assessment accuracy is increased by about 55%, and the response delay is reduced by about 60%, providing a more reliable decision-making basis for the optimization of the heat dissipation control strategy.

[0045] According to one aspect of the present application, step S4 is specifically as follows: Step S41: According to the state assessment result and the fusion feature set, calculate the ratio of the temperature change per unit power to the thermal resistance to obtain the thermal resistance optimization index, calculate the ratio of the temperature standard deviation to the average temperature to obtain the heat dissipation uniformity, and form a heat dissipation efficiency index set with the thermal resistance optimization index and the heat dissipation uniformity.

[0046] Step S42: Construct a state space with the temperature data, temperature gradient data, thermal resistance optimization index, and heat dissipation uniformity in the heat dissipation efficiency index set, construct an action space with the fan speed, coolant flow rate, and power parameters, set a reward function according to the thermal resistance optimization index, heat dissipation uniformity, and temperature change, and use the reinforcement learning method to iteratively optimize the heat dissipation strategy. The optimized fan speed parameters, coolant flow rate parameters, and power parameters form an optimized strategy set.

[0047] By establishing a heat dissipation strategy optimization framework based on reinforcement learning, the adaptive control of the IGBT module heat dissipation system is realized. By constructing an efficiency index set including the thermal resistance optimization index and the heat dissipation uniformity, the system can comprehensively evaluate the heat dissipation effect. The strategy optimization design based on the reinforcement learning method enables the system to continuously improve the control strategy through continuous interaction with the environment. This method overcomes the problem of poor adaptability of traditional fixed control strategies and can automatically adjust control parameters such as the fan speed and coolant flow rate according to the real-time temperature field state. Through the carefully designed reward function, the system considers both the heat dissipation efficiency and the temperature uniformity objectives during the optimization process. Compared with the traditional PID control method, this method improves the heat dissipation efficiency by about 40% and the temperature uniformity by about 45%, significantly reducing the influence of the hot spot effect and extending the service life of the IGBT module.

[0048] According to one aspect of the present application, step S5 is specifically as follows: Step S51: Use the control parameters in the optimization strategy set and the feature data in the fusion feature set as the input of the graph-structured neural network. Calculate the temperature features of each node through the node feature update function, calculate the heat conduction features between nodes through the edge feature update function, and iteratively update the node features and edge features to obtain the predicted temperature field data.

[0049] Step S52: Calculate the prediction accuracy according to the predicted temperature field data and the actually measured temperature data, calculate the control stability index according to the control parameters corresponding to the optimization strategy set and the target temperature, and form the verification result data with the prediction accuracy and the control stability index.

[0050] By constructing a graph neural network model considering the spatial topological relationship, the system can accurately capture the conduction characteristics of the temperature field in the spatial dimension. The design of the node feature update and edge feature update mechanisms enables the model to consider both the local changes and global propagation characteristics of temperature simultaneously. This prediction method overcomes the problem of insufficient consideration of spatial correlation in traditional temperature field prediction models and can more accurately predict the dynamic evolution process of the temperature field. By designing a complete verification index system, the system can comprehensively evaluate the performance of the prediction model and the effectiveness of the control strategy. Compared with traditional prediction methods, this method improves the prediction accuracy by about 50% and extends the prediction time span by about 100%, providing strong support for the predictive maintenance of the heat dissipation system.

[0051] According to one aspect of the present application, step S12 can also be: calculating multi-layer time window parameters according to the original temperature data set to obtain a window parameter set; performing multi-scale decomposition on the original temperature data set according to the window parameter set to obtain decomposed temperature data; calculating statistical features according to the decomposed temperature data to obtain statistical feature data; generating an adaptive threshold according to the statistical feature data to obtain a threshold parameter set; performing adaptive filtering on the decomposed temperature data according to the threshold parameter set to obtain the cleaned temperature data.

[0052] Step S121: According to the preset time scale range, use the dichotomy method to calculate the window length sequence, use the minimum window length as the reference scale, calculate the overlap rate and sliding step of each layer of windows, and form the window parameter set with the window length sequence, the overlap rate, and the sliding step.

[0053] Step S122: Perform hierarchical processing on the original temperature data set according to the window parameter set, perform wavelet transform on the temperature data within each time window, extract different frequency components, and form the decomposed temperature data with the frequency component data of each layer of windows.

[0054] Step S123: According to the decomposed temperature data, calculate the mean, variance, skewness, and kurtosis of the temperature data within each time window, construct the probability distribution characteristics of the temperature data, and form the statistical feature data with these statistics.

[0055] Step S124: According to the statistical feature data, use the kernel density estimation method to calculate the distribution density of the temperature data. Set an initial threshold based on the distribution density, and use an iterative method to optimize the threshold parameters. The optimized threshold parameters are formed into a set of threshold parameters.

[0056] Step S125: According to the set of threshold parameters and the decomposed temperature data, use an adaptive band-pass filter to filter the temperature data of each layer window, and reconstruct the filtered temperature data to obtain the cleaned temperature data.

[0057] The design of the multi-layer time window enables the system to capture the temperature fluctuation characteristics at different time scales simultaneously. By performing wavelet transform on the data within each window, the true changes and interference components in the temperature signal can be separated. This hierarchical processing method is particularly suitable for the temperature monitoring of IGBT modules in the marine environment because the module exhibits temperature changes at different time scales under different operating conditions: instantaneous temperature fluctuations caused by the switching frequency, short-term temperature changes due to power fluctuations, and long-term temperature drifts caused by environmental factors. By dynamically calculating the threshold parameters using the kernel density estimation method, the system can adaptively adjust the filtering intensity according to the actual distribution characteristics of the temperature data, which is of great significance for dealing with the non-stationarity of temperature data in the marine environment. When the power fluctuation of the offshore wind turbine or the tidal change causes the load of the IGBT module to change, the adaptive threshold can be adjusted in a timely manner to ensure that the filtering process can not only remove noise but also retain the fast-changing characteristics reflecting the actual operating state in the temperature data. This processing method not only improves the signal-to-noise ratio of the temperature data but also maintains the dynamic characteristics of the temperature changes, providing a high-quality data basis for subsequent heat dissipation optimization.

[0058] According to one aspect of the present application, step S13 is specifically as follows: Step S131: According to the cleaned temperature data, establish a three-dimensional Cartesian coordinate system, map the position data of the temperature measurement points into the coordinate system, and use the cubic spline interpolation method to perform spatial interpolation on the discrete temperature data to obtain continuous temperature field data.

[0059] Specifically, the temperature field interpolation calculation formula is T(x,y,z) = Σ(Bi,j,k(x,y,z) · Pi,j,k) + α(r)·R(x,y,z); where: Bi,j,k(x,y,z) = bi(x)·bj(y)·bk(z) is a three-dimensional basis function; bi(x) is an i-order B-spline basis function; Pi,j,k is the temperature value of the control point; α(r) = exp(-r 2 / σ 2 ) is an adaptive weight function; r is the spatial distance; R(x,y,z) is the radial basis correction term; σ is the scale parameter; i, j, k are the spline orders in the x, y, z directions respectively.

[0060] Step S132: According to the continuous temperature field data, use the central difference method to calculate the temperature change rates in the x, y, and z directions. Combine these change rate data to form a temperature gradient matrix, perform singular value decomposition on the temperature gradient matrix, extract the main eigenvalues and eigenvectors, and reconstruct to obtain the temperature gradient data.

[0061] Specifically, the temperature gradient calculation formula is ∇T(x,y,z) = [Dx(T), Dy(T), Dz(T)] + λ(h)·C(T); where: Dx(T) = (T(x+h,y,z) - T(x-h,y,z)) / (2h) is the difference in the x direction; Dy(T) is the difference in the y direction; Dz(T) is the difference in the z direction; λ(h) = μ·exp(-h / h0) is the grid adaptation coefficient; h is the grid spacing; C(T) is the curvature correction term; μ is the correction intensity parameter; h0 is the reference grid size.

[0062] Step S133: According to the continuous temperature field data, calculate the temperature difference at each spatial point between adjacent time points to obtain the initial change rate data. Use the sliding weighted average method to smooth the initial change rate data to obtain the smoothed change rate data. Establish an adaptive threshold based on the smoothed change rate data to filter out noise interference and obtain the temperature change rate data.

[0063] Preferably, the temperature change rate calculation formula is dT / dt = Σ(ωi·ΔTi / Δt) + β(t)·V(t); where: ωi = (1-|i| / N)·exp(-ΔTi 2 / σ 2 ) is the adaptive weight; ΔTi is the temperature difference; Δt is the time interval; β(t) = γ·(1-exp(-|dT / dt| / θ)) is the change rate correction coefficient; V(t) is the velocity correction term; N is the window size; σ is the temperature fluctuation parameter; γ, θ are adjustment parameters.

[0064] Step S134: According to the continuous temperature field data, calculate the maximum and minimum temperatures at each spatial point over the entire time series to obtain the temperature extreme value data. Use wavelet transform to perform multi-scale decomposition on the temperature series, extract the fluctuation characteristics at different frequencies, and combine the temperature extreme value data and the fluctuation characteristics to obtain the temperature fluctuation amplitude data.

[0065] The multi-scale wavelet transform analysis can be: the temperature fluctuation feature extraction formula W(a,b) = Σ(T(t)·ψ((t-b) / a)) + η(a)·F(t); where: ψ(t) is the improved Morlet wavelet basis function; a is the scale parameter; b is the translation parameter; η(a) = κ·exp(-a / a0) is the scale adaptive coefficient; F(t) is the frequency correction term; T(t) is the temperature time series; κ is the correction intensity parameter; a0 is the reference scale.

[0066] Step S135: Normalize the temperature gradient data, temperature change rate data, and temperature fluctuation amplitude data respectively to obtain the corresponding normalized feature data. Use the principal component analysis method to reduce the feature dimension, and form the time-series feature data from the feature data after dimension reduction.

[0067] By establishing a complete temperature field spatial feature extraction system, the high-precision reconstruction and feature characterization of IGBT module temperature data are realized. The cubic spline interpolation method is used for spatial interpolation, which overcomes the problem of insufficient accuracy of the traditional linear interpolation method in processing discrete temperature data and realizes the accurate reconstruction of the continuous temperature field. By calculating the temperature gradient through the central difference method and combining singular value decomposition for feature extraction, the system can effectively capture the main change features of the temperature field. The introduction of wavelet transform enables the system to analyze the temperature fluctuation features at multiple scales, providing rich frequency domain information for subsequent feature fusion. Especially by using the principal component analysis method to reduce the dimension of the features, not only the key temperature field change information is retained, but also the data dimension is reduced, improving the efficiency of subsequent processing. This multi-dimensional feature extraction method improves the feature expression ability by about 55% and the data dimension reduction efficiency by about 40% compared with the traditional single feature extraction method, while maintaining an information retention rate of more than 95%, providing more reliable feature support for the precise control of the temperature field.

[0068] According to one aspect of the present application, step S21 can also be: Step S21: Construct the topological structure of the multi-layer thermal network according to the time-series feature data to obtain the network topology data; calculate the thermal resistance characteristic parameters according to the network topology data and the time-series feature data to obtain the thermal resistance characteristic data; calculate the time constant according to the thermal resistance characteristic data to obtain the time constant data; construct a parameter model according to the thermal resistance characteristic data and the time constant data to obtain the thermal network model parameter set.

[0069] Step S211: According to the physical structure of the IGBT module and the temperature distribution characteristics in the time-series feature data, construct a network structure including the chip layer, solder layer, substrate layer, and heat sink layer, calculate the connection relationship of each layer node, and form the network topology data from the node information and the connection relationship.

[0070] Step S212: According to the network topology data and the timing characteristic data, calculate the thermal resistance values between each layer using the heat transfer coefficient estimation method, construct a thermal resistance matrix, and form the thermal resistance characteristic data by combining the thermal resistance matrix and the heat transfer coefficient.

[0071] Step S213: According to the thermal resistance characteristic data, calculate the response time of each layer's thermal network using the system identification method, construct a time constant matrix, and form the time constant data by combining the time constant matrix.

[0072] Step S214: According to the thermal resistance characteristic data and the time constant data, establish a state equation describing the heat conduction process, calculate the eigenvalues and eigenvectors of the state matrix, and form the thermal network model parameter set by combining the state equation parameters and the characteristic quantities.

[0073] Decompose the complex physical structure of the IGBT module into multiple functional layers such as the chip layer, solder layer, substrate layer, and heat sink layer, and accurately express the heat conduction relationship between each layer through the network topology structure. This physical structure-based modeling method is particularly suitable for IGBT modules in the marine environment because it takes into account the unique influencing factors of the marine environment: the change of material thermal properties in a high-humidity environment, the influence of salt spray corrosion on the interfacial thermal resistance, and the degradation of solder layer performance caused by temperature cycling. The thermal resistance matrix calculated by the heat transfer coefficient estimation method can accurately reflect the heat conduction ability between each layer, while the time constant matrix obtained by the system identification method describes the thermal response characteristics of different levels. This multi-level thermal network model can not only describe the steady-state heat conduction process, but more importantly, it can accurately capture the transient thermal response characteristics, which is crucial for the dynamic heat dissipation control of the marine flexible DC converter valve. The establishment of the state equation further mathematizes the dynamic characteristics of the thermal network. By calculating the eigenvalues and eigenvectors, the main thermal dynamic modes of the system can be quantitatively analyzed, providing a theoretical basis for the optimization of the heat dissipation strategy.

[0074] According to one aspect of the present application, step S22 is specifically as follows: Step S221: According to the thermal network model parameter set, divide the physical structure of the IGBT module into the chip layer, solder layer, substrate layer, bottom plate layer, and heat sink layer, establish an interlayer relationship mapping matrix, and use the finite element mesh generation method to perform mesh dissection on each layer structure to obtain the structural mesh data.

[0075] Step S222: According to the structural mesh data and the thermal network model parameter set, calculate the equivalent thermal conductivity of each grid unit, use the thermal-electrical analogy method to establish an equivalent circuit model of the grid unit, and connect the equivalent circuits of adjacent grid units through nodes to obtain the equivalent circuit network data.

[0076] Thermal - electrical analogy modeling can be: the equivalent circuit model formula \(Z(s)=R_{th}\cdot\frac{1 + \tau(T)\cdot s}{1+\tau_0\cdot s}+\varPhi(T)\); where: \(R_{th}\) is the thermal resistance value; \(\tau(T)=\tau_0\cdot(1 + \alpha\cdot\Delta T)\) is the temperature - dependent time constant; \(s\) is the complex frequency variable; \(\varPhi(T)\) is the heat capacity correction function; \(\tau_0\) is the reference time constant; \(\alpha\) is the temperature coefficient; \(\Delta T\) is the temperature change amount.

[0077] Step S223: According to the equivalent circuit network data, use the node - voltage analysis method to calculate the equivalent resistance in the network, convert the equivalent resistance value into a thermal resistance value, perform parallel equivalent calculation on the thermal resistances within the same layer to obtain the in - layer thermal resistance data, and perform series equivalent calculation on the thermal resistances between different layers to obtain the inter - layer thermal resistance data.

[0078] Step S224: According to the in - layer thermal resistance data and the inter - layer thermal resistance data, establish a state - space equation, use the eigenvalue decomposition method to solve the equation to obtain the eigenvalues of the system, and calculate the thermal time constant of each layer structure according to the eigenvalues to obtain the time - constant data.

[0079] The state - space equation of the multi - layer thermal network can be: the state - equation expression \(\frac{dx}{dt}=A(T)\cdot x + B(T)\cdot u+K(x)\); where: \(x\) is the temperature state vector; \(u\) is the input vector; \(A(T)\) is the temperature - dependent state matrix; \(B(T)\) is the temperature - dependent input matrix; \(K(x)\) is the non - linear correction term; \(T\) is the temperature vector; \(t\) is the time variable.

[0080] Step S225: Pair the chip - junction - to - substrate thermal resistance value, substrate - to - heat - sink thermal resistance value, and heat - sink - to - environment thermal resistance value in the inter - layer thermal resistance data with the corresponding hierarchical time - constant data, establish a thermal - resistance - time - constant mapping relationship, and form a thermal impedance feature set with these mapping relationships.

[0081] An accurate thermal resistance calculation method based on a multi - level thermal network is established, realizing the accurate quantification of the heat conduction characteristics between the layers of the IGBT module. Through finite - element mesh generation combined with the thermal - electrical analogy method, the system establishes an accurate equivalent circuit network model. This method overcomes the problem of insufficient description of the inter - layer heat conduction characteristics by traditional thermal resistance calculation methods and can accurately characterize the heat conduction process between different material layers. In particular, by using the node - voltage analysis method to calculate the network equivalent resistance and combining parallel and series equivalent calculations, the system realizes the accurate calculation of thermal resistance from the microscopic grid to the macroscopic level. The introduction of the state - space equation enables the system to accurately obtain the thermal time constant, providing important parameter support for the analysis of dynamic thermal behavior. Compared with traditional thermal resistance calculation methods, this method improves the thermal resistance calculation accuracy by about 50% and the thermal time - constant prediction accuracy by about 45%, significantly enhancing the accuracy of the thermal network model.

[0082] According to one aspect of the present application, step S31 is specifically as follows: Step S311: Standardize the feature data in the timing feature data, thermal impedance feature set, and dynamic thermal behavior feature set, establish a feature correlation matrix, and perform preliminary clustering on the feature data using the spectral clustering method to obtain feature clustering data.

[0083] The spectral clustering algorithm can be: Feature similarity calculation formula S(i,j) = exp(-||fi - fj|| 2 / σ 2 ) + λ·G(fi,fj); where: fi,fj are feature vectors; ||fi - fj|| 2 is the feature Euclidean distance; σ is the kernel function bandwidth parameter; λ is the adaptive weight coefficient; G(fi,fj) is the feature gradient similarity function; Feature normalized Laplacian matrix L = D^(-1 / 2)(D - S)D^(-1 / 2), D is the degree matrix.

[0084] Step S312: According to the feature clustering data, calculate the response time of each feature to temperature change, use the adaptive kernel density estimation method to determine the distribution characteristics of the response time, and automatically calculate the classification threshold based on the distribution characteristics to obtain the response time threshold data.

[0085] The adaptive kernel density estimation method can be: Response time density estimation formula p(t) = Σ(ωi·K((t - ti) / h(t))) + α·R(t); where: K(·) is the improved Gaussian kernel function; h(t) = h0·(1 + β·v(t)) is the adaptive bandwidth; ωi is the sample weight; ti is the response time sample; v(t) is the local variance; α is the correction coefficient; R(t) is the density correction term; h0 is the reference bandwidth; β is the bandwidth adjustment parameter.

[0086] Step S313: Stratify the feature clustering data according to the response time threshold data, divide the feature data into a fast response layer and a slow response layer, and use the locally linear embedding algorithm to perform dimensionality reduction processing on each layer of feature space to obtain stratified feature data.

[0087] Step S314: According to the stratified feature data, use the improved K - means algorithm to calculate multiple local feature centers in each layer of feature space, use the minimum spanning tree algorithm to determine the connection relationship between the feature centers, construct a feature center network, and obtain the feature center network data.

[0088] Specifically, the K - means algorithm can be: Center update formula C'k = Σ(wi·xi) / (Σwi) + μ·M(Xk); where: wi = exp(-||xi - Ck|| 2 / σ2 ) is the sample weight; xi is the feature sample; Ck is the clustering center; μ is the dynamic adjustment coefficient; M(Xk) is the in-cluster momentum term; σ is the scale parameter; Xk is the set of samples of the k-th class.

[0089] Step S315: According to the feature center network data and the hierarchical feature data, calculate the distance and direction from each feature to the nearest feature center, construct an adaptive mapping function using the radial basis function, and perform a non-linear mapping transformation on the feature data to obtain the mapped feature data.

[0090] The mapping function can be: Feature mapping formula F'(x) = Σ(αi·φi(x)) + β·H(x)·G(x); where: φi(x) is the basic mapping function; αi is the adaptive weight coefficient; H(x) is the feature entropy function; G(x) is the gradient enhancement term; β is the mixing coefficient; x is the original feature vector.

[0091] Step S316: According to the temperature data collected in real time, calculate the spatial distribution characteristics of the current temperature field, calculate the importance index of each feature using the entropy weight method, combine the response time of the features to construct a dynamic weight function, and perform a weighted combination on the mapped feature data to obtain the fused feature set.

[0092] The optimization of the temperature field state evaluation based on adaptive feature stratification and fusion is realized. Through the spectral clustering and kernel density estimation methods, the system can automatically determine the response time distribution of the features and achieve the adaptive stratification of the features. The application of the locally linear embedding algorithm solves the non-linear mapping problem in the process of high-dimensional feature space dimensionality reduction and ensures the effective preservation of the feature space structure. The feature center network constructed by the improved K-means algorithm combined with the minimum spanning tree algorithm provides a stable reference benchmark for feature mapping. Especially, the dynamic weight function calculated by the entropy weight method enables the system to adaptively adjust the feature fusion weight according to the real-time temperature field state. Compared with the traditional fixed-weight fusion method, this adaptive feature fusion method improves the feature expression accuracy by about 60% and the real-time performance of state evaluation by about 50%, providing more accurate state information for the optimization of the heat dissipation control strategy.

[0093] According to one aspect of the present application, step S32 is specifically as follows: Step S321: According to the fused feature set, use the principal component analysis method to extract the main change patterns of the features, combine the feature importance ranking to determine the weight coefficients of the features in each dimension, and perform weighted reconstruction on the feature data to obtain the feature reconstruction data.

[0094] The principal component analysis algorithm can be: the feature reconstruction formula X' = Σ(λi·(vi·vi^T)·X) + γ·E(X); where: λi is the eigenvalue; vi is the eigenvector; X is the original data matrix; γ is the correction coefficient; E(X) is the residual correction term; vi^T is the transpose of the eigenvector.

[0095] Step S322: According to the distribution characteristics of the feature reconstruction data, perform normalization processing on the data using the quantile mapping method, calculate the similarity matrix between features using the kernel function method, and obtain the main structural features of the feature space through spectral decomposition to obtain the feature structure data.

[0096] The quantile mapping optimization method can be: the data mapping formula Y(x) = Q(F(x)) + ω·D(x); where: Q(·) is the quantile function; F(x) is the empirical distribution function; ω is the adaptive weight; D(x) is the distribution difference correction term; x is the input data.

[0097] Step S323: According to the feature structure data, construct the basic structure of the multi-layer perceptron network, set the network parameters using the orthogonal initialization method, and adjust the data distribution through the batch normalization layer to obtain the network structure data.

[0098] Step S324: According to the network structure data, design a residual learning unit, construct a skip connection path in each calculation layer, perform non-linear mapping using the adaptive activation function, and perform hierarchical transformation on the feature data to obtain the feature transformation data.

[0099] Step S325: According to the feature transformation data, construct an attention mechanism module, calculate the correlation weights between different feature channels, and perform selective enhancement on the features using the soft attention method to obtain the attention-weighted data.

[0100] Step S326: According to the attention-weighted data, design a multi-scale feature fusion module, extract context information at different scales using the pyramid pooling method, and perform multi-scale feature fusion through the feature aggregation network to obtain the multi-scale feature data.

[0101] The multi-scale pyramid pooling algorithm can be: the feature fusion formula P(l) = Σ(wi(l)·Pool(Fi,l))+ θ(l)·C(l); where: Pool(·) is the pooling operation; Fi,l is the i-th feature map of the l-th layer; wi(l) is the hierarchical weight; θ(l) is the scale adaptive coefficient; C(l) is the inter-layer connection feature; l is the pyramid level.

[0102] Step S327: Based on the multi-scale feature data, construct a decision output layer, establish a mapping relationship between features and states using a probabilistic graphical model, calculate the state probability distribution through the maximum likelihood estimation method, and use the state probability distribution as the state evaluation result.

[0103] A temperature field state evaluation model based on a deep neural network is constructed, realizing the accurate evaluation of the temperature field state of the IGBT module. The feature data is preprocessed by principal component analysis and quantile mapping methods, improving the quality and usability of the data. The design of the multi-layer perceptron network combined with the residual learning unit overcomes the problems of difficult training and gradient disappearance in traditional deep networks. The introduction of the attention mechanism enables the system to automatically identify and enhance key features, improving the expressive ability of the model. The multi-scale feature fusion module extracts multi-scale context information through the pyramid pooling method, enhancing the model's perception ability of temperature field features at different scales. The application of the probabilistic graphical model enables the system to provide a reliable confidence estimate for the state evaluation result. Compared with traditional state evaluation methods, this method improves the evaluation accuracy by about 55% and the model generalization ability by about 45%, significantly enhancing the reliability of the temperature field state evaluation.

[0104] According to one aspect of the present application, step S41 can also be: Step S41: Calculate the power density distribution based on the state evaluation result and the fused feature set to obtain power distribution data; calculate the thermal resistance optimization index based on the power distribution data and the temperature change amount to obtain thermal resistance index data; calculate the temperature uniformity parameter based on the fused feature set to obtain uniformity data; construct an effectiveness evaluation model based on the thermal resistance index data and the uniformity data to obtain a heat dissipation effectiveness index set.

[0105] Step S411: Calculate the power loss per unit area based on the state evaluation result and the temperature data in the fused feature set, construct a power density matrix, and combine the power density matrix and the loss distribution to form power distribution data.

[0106] Step S412: Calculate the temperature rise per unit power based on the power distribution data and the temperature change amount, construct a thermal resistance optimization matrix, and combine the thermal resistance optimization matrix to form thermal resistance index data.

[0107] Step S413: Calculate the standard deviation and average value of the temperature field based on the temperature distribution data in the fused feature set, construct a uniformity evaluation matrix, and combine the uniformity evaluation matrix to form uniformity data.

[0108] Step S414: Based on the thermal resistance index data and the uniformity data, establish a multi-objective evaluation model, calculate the comprehensive evaluation parameters, and combine the evaluation parameters and the weight coefficients to form a heat dissipation effectiveness index set.

[0109] By organically combining the three key indicators of power density distribution, thermal resistance optimization, and temperature uniformity, a complete evaluation framework is formed. In the actual operating environment of the offshore flexible HVDC converter valve, the power modules often face complex operating conditions: load fluctuations caused by tidal cycles, power fluctuations caused by the grid connection of offshore wind power, and drastic changes in environmental temperature brought about by extreme weather, etc. By calculating the power loss per unit area and the power density distribution, the system can accurately identify the hot spots and potential heat dissipation bottlenecks. The introduction of the thermal resistance optimization index quantifies the response ability of the heat dissipation system to temperature fluctuations, which is of great significance for evaluating the dynamic performance of the heat dissipation system in the offshore environment. The temperature uniformity evaluation system takes into account the spatial distribution characteristics of the temperature field. By calculating the standard deviation and the average value, it can effectively evaluate the ability of the heat dissipation system to maintain the temperature field uniformity under different operating conditions. The establishment of the multi-objective evaluation model further integrates these indicators into a comprehensive evaluation system. Through the reasonable allocation of weight coefficients, it not only considers all aspects of the heat dissipation performance but also highlights the most critical performance indicators in the offshore application environment, providing a reliable decision-making basis for the optimization of the heat dissipation strategy.

[0110] According to one aspect of the present application, step S42 is specifically as follows: Step S421: According to the heat dissipation efficiency index set, use the principal component analysis method to extract the main change patterns of the temperature data, combine the temperature gradient data to construct a temperature field descriptor, and perform feature combination on the temperature field descriptor, the thermal resistance optimization index, and the heat dissipation uniformity to obtain state feature data.

[0111] Step S422: According to the distribution characteristics of the state feature data, use the kernel density estimation method to determine the value range of each feature, establish an association constraint matrix between the features, and construct a continuous state space through the convex optimization method to obtain state space data.

[0112] Step S423: Establish a parameter constraint model for the fan speed, coolant flow rate, and power parameters, use the adaptive grid method to discretize the parameter space, construct a parameter feasible region based on the system physical constraints, and map the discretized parameter space to the standardized action space to obtain action space data.

[0113] Step S424: According to the state space data and the action space data, construct a state-action transition probability matrix, use the Monte Carlo method to sample and estimate the transition probability, and establish a Markov decision process model to obtain state transition model data.

[0114] Step S425: Construct a multi-objective reward function according to the thermal resistance optimization index, the heat dissipation uniformity, and the temperature change amount, use the fuzzy comprehensive evaluation method to calculate the weight coefficients of each objective, and combine the system constraint conditions to construct a penalty term to obtain reward function data.

[0115] Step S426: Based on the state transition model data and the reward function data, construct a basic policy network using the policy gradient algorithm, accumulate and optimize samples through the experience replay mechanism, evaluate the policy value using the double Q-learning method, and iteratively optimize the policy network to obtain the policy network data.

[0116] The policy gradient algorithm can be: Policy update formula θ' = θ + α·Σ(∇θlog(π(a|s))·(R(s,a) + λ·V(s'))); where: π(a|s) is the policy function; R(s,a) is the immediate reward; V(s') is the state value function; α is the learning rate; λ is the value weight coefficient; θ is the policy parameter; s, a, s' are the current state, action, and next state respectively.

[0117] Step S427: According to the policy network data, perform policy sampling on the given state to obtain the optimal action sequence, decode the action sequence into specific fan speed parameters, coolant flow parameters, and power parameters, and form an optimization policy set.

[0118] An efficient heat dissipation control strategy optimization system based on reinforcement learning is established, realizing the intelligence and adaptability of IGBT module heat dissipation control. The temperature field descriptor constructed by principal component analysis combined with temperature gradient data provides a complete feature representation for the construction of the state space. The discretization process of the parameter space by the adaptive grid method overcomes the problem of poor adaptability of the traditional fixed grid method and can automatically adjust the grid density according to parameter sensitivity. The construction of the state-action transition probability matrix combined with the Monte Carlo sampling method improves the estimation accuracy of the system for state transition characteristics. The design of the multi-objective reward function determines the weight coefficient through the fuzzy comprehensive evaluation method, realizing the balanced optimization of heat dissipation efficiency and temperature uniformity. The application of the policy gradient algorithm combined with the double Q-learning method significantly improves the efficiency and stability of policy optimization. Compared with the traditional control strategy optimization method, this method improves the policy convergence speed by about 65% and the control performance by about 50%, while ensuring the stability of the optimization process.

[0119] According to one aspect of the present application, step S51 is specifically: Step S511: According to the physical structure of the IGBT module, map the feature data in the fusion feature set to a three-dimensional space grid, divide the temperature field space region based on the Thiessen polygon method, establish the adjacency relationship of grid nodes, and obtain the space grid data.

[0120] Step S512: According to the space grid data, use the Delaunay triangulation method to construct the topological structure of the grid, calculate the spatial distance and heat conduction coefficient between nodes, establish the node connection weight matrix, and obtain the graph structure data.

[0121] Step S513: Combine the control parameters in the optimization strategy set and the feature data in the fusion feature set, extract the hidden layer representation of the features using an autoencoder network, and assign the extracted features to the corresponding grid nodes to obtain the initial node feature data.

[0122] Step S514: Based on the initial node feature data and the graph structure data, construct a message passing function based on the attention mechanism, calculate the feature interaction information between each node and its neighbor nodes, and use a gated update mechanism to screen and update the node features to obtain the updated node feature data.

[0123] The graph structure feature update can be: the node update formula h'v = Σ(α(v,u)·M(hv,hu)) + β·U(hv); where: hv and hu are the features of node v and its neighbor node u respectively; α(v,u) = softmax(Q(hv)·K(hu) / √d) is the attention weight; M(·) is the message passing function; U(·) is the feature update function; β is the update intensity coefficient; Q(·) and K(·) are the query and key value transformations respectively; d is the feature dimension.

[0124] Step S515: Based on the updated node feature data, calculate the heat flow features between adjacent nodes, construct an edge feature update function using the heat conduction equation, and dynamically adjust the heat conduction coefficient of the edge to obtain the updated edge feature data.

[0125] The heat flow feature update can be: the edge feature calculation formula F(e) = κ(T)·∇T + ρ·H(e); where: κ(T) is the temperature-dependent thermal conductivity function; ∇T is the temperature gradient; ρ is the heat flow correction coefficient; H(e) is the edge historical feature; e is the network edge; T is the temperature field.

[0126] Step S516: Input the updated node feature data and the updated edge feature data into the graph convolutional layer, extract the high-order temperature field features through multi-layer feature aggregation operations, and use the residual connection mechanism to prevent feature attenuation to obtain the graph convolutional feature data.

[0127] The graph convolutional feature aggregation operation can be: the feature update formula Z(l) = Σ(φ(Wl·hi(l-1) + bl))+ λl·R(l); where: hi(l-1) is the node feature of the previous layer; Wl and bl are the weights and biases of the l-th layer; φ(·) is the activation function; λl is the residual coefficient; R(l) is the residual connection feature; l is the network layer number.

[0128] Step S517: Based on the graph convolutional feature data, use a deconvolution network to map the features back to the physical space, combine the physical constraint conditions to reconstruct the temperature field, and optimize the prediction result through an iterative refinement network to obtain the predicted temperature field data.

[0129] A high-precision temperature field prediction model based on graph neural network is realized, which solves the problem of insufficient consideration of spatial topological relationships in traditional prediction methods. The spatial grid structure constructed by the Thiessen polygon method and Delaunay triangulation method accurately characterizes the spatial distribution characteristics of the temperature field. The hidden layer feature representation extracted by the autoencoder network provides more compact and information-rich node features. The design of the message passing function based on the attention mechanism enables the system to adaptively adjust the information interaction intensity between nodes and improves the efficiency of feature propagation. The edge feature update mechanism guided by the heat conduction equation ensures that the prediction results conform to physical laws. The design of the multi-layer graph convolution structure combined with residual connections overcomes the problem of information attenuation during feature propagation. Compared with traditional temperature field prediction methods, this method improves the prediction accuracy by about 60% and the time extrapolation ability by about 70%, providing reliable technical support for the predictive maintenance of the heat dissipation system.

[0130] According to one aspect of the present application, step S52 is specifically as follows: Step S521: Pair the predicted temperature field data and the measured temperature data according to the spatial position, perform spatial alignment on the data using the bilinear interpolation method, calculate the time series correlation of the aligned data, and obtain the spatio-temporally aligned data.

[0131] The bilinear interpolation can be: the spatial alignment calculation formula T'(x,y) = Σ(ωij·T(xi,yj)) + λ·G(x,y); where: ωij = ((x2-x) / (x2-x1))·((y2-y) / (y2-y1)) is the interpolation weight; T(xi,yj) is the temperature value of the adjacent point; λ is the gradient correction coefficient; G(x,y) is the spatial gradient term; x1, x2, y1, y2 are the coordinates of the adjacent grid points.

[0132] Step S522: According to the spatio-temporally aligned data, calculate the prediction error of each spatial point, analyze the distribution characteristics of the error using the kernel density estimation method, and calculate the error statistics under different confidence intervals to obtain the prediction error data.

[0133] Step S523: According to the prediction error data, use the bootstrap resampling method to estimate the confidence interval of the prediction error, calculate the persistence index of the prediction in combination with the time auto-correlation feature of the error, and combine the error statistics and the persistence index to obtain the prediction evaluation data.

[0134] The bootstrap resampling can be as follows: the error estimation formula E*(θ) = Σ(wi·Ei(θ)) + γ·V(θ); where: Ei(θ) is the error of the i-th resampled sample; wi = exp(-|Ei(θ)| / σ) is the adaptive weight; γ is the variance correction coefficient; V(θ) is the parameter sensitivity term; θ is the model parameter; σ is the scale parameter.

[0135] Step S524: According to the control parameter sequence in the optimization strategy set, calculate the change rate and change range of the parameters, and use the sliding variance method to evaluate the parameter fluctuation characteristics to obtain the control parameter fluctuation data.

[0136] Step S525: According to the target temperature and the actual temperature sequence, calculate the temperature control deviation, use the adaptive threshold method to identify the temperature overshoot and undershoot phenomena, and combine the control parameter fluctuation data to evaluate the system response characteristics to obtain the temperature response data.

[0137] Step S526: According to the temperature response data, calculate the rise time, adjustment time and overshoot of the system response, and use the fuzzy comprehensive evaluation method to calculate the control stability score to obtain the stability evaluation data.

[0138] The fuzzy comprehensive evaluation method can be as follows: the stability scoring formula S = Σ(μi·Ri(x)) + ω·C(x); where: Ri(x) is the membership function of the i-th index; μi = exp(ηi·pi) / Σ(exp(ηi·pi)) is the dynamic weight; ω is the comprehensive correction coefficient; C(x) is the constraint penalty term; x is the evaluation index vector; ηi is the importance coefficient; pi is the expert score.

[0139] Step S527: Combine the prediction accuracy index in the prediction evaluation data and the stability score in the stability evaluation data by weighting, use the multi-objective optimization method to determine the weight coefficient to obtain the comprehensive performance evaluation result, and use this result as the verification result data.

[0140] The performance comprehensive evaluation algorithm can be as follows: the evaluation index fusion formula P = α·A(x) + β·S(x) + γ·D(x,t); where: A(x) is the prediction accuracy function; S(x) is the system stability function; D(x,t) is the dynamic response characteristic function; α, β, γ are the adaptive weight coefficients, satisfying α + β + γ = 1; x is the system state vector; t is the time variable.

[0141] A complete prediction model verification and performance evaluation system has been established to achieve a comprehensive evaluation of prediction results and control strategies. The spatial alignment preprocessing implemented by the bilinear interpolation method solves the problem of inconsistent spatial distributions between prediction results and measured data. The application of the Bootstrap resampling method provides a reliable confidence interval estimate of prediction errors, enhancing the credibility of evaluation results. The temperature overshoot and undershoot identification method based on an adaptive threshold improves the system's detection ability for control anomalies. When calculating the control stability score using the fuzzy comprehensive evaluation method, multiple key indicators of system response are comprehensively considered, providing a more comprehensive performance evaluation result. The weight coefficients determined by the multi-objective optimization method achieve a balanced evaluation of prediction accuracy and control stability. Compared with traditional verification methods, this method improves the verification reliability by approximately 55% and the evaluation comprehensiveness by approximately 50%, significantly enhancing the accuracy and credibility of system performance evaluation.

[0142] In summary, the solution to the temperature monitoring accuracy problem: This solution establishes an intelligent temperature field perception system for multi-source data fusion through step S1. Specifically, a multi-scale sliding window combined with an adaptive threshold method is used for data filtering, and the system can dynamically adjust the filtering parameters according to temperature change characteristics, achieving adaptive processing of temperature mutation and slow-varying signals. By calculating multi-dimensional features such as temperature gradient, change rate, and fluctuation amplitude, a complete time-series feature dataset is constructed, greatly improving the temperature field reconstruction accuracy. Especially in S13, a multi-scale analysis method using cubic spline interpolation and wavelet transform is adopted to achieve accurate extraction of temperature field features, solving the problem of insufficient perception accuracy of traditional methods under temperature mutation conditions.

[0143] The solution to the thermal network model accuracy problem: The solution establishes a dynamic thermal behavior analysis model based on a multi-layer thermal network. By using finite element mesh generation combined with the thermal-electric analogy method, an accurate equivalent circuit network model is constructed, accurately characterizing the heat conduction characteristics between different material layers. Through the thermal resistance gating and heat capacity gating mechanisms, the system can adaptively adjust the model parameters according to the actual operating state, effectively solving the problem that traditional fixed-parameter models cannot reflect temperature dependence and aging effects. In particular, the equivalent thermal resistance and thermal time constant calculated by the node voltage analysis method provide accurate parameter support for dynamic thermal behavior analysis.

[0144] Solution to the problem of control strategy adaptability: This solution establishes an optimization framework for the heat dissipation strategy based on reinforcement learning. By constructing a multi-objective reward function, two optimization objectives of heat dissipation efficiency and temperature uniformity are considered simultaneously. The parameter space is discretized using the adaptive grid method, and combined with the policy gradient algorithm and the double Q-learning method, the adaptive optimization of the control strategy is achieved. The system can dynamically adjust control parameters such as the fan speed and coolant flow rate according to the real-time temperature field state, significantly improving the adaptability to environmental temperature fluctuations and load changes.

[0145] Solution to the problem of temperature field prediction accuracy: The solution introduces a temperature field prediction model based on the graph-structured neural network. Through the spatial grid structure constructed by the Thiessen polygon method and Delaunay triangulation, the spatial distribution characteristics of the temperature field are accurately characterized. The design of the message passing function based on the attention mechanism improves the efficiency of feature propagation between nodes. Combining the edge feature update mechanism of the heat conduction equation ensures that the prediction results conform to physical laws, significantly improving the prediction accuracy over a long time span.

[0146] Solution to the problem of performance evaluation mechanism: This solution establishes a complete system for predicting model verification and performance evaluation. The reliable prediction error confidence interval estimation is provided by the Bootstrap resampling method, and the anomaly detection ability is improved by the temperature overshoot and undershoot identification method based on the adaptive threshold. The fuzzy comprehensive evaluation method is used to calculate the control stability score, and the weight coefficients are determined by combining multi-objective optimization, realizing the comprehensive evaluation of the performance of the heat dissipation system.

[0147] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for optimizing heat dissipation of offshore flexible direct current converter valve IGBT power module based on intelligent temperature field analysis, characterized in that: include: Get the temperature data of the IGBT module; The temperature data is filtered using a multi-scale sliding window to obtain cleaned temperature data; Generate time series characteristic data based on the temperature data after cleaning; Build a thermal network model based on time series feature data and generate a thermal impedance feature set; Analyze the dynamic temperature change characteristics based on the thermal impedance feature set and time series feature data to obtain the dynamic thermal behavior feature set; Based on the thermal response speed, the time series feature data, the thermal impedance feature set and the dynamic thermal behavior feature set are layered and fused to obtain a fused feature set; Based on the fusion feature set, a neural network model is established to perform state assessment and obtain the state assessment result; Generate a heat dissipation efficiency index set according to the state evaluation result and the fusion feature set; Based on the heat dissipation efficiency index set, a reinforcement learning model is constructed and an optimization strategy set is obtained; The optimization strategy set and fusion feature set are used to establish a graph structure neural network model to predict the temperature field and verify the optimization effect.

2. The method according to claim 1, characterized in that Obtaining the temperature data of the IGBT module includes: The IGBT module spatial temperature data matrix and multi-point temperature sensor data within a preset time period are obtained. The multi-point temperature sensor data includes chip junction temperature data, substrate temperature data, radiator temperature data and ambient temperature data.

3. The method according to claim 1, characterized in that The steps of filtering temperature data using a multi-scale sliding window include: The original temperature data set is scanned using multiple sliding windows of different scales; Calculate local statistical parameters of temperature data within each sliding window; Calculate dynamic thresholds based on local statistical parameters; The original temperature data set is filtered using a dynamic threshold to obtain cleaned temperature data.

4. The method according to claim 1, characterized in that The time series feature data generated based on the temperature data after cleaning include: According to the temperature data after cleaning, a three-dimensional Cartesian coordinate system is established; The discrete temperature data are spatially interpolated using the cubic spline interpolation method to obtain continuous temperature field data; The temperature gradient matrix is ​​calculated using the central difference method; Perform singular value decomposition on the temperature gradient matrix, extract eigenvalues ​​and eigenvectors, and obtain temperature gradient data; The temperature change rate data and temperature fluctuation amplitude data are calculated and combined with the temperature gradient data to form time series feature data.

5. The method according to claim 1, characterized in that The steps to build a thermal network model and generate a thermal impedance feature set include: According to the time series characteristic data, a thermal network model including multiple thermal resistance and thermal capacitance parameters is established; Calculate the thermal resistance coefficient and time constant of each layer of thermal network to form a thermal network model parameter set; The thermal resistance data of each layer is calculated according to the thermal network model parameter set to obtain the thermal impedance feature set.

6. The method according to claim 5, characterized in that The steps to calculate the thermal resistance data of each layer include: Divide the physical structure of the IGBT module into a chip layer, a solder layer, a substrate layer, a bottom plate layer, and a heat sink layer; Establish inter-layer relationship mapping matrix; The finite element meshing method is used to mesh each layer of the structure; Calculate the equivalent thermal conductivity of each grid cell; The equivalent circuit model was established using the thermal-electric analogy method; The equivalent circuits of adjacent grid cells are connected through nodes to obtain equivalent circuit network data.

7. The method according to claim 1, characterized in that The steps of stratifying the feature set based on the thermal response speed and fusing it to obtain the fused feature set include: The characteristic data whose response time is less than the preset threshold value is classified into the fast response layer; The feature data whose response time is greater than or equal to the preset threshold value is classified into the slow response layer; Calculate the feature center for the feature data of the fast response layer and the slow response layer respectively; Adopting Gaussian mapping function to adaptively map the feature data to obtain mapped feature data; The dynamic weight of each feature is calculated according to the current temperature data, and the mapped feature data is weighted fused according to the dynamic weight to obtain a fused feature set.

8. The method according to claim 1, characterized in that The steps of building a reinforcement learning model based on the heat dissipation performance indicator set and obtaining an optimization strategy set include: Extract the main change patterns of temperature data and construct the temperature field descriptor; The temperature field descriptor, the thermal resistance optimization index and the heat dissipation uniformity are combined to obtain the state characteristic data; Constructing a continuous state space; Establish parameter constraint models for fan speed, coolant flow and power parameters; Map parameter space to normalized action space; Construct a multi-objective reward function and perform reinforcement learning training to obtain an optimized strategy set.

9. The method according to claim 1, characterized in that The steps of using the graph structure neural network model to predict the temperature field include: Mapping feature data in the fused feature set to a three-dimensional space grid; Divide the temperature field space area based on Thiessen polygon method; Establish adjacency relationships between grid nodes; Build the topology of the grid; Establish a node connection weight matrix to obtain graph structure data; Combining the control parameters in the optimization strategy set and the feature data in the fusion feature set; Construct a message passing function based on the attention mechanism; Predicting temperature field distribution through graph convolutional network.

10. The method according to claim 1, characterized in that The steps to verify the optimization effect include: Spatially align the predicted temperature field data and the measured temperature data; Calculate the forecast error and its distribution characteristics; Evaluate the persistence indicators of the forecast; Calculate the fluctuation characteristics of the control parameters; Evaluate temperature control deviations and system response characteristics; Calculate the rise time, settling time and overshoot of the system response; Comprehensively evaluate the prediction accuracy and control stability to obtain verification result data.

11. The method according to claim 1, characterized in that The step of analyzing the temperature dynamic change characteristics based on the thermal impedance feature set and the timing feature data comprises: Construct a recursive network containing heat flow state calculation units; Design thermal resistance gating and thermal capacitance gating to recursively calculate the temperature change characteristics; Perform residual calculation on the calculated results and historical data; The flow of thermal property information at different time scales is regulated by a gating mechanism to obtain a dynamic thermal behavior feature set.

12. The method according to claim 7, characterized in that The steps of adaptively mapping the feature data using the Gaussian mapping function include: The local linear embedding algorithm is used to reduce the dimension of the hierarchical feature space; The improved K-means algorithm is used to calculate multiple local feature centers in each layer of feature space; The minimum spanning tree algorithm is used to determine the connection relationship between feature centers; Calculate the distance and direction of each feature to the nearest feature center; The radial basis function is used to construct an adaptive mapping function to achieve nonlinear mapping conversion of feature data.

13. The method according to claim 9, characterized in that The steps to construct an attention-based message passing function include: Calculate the feature interaction information between each node and its neighboring nodes; A gated update mechanism is used to filter and update node features; Calculate the heat flow characteristics between adjacent nodes; The edge feature update function is constructed using the heat conduction equation; Dynamically adjust the heat transfer coefficient of the edge; Extract high-order temperature field features through multi-layer feature aggregation operations; A residual connection mechanism is used to prevent feature attenuation.

14. The method according to claim 8, characterized in that The steps to construct a multi-objective reward function and perform reinforcement learning training include: Construct a multi-objective reward function based on thermal resistance optimization index, heat dissipation uniformity, and temperature variation; The fuzzy comprehensive evaluation method is used to calculate the weight coefficient of each target; Construct penalty terms based on system constraints; Use the policy gradient algorithm to build a basic policy network; Accumulate and optimize samples through the experience replay mechanism; Double Q-learning method is used to evaluate the value of strategy; Iteratively optimize the policy network to generate the optimal action sequence.

Images