A method for predicting junction temperature of IGBT in inverter based on neural network

Abstract

The present invention provides a method for predicting the junction temperature of an inverter IGBT based on a neural network, which relates to the field of neural network technology, including: obtaining the circuit structure and operating parameters of a subway traction drive system, and constructing a dynamic thermal network model; establishing a simulation condition set, performing simulation acquisition of a real-time parameter set, and establishing a sample data set after preprocessing; establishing a neural network with multiple input layers, executing training to construct an IGBT junction temperature prediction model, each input layer is mapped with a feature extraction subnetwork, and feature extraction is performed using the feature extraction subnetwork, and neural network training is performed based on the feature extraction results; data prediction is performed using the IGBT junction temperature prediction model, and a junction temperature time series is established. The present invention solves the technical problem that the prior art is difficult to capture the dynamic changes of junction temperature under complex operating conditions, especially when the load fluctuation is large or multiple operating conditions are switched, the prediction error of the traditional model increases significantly, resulting in insufficient accuracy of junction temperature prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of neural network, and in particular to a method for predicting junction temperature of an IGBT of a frequency converter based on a neural network. Background Art

[0002] The insulated gate bipolar transistor (IGBT) power module integrates the advantages of insulated gate field effect transistors and power transistors, and has the advantages of high switching frequency and small driving power. It is a key component of the subway traction transmission system. The power module has a "sandwich" structure. The internal structure section from top to bottom is composed of seven layers of materials, including IGBT chip and diode chip, welding layer, ceramic copper clad plate (including upper copper layer, insulating ceramic layer, lower copper layer), welding layer and bottom plate.

[0003] In practical applications, the IGBT modules of subway converters usually have a switching frequency of kHz and a high-power operation state, which generates large heat loss during operation, causing the junction temperature of the chip to rise and fluctuate, which may cause device failure in severe cases. There are four methods for monitoring the junction temperature of IGBT modules: optical measurement method, physical contact method, thermal model method and temperature-sensitive electrical parameter method.

[0004] In view of the following problems in the existing high-power high-voltage inverter IGBT junction temperature prediction methods: 1) considering multiple influencing factors such as operating conditions switching frequency and gate resistance; 2) complex calculations. Therefore, a macro-micro attention long short-term memory recursive neural network IGBT junction temperature prediction method based on a macro-micro attention long short-term memory recursive neural network that takes into account both self-learning ability and ease of use is very important. Summary of the invention

[0005] The present application provides a method for predicting the junction temperature of an inverter IGBT based on a neural network, aiming to solve the technical problem that the existing technology is difficult to capture the dynamic changes of junction temperature under complex operating conditions, especially when the load fluctuates greatly or multiple operating conditions are switched, the prediction error of the traditional model increases significantly, resulting in insufficient accuracy of junction temperature prediction.

[0006] The present application discloses a method for predicting junction temperature of an inverter IGBT based on a neural network. The method comprises: after establishing digital communication with a subway traction drive system, obtaining the circuit structure and operating parameters of the subway traction drive system, and using MATLAB / Simulink software to construct a dynamic thermal network model of the inverter of the subway traction drive system; establishing a simulation condition set of the subway traction drive system, performing simulation based on the dynamic thermal network model with the simulation condition set and environmental parameters, collecting a real-time parameter set, the real-time parameter set including on-state voltage drop, collector current, ambient temperature, radiator temperature, and load fluctuation, and preprocessing the real-time parameter set. After that, a sample data set is established; a neural network with multiple input layers is established, and the sample data set is input into the neural network with multiple input layers for training, so as to construct an IGBT junction temperature prediction model, wherein the multiple input layers include a on-state voltage drop input layer, a collector current input layer, an ambient temperature input layer, a heat sink temperature input layer and a load fluctuation input layer, and each input layer is mapped with a feature extraction sub-network, and before the neural network performs training, the feature extraction sub-network is used to extract features of the corresponding feature data, and the neural network is trained according to the feature extraction results; the IGBT junction temperature prediction model is used to perform data prediction, and a junction temperature time series is established.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] The use of neural network models, including multi-input layer networks and attention mechanisms, can efficiently capture the complex nonlinear relationship between input parameters and IGBT junction temperature, improve the accuracy of prediction, and establish multi-operating condition simulation data sets. The model can be widely adapted to different operating environments and is suitable for a variety of practical engineering applications. Compared with traditional physical measurement methods or complex thermal model methods, this method only needs to collect a small number of parameters such as on-state voltage drop and collector current for prediction, avoiding the need for complex hardware or experimental equipment. By cleaning, normalizing and time series segmenting the real-time parameter set, a high-quality data set is generated, which is convenient for neural network learning and using MAT LAB / Simulink software is used to establish a thermal network model, which simplifies the modeling and training process and can be directly applied to actual projects. The dynamic thermal resistance formula is used to accurately simulate the time dynamics and power input response of the IGBT junction temperature, which improves the authenticity of the heat transfer simulation. The multi-input layer structure allows independent modeling of features from different sources, improving the model's ability to understand multi-dimensional data. At the same time, the feature extraction sub-network is used to optimize feature expression and improve prediction accuracy. The IGBT junction temperature prediction model is used to predict data and establish a junction temperature time series, which can not only monitor the junction temperature in real time, but also predict future trends, providing guidance for equipment maintenance and optimized operation.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic flow chart of a method for predicting junction temperature of an inverter IGBT based on a neural network is provided for an embodiment of the present application;

[0011] Figure 2 A schematic flow chart of establishing a sample data set in a method for predicting junction temperature of an inverter IGBT based on a neural network is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0012] The embodiment of the present application solves the technical problem that the prior art is difficult to capture the dynamic changes of junction temperature under complex operating conditions, especially when the load fluctuates greatly or multiple operating conditions are switched, the prediction error of the traditional model increases significantly, resulting in insufficient accuracy of junction temperature prediction, by providing a method for predicting the junction temperature of an inverter IGBT based on a neural network.

[0013] After introducing the basic principles of the present application, various non-limiting implementation methods of the present application will be specifically described below in conjunction with the drawings in the specification.

[0014] like Figure 1 As shown, the embodiment of the present application provides a method for predicting junction temperature of an inverter IGBT based on a neural network, the method comprising:

[0015] After establishing digital communication with the subway traction drive system, the circuit structure and operating parameters of the subway traction drive system are obtained, and the dynamic thermal network model of the subway traction drive system inverter is constructed using MATLAB / Simulink software.

[0016] Digital communication with the subway traction drive system can be established through industrial protocols. Through this communication, information related to circuit operation can be obtained in real time, including circuit structure and operating parameters. The circuit structure refers to the inverter circuit design used in the subway traction system, which is composed of circuit elements (IGBT), capacitors, inductors, transformers and other components. The circuit topology will affect the generation and propagation of heat; operating parameters include the operating frequency of the inverter, the collector current and voltage of the IGBT module, the power loss of the IGBT, the ambient temperature, etc.

[0017] MATLAB / Simulink is an engineering simulation software widely used in dynamic system modeling and simulation. In this embodiment, MATLAB / Simulink is used to construct a dynamic thermal network model of the inverter of the subway traction drive system.

[0018] The dynamic thermal network model of the inverter is used to describe the process of heat generation, transfer and dissipation in the circuit. The model is based on the concept of thermal resistance, that is, the heat flow is determined by the thermal resistance between the power loss of the circuit element (IGBT) and the heat conduction path between it and the surrounding environment (such as heat sink, air, etc.). The model includes static thermal resistance and dynamic thermal resistance. Among them, the static thermal resistance is related to the thermal conduction characteristics of the material and structure and is fixed. The dynamic thermal resistance changes dynamically with time and power input, characterizing the resistance of heat in the system and affecting the response speed and change amplitude of the system temperature.

[0019] A simulation condition set of the subway traction drive system is established, and a simulation based on a dynamic thermal network model is performed using the simulation condition set and environmental parameters. A real-time parameter set is collected, including on-state voltage drop, collector current, ambient temperature, radiator temperature, and load fluctuation. After preprocessing the real-time parameter set, a sample data set is established.

[0020] The simulation condition set is the input scenario for the operation of the dynamic thermal network model, including different loads and operating conditions of the subway traction drive system. According to the operating characteristics of the subway traction system, the following main conditions are divided. Specifically, light load condition: the motor works under low load conditions, the output power is small, and the heat loss is low; half load condition: the motor output power reaches half of the rated power, and the heat loss is medium; full load condition: the motor works under rated power conditions, and the heat loss is large; overload condition (optional): test the junction temperature fluctuation under extreme conditions.

[0021] Environmental parameters reflect the impact of environmental conditions on thermal performance in the simulation, including ambient temperature, radiator temperature, operating condition duration, etc. The ambient temperature can be set to a fixed value or a dynamically changing value, such as 30°C, 40°C, etc. The radiator temperature is affected by the heat dissipation conditions and the environment.

[0022] The dynamic thermal network model is run through MATLAB / Simulink to simulate the above working conditions and simulate the junction temperature changes of the IGBT module under various working conditions. Real-time parameter sets are collected during the simulation process, including on-state voltage drop: the real-time value of the on-state voltage drop of the IGBT module; collector current: the load current of the IGBT module; ambient temperature: the impact of the environment on the cooling system; radiator temperature: the fluctuation of the radiator temperature; load fluctuation: the dynamic change of the load.

[0023] After data collection is completed, the real-time parameter set is cleaned and standardized to ensure data quality. Data cleaning includes removing outliers, noise data, and smoothing and filling missing values. Standardization uses normalization methods to scale different parameters to the range of [0,1]. The processed data is used as a sample data set for neural network model training.

[0024] A neural network with multiple input layers is established, and sample data sets are input into the neural network with multiple input layers for training, so as to construct an IGBT junction temperature prediction model, wherein the multiple input layers include an on-state voltage drop input layer, a collector current input layer, an ambient temperature input layer, a heat sink temperature input layer and a load fluctuation input layer, and each input layer is mapped with a feature extraction sub-network. Before the neural network performs training, the feature extraction sub-network is used to extract features of the corresponding feature data, and the neural network is trained according to the feature extraction results.

[0025] A neural network with multiple input layers is established. The multi-input layer neural network is designed to process input data with a variety of different characteristics. Each input layer corresponds to a specific input feature. The multiple input layers include a forward voltage drop input layer, a collector current input layer, an ambient temperature input layer, a heat sink temperature input layer, and a load fluctuation input layer. Each input layer is mapped to a feature extraction subnetwork. The function of the subnetwork is to extract a high-dimensional representation of the input features, such as generating a more representative feature vector through dimensionality reduction or nonlinear transformation.

[0026] Before the neural network is trained, the sample data set is input into the corresponding input layer, and the input data is processed by the feature extraction sub-network to extract the high-dimensional representation of each feature. After processing, each sub-network outputs a feature vector, which is used as the input for the next step. The feature vectors extracted by each feature extraction sub-network are integrated in the fusion layer. For example, the splicing method is used to directly splice all feature vectors to form a large feature vector, or through weighted summation, according to the weight of each feature, it is synthesized into a comprehensive feature representation. The obtained fused feature vector will be used as the input of the main network for predicting the junction temperature.

[0027] The neural network is trained and verified based on the feature extraction results. The batch gradient descent method can be used to optimize the loss function to build an accurate IGBT junction temperature prediction model. This design can effectively process complex and multi-dimensional input data and has strong generalization capabilities.

[0028] The IGBT junction temperature prediction model is used to perform data prediction and establish a junction temperature time series.

[0029] By using the trained IGBT junction temperature prediction model, the junction temperature time series of the IGBT module is generated according to the real-time input data. Through the analysis and visualization of the time series, the real-time monitoring and prediction of the operating status of the IGBT module can be achieved, providing important support for the reliable operation of high-power and high-voltage inverters.

[0030] Furthermore, the method of constructing a dynamic thermal network model of a subway traction drive system inverter using MATLAB / Simulink software further includes:

[0031] The dynamic thermal network model includes static thermal resistance and dynamic thermal resistance, and the dynamic thermal resistance is constructed by fitting the following formula:

[0032] ;

[0033] in, Characterize the dynamic thermal resistance, over time t and power input a is the initial response amplitude of thermal resistance, b is the thermal resistance attenuation coefficient, t represents the time variable, c is the dynamic response coefficient of power input, is the power input.

[0034] Specifically, the dynamic thermal network model includes static thermal resistance and dynamic thermal resistance. The static thermal resistance is related to the thermal conductivity characteristics of materials and structures and is fixed. The dynamic thermal resistance changes dynamically with time and power input, characterizing the resistance of heat in the system and affecting the response speed and change amplitude of the system temperature.

[0035] The dynamic thermal resistance is constructed by fitting the following formula:

[0036] ;

[0037] Among them, the first part Describes the time dependence of thermal resistance, a is the initial response amplitude of thermal resistance, when time t=0, the initial value of thermal resistance is a, Describes the exponential decay characteristics of thermal resistance over time. b is the thermal resistance decay coefficient. If b is large, the thermal resistance will decay rapidly. If b is small, the thermal resistance decays slowly. This part indicates that as time goes by, the thermal resistance gradually decreases from the initial value, simulating the process of heat transfer to a stable state.

[0038] Part 2 Describes the power dependence of thermal resistance. c is the dynamic response coefficient of power input. A larger c indicates a greater impact of power input on thermal resistance, and a smaller c indicates a smaller impact of power input. is the power input. This part indicates that when the power input increases, the thermal resistance also changes, reflecting the relationship between thermal resistance and power input.

[0039] Combining these two parts, the dynamic thermal resistance formula is obtained to predict the dynamic change of thermal resistance with time and power input. Specifically, at the beginning of operation, the thermal resistance is dominated by the initial response amplitude a of the thermal resistance, reflecting the initial heat transfer resistance. As time goes by, the thermal resistance gradually decays to a lower level. At the same time, the power input The change of will adjust the thermal resistance value in real time to reflect the dynamic impact of load and power conditions on heat transfer. In this way, the entire dynamic thermal network model can more accurately simulate the complex characteristics of the heat transfer process and provide support for the temperature prediction of the system.

[0040] Furthermore, the extracting features of the corresponding feature data using the feature extraction subnetwork and training the neural network according to the feature extraction results also includes:

[0041] Adaptive feature clustering is performed on the feature extraction results, and feature fusion is performed based on the adaptive feature clustering results to construct fused features; a hybrid attention channel is called: A: the hybrid attention channel is used to perform temporal dynamic analysis of a single feature of the feature extraction result, key points of feature temporal changes are extracted, and a weighted output of intra-feature attention is established according to the key points of feature temporal changes; B: the hybrid attention channel is used to perform feature interaction importance analysis of the feature extraction result, and an interactive output of attention between features is established according to the analysis result; C: the hybrid attention channel is used to perform long-term dependency enhancement of the feature extraction result, and a temporal output of temporally enhanced attention is established according to the long-term dependency enhancement result; attention fusion is performed based on the weighted output of intra-feature attention, the interactive output of attention between features, and the temporal output of temporally enhanced attention, and a comprehensive feature weight is established; neural network training is performed according to the fused features, feature extraction results, and comprehensive feature weights.

[0042] Receive the feature extraction results from the feature extraction subnetwork, including high-dimensional vectors of features such as on-state voltage drop and collector current, and use an adaptive clustering algorithm, such as the K-means clustering method, to dynamically adjust the number or position of cluster centers. The goal is to group the features according to their similarity so that the features in the same group have similar patterns. After clustering, multiple feature clusters are generated, and each cluster contains similar feature representations.

[0043] The multiple feature clusters after clustering are fused. Specifically, operations are performed on the feature vectors in the same cluster, such as summation and weighted average, to generate cluster center features. All cluster center features are integrated to form the final fused feature vector. The fused feature vector is more representative and can effectively reduce the impact of noise and redundant features, providing high-quality input for subsequent analysis.

[0044] The hybrid attention channel is called, which integrates temporal dynamic analysis, feature interaction analysis and long-term dependency enhancement to improve the quality of feature extraction and prediction ability.

[0045] Temporal dynamic analysis is used to analyze the trend of a single feature changing over time and to extract key change points from the time series. Specifically, a single feature such as on-state voltage drop, collector current, etc. may change significantly over time, and some change points are crucial to prediction. The goal of temporal dynamic analysis is to find these important time points through the attention mechanism. Specifically, for the input time series data, the attention weight of each time point is calculated and weighted by the attention weight. The larger the weight of the time point, the greater the contribution of its feature to the prediction result. Finally, the weighted result of the attention within the feature is output, which represents the dynamic influence of a single feature in the time dimension.

[0046] Feature interaction importance analysis is used to analyze the interaction relationship between different features, explore the interaction between features and their importance to the prediction results. Specifically, in many scenarios, a single feature may not be sufficient to explain the prediction target. For example, the on-state voltage drop and collector current have a synergistic effect in certain situations. Through the mixed attention mechanism, the interaction weights between features are calculated to measure the impact of one feature on another feature, and a feature interaction weight matrix is ​​generated to represent the interaction strength between features. According to the matrix, the interaction output of the attention between features is established to reflect the synergistic relationship between features.

[0047] Long-term dependency enhancement is used to enhance the model's ability to utilize historical data by analyzing long-term dependencies in time series. Specifically, the performance of certain features at the current moment may be affected by the state at earlier time points. For example, early changes in load fluctuations may have a significant impact on subsequent junction temperatures. The dependency weights of the current time point and multiple time points in the past are calculated through a hybrid attention mechanism, where closer time points have higher weights and more distant time points have weights that decay over time. The temporal output of temporal enhanced attention is established based on the analysis results.

[0048] Attention fusion combines the weighted output of intra-feature attention, the interactive output of inter-feature attention, and the temporal output of temporal enhanced attention. For example, adjustable hyperparameters are preset, which represent the weight of each attention mechanism. The adjustable hyperparameters are used to perform weighted summation on them to generate a globally representative comprehensive feature weight from multiple dimensions. The comprehensive feature weight is a multidimensional vector that integrates the contributions of different attention mechanisms and is used to guide subsequent neural network training.

[0049] The fusion features and feature extraction results are used as input data, and the actual IGBT junction temperature is used as output data for neural network training. By inputting fusion features and feature weights, the neural network can efficiently learn complex nonlinear relationships during training, thereby achieving accurate IGBT junction temperature prediction. This process combines the advantages of deep learning and attention mechanism, ensuring the sensitivity and accuracy of the model to multi-dimensional input data, and providing reliable support for practical engineering applications.

[0050] Furthermore, the step of establishing a weighted output of intra-feature attention according to the key points of feature temporal changes also includes:

[0051] The weighted output of the intra-feature attention is calculated by the formula as follows:

[0052] ;

[0053] ;

[0054] in, represents the attention weight of feature i at time t, The query matrix representing feature i, represents the representation of feature i at time t, The weighted output of the intra-feature attention representing feature i.

[0055] Specifically, the weighted output of the intra-feature attention is calculated by the formula as follows:

[0056] ;

[0057] ;

[0058] Among them, in the attention weight calculation, the dot product Calculate the relevance score of feature i at time t, and then convert it into a probability distribution through the softmax function to clarify the relative importance of each time point; in the weighted output calculation of the intra-feature attention, according to the attention weight Assign importance weights to the values ​​of feature i at different time points, and finally output a time-weighted comprehensive feature representation .

[0059] In general, the attention mechanism automatically learns the relative importance of feature i at different time points For example, in the change of on-state voltage drop over time, some moments may better reflect load changes or abnormal behavior, and then the representation of feature i at all time points Perform weighted summation to generate a more representative comprehensive representation of the time series ,This attention mechanism enables the model to pay more attention to the time points that are critical to the ,prediction task, while reducing the interference of irrelevant time points and ,enhancing the feature expression capability.

[0060] Furthermore, the method of performing feature interaction importance analysis of feature extraction results using the hybrid attention channel and establishing an interactive output of attention between features according to the analysis results also includes:

[0061] The feature interaction weights are established based on the feature interaction importance analysis as follows:

[0062] ;

[0063] in, Characterizes the interaction weight between feature i and feature j, Representation query matrix, and are the representation vectors of the i-th feature and the j-th feature respectively.

[0064] All feature interaction weights are weighted and the interactive output of attention between features is established as follows:

[0065] ;

[0066] in, Represents the interactive output of attention between features, Characteristic pair joint representation of .

[0067] Specifically, feature interaction importance analysis is used to capture the synergistic relationship between features. For example, on-state voltage drop and collector current may have a strong correlation under certain operating conditions, and this interaction is crucial to the prediction results.

[0068] The feature interaction weights are established based on the feature interaction importance analysis as follows:

[0069] ;

[0070] Among them, the vector representations of the two features are interacted item by item through the outer product operation to generate a high-dimensional vector that represents their association relationship in different dimensions. By learning to assign weights to the interactions between different features, that is, giving higher weights to strong interactions and downgrading weak interactions, the higher the final weight, the greater the contribution of the interaction of the feature pair to the prediction. The interaction weights of all feature pairs are normalized through softmax, thereby efficiently capturing the correlation between features and providing richer information for the prediction task.

[0071] All feature interaction weights are weighted and the interactive output of attention between features is established as follows:

[0072] ;

[0073] This formula describes how to use feature interaction weights For all feature pairs Weighted to generate interactive output of attention between features , which quantifies the synergistic relationships between features and integrates them into a joint representation, where Calculate feature pairs The joint representation reflects the combined characteristics of the two, through the interaction weight The joint representation of feature pairs is weighted so that feature pairs with high correlation contribute more, while feature pairs with low correlation contribute less. The weighted joint representation of is summed to generate a comprehensive vector , represents the interactive relationship of the entire feature set. The summary result not only contains the unique information of each feature pair, but also integrates the global synergistic relationship between features.

[0074] Furthermore, the method of performing long-term dependency enhancement of feature extraction results using a hybrid attention channel and establishing a temporal output of temporally enhanced attention according to the long-term dependency enhancement results further includes:

[0075] The time output of time-enhanced attention is established through the formula:

[0076] ;

[0077] ;

[0078] in, represents the enhanced features at time point t, s represents the time scale, k represents the time offset, is the time offset weight, which represents the degree of dependence between time point t and historical time point t-k at time scale s. , Characterize the time decay factor, based on different time scale s configurations, Characterizes the representation of feature i at time t-k, Represents the attention weight at time point t, , Representing time enhances the temporal output of attention.

[0079] Specifically, the time series information is enhanced through the attention mechanism to capture the dependency between the current time point and the historical time point. The formula is divided into two main parts, namely, calculating the enhanced features of each time point t , and synthesize the enhanced features of each time point to generate the time output of time-enhanced attention , the formula is as follows:

[0080] ;

[0081] ;

[0082] In the calculation of enhanced features at time point t, is the time offset weight, which is used to measure the correlation between time point t and historical time point t-k, and is defined as , the influence of time scale s on the degree of dependence is controlled by the time decay factor, and the enhanced features of time point t are finally calculated , Combining the information of the current time point with the dependencies of historical time points allows the model to capture dependencies in different time ranges and thus model complex temporal relationships.

[0083] Time Enhancement of Attention Output During the calculation, the enhanced features for all time points t Perform weighted summation to obtain the final time-enhanced attention output , It is a global enhanced representation of time series, which integrates the information of current and historical time points and can more comprehensively describe the dynamic changes of time series. This formula captures the dynamic dependencies of time series through time offset weights and attention mechanisms, providing a powerful tool for complex time series modeling.

[0084] Furthermore, if Figure 2 As shown, after the real-time parameter set is preprocessed, a sample data set is established, which also includes:

[0085] The real-time parameter set is cleaned to remove outliers and noise data, and missing values ​​are filled in smoothly. The real-time parameter set after data cleaning is normalized and the data normalization results are established. The normalization results are segmented into time series according to the load fluctuation frequency, and a sample data set is established based on the time series segmentation results.

[0086] Data cleaning is performed on the real-time parameter set. The goal of data cleaning is to improve the quality and consistency of the data, remove outliers and noise data, fill in missing values, and ensure data integrity. Specifically, outliers refer to extreme data points that are beyond the normal range or do not meet expectations. They may be caused by sensor failure or acquisition errors. The reasonable range of each parameter can be defined based on prior knowledge or historical data, and values ​​outside the range are marked as outliers and removed; noise is the inevitable random interference in the acquisition process, which can be removed by smoothing methods to smooth the data and reduce random fluctuations introduced during the acquisition process; missing values ​​can be filled using the mean filling method, that is, the local mean of the data (such as the mean of the sliding window) is used for smoothing to fill in the gaps in the collected data and avoid modeling errors due to missing values.

[0087] Data normalization is to scale the eigenvalues ​​of different ranges to the same range to eliminate dimensional differences and improve training efficiency. The minimum-maximum normalization method can be used to normalize the real-time parameter set after data cleaning, scale the data to the range of [0,1], and establish the data normalization processing result.

[0088] The normalized data is divided into several time periods according to the load fluctuation frequency, such as every 1 second and every 5 seconds, to reflect the dynamic changes of the load. Each time series segment after segmentation contains multi-dimensional input features for subsequent modeling. The time series data is used as input, and the target value corresponding to the time period (i.e., IGBT junction temperature) is used as output to establish a sample data set. This ensures that the model can effectively learn the time dynamic characteristics and the impact of load fluctuations on the system, thereby providing a solid foundation for accurately predicting the IGBT junction temperature.

[0089] Furthermore, the method further comprises:

[0090] Abnormal fluctuations of the junction temperature time series are identified to establish a first abnormal result; a junction temperature trend is predicted based on the junction temperature time series to establish a second abnormal result; an early warning signal is matched according to the first abnormal result and the second abnormal result, and the early warning signal is reported.

[0091] Abnormal fluctuations in the junction temperature time series include abnormal sudden increase, abnormal sudden drop, excessive fluctuation, etc., among which, abnormal sudden increase is manifested as a rapid increase in junction temperature in a short period of time, exceeding the set threshold; abnormal sudden drop is manifested as a rapid decrease in junction temperature in a short period of time; excessive fluctuation is manifested as the change in junction temperature over time is significantly higher than the normal range. Based on the preset threshold, these abnormal fluctuations are identified and the first abnormal result is established, including the abnormal time point and the corresponding junction temperature value.

[0092] The junction temperature trend is predicted based on the junction temperature time series. For example, after the junction temperature time series is processed for stationarity, an autoregressive integral sliding average model is used for trend prediction to analyze the trend of the junction temperature over time and predict whether the future junction temperature change will exceed the safety range. The second abnormal result is established based on the trend prediction value and whether it exceeds the safety threshold.

[0093] Combined with the first abnormal result and the second abnormal result, the warning signal is matched. Specifically, if a fluctuation abnormality is detected at the current time point according to the first abnormal result, a short-term warning is triggered to indicate the impending risk; if it is predicted according to the second abnormal result that the junction temperature will exceed the safety range in the future, a long-term warning is triggered to indicate potential future risks.

[0094] The warning signal is reported, for example, the warning information is displayed in real time on the monitoring interface, or an audible and visual alarm or control signal is triggered, so as to realize real-time monitoring and intelligent warning of the operating status of the IGBT module, providing effective guarantee for the safe operation of the system.

[0095] In summary, the inverter IGBT junction temperature prediction method based on neural network provided in the embodiment of the present application has the following technical effects:

[0096] The use of neural network models, including multi-input layer networks and attention mechanisms, can efficiently capture the complex nonlinear relationship between input parameters and IGBT junction temperature, improve the accuracy of prediction, and establish multi-operating condition simulation data sets. The model can be widely adapted to different operating environments and is suitable for a variety of practical engineering applications. Compared with traditional physical measurement methods or complex thermal model methods, this method only needs to collect a small number of parameters such as on-state voltage drop and collector current for prediction, avoiding the need for complex hardware or experimental equipment. By cleaning, normalizing and time series segmenting the real-time parameter set, a high-quality data set is generated, which is convenient for neural network learning and using MAT LAB / Simulink software is used to establish a thermal network model, which simplifies the modeling and training process and can be directly applied to actual projects. The dynamic thermal resistance formula is used to accurately simulate the time dynamics and power input response of the IGBT junction temperature, which improves the authenticity of the heat transfer simulation. The multi-input layer structure allows independent modeling of features from different sources, improving the model's ability to understand multi-dimensional data. At the same time, the feature extraction sub-network is used to optimize feature expression and improve prediction accuracy. The IGBT junction temperature prediction model is used to predict data and establish a junction temperature time series, which can not only monitor the junction temperature in real time, but also predict future trends, providing guidance for equipment maintenance and optimized operation.

[0097] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting junction temperature of IGBT of inverter based on neural network, characterized in that: The method comprises: After establishing digital communication with the subway traction drive system, the circuit structure and operating parameters of the subway traction drive system are obtained, and the dynamic thermal network model of the subway traction drive system inverter is constructed using MATLAB / Simulink software; Establish a simulation condition set of the subway traction drive system, perform simulation based on the dynamic thermal network model with the simulation condition set and environmental parameters, collect real-time parameter sets, the real-time parameter sets include on-state voltage drop, collector current, ambient temperature, radiator temperature and load fluctuation, and establish a sample data set after preprocessing the real-time parameter sets; Establish a neural network with multiple input layers, and input the sample data set into the neural network with multiple input layers for training, and construct an IGBT junction temperature prediction model, wherein the multiple input layers include an on-state voltage drop input layer, a collector current input layer, an ambient temperature input layer, a heat sink temperature input layer, and a load fluctuation input layer, and each input layer is mapped with a feature extraction sub-network. Before the neural network performs training, the feature extraction sub-network is used to extract features of the corresponding feature data, and the neural network is trained according to the feature extraction results; Use the IGBT junction temperature prediction model to predict data and establish a junction temperature time series; The method of using MATLAB / Simulink software to construct a dynamic thermal network model of a subway traction drive system inverter also includes: The dynamic thermal network model includes static thermal resistance and dynamic thermal resistance, and the dynamic thermal resistance is constructed by fitting the following formula: ； in, Characterize the dynamic thermal resistance, over time t and power input a is the initial response amplitude of thermal resistance, b is the thermal resistance attenuation coefficient, t represents the time variable, c is the dynamic response coefficient of power input, is the power input.

2. The method for predicting junction temperature of IGBT of inverter based on neural network according to claim 1, characterized in that: The feature extraction sub-network is used to extract features of the corresponding feature data, and the neural network training is performed according to the feature extraction results, and further includes: Performing adaptive feature clustering on the feature extraction results, performing feature fusion based on the adaptive feature clustering results, and constructing fusion features; Call the mixed attention channel: A: Use the hybrid attention channel to perform temporal dynamic analysis of a single feature of the feature extraction result, extract the key points of feature temporal change, and establish the weighted output of intra-feature attention based on the key points of feature temporal change; B: Use the hybrid attention channel to perform feature interaction importance analysis of the feature extraction results, and establish the interactive output of the attention between features based on the analysis results; C: Use the hybrid attention channel to perform long-term dependency enhancement of the feature extraction results, and build the temporal output of temporal enhanced attention based on the long-term dependency enhancement results; Attention fusion is performed based on the weighted output of intra-feature attention, the interactive output of inter-feature attention, and the temporal output of temporal enhanced attention to establish comprehensive feature weights; The neural network is trained based on the fusion features, feature extraction results, and comprehensive feature weights.

3. The method for predicting junction temperature of IGBT of inverter based on neural network as claimed in claim 2, characterized in that: The step of establishing a weighted output of intra-feature attention according to the key points of feature temporal changes also includes: The weighted output of the intra-feature attention is calculated by the formula as follows: ； ； in, represents the attention weight of feature i at time t, The query matrix representing feature i, Characterizes the representation of feature i at time t, The weighted output of the intra-feature attention representing feature i.

4. The method for predicting junction temperature of IGBT of inverter based on neural network as claimed in claim 3, characterized in that: The method further comprises: performing feature interaction importance analysis of feature extraction results using a hybrid attention channel, and establishing an interactive output of attention between features according to the analysis results, and further comprising: The feature interaction weights are established based on the feature interaction importance analysis as follows: ； in, Characterizes the interaction weight between feature i and feature j, Representation query matrix, and are the representation vectors of the i-th feature and the j-th feature respectively; All feature interaction weights are weighted and the interactive output of attention between features is established as follows: ； in, Represents the interactive output of attention between features, Characteristic pair joint representation of .

5. The method for predicting junction temperature of IGBT of inverter based on neural network as claimed in claim 4, characterized in that: The method further comprises: performing long-term dependency enhancement of feature extraction results using a hybrid attention channel, and establishing a temporal output of temporally enhanced attention according to the long-term dependency enhancement results. The time output of time-enhanced attention is established through the formula: ； ； in, represents the enhanced features at time point t, s represents the time scale, k represents the time offset, is the time offset weight, which represents the degree of dependence between time point t and historical time point t-k at time scale s. , Characterize the time decay factor, based on different time scale s configurations, Characterizes the representation of feature i at time t-k, Represents the attention weight at time point t, , Representing time enhances the temporal output of attention.

6. The method for predicting junction temperature of IGBT of inverter based on neural network according to claim 1, characterized in that: After the real-time parameter set is preprocessed, a sample data set is established, which further includes: Perform data cleaning on the real-time parameter set to remove outliers and noise data, and smoothly fill in missing values; The real-time parameter set after data cleaning is normalized and the data normalization results are established. The normalization results are segmented into time series according to the load fluctuation frequency, and a sample data set is established based on the time series segmentation results.

7. The method for predicting junction temperature of IGBT of inverter based on neural network according to claim 1, characterized in that: The method further comprises: Performing abnormal fluctuation identification on the junction temperature time series to establish a first abnormal result; Based on the junction temperature time series, the junction temperature trend is predicted and the second abnormal result is established; A warning signal is matched according to the first abnormal result and the second abnormal result, and the warning signal is reported.

Images