SiC MOSFET array-oriented equalization control method and system, electronic equipment and medium

Abstract

The invention provides a balance control method and system for a SiC MOSFET array, electronic equipment and a medium, and relates to the technical field of power electronics. Temperature data of all devices are collected in real time through a thin film thermocouple array on a SiC MOSFET array substrate; dynamically matching the temperature data with a database for storing thermal characteristic curves of similar devices to obtain associated data of temperature and current; generating a temperature characteristic graph based on the temperature data, inputting the graph into a convolutional neural network, and outputting a risk index; and finally, dynamically adjusting the duty ratio of the gate driving signal by adopting reinforcement learning based on the associated data, the risk index and the current deviation between the adjacent devices through the intelligent agent corresponding to each device so as to realize the equalization control of the array. According to the invention, the equalization control of the SiC MOSFET array can be accurately realized.

Description

Technical Field

[0001] This application relates to the field of power electronics technology, and in particular to a method, system, electronic device and dielectric for equalization control of SiC MOSFET arrays. Background Technology

[0002] In high-power power electronic devices such as inverters and industrial converters, SiC MOSFET arrays are widely used due to their high-frequency and high-efficiency characteristics. However, uneven temperature distribution and current stress differences among the devices in the array can lead to local overheating or current imbalance, and accelerate device aging or even cause thermal runaway. Therefore, precise equalization control is required to achieve coordinated stability of temperature and current of each device to meet the technical requirements of high reliability and long life of the equipment.

[0003] Currently, the mainstream solution for balanced control of SiC MOSFET arrays adopts a centralized controller architecture. Specifically, the temperature and current data of each device are collected by sensors, and the central controller calculates the adjustment amount based on the preset temperature and current threshold model. Then, it uniformly issues the adjustment command of the duty cycle of the gate drive signal to balance the operating status of each device in the array.

[0004] However, this centralized approach has the following drawbacks: the central controller needs to process massive amounts of distributed data, which can lead to delays in adjustment due to data transmission and centralized computation, making it difficult to adapt to the dynamic changes of components; furthermore, the adjustment strategy based on a single threshold model may amplify current deviations due to local adjustment conflicts, and lacks a hierarchical assessment of thermal runaway risk. Relying solely on threshold judgment can easily lead to misjudgment of risk levels, resulting in insufficient control accuracy. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, electronic device, and medium for equalization control of SiC MOSFET arrays, in order to solve the problem of poor equalization control effect of SiC MOSFET arrays in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for equalization control of SiC MOSFET arrays, comprising: Temperature data of each device in the SiC MOSFET array is collected in real time using a thin-film thermocouple array arranged on the substrate of the SiC MOSFET array. The temperature data is dynamically matched with a database containing thermal characteristic curves of similar devices to obtain the correlation data between the device's temperature and current. Based on the temperature data of each device, a three-dimensional temperature field corresponding to the SiC MOSFET array is constructed, and a temperature feature map corresponding to the three-dimensional temperature field is generated. The temperature feature map is input into a convolutional neural network. Through multi-layer convolution and pooling operations in the convolutional neural network, key information in the temperature feature map is extracted and hierarchical operations are performed to output a risk index. By using the intelligent agents corresponding to each device in the SiC MOSFET array, based on the correlation data between temperature and current, the risk index, and the current deviation between adjacent devices, reinforcement learning technology is used to dynamically adjust the duty cycle of the gate drive signal of the corresponding device to achieve balanced control of the SiC MOSFET array.

[0007] Optionally, the step of extracting key information from the temperature feature map and performing hierarchical operations through multiple convolution and pooling operations in the convolutional neural network to output a risk index includes: The temperature feature map is convolved by the first convolutional layer in the convolutional neural network to extract a first feature map that represents the temperature value and temperature change gradient of the local area. The first feature map is then dimensionality-reduced and filtered by the first pooling layer to retain local features that meet the first preset condition, and a second feature map is output. The second feature map is convolved by the second convolutional layer in the convolutional neural network. Correlation analysis is performed on the local features of adjacent spatial regions to extract a third feature map that represents the temperature linkage relationship between regions. The third feature map is then reduced in dimensionality by the second pooling layer to filter out the target linkage features that represent the temperature field regional correlation pattern and output a fourth feature map. The fourth feature map is convolved by the third convolutional layer in the convolutional neural network. The target linkage features between spatially non-directly adjacent regions are abstracted and fused in a higher order to extract deep feature maps. The deep feature maps are then reduced in dimensionality by the third pooling layer to retain deep features that meet the second preset condition and output the fifth feature map. The risk index is obtained by weighted summation of the deep features in the fifth feature map through the output layer of the convolutional neural network.

[0008] Optionally, the step of performing a convolution operation on the fourth feature map through the third convolutional layer in the convolutional neural network to perform high-order abstraction and fusion of target linkage features between spatially non-directly adjacent regions and extract a deep feature map includes: The third convolutional layer uses a convolutional kernel of a preset size to set a corresponding receptive field on the fourth feature map. Each receptive field covers at least two image regions that are not directly adjacent in spatial location. Within each receptive field, feature combination and high-order abstraction are performed on multiple target linkage features to mine deep correlation patterns in thermal state changes of different regions and generate high-order feature maps corresponding to each receptive field. The high-order feature maps corresponding to multiple receptive fields are globally integrated and spliced ​​together to form a unified global feature representation; The global feature representation is aggregated along its feature dimensions to form a deep feature map that characterizes the intrinsic correlations of the global temperature field.

[0009] Secondly, this application provides a balanced control system for SiC MOSFET arrays, including: The acquisition module is used to acquire the temperature data of each device in the SiC MOSFET array in real time using a thin-film thermocouple array arranged on the substrate of the SiC MOSFET array. The matching module is used to dynamically match the temperature data with a database that stores the thermal characteristic curves of corresponding similar devices to obtain the correlation data between the device's temperature and current. The construction module is used to construct a three-dimensional temperature field corresponding to the SiC MOSFET array based on the temperature data of each device, and generate a temperature feature map corresponding to the three-dimensional temperature field. The extraction module is used to input the temperature feature map into a convolutional neural network, extract key information from the temperature feature map through multi-layer convolution and pooling operations in the convolutional neural network, perform hierarchical operations, and output a risk index. The adjustment module is used to dynamically adjust the duty cycle of the gate drive signal of the corresponding device by using reinforcement learning technology based on the correlation data between temperature and current, the risk index and the current deviation between adjacent devices, through the intelligent agents corresponding to each device in the SiC MOSFET array, so as to achieve balanced control of the SiC MOSFET array.

[0010] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, configured to execute the computer program, implement the steps of an equalization control method for a SiCMOSFET array as described in the first aspect above.

[0011] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the equalization control method for a SiC MOSFET array as described in the first aspect above.

[0012] The equalization control method for SiC MOSFET arrays provided in this application has the following advantages: This application utilizes a thin-film thermocouple array to accurately and in real-time acquire temperature data for each device in a SiC MOSFET array. The temperature data is then dynamically matched with a thermal characteristic curve database to obtain temperature-current correlation data, which intuitively reflects the relationship between device temperature changes and current load. Next, a three-dimensional temperature field is constructed based on the temperature data, generating a temperature feature map. This transforms discrete temperature data into a structured and visualized temperature distribution representation, thus fully presenting the overall temperature spatial distribution characteristics of the array. Subsequently, multi-layer convolution and pooling operations are performed using a convolutional neural network to output a risk index, enabling in-depth mining of key information from the temperature feature map. This allows for a quantitative assessment of thermal runaway risk and provides precise indicators for risk warning. Finally, distributed dynamic control is achieved by adjusting the gate drive signal duty cycle based on the correlation data, risk index, and current deviation between adjacent devices using the corresponding intelligent agents for each device. This effectively balances the temperature and current states of each device, preventing local imbalances or thermal runaway and ensuring the overall balanced operation of the array.

[0013] Furthermore, this application extracts local temperature values ​​and temperature change gradients through the first convolutional layer and the first pooling layer of a convolutional neural network, screening for significant local features; it then performs correlation analysis on local features of adjacent spatial regions through the second convolutional layer and the second pooling layer to extract temperature linkage relationships between regions; finally, it sets up multi-scale receptive fields through the third convolutional layer to perform high-order abstraction and fusion of target linkage features in non-directly adjacent spatial regions, followed by global integration and feature aggregation to uncover the intrinsic correlation of the global temperature field; and finally, it outputs a risk index by weighted summation of deep features. This method achieves hierarchical thermal risk quantification from local to global and from appearance to essence, providing a precise basis for equilibrium control.

[0014] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an equalization control method for SiC MOSFET arrays provided in this application embodiment; Figure 2 Provided for the embodiments of this application Figure 1 Flowchart of S104 in China; Figure 3A schematic diagram of a specific implementation of a balanced control system for SiC MOSFET arrays provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] Existing equalization control schemes for SiC MOSFET arrays have the following drawbacks in practical applications: the central controller needs to process massive amounts of distributed data, which can easily lead to adjustment lag due to data transmission and centralized computing delays, making it difficult to adapt to the dynamic changes of the devices; furthermore, adjustment strategies based on a single threshold model may amplify current deviations due to local adjustment conflicts, and lack hierarchical assessment of thermal runaway risks. Relying solely on threshold judgments can easily lead to misjudgment of risk levels, resulting in insufficient control accuracy.

[0018] To address the aforementioned issues, this application proposes a balanced control method for SiC MOSFET arrays. The core idea of ​​this method is as follows: First, a dedicated temperature sensor collects the temperature of each device in real time and, combined with the device's own thermal characteristic data, establishes a correlation between temperature and current. Then, the temperature distribution is transformed into an intuitive temperature feature map, and a convolutional neural network capable of deep analysis of temperature distribution is used to extract key information, thereby accurately assessing the risk of thermal runaway. Finally, an intelligent adjustment unit corresponding to each device dynamically adjusts its own operating parameters based on the temperature-current correlation data, risk assessment results, and current deviations of adjacent devices. This approach avoids the latency problem of centralized processing through distributed adjustment, captures the mutual influence between devices through deep analysis of temperature distribution, and reduces misjudgments through refined risk assessment, thereby achieving more accurate and timely balanced control of the SiC MOSFET array.

[0019] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The core of this application is to provide a balanced control method for SiC MOSFET arrays, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Using a thin-film thermocouple array arranged on the substrate of the SiC MOSFET array, the temperature data of each device in the SiC MOSFET array is collected in real time.

[0021] Among them, SiC MOSFET array refers to a collection of devices composed of multiple silicon carbide metal-oxide-semiconductor field-effect transistors. This array is used to realize power conversion of high-power power electronic devices. The substrate is the basic component that supports the array and can provide mechanical support and heat dissipation path for the devices. A thin-film thermocouple array refers to a miniature temperature sensing device deployed on a substrate. This array can consist of multiple pairs of thermocouples and can simultaneously detect the temperature at multiple points. Temperature data refers to the numerical information collected by the sensor that reflects the real-time temperature of each SiC MOSFET device.

[0022] In step S101, firstly, a thin-film thermocouple array is arranged at a preset position on the substrate surface of the SiC MOSFET array to ensure that each device corresponds to at least one thermocouple, so that the sensor can directly sense the temperature change of the device.

[0023] Secondly, the thin-film thermocouple array is connected to the data acquisition circuit. When the device is running, the thermocouples in the thin-film thermocouple array generate thermoelectric potential signals due to temperature differences, and transmit them to the data acquisition circuit. The data acquisition circuit converts the thermoelectric potential signals into corresponding digital temperature values.

[0024] Finally, the data acquisition circuit continuously acquires the temperature data of all thermocouples at a preset frequency and transmits this data to the subsequent processing unit in real time, thereby forming a dynamic temperature record for each device.

[0025] S102. Dynamically match the temperature data with a database that stores thermal characteristic curves of similar devices to obtain the correlation data between the device's temperature and current.

[0026] Among them, "same type devices" refers to SiC MOSFETs with matching models and specifications; "database" refers to a collection that stores the thermal characteristic curves of various devices. The thermal characteristic curve contains temperature records of the corresponding device under different currents. The temperature records are arranged in ascending order of current and include multiple temperature ranges. Each temperature range corresponds to a unique current range on the thermal characteristic curve. The correlation data between temperature and current refers to the mapping relationship between device temperature and the corresponding current range.

[0027] In one specific implementation, step S102 includes the following steps: Step 1021: For each device in the SiC MOSFET array, select thermal characteristic curves that match the model and specifications of the device from the database.

[0028] Specifications may include voltage rating, current capacity, and package type.

[0029] In step 1021, for each device in the SiC MOSFET array, search keywords are first constructed based on the core parameters such as the device model and specifications. Then, the thermal characteristic curves corresponding to the keywords are selected from the database through a precise matching algorithm. The thermal characteristic curves are generated by multiple experiments on similar devices under a standardized test environment. The curves are plotted with current on the horizontal axis and temperature on the vertical axis, and record the complete temperature change of the device from startup to rated load in order of increasing current. The thermal characteristic curves are then divided into multiple continuous temperature ranges, and each temperature range corresponds to a unique current range through the monotonicity of the thermal characteristic curve, so as to ensure that the mapping relationship between temperature and current is deterministic.

[0030] For example, in the SiC MOSFET array of an industrial converter, for the target model and known specifications of device C, thermal characteristic curves of the same model and specifications are selected from the database. Based on the test data of the device under the reference ambient temperature of 20℃ and standard heat dissipation conditions, the thermal characteristic curve is calibrated as the temperature corresponding to the current range of 25A-35A is 110℃-130℃, and the temperature corresponding to the current range of 35A-45A is 130℃-150℃.

[0031] Step 1022: Based on the temperature data of the device, perform interval positioning on the corresponding thermal characteristic curve of the device to determine the corresponding current range.

[0032] In step 1022, by comparing the real-time temperature data with the upper and lower limits of each temperature range of the thermal characteristic curve, the target range in which the temperature falls is determined. Then, based on the pre-set correspondence between the temperature range and the current range of the thermal characteristic curve, the current range bound to the temperature range is extracted.

[0033] For example, if the temperature of device C is 125℃ at a certain moment, and its temperature is located in the range of 110℃-130℃ on the thermal characteristic curve, then the corresponding current range of 25A-35A can be obtained.

[0034] Step 1023: Combine the operating environment parameters of the device to dynamically adjust the current range so that the adjusted current range and the temperature data of the device form a unique correspondence, thus forming the correlation data between the temperature and current of the device.

[0035] The operating environment parameters may include the heat dissipation conditions of the device, the ambient temperature, the cumulative operating time, etc. The heat dissipation conditions may include heat dissipation efficiency and a corresponding heat dissipation correction factor; the ambient temperature corresponds to a temperature correction factor, and the cumulative operating time corresponds to an aging correction factor.

[0036] In step 1023, for each environmental parameter, a corresponding correction coefficient can be calculated based on a preset correction model. Then, all correction coefficients are merged to obtain a comprehensive correction coefficient. Next, the upper and lower limits of the range are multiplied by the comprehensive correction coefficient to obtain the adjusted current range, thereby determining the temperature-current correlation data suitable for the current operating conditions. This correlation data indicates that, under the current environmental conditions, the current required for the device to reach the temperature indicated by the temperature data should be within the actual current range.

[0037] It should be noted that this embodiment does not limit the mathematical expressions of the correction models corresponding to each operating environment parameter or the fusion method of multiple correction coefficients.

[0038] This application achieves adaptive calibration of temperature and current correlation data under different operating conditions by dynamically matching real-time temperature data with the thermal characteristic curves of similar devices and calculating a comprehensive correction coefficient based on operating environment parameters, thereby adjusting the current range. This provides a precise and reliable mapping basis for subsequent equalization control.

[0039] S103. Based on the temperature data of each device, construct a three-dimensional temperature field corresponding to the SiC MOSFET array, and generate a temperature feature map corresponding to the three-dimensional temperature field.

[0040] Among them, the temperature feature map refers to a structured image transformed from a three-dimensional temperature field, used to intuitively present the spatial distribution characteristics of temperature.

[0041] In one specific implementation, step S103 involves constructing a three-dimensional temperature field corresponding to the SiC MOSFET array based on the temperature data of each device, including the following steps: Step 1031: Determine the three-dimensional spatial coordinates of each device in the SiC MOSFET array, and associate the temperature data of each device with the three-dimensional spatial coordinates to form a discrete dataset.

[0042] Among them, the three-dimensional spatial coordinates refer to the three-dimensional position information of each device on the substrate of the SiC MOSFET array, and the discrete dataset refers to the data set formed by associating the three-dimensional spatial coordinates of each device with its corresponding temperature data.

[0043] In step 1031, the three-dimensional spatial coordinates of each device in the SiC MOSFET array on the substrate are first determined, and a three-dimensional coordinate system is established with the lower left corner of the substrate as the origin. The x-axis extends laterally along the substrate, the y-axis extends longitudinally along the substrate, and the z-axis is perpendicular to the substrate surface. That is, the midpoint of the contact surface between the device and the substrate is taken as the z-axis coordinate point. Then, a laser positioning instrument is used to accurately measure the coordinate values ​​of each device, for example, with an accuracy of 0.1 mm. The real-time temperature data of each device is then associated with the corresponding three-dimensional spatial coordinates through a known data binding algorithm to form a discrete dataset containing three-dimensional spatial coordinates and temperature data. Each data point is accompanied by a collection timestamp to ensure the consistency of spatiotemporal information.

[0044] Step 1032: Based on the discrete dataset, construct an initial three-dimensional model with the physical boundary of the substrate as the modeling range.

[0045] The initial three-dimensional model refers to a three-dimensional structure built based on a discrete dataset and containing temperature information of the device's location.

[0046] In step 1032, the discrete dataset is spatially reconstructed using existing 3D modeling tools. Specifically, the coordinate points in the discrete dataset are mapped to a 3D mesh using the actual size of the substrate as the boundary, with each coordinate point corresponding to a node in the mesh. Temperature data is stored as a node attribute. During the modeling process, the physical contour information of the device is preserved, and the device position is represented by a cube module in the model. The vertex coordinates of the module are aligned with the 3D spatial coordinates of the device to form an initial 3D model. Areas in the model where no device is placed are represented by blank meshes.

[0047] Step 1033: For the blank areas in the initial three-dimensional model where no devices are arranged, calculate the temperature value of the blank areas based on the temperature data of the existing devices adjacent to the blank areas and the spatial distance between the blank areas and the existing devices, so as to fill the blank areas in the initial three-dimensional model and form a three-dimensional temperature field.

[0048] The blank area refers to the space area in the initial three-dimensional model where no devices are placed, and the three-dimensional temperature field refers to the complete three-dimensional model containing the temperature information of all space areas of the array after filling in the temperature of the blank area.

[0049] In step 1033, for blank areas in the initial 3D model, such as gaps between devices, substrate edge areas, etc., an interpolation neighborhood radius is set, and for each blank grid, all nodes of existing devices within the radius are searched as reference points; then, the spatial distance between the reference points and the blank grid is calculated using a known type of 3D distance formula, and weights are assigned, with the weight increasing as the distance increases. Then, the weighted summation formula was used. Calculate the temperature values ​​for the blank grid, where, This represents the temperature value of the blank area, and n represents the number of adjacent devices involved in the calculation. This represents the temperature data of the i-th adjacent device. The value of the i-th adjacent device is not the focus of this application, so relevant technologies can be consulted and will not be elaborated here. Then, the calculation results of all blank areas are filled into the model to form a continuous three-dimensional temperature field covering the entire space of the array. The three-dimensional temperature field is then transformed into a temperature feature map through a visualization algorithm, and the temperature values ​​are mapped with different colors. The pixel size of the feature map is consistent with the grid precision of the three-dimensional model, which can intuitively present the spatial distribution differences of temperature.

[0050] This application can transform discrete temperature data into structured information that reflects the overall spatial distribution of temperature in the array, fully presenting the temperature change trend in space. This provides comprehensive and intuitive input data for subsequent convolutional neural network extraction of key information, ensuring the accuracy of deep feature extraction and thus guaranteeing the accuracy and reliability of thermal risk quantification assessment.

[0051] S104. Input the temperature feature map into a convolutional neural network. Through multi-layer convolution and pooling operations in the convolutional neural network, extract key information from the temperature feature map and perform hierarchical operations to output a risk index.

[0052] Convolutional neural networks (CNNs) are deep learning models that include multi-layered processing structures. They extract data features through convolution operations and filter key features through pooling operations. CNNs are pre-trained networks, and this embodiment does not limit their structural design, parameter design, or training process. They can be set according to actual conditions.

[0053] The key information consists of quantitative indicators that characterize the essential thermal state of the SiC MOSFET array, which are abstracted layer by layer from the temperature characteristic map. These indicators include local features, target linkage features, and deep features. The risk index refers to the risk index corresponding to the entire SiC MOSFET array.

[0054] In one specific implementation, such as Figure 2 As shown, step S104 includes the following steps: Step 1041: Perform a convolution operation on the temperature feature map through the first convolutional layer in the convolutional neural network to extract a first feature map that represents the temperature value and temperature change gradient of the local area. Perform dimensionality reduction and filtering on the first feature map through the first pooling layer to retain local features that meet the first preset conditions and output a second feature map.

[0055] In the temperature feature map, each pixel corresponds to a spatial region on the substrate, and the pixel value represents the temperature value of that region. The first convolutional layer refers to the convolutional layer in the convolutional neural network used for preliminary feature extraction. It uses a convolutional kernel of a preset size to slide across the temperature feature map to extract local information. The first feature map refers to the feature map output after processing by the first convolutional layer, where each channel corresponds to a type of local feature. The first pooling layer refers to the pooling layer used to perform dimensionality reduction on the first feature map, which reduces the amount of data and retains significant features through downsampling; the first preset condition refers to the quantification standard used to filter local features, for example, retaining local features whose influence on the thermal state is greater than a preset threshold. Local features refer to quantitative information extracted from temperature feature maps that characterizes the temperature values ​​of a single device or a very small neighborhood and the degree of drastic temperature changes in the surrounding area.

[0056] In step 1041, the temperature feature map generated in step S103 is first input into the first convolutional layer. The first convolutional layer has multiple convolutional kernels of preset size. Each convolutional kernel performs a sliding convolution operation on the temperature feature map with a preset stride. For each sliding position, the convolutional kernel is multiplied element-wise with the local region it covers and summed to generate the feature value at that position. Then, through the parallel operation of multiple convolutional kernels, different types of local features are extracted. These local features include the temperature numerical features of each region and the temperature gradient features in the horizontal, vertical, and diagonal directions. The outputs of all convolutional kernels are stacked in the channel dimension to form the first feature map.

[0057] Then, the first feature map is input into the first pooling layer. The first pooling layer uses a pooling window of a preset size and slides on the first feature map with a preset step size. Within each pooling window, significant feature values ​​are extracted according to a preset pooling method, such as max pooling or average pooling. Through the pooling operation, on the one hand, the size of the feature map is reduced, thus reducing the amount of subsequent computation; on the other hand, local features with high response intensity and significant impact on local thermal state are selected. The feature map output by the pooling layer is the second feature map.

[0058] It should be noted that the embodiments of this application do not impose specific limitations on the structural design and parameter design of the first convolutional layer and the first pooling layer.

[0059] For example, in the equalization control process of the SiC MOSFET array in an industrial converter, step S103 generates a temperature feature map with a resolution of 32×32 pixels, where each pixel corresponds to a 1mm×1mm area on the substrate; the first convolutional layer sets 8 convolutional kernels of size 3×3 with a stride of 1 and a padding of 1 to extract a first feature map of size 32×32×8; the first pooling layer uses 2×2 max pooling with a stride of 2 to perform dimensionality reduction filtering on the first feature map and output a second feature map of size 16×16×8; in this example, the first preset condition is achieved through max pooling, that is, only the feature value with the highest response intensity is retained in each 2×2 local region. For example, max pooling retains the maximum value 128, retains 128 in the 2×2 local region [120, 125, 118, 128], and discards 120, 125, and 118, that is, it retains the point with the highest temperature in the region and its feature information, thereby filtering out the most significant local features in each local region.

[0060] Step 1042: Perform convolution operation on the second feature map through the second convolutional layer in the convolutional neural network, perform correlation analysis on the local features of adjacent spatial regions, extract the third feature map to characterize the temperature linkage relationship between regions, perform dimensionality reduction operation on the third feature map through the second pooling layer, filter out the target linkage features to characterize the temperature field regional correlation pattern, and output the fourth feature map.

[0061] The second convolutional layer refers to the convolutional layer in the convolutional neural network used to perform spatial correlation analysis on local features. By increasing the size of the convolutional kernel or the dilation rate, the receptive field is expanded so that each convolutional operation can cover multiple adjacent spatial regions, thereby capturing the feature correlation between regions. The third feature map refers to the feature map output after processing by the second convolutional layer, where each channel corresponds to a type of temperature linkage relationship between regions. The second pooling layer is used to perform dimensionality reduction on the third feature map. It reduces the amount of data by downsampling and filters out stable regional association patterns. The target linkage feature refers to the quantitative feature selected from the third feature map that can stably characterize the temperature change correlation between adjacent regions, and is used to reflect the synergy or difference of temperature distribution on the heat conduction path; the fourth feature map refers to the feature map output after processing by the second pooling layer that contains the target linkage feature.

[0062] In step 1042, the second feature map output from step 1041 is first input into the second convolutional layer. The second convolutional layer contains multiple convolutional kernels of preset sizes, each larger than the kernels in the first convolutional layer (e.g., 5×5 or 7×7), or using dilated convolution to expand the receptive field. The convolutional kernels in the second convolutional layer perform sliding convolution operations on the second feature map with a preset stride. Each kernel simultaneously covers multiple adjacent spatial regions. Through convolution operations, the local features of these adjacent regions are weighted and combined, and correlation analysis is performed to uncover the linkage relationship of temperature changes between regions. This linkage relationship includes the synchronicity of temperature changes in adjacent regions, the proportional relationship of change amplitude, and the similarity of change trends. Multiple convolutional kernels operate in parallel, each extracting one type of inter-regional temperature linkage relationship. The outputs of all convolutional kernels are stacked along the channel dimension to form the third feature map.

[0063] Then, the third feature map is input into the second pooling layer. The second pooling layer uses a pooling window of a preset size and slides across the third feature map with a preset step size. Within each pooling window, significant feature values ​​are extracted according to a preset pooling method. The pooling operation reduces the size of the feature map, thus reducing subsequent computation; it also filters out target linkage features that are spatially representative and can stably reflect the correlation patterns of the temperature field regions, discarding occasional or unstable correlation patterns. The feature map output by the pooling layer is the fourth feature map, which contains regional linkage features that have a significant impact on the overall thermal state.

[0064] For example, following the example of step 1041, step 1041 outputs a second feature map with a size of 16×16×8. The second convolutional layer sets 16 convolutional kernels with a size of 5×5, a stride of 1, and padding of 2, and extracts a third feature map with a size of 16×16×16. Each channel represents different types of temperature linkage relationships between regions, such as temperature synchronicity between adjacent regions, thermal diffusion effect, and thermal aggregation effect.

[0065] The second pooling layer uses 3×3 max pooling with a stride of 2 to perform dimensionality reduction filtering on the third feature map, outputting a fourth feature map with a size of 8×8×16. The pooling operation retains the strongest linkage features within each 3×3 region, thus filtering out regional correlation patterns that can exist stably over a large range, forming the target linkage feature. This target linkage feature provides the foundation for subsequent steps to extract the intrinsic correlation of the global temperature field.

[0066] Step 1043: Perform convolution operation on the fourth feature map through the third convolutional layer in the convolutional neural network, perform high-order abstraction and fusion on the target linkage features between spatially non-directly adjacent regions, extract deep feature maps, perform dimensionality reduction and screening on the deep feature maps through the third pooling layer, retain deep features that meet the second preset conditions, and output the fifth feature map.

[0067] The third convolutional layer refers to the convolutional layer in the convolutional neural network used for high-order abstraction and global fusion of target linkage features. By setting convolutional kernels of different sizes or dilated convolutions of different dilation rates, a receptive field that can cover spatially non-directly adjacent regions can be formed. Deep feature maps refer to feature maps that represent the intrinsic correlations of the global temperature field after being abstracted by the third convolutional layer. The third pooling layer is a pooling layer used to perform dimensionality reduction on deep feature maps. It filters out deep features that play a dominant role in the overall thermal state through downsampling. The second preset condition refers to the quantitative standard used to screen deep features; the fifth feature map refers to the feature map output after processing by the third pooling layer, which contains deep dominant features.

[0068] In step 1043, the fourth feature map is convolved by the third convolutional layer in the convolutional neural network to perform high-order abstraction and fusion of target linkage features between spatially non-directly adjacent regions, and to extract deep feature maps, including the following steps: Step a1: Using a convolution kernel of a preset size in the third convolutional layer, set a corresponding receptive field on the fourth feature map. Each receptive field covers at least two image regions that are not directly adjacent in spatial location.

[0069] The receptive field refers to the area that a single neuron in a convolutional layer can cover on the input feature map. In this step, it can refer to the convolutional range that can simultaneously cover at least two spatially non-directly adjacent image regions. Non-directly adjacent image regions refer to two or more regions that are not adjacent in spatial location and are separated by other regions.

[0070] In step a1, a large-size convolutional kernel or dilated convolution is used to expand the area covered by each convolutional operation to simultaneously include at least two spatially non-directly adjacent image regions. The convolutional kernel slides on the fourth feature map with a preset stride, and each sliding position corresponds to a receptive field. Each receptive field covers multiple spatial positions, which may have intervals between them, thereby achieving synchronous coverage of spatially separated regions.

[0071] For example, following the example of step 1042, the fourth feature map has a size of 8×8×16, with each spatial location corresponding to an area of ​​approximately 4mm×4mm on the substrate. The third convolutional layer uses a 7×7 convolutional kernel with a stride of 1 and a padding of 3, so that each receptive field covers 7×7 spatial locations, corresponding to an area of ​​approximately 28mm×28mm on the substrate. Since the total size of the fourth feature map is 8×8, the 7×7 receptive field almost covers the entire temperature field, and can simultaneously cover multiple spatially non-directly adjacent areas, such as simultaneously covering the upper left corner area and the lower right corner area, which are spatially separated by approximately 28mm.

[0072] Alternatively, a 3×3 dilated convolution with an expansion rate of 3 can also be used to achieve a 7×7 receptive field coverage.

[0073] Step a2: Within each receptive field, feature combination and high-order abstraction are performed on multiple target linkage features to mine deep correlation patterns in thermal state changes of different regions and generate high-order feature maps corresponding to each receptive field.

[0074] Among them, feature combination and higher-order abstraction refer to the nonlinear combination and transformation of the linked features of multiple targets within the same receptive field to extract higher-level and more abstract feature representations. Deep correlation patterns refer to the intrinsic connections between spatially separated regions in terms of thermal state changes, such as long-distance thermal coupling effects, heat propagation paths, and the symmetry or periodicity of temperature fields.

[0075] In step a2, the target linkage features at different spatial locations within the receptive field are weighted and nonlinearly transformed using the weight parameters of the convolution kernel to extract higher-order statistical relationships between these features. These higher-order relationships reflect the inherent connections in thermal state changes between spatially separated regions, such as the synchronicity of temperature changes in distant regions, the regularity of heat transfer along specific paths, and the symmetry or periodicity of temperature field distribution. Each receptive field generates a feature value or feature vector after convolution. The convolution results of all receptive fields are arranged according to spatial location to form a higher-order feature map with the same spatial size as the fourth feature map.

[0076] For example, within the 7×7 receptive field set in step a1, multiple target linkage features from position (1,1) to (7,7) in the fourth feature map are covered. These features include the temperature synchronization feature of the upper left region (0.9), the thermal diffusion effect feature of the central region (0.85), and the thermal aggregation effect feature of the lower right region (0.8), etc.

[0077] The weights of the aforementioned convolutional kernels are pre-trained to identify specific deep association patterns. For example, a certain convolutional kernel is specifically designed to identify diagonal thermal coupling patterns. When high-temperature-related features appear simultaneously in the upper left and lower right regions, the convolution outputs a high response value of 0.85, indicating that the deep association pattern of diagonal thermal coupling has been detected. In this step, 16 convolutional kernels operate in parallel, each generating an 8×8 high-order feature map, which are stacked to form a high-order feature map of size 8×8×16. Each feature map corresponds to a type of deep association pattern.

[0078] Step a3: Globally integrate the high-order feature maps corresponding to the multiple receptive fields and splice them together to form a unified global feature representation.

[0079] Global integration refers to aligning and stitching together the high-order feature maps generated by all receptive fields in the spatial dimension to form a unified feature representation covering the entire temperature field; global feature representation refers to the integrated feature map containing global long-range correlation information.

[0080] In step a3, after confirming that all higher-order feature maps have the same spatial dimensions and coordinate system, they are stitched together along the channel dimension. The spatial dimensions of the stitched feature map remain unchanged, and the number of channels is equal to the sum of the number of channels in each higher-order feature map. Through this stitching operation, the local higher-order information originally scattered in different feature maps is integrated into the same feature representation, so that each spatial location contains correlation information from different receptive fields and different scales, thereby achieving a complete representation of the global temperature field.

[0081] For example, step a2 generates 16 high-order feature maps of size 8×8, corresponding to 16 different types of deep association modes, such as diagonal thermal coupling mode, horizontal thermal propagation mode, vertical thermal propagation mode, and temperature field symmetry mode. These 16 feature maps are concatenated along the channel dimension to generate a global feature representation of size 8×8×16. After concatenation, each spatial location has a 16-dimensional feature vector, which contains the response value of that location in the 16 different deep association modes. For example, at position (4, 4), the feature vector is a 16-dimensional vector [0.85, 0.32, 0.45, 0.78, ...], indicating that the response of that location is 0.85 in the diagonal thermal coupling mode, 0.32 in the horizontal thermal propagation mode, and so on.

[0082] Step a4: Perform feature dimension aggregation on the global feature representation to form a deep feature map that characterizes the intrinsic correlation of the global temperature field.

[0083] Among them, the feature dimension aggregation operation refers to the fusion and compression of global feature representations in the channel dimension, classifying and merging similar or related features to form a compact deep feature map; the intrinsic correlation of the global temperature field refers to the essential features that reflect the spatial structure and evolution trend of the temperature field of the entire SiC MOSFET array.

[0084] In step a4, multiple features are fused and compressed along the channel dimension using methods such as 1×1 convolution or channel attention mechanisms. For example, 1×1 convolution can weight and combine features from different channels, fusing multiple related features into fewer comprehensive features; or, channel attention mechanisms can learn the importance weights of each channel feature, enhancing important features and suppressing secondary features. After the aggregation operation, the spatial size of the output feature map remains unchanged, while the number of channels is reduced, forming a deep feature map. This feature map represents the intrinsic correlation of the global temperature field, including essential features such as the spatial structure and evolution trend of the temperature field.

[0085] For example, for a global feature representation of size 8×8×16, eight 1×1 convolutional kernels are used for convolution operations. Each 1×1 convolutional kernel performs a weighted summation along the channel dimension, fusing the 16 input channels into one output channel. The eight convolutional kernels operate in parallel to generate a deep feature map of size 8×8×8. For instance, a certain 1×1 convolutional kernel fuses four related features—diagonal thermal coupling mode, horizontal thermal propagation mode, vertical thermal propagation mode, and temperature field symmetry mode—into a comprehensive feature representing the spatial structure of the temperature field. The eight channels of the final deep feature map represent different types of intrinsic correlations in the global temperature field, such as global thermal coupling strength, dominant direction of thermal propagation, temperature field symmetry index, and distribution of heat accumulation regions.

[0086] Step 1044: Through the output layer of the convolutional neural network, the deep features in the fifth feature map are weighted and summed to obtain the risk index.

[0087] The output layer refers to the last layer of the convolutional neural network, which is used to map the extracted deep features to the final risk assessment result. It is usually composed of a global average pooling layer and a fully connected layer, or directly composed of a fully connected layer. The weighting coefficients in the weighted summation process can be predetermined through the training process of the convolutional neural network; the risk index is a quantitative assessment value of the possibility of thermal runaway of the entire SiC MOSFET array, used to characterize the degree of risk under the current array temperature distribution state.

[0088] In step 1044, the fifth feature map output in step 1043 is first flattened or subjected to global average pooling to transform it into a one-dimensional feature vector. Specifically, if flattening is used, the feature values ​​of all spatial locations and all channels in the fifth feature map are arranged sequentially into a one-dimensional vector; if global average pooling is used, the average value of all spatial locations in the feature map of each channel is calculated to obtain a one-dimensional vector with a length equal to the number of channels. Both methods can compress spatial information into vector form, which is convenient for subsequent fully connected layer processing.

[0089] Then, the one-dimensional vector is input into the fully connected layer of the output layer. Each neuron in the fully connected layer is connected to all components of the input feature vector, and each connection corresponds to a weight coefficient. The fully connected layer performs a weighted summation operation, multiplying each component of the input feature vector by its corresponding weight coefficient and summing the results, then adding a bias term to obtain a preliminary weighted sum. This weighted sum is then passed through an activation function, such as the sigmoid function or the softmax function, for a non-linear transformation, mapping the output value to a preset range, ultimately yielding a quantified risk index.

[0090] It should be noted that the weights and biases in the fully connected layers are automatically learned through backpropagation during the training of the convolutional neural network. During training, a large number of historical temperature feature maps labeled with actual thermal runaway risks are used as training samples. The weights are continuously adjusted through optimization algorithms to minimize the error between the network's output risk index and the actual risk label. After training, these weights are fixed and used for calculating the risk index in practical applications.

[0091] This application enables temperature characteristic analysis from local to global perspectives, thereby allowing the risk index to fully reflect the thermal state of the array and providing accurate and reliable quantitative basis for subsequent risk-based equilibrium control.

[0092] S105. Through the intelligent agents corresponding to each device in the SiC MOSFET array, based on the correlation data between temperature and current, the risk index, and the current deviation between adjacent devices, reinforcement learning technology is used to dynamically adjust the duty cycle of the gate drive signal of the corresponding device to achieve balanced control of the SiC MOSFET array.

[0093] Among them, the intelligent agent refers to an independent control unit configured for each device in the SiC MOSFET array, which is used to make autonomous decisions and adjust the operating parameters of the corresponding device based on local information and neighbor information; Reinforcement learning is a machine learning method that optimizes decisions based on feedback by interacting with the environment. It can be implemented through a reinforcement learning model. It should be noted that this embodiment does not limit the structural design and training process of the reinforcement learning model. The specific implementation process can be referred to steps 1051 to 1052 below, which will not be repeated here.

[0094] The duty cycle of the gate drive signal, also known as the duty cycle, refers to the ratio of the duration of the high level in the gate drive signal to the period, and is used to adjust the conduction level of the device.

[0095] The current deviation between adjacent devices refers to the degree of difference in real-time current values ​​between any two adjacent devices during the operation of a SiC MOSFET array.

[0096] In one specific implementation, step S105 includes the following steps: Step 1051: Each intelligent agent acquires the temperature and current correlation data and the risk index of the corresponding device, and acquires the real-time current value of the adjacent device through real-time communication to determine its current state space; the state space includes at least: the temperature and current matching degree, the risk index of the device, and the current deviation between the device and the adjacent device.

[0097] Among them, adjacent devices refer to other SiC MOSFET devices that are directly adjacent to the current device in terms of spatial location; real-time current value refers to the current value of the device being turned on, which is collected in real time by a current sensor; The state space refers to the set of multidimensional information used to describe the current state of an agent, which serves as the input for the decision-making of a reinforcement learning model. Temperature-current matching degree refers to the degree of agreement between the real-time current value of a device and the ideal current value determined from associated data based on the current temperature. The risk index of this device refers to the local risk component related to this device in the risk index output in step S104, or the overall risk index reflected in this device. The current deviation between this device and its neighboring devices refers to the degree of difference in real-time current values ​​between the current device and each of its neighboring devices, which can be expressed as the maximum deviation value, the average deviation value, or the deviation range.

[0098] Before step 1044, each spatial location of the fifth feature map corresponds to a region on the substrate. The local risk contribution value of each location can be obtained by weighting the deep feature vectors at that location. By performing a weighted summation of the spatial dimensions of the fifth feature map, both the overall risk index and the local risk components of each spatial location can be obtained simultaneously. The agent can obtain the local risk component of its corresponding device location from the network as the risk index of its device.

[0099] Alternatively, the overall risk index can be proportionally allocated based on the operating parameters of each device to obtain a local risk index; or, an independent local risk assessment branch can be configured for each device. Therefore, this embodiment does not limit the specific method of obtaining the local risk components, as long as each agent can obtain a quantitative index reflecting the degree of local risk of its corresponding device.

[0100] For example, the agent corresponding to device A has determined the current state space as follows: the temperature and current matching degree is 94%, the risk index of this device is 7.5, and the maximum current deviation from the adjacent device is 2A.

[0101] Step 1052: The agent inputs the state space into the reinforcement learning model to initially determine the initial adjustment action.

[0102] Specifically, step 1052 may include the following process: the reinforcement learning model evaluates the expected cumulative reward that can be obtained by performing different adjustment actions according to the state space, and selects the action with the largest expected reward as the initial adjustment action, which includes the adjustment direction and adjustment magnitude.

[0103] In step 1052, after receiving the state space, the reinforcement learning model begins to analyze the current state and determines all possible adjustment actions as candidate actions. These candidate actions include adjustment magnitudes in different directions. For each candidate action, the model calculates the expected cumulative reward that can be obtained after executing the action through forward propagation. The formula for calculating the expected cumulative reward takes into account both immediate reward and future reward. Among them, the immediate reward is positively correlated with the improvement of temperature balancing effect, the reduction of risk index, and the reduction of current deviation after the action is executed. The future reward is discounted to the present through a discount factor so that the action generated by the agent in the current state can obtain the maximum long-term benefit. It should be noted that the expression of the formula for calculating the expected cumulative reward can be referred to relevant technologies, and will not be repeated here.

[0104] After calculating the expected cumulative reward of all candidate actions, the reinforcement learning model uses an existing greedy strategy or an ε-greedy strategy to select the action with the largest expected cumulative reward as the initial adjustment action. This selected action, as the initial adjustment action, contains a clear adjustment direction and adjustment magnitude.

[0105] For example, the agent inputs this state space into the reinforcement learning model. After receiving the state space, the reinforcement learning model begins to analyze the current state. The model identifies all possible adjustment actions as candidate actions, including options such as increasing the gate drive signal duty cycle by 1%, 2%, 3%, decreasing by 1%, 2%, 3%, or keeping it unchanged.

[0106] For each candidate action, the model performs forward propagation calculations through its internal policy network to evaluate the expected cumulative reward obtained after executing the action. If the model obtains the following expected cumulative rewards for each candidate action: a 2% reduction in duty cycle yields an expected cumulative reward of 0.85, a 1% reduction in duty cycle yields an expected cumulative reward of 0.72, keeping the duty cycle unchanged yields an expected cumulative reward of 0.45, a 1% increase in duty cycle yields an expected cumulative reward of 0.30, a 2% increase in duty cycle yields an expected cumulative reward of 0.18, and the expected cumulative rewards for other actions are all below 0.85, then the reinforcement learning model adopts a greedy strategy to select the action with the largest expected cumulative reward as the initial adjustment action. Therefore, a 2% reduction in duty cycle is chosen as the initial adjustment action, which includes a clear adjustment direction of reducing the duty cycle with an adjustment magnitude of 2%.

[0107] Step 1053: Based on the information interaction between the intelligent agent and the neighboring intelligent agents, exchange their respective preliminary adjustment actions. When a conflict is detected between the preliminary adjustment actions of the two parties, the preliminary adjustment action of the agent is modified according to the preset negotiation rules or the cooperation mechanism in the reinforcement learning model to generate the target adjustment action.

[0108] The pre-defined negotiation rules or cooperation mechanism in the reinforcement learning model can refer to the following: when the local risk indices of multiple agents are all higher than the pre-defined risk index threshold but there are conflicts in the adjustment actions, the overall risk index can be used as a reference benchmark for global coordination to determine which area's adjustment needs should be prioritized.

[0109] For example, the agent of device A determines its initial adjustment action as reducing the duty cycle by 2% through a reinforcement learning model. Simultaneously, the agent of neighboring device B, based on its own state space, determines its initial adjustment action as increasing the duty cycle by 1.5% through its reinforcement learning model. The agent of device A exchanges its initial adjustment actions with the agents of neighboring devices B, C, and D via a real-time communication network.

[0110] When the agent of device A receives the initial adjustment action from device B, it detects a conflict between the two adjustment actions. The specific analysis is as follows: device A needs to reduce its duty cycle, which will cause its current to decrease; device B needs to increase its duty cycle, which will cause its current to increase. The simultaneous decrease in current and increase in current of two adjacent devices may further widen the current deviation between them, and even trigger local current oscillations, affecting the stable operation of the array.

[0111] According to the preset negotiation rules, when it is detected that the initial adjustment actions of adjacent intelligent agents are in opposite directions and may exacerbate the current imbalance, the two parties initiate a negotiation mechanism. The specific negotiation rules are as follows: compare the local risk indices of the two devices. The device with the higher local risk index has a higher adjustment priority, and its initial adjustment action remains unchanged; the device with the lower local risk index needs to correct its initial adjustment action by halving its adjustment magnitude.

[0112] Currently, device A has a local risk index of 7.5, while device B has a local risk index of 6.2. Therefore, device A has a higher priority than device B. According to the negotiation rules, the agent of device A maintains its initial adjustment action, which is still to reduce the duty cycle by 2%. After receiving the negotiation result, the agent of device B modifies its initial adjustment action from increasing the duty cycle by 1.5% to increasing the duty cycle by 0.75%, in order to reduce the degree of conflict with the adjustment action of device A.

[0113] After both agents confirm that the revised adjustment actions do not conflict, they each designate the revised actions as the target adjustment actions and prepare to execute them.

[0114] Step 1054: Adjust the action according to the target, adjust the duty cycle of the gate drive signal of the corresponding device through the intelligent agent, observe the state change after the adjustment after a preset period, quantify the state change into a reward value and feed it back to the reinforcement learning model to update the model parameters and realize the balanced control of the SiC MOSFET array.

[0115] For example, the agent of device A, based on the target adjustment action determined in step 1053, sends a control command to the gate drive circuit of its corresponding device A, reducing the duty cycle of the gate drive signal from the current 20% to 18%. After the duty cycle adjustment is completed, the agent starts a timer and waits for the preset observation period to end. For example, in this embodiment, the preset observation period is 100ms, which is sufficient for the device to respond to the duty cycle change and reach a new steady state.

[0116] 100ms later, the agent re-acquires the operating status of device A. The temperature sensor reads that the temperature of device A has decreased from 112℃ to 106℃, and the current sensor reads that the current of device A has decreased from 19A to 17.2A. The agent reads the ideal current value of 17.5A at the current temperature from its local memory and recalculates the temperature-current matching degree to 98%. Simultaneously, the agent obtains the updated current values ​​of neighboring devices through the real-time communication network: device B's current increases from 17A to 17.3A, device C's current remains unchanged at 18A, and device D's current remains unchanged at 20A.

[0117] The current deviations from adjacent devices are recalculated: the deviation from device B is |17.2-17.3|=0.1A, the deviation from device C is |17.2-18|=0.8A, and the deviation from device D is |17.2-20|=2.8A, with a maximum current deviation of 2.8A. The agent can also determine the updated overall risk index to be 7.6, and extract the local risk index of the region where the device is located as 6.5.

[0118] The agent quantifies the adjusted state changes into reward values, which are then fed back to the reinforcement learning model. The reward value calculation considers the following factors: the local risk index decreases from 7.5 to 6.5, the temperature decreases from 112℃ to 106℃, the maximum current deviation increases from 2A to 2.8A, and the temperature-current matching accuracy improves from 94% to 98%. For example, the agent quantifies each state change into a corresponding reward component according to a preset reward value mapping rule. The mapping rule uses a piecewise linear function, as follows: mapping of risk index reduction to reward component: 0.2 reward for every 0.5 reduction; mapping of temperature reduction to reward component: 0.25 reward for every 5℃ reduction; mapping of current deviation increase to penalty component: 0.1 penalty for every 0.4A increase; mapping of matching degree improvement to reward component: 0.05 reward for every 2 percentage point improvement.

[0119] Therefore, the risk reduction reward for device A is 0.4, the temperature improvement reward is 0.3, the current deviation penalty is -0.2, and the matching improvement reward is 0.1. The total reward for this adjustment is 0.4 + 0.3 - 0.2 + 0.1 = 0.6.

[0120] The agent packages the complete empirical data from this adjustment into a training sample. This sample includes: the state space before adjustment with a matching degree of 94%, a local risk index of 7.5, and a maximum current deviation of 2A; the target adjustment action performed is to reduce the duty cycle by 2%, resulting in a reward value of 0.6; and the new state space after adjustment with a matching degree of 98%, a local risk index of 6.5, and a maximum current deviation of 2.8A. The agent sends this sample to the training module of the reinforcement learning model for online model updates.

[0121] After receiving the training sample, the reinforcement learning model adjusts the parameters of its internal policy network through backpropagation. This allows the network to more accurately predict the expected cumulative reward for each action when encountering a similar state in the future. Through this closed-loop mechanism of adjustment, observation, feedback, and update, the reinforcement learning model of each agent continuously optimizes its decision-making strategy, achieving adaptive balanced control of the entire SiC MOSFET array.

[0122] This application enables distributed and precise control based on real-time data by having each device's intelligent agent autonomously analyze its state, formulate preliminary adjustment actions, and coordinate conflicts with adjacent intelligent agents to dynamically adjust the duty cycle of the gate drive signal. This prioritizes reducing the risk of thermal runaway while also balancing current deviations. Furthermore, the adjustment experience accumulated through reinforcement learning continuously optimizes decisions to maintain a balanced temperature and current for each device in the SiC MOSFET array, thereby improving the stability and reliability of the array's operation.

[0123] Figure 3 This is a schematic diagram illustrating a specific implementation of an equalization control system for SiC MOSFET arrays provided in this application. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire the temperature data of each device in the SiC MOSFET array in real time using the thin-film thermocouple array arranged on the substrate of the SiC MOSFET array.

[0124] The matching module 32 is used to dynamically match the temperature data with a database that stores the thermal characteristic curves of corresponding similar devices to obtain the correlation data between the temperature and current of the device.

[0125] The construction module 33 is used to construct a three-dimensional temperature field corresponding to the SiC MOSFET array based on the temperature data of each device, and generate a temperature feature map corresponding to the three-dimensional temperature field.

[0126] The extraction module 34 is used to input the temperature feature map into the convolutional neural network, extract key information from the temperature feature map and perform hierarchical operations through multi-layer convolution and pooling operations in the convolutional neural network, and output the risk index.

[0127] The adjustment module 35 is used to dynamically adjust the duty cycle of the gate drive signal of the corresponding device by using reinforcement learning technology based on the correlation data between temperature and current, the risk index and the current deviation between adjacent devices, through the intelligent agents corresponding to each device in the SiC MOSFET array, so as to achieve balanced control of the SiC MOSFET array.

[0128] An equalization control system for SiC MOSFET arrays according to an embodiment of this application is used to implement the aforementioned equalization control method for SiC MOSFET arrays. Therefore, the specific implementation of the equalization control system for SiC MOSFET arrays can be found in the embodiment section of the equalization control method for SiC MOSFET arrays above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0129] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the equalization control method for SiC MOSFET array described above.

[0130] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described equalization control methods for SiC MOSFET arrays.

[0131] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0132] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the equalization control method for SiC MOSFET arrays described above.

[0133] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0134] The foregoing has provided a detailed description of the equalization control method, system, electronic device, and dielectric for SiC MOSFET arrays provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for equalization control of SiC MOSFET arrays, characterized in that, include: Temperature data of each device in the SiC MOSFET array is collected in real time using a thin-film thermocouple array arranged on the substrate of the SiC MOSFET array. The temperature data is dynamically matched with a database containing thermal characteristic curves of similar devices to obtain the correlation data between the device's temperature and current. Based on the temperature data of each device, a three-dimensional temperature field corresponding to the SiC MOSFET array is constructed, and a temperature feature map corresponding to the three-dimensional temperature field is generated. The temperature feature map is input into a convolutional neural network. Through multi-layer convolution and pooling operations in the convolutional neural network, key information in the temperature feature map is extracted and hierarchical operations are performed to output a risk index. By using the intelligent agents corresponding to each device in the SiC MOSFET array, based on the correlation data between temperature and current, the risk index, and the current deviation between adjacent devices, reinforcement learning technology is used to dynamically adjust the duty cycle of the gate drive signal of the corresponding device to achieve balanced control of the SiC MOSFET array.

2. The method according to claim 1, characterized in that, The process involves extracting key information from the temperature feature map through multiple convolutional and pooling operations in the convolutional neural network, performing hierarchical calculations, and outputting a risk index, including: The temperature feature map is convolved by the first convolutional layer in the convolutional neural network to extract a first feature map that represents the temperature value and temperature change gradient of the local area. The first feature map is then dimensionality-reduced and filtered by the first pooling layer to retain local features that meet the first preset condition, and a second feature map is output. The second feature map is convolved by the second convolutional layer in the convolutional neural network. Correlation analysis is performed on the local features of adjacent spatial regions to extract a third feature map that represents the temperature linkage relationship between regions. The third feature map is then reduced in dimensionality by the second pooling layer to filter out the target linkage features that represent the temperature field regional correlation pattern and output a fourth feature map. The fourth feature map is convolved by the third convolutional layer in the convolutional neural network. The target linkage features between spatially non-directly adjacent regions are abstracted and fused in a higher order to extract deep feature maps. The deep feature maps are then reduced in dimensionality by the third pooling layer to retain deep features that meet the second preset condition and output the fifth feature map. The risk index is obtained by weighted summation of the deep features in the fifth feature map through the output layer of the convolutional neural network.

3. The method according to claim 2, characterized in that, The step of performing a convolution operation on the fourth feature map through the third convolutional layer in the convolutional neural network, and performing high-order abstraction and fusion of target linkage features between spatially non-directly adjacent regions to extract deep feature maps includes: The third convolutional layer uses a convolutional kernel of a preset size to set a corresponding receptive field on the fourth feature map. Each receptive field covers at least two image regions that are not directly adjacent in spatial location. Within each receptive field, feature combination and high-order abstraction are performed on multiple target linkage features to mine deep correlation patterns in thermal state changes of different regions and generate high-order feature maps corresponding to each receptive field. The high-order feature maps corresponding to multiple receptive fields are globally integrated and spliced ​​together to form a unified global feature representation; The global feature representation is aggregated along its feature dimensions to form a deep feature map that characterizes the intrinsic correlations of the global temperature field.

4. The method according to claim 1, characterized in that, The method involves using an intelligent agent corresponding to each device in the SiC MOSFET array, based on the correlation data between temperature and current, the risk index, and the current deviation between adjacent devices, to dynamically adjust the duty cycle of the gate drive signal of the corresponding device using reinforcement learning technology, thereby achieving balanced control of the SiC MOSFET array. This includes: Each of the intelligent agents acquires the correlation data of temperature and current of the corresponding device, the risk index, and obtains the real-time current value of adjacent devices through real-time communication to determine its own current state space; the state space includes at least: the matching degree of temperature and current, the risk index of the device, and the current deviation between the device and adjacent devices. The agent inputs the state space into the reinforcement learning model to initially determine the initial adjustment actions; Based on the information interaction between the intelligent agent and neighboring intelligent agents, they exchange their initial adjustment actions. When a conflict is detected between the initial adjustment actions of the two parties, the initial adjustment action of the agent is modified according to the preset negotiation rules or the cooperation mechanism in the reinforcement learning model to generate the target adjustment action. According to the target adjustment action, the agent adjusts the duty cycle of the gate drive signal of the corresponding device, observes the state change after the adjustment after a preset period, quantifies the state change into a reward value and feeds it back to the reinforcement learning model to update the model parameters and realize the balanced control of the SiC MOSFET array.

5. The method according to claim 4, characterized in that, The agent inputs the state space into the reinforcement learning model to initially determine preliminary adjustment actions, including: The reinforcement learning model evaluates the expected cumulative reward that can be obtained by performing different adjustment actions based on the state space, and selects the action with the largest expected reward as the initial adjustment action. The initial adjustment action includes the adjustment direction and adjustment magnitude.

6. The method according to claim 1, characterized in that, The step of dynamically matching the temperature data with a database storing thermal characteristic curves of corresponding similar devices to obtain the correlation data between the device's temperature and current includes: For each device in the SiC MOSFET array, thermal characteristic curves matching the model and specifications of the device are selected from the database; Based on the temperature data of the device, the corresponding current range is determined by locating the corresponding current range on the thermal characteristic curve of the device. Based on the operating environment parameters of the device, the current range is dynamically adjusted so that the adjusted current range has a unique correspondence with the temperature data of the device, thus forming the correlation data between the device's temperature and current.

7. The method according to claim 1, characterized in that, The construction of the three-dimensional temperature field corresponding to the SiC MOSFET array based on the temperature data of each device includes: The three-dimensional spatial coordinates of each device in the SiC MOSFET array are determined, and the temperature data of each device is associated with the three-dimensional spatial coordinates to form a discrete dataset; Based on the discrete dataset, an initial three-dimensional model is constructed with the physical boundaries of the substrate as the modeling range; For the blank areas in the initial three-dimensional model where no devices are placed, the temperature value of the blank areas is calculated based on the temperature data of the existing devices adjacent to the blank areas and the spatial distance between the blank areas and the existing devices, so as to fill the blank areas in the initial three-dimensional model and form a three-dimensional temperature field.

8. A balanced control system for SiC MOSFET arrays, characterized in that, include: The acquisition module is used to acquire the temperature data of each device in the SiC MOSFET array in real time using a thin-film thermocouple array arranged on the substrate of the SiC MOSFET array. The matching module is used to dynamically match the temperature data with a database that stores the thermal characteristic curves of corresponding similar devices to obtain the correlation data between the device's temperature and current. The construction module is used to construct a three-dimensional temperature field corresponding to the SiC MOSFET array based on the temperature data of each device, and generate a temperature feature map corresponding to the three-dimensional temperature field. The extraction module is used to input the temperature feature map into a convolutional neural network, extract key information from the temperature feature map through multi-layer convolution and pooling operations in the convolutional neural network, perform hierarchical operations, and output a risk index. The adjustment module is used to dynamically adjust the duty cycle of the gate drive signal of the corresponding device by using reinforcement learning technology based on the correlation data between temperature and current, the risk index and the current deviation between adjacent devices, through the intelligent agents corresponding to each device in the SiC MOSFET array, so as to achieve balanced control of the SiC MOSFET array.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a balanced control method for a SiCMOSFET array as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables a balanced control method for SiC MOSFET arrays as described in any one of claims 1 to 7.

Images