Dynamic monitoring method and system for multi-physics coupling effect of third-generation semiconductor device

Abstract

The invention provides a third-generation semiconductor device multi-physical field coupling effect dynamic monitoring method and system, and relates to the technical field of semiconductors, and the method comprises the steps: collecting the data of a temperature field, an electric field, a magnetic field and a stress field through a complementary sensor array, constructing a feature tensor, inputting the feature tensor into a pre-trained graph neural network, and extracting a coupling effect dynamic evolution rule; on-line analysis and prediction are carried out by using an adaptive time window and a recurrent neural network, a multi-field coupling correlation degree evaluation model is established, an early warning level is determined according to the deviation between the coupling correlation degree and an early warning threshold value, and device risk evaluation and maintenance suggestions are generated, so that early warning of the failure risk of the semiconductor device is realized.

Description

Technical Field

[0001] The present invention relates to semiconductor technology, and in particular to a method and system for dynamically monitoring multi-physical field coupling effects of third-generation semiconductor devices. Background Art

[0002] Third-generation semiconductor materials (such as gallium nitride, silicon carbide, and gallium oxide) demonstrate enormous potential for application in high-frequency, high-temperature, high-power, and high-efficiency electronic devices due to their wide bandgap, high breakdown electric field, high thermal conductivity, and high electron saturation velocity. They are gradually replacing traditional silicon-based semiconductor materials. Third-generation semiconductor devices are widely used in aerospace, power electronics, communications equipment, and new energy vehicles.

[0003] During the operation of third-generation semiconductor devices, multiple physical fields—temperature, electric, magnetic, and stress fields—coexist and interact, forming complex coupling effects. These coupling effects directly impact device performance stability and reliability, and can even cause device failure under extreme operating conditions. Therefore, real-time monitoring and prediction of these multi-field coupling effects are crucial for ensuring the safe operation of third-generation semiconductor devices.

[0004] Traditional monitoring methods usually only focus on a single physical field parameter, such as monitoring only temperature or current, and are unable to fully capture the complex coupling relationship between multiple physical fields, resulting in incomplete assessment of device status and inability to accurately predict potential failure risks.

[0005] Existing technologies lack the ability to perform real-time dynamic analysis of multi-physical field data and mostly use static threshold judgment methods, which are difficult to adapt to the dynamic changing characteristics of third-generation semiconductor devices under different working conditions. They are prone to missed reports or false reports, affecting the device's service life and system reliability.

[0006] Existing monitoring systems fail to effectively establish coupling correlation models between multiple physical fields, cannot quantitatively analyze the interaction mechanisms between various physical fields, and find it difficult to fundamentally understand the physical nature of device failures. Therefore, they are unable to provide targeted preventive maintenance strategies, increasing maintenance costs and downtime risks. Summary of the Invention

[0007] The embodiments of the present invention provide a method and system for dynamically monitoring multi-physical field coupling effects of third-generation semiconductor devices, which can solve the problems in the prior art.

[0008] A first aspect of an embodiment of the present invention provides a method for dynamically monitoring multi-physical field coupling effects in a third-generation semiconductor device, comprising: Collecting temperature field data, electric field data, magnetic field data, and stress field data in third-generation semiconductor devices, and synchronously acquiring the time-series variation characteristics of each physical field data in the temperature field data, the electric field data, the magnetic field data, and the stress field data through a complementary sensor array; Constructing a feature tensor of the multi-physics field coupling effect according to the temporal variation characteristics, inputting the feature tensor into a pre-trained graph neural network model, and extracting the dynamic evolution law of the multi-physics field coupling effect based on the graph neural network model; An adaptive time window sliding mechanism is used to perform online analysis on the dynamic evolution law, a recursive neural network is used to dynamically predict the multi-physics field coupling effect, and a multi-dimensional warning threshold is generated according to the prediction results of the multi-physics field coupling effect; Constructing a multi-field coupling correlation evaluation model based on the dynamic evolution law, inputting the temperature field data, the electric field data, the magnetic field data, and the stress field data into the multi-field coupling correlation evaluation model to obtain the coupling correlation between the physical fields; According to the temporal variation trend of the coupling correlation and the multidimensional warning threshold, the deviation between the coupling correlation and the multidimensional warning threshold is calculated, the warning level is determined, and a targeted device failure risk assessment report and preventive maintenance recommendations are generated based on the warning level.

[0009] The dynamic evolution law is analyzed online using an adaptive time window sliding mechanism, and the multi-physics field coupling effect at different time scales is dynamically predicted using a recursive neural network. The multi-dimensional warning threshold is generated based on the prediction results of the multi-physics field coupling effect, including: Building an adaptive time window based on the continuous sampling data of the dynamic evolution law, wherein the length of the adaptive time window is dynamically adjusted according to the change rate of the dynamic evolution law; Performing sliding sampling on the dynamic evolution law using the adaptive time window, and obtaining characteristic data of multi-physical field coupling effects at different time points through the sliding sampling, wherein the characteristic data includes physical field intensity values ​​and coupling correlations; Grouping the feature data according to different time scales to form multiple time scale sequences, and inputting the multiple time scale sequences into a pre-trained recursive neural network; Extracting the time series features of the multiple time scale sequences through the hidden layer state of the recurrent neural network, predicting the evolution trend of the multi-physics field coupling effect based on the time series features, and obtaining prediction results at different time scales; The warning reference value of each physical field is calculated according to the prediction result, and a multi-dimensional warning threshold is generated based on the warning reference value combined with the historical fluctuation range of each physical field. The multi-dimensional warning threshold is updated in real time as the prediction result changes.

[0010] A multi-field coupling correlation evaluation model is constructed based on the dynamic evolution law, and the temperature field data, the electric field data, the magnetic field data, and the stress field data are input into the multi-field coupling correlation evaluation model to obtain the coupling correlation between the physical fields, including: Calculating sampling time deviations among the temperature field data, the electric field data, the magnetic field data, and the stress field data, and performing data synchronization processing according to the sampling time deviations; performing dimension conversion on the synchronized temperature field data, the electric field data, the magnetic field data, and the stress field data to obtain normalized data; Calculating the change rate of the normalized data at adjacent sampling moments, performing a weighted combination of the change rate and the time integral of the normalized data to construct a dynamic evolution feature; The information entropy of each physical field is calculated based on the normalized data, and a coupling weight matrix is ​​constructed according to the information entropy; the second-order statistics of the dynamic evolution characteristics are calculated within a sliding time window, and the local coupling coefficient between the physical fields is calculated in combination with the coupling weight matrix, and the local coupling coefficient represents the interaction strength between the physical fields; the second-order statistics and the local coupling coefficient are weightedly superimposed to obtain the coupling correlation between the physical fields, and the coupling correlation represents the degree of coupling of the multi-physical fields as a whole.

[0011] Calculating the information entropy of each physical field based on the normalized data, constructing a coupling weight matrix according to the information entropy; calculating the second-order statistics of the dynamic evolution characteristics within a sliding time window, and calculating the local coupling coefficient between the physical fields in combination with the coupling weight matrix includes: Calculating probability distributions for the normalized multiple physical field data sequences respectively, calculating the information entropy of each physical field data sequence based on the probability distribution, and obtaining an information entropy set of the multiple physical field data sequences; The information entropy set is constructed as a coupling weight matrix, wherein the elements of the coupling weight matrix are determined by the ratio of the product of the information entropies of the corresponding two physical fields to the sum of the products of the information entropies of all physical fields; Calculating, within a sliding time window of a preset length, second-order statistics of the dynamic evolution characteristics of the plurality of physical field data sequences, the second-order statistics including an autocorrelation function of the physical field data sequence and a cross-correlation function between different physical field data sequences; The local coupling coefficient between the physical fields is calculated based on the coupling weight matrix, the autocorrelation function and the cross-correlation function.

[0012] Calculating the deviation between the coupling correlation and the multi-dimensional warning threshold according to the temporal variation trend of the coupling correlation and the multi-dimensional warning threshold, determining a warning level, and generating a targeted component failure risk assessment report and preventive maintenance recommendations based on the warning level include: Calculating a standardized deviation between the coupling correlation degree and the multidimensional warning threshold, constructing a dynamic weight coefficient based on the standardized deviation, and accumulating the product of the dynamic weight coefficient and the standardized deviation to obtain a comprehensive deviation index; Inputting the comprehensive deviation index into a preset membership function, calculating the membership value of the comprehensive deviation index for different risk levels, and performing weighted summation of the membership value and the corresponding risk level value to obtain a fuzzy risk degree; Determining a warning level based on the current value and rate of change of the fuzzy risk, wherein the warning level is dynamically adjusted by an adaptive mapping function that is updated in real time according to the changing trend of the fuzzy risk; The coupling correlation degree, the standardized deviation, the fuzzy risk degree, and historical fault characteristics are subjected to multimodal feature fusion to construct a fused feature vector; a device failure risk assessment report is obtained based on the fused feature vector and the warning level, and preventive maintenance recommendations are calculated based on the device failure risk assessment report and maintenance costs.

[0013] Obtaining a device failure risk assessment report based on the fused feature vector and the warning level, and calculating a preventive maintenance recommendation based on the device failure risk assessment report and the maintenance cost includes: Inputting the fused feature vector into a preset multi-physics field coupling state assessment model to obtain a device state assessment result; Constructing an adaptive warning threshold based on the device status assessment result, wherein the adaptive warning threshold is determined by the sum of the product of a baseline threshold and a dynamic adjustment coefficient and a correction term, the dynamic adjustment coefficient is dynamically updated according to the device status assessment result, and the correction term is calculated based on historical warning data; Inputting the device status assessment result, the adaptive warning threshold and the digital twin status data into the multi-physics field coupling status assessment model to obtain a device failure risk assessment report; Constructing a cost-effectiveness strategy based on the device failure risk assessment report, the maintenance cost data, and the failure risk cost data, wherein an objective function of the cost-effectiveness strategy is a minimum value of the sum of the maintenance cost and the failure risk cost multiplied by the failure probability; A maintenance strategy optimization result is generated based on the cost-effectiveness strategy, wherein the maintenance strategy optimization result satisfies the constraint condition that the equipment risk assessment result is less than a preset risk threshold, and a preventive maintenance recommendation is obtained.

[0014] A second aspect of an embodiment of the present invention provides a dynamic monitoring system for multi-physical field coupling effects of a third-generation semiconductor device, comprising: The first unit is used to collect temperature field data, electric field data, magnetic field data and stress field data in the third-generation semiconductor device, and synchronously obtain the time series change characteristics of each physical field data in the temperature field data, the electric field data, the magnetic field data and the stress field data through a complementary sensor array; The second unit is used to construct a feature tensor of the multi-physics field coupling effect according to the time series change characteristics, input the feature tensor into a pre-trained graph neural network model, and extract the dynamic evolution law of the multi-physics field coupling effect based on the graph neural network model; The third unit is used to perform online analysis of the dynamic evolution law using an adaptive time window sliding mechanism, dynamically predict the multi-physics field coupling effect at different time scales through a recursive neural network, and generate a multi-dimensional warning threshold based on the prediction results of the multi-physics field coupling effect; A fourth unit is configured to construct a multi-field coupling correlation evaluation model based on the dynamic evolution law, input the temperature field data, the electric field data, the magnetic field data, and the stress field data into the multi-field coupling correlation evaluation model, and obtain the coupling correlation between the physical fields; The fifth unit is used to calculate the deviation between the coupling correlation and the multidimensional warning threshold according to the time series change trend of the coupling correlation and the multidimensional warning threshold, determine the warning level, and generate a targeted device failure risk assessment report and preventive maintenance recommendations based on the warning level.

[0015] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0016] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0017] The beneficial effects of this application are as follows: The present invention uses a complementary sensor array to synchronously collect temperature field, electric field, magnetic field and stress field data to achieve comprehensive monitoring of the multi-physical field coupling effects of third-generation semiconductor devices, overcoming the limitations of traditional single physical field monitoring methods and significantly improving monitoring accuracy and comprehensiveness.

[0018] The present invention combines graph neural networks and recursive neural networks to construct a multi-field coupling correlation evaluation model, which can accurately capture the complex coupling relationships and dynamic evolution laws between physical fields. It uses an adaptive time window sliding mechanism for online analysis, realizes real-time prediction of device performance changes, and provides an effective means for early identification of device anomalies.

[0019] The present invention calculates the warning level based on the deviation of multi-dimensional warning thresholds and coupling correlation, and automatically generates a device failure risk assessment report and preventive maintenance recommendations, effectively reducing the risk of device failure, extending the service life of third-generation semiconductor devices, and improving the overall reliability of the system, with important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the process of a method for dynamically monitoring multi-physical field coupling effects of a third-generation semiconductor device according to an embodiment of the present invention; Figure 2 This is a dynamic analysis flow chart of the multi-physics field coupling intelligent early warning system according to an embodiment of the present invention; Figure 3 Schematic diagram showing performance comparison of multi-field coupling correlation evaluation models according to an embodiment of the present invention; Figure 4 This is a flowchart of equipment risk assessment and preventive maintenance decision-making based on multi-dimensional feature fusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0022] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0023] Figure 1 FIG. 1 is a flow chart of a method for dynamically monitoring multi-physical field coupling effects of a third-generation semiconductor device according to an embodiment of the present invention. Figure 1 As shown, the method includes: Collecting temperature field data, electric field data, magnetic field data, and stress field data in third-generation semiconductor devices, and synchronously acquiring the time-series variation characteristics of each physical field data in the temperature field data, the electric field data, the magnetic field data, and the stress field data through a complementary sensor array; Constructing a feature tensor of the multi-physics field coupling effect according to the temporal variation characteristics, inputting the feature tensor into a pre-trained graph neural network model, and extracting the dynamic evolution law of the multi-physics field coupling effect based on the graph neural network model; An adaptive time window sliding mechanism is used to perform online analysis on the dynamic evolution law, a recursive neural network is used to dynamically predict the multi-physics field coupling effect, and a multi-dimensional warning threshold is generated according to the prediction results of the multi-physics field coupling effect; Constructing a multi-field coupling correlation evaluation model based on the dynamic evolution law, inputting the temperature field data, the electric field data, the magnetic field data, and the stress field data into the multi-field coupling correlation evaluation model to obtain the coupling correlation between the physical fields; According to the temporal variation trend of the coupling correlation and the multidimensional warning threshold, the deviation between the coupling correlation and the multidimensional warning threshold is calculated, the warning level is determined, and a targeted device failure risk assessment report and preventive maintenance recommendations are generated based on the warning level.

[0024] In an optional embodiment, an adaptive time window sliding mechanism is used to perform online analysis of the dynamic evolution law, and a recursive neural network is used to dynamically predict the multi-physics field coupling effect at different time scales. The multi-dimensional warning threshold is generated based on the prediction results of the multi-physics field coupling effect, including: Building an adaptive time window based on the continuous sampling data of the dynamic evolution law, wherein the length of the adaptive time window is dynamically adjusted according to the change rate of the dynamic evolution law; Performing sliding sampling on the dynamic evolution law using the adaptive time window, and obtaining characteristic data of multi-physical field coupling effects at different time points through the sliding sampling, wherein the characteristic data includes physical field intensity values ​​and coupling correlations; Grouping the feature data according to different time scales to form multiple time scale sequences, and inputting the multiple time scale sequences into a pre-trained recursive neural network; Extracting the time series features of the multiple time scale sequences through the hidden layer state of the recurrent neural network, predicting the evolution trend of the multi-physics field coupling effect based on the time series features, and obtaining prediction results at different time scales; The warning reference value of each physical field is calculated according to the prediction result, and a multi-dimensional warning threshold is generated based on the warning reference value combined with the historical fluctuation range of each physical field. The multi-dimensional warning threshold is updated in real time as the prediction result changes.

[0025] like Figure 2 As shown, the method includes: A technical solution for predicting multi-physics field coupling effects using a recurrent neural network. In this implementation, the system first constructs an adaptive time window based on continuously sampled data that dynamically evolves. Specifically, the system determines the length of the time window by calculating the rate of change of the sampled data over a continuous time period.

[0026] When a rapid change in a physical field parameter is detected, the window length is automatically shortened, for example, from the standard 60 minutes to 30 minutes. When the parameter changes slowly, the window length is correspondingly extended, for example, from 60 minutes to 120 minutes. This dynamic adjustment mechanism ensures that the system can capture rapidly changing physical field characteristics while conserving computing resources during the stable phase.

[0027] Using the constructed adaptive time window, the system performs sliding sampling of the dynamic evolution law. In actual applications, the sliding step is set to 25% of the window length. For example, for a 60-minute window, each sliding step is 15 minutes. Within each sliding window, the system extracts the intensity value and coupling correlation of the physical field as characteristic data. The physical field intensity value is directly obtained from the sensor data, such as the Celsius value of the temperature field, the Pascal value of the pressure field, and the MPa value of the stress field. The coupling correlation is obtained by calculating the correlation coefficient between different physical fields. For example, the coupling correlation between the temperature field and the stress field may fluctuate between 0.75 and 0.92, indicating that the two have a strong mutual influence.

[0028] The collected feature data is grouped according to different time scales. In this example, three time scales are set: short-term (5 minutes), medium-term (30 minutes), and long-term (120 minutes). The short-term scale is mainly used to capture instantaneous changes in the physical field, the medium-term scale is used to identify fluctuation trends in the physical field, and the long-term scale is used to analyze the overall evolution of the physical field.

[0029] For each time scale, the system constructs a corresponding feature sequence. For example, for the temperature field, the short-term scale may contain sampling values ​​every 5 seconds, forming a sequence of 60 data points; the medium-term scale may contain sampling values ​​every 30 seconds, forming 60 data points; and the long-term scale may contain sampling values ​​every 2 minutes, also forming a sequence of 60 data points.

[0030] The constructed multiple timescale sequences are fed into a pretrained recurrent neural network for processing. This recurrent neural network uses a long short-term memory (LSTM) architecture and consists of an input layer, hidden layers, and an output layer. The number of input layer nodes matches the feature data dimensions, the hidden layer is set to 128 neurons, and the output layer dimensions correspond to the prediction target.

[0031] The network training used 100,000 sets of physical field data collected historically, optimized the network parameters through the back-propagation algorithm, used the mean square error as the loss function, set the learning rate to 0.001, and iterated 200 rounds. The final prediction accuracy on the validation set reached 92.5%.

[0032] Recurrent neural networks extract temporal features of time-scale sequences through their hidden states. The hidden states can capture both long-term dependencies and short-term variations in the data sequence, which is particularly important for predicting multi-physics coupling effects.

[0033] In the coupled analysis of pressure and stress fields, the hidden state can identify the time delay effect of stress changes caused by increased pressure. This delay typically varies by 5-15 seconds under different operating conditions. Based on the extracted time series features, the system predicts the evolution trend of the multi-physics field coupling effect, providing short-term (next 5-15 minutes), medium-term (next 1-2 hours), and long-term (next 4-8 hours) forecasts. The forecast results include the intensity values ​​of each physical field and the changing trend of the coupling correlation.

[0034] Based on the prediction results, the system calculates a warning baseline value for each physical field. This warning baseline takes into account the degree of deviation between the predicted intensity value of the physical field and the historical average. For example, if the predicted value of the temperature field exceeds the historical average by 20%, and the predicted upward trend lasts for more than 30 minutes, the warning baseline value for the temperature field will be set to 1.15 times the current value. Based on the calculated warning baseline value, the system combines the historical fluctuation range of each physical field to generate multi-dimensional warning thresholds.

[0035] The historical fluctuation range is determined by counting the standard deviation of the physical field values ​​in the past three months, and is usually taken as 2-3 times the standard deviation. The final multi-dimensional warning threshold includes the warning baseline value plus or minus the adjusted historical fluctuation range. These thresholds are updated in real time as the prediction results change, and the update frequency matches the shortest prediction time scale, which is once every 5 minutes in this embodiment. When the monitored physical field parameters exceed the warning threshold, the system will trigger an alarm of the corresponding level and activate the preset emergency plan.

[0036] In an optional embodiment, a multi-field coupling correlation evaluation model is constructed based on the dynamic evolution law, and the temperature field data, the electric field data, the magnetic field data, and the stress field data are input into the multi-field coupling correlation evaluation model to obtain the coupling correlation between the physical fields, including: Calculating sampling time deviations among the temperature field data, the electric field data, the magnetic field data, and the stress field data, and performing data synchronization processing according to the sampling time deviations; performing dimension conversion on the synchronized temperature field data, the electric field data, the magnetic field data, and the stress field data to obtain normalized data; Calculating the change rate of the normalized data at adjacent sampling moments, performing a weighted combination of the change rate and the time integral of the normalized data to construct a dynamic evolution feature; The information entropy of each physical field is calculated based on the normalized data, and a coupling weight matrix is ​​constructed according to the information entropy; the second-order statistics of the dynamic evolution characteristics are calculated within a sliding time window, and the local coupling coefficient between the physical fields is calculated in combination with the coupling weight matrix, and the local coupling coefficient represents the interaction strength between the physical fields; the second-order statistics and the local coupling coefficient are weightedly superimposed to obtain the coupling correlation between the physical fields, and the coupling correlation represents the degree of coupling of the multi-physical fields as a whole.

[0037] During the data collection phase, the equipment monitoring system collects temperature field data, electric field data, magnetic field data, and stress field data through sensors arranged at different locations.

[0038] Temperature field data is collected by multiple thermocouples with a measurement range of -50°C to 1200°C and a sampling frequency of 100Hz; electric field data is collected by electric field strength sensors with a measurement range of 0-500kV / m and a sampling frequency of 200Hz; magnetic field data is collected by Hall sensors with a measurement range of 0-2000μT and a sampling frequency of 150Hz; stress field data is collected by strain gauges with a measurement range of 0-1000MPa and a sampling frequency of 120Hz.

[0039] Perform synchronization processing on the collected data and calculate the sampling time deviation. For example, in a certain measurement, the sampling time of the temperature field data is 10:00:00.000, the sampling time of the electric field data is 10:00:00.005, the sampling time of the magnetic field data is 10:00:00.003, and the sampling time of the stress field data is 10:00:00.008. The calculated deviations from the reference time are 0ms, 5ms, 3ms, and 8ms, respectively.

[0040] The data were resampled using linear interpolation based on the time offset to obtain a time-aligned dataset. In this process, all data were unified to a sampling frequency of 100 Hz, and the sampling time of the temperature field was selected as the benchmark for interpolation of the other physical field data.

[0041] After synchronization processing, the data is dimensionalized and normalized. Taking temperature field data as an example, the original data range is 25.5°C to 78.3°C. By subtracting the minimum value and dividing by the difference between the maximum and minimum values, the data is converted to the range of 0-1, resulting in normalized temperature data such as 0.135, 0.246, 0.358, etc.

[0042] The electric field data were converted from the original values ​​of 5.8 kV / m to 125.3 kV / m to normalized values ​​of 0.041 to 0.892; the magnetic field data were converted from the original values ​​of 87.5 μT to 1245.8 μT to normalized values ​​of 0.072 to 0.967; and the stress field data were converted from the original values ​​of 25.6 MPa to 487.9 MPa to normalized values ​​of 0.056 to 0.853.

[0043] Calculate the rate of change of the normalized data at adjacent sampling points. For the temperature field, for example, the temperature value at time t is 0.246, and the temperature value at time t+1 is 0.251, resulting in a rate of change of 0.005. For 10 consecutive sampling points, the sequence of the rate of change of the temperature field is 0.005, 0.008, -0.003, 0.002, 0.007, 0.001, -0.006, 0.004, 0.009, and 0.002. The time integral of the normalized data is also calculated. For a data window of 100 sampling points, the time integral of the temperature field is 24.37, the time integral of the electric field is 38.56, the time integral of the magnetic field is 45.92, and the time integral of the stress field is 31.28.

[0044] The dynamic evolution signature is constructed by weighting the rate of change and the time integral. The weight of the rate of change is set to 0.7, and the weight of the time integral is set to 0.3. The dynamic evolution signature is calculated for each physical field.

[0045] For the temperature field, the dynamic evolution characteristics at a certain moment are 0.7×0.005+0.3×24.37=7.315; for the electric field, the dynamic evolution characteristics are 0.7×0.012+0.3×38.56=11.576; for the magnetic field, the dynamic evolution characteristics are 0.7×0.008+0.3×45.92=13.782; for the stress field, the dynamic evolution characteristics are 0.7×0.015+0.3×31.28=9.395.

[0046] The information entropy of each physical field was calculated based on the normalized data. The data was divided into 10 intervals at 0.1 intervals, and the percentage of data points in each interval was calculated to obtain a probability distribution. The probability of each interval for the temperature field was 0.05, 0.12, 0.18, 0.25, 0.15, 0.08, 0.07, 0.04, 0.03, and 0.03, respectively, resulting in an information entropy of 2.15. The information entropy for the electric field was 1.98, the information entropy for the magnetic field was 1.85, and the information entropy for the stress field was 2.05. A coupling weight matrix was constructed based on the information entropy: the weights for the temperature field and electric field were 0.15, the weights for the temperature field and magnetic field were 0.18, the weights for the temperature field and stress field were 0.22, the weights for the electric field and magnetic field were 0.17, the weights for the electric field and stress field were 0.13, and the weights for the magnetic field and stress field were 0.15.

[0047] Second-order statistics of the dynamic evolution characteristics were calculated within a sliding time window. The window size was set to 30 sampling points, and the sliding step was set to 10 sampling points. In the first window, the mean of the dynamic evolution characteristics of the temperature field was 7.52, and the variance was 0.086; the mean of the dynamic evolution characteristics of the electric field was 11.43, and the variance was 0.157; the mean of the dynamic evolution characteristics of the magnetic field was 13.65, and the variance was 0.109; and the mean of the dynamic evolution characteristics of the stress field was 9.28, and the variance was 0.132.

[0048] The local coupling coefficients between the physical fields are calculated using the coupling weight matrix. For the temperature and electric fields, the second-order statistic has a correlation coefficient of 0.78, which, when multiplied by a weight of 0.15, yields a local coupling coefficient of 0.117. For the temperature and magnetic fields, the second-order statistic has a correlation coefficient of 0.65, which, when multiplied by a weight of 0.18, yields a local coupling coefficient of 0.117. For the temperature and stress fields, the second-order statistic has a correlation coefficient of 0.82, which, when multiplied by a weight of 0.22, yields a local coupling coefficient of 0.180. The local coupling coefficients between other pairs of physical fields are calculated similarly.

[0049] The second-order statistics and the local coupling coefficient are weighted and superimposed to obtain the coupling correlation between each physical field. The weight of the second-order statistics is set to 0.6, and the weight of the local coupling coefficient is set to 0.4. For the temperature field and electric field, the coupling correlation is 0.6×0.78+0.4×0.117=0.515; for the temperature field and magnetic field, the coupling correlation is 0.6×0.65+0.4×0.117=0.437; and for the temperature field and stress field, the coupling correlation is 0.6×0.82+0.4×0.180=0.564. In this way, the coupling correlation between all pairs of physical fields is calculated, and a complete coupling correlation matrix is ​​constructed to characterize the degree of coupling between the multi-physical fields as a whole.

[0050] Figure 3This is a schematic diagram showing the performance comparison of the multi-field coupling correlation evaluation model according to an embodiment of the present invention: This figure compares the coupling identification performance between different physical field combinations, including the recognition accuracy of the present technical solution, traditional cross-correlation analysis, and static coupling models. In temperature-electric field coupling identification, the present technical solution achieved an accuracy of 92.5%, significantly outperforming the 78.4% of the traditional cross-correlation analysis method and the 70.2% of the static coupling model. In electric-magnetic field coupling identification, the present solution performed most outstandingly, achieving a top accuracy of 94.8%, significantly superior to the 82.5% of the traditional method and the 73.8% of the static model. In magnetic-stress field coupling identification, the present solution maintained a high accuracy of 90.5%, compared to 76.3% for traditional methods and 68.9% for static models. In temperature-stress field coupling identification, the present solution maintained an excellent performance of 91.2%, compared to 80.6% for traditional methods and 71.5% for static models. Overall, this technical solution maintains a high accuracy rate of over 90% in all coupled field identifications, an average improvement of approximately 13 percentage points over traditional methods and approximately 20 percentage points over static models. This fully demonstrates the solution's advancement and effectiveness in multi-physics field coupled identification. The solution's advantages are particularly evident in identifying complex coupled relationships such as electric and magnetic fields.

[0051] In an optional embodiment, calculating the information entropy of each physical field based on the normalized data, constructing a coupling weight matrix according to the information entropy; calculating the second-order statistics of the dynamic evolution characteristics within a sliding time window, and calculating the local coupling coefficient between the physical fields in combination with the coupling weight matrix includes: Calculating probability distributions for the normalized multiple physical field data sequences respectively, calculating the information entropy of each physical field data sequence based on the probability distribution, and obtaining an information entropy set of the multiple physical field data sequences; The information entropy set is constructed as a coupling weight matrix, wherein the elements of the coupling weight matrix are determined by the ratio of the product of the information entropies of the corresponding two physical fields to the sum of the products of the information entropies of all physical fields; Calculating, within a sliding time window of a preset length, second-order statistics of the dynamic evolution characteristics of the plurality of physical field data sequences, the second-order statistics including an autocorrelation function of the physical field data sequence and a cross-correlation function between different physical field data sequences; The local coupling coefficient between the physical fields is calculated based on the coupling weight matrix, the autocorrelation function and the cross-correlation function.

[0052] The collected data of multiple physical fields are preprocessed and normalized, and then the information entropy of each physical field is calculated and a coupling weight matrix is ​​constructed. Finally, the second-order statistics of the dynamic evolution characteristics are calculated within the sliding time window, and the local coupling coefficients between the physical fields are calculated in combination with the coupling weight matrix.

[0053] For multiple collected physical field data sequences, such as temperature, pressure, and velocity fields, preprocessing is required to eliminate noise and outliers. This preprocessing step includes using median filtering to remove outliers, applying wavelet transforms for noise reduction, and using interpolation methods to handle missing data. Specifically, if the original temperature field data contains an outlier of 95°C (the surrounding data averages 25°C), this outlier is corrected to the median of the adjacent data, approximately 26°C, through a 5-point median filter.

[0054] After preprocessing, the data needs to be normalized to make different physical quantities comparable. This normalization method uses min-max normalization, mapping each physical field data point to the interval [0, 1]. For example, for temperature field data ranging from 20°C to 35°C, min-max normalization maps 20°C to 0, 35°C to 1, and 25°C to approximately 0.33.

[0055] For the normalized physical field data, calculate the information entropy of each physical field. First, estimate the probability distribution of each physical field data, which can be achieved by constructing a histogram. Divide the data range [0, 1] into multiple equal-width intervals, such as 10 intervals, each with a width of 0.1. Count the number of data points in each interval and divide it by the total number of data points to obtain the probability of that interval. For example, for the normalized pressure field data, if there are 150 data points in the interval [0, 0.1] and a total of 1000 data points, then the probability of this interval is 0.15.

[0056] Information entropy is calculated based on the probability of each interval. For each physical field, the information entropy of the field is obtained by summing the product of the probability of each interval and its natural logarithm and taking the negative value. For example, if the interval probabilities of the temperature field are [0.05, 0.12, 0.15, 0.2, 0.18, 0.13, 0.08, 0.05, 0.03, 0.01], the calculated information entropy is approximately 2.16.

[0057] After obtaining the information entropy of all physical fields, a coupling weight matrix is ​​constructed. Each element of this matrix represents the weight relationship between two physical fields, calculated by dividing the product of the information entropy of the two physical fields by the sum of the products of the information entropy of all physical fields. For example, if there are three physical fields A, B, and C, whose information entropies are 2.16, 1.85, and 2.35 respectively, then the coupling weight between A and B is: (2.16×1.85) / (2.16×1.85+2.16×2.35+1.85×2.35)=0.3.

[0058] Compute second-order statistics of the physical field data series within a sliding time window of a preset length. Set the time window length to 300 data points and the sliding step size to 50 data points. For the data within each window, calculate the autocorrelation function of each physical field and the cross-correlation function between different physical fields.

[0059] The autocorrelation function describes the correlation between data from the same physical field at different time points. It is obtained by calculating the correlation coefficient between the data sequence and itself at different time delays. For example, for temperature field data within a sliding window, the autocorrelation coefficient is calculated for delays of 0 to 50 time steps. The resulting autocorrelation function curve can reflect the periodic characteristics and memory effect of the temperature field.

[0060] The cross-correlation function measures the correlation between different physical fields. It is calculated by calculating the correlation coefficient between a data series from one physical field and another at different time delays. For example, the cross-correlation function between the temperature field and the pressure field can reflect the degree of influence of temperature changes on pressure changes and the lag time. If the cross-correlation coefficient reaches a maximum of 0.75 at a delay of 15 time steps, it indicates that the temperature change has a significant impact on the pressure after approximately 15 time steps.

[0061] The local coupling coefficient between the physical fields is calculated based on the coupling weight matrix, the autocorrelation function, and the cross-correlation function. The local coupling coefficient reflects the coupling strength between the two physical fields within a specific time window. It is calculated by multiplying the coupling weight by the maximum value of the normalized cross-correlation function. If the coupling weight between the temperature field and the pressure field is 0.3 and the maximum cross-correlation function is 0.75, the local coupling coefficient is: 0.3×0.75=0.225.

[0062] This method can be applied to industrial production environment monitoring, identifying key influencing factors by analyzing the coupling relationships between different physical fields. For example, in chemical production processes, multiple physical fields such as temperature, pressure, and flow rate have complex coupling relationships. Using this method, we calculated that the local coupling coefficient between temperature and reaction rate is 0.68, and the local coupling coefficient between pressure and reaction rate is 0.42, indicating that temperature has a more significant impact on the reaction process, and that temperature parameters should be controlled.

[0063] In an optional embodiment, based on the temporal variation trend of the coupling correlation and the multi-dimensional warning threshold, calculating the deviation between the coupling correlation and the multi-dimensional warning threshold, determining a warning level, and generating a targeted device failure risk assessment report and preventive maintenance recommendations based on the warning level include: Calculating a standardized deviation between the coupling correlation degree and the multidimensional warning threshold, constructing a dynamic weight coefficient based on the standardized deviation, and accumulating the product of the dynamic weight coefficient and the standardized deviation to obtain a comprehensive deviation index; Inputting the comprehensive deviation index into a preset membership function, calculating the membership value of the comprehensive deviation index for different risk levels, and performing weighted summation of the membership value and the corresponding risk level value to obtain a fuzzy risk degree; Determining a warning level based on the current value and rate of change of the fuzzy risk, wherein the warning level is dynamically adjusted by an adaptive mapping function that is updated in real time according to the changing trend of the fuzzy risk; The coupling correlation degree, the standardized deviation, the fuzzy risk degree, and historical fault characteristics are subjected to multimodal feature fusion to construct a fused feature vector; a device failure risk assessment report is obtained based on the fused feature vector and the warning level, and preventive maintenance recommendations are calculated based on the device failure risk assessment report and maintenance costs.

[0064] like Figure 4 As shown, the method includes: The system calculates the deviation between the coupling correlation and the multidimensional warning thresholds through standardization to make the indicators of different dimensions comparable. Specifically, if the coupling correlation of a temperature monitoring point is 0.78 and the corresponding warning threshold is 0.65, the system divides the difference between the two by the threshold standard deviation (for example, 0.1) to obtain a standardized deviation of 1.3. Similarly, the standardized deviation of vibration data might be 0.9, and the standardized deviation of current data might be -0.5. These standardized deviations reflect the degree to which each parameter deviates from the safety threshold.

[0065] The dynamic weight coefficients are constructed based on the importance of the parameters and their current deviation status. For example, when the temperature deviation is large, its weight should be increased accordingly. The system uses a soft maximization function to process the standardized deviations, giving the dynamic weights of temperature, vibration, and current 0.45, 0.35, and 0.2, respectively. These weights are multiplied by the corresponding standardized deviations and then accumulated to obtain the comprehensive deviation index: (1.3×0.45)+(0.9×0.35)+(-0.5×0.2)=0.585+0.315-0.1=0.8.

[0066] In the membership function calculation, the system presets membership functions for three risk levels: low risk (0-0.4), medium risk (0.3-0.7), and high risk (0.6-1.0). For a comprehensive deviation index of 0.8, the membership function calculation results in a low risk membership of 0.0, a medium risk membership of 0.5, and a high risk membership of 0.8. Each membership is weighted and summed with the risk level value (low = 1, medium = 2, high = 3): (0.0×1)+(0.5×2)+(0.8×3)=0+1+2.4=3.4, and the fuzzy risk is 3.4.

[0067] The determination of the warning level takes into account not only the current value of the fuzzy risk but also its rate of change. For example, if the current fuzzy risk is 3.4 and the previous value was 3.0, the rate of change is (3.4 - 3.0) / 3.0 = 13.3%. The system uses an adaptive mapping function to convert the fuzzy risk and its rate of change into a warning level. This function dynamically adjusts based on historical data. For example, if the fuzzy risk is high and the rate of change exceeds 10%, the warning level may be raised by one level. In this example, a fuzzy risk of 3.4 corresponds to a "medium warning," and considering the 13.3% rate of change, the final level is determined to be a "medium-high warning."

[0068] To generate a comprehensive risk assessment, the system performs multimodal feature fusion on various indicators. This fusion vector is formed by combining the coupling correlation vector (0.78, 0.72, 0.58), the standardized deviation vector (1.3, 0.9, -0.5), and the fuzzy risk score of 3.4 with historical failure characteristics (e.g., equipment operating time of 8760 hours and a historical failure frequency of 0.002 per hour). This feature vector is then used for feature extraction and correlation analysis to identify key influencing factors.

[0069] Based on the fused feature vectors and the "medium to high warning" level, the system generates a component failure risk assessment report, including a comprehensive risk index, possible failure modes, and their probability of occurrence. For example, the assessment report might indicate a 15% probability of failure within the next 30 days, with motor bearing wear being the most likely cause, accounting for 65% of the total failure risk.

[0070] Preventive maintenance recommendations were made based on the assessment report and maintenance cost calculations. The system analyzed the cost of bearing replacement (including 1,000 yuan for parts and 500 yuan for labor) compared to the potential losses from the failure (20,000 yuan per day of downtime, with an estimated repair time of two days). The system recommended that the bearing be replaced during the next scheduled downtime, saving an estimated 35,000 yuan in potential losses. Additionally, it recommended increasing the frequency of vibration monitoring from weekly to daily to detect abnormal changes earlier.

[0071] Through the above technical implementation, the present invention can accurately assess the risk of device failure and provide cost-effective preventive maintenance recommendations, significantly improving the reliability and economic benefits of industrial equipment.

[0072] In an optional embodiment, obtaining a device failure risk assessment report based on the fused feature vector and the warning level, and calculating a preventive maintenance recommendation based on the device failure risk assessment report and the maintenance cost includes: Inputting the fused feature vector into a preset multi-physics field coupling state assessment model to obtain a device state assessment result; Constructing an adaptive warning threshold based on the device status assessment result, wherein the adaptive warning threshold is determined by the sum of the product of a baseline threshold and a dynamic adjustment coefficient and a correction term, the dynamic adjustment coefficient is dynamically updated according to the device status assessment result, and the correction term is calculated based on historical warning data; Inputting the device status assessment result, the adaptive warning threshold and the digital twin status data into the multi-physics field coupling status assessment model to obtain a device failure risk assessment report; Constructing a cost-effectiveness strategy based on the device failure risk assessment report, the maintenance cost data, and the failure risk cost data, wherein an objective function of the cost-effectiveness strategy is a minimum value of the sum of the maintenance cost and the failure risk cost multiplied by the failure probability; A maintenance strategy optimization result is generated based on the cost-effectiveness strategy, wherein the maintenance strategy optimization result satisfies the constraint condition that the equipment risk assessment result is less than a preset risk threshold, and a preventive maintenance recommendation is obtained.

[0073] The fused feature vectors obtained through data processing are input into a pre-defined multi-physics field coupling state assessment model. This model is a deep learning model built based on the coupling relationships between physical fields such as mechanical stress, thermal stress, and electromagnetic field distribution. In its implementation, the model uses a multi-layer neural network structure consisting of an input layer, multiple hidden layers, and an output layer.

[0074] The input layer receives a fused feature vector, which contains the numerical values ​​of multi-source sensor data such as device temperature, vibration, current, and voltage after noise reduction, normalization, and feature extraction. The hidden layer transforms and combines these features using a nonlinear activation function. The output layer generates a device status assessment result, including the health index of each key component, the estimated remaining service life, and the failure probability distribution. For example, after the fused feature vector of a high-voltage switchgear device is input into the model, the output status assessment result shows a contact wear health index of 0.78, an estimated remaining service life of 189 days, and a probability of 0.13 for poor contact, the primary failure mode.

[0075] Based on the obtained equipment status assessment results, an adaptive warning threshold is constructed. This adaptive warning threshold is determined by multiplying the baseline threshold by the dynamic adjustment coefficient and adding a correction factor. The baseline threshold is selected from a preset threshold library based on the equipment type and operating environment. For example, for high-voltage switchgear, the temperature baseline threshold is 75°C.

[0076] The dynamic adjustment coefficient is updated based on the equipment status assessment results. When the equipment status assessment results indicate good health, the adjustment coefficient increases, raising the warning threshold accordingly. Conversely, the warning threshold is lowered, prompting an earlier warning. A correction factor is calculated based on historical warning data, primarily considering historical warning accuracy and false alarm rates. The correction value is calculated using a weighted average method. In a practical application, the temperature baseline threshold of a transformer equipment is 80°C. Based on its health index of 0.92, the dynamic adjustment coefficient is calculated to be 1.05. The correction factor derived from analysis of historical warning data is -2°C, resulting in a final adaptive temperature warning threshold of 82°C.

[0077] The device status assessment results, adaptive warning thresholds, and digital twin status data are input into the multi-physics coupled state assessment model. The digital twin status data is a simulation of the actual device status using the physical model, including operating parameters and status change trends under different operating conditions. The model compares the actual device status with the digital twin simulation and, in combination with the adaptive warning thresholds, generates a device failure risk assessment report.

[0078] The report includes risk level, failure probability, impact scope, failure mode analysis, and risk evolution trends. For example, a risk assessment report for a power transformer shows that its overall risk level is medium (Level 3), the probability of failure of the main winding insulation within the next 30 days is 0.08, the impact scope includes three downstream distribution stations, the failure mode is insulation breakdown, and the risk is showing a slowly increasing trend.

[0079] Build a cost-effectiveness strategy based on component failure risk assessment reports, maintenance cost data, and failure risk cost data. Maintenance cost data includes labor costs, parts costs, and equipment downtime losses. For example, for a certain model of circuit breaker, the regular maintenance cost is 5,000 yuan, the contact replacement cost is 8,000 yuan, and the equipment downtime loss is 2,000 yuan per hour.

[0080] Failure risk cost data includes equipment losses, production losses, and safety accident losses. For example, a complete circuit breaker failure could result in a total loss of 500,000 yuan. The objective function of the cost-effectiveness strategy is to minimize the sum of the maintenance cost and the failure risk cost multiplied by the failure probability. Numerical optimization methods such as genetic algorithms or particle swarm optimization are used to find the minimum point of this objective function and determine the optimal maintenance strategy.

[0081] Generate maintenance strategy optimization results based on a cost-effectiveness strategy. During the optimization process, constraints are set such that the equipment risk assessment result must be less than a preset risk threshold, such as a risk level of no more than Level 2 or a failure probability of no more than 0.05. An iterative optimization algorithm, while satisfying these constraints, finds a maintenance plan that minimizes the weighted sum of maintenance cost and risk cost, generating a preventive maintenance recommendation. This recommendation includes a maintenance schedule, a maintenance item list, required resource allocation, and risk control measures.

[0082] For a certain power distribution equipment, the optimization results recommend preventive maintenance within 15 days, including replacing aging relays, cleaning insulation components, and adjusting protection device parameters. This requires two maintenance personnel and three sets of professional tools. The maintenance time is expected to be four hours, and a backup system must be activated during maintenance. After maintenance, the expected risk level is reduced to Level 1, and the equipment service life is expected to be extended by eight months.

[0083] A second aspect of an embodiment of the present invention provides a dynamic monitoring system for multi-physical field coupling effects of a third-generation semiconductor device, comprising: The first unit is used to collect temperature field data, electric field data, magnetic field data and stress field data in the third-generation semiconductor device, and synchronously obtain the time series change characteristics of each physical field data in the temperature field data, the electric field data, the magnetic field data and the stress field data through a complementary sensor array; The second unit is used to construct a feature tensor of the multi-physics field coupling effect according to the time series change characteristics, input the feature tensor into a pre-trained graph neural network model, and extract the dynamic evolution law of the multi-physics field coupling effect based on the graph neural network model; The third unit is used to perform online analysis of the dynamic evolution law using an adaptive time window sliding mechanism, dynamically predict the multi-physics field coupling effect at different time scales through a recursive neural network, and generate a multi-dimensional warning threshold based on the prediction results of the multi-physics field coupling effect; A fourth unit is configured to construct a multi-field coupling correlation evaluation model based on the dynamic evolution law, input the temperature field data, the electric field data, the magnetic field data, and the stress field data into the multi-field coupling correlation evaluation model, and obtain the coupling correlation between the physical fields; The fifth unit is used to calculate the deviation between the coupling correlation and the multidimensional warning threshold according to the time series change trend of the coupling correlation and the multidimensional warning threshold, determine the warning level, and generate a targeted device failure risk assessment report and preventive maintenance recommendations based on the warning level.

[0084] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0085] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0086] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic monitoring method for multi-physical field coupling effects of third-generation semiconductor devices, characterized in that: include: Collecting temperature field data, electric field data, magnetic field data, and stress field data in third-generation semiconductor devices, and synchronously acquiring the time-series variation characteristics of each physical field data in the temperature field data, the electric field data, the magnetic field data, and the stress field data through a complementary sensor array; Constructing a feature tensor of the multi-physics field coupling effect according to the temporal variation characteristics, inputting the feature tensor into a pre-trained graph neural network model, and extracting the dynamic evolution law of the multi-physics field coupling effect based on the graph neural network model; An adaptive time window sliding mechanism is used to perform online analysis of the dynamic evolution law, a recursive neural network is used to dynamically predict the multi-physics field coupling effect at different time scales, and a multi-dimensional warning threshold is generated based on the prediction results of the multi-physics field coupling effect; Based on the dynamic evolution law, a multi-field coupling correlation evaluation model is constructed, and the temperature field data, the electric field data, the magnetic field data, and the stress field data are input into the multi-field coupling correlation evaluation model to obtain the coupling correlation between the physical fields; According to the temporal variation trend of the coupling correlation and the multidimensional warning threshold, the deviation between the coupling correlation and the multidimensional warning threshold is calculated, the warning level is determined, and a targeted device failure risk assessment report and preventive maintenance recommendations are generated based on the warning level.

2. The method according to claim 1, characterized in that The dynamic evolution law is analyzed online using an adaptive time window sliding mechanism, and the multi-physics field coupling effect at different time scales is dynamically predicted using a recursive neural network. The multi-dimensional warning threshold is generated based on the prediction results of the multi-physics field coupling effect, including: Building an adaptive time window based on the continuous sampling data of the dynamic evolution law, wherein the length of the adaptive time window is dynamically adjusted according to the change rate of the dynamic evolution law; Performing sliding sampling on the dynamic evolution law using the adaptive time window, and obtaining characteristic data of multi-physical field coupling effects at different time points through the sliding sampling, wherein the characteristic data includes physical field intensity values ​​and coupling correlations; Grouping the feature data according to different time scales to form multiple time scale sequences, and inputting the multiple time scale sequences into a pre-trained recursive neural network; Extracting the time series features of the multiple time scale sequences through the hidden layer state of the recurrent neural network, predicting the evolution trend of the multi-physics field coupling effect based on the time series features, and obtaining prediction results at different time scales; The warning reference value of each physical field is calculated according to the prediction result, and a multi-dimensional warning threshold is generated based on the warning reference value combined with the historical fluctuation range of each physical field. The multi-dimensional warning threshold is updated in real time as the prediction result changes.

3. The method according to claim 1, characterized in that Based on the dynamic evolution law, a multi-field coupling correlation evaluation model is constructed, and the temperature field data, the electric field data, the magnetic field data, and the stress field data are input into the multi-field coupling correlation evaluation model to obtain the coupling correlation between the physical fields, including: Calculating sampling time deviations among the temperature field data, the electric field data, the magnetic field data, and the stress field data, and performing data synchronization processing according to the sampling time deviations; performing dimension conversion on the synchronized temperature field data, the electric field data, the magnetic field data, and the stress field data to obtain normalized data; Calculating the change rate of the normalized data at adjacent sampling moments, performing a weighted combination of the change rate and the time integral of the normalized data to construct a dynamic evolution feature; The information entropy of each physical field is calculated based on the normalized data, and a coupling weight matrix is ​​constructed according to the information entropy; the second-order statistics of the dynamic evolution characteristics are calculated within a sliding time window, and the local coupling coefficient between the physical fields is calculated in combination with the coupling weight matrix, and the local coupling coefficient represents the interaction strength between the physical fields; the second-order statistics and the local coupling coefficient are weightedly superimposed to obtain the coupling correlation between the physical fields, and the coupling correlation represents the degree of coupling of the multi-physical fields as a whole.

4. The method according to claim 3, characterized in that Calculating the information entropy of each physical field based on the normalized data, constructing a coupling weight matrix according to the information entropy; calculating the second-order statistics of the dynamic evolution characteristics within a sliding time window, and calculating the local coupling coefficient between the physical fields in combination with the coupling weight matrix includes: Calculating probability distributions for the normalized multiple physical field data sequences respectively, calculating the information entropy of each physical field data sequence based on the probability distribution, and obtaining an information entropy set of the multiple physical field data sequences; The information entropy set is constructed as a coupling weight matrix, wherein the elements of the coupling weight matrix are determined by the ratio of the product of the information entropies of the corresponding two physical fields to the sum of the products of the information entropies of all physical fields; Calculating, within a sliding time window of a preset length, second-order statistics of the dynamic evolution characteristics of the plurality of physical field data sequences, the second-order statistics including an autocorrelation function of the physical field data sequence and a cross-correlation function between different physical field data sequences; The local coupling coefficient between physical fields is calculated based on the coupling weight matrix, the autocorrelation function and the cross-correlation function.

5. The method according to claim 1, wherein Calculating the deviation between the coupling correlation and the multi-dimensional warning threshold according to the temporal variation trend of the coupling correlation and the multi-dimensional warning threshold, determining a warning level, and generating a targeted component failure risk assessment report and preventive maintenance recommendations based on the warning level include: Calculating a standardized deviation between the coupling correlation degree and the multidimensional warning threshold, constructing a dynamic weight coefficient based on the standardized deviation, and accumulating the product of the dynamic weight coefficient and the standardized deviation to obtain a comprehensive deviation index; Inputting the comprehensive deviation index into a preset membership function, calculating the membership value of the comprehensive deviation index for different risk levels, and performing weighted summation of the membership value and the corresponding risk level value to obtain a fuzzy risk degree; Determining a warning level based on the current value and rate of change of the fuzzy risk, wherein the warning level is dynamically adjusted by an adaptive mapping function that is updated in real time according to the changing trend of the fuzzy risk; The coupling correlation degree, the standardized deviation, the fuzzy risk degree, and historical fault characteristics are subjected to multimodal feature fusion to construct a fused feature vector; a device failure risk assessment report is obtained based on the fused feature vector and the warning level, and preventive maintenance recommendations are calculated based on the device failure risk assessment report and maintenance costs.

6. The method according to claim 5, characterized in that Obtaining a device failure risk assessment report based on the fused feature vector and the warning level, and calculating a preventive maintenance recommendation based on the device failure risk assessment report and the maintenance cost includes: Inputting the fused feature vector into a preset multi-physics field coupling state assessment model to obtain a device state assessment result; Constructing an adaptive warning threshold based on the device status assessment result, wherein the adaptive warning threshold is determined by the sum of a product of a baseline threshold and a dynamic adjustment coefficient and a correction term, the dynamic adjustment coefficient is dynamically updated according to the device status assessment result, and the correction term is calculated based on historical warning data; Inputting the device status assessment result, the adaptive warning threshold and the digital twin status data into the multi-physics field coupling status assessment model to obtain a device failure risk assessment report; Constructing a cost-effectiveness strategy based on the device failure risk assessment report, the maintenance cost data, and the failure risk cost data, wherein an objective function of the cost-effectiveness strategy is a minimum value of the sum of the maintenance cost and the failure risk cost multiplied by the failure probability; A maintenance strategy optimization result is generated based on the cost-effectiveness strategy, wherein the maintenance strategy optimization result satisfies the constraint condition that the equipment risk assessment result is less than a preset risk threshold, and a preventive maintenance recommendation is obtained.

7. A third-generation semiconductor device multi-physics field coupling effect dynamic monitoring system, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to collect temperature field data, electric field data, magnetic field data and stress field data in the third-generation semiconductor device, and synchronously obtain the time series change characteristics of each physical field data in the temperature field data, the electric field data, the magnetic field data and the stress field data through a complementary sensor array; The second unit is used to construct a feature tensor of the multi-physics field coupling effect according to the time series change characteristics, input the feature tensor into a pre-trained graph neural network model, and extract the dynamic evolution law of the multi-physics field coupling effect based on the graph neural network model; The third unit is used to perform online analysis of the dynamic evolution law using an adaptive time window sliding mechanism, dynamically predict the multi-physics field coupling effect at different time scales through a recursive neural network, and generate a multi-dimensional warning threshold based on the prediction results of the multi-physics field coupling effect; A fourth unit is configured to construct a multi-field coupling correlation evaluation model based on the dynamic evolution law, input the temperature field data, the electric field data, the magnetic field data, and the stress field data into the multi-field coupling correlation evaluation model, and obtain the coupling correlation between the physical fields; The fifth unit is used to calculate the deviation between the coupling correlation and the multidimensional warning threshold according to the time series change trend of the coupling correlation and the multidimensional warning threshold, determine the warning level, and generate a targeted device failure risk assessment report and preventive maintenance recommendations based on the warning level.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images