A method for monitoring junction temperature of an IGBT module of a high-voltage frequency converter

Abstract

The application relates to the technical field of high-power high-voltage frequency converters, and discloses a high-voltage frequency converter IGBT module junction temperature monitoring method, which comprises the following steps: constructing an electro-thermal coupling model by adopting a Foster type equivalent thermal network based on the electrical characteristics and structural parameters of an IGBT module; simulating under different operating conditions based on the electro-thermal coupling model, obtaining multiple groups of IGBT module junction temperature data and corresponding electrical data, and constructing a training sample set; constructing a long short-term memory recurrent neural network based on a macro-micro attention mechanism, and training the long short-term memory recurrent neural network by using the training sample set to obtain an IGBT module junction temperature monitoring model; and inputting real-time electrical data into the trained IGBT module junction temperature monitoring model to output an IGBT junction temperature monitoring value. The application improves the generalization and engineering applicability of the model while ensuring the junction temperature monitoring precision.

Description

Technical Field

[0001] This invention relates to the field of high-power high-voltage frequency converter technology, specifically to a method for monitoring the junction temperature of IGBT modules in high-voltage frequency converters. Background Technology

[0002] Insulated-Gate Bipolar Transistor (IGBT) power modules integrate the advantages of insulated-gate field-effect transistors and power transistors, offering high switching frequencies and low drive power, making them a key component of high-power, high-voltage frequency converters. This power module has a "sandwich" structure, with its internal cross-section consisting of, from top to bottom, the IGBT chip and diode chip, chip solder layer, ceramic copper-clad laminate (including upper copper layer, insulating ceramic layer, and lower copper layer), substrate solder layer, and base plate—a total of seven layers. According to industry surveys, approximately 34% of failures are caused by power devices, and 60% of device failures are temperature-related. In practical applications, high-power frequency converter IGBT modules typically operate at kHz-level switching frequencies and high power, generating significant heat loss during operation, causing rises and fluctuations in chip junction temperature, which can lead to device failure in severe cases. Therefore, effective monitoring of the junction temperature is crucial for ensuring stable system operation and extending device service life.

[0003] There are four methods for monitoring IGBT module junction temperature: optical measurement, physical contact, thermal modeling, and temperature-sensitive electrical parameter methods. However, these methods still have significant limitations in practical applications: 1) The thermal network modeling method requires the establishment of a highly complex electrothermal coupling model, making parameter extraction difficult and heavily dependent on the physical structure; 2) Traditional temperature-sensitive electrical parameter methods often require the introduction of additional sensors or detection circuits, which not only increases system complexity and cost but also easily introduces interference. Therefore, existing methods struggle to simultaneously meet the requirements of engineering practicality, high accuracy, and real-time performance, necessitating a junction temperature monitoring method for IGBT modules that combines high accuracy with ease of engineering use. Summary of the Invention

[0004] To address the aforementioned shortcomings in the prior art, this invention provides a method for monitoring the junction temperature of IGBT modules in high-voltage frequency converters.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] A method for monitoring the junction temperature of IGBT modules in a high-voltage frequency converter includes the following steps:

[0007] Based on the electrical characteristics and structural parameters of the IGBT module, an electrothermal coupling model is constructed using a Foster-type equivalent thermal network.

[0008] Based on the electrothermal coupling model, simulations were performed under different operating conditions to obtain multiple sets of IGBT module junction temperature data and corresponding electrical data, and a training sample set was constructed.

[0009] A long short-term memory recurrent neural network based on macro-micro attention mechanism is constructed, which includes macro attention mechanism and micro attention mechanism. The macro attention mechanism is used to adaptively focus on key time period features in the time dimension, and the micro attention mechanism is used to adaptively weight electrical features in the feature dimension. The junction temperature monitoring model of IGBT module is obtained by training with training sample set.

[0010] The trained IGBT module junction temperature monitoring model is used as input for real-time electrical data and outputs IGBT junction temperature monitoring values.

[0011] Optionally, based on the electrical characteristics and structural parameters of the IGBT module, an electrothermal coupling model is constructed using a Foster-type equivalent thermal network, including:

[0012] Based on the electrical characteristics and structural parameters of IGBT modules, a Foster-type equivalent thermal network model is established to simulate the heat transfer path from the junction to the environment using a thermal resistance-thermal capacity network.

[0013] Construct a power loss model to calculate the conduction loss and switching loss of the IGBT module;

[0014] Using collector-emitter voltage, collector current, switching frequency, modulation ratio, and ambient temperature as model inputs, and IGBT junction temperature as model output, electrothermal interaction is achieved through the coupling of the electrical system and thermal network in Simulink.

[0015] Optionally, when extracting the thermal network parameters of the Foster-type equivalent thermal network model, the thermal resistance and heat capacity parameters of each order of the Foster network are obtained by fitting the transient thermal resistance curve, and the multi-level thermal resistance and heat capacity values ​​of the junction, shell heat sink, and heat sink environment are determined.

[0016] Optionally, when calculating the conduction loss of the IGBT module, the conduction loss is calculated based on the functional relationship between the collector current and the saturation voltage drop, and the conduction voltage drop at different operating points is obtained by interpolation of the characteristic curve.

[0017] When calculating the switching losses of an IGBT module, a polynomial fitting method is used to establish an estimation model of how the losses change with current, voltage, and temperature, based on the relationship between switching energy and collector current.

[0018] Optionally, based on the electrothermal coupling model, simulations are performed under different operating conditions to obtain multiple sets of IGBT module junction temperature data and corresponding electrical data, constructing a training sample set, including:

[0019] Dynamic simulations were performed under various operating conditions, including no-load, half-load, full-load and other typical load states, using an electrothermal coupling model. Multiple sets of junction temperature timing data of IGBT modules and their corresponding electrical characteristic parameters were obtained, and the obtained data were used to form a training sample set.

[0020] Optionally, the long short-term memory recurrent neural network based on the macro-micro attention mechanism calculates the mean of the input electrical data by time period, and calculates the macro attention weight of each time period based on the macro query vector to amplify the features of important time periods; then, for the input data and recurrent data at different times, the micro attention weight is calculated using the micro query vector to enhance the sensitivity to key instantaneous features; finally, the weights of the input and recurrent data are amplified simultaneously based on the macro and micro weights, and the weighted features are input into the long short-term memory recurrent neural network for the fusion of long and short-term features.

[0021] Optionally, the mean of the input electrical data is calculated by time period, and the macro attention weight for each time period is calculated based on the macro query vector to amplify the features of important time periods, including:

[0022] The electrical data at continuous time points are divided into several time periods. The average value of each parameter is calculated for all electrical data in each time period. The average value of each parameter is used to construct a macroscopic feature matrix.

[0023] Generate a query vector based on the hidden state of the previous time period, and match the macro feature matrix of each time period with the generated query vector to calculate the relevance score;

[0024] The relevance scores for all time periods are normalized to obtain the macro attention weight for each time period.

[0025] Optionally, micro-attention weights are calculated using micro-query vectors for both the time-based input data and the recursive data to enhance sensitivity to key instantaneous features, including:

[0026] Construct a micro-feature matrix for each time point within each time period;

[0027] Generate a query vector based on the hidden state of the previous time point, and match the micro-feature matrix of each time point with the generated query vector to calculate the relevance score;

[0028] The relevance scores at all time points are normalized to obtain the micro-attention weights at each time point.

[0029] Optionally, the input and recursive data weights are simultaneously amplified based on macro and micro weights, and the weighted features are input into a long short-term memory recursive neural network to fuse long and short-term features, including:

[0030] The input gate and forget gate are calculated by weighting the input data, the hidden state of the previous time step, and the attention weights.

[0031] Based on the weighted new input and the weighted past state, generate new information candidate content;

[0032] The old cell state is forgotten according to the forget gate ratio, and new candidate cell states are added according to the input gate ratio to form a new cell state;

[0033] Control the output ratio from the cell state to the hidden state, calculate the output gate, and update the hidden state.

[0034] The final hidden state is mapped through a fully connected layer to obtain the final junction temperature monitoring value.

[0035] Optionally, when training the IGBT module junction temperature monitoring model, a thermal network model is used to calculate the true junction temperature, specifically:

[0036] ;

[0037] in, For the junction temperature, For the power loss of IGBT, The junction-to-shell thermal resistance of the frequency converter. This refers to the thermal resistance from the inverter's casing to the heat sink. The thermal resistance from the inverter's heat sink to the environment. The ambient temperature.

[0038] The present invention has the following beneficial effects:

[0039] This invention first establishes the electrothermal coupling relationship of IGBT modules using a Foster thermal network model to obtain IGBT junction temperature and electrical parameter samples under multiple operating conditions. Then, it constructs a recurrent neural network model integrating macro- and micro-attention mechanisms to achieve junction temperature monitoring. This method only requires the acquisition of five types of signals: collector-emitter voltage, collector current, switching frequency, modulation ratio, and ambient temperature. It eliminates the need for additional sensors or complex thermal parameter identification, and its data-driven approach avoids the multi-parameter coupling and model mismatch problems found in traditional methods. The macro- and micro-attention mechanisms can adaptively capture key information from both the time and feature dimensions, enhancing the ability to characterize junction temperature trends. While ensuring junction temperature monitoring accuracy, it improves the model's generalization and engineering applicability, making it suitable for real-time status monitoring and health management of IGBT modules in high-power high-voltage frequency converters. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of a method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to the present invention.

[0041] Figure 2This is a schematic diagram of the electrothermal coupling model of the IGBT module in this invention;

[0042] Figure 3 This is a schematic diagram of the IGBT black box model in this invention;

[0043] Figure 4 This is a schematic diagram of the macro-micro attention mechanism in this invention;

[0044] Figure 5 This is a diagram of the hidden layer structure of the long short-term memory recurrent neural network in this invention;

[0045] Figure 6 This is a flowchart of the macro-micro attention long short-term memory recurrent neural network in this invention. Detailed Implementation

[0046] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0047] Given the following issues with existing high-power high-voltage frequency converter IGBT module junction temperature monitoring:

[0048] 1) Traditional thermal network modeling methods require the establishment of highly complex electrothermal coupling models, which are difficult to extract parameters and heavily dependent on physical structures;

[0049] 2) Traditional temperature-sensitive electrical parameter methods require the introduction of additional sensors or detection circuits, which increases system complexity and cost and is prone to interference.

[0050] This invention addresses the challenge of accurate, real-time, and non-invasive junction temperature monitoring of IGBT modules in high-power high-voltage frequency converters by proposing a junction temperature monitoring method based on a macro-micro attention long short-term memory recurrent neural network. This method requires only the acquisition of five types of signals, eliminating the need for additional sensors or complex thermal parameter identification. Through data-driven and attention-based adaptive capture of key information, it ensures both accuracy and real-time performance in junction temperature monitoring.

[0051] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for monitoring the junction temperature of a high-voltage inverter IGBT module, comprising the following steps S1 to S4:

[0052] S1. Based on the electrical characteristics and structural parameters of the IGBT module, an electrothermal coupling model is constructed using a Foster-type equivalent thermal network;

[0053] In an optional embodiment of the present invention, step S1, based on the electrical characteristics and structural parameters of the IGBT module, constructs an electrothermal coupling model using a Foster-type equivalent thermal network, including:

[0054] Based on the electrical characteristics and structural parameters of IGBT modules, a Foster-type equivalent thermal network model is established to simulate the heat transfer path from the junction to the environment using a thermal resistance-thermal capacity network.

[0055] Construct a power loss model to calculate the conduction loss and switching loss of the IGBT module;

[0056] Using collector-emitter voltage, collector current, switching frequency, modulation ratio, and ambient temperature as model inputs, and IGBT junction temperature as model output, electrothermal interaction is achieved through the coupling of the electrical system and thermal network in Simulink.

[0057] When extracting the thermal network parameters of the Foster-type equivalent thermal network model, the thermal resistance and thermal capacity parameters of each order of the Foster network are obtained by fitting the transient thermal resistance curve, and the multi-level thermal resistance and thermal capacity values ​​of the junction, shell heat sink, and heat sink environment are determined.

[0058] When calculating the conduction loss of an IGBT module, the conduction loss is calculated based on the functional relationship between collector current and saturation voltage drop. The conduction voltage drop at different operating points is obtained by interpolation of the characteristic curve.

[0059] When calculating the switching losses of an IGBT module, a polynomial fitting method is used to establish an estimation model of how the losses change with current, voltage, and temperature, based on the relationship between switching energy and collector current.

[0060] This embodiment is based on the electrical characteristics and structural parameters of the IGBT module. Using MATLAB / Simulink software, an electrothermal coupling model of the high-power high-voltage inverter IGBT module is constructed using a Foster-type equivalent thermal network. The electrothermal coupling model is built using a fourth-order Foster-type equivalent thermal network structure. Its thermal resistance and thermal capacity parameters are obtained by fitting the transient thermal impedance curve provided in the IGBT module datasheet, used to simulate the dynamic response of the IGBT junction temperature under changes in electrical load and ambient temperature. The input variables of the model include collector-emitter voltage, collector current, switching frequency, modulation ratio, and ambient temperature, while the output variable is the junction temperature of the IGBT chip.

[0061] like Figure 2 As shown, the process of constructing the electrothermal coupling model of the IGBT module in this embodiment is as follows:

[0062] (1) Establish an electrothermal coupling model framework. Based on the electrical characteristics and structural parameters of the IGBT module, build a Foster-type equivalent thermal network model in the MATLAB / Simulink environment. The model takes the thermal resistance-thermal capacity network as the core to simulate the heat transfer path from the junction to the environment.

[0063] (2) Extract thermal network parameters. Based on the transient thermal impedance curve provided in the IGBT module product datasheet, fit and obtain the thermal resistance and thermal capacity parameters of each order of the Foster network, and determine the multi-level thermal resistance and thermal capacity values ​​of the junction, shell heat sink, heat sink environment, etc.

[0064] (3) Construct a power loss model. The conduction loss is calculated based on the functional relationship between collector current and saturation voltage drop. The conduction voltage drop at different operating points is obtained by interpolation of the characteristic curve output from the datasheet. The switching loss is based on the relationship between switching energy and collector current provided in the datasheet. A polynomial fitting method is used to establish an estimation model of the loss as a function of current, voltage and temperature.

[0065] (4) Define the model input and output variables. The model input variable is the loss value, and the output variable is the IGBT junction temperature. Electrothermal interaction is achieved through the coupling of the electrical system and the thermal network in Simulink.

[0066] like Figure 3 As shown, this embodiment is based on the analysis of the electrothermal coupling mechanism, and determines that the input of the IGBT black box model is five easily monitored electrical variables: collector-emitter voltage, collector current, switching frequency, modulation ratio and ambient temperature; the output is the IGBT junction temperature, and a nonlinear mapping relationship of multiple inputs and single output is established.

[0067] S2. Based on the electrothermal coupling model, simulations are performed under different operating conditions to obtain multiple sets of IGBT module junction temperature data and corresponding electrical data, and a training sample set is constructed.

[0068] In an optional embodiment of the present invention, step S2 is based on an electrothermal coupling model, performing simulations under different operating conditions to obtain multiple sets of IGBT module junction temperature data and corresponding electrical data, and constructing a training sample set, including:

[0069] Dynamic simulations were performed under various operating conditions, including no-load, half-load, full-load and other typical load states, using an electrothermal coupling model. Multiple sets of junction temperature timing data of IGBT modules and their corresponding electrical characteristic parameters were obtained, and the obtained data were used to form a training sample set.

[0070] This embodiment balances simulation accuracy and computational efficiency by configuring the solver and simulation parameters, setting the simulation duration and sampling step size in Simulink; running multi-condition electrothermal co-simulation, running simulation under different load conditions (no load, half load, full load), recording IGBT junction temperature dynamic response data and corresponding electrical parameters, and generating a standard sample set for neural network training.

[0071] S3. Construct a long short-term memory recurrent neural network based on macro-micro attention mechanism, which includes macro attention mechanism and micro attention mechanism. Macro attention mechanism is used to adaptively focus on key time period features in time dimension, and micro attention mechanism is used to adaptively weight electrical features in feature dimension; and use training sample set to train IGBT module junction temperature monitoring model.

[0072] In an optional embodiment of the present invention, step S3 calculates the mean of the input electrical data by time period based on the long short-term memory recurrent neural network with macro-micro attention mechanism, and calculates the macro attention weight of each time period based on macro query vector to amplify the features of important time periods; then, for the input data and recurrent data at different times, the micro attention weight is calculated using micro query vector to enhance the sensitivity to key instantaneous features; finally, the weights of the input and recurrent data are amplified simultaneously based on macro and micro weights, and the weighted features are input into the long short-term memory recurrent neural network for the fusion of long and short-term features.

[0073] Specifically, the input electrical data is averaged over time periods, and macro-level attention weights for each time period are calculated based on the macro-level query vector to amplify the characteristics of important time periods, including:

[0074] The electrical data at continuous time points are divided into several time periods. The average value of each parameter is calculated for all electrical data in each time period. The average value of each parameter is used to construct a macroscopic feature matrix.

[0075] Generate a query vector based on the hidden state of the previous time period, and match the macro feature matrix of each time period with the generated query vector to calculate the relevance score;

[0076] The relevance scores for all time periods are normalized to obtain the macro attention weight for each time period.

[0077] For both time-based input data and recursive data, micro-level attention weights are calculated using micro-query vectors to enhance sensitivity to key instantaneous features, including:

[0078] Construct a micro-feature matrix for each time point within each time period;

[0079] Generate a query vector based on the hidden state of the previous time point, and match the micro-feature matrix of each time point with the generated query vector to calculate the relevance score;

[0080] The relevance scores at all time points are normalized to obtain the micro-attention weights at each time point.

[0081] Based on the simultaneous amplification of input and recursive data weights using macro and micro weights, the weighted features are input into a long short-term memory recursive neural network for the fusion of long and short-term features, including:

[0082] The input gate and forget gate are calculated by weighting the input data, the hidden state of the previous time step, and the attention weights.

[0083] Based on the weighted new input and the weighted past state, generate new information candidate content;

[0084] The old cell state is forgotten according to the forget gate ratio, and new candidate cell states are added according to the input gate ratio to form a new cell state;

[0085] Control the output ratio from the cell state to the hidden state, calculate the output gate, and update the hidden state.

[0086] The final hidden state is mapped through a fully connected layer to obtain the final junction temperature monitoring value.

[0087] The long short-term memory recurrent neural network based on macro-micro attention mechanism constructed in this embodiment is a multi-input single-output structure; wherein, the input layer contains 5 neurons and the output layer contains 1 neuron; the activation function of the network adopts the Tanh function, and the optimization algorithm in the model training process adopts the Adam gradient descent algorithm to minimize the error between the junction temperature prediction value and the output value of the electrothermal coupling model.

[0088] like Figure 4 As shown, this embodiment employs a macroscopic attention mechanism to evaluate the impact of overall characteristics across different time periods on junction temperature trends, and a microscopic attention mechanism to enhance the characterization of instantaneous junction temperature changes caused by fluctuations in electrical parameters at critical moments. Specifically: First, the mean of the input data is calculated for each time period, and the macroscopic attention coefficient for each time period is calculated based on the macroscopic query vector to amplify the characteristics of important time periods; then, for the input data and recursive data at different times, the microscopic attention coefficient is calculated using the microscopic query vector to enhance the sensitivity to key instantaneous features; finally, the weights of the input and recursive data are simultaneously amplified based on the macroscopic and microscopic coefficients, and the weighted features are input into the macroscopic and microscopic attention long short-term memory recursive neural network to achieve the fusion of long and short-term features.

[0089] like Figure 5As shown, this embodiment uses the hidden layer structure diagram of a Long Short-Term Memory recurrent neural network, specifically:

[0090] The structure consists of an input gate, a forget gate, an output gate, and a cell state. It selectively updates and transmits information through a gating mechanism. The input gate controls the inflow of new information, the forget gate determines the degree of retention of historical information, the output gate regulates the output of the current hidden state, and the cell state spans time steps to achieve long-term dependent modeling and transmission.

[0091] like Figure 6 As shown, the workflow of the macro-micro attention long short-term memory recurrent neural network is as follows:

[0092] Step one involves monitoring five key parameters (collector-emitter voltage, collector current, switching frequency, modulation ratio, and ambient temperature) and the IGBT junction temperature based on an electrothermal coupling model. The data from consecutive time points are then arranged and continuously recorded. T At each time point, all input data are arranged into a matrix of shape [value]. T A row of 5 columns is represented as:

[0093] ;

[0094] In the formula, X is the input matrix, representing the... T All data collected at consecutive time points; x t For the first t The input feature vector at each time step; T The total number of time steps, i.e., the length of historical data points input into the network in one cycle; each time step has D Input features ( D =5: Collector-emitter voltage, instantaneous collector current, switching frequency, modulation ratio, ambient temperature); [...] T This is a transpose.

[0095] Output junction temperature is This refers to the IGBT junction temperature monitored by the model. This process is a supervised learning process; the input is the electrical parameter matrix X, and the output is the monitored junction temperature. The label is the truth value and temperature y.

[0096] Step two, macro-level attention calculation, will continuously... T The data at a given time point is divided into several time periods. For all data within each time period, calculations are performed. T The model calculates the average value of each parameter at each time point, and uses five average values ​​to summarize the overall characteristics of each time period. A query vector is generated based on the hidden state of the previous time step. The overall feature vector of each time period is matched with the generated query vector to calculate the scoring function.

[0097] The query vector is:

[0098] q t = W q h t-1 ;

[0099] In the formula, q t For at any time t The generated query vector; W q The weight matrix projects the hidden state into the query space; h t-1 For at any time t -1 is a hidden state; H is the dimension of the hidden layer unit.

[0100] The scoring function is:

[0101] ;

[0102] In the formula, To score relevance, calculate an input vector. With query vector q The similarity between them; Let be the feature vector of the i-th time period; This is the scaling factor.

[0103] The scores for all time periods are transformed using the softmax function, resulting in a probability distribution with a sum of 1. The final weight value for each time period is the macro-attention weight.

[0104] The normalized attention weights are:

[0105] ;

[0106] In the formula, For the first i Macro-level attention weights for each time period; the softmax function transforms a set of arbitrary ratings into a probability distribution; exp() is the natural exponential function; This represents the average macro-level attention coefficient obtained over the entire time dimension. For the input dataset X t The Middle j The mean of a vector.

[0107] After macro-level attention identifies key time periods, the micro-level attention mechanism further generates query vectors at specific time points based on previous information. For each parameter at each time point, the relevance score between its value and the micro-level vector is calculated. Using the softmax function, the scores of the five parameters at a given time point are converted into a weight distribution with a sum of 1. This five-dimensional weight vector is the micro-level attention weight.

[0108] Micro-attention calculation involves weighting each of the five input features:

[0109] ;

[0110] In the formula, For at any time t The micro-attention weight vector; The activation function compresses the output to the (0,1) interval, serving as a gating mechanism; This is the weight matrix; To facilitate the splicing operation, the hidden state will be... and current input Concatenate them to form a growth vector.

[0111] Step three involves combining macro- and micro-level attention mechanisms with long short-term memory (LSM) neural networks. The LSM internal structure (input gate, forget gate, output gate) controls the flow of information. When calculating the gates, the weighted result of the input data, the hidden state from the previous time step, and the attention weights is used. The input gate determines which new information is stored in the memory cell. The input gate is represented as:

[0112] ;

[0113] In the formula, The input gate vector is used to completely retain values ​​close to 1 and completely ignore values ​​close to 0. and is a trainable weight matrix; For the current input x t Macro attention weight Scaling; Previous hidden state h t-1 Micro-attention weights Scaling element by element; The threshold value is the input gate threshold for the hidden layer.

[0114] Step four, the forget gate determines whether to forget old information from the past. The forget gate is represented as:

[0115] ;

[0116] In the formula, This is the forgetting gate vector, which controls the past cell states. The number that is retained; , , The weight matrix corresponding to the forget gate and the hidden layer forget gate threshold.

[0117] Step 5: Based on the weighted new input and the weighted past state, generate new information candidate content. The candidate cell state is represented as follows:

[0118] ;

[0119] In the formula, This represents the candidate cell state vector; tanh It is the hyperbolic tangent activation function; , , The weight matrix corresponds to the candidate state and the threshold of the hidden layer storage cell unit.

[0120] Old memories Forgotten according to the forgetting gate ratio, plus new candidate memories according to the input gate ratio. This forms new memories. Cell state updates are represented as:

[0121] ;

[0122] In the formula, For at any time t Updated cell state; Forgotten parts of past memories; To learn some new information.

[0123] Step six, control the output ratio from the cell state to the hidden state, the output gate is represented as:

[0124] ;

[0125] Hidden state update:

[0126] ;

[0127] In the formula, The output gate vector controls the cell state. The information will be output to the hidden state. Quantity; , , The corresponding weight matrix of the output gate and the threshold of the hidden layer output gate; For at any time t The updated hidden state.

[0128] By passing the hidden state of the final output from the Long Short-Term Memory (LSTM) layer through a fully connected layer, the final junction temperature monitoring value can be obtained. The junction temperature monitoring value is represented as:

[0129] ;

[0130] In the formula, This represents the final junction temperature monitoring value for the model. For the last time step T The hidden state; is a trainable weight matrix; This is a bias term.

[0131] The loss function used in this embodiment during model training is expressed as follows:

[0132] ;

[0133] In the formula, L This represents the loss value, i.e., the degree of inaccuracy in the model's predictions. N This represents the number of samples in the training batch. For the first i The true closing temperature of each sample; Let be the monitoring junction temperature of the i-th sample.

[0134] Parameter updates are represented as:

[0135] ;

[0136] In the formula, For all trainable parameters in the model; For the update operation, subtract the calculated gradient from the current value of the parameter; The learning rate controls the step size for each parameter update; For loss function L Regarding parameters The partial derivatives of .

[0137] In this embodiment, when training the IGBT module junction temperature monitoring model, a thermal network model is used to calculate the true junction temperature, specifically:

[0138] ;

[0139] in, For the junction temperature, For the power loss of IGBT, The junction-shell thermal resistance of the frequency converter. This refers to the thermal resistance from the inverter's casing to the heat sink. The thermal resistance from the inverter's heat sink to the environment. The ambient temperature.

[0140] S4. Using the trained IGBT module junction temperature monitoring model, input real-time electrical data and output IGBT junction temperature monitoring values.

[0141] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0142] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0143] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0144] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0145] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter, characterized in that, Includes the following steps: Based on the electrical characteristics and structural parameters of the IGBT module, an electrothermal coupling model is constructed using a Foster-type equivalent thermal network. Based on the electrothermal coupling model, simulations were performed under different operating conditions to obtain multiple sets of IGBT module junction temperature data and corresponding electrical data, and a training sample set was constructed. A long short-term memory recurrent neural network based on macro-micro attention mechanism is constructed. The macro-micro attention mechanism includes macro attention mechanism and micro attention mechanism. The macro attention mechanism is used to adaptively focus on key time period features in the time dimension, and the micro attention mechanism is used to adaptively weight electrical features in the feature dimension. The IGBT module junction temperature monitoring model was trained using the training sample set; the long short-term memory recurrent neural network based on the macro-micro attention mechanism calculated the mean of the input electrical data by time period, and calculated the macro attention weight of each time period based on the macro query vector to amplify the features of important time periods. Then, for the input data and recursive data at each time step, the hidden state of the previous time step is concatenated with the input data of the current time step, and the micro-attention weights are calculated through the weight matrix and activation function to enhance the sensitivity to key instantaneous features. Finally, the input and recursive data are scaled synchronously based on macro and micro weights, and the weighted features are input into the long short-term memory recursive neural network to fuse long and short-term features; The process involves calculating the mean of the input electrical data by time period and calculating the macro attention weight of each time period based on the macro query vector to amplify the features of important time periods. This includes dividing the electrical data at consecutive time points into several time periods, calculating the average value of each parameter for all electrical data in each time period, and constructing a macro feature matrix based on the average value of each parameter. Generate a query vector based on the hidden state of the previous time period, match the macro feature matrix of each time period with the generated query vector to calculate the relevance score, and normalize the relevance scores of all time periods to obtain the macro attention weight of each time period. The trained IGBT module junction temperature monitoring model is used as input for real-time electrical data and outputs IGBT junction temperature monitoring values.

2. The method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to claim 1, characterized in that, Based on the electrical characteristics and structural parameters of the IGBT module, an electrothermal coupling model is constructed using a Foster-type equivalent thermal network, including: Based on the electrical characteristics and structural parameters of IGBT modules, a Foster-type equivalent thermal network model is established to simulate the heat transfer path from the junction to the environment using a thermal resistance-thermal capacity network. Construct a power loss model to calculate the conduction loss and switching loss of the IGBT module; Using collector-emitter voltage, collector current, switching frequency, modulation ratio, and ambient temperature as model inputs, and IGBT junction temperature as model output, electrothermal interaction is achieved through the coupling of the electrical system and thermal network in Simulink.

3. The method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to claim 2, characterized in that, When extracting the thermal network parameters of the Foster-type equivalent thermal network model, the thermal resistance and thermal capacity parameters of each order of the Foster network are obtained by fitting the transient thermal resistance curve, and the multi-level thermal resistance and thermal capacity values ​​of the junction, shell radiator, and radiator environment are determined.

4. The method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to claim 2, characterized in that, When calculating the conduction loss of an IGBT module, the conduction loss is calculated based on the functional relationship between collector current and saturation voltage drop. The conduction voltage drop at different operating points is obtained by interpolation of the characteristic curve. When calculating the switching losses of an IGBT module, a polynomial fitting method is used to establish an estimation model of how the losses change with current, voltage, and temperature, based on the relationship between switching energy and collector current.

5. The method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to claim 1, characterized in that, Based on the electrothermal coupling model, simulations were performed under different operating conditions to obtain multiple sets of IGBT module junction temperature data and corresponding electrical data, and a training sample set was constructed, including: Dynamic simulations were performed under various operating conditions, including no-load, half-load, full-load and other typical load states, using an electrothermal coupling model. Multiple sets of junction temperature timing data of IGBT modules and their corresponding electrical characteristic parameters were obtained, and the obtained data were used to form a training sample set.

6. The method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to claim 1, characterized in that, The input and recursive data are simultaneously scaled based on macro and micro weights, and the weighted features are then input into a long short-term memory recursive neural network for the fusion of long and short-term features, including: The input gate and forget gate are calculated by weighting the input data, the hidden state of the previous time step, and the attention weights. Based on the weighted new input and the weighted past state, generate new information candidate content; The old cell state is forgotten according to the forget gate ratio, and new candidate cell states are added according to the input gate ratio to form a new cell state; Control the output ratio from the cell state to the hidden state, calculate the output gate, and update the hidden state. The final hidden state is mapped through a fully connected layer to obtain the final junction temperature monitoring value.

7. The method for monitoring the junction temperature of an IGBT module in a high-voltage frequency converter according to claim 1, characterized in that, When training the junction temperature monitoring model for IGBT modules, a thermal network model is used to calculate the true junction temperature, specifically: ； in, For the junction temperature, For the power loss of IGBT, The junction-to-shell thermal resistance of the frequency converter. This refers to the thermal resistance from the inverter's casing to the heat sink. The thermal resistance from the inverter's heat sink to the environment. The ambient temperature.

Images