A semiconductor device loss intelligent calculation method and system

Abstract

The application provides a kind of semiconductor device loss intelligent calculation method and system, the method first real-time acquisition semiconductor device's operating electrical parameter and the temperature parameter of surrounding environment multiple temperature measuring points, form synchronous operation data set with time stamp.Based on the principle of semiconductor physics, a parameterized intrinsic loss model is established to describe the relationship between conduction loss and switching loss as a function of electrical parameters and junction temperature.A parameterized thermal coupling effect model is established based on the physical layout of the control cabinet and the heat transfer path, with the thermal resistance parameter to be identified describing the heat transfer relationship between the device and other heat sources.Parameter identification algorithm is used to update the model parameters using the synchronous operation data set, resulting in an adaptive model parameter set that reflects the current operating conditions and coupling state.Combining the synchronous operation data set and the adaptive model parameter set, the real-time loss value of the semiconductor device is calculated through the intrinsic loss model and the thermal coupling effect model, achieving accurate loss evaluation under actual operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device technology, specifically relating to a method and system for intelligent calculation of semiconductor device losses. Background Technology

[0002] In the field of modern power electronics systems and industrial control equipment, accurately assessing and monitoring the operating status and performance parameters of key components has always been a major technical challenge. This is especially true in high-power, high-frequency applications, where thermal management and energy efficiency optimization are paramount. Traditional methods often rely on simplified theoretical calculations or offline test data, which fail to accurately reflect the true operating conditions of components in real-world environments. This leads to excessive design redundancy or increased reliability risks. The most critical technical problem lies in the lack of a method capable of accurately assessing the performance parameters of key components in real-time, taking into account the influence of complex environmental factors under actual operating conditions.

[0003] The core challenge of this problem lies in the complexity and dynamism of the actual working environment. In real control systems, the performance of components is affected by a combination of factors, including changes in electrical parameters, temperature fluctuations, load variations, and interactions with surrounding components. These factors are interconnected and dynamically changing, making it difficult for traditional methods based on static models or simplified assumptions to accurately reflect the actual operating state of the components. Especially in modern control equipment with high-density integration, the thermal coupling effect between components is even more significant, further increasing the complexity of performance evaluation. Summary of the Invention

[0004] This invention provides a method and system for intelligent calculation of semiconductor device losses, in order to solve the problem of inaccurate calculation of semiconductor device losses in control systems.

[0005] In a first aspect, the present invention provides a method for intelligent calculation of semiconductor device losses, the method comprising the following steps:

[0006] Real-time acquisition of the operating electrical parameters of semiconductor devices in the control cabinet and the temperature parameters of multiple preset temperature measurement points in the vicinity of the semiconductor devices to form a synchronous operation dataset containing timestamps;

[0007] A parameterized bulk loss model for semiconductor devices is established based on semiconductor physics principles. The bulk loss model includes device parameters to be identified that describe the changes in conduction and switching losses of semiconductor devices with operating electrical parameters and junction temperature.

[0008] A parameterized thermal coupling effect model is established based on the physical layout and heat transfer path of the control cabinet. The thermal coupling effect model uses the thermal resistance parameter to be identified to describe the heat transfer relationship between semiconductor devices and other heat sources in the control cabinet.

[0009] A parameter identification algorithm is used to identify and update the parameters of the device to be identified and the thermal resistance parameters to be identified using a synchronous running dataset, thereby obtaining an adaptive model parameter set that reflects the current operating condition and coupling state of the control cabinet;

[0010] By combining the synchronous running dataset and the adaptive model parameter set, and by calculating the real-time loss value of the semiconductor device through the body loss model and the thermal coupling effect model, the real-time loss value of the semiconductor device is obtained.

[0011] Optionally, the real-time acquisition of the operating electrical parameters of the semiconductor devices inside the control cabinet and the temperature parameters of multiple preset temperature measurement points in the vicinity of the semiconductor devices to form a synchronous operation dataset including timestamps includes the following steps:

[0012] The operating electrical parameters of semiconductor devices within the control cabinet are captured by pre-configured high-bandwidth current sensors and high-impedance voltage sensors.

[0013] Temperature sensors are installed at predetermined environmental locations on the surface of the semiconductor device's housing and heat sink, as well as inside the control cabinet, to collect the temperature parameters of the semiconductor device.

[0014] A precise time protocol is used to maintain time consistency for all operating electrical and temperature parameters;

[0015] The sampling rate is set according to the dynamic characteristics of the operating electrical and temperature parameters;

[0016] The collected operating electrical and temperature parameters are digitally filtered, and amplitude correction is performed based on sensor calibration data.

[0017] The preprocessed operating electrical and temperature parameters are aligned to a unified time axis through interpolation to form a synchronized operating dataset.

[0018] Optionally, the establishment of a parameterized bulk loss model for semiconductor devices based on semiconductor physics principles includes the following steps:

[0019] The on-state voltage drop of a semiconductor device is expressed as the sum of a zero-current voltage drop term and an equivalent on-resistance term to construct a parameterized on-state loss model. Both the zero-current voltage drop and the equivalent on-resistance are parameterized as polynomial functions of the junction temperature.

[0020] A parameterized switching loss model is constructed by expressing the single turn-on energy and single turn-off energy of a semiconductor device as power-law functions of junction temperature, switching current, and blocking voltage, respectively, combined with polynomial interpolation.

[0021] The junction temperature correlation coefficient in the conduction loss model and the correlation coefficients of junction temperature, switching current and blocking voltage in the switching loss model are collectively defined as the set of parameters of the device to be identified.

[0022] Initial estimates are set for the parameters of the device to be identified based on typical curves extracted from the device datasheets of semiconductor devices.

[0023] The conduction loss model and the switching loss model are integrated into a bulk loss model for semiconductor devices. The bulk loss model takes operating electrical parameters, temperature parameters, and parameters of the device to be identified as model inputs.

[0024] Optionally, establishing a parameterized thermal coupling effect model based on the physical layout and heat transfer path of the control cabinet includes the following steps:

[0025] Based on the physical layout of the control cabinet, identify all major heat-generating devices and key heat dissipation paths within the control cabinet, and treat all major heat-generating devices as thermal network nodes.

[0026] By combining the principles of heat transfer and connecting all thermal network nodes through key heat dissipation paths, a thermal resistance network topology for the control cabinet is constructed.

[0027] The thermal resistance parameters to be identified are defined based on the heat transfer effect of the thermal resistance network topology. The thermal resistance parameters to be identified include the self-thermal resistance used to represent the heat transfer from the thermal network node itself to the thermal resistance network topology, and the mutual thermal resistance used to represent the heat generated by one thermal network node to another thermal network node.

[0028] A parameterized thermal coupling effect model is constructed based on the steady-state thermal balance equation and combined with the thermal resistance network topology and the thermal resistance parameters to be identified.

[0029] Optionally, the step of using a parameter identification algorithm and utilizing a synchronously running dataset to identify and update the parameters of the device to be identified and the thermal resistance parameters to be identified, to obtain an adaptive model parameter set reflecting the current operating condition and coupling state of the control cabinet, includes the following steps:

[0030] The bulk loss model and the thermal coupling effect model are integrated into a nonlinear state-space equation. The state vector in the nonlinear state-space equation includes the junction temperature and case temperature of the semiconductor device.

[0031] The device parameters and thermal resistance parameters to be identified are augmented into the state vector using slow time-varying characteristics;

[0032] Initialize the state vector and the estimated values ​​of the parameters to be identified, as well as the covariance matrix;

[0033] At each sampling time, the prediction and update steps of the extended Kalman filter are executed. The Kalman gain is calculated using the synchronous running dataset as the measurement value, the predicted values ​​of the state vector and the parameters to be identified are corrected, and the posterior optimal estimate at the current time is obtained.

[0034] The updated state vector is output periodically, and the adaptive model parameter set is extracted from the state vector.

[0035] Optionally, the step of combining the synchronous running dataset and the adaptive model parameter set, and calculating the real-time loss value of the semiconductor device through the bulk loss model and the thermal coupling effect model, includes the following steps:

[0036] The junction temperature estimate of the semiconductor device at the current moment is obtained from the posterior state estimate of the extended Kalman filter output;

[0037] Substitute the junction temperature estimate, the device parameters in the adaptive model parameter set updated at the current time, and the operating electrical parameters in the synchronous running dataset into the body loss model to calculate the real-time conduction loss and real-time switching loss of the semiconductor device.

[0038] The junction temperature estimate, the thermal resistance parameters in the adaptive model parameter set updated at the current time, and the temperature parameters in the synchronous running dataset are substituted into the thermal coupling effect model to calculate the real-time thermal effect loss of the semiconductor device.

[0039] The real-time loss value of the target semiconductor device is obtained by combining real-time conduction loss, real-time switching loss and real-time thermal effect loss.

[0040] Optionally, the method further includes the following steps:

[0041] Real-time calculation of prediction error in the extended Kalman filter update step;

[0042] Statistical analysis was performed on the prediction errors at multiple consecutive sampling times to calculate the mean and variance of the error sequence.

[0043] When the mean and / or variance of the error sequence exceed a preset threshold, it is determined that a model mismatch or a sudden change in operating conditions has occurred.

[0044] In response to model mismatch or sudden changes in operating conditions, the terms corresponding to the parameter states in the process noise covariance matrix of the extended Kalman filter are increased to accelerate the parameter adaptation rate.

[0045] Optionally, the method further includes the following steps:

[0046] The system continuously accumulates the synchronous running dataset and the corresponding real-time loss value. When the accumulated data reaches a preset data threshold, a composite feature set is extracted from the accumulated synchronous running dataset.

[0047] The target value of loss residual is defined based on the dynamic cross-influence of device parameters and temperature parameters;

[0048] A pre-defined loss residual prediction model is trained using a composite feature set and the corresponding loss residual target value, so that the loss residual prediction model learns the nonlinear mapping relationship between the composite feature set and the loss residual.

[0049] After the loss residual prediction model is trained, if the real-time loss value is calculated, the real-time composite feature is extracted from the synchronous running dataset corresponding to the real-time loss value.

[0050] The real-time composite features are input into the loss residual prediction model, and the current loss residual prediction value is output in real time through the loss residual prediction model.

[0051] The current loss residual prediction value is fused with the real-time loss value to obtain the loss calculation result of the semiconductor device after residual compensation.

[0052] In a second aspect, the present invention also provides a semiconductor device loss intelligent calculation system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the semiconductor device loss intelligent calculation method as described in the first aspect.

[0053] Thirdly, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the intelligent calculation method for semiconductor device losses according to the first aspect.

[0054] The beneficial effects of this invention are:

[0055] This invention significantly improves the accuracy and practicality of semiconductor device loss calculation by innovatively integrating real-time data acquisition, physical model construction, and parameter adaptation techniques, solving the technical problem of large evaluation deviations under actual operating conditions in traditional methods. Compared with traditional methods that rely on static parameters and simplified environmental assumptions, this invention establishes a bulk loss model based on semiconductor physics principles, which can accurately describe the dynamic characteristics of device conduction and switching losses as electrical parameters and junction temperature change, making loss calculations more consistent with actual operating conditions. In particular, in the handling of thermal coupling effects, this invention breaks through by constructing a parameterized thermal coupling model that considers the influence of control cabinet physical layout and heat transfer paths, effectively capturing the complex interaction between semiconductor devices and surrounding heat sources, and overcoming the systematic errors caused by traditional methods ignoring environmental influences. By introducing a parameter identification algorithm and continuously updating model parameters using a synchronously running dataset, this invention achieves dynamic adaptation to device characteristics and environmental changes, ensuring that loss calculation results always maintain high accuracy and accurately reflect the actual state of the device even under long-term operation and changing operating conditions. This real-time and accurate loss calculation capability provides a reliable basis for equipment optimization design, energy efficiency management, and predictive maintenance, effectively extending equipment lifespan, reducing operating costs, and improving system reliability. It is of great value for enhancing the overall performance of modern power electronic systems and industrial control equipment. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the intelligent calculation method for semiconductor device losses in one embodiment of this application. Detailed Implementation

[0057] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0058] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0059] Figure 1 This is a flowchart illustrating a method for intelligently calculating semiconductor device losses in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps. For example Figure 1 As shown, the intelligent loss calculation method for semiconductor devices disclosed in this invention specifically includes the following steps:

[0060] S101. Real-time acquisition of the operating electrical parameters of semiconductor devices in the control cabinet and the temperature parameters of multiple preset temperature measurement points in the vicinity of the semiconductor devices, forming a synchronous operation dataset containing timestamps.

[0061] The control cabinet incorporates high-bandwidth current sensors and high-impedance voltage sensors to accurately capture real-time current and voltage waveforms from semiconductor devices such as IGBTs and MOSFETs. Current sensors typically employ the Hall effect or Rogowski coil technology, enabling the measurement of large currents without interfering with normal circuit operation. Voltage sensors utilize high-impedance voltage divider networks, ensuring measurement accuracy without affecting circuit performance. Temperature sensors, which can be thermocouples, thermistors, or infrared sensors, are strategically placed on the semiconductor device housings, heat sinks, and at critical locations within the control cabinet to collect device surface temperatures and ambient temperatures. To ensure data consistency, all sensors are synchronized using precise time protocols (such as PTP or IEEE 1588), ensuring accurate timestamps on the measurement data. Depending on the dynamic characteristics of the semiconductor devices, the sampling rate for electrical parameters is typically set from tens of kHz to several MHz, while the sampling rate for temperature parameters, which change more slowly, can be set from several Hz to tens of Hz. The acquired raw data undergoes digital filtering to eliminate high-frequency noise and electromagnetic interference, and amplitude correction is performed based on sensor calibration data to ensure measurement accuracy. Finally, the processed electrical and temperature parameters are aligned to a unified time axis through interpolation, forming a synchronously running dataset with timestamps. This synchronous data acquisition method ensures the accuracy of the temporal relationship between electrical and temperature parameters in subsequent analysis, providing a reliable data foundation for accurately calculating semiconductor device losses.

[0062] S102. Establish a parameterized bulk loss model for semiconductor devices based on semiconductor physics principles. The bulk loss model includes device parameters to be identified that describe the changes in conduction and switching losses of semiconductor devices with operating electrical parameters and junction temperature.

[0063] The inherent losses of semiconductor devices mainly consist of two parts: conduction losses and switching losses. The conduction loss model is based on the current-voltage characteristics of the semiconductor device in the conduction state. Specifically, the conduction voltage drop of the semiconductor device is expressed as the sum of the zero current drop term and the equivalent on-resistance term, i.e., V(I,Tj)=V0(Tj)+R(Tj)×I, where V0 represents the zero current drop, R represents the equivalent on-resistance, I is the current flowing through the device, and Tj is the junction temperature. Both the zero current drop V0 and the equivalent on-resistance R are parameterized as polynomial functions of the junction temperature Tj, for example, V0(Tj)=a0+a1×Tj+a2×Tj2, R(Tj)=b0+b1×Tj+b2×Tj2, where a0, a1, a2, b0, b1, and b2 are the parameters of the device to be identified. Conduction loss is calculated as the product of conduction voltage and current, i.e., Pcond = V(I,Tj) × I. The switching loss model is based on the energy loss characteristics of the device during turn-on and turn-off. The single turn-on energy Eon and single turn-off energy Eoff are expressed as a combination of power-law functions of junction temperature, switching current, and blocking voltage, and polynomial interpolation, such as Eon(I,V,Tj) = k1 × I^α × V^β × f(Tj), Eoff(I,V,Tj) = k2 × I^γ × V^δ × g(Tj), where k1, k2, α, β, γ, and δ are parameters to be identified, and f(Tj) and g(Tj) are polynomial functions of junction temperature. Switching loss is calculated as the single switching energy multiplied by the switching frequency, i.e., Psw = (Eon + Eoff) × fsw. Typical curves are extracted from the datasheets of semiconductor devices to set initial estimates for the parameters of the device to be identified. Then, the conduction loss model and the switching loss model are integrated to form a complete bulk loss model. This model takes operating electrical parameters, temperature parameters, and the parameters of the device to be identified as inputs, and can accurately reflect the loss characteristics of semiconductor devices under different operating conditions, laying a theoretical foundation for subsequent parameter identification and loss calculation.

[0064] S103. Based on the physical layout and heat transfer path of the control cabinet, a parameterized thermal coupling effect model is established. The thermal coupling effect model uses the thermal resistance parameter to be identified to describe the heat transfer relationship between semiconductor devices and other heat sources in the control cabinet.

[0065] The establishment of the thermal coupling effect model first requires identifying all major heat-generating devices and key heat dissipation paths based on the physical layout of the control cabinet. The main heat-generating devices within the control cabinet include semiconductor power devices (such as IGBT modules and rectifier diodes), reactors, and transformers. The heat generated by these devices during operation is transferred through conduction, convection, and radiation. All major heat-generating devices are treated as nodes in a thermal network. Based on the physical relationships between the devices and the heat dissipation paths, a thermal resistance network topology for the control cabinet is constructed. In the thermal resistance network, each node represents the temperature of a heat source, and nodes are connected by thermal resistance, which characterizes the ease of heat transfer. The thermal resistance network includes two key parameters: self-thermal resistance and mutual thermal resistance. Self-thermal resistance describes how the heat generated by a device itself affects its temperature, while mutual thermal resistance describes how the heat generated by one device affects the temperature of other devices. For example, for two thermal network nodes i and j, the temperature rise ΔTi of node i can be expressed as a function of its own thermal resistance Pi and the thermal resistance Pj of other nodes: ΔTi = Rth_ii × Pi + Rth_ij × Pj, where Rth_ii is the self-thermal resistance of node i, and Rth_ij is the mutual thermal resistance between node j and node i. Based on the steady-state thermal balance equation, a thermal coupling matrix equation containing all nodes is constructed: [ΔT1; ΔT2; ...; ΔTn] = [Rth_11, Rth_12, ..., Rth_1n; Rth_21, Rth_22, ..., Rth_2n; ...; Rth_n1, Rth_n2, ..., Rth_nn] × [P1; P2; ...; Pn], where n is the number of thermal network nodes. These initial values ​​of thermal resistance parameters can be estimated using thermal simulation software, but due to the complexity and uncertainty of actual heat dissipation conditions, they need to be adjusted online using parameter identification algorithms. The established thermal coupling effect model can accurately describe the thermal interaction between various devices in the control cabinet, providing a thermal environment basis for semiconductor device junction temperature estimation and loss calculation.

[0066] S104. Using a parameter identification algorithm and a synchronous running dataset, identify and update the parameters of the device to be identified and the thermal resistance parameters to be identified, to obtain an adaptive model parameter set that reflects the current operating condition and coupling state of the control cabinet.

[0067] The parameter identification process first integrates the bulk loss model and the thermal coupling effect model into a nonlinear state-space equation. The state vector includes the junction temperature Tj and case temperature Tc of the semiconductor device, as well as the device parameters and thermal resistance parameters to be identified. The state equation describes the evolution of the state vector over time, while the observation equation describes the relationship between the measured values ​​and the state vector. Considering that the semiconductor device parameters and thermal resistance parameters change slowly over a short period of time, these parameters are augmented into the state vector as state variables with slow time-varying characteristics. The estimated values ​​of the state vector and the parameters to be identified are initialized, typically based on the device datasheet and thermal simulation results; simultaneously, the state estimation error covariance matrix P is initialized to reflect the uncertainty of the initial estimate. At each sampling time, an extended Kalman filter prediction step is executed, predicting the current state based on the state estimate from the previous time step and the system dynamic model; then, an update step is executed, comparing the measured values ​​and predicted values ​​in the synchronously running dataset, calculating the Kalman gain, correcting the predicted values ​​of the state vector and the parameters to be identified, and obtaining the posterior optimal estimate for the current time step. The core of Extended Kalman Filtering (EKF) lies in linearizing nonlinear systems. The computational process includes state prediction equations, covariance prediction equations, Kalman gain calculation, state update equations, and covariance update equations. To address model mismatch or sudden changes in operating conditions, prediction errors are calculated and statistically analyzed in real time. When the mean or variance of the error sequence exceeds a preset threshold, the corresponding parameter state term in the process noise covariance matrix is ​​increased, accelerating the parameter adaptation rate. The updated state vector is periodically output, from which the adaptive model parameter set, including updated device parameters and thermal resistance parameters, is extracted. This parameter identification method can track changes in semiconductor device characteristics and the thermal environment in real time, ensuring the model always accurately describes the actual system.

[0068] S105. Combine the synchronous running dataset and the adaptive model parameter set, and calculate the real-time loss value of the semiconductor device through the body loss model and the thermal coupling effect model.

[0069] The real-time loss calculation process first obtains the estimated junction temperature Tj of the semiconductor device at the current moment from the posterior state estimate output by the extended Kalman filter. Junction temperature is the temperature of the PN junction inside the semiconductor device, directly affecting its conduction and switching characteristics, but it is usually not directly measurable and needs to be obtained through state estimation. The estimated junction temperature, the device parameters in the updated adaptive model parameter set at the current moment, and the operating electrical parameters (current I and voltage V) in the synchronous running dataset are substituted into the body loss model to calculate the real-time conduction loss Pcond and real-time switching loss Psw of the semiconductor device. The conduction loss calculation formula is Pcond = V(I,Tj) × I, where V(I,Tj) is the conduction voltage, determined by the zero current drop and the equivalent on-resistance; the switching loss calculation formula is Psw = (Eon + Eoff) × fsw, where Eon and Eoff are the single turn-on energy and turn-off energy, respectively, and fsw is the switching frequency. Simultaneously, the estimated junction temperature, the thermal resistance parameters in the updated adaptive model parameter set, and the temperature parameters in the synchronous running dataset are substituted into the thermal coupling effect model to calculate the real-time thermal effect loss Pth of the semiconductor device, which is the additional loss caused by the thermal coupling effect of other heat sources. Finally, combining the real-time conduction loss, real-time switching loss, and real-time thermal effect loss, the real-time loss value Ptotal = Pcond + Psw + Pth of the target semiconductor device is comprehensively calculated. This loss calculation method not only considers the electrical characteristics of the semiconductor device itself but also the influence of the thermal environment, accurately reflecting the loss state of the device under actual operating conditions. The real-time loss value can be used for thermal management of control cabinets, device lifetime prediction, and system efficiency optimization, providing data support for safe equipment operation and preventive maintenance. By continuously accumulating the synchronous running dataset and the corresponding real-time loss value, a loss residual prediction model can also be trained to further improve the accuracy of loss calculation.

[0070] In one embodiment, the real-time acquisition of the operating electrical parameters of semiconductor devices inside the control cabinet and the temperature parameters of multiple preset temperature measurement points in the vicinity of the semiconductor devices to form a synchronous operation dataset including timestamps includes the following steps:

[0071] The operating electrical parameters of semiconductor devices within the control cabinet are captured by pre-configured high-bandwidth current sensors and high-impedance voltage sensors.

[0072] Temperature sensors are installed at predetermined environmental locations on the surface of the semiconductor device's housing and heat sink, as well as inside the control cabinet, to collect the temperature parameters of the semiconductor device.

[0073] A precise time protocol is used to maintain time consistency for all operating electrical and temperature parameters;

[0074] The sampling rate is set according to the dynamic characteristics of the operating electrical and temperature parameters;

[0075] The collected operating electrical and temperature parameters are digitally filtered, and amplitude correction is performed based on sensor calibration data.

[0076] The preprocessed operating electrical and temperature parameters are aligned to a unified time axis through interpolation to form a synchronized operating dataset.

[0077] In this embodiment, high-bandwidth current sensors and high-impedance voltage sensors are installed in the control cabinet to capture the operating electrical parameters of semiconductor devices in real time. The current sensors typically employ the Hall effect principle or Rogowski coil technology, enabling the measurement of large currents without interfering with the circuit. For example, a Hall effect current sensor utilizes the magnetic field generated by current passing through a conductor acting on a Hall element, outputting a voltage signal proportional to the current, with a bandwidth reaching hundreds of kHz, suitable for capturing rapid current changes in semiconductor devices. The voltage sensors employ a high-impedance voltage divider network design, with an input impedance typically greater than 1 MΩ, ensuring measurement accuracy without causing a load effect on the circuit under test. These sensors are installed at critical locations on the semiconductor devices, such as the collector-emitter and gate-emitter positions of IGBT modules, ensuring accurate capture of electrical parameters during switching transients and conduction states. The sensor output signals are transmitted to the data acquisition system via shielded cables to reduce external electromagnetic interference and ensure signal integrity and accuracy.

[0078] Temperature sensors are strategically placed at critical locations within semiconductor devices to comprehensively monitor their temperature distribution. Thermocouples or thermistors are mounted on the device housing surface to directly measure the housing temperature. These sensors are typically miniaturized, such as TO-92 packaged thermistors or K-type thermocouples with a diameter not exceeding 0.5 mm, ensuring no impact on the device's heat dissipation performance. Multiple temperature sensors are deployed at different locations on the heat sink surface, forming a temperature gradient monitoring network. Sensors are typically placed in hotspot and edge areas to capture the heat sink's temperature distribution characteristics. Simultaneously, ambient temperature sensors are installed at critical environmental locations within the control cabinet, such as air inlets, outlets, and the air area surrounding power devices. These sensors typically have a measurement range of -40℃ to 150℃, with an accuracy of ±0.5℃, and use standardized output signals such as 4-20mA or 0-10V for easy integration with data acquisition systems. A well-planned layout of temperature sensors provides information on the device's temperature field distribution, offering a data foundation for analyzing thermal coupling effects.

[0079] To ensure time synchronization of all sensor data, a precise time protocol is used to synchronize the data acquisition system. Specifically, a time synchronization system based on the IEEE 1588 Precise Time Protocol (PTP) or Network Time Protocol (NTP) is deployed within the control cabinet. The master clock server distributes time information to all data acquisition devices via the network, and each device calculates its local clock correction value based on the received time information and transmission delay. The PTP protocol achieves microsecond-level synchronization accuracy and is suitable for acquiring electrical parameters with strict time requirements; while the NTP protocol achieves millisecond-level synchronization accuracy and is suitable for acquiring slowly changing temperature parameters. Each acquisition device appends a current timestamp when recording data to ensure accurate timing relationships.

[0080] Different sampling rate strategies are set to suit the dynamic characteristics of different parameters. Electrical parameters such as current and voltage change rapidly, especially during switching transients, with change times ranging from a few microseconds to tens of microseconds. Therefore, the sampling rate for electrical parameters is set relatively high, typically in the range of 50kHz to 5MHz. For example, for an IGBT module with a switching frequency of 10kHz, the sampling rate for electrical parameters should be set to at least 200kHz to ensure that more than 20 data points are collected per switching cycle, accurately capturing the switching transient process. Temperature parameters change relatively slowly, with thermal time constants typically on the order of seconds or minutes. Therefore, the sampling rate for temperature parameters can be set relatively low, typically in the range of 1Hz to 100Hz. Appropriate sampling rate settings ensure the capture of critical dynamic changes while avoiding the generation of excessive redundant data.

[0081] Raw acquired data typically contains various noises and interferences, requiring digital filtering to improve signal quality. For electrical parameters, a low-pass filter is used to remove high-frequency noise, with the cutoff frequency set to 1 / 4 to 1 / 10 of the sampling rate to ensure no loss of effective signal components. For data from switching transient processes, a small-window median filter can be used to remove spike interference. Temperature parameters are processed using moving average filtering or Kalman filtering to smooth the temperature curve and suppress random fluctuations. The filtered data then needs amplitude correction based on sensor calibration data. Amplitude correction converts the sensor output signal into an actual physical quantity, such as converting a voltage signal into an actual current or temperature value. The correction formula typically includes linear and nonlinear terms, considering the sensor's sensitivity, zero-point offset, and nonlinear characteristics.

[0082] Since electrical and temperature parameters are collected using different sampling rates, these asynchronous data need to be aligned to a unified time axis. First, the time resolution of the unified time axis is determined, typically depending on the needs of subsequent analysis; higher time resolution, such as 10μs to 100μs, is required for loss calculations. Then, time interpolation is performed on the parameter data. For electrical parameters, due to the higher sampling rate, linear interpolation or cubic spline interpolation can be used to downsample to the unified time axis. For temperature parameters, due to the lower sampling rate, interpolation methods are needed to improve the time resolution. Commonly used methods include linear interpolation, polynomial interpolation, or piecewise cubic Hermite interpolation, the latter of which maintains the smoothness and monotonicity of the data. The interpolation process needs to consider the physical characteristics of each parameter, such as the continuity of temperature changes and the rapid changes in current during switching transients. The resulting synchronized dataset contains the values ​​of all parameters at a unified time point, with each data point bearing a precise timestamp, providing a complete and consistent data foundation for subsequent loss calculations and model identification.

[0083] In one embodiment, establishing a parameterized bulk loss model for a semiconductor device based on semiconductor physics principles includes the following steps:

[0084] The on-state voltage drop of a semiconductor device is expressed as the sum of a zero-current voltage drop term and an equivalent on-resistance term to construct a parameterized on-state loss model. Both the zero-current voltage drop and the equivalent on-resistance are parameterized as polynomial functions of the junction temperature.

[0085] A parameterized switching loss model is constructed by expressing the single turn-on energy and single turn-off energy of a semiconductor device as power-law functions of junction temperature, switching current, and blocking voltage, respectively, combined with polynomial interpolation.

[0086] The junction temperature correlation coefficient in the conduction loss model and the correlation coefficients of junction temperature, switching current and blocking voltage in the switching loss model are collectively defined as the set of parameters of the device to be identified.

[0087] Initial estimates are set for the parameters of the device to be identified based on typical curves extracted from the device datasheets of semiconductor devices.

[0088] The conduction loss model and the switching loss model are integrated into a bulk loss model for semiconductor devices. The bulk loss model takes operating electrical parameters, temperature parameters, and parameters of the device to be identified as model inputs.

[0089] In this embodiment, the voltage drop of a semiconductor device in the on-state can be decomposed into two key components: the zero-current voltage drop and the current-dependent voltage drop. The zero-current voltage drop represents the voltage threshold of the device at extremely low current, similar to the on-state voltage of a PN junction; while the current-dependent voltage drop can be represented by multiplying the current by the equivalent on-resistance. This decomposition method is based on the physical characteristics of semiconductors and can accurately describe the conduction behavior of power devices such as IGBTs or MOSFETs. Specifically, the on-state voltage drop is expressed as V(I,Tj) = V0(Tj) + R(Tj) × I, where V0 is the zero-current voltage drop, R is the equivalent on-resistance, I is the current through the device, and Tj is the junction temperature. To describe the effect of temperature on these parameters, both V0 and R are expressed as polynomial functions of junction temperature, such as second-order polynomials: V0(Tj)=a0+a1×Tj+a2×Tj2, R(Tj)=b0+b1×Tj+b2×Tj2, where a0, a1, a2, b0, b1, and b2 are the parameters to be identified.

[0090] Energy losses in semiconductor devices during switching are influenced by multiple factors, including switching current, blocking voltage, and junction temperature. The turn-on energy (Eon) and turn-off energy (Eoff) are key parameters describing these losses. By analyzing the physical mechanism of the switching process, the single turn-on and turn-off energies can be expressed as functions of these factors. Specifically, this is achieved using a combination of power-law functions and polynomial interpolation: Eon(I,V,Tj)=k1×I^α×V^β×f(Tj), Eoff(I,V,Tj)=k2×I^γ×V^δ×g(Tj). Here, I is the switching current, V is the blocking voltage, k1, k2, α, β, γ, and δ are parameters describing the effects of current and voltage, and f(Tj) and g(Tj) are polynomial functions of junction temperature, such as f(Tj)=c0+c1×Tj+c2×Tj². The power-law function part describes the nonlinear relationship between switching energy and current and voltage. Typically, α and γ are close to 1, and β and δ are close to 1 or slightly larger. The polynomial function part describes the effect of junction temperature on switching energy. As the junction temperature increases, the turn-on energy usually increases while the turn-off energy decreases.

[0091] The unknown coefficients in the models established in the first two steps are combined into a complete set of parameters for the device to be identified. This set includes coefficients (a0, a1, a2, b0, b1, b2) describing the zero-current voltage drop and equivalent on-resistance as a function of junction temperature in the conduction loss model, and coefficients (k1, k2, α, β, γ, δ, c0, c1, c2, etc.) describing the single-cycle turn-on energy and turn-off energy as a function of junction temperature, switching current, and blocking voltage in the switching loss model. These parameters together constitute a high-dimensional parameter vector θ = [a0, a1, a2, b0, b1, b2, k1, k2, α, β, γ, δ, c0, c1, c2, ...]. The dimension of the parameter set depends on the order of the chosen polynomial and the complexity of the model, typically between 10 and 20. These parameters have clear physical meanings, corresponding to different physical characteristics of semiconductor devices, such as PN junction characteristics, drift region resistance, and carrier lifetime. These parameters are collectively defined as parameters to be identified, which facilitates online estimation and updating using a unified parameter identification algorithm, enabling the model to adapt to changes in device characteristics.

[0092] To ensure the parameter identification algorithm converges quickly to a reasonable solution, appropriate initial estimates need to be set for the parameters of the device to be identified. These initial values ​​are obtained by analyzing typical curves provided in the semiconductor device datasheets. Datasheets typically contain characteristic curves under various operating conditions, such as output characteristic curves (Vce-Ic or Vds-Id) at different junction temperatures, and curves showing the relationship between switching energy loss and current. From the output characteristic curves, the relationship between on-state voltage and current at different junction temperatures can be extracted. Linear fitting yields the values ​​of zero-current voltage drop V0 and equivalent on-resistance R at different junction temperatures, and polynomial fitting provides the initial values ​​for the coefficients of V0(Tj) and R(Tj). From the switching energy loss curves, the relationship between Eon, Eoff, and current and voltage can be analyzed in a double logarithmic coordinate system to determine the initial values ​​of the power-law exponents α, β, γ, and δ. Furthermore, the initial values ​​of the coefficients of the temperature influence functions f(Tj) and g(Tj) are determined using test data at different junction temperatures. These initial parameters allow the model to roughly reflect the basic characteristics of the device before identification begins.

[0093] The previously established conduction loss model and switching loss model are integrated into a complete bulk loss model. Conduction loss is calculated as the product of conduction voltage and current: Pcond = V(I,Tj) × I = [V0(Tj) + R(Tj) × I] × I, where V(I,Tj) is the conduction voltage drop, I is the current, and Tj is the junction temperature. Switching loss is calculated as the energy of a single switch multiplied by the switching frequency: Psw = [Eon(I,V,Tj) + Eoff(I,V,Tj)] × fsw, where Eon and Eoff are the energy of a single turn-on and turn-off, respectively, and fsw is the switching frequency. The total bulk loss is the sum of conduction loss and switching loss: Ptotal = Pcond + Psw. This integrated model takes operating electrical parameters (current I, voltage V, switching frequency fsw), temperature parameters (junction temperature Tj), and device parameters (parameter vector θ) as inputs, and outputs the real-time loss values ​​of the semiconductor device. The establishment of the bulk loss model enables accurate calculation of semiconductor device losses under known operating conditions and device parameters, providing a foundation for subsequent thermal analysis and lifetime prediction.

[0094] In one implementation, establishing a parameterized thermal coupling effect model based on the physical layout and heat transfer path of the control cabinet includes the following steps:

[0095] Based on the physical layout of the control cabinet, identify all major heat-generating devices and key heat dissipation paths within the control cabinet, and treat all major heat-generating devices as thermal network nodes.

[0096] By combining the principles of heat transfer and connecting all thermal network nodes through key heat dissipation paths, a thermal resistance network topology for the control cabinet is constructed.

[0097] The thermal resistance parameters to be identified are defined based on the heat transfer effect of the thermal resistance network topology. The thermal resistance parameters to be identified include the self-thermal resistance used to represent the heat transfer from the thermal network node itself to the thermal resistance network topology, and the mutual thermal resistance used to represent the heat generated by one thermal network node to another thermal network node.

[0098] A parameterized thermal coupling effect model is constructed based on the steady-state thermal balance equation and combined with the thermal resistance network topology and the thermal resistance parameters to be identified.

[0099] In this embodiment, the distribution of heat sources and heat dissipation paths within the control cabinet significantly impact the temperature of semiconductor devices. First, all major heat-generating devices need to be identified through the control cabinet's design drawings and on-site surveys. These include power semiconductor devices (such as IGBT modules and rectifier diodes), reactors, transformers, power modules, and braking resistors. These devices generate a significant amount of heat during operation, becoming critical nodes in the thermal network. For example, a typical inverter control cabinet contains 6 IGBT modules, 12 rectifier diodes, 3 reactors, and 2 power modules. Second, key heat dissipation paths need to be identified, including conduction paths (such as the heat transfer path from the device to the heat sink), convection paths (such as air duct design and airflow direction), and radiation paths. Heat dissipation paths can be identified by acquiring thermal images during equipment operation using a thermal imaging camera to observe the direction of heat flow and temperature gradient distribution. Finally, each major heat-generating device is defined as a node in the thermal network, preparing for the subsequent construction of the thermal resistance network. Next, the heat conduction paths between physically contacting devices are analyzed, such as the heat conduction connections between IGBT modules sharing the same heat sink. Then, the heat convection paths formed by airflow are analyzed, such as the heat transfer between upstream and downstream devices via fan airflow. Finally, consider the thermal radiation paths, especially the radiation impact of high-temperature devices on surrounding low-temperature devices. In practical implementation, these connections can be represented using a thermal network diagram, where nodes represent heat-generating devices and edges represent heat transfer paths. For example, the six IGBT modules in a three-phase inverter form a fully connected subnetwork, while the connection to the power supply module is through a weaker thermal connection.

[0100] Based on the thermal resistance network topology, two key thermal resistance parameters are defined: self-thermal resistance and mutual thermal resistance. Self-thermal resistance describes how the heat generated by a device affects its temperature, reflecting the thermal resistance characteristics from the device to the heat dissipation environment. For example, the self-thermal resistance of an IGBT module includes the series thermal resistances from the chip to the substrate, the substrate to the heat sink, and the heat sink to the environment. Mutual thermal resistance describes how the heat generated by one device affects the temperature of other devices, reflecting the strength of thermal coupling between devices. For example, the mutual thermal resistance between two adjacent IGBT modules depends on their distance from the heat sink and the thermal conductivity of the heat sink. For n thermal network nodes, a total of n self-thermal resistances and n(n-1) mutual thermal resistances need to be defined, forming an n×n thermal resistance matrix. The initial values ​​of these thermal resistance parameters can be estimated using thermal simulation software, but due to the complexity and uncertainty of actual heat dissipation conditions, they need to be defined as parameters to be identified and updated using online data. This parameterization method can accurately describe the complex thermal coupling effects within the control cabinet.

[0101] Based on the first law of thermodynamics and steady-state thermal equilibrium conditions, a parameterized thermal coupling effect model is constructed. Under steady-state conditions, the heat generated by each thermal network node is equal to the sum of the heat transferred out through each thermal resistance path. For node i, its temperature rise ΔTi (relative to ambient temperature) can be expressed as a function of its own loss Pi and the losses of other nodes Pj: ΔTi=Rth_ii×Pi+∑(Rth_ij×Pj), where j≠i, Rth_ii is the self-thermal resistance of node i, and Rth_ij is the mutual thermal resistance between node j and node i. Combining the equations of all nodes, a thermal coupling matrix equation is formed: [ΔT1;ΔT2;...;ΔTn]=[Rth]×[P1;P2;...;Pn], where [Rth] is an n×n thermal resistance matrix containing all self-thermal resistances and mutual thermal resistances. This matrix equation describes the thermal coupling relationship between all thermal network nodes in the control cabinet and can predict the temperature of each node under a given loss distribution. In practical applications, the temperature of semiconductor devices can be calculated by solving this system of equations, or unknown thermal resistance parameters can be deduced by measuring the temperature and known losses. This parameterized thermal coupling model provides a theoretical basis for temperature prediction and thermal management of semiconductor devices.

[0102] In one implementation, the process of using a parameter identification algorithm and utilizing a synchronously running dataset to identify and update the parameters of the device to be identified and the thermal resistance parameters to be identified, thereby obtaining an adaptive model parameter set reflecting the current operating condition and coupling state of the control cabinet, includes the following steps:

[0103] The bulk loss model and the thermal coupling effect model are integrated into a nonlinear state-space equation. The state vector in the nonlinear state-space equation includes the junction temperature and case temperature of the semiconductor device.

[0104] The device parameters and thermal resistance parameters to be identified are augmented into the state vector using slow time-varying characteristics;

[0105] Initialize the state vector and the estimated values ​​of the parameters to be identified, as well as the covariance matrix;

[0106] At each sampling time, the prediction and update steps of the extended Kalman filter are executed. The Kalman gain is calculated using the synchronous running dataset as the measurement value, the predicted values ​​of the state vector and the parameters to be identified are corrected, and the posterior optimal estimate at the current time is obtained.

[0107] The updated state vector is output periodically, and the adaptive model parameter set is extracted from the state vector.

[0108] In this implementation, the bulk loss model and the thermal coupling effect model need to be integrated into a unified state-space representation to facilitate the application of state estimation algorithms. The state-space equations consist of two parts: state equations and observation equations. The junction temperature Tj and case temperature Tc of the semiconductor device are chosen as the core state variables for the state vector because the junction temperature determines the electrical characteristics of the device, while the case temperature is a measurable physical quantity. The state equations describe the evolution of the junction and case temperatures over time, based on fundamental thermodynamic principles: the rate of change of the junction temperature is related to the bulk power loss and heat transfer from the junction to the case temperature; the rate of change of the case temperature is related to the heat transferred from the junction, the device's own power loss, and heat transfer from the case to the environment.

[0109] In practical implementation, a first-order thermal differential equation is used to describe the temperature dynamics, where heat capacity and thermal resistance are key parameters. The observation equation describes the relationship between measured values ​​(such as case temperature and ambient temperature) and state variables. Because the bulk loss model includes the nonlinear influence of junction temperature on conduction and switching characteristics, the entire state-space equation exhibits significant nonlinearity, which is why a nonlinear state estimation method is needed subsequently. To simultaneously estimate state variables and model parameters, a parameter augmentation technique is used, adding the device parameters and thermal resistance parameters to be identified as additional state variables to the state vector. The augmented state vector includes the original state variables (junction temperature Tj, case temperature Tc) and all parameters to be identified (such as conduction loss parameters a0-a2, b0-b2, switching loss parameters k1, k2, α, β, etc., and self-thermal resistance and mutual thermal resistance parameters in the thermal resistance network). These parameters change slowly physically and can be modeled as slow time-varying processes; that is, the dynamic equation of the parameters can be expressed as parameter values ​​plus small perturbations. For example, for parameter θ, its dynamic equation can be expressed as θ(k+1)=θ(k)+w(k), where w(k) is the small-variance process noise, reflecting the slow-changing characteristics of the parameter. This augmented state-space method transforms the parameter estimation problem into a state estimation problem, enabling the simultaneous estimation of temperature state and model parameters using mature state estimation techniques (such as Kalman filtering).

[0110] Before starting the parameter identification process, it is necessary to reasonably initialize the state vector and the estimated values ​​of the parameters to be identified, as well as their covariance matrices. The initial estimates of the state variables (junction temperature Tj, case temperature Tc) can be set based on the currently measured ambient temperature and the initial operating state of the device. For example, when the device is first started, it can be assumed that the junction temperature and case temperature are close to the ambient temperature. The initial estimates of the parameters to be identified are based on typical values ​​extracted from the device datasheet. For example, the conduction loss parameter can be obtained by fitting the output characteristic curve, and the thermal resistance parameter can be set from thermal simulation results or empirical values. The covariance matrix P reflects the uncertainty of the initial estimates. The diagonal element Pii represents the variance of the i-th state or parameter estimate, and the off-diagonal element Pij represents the covariance of the i-th and j-th state or parameter estimates. Typically, a smaller initial covariance for the temperature state (e.g., 1-5℃²) indicates a better grasp of the initial temperature; while a larger initial covariance for the parameters (e.g., 10%-30% of the parameter estimate) indicates higher uncertainty regarding the initial parameter values, allowing the algorithm to make larger adjustments during the identification process. Proper initialization settings can accelerate algorithm convergence and improve estimation accuracy. Extended Kalman Filter (EKF) is an effective method for state estimation of nonlinear systems. Its core idea is to linearize the nonlinear system at the current estimation point and then apply the standard Kalman filter algorithm. EKF includes two key steps: prediction and update. In the prediction step, based on the state estimate from the previous time step and the system dynamic model, the prior state estimate and error covariance matrix for the current time step are predicted. Specifically, the predicted temperature state is calculated using the nonlinear state equation, while the parameter state is assumed to be a slow time-varying process; simultaneously, the state transition matrix (system Jacobian matrix) is calculated to predict the propagation of the error covariance matrix. In the update step, the prior estimate is corrected using the current measurement data (such as shell temperature, ambient temperature, etc.). First, the measured predicted value and measurement Jacobian matrix are calculated, then the Kalman gain is calculated, and finally, the state estimate and error covariance matrix are updated. The Kalman gain determines the contribution weight of measurement information to the state update; it depends on the relative magnitudes of the prediction error covariance and the measurement noise covariance. Through this prediction-update cycle, EKF is able to provide optimal estimates of state and parameters while taking into account system nonlinearity and measurement noise.

[0111] The Extended Kalman Filter (EKF) algorithm updates the state vector at each sampling time. However, considering computational efficiency and storage requirements, it typically outputs and saves the updated state vector at predetermined intervals (e.g., per second or per minute). From the updated state vector, the currently estimated junction temperature Tj and case temperature Tc, as well as the latest estimates of all parameters to be identified, can be directly extracted, forming an adaptive model parameter set. This parameter set includes conduction loss parameters (a0-a2, b0-b2, etc.) and switching loss parameters (k1, k2, α, β, etc.) from the bulk loss model, and thermal resistance parameters (self-heating resistance and mutual thermal resistance) from the thermal coupling effect model. These parameters are adaptively updated as the system operating state changes, accurately reflecting the characteristics and thermal environment of the semiconductor device under current operating conditions. For example, as the device ages, the on-resistance increases, which is reflected in the increase of parameter b0; as heat dissipation conditions change (e.g., dust accumulation), the thermal resistance parameter will adjust accordingly. The output of this adaptive parameter set provides an accurate model basis for subsequent loss calculations, temperature predictions, and lifetime assessments, greatly improving the reliability of these analyses.

[0112] In one implementation, the real-time loss value of the semiconductor device is calculated by combining a synchronously running dataset and an adaptive model parameter set, and using a bulk loss model and a thermal coupling effect model, including the following steps:

[0113] The junction temperature estimate of the semiconductor device at the current moment is obtained from the posterior state estimate of the extended Kalman filter output;

[0114] Substitute the junction temperature estimate, the device parameters in the adaptive model parameter set updated at the current time, and the operating electrical parameters in the synchronous running dataset into the body loss model to calculate the real-time conduction loss and real-time switching loss of the semiconductor device.

[0115] The junction temperature estimate, the thermal resistance parameters in the adaptive model parameter set updated at the current time, and the temperature parameters in the synchronous running dataset are substituted into the thermal coupling effect model to calculate the real-time thermal effect loss of the semiconductor device.

[0116] The real-time loss value of the target semiconductor device is obtained by combining real-time conduction loss, real-time switching loss and real-time thermal effect loss.

[0117] In this embodiment, after completing the prediction and update steps at each sampling time, the extended Kalman filter algorithm outputs the posterior state estimate for the current time, which includes the junction temperature estimate of the semiconductor device. The junction temperature refers to the temperature of the PN junction inside the semiconductor chip, which is a key parameter determining the electrical characteristics of the device, but it cannot be directly measured because it is located inside the device. To extract the junction temperature estimate from the output state vector of the extended Kalman filter, it is only necessary to obtain the state component corresponding to the junction temperature. For example, if the first element of the state vector represents the junction temperature, then the junction temperature estimate Tj = x (1), where x is the posterior state estimate vector. The accuracy of the junction temperature estimate depends on the model accuracy and the quality of the measurement data, and it can usually achieve an accuracy within ±5℃. Using the obtained junction temperature estimate Tj, combined with the device parameters in the adaptive model parameter set and the electrical parameters in the synchronous running dataset, the real-time conduction loss and switching loss of the semiconductor device are calculated. The calculation of conduction loss first obtains the current I from the synchronous operation dataset. Then, combining the estimated junction temperature Tj and the conduction parameters (a0-a2, b0-b2, etc.) in the adaptive parameter set, the conduction voltage V(I,Tj) = V0(Tj) + R(Tj) × I is calculated, where V0(Tj) = a0 + a1 × Tj + a2 × Tj2, and R(Tj) = b0 + b1 × Tj + b2 × Tj2. The conduction loss is equal to the product of the conduction voltage and the current: Pcond = V(I,Tj) × I. The calculation of switching loss requires obtaining the current I, voltage V, and switching frequency fsw from the synchronous operation dataset. Then, combining the estimated junction temperature Tj and the switching parameters (k1, k2, α, β, etc.) in the adaptive parameter set, the single-cycle turn-on energy Eon and turn-off energy Eoff are calculated. The switching loss is equal to the sum of the single-cycle switching energies multiplied by the switching frequency: Psw = (Eon + Eoff) × fsw.

[0118] Thermal effect loss refers to the additional thermal burden borne by the target semiconductor device due to the thermal coupling effect of other heat sources within the control cabinet. To calculate thermal effect loss, temperature data from each measurement point is first obtained from the synchronous operation dataset, including the ambient temperature around the target device and the temperatures of other heat sources. Then, combined with the junction temperature estimate Tj obtained in the first step and the thermal resistance parameters (self-thermal resistance and mutual thermal resistance) in the adaptive model parameter set, the thermal coupling effect model is applied to calculate the thermal effect loss. The specific calculation process is based on the thermal balance equation, analyzing the heat exchange between the target device and other heat sources. For example, if the target device is node i, its received thermal effect loss can be expressed as Pth_i = ∑(ΔTj / Rth_ji), where j represents other heat source nodes, ΔTj is the temperature rise of node j, and Rth_ji is the mutual thermal resistance from node j to node i. The final step is to synthesize the various losses calculated above to obtain the total real-time loss value of the target semiconductor device. The comprehensive calculation considers three components: conduction loss Pcond, switching loss Psw, and thermal effect loss Pth, with the total loss value Ptotal = Pcond + Psw + Pth.

[0119] In one embodiment, the method further includes the following steps:

[0120] Real-time calculation of prediction error in the extended Kalman filter update step;

[0121] Statistical analysis was performed on the prediction errors at multiple consecutive sampling times to calculate the mean and variance of the error sequence.

[0122] When the mean and / or variance of the error sequence exceed a preset threshold, it is determined that a model mismatch or a sudden change in operating conditions has occurred.

[0123] In response to model mismatch or sudden changes in operating conditions, the terms corresponding to the parameter states in the process noise covariance matrix of the extended Kalman filter are increased to accelerate the parameter adaptation rate.

[0124] In this embodiment, the prediction error, i.e., the difference between the measured value and the predicted value, needs to be calculated in each update step of the extended Kalman filter. The prediction error calculation process first obtains the actual measured value at the current moment from the synchronously running dataset, such as shell temperature and ambient temperature. Then, based on the prior state estimate at the current moment (including junction temperature, shell temperature, and model parameters), the predicted measured value is calculated through the observation equation. The prediction error is the difference between the actual measured value and the predicted measured value. For example, if the actual measured shell temperature at the current moment is 75℃, while the shell temperature predicted based on the model is 72℃, then the prediction error is 3℃. This prediction error is not only used for the state update calculation of the Kalman filter but is also an important indicator for evaluating model performance and system state. The prediction error originates from various factors, including model structure errors, parameter errors, measurement noise, and changes in system operating conditions. By monitoring the prediction error in real time, model mismatch or sudden changes in operating conditions can be detected in a timely manner, providing a basis for subsequent adaptive adjustments. To evaluate the prediction performance of the system and detect anomalies, statistical analysis of the prediction error at multiple consecutive sampling moments is required.

[0125] In practice, a sliding time window is selected, such as the most recent 100 sampling points or the data from the most recent 10 seconds, and the statistical characteristics of the prediction error within this period are calculated. First, the mean of the error sequence is calculated, which is the arithmetic mean of all prediction errors within the window, reflecting the systematic bias of the prediction. For example, if the mean of the shell temperature prediction error within 10 seconds is 2.5℃, it indicates that the model has a tendency to continuously overestimate or underestimate. Second, the variance of the error sequence is calculated, which is the average of the squared deviations of the prediction errors within the window, reflecting the volatility of the prediction. For example, if the variance of the shell temperature prediction error is 4℃², it indicates that the prediction uncertainty is relatively large. In addition, other statistical indicators can be calculated, such as the maximum absolute value of the error, the skewness and kurtosis of the error distribution, etc., to comprehensively evaluate the prediction performance. Based on the error statistical indicators calculated in the previous step, reasonable thresholds are set for anomaly detection. Threshold settings are usually based on historical data analysis under normal operating conditions; for example, the mean threshold can be set to 3 times the normal error mean, and the variance threshold can be set to 2 times the normal error variance. When the mean of the error sequence exceeds the threshold, it indicates that there is a systematic bias in the model, which is due to continuous changes in device parameters (such as aging) or a significant shift in the operating point.

[0126] Once model mismatch or sudden changes in operating conditions are detected, the parameters of the Extended Kalman Filter (EKF) need to be adjusted to accelerate model adaptation. A key adjustment is to increase the diagonal elements in the process noise covariance matrix Q that are related to the parameters to be identified. The process noise covariance matrix describes the degree of random variation of state variables, and its diagonal element Qi represents the process noise variance of the i-th state variable. Increasing the process noise variance of the parameter states allows for greater variation in parameter estimates, thereby accelerating the parameter adaptation rate. In practice, different parameters can be adjusted specifically according to the type of anomaly detected. For example, if the shell temperature prediction error increases, the process noise variance corresponding to the thermal resistance parameter can be increased; if the switching loss prediction is inaccurate, the process noise variance corresponding to the switching loss parameter can be increased. The increase can be set to 2-10 times the original value, depending on the degree of anomaly. This adaptive adjustment mechanism enables the EKF to respond quickly to system changes and rapidly reconverge to accurate parameter estimates after sudden changes in operating conditions, improving the model's adaptability and robustness in dynamic environments. Once the system stabilizes, the process noise covariance matrix can be gradually restored to normal levels to maintain the stability of parameter estimates.

[0127] In one embodiment, the method further includes the following steps:

[0128] The system continuously accumulates the synchronous running dataset and the corresponding real-time loss value. When the accumulated data reaches a preset data threshold, a composite feature set is extracted from the accumulated synchronous running dataset.

[0129] The target value of loss residual is defined based on the dynamic cross-influence of device parameters and temperature parameters;

[0130] A pre-defined loss residual prediction model is trained using a composite feature set and the corresponding loss residual target value, so that the loss residual prediction model learns the nonlinear mapping relationship between the composite feature set and the loss residual.

[0131] After the loss residual prediction model is trained, if the real-time loss value is calculated, the real-time composite feature is extracted from the synchronous running dataset corresponding to the real-time loss value.

[0132] The real-time composite features are input into the loss residual prediction model, and the current loss residual prediction value is output in real time through the loss residual prediction model.

[0133] The current loss residual prediction value is fused with the real-time loss value to obtain the loss calculation result of the semiconductor device after residual compensation.

[0134] In this embodiment, during system operation, the synchronous running dataset and corresponding calculated real-time loss values ​​are continuously saved to form a historical database. When the accumulated data reaches a preset threshold (e.g., 10,000 samples or data from 7 consecutive days of operation), the feature extraction process begins. A composite feature set, including static and dynamic features, is extracted from the accumulated data. Static features include basic parameters such as current amplitude, voltage amplitude, power factor, and ambient temperature; dynamic features include parameters describing the dynamic behavior of the system, such as current change rate, temperature gradient, and voltage fluctuation index; and cross-features include indicators reflecting the interaction of parameters, such as the product of current and temperature, and the combination of voltage and switching frequency. These composite features can comprehensively capture the operating characteristics of semiconductor devices under different operating conditions, providing rich input information for subsequent residual prediction models. The target value of the loss residual is defined as the difference between the actual loss and the loss calculated by the model, reflecting the complex nonlinear relationships that the model fails to fully capture. Specifically, the actual loss of the semiconductor device is first measured under typical operating conditions using high-precision measuring equipment (such as a calorimeter or precision electrical parameter analyzer), or the actual loss is indirectly calculated through the thermal balance equation. The synchronously running dataset is then input into the bulk loss model and the thermal coupling effect model to calculate the theoretical loss value. The loss residual is the difference between the actual loss and the theoretical loss. This residual includes the dynamic cross-influence of device parameters and temperature parameters that the model fails to describe, such as the nonlinear effect of temperature on switching characteristics and the effect of current waveform on heat distribution, providing a learning target for the residual prediction model.

[0135] Choose a suitable machine learning model for nonlinear mappings, such as gradient boosting decision trees, deep neural networks, or support vector regression, as the loss residual prediction model. Use the composite feature set extracted in the first step as input and the loss residual target value defined in the second step as output, and train the model using supervised learning methods. Cross-validation is used during training to evaluate model performance and avoid overfitting. Model hyperparameters are optimized using grid search or Bayesian optimization methods. The trained model can learn complex nonlinear mapping relationships and capture subtle effects that are difficult to describe by traditional physical models, such as the impact of uneven temperature distribution on loss and additional losses caused by current harmonic components, thus compensating for the shortcomings of physical models. When real-time loss values ​​need to be calculated, first extract the same real-time composite features defined during training from the current synchronously running dataset. Maintain consistency in the feature extraction process to ensure that training and prediction use the same feature space. Real-time feature extraction needs to be implemented efficiently, typically using sliding window techniques to process time-series data and calculate various statistical features. For computationally complex features, approximate calculations or simplified algorithms can be used to improve processing speed. After feature extraction, perform necessary normalization or standardization to ensure that the feature distribution is consistent with the training data, ensuring the accuracy of model predictions. This step transforms the raw running data into structured feature vectors, preparing for the next step of residual prediction.

[0136] The extracted real-time composite features are input into a trained loss residual prediction model. Based on the learned nonlinear mapping, the model outputs a predicted loss residual value for the current operating condition. The prediction process is fast, typically completed within milliseconds, meeting real-time computing requirements. If a deep learning model is used, inference speed can be optimized through model quantization or pruning techniques. The predicted loss residual reflects additional losses or overestimations that the physical model failed to capture; its sign and magnitude provide important information about the physical model's accuracy. The residual prediction value can also include a confidence interval, indicating the prediction uncertainty. High prediction uncertainty can trigger warnings or adopt a conservative estimation strategy to ensure system safety. The final step is to fuse the real-time loss value calculated by the physical model with the loss residual predicted by the machine learning model to obtain a more accurate loss estimate. The fusion method is direct addition: compensated loss = real-time loss value + predicted loss residual value. A positive residual indicates that the physical model underestimated the actual loss; a negative residual indicates that the physical model overestimated the actual loss. In some cases, a weighted fusion strategy can be used to adjust the weights of the residuals based on the similarity of operating conditions or the confidence level of the prediction. The loss calculation results after residual compensation combine the theoretical basis of the physical model and the adaptive capability of the data-driven model, which can more accurately reflect the actual loss of semiconductor devices under complex operating conditions, improve the accuracy of loss estimation, and can typically reduce the error by 30%-50%.

[0137] The present invention also discloses a semiconductor device loss intelligent calculation system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the semiconductor device loss intelligent calculation method as described in any of the above embodiments.

[0138] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.

[0139] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of internal storage units and external storage devices of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0140] The present invention also discloses a computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to be configured to perform the intelligent calculation method for semiconductor device losses described in any of the above embodiments.

[0141] The computer program can be stored in a machine-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The machine-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the machine-readable medium includes, but is not limited to, the above-mentioned components.

[0142] The intelligent calculation method for semiconductor device loss in the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0143] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0144] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for intelligent calculation of semiconductor device losses, characterized in that, Includes the following steps: Real-time acquisition of the operating electrical parameters of semiconductor devices in the control cabinet and the temperature parameters of multiple preset temperature measurement points in the vicinity of the semiconductor devices to form a synchronous operation dataset containing timestamps; A parameterized bulk loss model for semiconductor devices is established based on semiconductor physics principles. The bulk loss model includes device parameters to be identified that describe the changes in conduction and switching losses of semiconductor devices with operating electrical parameters and junction temperature. A parameterized thermal coupling effect model is established based on the physical layout and heat transfer path of the control cabinet. The thermal coupling effect model uses the thermal resistance parameter to be identified to describe the heat transfer relationship between semiconductor devices and other heat sources in the control cabinet. A parameter identification algorithm is used to identify and update the parameters of the device to be identified and the thermal resistance parameters to be identified using a synchronous running dataset, thereby obtaining an adaptive model parameter set that reflects the current operating condition and coupling state of the control cabinet; By combining the synchronous running dataset and the adaptive model parameter set, and by calculating the real-time loss value of the semiconductor device through the body loss model and the thermal coupling effect model, the real-time loss value of the semiconductor device is obtained.

2. The intelligent calculation method for semiconductor device losses according to claim 1, characterized in that, The process of acquiring the operating electrical parameters of semiconductor devices inside the control cabinet in real time, as well as the temperature parameters of multiple preset temperature measurement points in the vicinity of the semiconductor devices, to form a synchronous operating dataset containing timestamps includes the following steps: The operating electrical parameters of semiconductor devices within the control cabinet are captured by pre-configured high-bandwidth current sensors and high-impedance voltage sensors. Temperature sensors are installed at predetermined environmental locations on the surface of the semiconductor device's housing and heat sink, as well as inside the control cabinet, to collect the temperature parameters of the semiconductor device. A precise time protocol is used to maintain time consistency for all operating electrical and temperature parameters; The sampling rate is set according to the dynamic characteristics of the operating electrical and temperature parameters; The collected operating electrical and temperature parameters are digitally filtered, and amplitude correction is performed based on sensor calibration data. The preprocessed operating electrical and temperature parameters are aligned to a unified time axis through interpolation to form a synchronized operating dataset.

3. The intelligent calculation method for semiconductor device losses according to claim 1, characterized in that, The establishment of a parameterized bulk loss model for semiconductor devices based on semiconductor physics principles includes the following steps: The on-state voltage drop of a semiconductor device is expressed as the sum of a zero-current voltage drop term and an equivalent on-resistance term to construct a parameterized on-state loss model. Both the zero-current voltage drop and the equivalent on-resistance are parameterized as polynomial functions of the junction temperature. A parameterized switching loss model is constructed by expressing the single turn-on energy and single turn-off energy of a semiconductor device as power-law functions of junction temperature, switching current, and blocking voltage, respectively, combined with polynomial interpolation. The junction temperature correlation coefficient in the conduction loss model and the correlation coefficients of junction temperature, switching current and blocking voltage in the switching loss model are collectively defined as the set of parameters of the device to be identified. Initial estimates are set for the parameters of the device to be identified based on typical curves extracted from the device datasheets of semiconductor devices. The conduction loss model and the switching loss model are integrated into a bulk loss model for semiconductor devices. The bulk loss model takes operating electrical parameters, temperature parameters, and parameters of the device to be identified as model inputs.

4. The intelligent calculation method for semiconductor device losses according to claim 1, characterized in that, The establishment of a parameterized thermal coupling effect model based on the physical layout and heat transfer path of the control cabinet includes the following steps: Based on the physical layout of the control cabinet, identify all major heat-generating devices and key heat dissipation paths within the control cabinet, and treat all major heat-generating devices as thermal network nodes. By combining the principles of heat transfer and connecting all thermal network nodes through key heat dissipation paths, a thermal resistance network topology for the control cabinet is constructed. The thermal resistance parameters to be identified are defined based on the heat transfer effect of the thermal resistance network topology. The thermal resistance parameters to be identified include the self-thermal resistance used to represent the heat transfer from the thermal network node itself to the thermal resistance network topology, and the mutual thermal resistance used to represent the heat generated by one thermal network node to another thermal network node. A parameterized thermal coupling effect model is constructed based on the steady-state thermal balance equation and combined with the thermal resistance network topology and the thermal resistance parameters to be identified.

5. The intelligent calculation method for semiconductor device losses according to claim 1, characterized in that, The process of using a parameter identification algorithm and a synchronously running dataset to identify and update the parameters of the device to be identified and the thermal resistance parameters to be identified, in order to obtain an adaptive model parameter set reflecting the current operating condition and coupling state of the control cabinet, includes the following steps: The bulk loss model and the thermal coupling effect model are integrated into a nonlinear state-space equation. The state vector in the nonlinear state-space equation includes the junction temperature and case temperature of the semiconductor device. The device parameters and thermal resistance parameters to be identified are augmented into the state vector using slow time-varying characteristics; Initialize the state vector and the estimated values ​​of the parameters to be identified, as well as the covariance matrix; At each sampling time, the prediction and update steps of the extended Kalman filter are executed. The Kalman gain is calculated using the synchronous running dataset as the measurement value, the predicted values ​​of the state vector and the parameters to be identified are corrected, and the posterior optimal estimate at the current time is obtained. The updated state vector is output periodically, and the adaptive model parameter set is extracted from the state vector.

6. The intelligent calculation method for semiconductor device losses according to claim 5, characterized in that, The process of combining the synchronously running dataset and the adaptive model parameter set, and calculating the real-time loss value of the semiconductor device through the bulk loss model and the thermal coupling effect model, includes the following steps: The junction temperature estimate of the semiconductor device at the current moment is obtained from the posterior state estimate of the extended Kalman filter output; Substitute the junction temperature estimate, the device parameters in the adaptive model parameter set updated at the current time, and the operating electrical parameters in the synchronous running dataset into the body loss model to calculate the real-time conduction loss and real-time switching loss of the semiconductor device. The junction temperature estimate, the thermal resistance parameters in the adaptive model parameter set updated at the current time, and the temperature parameters in the synchronous running dataset are substituted into the thermal coupling effect model to calculate the real-time thermal effect loss of the semiconductor device. The real-time loss value of the target semiconductor device is obtained by combining real-time conduction loss, real-time switching loss and real-time thermal effect loss.

7. The intelligent calculation method for semiconductor device losses according to claim 5, characterized in that, The method further includes the following steps: Real-time calculation of prediction error in the extended Kalman filter update step; Statistical analysis was performed on the prediction errors at multiple consecutive sampling times to calculate the mean and variance of the error sequence. When the mean and / or variance of the error sequence exceed a preset threshold, it is determined that a model mismatch or a sudden change in operating conditions has occurred. In response to model mismatch or sudden changes in operating conditions, the terms corresponding to the parameter states in the process noise covariance matrix of the extended Kalman filter are increased to accelerate the parameter adaptation rate.

8. The intelligent calculation method for semiconductor device losses according to claim 1, characterized in that, The method further includes the following steps: The synchronous running dataset and its corresponding real-time loss value are continuously accumulated. When the accumulated data volume reaches a preset data volume threshold, a composite feature set is extracted from the accumulated synchronous running dataset. The target value of loss residual is defined based on the dynamic cross-influence of device parameters and temperature parameters; A pre-defined loss residual prediction model is trained using a composite feature set and the corresponding loss residual target value, so that the loss residual prediction model learns the nonlinear mapping relationship between the composite feature set and the loss residual. After the loss residual prediction model is trained, if the real-time loss value is calculated, the real-time composite feature is extracted from the synchronous running dataset corresponding to the real-time loss value. The real-time composite features are input into the loss residual prediction model, and the current loss residual prediction value is output in real time through the loss residual prediction model. The current loss residual prediction value is fused with the real-time loss value to obtain the loss calculation result of the semiconductor device after residual compensation.

9. A semiconductor device loss intelligent calculation system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent calculation method for semiconductor device losses as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the intelligent calculation method for semiconductor device losses according to any one of claims 1 to 8.

Images