Power device packaging stress monitoring and early warning method and system

Abstract

The invention discloses a power device packaging stress monitoring and early warning method and system, and relates to the technical field of packaging stress monitoring, and the method comprises the steps: constructing an integrated stress monitoring network of a power device, and obtaining multi-source stress data; the historical packaging health state of the power device is introduced, parallel monitoring is carried out on the multi-source stress data, and bidirectional stress state information is obtained; extracting a plurality of characteristic parameters to carry out evolution prediction on the power device, and constructing a packaging stress evolution prediction trend; carrying out risk assessment on the packaging stress of the power device, and constructing a multi-stage early warning signal; and carrying out identification verification on the multi-source stress data, and updating the multi-stage early warning signal according to a verification result. According to the invention, the technical problems of lack of full-process coverage, insufficient real-time performance and poor early warning accuracy of monitoring of the packaging stress of the power device in the prior art are solved, and the technical effects of realizing full-process accurate management and control of the packaging stress of the power device and improving the real-time performance of packaging stress monitoring and the accuracy of early warning are achieved.

Description

Technical Field

[0001] This invention relates to the field of packaging stress monitoring technology, and more specifically to a method and system for monitoring and early warning of packaging stress in power devices. Background Technology

[0002] In the field of power electronics, power devices, as core components for power conversion and control, are widely used in critical scenarios such as new energy vehicles, rail transit, and smart grids. Their packaging structures must withstand complex conditions such as high temperatures, vibration, and electromagnetic forces over long periods, easily leading to accumulated packaging stress. If the stress exceeds the material's tolerance limit, it can cause bond wire detachment, package cracking, increased thermal resistance, and even device failure, directly affecting the reliability and lifespan of the entire power system. Currently, the industry primarily uses traditional single-point sensors, such as strain gauges and thermocouples combined with simple threshold judgments, for monitoring and early warning of packaging stress in power devices. This approach has significant technical limitations: Firstly, single-point monitoring cannot cover the entire stress distribution of the package, easily missing key stress concentration areas, such as the chip-substrate connection edge. Furthermore, the data acquisition dimension is singular, failing to synchronously correlate with key factors affecting stress evolution, such as temperature and load, resulting in a lack of continuous and comprehensive monitoring throughout the entire process. Secondly, early warning mechanisms often rely on fixed stress thresholds, failing to combine them with the device's historical health status, such as past stress cycle counts and fatigue damage levels, for dynamic analysis. This makes it difficult to accurately predict stress evolution trends, often resulting in false or missed warnings, and failing to provide effective support for operation and maintenance decisions.

[0003] Existing technologies lack full-process coverage for monitoring the packaging stress of power devices, resulting in insufficient real-time performance and poor early warning accuracy. Summary of the Invention

[0004] This application provides a method and system for monitoring and warning of power device packaging stress, which addresses the technical problems in the prior art where the monitoring of power device packaging stress lacks full-process coverage, has insufficient real-time performance, and poor warning accuracy.

[0005] In view of the above problems, this application provides a method and system for monitoring and early warning of stress in power device packaging.

[0006] A first aspect of this application provides a method for monitoring and early warning of package stress in power devices, the method comprising: An integrated stress monitoring network for power devices is constructed to collect operational data and obtain multi-source stress data. Historical package health status of the power devices is introduced, and the multi-source stress data is monitored in parallel based on this historical package health status to obtain bidirectional stress state information. Feature analysis is performed based on the bidirectional stress state information to extract multiple feature parameters for power device evolution prediction, constructing a package stress evolution prediction trend. Risk assessment of the power device package stress is conducted according to the package stress evolution prediction trend, and multi-level early warning signals are constructed. The multi-level early warning signals are sent to the device terminal to identify and verify the multi-source stress data, and the multi-level early warning signals are updated based on the verification results.

[0007] A second aspect of this application provides a power device package stress monitoring and early warning system, the system comprising: The system includes a multi-source stress data acquisition module, which constructs an integrated stress monitoring network for power devices to collect operational data and obtain multi-source stress data; a bidirectional stress state information acquisition module, which incorporates the historical package health status of power devices and performs parallel monitoring of multi-source stress data based on the historical package health status to obtain bidirectional stress state information; a prediction trend construction module, which performs feature analysis based on the bidirectional stress state information, extracts multiple feature parameters to predict the evolution of power devices, and constructs a package stress evolution prediction trend; a multi-level early warning signal construction module, which performs risk assessment of the package stress of power devices according to the package stress evolution prediction trend and constructs multi-level early warning signals; and a multi-level early warning signal update module, which sends the multi-level early warning signals to the device terminal to identify and verify the multi-source stress data and updates the multi-level early warning signals based on the verification results.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: An integrated stress monitoring network for power devices is constructed to collect operational data and obtain multi-source stress data. Historical package health status of the power devices is incorporated, and the multi-source stress data is monitored in parallel to obtain bidirectional stress state information. Multiple feature parameters are extracted to predict the evolution of the power devices and construct a package stress evolution prediction trend. Risk assessment of the package stress of the power devices is performed, and multi-level early warning signals are constructed. The multi-source stress data is identified and verified, and the multi-level early warning signals are updated based on the verification results. This achieves precise control over the entire process of power device package stress, improving the real-time performance of package stress monitoring and the accuracy of early warning. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic flowchart of a power device packaging stress monitoring and early warning method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the power device packaging stress monitoring and early warning system provided in an embodiment of this application.

[0011] Figure labeling: Multi-source stress data acquisition module 10, bidirectional stress state information acquisition module 20, prediction trend construction module 30, multi-level early warning signal construction module 40, multi-level early warning signal update module 60. Detailed Implementation

[0012] This application provides a method and system for monitoring and warning of power device packaging stress, which addresses the technical problems in the prior art where the monitoring of power device packaging stress lacks full-process coverage, has insufficient real-time performance, and poor warning accuracy.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] Example 1, as Figure 1 As shown, this application provides a method for monitoring and early warning of stress in power device packaging, the method comprising: Step S100: Construct an integrated stress monitoring network for power devices to collect operational data and obtain multi-source stress data.

[0015] Specifically, by combining embedding and surface mounting, various miniature sensors adapted to different stress monitoring needs are deployed at key locations inside and on the surface of the power device package to form a distributed sensing array. Among them, the Hall-canceling packaged stress sensor is deployed: This sensor adopts a core design of Hall element, compensation coil, and differential signal processing circuit. The compensation coil is symmetrically wound around the Hall element. When working, a compensation current opposite to the direction of the external electromagnetic field is passed through it. By adjusting the current magnitude, a canceling magnetic field is generated, which can neutralize more than 90% of the external electromagnetic interference. The differential signal processing circuit is constructed using an instrumentation amplifier to differentially amplify the weak stress signal output by the Hall element, while suppressing common-mode noise and improving the signal-to-noise ratio to over 60dB. The sensor is integrated using MEMS (Micro-Electro-Mechanical Systems) technology: When embedded on the chip surface, a low-temperature bonding process is used to directly bond the sensor chip to a reserved area at the edge of the active region of the power chip, with the bonding layer thickness controlled at 5~10μm to ensure stress transmission efficiency; when embedded inside the chip, a micro-cavity is reserved during the power chip manufacturing stage using through-silicon via (TSV) technology, and the sensor die is fixed to the cavity using flip-chip bonding, with the gaps filled with low-modulus polyimide material to avoid interference from the sensor's own packaging on the chip's stress field, ultimately achieving accurate capture of the first mechanical stress generated by the mechanical action of the packaging structure. Piezoresistive stress sensor deployment: Near areas prone to stress due to differences in the thermal expansion coefficients of materials, such as lead frames or direct copper-clad substrates (DBC), a piezoresistive stress sensor is fixed using high-temperature epoxy resin adhesive or laser welding. The sensor's sensitive gate direction is aligned with the principal stress direction, specifically monitoring the second mechanical stress caused by thermal mismatch. Fiber Bragg grating (FBG) sensor deployment: Sensors are fixed in the heat dissipation path of the power device using either metallization or high-temperature ceramic encapsulation. In metallization, the FBG is encapsulated within a stainless steel capillary tube, sealed at both ends by laser welding. The outer wall of the capillary is coated with thermally conductive silicone grease and then embedded in a pre-reserved slot in the heat dissipation substrate. In high-temperature ceramic encapsulation, the FBG is embedded in a groove in an alumina ceramic sheet, sintered with high-temperature resistant glass powder, and then secured to the thermal interface with screws. Both encapsulation methods allow the sensor to withstand an operating temperature range of -50℃ to 300℃, ensuring long-term stable acquisition of thermal stress caused by temperature changes. For monitoring dynamic stress, a suitable stress wave sensor, such as a piezoelectric stress wave sensor with a resonant frequency of 500kHz to 2MHz, can be deployed as a supplement. This sensor is ultrasonically welded to the encapsulation shell surrounding the power chip to acquire dynamic stress data based on stress waves.Subsequently, the various sensors in the aforementioned distributed sensing array are integrated via I2C or SPI data bus to construct an integrated stress monitoring network covering the internal-surface-heat dissipation path. After the network is started, it synchronously collects the first mechanical stress data output by the Hall cancelling sensor, the second mechanical stress data output by the piezoresistive sensor, the thermal stress data output by the fiber optic grating sensor, and the dynamic stress data output by the stress wave sensor. These stress data from different sources and of different types are then structurally integrated to form multi-source stress data that includes thermal stress data, mechanical stress data, first mechanical stress, second mechanical stress, and dynamic stress data based on stress waves, providing comprehensive data support for subsequent monitoring and analysis.

[0016] Step S200: Introduce the historical package health status of the power device, and perform parallel monitoring of multi-source stress data based on the historical package health status to obtain bidirectional stress status information.

[0017] Specifically, a pre-constructed historical package health status of power devices is introduced. This status is obtained by acquiring full-lifecycle stress simulation data through accelerated aging simulation of the entire lifecycle of the power devices, combined with a failure case library formed by failure analysis, and retrieved actual maintenance logs. After being time-series organized and stored, a historical package status database is constructed and extracted, which can serve as a reference benchmark for real-time monitoring. Subsequently, a parallel processing architecture for local and global monitoring is constructed. Based on the above-mentioned historical package health status, multi-scale synchronous monitoring is carried out on the acquired multi-source stress data: on the one hand, local monitoring is carried out, focusing on key local areas of the power device package, such as the chip-to-substrate connection and wire bonding. Combining the stress failure patterns of such local areas in history, the distribution and variation characteristics of multi-source stress data in local areas are analyzed to obtain local multi-scale information; on the other hand, global monitoring is carried out, starting from the overall structure of the power device package, referring to the normal range of global stress distribution in the historical package health status, judging the transmission path and correlation influence of multi-source stress data on the overall package, and obtaining global multi-scale information. Subsequently, spatiotemporal registration is performed on local multi-scale information and global multi-scale information to ensure consistency between the two in the time dimension (i.e., data acquisition sequence) and the spatial dimension (i.e., encapsulation location coordinates). Then, a complete encapsulation stress state map is constructed through feature-level fusion. Key stress concentration areas in the map are identified and backtracked to the parallel processing architecture for cross-validation, ultimately forming bidirectional stress state information covering both local fine features and global overall features.

[0018] Step S300: Based on the bidirectional stress state information, perform feature analysis, extract multiple feature parameters to predict the evolution of power devices, and construct a prediction trend for the evolution of packaging stress.

[0019] Specifically, a multi-dimensional feature analysis system is constructed, comprising a time-domain feature analysis layer, a frequency-domain feature analysis layer, and a time-frequency-domain feature analysis layer. The acquired bidirectional stress state information, including local and global stress data, is simultaneously input into the three analysis layers: the time-domain feature analysis layer extracts the static and dynamic variation characteristics of stress over time by calculating parameters such as the mean, variance, and peak value of the stress data, obtaining time-domain feature parameters; the frequency-domain feature analysis layer transforms the stress data to the frequency domain using algorithms such as Fourier transform, analyzes the stress distribution patterns in different frequency bands, and generates frequency-domain feature parameters; the time-frequency-domain feature analysis layer captures the coupling characteristics of stress data in both time and frequency dimensions using methods such as wavelet transform, obtaining time-frequency-domain feature parameters. Next, linear dimensionality reduction is performed on the three types of feature parameters to simplify the data dimensions. Importance scores are then generated for each dimensionality reduction result through importance assessment. The top N key feature parameters are selected based on their scores and recursively eliminated to form an optimized combination of feature parameters. Subsequently, deep learning training was conducted based on this optimized combination to generate feature learning results that can accurately characterize the stress change law. These results were progressively transferred to the stress simulation data of the entire life cycle of power devices. Combined with the stress evolution law in the historical package health state, the stress change trend of the package in the current and future period was predicted. Finally, a complete package stress evolution prediction trend reflecting the power device package stress from the current state to the potential failure state was constructed.

[0020] Step S400: Perform a risk assessment of the power device packaging stress according to the predicted trend of packaging stress evolution, and construct a multi-level early warning signal.

[0021] Specifically, based on the constructed packaging stress evolution prediction trend, risk assessment is carried out in multiple dimensions: The first step is to conduct stress over-limit analysis, comparing the stress values ​​at each time point in the prediction trend, such as the chip-substrate interface stress and the wire bonding stress, with the safety thresholds determined in the historical packaging health status, such as the chip interface stress threshold of 150MPa and the wire bonding stress threshold of 80MPa. A statistical model is used to calculate the probability of stress exceeding the threshold within a preset future period, such as 24 hours, such as a 60% probability of exceeding the threshold in a certain period, the magnitude of the exceedance, such as exceeding the threshold by 20MPa and the duration, such as a single exceedance lasting 1.5 hours. A stress over-limit risk index of 0-100 points is generated using a weighted formula of probability × 0.4 + magnitude × 0.3 + duration × 0.3. The higher the index, the greater the risk of exceeding the limit. The first step is to extract the stress change rate from the predicted trend, such as a stress increase of 5 MPa per hour, and the fluctuation frequency, such as a stress fluctuation occurring once every 10 minutes. Combined with the correlation between the stress change rate > 8 MPa / hour and the fluctuation frequency > 1 time / 5 minutes in historical failure cases, the device failure probability increases by 3 times. The stress change risk is divided into three levels: "low" (change rate ≤ 2 MPa / hour, fluctuation frequency ≤ 1 time / 30 minutes), "medium" (2 MPa / hour < change rate ≤ 5 MPa / hour, 1 time / 30 minutes < fluctuation frequency ≤ 1 time / 10 minutes), and "high" (change rate > 5 MPa / hour, fluctuation frequency > 1 time / 10 minutes). Stress change risk data containing risk level and corresponding feature descriptions are generated. The third step involves cumulative fatigue damage assessment. Based on the predicted stress cycle count (e.g., 120 thermal stress cycles within 24 hours) and stress amplitude variation (e.g., stress amplitude fluctuating between 30 and 50 MPa), calculations are performed using Miner's linear cumulative damage rule. For power device packaging materials such as epoxy resin and aluminum-based leadframes, the relationship between stress amplitude Δσ and cycle count N is clearly derived from JEDEC standard JESD22-A104 or official datasheets provided by device manufacturers. For example, the SN curve formula for epoxy resin is Δσ = 500 × N^(-0.15). During calculation, the allowable cycle count N for each stress cycle is first determined based on the SN curve. i Then, according to the formula D=Σ(n i / N i ), where n iLet D be the actual number of cycles and D be the cumulative damage. The cumulative damage to the packaging structure is calculated. A value of D < 0.3 indicates low damage, 0.3 ≤ D < 0.7 indicates medium damage, and D ≥ 0.7 indicates high damage. Based on this, fatigue damage risk data is generated. Subsequently, the stress over-limit risk index, stress change risk data, and fatigue damage risk data are weighted and calculated using a multi-dimensional fusion method to obtain the composite risk index of power device packaging stress. The weighting principle is clearly defined as follows: stress over-limit risk weight 0.5, stress change risk weight 0.2, and fatigue damage risk weight 0.3. These weight values ​​are based on historical failure data statistics. Analysis of 500 past power device packaging failure cases revealed that failures directly caused by stress over-limit accounted for 52%, failures caused by stress changes accounted for 18%, and failures caused by fatigue damage accounted for 30%. This weighting scheme was determined after normalization. The calculation first converts the stress change risk data, low, medium, and high level fatigue damage risk data, and low, medium, and high damage into corresponding scores: low = 30 points, medium = 60 points, and high = 90 points. Then, the specific values ​​from 0 to 100 points are calculated using the formula: Composite Risk Index = Stress Exceedance Risk Index × 0.5 + Stress Change Risk Score × 0.2 + Fatigue Damage Risk Score × 0.3. Finally, risk levels are classified based on the numerical range of the composite risk index, and multi-level early warning signals are constructed: an index of 0-30 indicates low risk, corresponding to a blue warning, and it is recommended to conduct stress monitoring at a regular cycle, such as once a month; an index of 31-60 indicates medium risk, corresponding to a yellow warning, and it is recommended to shorten the monitoring cycle, such as once every 3 days, and investigate the causes of abnormal stress; an index of 61-85 indicates high risk, corresponding to an orange warning, and it is recommended to immediately stop the machine and check the packaging structure, such as checking whether the wire bonds are loose and developing a maintenance plan; an index of 86-100 indicates extremely high risk, corresponding to a red warning, and it is recommended to urgently replace the device to avoid sudden failure. At the same time, failure precursor data should be recorded for subsequent risk model optimization to ensure that the potential threat of packaging stress can be accurately transmitted according to the risk level.

[0022] Step S500: Send the multi-level early warning signal to the device terminal to identify and verify the multi-source stress data, and update the multi-level early warning signal according to the verification result.

[0023] Specifically, a two-way communication mechanism with the equipment terminals is established. Based on this mechanism, multi-level early warning signals, including warning intensities and response suggestions for different risk levels, are sent to the corresponding equipment terminals. Simultaneously, key segments of multi-source stress data are transmitted for terminal verification. After receiving the early warning signal, the equipment terminal obtains confirmation feedback through the two-way communication mechanism, generating a terminal confirmation signal. Based on this, the issued multi-level early warning signals are initially labeled, forming multi-level early warning label signals with confirmation labels. Subsequently, the equipment terminal cross-validates the multi-source stress data associated with the multi-level early warning label signals based on locally stored power device operating data. It compares the consistency between the real-time stress data collected by the terminal and the stress data sent by the system, and verifies the rationality of the early warning levels by combining historical fault records on the terminal side. The verification results are then generated and sent back to the system. After receiving the verification results, the system conducts multi-dimensional fuzzy dynamic analysis based on the multi-level early warning indicator signals: It divides the verification results into multiple dynamic credibility levels, such as high credibility, medium credibility, and low credibility, and analyzes the temporal changes of the verification results to construct a credibility trend chart. It matches the verification results with the credibility levels to generate a credibility radar chart, and then uses the early warning indicator signals as an index to retrieve and compare them to generate a credibility heatmap. The system combines these three types of charts to determine the specific credibility of each early warning signal. Finally, it updates the multi-level early warning signals according to the dynamic credibility levels through a weighted game, adjusting the weight distribution ratio of early warning signals with different credibility levels. Based on this ratio, it verifies and optimizes the original multi-level early warning signals, such as increasing the strength of high-credibility early warning signals and correcting the level of low-credibility early warning signals. Ultimately, it generates and outputs more accurate multi-level optimized early warning signals, completing the closed-loop iteration of the early warning signals.

[0024] In one possible implementation, step S100 further includes: A distributed sensing array is formed by arranging a variety of miniature sensors inside the package and at key locations on the surface of the power device.

[0025] An integrated stress monitoring network is constructed by integrating data through the distributed sensor array: S1: A Hall-cancelled packaged stress sensor is used, which is integrated into the surface or inside the power chip of the power device.

[0026] S2: Place a piezoresistive stress sensor near the lead frame or direct copper-clad substrate of the power device.

[0027] S3: Arrange fiber Bragg grating sensors on the heat dissipation path of power devices.

[0028] Multi-dimensional stress data is acquired through the integrated stress monitoring network. S1: Hall-cancelled package stress sensor is used to measure the first mechanical stress of a power device package.

[0029] S2: A piezoresistive stress sensor is used to monitor the second mechanical stress caused by thermal mismatch.

[0030] S3: Monitor the thermal stress of power devices during long-term operation using fiber optic grating sensors.

[0031] The first mechanical stress, the second mechanical stress, and the thermal stress are integrated to construct the multi-source stress data.

[0032] Specifically, in constructing an integrated stress monitoring network for power devices, the first step is to deploy miniature sensors around the stress-sensitive areas of the power device package using precise deployment and multi-type adaptation methods, forming a distributed sensing array covering key areas. The specific sensor deployment and network integration steps are as follows: First, for the power chip, the core stress source, a Hall-cancelled packaged stress sensor is used, integrated onto the surface or inside the power chip through embedded mounting. This type of sensor can effectively cancel external electromagnetic interference and accurately capture stress signals generated by structural deformation during chip operation. Second, considering that the lead frame and direct copper-clad substrate are areas prone to stress concentration due to differences in material thermal expansion coefficients, piezoresistive stress sensors are mounted nearby to utilize the piezoresistive effect to sense local stress changes caused by thermal mismatch in real time. Third, focusing on the thermal stress generated by temperature fluctuations during long-term operation of the power device, fiber Bragg grating sensors are deployed at key heat transfer nodes along the heat dissipation path, such as the heat dissipation substrate and the bonding points of thermally conductive interface materials. Leveraging their resistance to high temperatures and electromagnetic interference, these sensors stably collect thermal stress data. After deploying the above-mentioned multiple types of sensors in key locations inside and on the surface of the package, the signal acquisition ports of all sensors are integrated through a data bus to achieve data interconnection and collaborative work between sensors. Ultimately, an integrated stress monitoring network is constructed that covers all critical areas from chip to connection structure to heat dissipation path and can synchronously acquire multi-dimensional stress signals.

[0033] After the integrated stress monitoring network is put into operation, targeted multi-dimensional stress data acquisition is carried out based on the functional characteristics of various types of sensors. Among them, Hall-cancelled packaged stress sensors integrated on or inside the power chip, with their advantage of anti-electromagnetic interference, accurately capture the basic stress generated by the power device package under mechanical action, such as structural deformation and external vibration transmission, which is the first mechanical stress. Piezoresistive stress sensors arranged near the lead frame or direct copper-clad substrate use the piezoresistive effect of materials to sense the interface stress caused by the difference in thermal expansion coefficients of different packaged materials in real time. This stress is the second mechanical stress caused by thermal mismatch. Fiber Bragg grating sensors deployed on the heat dissipation path rely on their high sensitivity to temperature changes and stable long-term working performance to continuously collect the thermal stress generated in the package structure of the power device during long-term operation due to temperature fluctuations, such as power cycling and changes in ambient temperature. After the three types of stress data were collected, the data on the first mechanical stress, the second mechanical stress, and the thermal stress were time-series calibrated to ensure that the data collection time was synchronized and the data was structured and integrated. The data was classified and labeled according to stress type, collection location, and collection time. Data noise and redundant information were removed, and finally, a multi-source stress data covering mechanical stress, including two types of stress caused by foundation and thermal mismatch, and thermal stress was constructed, providing a comprehensive and accurate data source for subsequent stress monitoring and analysis.

[0034] In one possible implementation, step S200 further includes: Step S210: Perform accelerated aging simulation on the power device according to the entire life cycle of the power device to obtain full life cycle stress simulation data.

[0035] Step S220: Perform failure analysis based on the full life cycle stress simulation data and construct a failure case library.

[0036] Step S230: Retrieve the maintenance log of the power device, store the full life cycle stress simulation data, the failure case library, and the maintenance log according to time sequence data, and construct a historical packaging status database.

[0037] Step S240: Traverse the historical package states to perform health analysis and extract the historical package health states of the power devices.

[0038] Specifically, the entire lifecycle of power devices is clearly defined, from factory initialization, normal operation, performance degradation to final failure. Combined with the actual application scenarios of the devices, such as the operating conditions of new energy vehicles and industrial converters, targeted accelerated aging test schemes are designed. By controlling ambient temperature and humidity, applying cyclic power loads, and simulating vibration and shock, the natural aging cycle of the devices is compressed, simulating the extreme and normal operating conditions they may face throughout their lifecycle. During the accelerated aging test, a pre-built integrated stress monitoring network is simultaneously activated to collect real-time packaging stress data of the devices at different aging stages, such as the initial stabilization stage, the mid-term degradation stage, and the late-stage critical failure stage. This includes the dynamic changes of various types of stress, such as thermal stress and mechanical stress. Finally, this data, covering all key nodes of the entire lifecycle and reflecting the evolution of stress during the aging process, is organized and calibrated to form complete full lifecycle stress simulation data, providing basic data support for subsequent failure analysis and health status assessment.

[0039] Based on the acquired full lifecycle stress simulation data, this study focuses on packaging failure phenomena of power devices during accelerated aging simulation, such as package cracking, wire bond detachment, substrate warping, and heat dissipation interface failure. In-depth failure analysis is conducted first through data tracing to locate the stress data nodes corresponding to each failure phenomenon, analyzing the stress change trends before failure, such as sudden increases in thermal stress, continuous exceeding of mechanical stress limits, and stress concentration areas, such as the chip-substrate connection edge, lead frame corners, and the coupling relationship of different stress types. Then, combining materials mechanics and packaging process principles, the fundamental causes of various failure phenomena are identified, such as material fatigue caused by long-term thermal stress cycling and structural damage caused by mechanical stress concentration, and the stress threshold under critical failure conditions is quantified. Finally, the analysis results are classified and archived according to a structured dimension of failure type, stress cause, stress change characteristics, critical failure threshold, and packaging damage morphology. Independent cases with data support and mechanism analysis are established for each type of failure phenomenon, ultimately integrating them to form a failure case library covering multiple failure scenarios that can be used for subsequent health status assessment and risk prediction.

[0040] The maintenance logs generated during actual application are retrieved from the power device's operation and maintenance management platform or terminal storage module. These logs must cover the device failure time, failure manifestations (e.g., abnormal package heating, sudden performance degradation), maintenance operations (e.g., replacement of package components, resoldering of leads), stress anomaly feedback before and after the failure, and performance recovery status after maintenance. Subsequently, the acquired full lifecycle stress simulation data, including stress time-series changes at different aging stages and a constructed failure case library containing time-series stress nodes corresponding to various failure phenomena, are processed with the retrieved maintenance logs to unify data format. Using timestamps as the core index, the three types of data are correlated and aligned chronologically. For example, the failure time corresponding to a specific maintenance record is matched and labeled with the stress changes during the same period in the simulation data and the time-series characteristics of similar failures in the failure case library. Finally, this time-series integrated and labeled multi-source data is stored in a pre-defined structured database architecture, constructing a historical package status database that includes three dimensions of information: simulated stress evolution, failure case reference, and actual maintenance feedback, clearly presented along a timeline. This provides a complete data foundation for subsequent extraction of historical package health status.

[0041] A time-series data traversal algorithm is employed to sequentially read full-lifecycle stress simulation data, failure case information, and maintenance logs from the historical encapsulation status database along the timeline. Simultaneously, a data cleaning module removes abnormal noise data, such as instantaneous sensor interference values ​​and log error messages, ensuring data validity. Subsequently, a health feature extraction model is activated. On one hand, statistical analysis algorithms calculate statistical indicators such as the mean, variance, and peak value of stress data during normal operation to determine the stress baseline range under healthy conditions, such as the thermal stress stability range and mechanical stress safety threshold. On the other hand, comparative analysis algorithms analyze the stress change trends before failure in failure cases, such as sudden stress increases. The stress characteristic thresholds corresponding to three health levels—healthy, sub-healthy, and critical failure—are calculated by comparing the stress level, fluctuation frequency, and normal stage data. Simultaneously, an association analysis algorithm is used to match the fault repair time and post-repair stress recovery data from the maintenance log with the aforementioned health levels to verify the actual adaptability of each level's characteristic thresholds. If the post-repair stress returns to within the health threshold, the level classification is deemed valid. Finally, a feature fusion algorithm integrates key parameters such as the health baseline range, graded characteristic thresholds, health decay rate, and the stress change slope calculated from the normal to sub-healthy stages to generate a structured historical encapsulated health status dataset, completing the extraction process.

[0042] In one possible implementation, step S200 further includes: Construct a parallel processing architecture for local and global monitoring.

[0043] Using the historical encapsulation health status as a reference baseline, the multi-source stress data is monitored at multiple scales through the parallel processing architecture: S1: Based on the historical encapsulation health status, perform local monitoring of the multi-source stress data to obtain local multi-scale information.

[0044] S2: Based on the historical encapsulation health status, perform global monitoring of the multi-source stress data to obtain global multi-scale information.

[0045] The local multi-scale information and the global multi-scale information are spatiotemporally registered to obtain registration data.

[0046] Based on the registration data, the local multi-scale information and the global multi-scale information are fused at the feature level to construct an encapsulation stress state map.

[0047] The stress concentration is identified by traversing the stress state spectrum of the package and the key stress concentration areas are determined.

[0048] The key stress concentration area is traced back to the parallel processing architecture for analysis to obtain the bidirectional stress state information.

[0049] Specifically, a local-global parallel processing architecture is first built. This architecture includes two independent and collaborative processing units. The local monitoring unit focuses on key local areas of the power device package, such as stress-sensitive parts like the chip-substrate connection edge, wire bonding points, and heat dissipation interfaces. The global monitoring unit covers the overall package structure, including the package shell, pins, and internal support structures. The two units achieve synchronous data calls and result sharing through a data interaction channel.

[0050] A standardized reference benchmark library is constructed, containing historical package health status data including health stress thresholds for various local regions of power devices, global stress distribution patterns, and stress characteristics at different health levels, providing a basis for data comparison for monitoring. Subsequently, relying on a parallel processing architecture where local and global monitoring units work collaboratively, multi-scale analysis is performed simultaneously on multi-source stress data, including thermal and mechanical stress. Specifically, the local monitoring unit uses health stress parameters of key local regions in the reference benchmark library, such as the power chip surface, wire bonding points, and the edge connecting the substrate and the housing, as comparison standards. It performs refined decomposition of the stress signals of corresponding local regions in the multi-source stress data, analyzing the deviation of real-time stress values ​​from the health benchmark and whether the rate of stress change exceeds the normal range from two dimensions: microscale (e.g., local stress gradient of the micron-level package structure) and mesoscale (e.g., uniformity of stress distribution in local components). Simultaneously, it captures the instantaneous fluctuations and short-term evolution trends of local stress, ultimately integrating these data to form a comprehensive analysis. The system includes local multi-scale information such as the location, degree, and proportion of stress types at local stress anomalies. The global monitoring unit, on the other hand, uses a healthy stress field model of the overall packaged structure in the reference library, such as the uniformity of stress distribution throughout the package and the efficiency of stress transmission across regions, as a reference. It performs macroscopic and mesoscopic analysis on the global stress signals in the multi-source stress data. At the macroscopic scale, it assesses the overall distribution of stress within the entire package, such as whether there are large areas of high stress and the coordinated change patterns of stress in different regions. At the mesoscopic scale, it analyzes the stress interaction between different functional modules, such as heat dissipation modules and lead-wire modules, to determine whether the global stress deviates from the healthy distribution range. This results in the generation of global multi-scale information including the range of global stress anomaly areas, overall stress trends, and cross-module stress correlation characteristics.

[0051] After completing the acquisition of local and global multi-scale information, a spatiotemporal registration algorithm is needed to eliminate the deviations between the two types of information in the time and space dimensions to ensure data consistency. The specific process is as follows: In the time dimension, local multi-scale information, such as chip surface stress data, and global multi-scale information, such as full-package stress field data, are first extracted. Each is collected with a timestamp. Time synchronization calibration algorithms, such as interpolation completion based on data acquisition frequency and time alignment based on key events, are used to correct the time difference caused by differences in sensor response speed and data transmission delay, so that the stress data at the same moment in the two types of information are accurately matched. In the spatial dimension, based on the three-dimensional structural model of the power device package, the micro / mesoscopic regions corresponding to the local multi-scale information, such as wire bonding points and chip edge coordinates, are mapped to the macroscopic coordinate system of the global multi-scale information. Spatial coordinate transformation algorithms, such as coordinate calibration based on package reference marker points and scaling matching based on structural size ratio, are used to eliminate the deviations in spatial positioning between the local area and the global structure, such as coordinate errors caused by the offset of local sensor installation positions. After time synchronization and spatial calibration, the stress characteristics in the local multi-scale information, such as the stress value of a certain bonding point, can be accurately correlated with the stress characteristics of the corresponding spatial location in the global multi-scale information, such as the distribution role of the bonding point in the full-encapsulation stress field. This results in registration data that is highly consistent in spatiotemporal dimensions and can be directly used for subsequent fusion analysis.

[0052] After obtaining spatiotemporally consistent registration data, feature-level fusion is performed on local and global multi-scale information based on the registration data to construct a complete packaging stress state map. The specific process is as follows: First, key stress features at the micro and mesoscopic levels are extracted from the registered local multi-scale information, including local stress-sensitive areas such as the stress peak, stress gradient change, stress concentration coefficient, and local stress fluctuation frequency and attenuation trend over time at the chip-substrate connection edge and wire bonding point. At the same time, core stress features at the macro and mesoscopic levels are extracted from the registered global multi-scale information, covering the uniformity of stress field distribution across the entire packaging range, stress transmission paths between different functional modules, the deviation rate between the global stress maximum value and average value, and the correlation coupling coefficient of cross-regional stress. Subsequently, a mapping relationship between local and global features is established through feature association algorithms. For example, the stress value of a local stress concentration point is bound to the stress distribution trend of the region where that point is located in the global stress field, clarifying the influence weight of local stress anomalies on global stress balance. Then, feature fusion models, such as weighted fusion and feature splicing fusion, are used to integrate the two types of features, embedding microscopic local stress details into the macroscopic global stress framework, forming a feature set that combines local precision and global integrity. Finally, using a three-dimensional structural model of the power device package as a carrier, the fused stress features are visualized according to spatial coordinates, with different colors used to mark stress levels (e.g., blue for low stress, red for high stress), arrows to indicate stress transmission direction, and special symbols to mark stress concentration points. Ultimately, a package stress state map that intuitively reflects the local-global stress distribution, variation trend, and correlation relationship inside the package is constructed.

[0053] A pre-defined stress concentration identification algorithm traverses the stress state map of the package. Using stress safety thresholds determined from historical package health states, including local stress limits and global stress distribution equilibrium thresholds, as core criteria, a multi-dimensional identification and analysis is conducted: First, stress values ​​corresponding to each spatial coordinate in the map are scanned point-by-point, and stress points exceeding the safety threshold are selected and marked as initial stress anomaly points. Then, regional clustering analysis is performed on the initial anomaly points, grouping spatially adjacent anomalies with the same stress anomaly type (e.g., both exceeding thermal stress limits or mechanical stress concentration) into anomaly regions. The average stress, maximum stress, and stress gradient change rate of each anomaly region are calculated. First, isolated anomalies caused by transient sensor interference are excluded. Then, the risk level of each anomaly area is assessed by combining the correlation between stress concentration and failure in the historical failure case library. For example, if an anomaly area is located at the edge of the chip-substrate connection and the stress gradient exceeds the failure threshold, it is identified as a high-risk area. Finally, from the high-risk anomaly areas, areas with significantly exceeded stress values, rapid stress change rates, and significant impact on core package functions such as heat dissipation and signal transmission are further screened out and identified as key stress concentration areas. Their specific spatial range, stress peak value, and stress type are highlighted in the map to provide clear targets for subsequent backtracking analysis.

[0054] After identifying key stress concentration areas, including their spatial coordinates, stress peak values, stress types, and risk levels, and backtracking them to a parallel processing architecture of local and global monitoring, the two units in the architecture conduct collaborative analysis based on their respective monitoring dimensions. On one hand, the local monitoring unit calls upon the original acquired signals of the key area in multi-source stress data, and combines them with the health stress characteristics of corresponding local parts in historical package health status to re-verify the micro-stress details of the key area. For example, it analyzes whether the stress concentration in the area is accompanied by microstructural deformation signals, such as stress wave changes in chip crack precursors, the specific magnitude and duration of stress values ​​exceeding the health threshold, and supplements the generation of local-level stress anomaly in-depth analysis results. On the other hand, the global monitoring unit, based on the spatial location of the key area, infers its impact on the overall package stress field. For example, it determines whether the stress in the area is transmitted to surrounding heat dissipation modules and lead frames, whether it causes an imbalance in the global stress distribution, and whether it triggers compensatory increases in stress in other areas. Combining historical global stress anomaly cases, it outputs the macroscopic impact assessment results of the key area on the overall package stress state. Finally, the parallel processing architecture integrates the microscopic analysis results of the local monitoring unit and the macroscopic evaluation results of the global monitoring unit through the data interaction channel, forming bidirectional stress state information that simultaneously contains the local fine features of the key stress concentration area and the global features of the area's impact on the overall stress of the package, thus fully presenting the local details and global correlation of the stress problem in the key area.

[0055] In one possible implementation, step S300 further includes: Step S310: Construct multiple feature analysis layers, including a time-domain feature analysis layer, a frequency-domain feature analysis layer, and a time-frequency-domain feature analysis layer.

[0056] Step S320: Synchronize the bidirectional stress state information to the time-domain feature analysis layer for calculation to obtain time-domain feature parameters.

[0057] Step S330: Synchronize the bidirectional stress state information to the frequency domain feature analysis layer for calculation to obtain frequency domain feature parameters.

[0058] Step S340: Synchronize the bidirectional stress state information to the time-frequency domain feature analysis layer for calculation to obtain time-frequency domain feature parameters.

[0059] Step S350: Based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters, perform dimensionality reduction and screening to obtain multiple feature parameters. Transfer these parameters to the full life-cycle stress simulation data for evolution prediction to generate a packaging stress evolution prediction trend.

[0060] Specifically, multiple feature analysis layers are designed and constructed, including a time-domain feature analysis layer, a frequency-domain feature analysis layer, and a time-frequency-domain feature analysis layer. Each layer is functionally independent yet interconnected. The time-domain feature analysis layer uses the time axis as its core dimension and is equipped with a sliding window processing module and a statistical calculation module for calculating indicators such as mean, variance, and peak value. It is specifically designed to capture the static distribution of stress data over time, such as the mean stress and dynamic fluctuations during stable operation, such as instantaneous stress pulse characteristics. The frequency-domain feature analysis layer uses the frequency axis as its core dimension and integrates a Fast Fourier Transform (FFT) module and a power spectrum analysis module. It is responsible for converting time-domain stress data into frequency-domain signals and analyzing the energy distribution of stress at different frequency bands, such as high-frequency fluctuations corresponding to local vibrations and low-frequency changes corresponding to slow temperature drift. The time-frequency-domain feature analysis layer takes into account both time and frequency dimensions and is equipped with a wavelet transform module and a time-frequency energy entropy calculation module. It can simultaneously capture frequency mutations of stress at specific time points, such as the frequency anomalies of stress waves before a fault and the time distribution patterns of specific frequency bands, such as stress at a certain frequency only appearing during high-load periods, filling the blind spots of single time-domain or frequency-domain analysis. Each feature analysis layer achieves data interoperability through a standardized data interface, ensuring that subsequent bidirectional stress state information can be synchronously input into each layer for parallel analysis, laying the foundation for comprehensive extraction of stress features.

[0061] The acquired bidirectional stress state information, including local and global stress data, is synchronously input into the time-domain feature analysis layer. This layer calculates the time-domain indices of the stress data through a sliding window algorithm and a statistical analysis model, including the mean, variance, peak, valley, pulse count, and stress change slope of stress in different time periods. Finally, it generates time-domain feature parameters that can reflect the static distribution and dynamic fluctuation characteristics of stress on the time axis.

[0062] The bidirectional stress state information is synchronously transmitted to the frequency domain feature analysis layer. This layer uses Fast Fourier Transform (FFT) to convert the time-domain stress data into a frequency-domain signal. By calculating indicators such as power spectral density, dominant frequency components, and harmonic content, it analyzes the energy distribution law of stress in different frequency bands, such as whether high-frequency stress fluctuations correspond to local structural vibrations, and generates frequency domain feature parameters that characterize the frequency characteristics of stress.

[0063] The bidirectional stress state information is synchronously entered into the time-frequency domain feature analysis layer. This layer uses wavelet transform algorithm to decompose the stress data into sub-signals with different time scales and different frequency components. By extracting features such as modulus maxima, energy entropy, and time-frequency clustering of wavelet coefficients, it captures the frequency mutation of stress at specific time points, such as the abnormal stress wave frequency and time-frequency coupling law before the fault, and obtains time-frequency domain feature parameters.

[0064] Finally, the three types of feature parameters in the time domain, frequency domain, and time-frequency domain are subjected to dimensionality reduction screening and evolution prediction according to the following process to generate the predicted trend of packaging stress evolution: First, linear dimensionality reduction is initiated, employing Principal Component Analysis (PCA) to compress the dimensionality of the three types of feature parameters. A 95% retention rate for principal components is explicitly set; this is achieved by calculating the variance contribution rate of each principal component and selecting the set of principal components whose sum of variance contribution rates reaches 95%. This significantly reduces data dimensionality and computational redundancy while preserving the core information strongly correlated with packaging stress changes in the feature parameters, avoiding the loss of key stress features due to dimensionality reduction. After PCA dimensionality reduction, a random forest feature importance assessment model is introduced. The dimensionality-reduced feature parameters are correlated with stress evolution labels in the historical packaging health states of power devices, such as stress growth stage, stable stage, and critical failure stage, for training. The model calculates the contribution of each dimensionality-reduced feature to the stress evolution stage identification task, generating an importance score for each feature. Subsequently, the dimensionality-reduced features are sorted from highest to lowest importance score, initially selecting the top N feature parameters. N is initially set to 20% of the total number of features after PCA dimensionality reduction, serving as the initial feature set for recursive elimination. Next, a recursive feature elimination (RFE) operation is performed, using the mean squared error (MSE) of the encapsulated stress evolution trend prediction model as the core evaluation metric for feature selection. The termination condition for recursive elimination is defined as follows: elimination stops when the number of features in the current feature set decreases to 10, regardless of changes in model accuracy; or, during feature removal, if the model's prediction accuracy, measured by MSE, decreases by more than 1% compared to the previous round after a certain round of feature removal, elimination stops immediately to avoid excessive removal leading to significant degradation in model performance. Following this termination condition, redundant features with the lowest importance and least contribution to model accuracy are gradually removed from the initial feature set, ultimately forming an optimized combination of feature parameters that balances dimensionality reduction and prediction accuracy. Subsequently, a deep learning prediction model, such as a Long Short-Term Memory (LSTM) network, is constructed based on this optimized feature parameter combination. Using stress time-series data from the historical encapsulated health status of power devices as training samples, model parameters, such as the learning rate and the number of hidden layer nodes, are iteratively optimized to enable the model to accurately learn the mapping relationship between feature parameter changes and stress evolution, generating feature learning results. Finally, the feature learning results are aligned in format and dimension through the data adaptation module and progressively migrated to the full life cycle stress simulation data, which includes stress benchmark data for the entire stage of the device from initialization to failure. Combining the historical patterns of stress stage-feature changes in the full life cycle stress simulation data, the numerical changes, stage transitions, and critical threshold approach of the current power device packaging stress in the future preset period are predicted. Finally, a packaging stress evolution prediction trend including stress time series change curves, key time nodes, and risk evolution paths is generated.

[0065] In one possible implementation, step S350 further includes: Step S351: Perform linear dimensionality reduction based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters, evaluate the importance of the dimensionality reduction results, and generate multiple importance scores.

[0066] Step S352: Filter the dimensionality reduction results according to the multiple importance scores, extract the top N feature parameters, recursively eliminate the top N feature parameters, and construct an optimized combination of feature parameters, where N is a positive integer greater than 0 and less than the number of dimensionality reduction results.

[0067] Step S353: Perform deep learning based on the optimized combination of the feature parameters to generate feature learning results.

[0068] Step S354: Progressively transfer the feature learning results to the full life cycle stress simulation data for evolution prediction, and generate the packaging stress evolution prediction trend.

[0069] Specifically, for the acquired time-domain feature parameters, such as mean stress, peak stress, and slope of change, and frequency-domain feature parameters, such as power spectral density, dominant frequency components, and time-frequency domain feature parameters, such as wavelet energy entropy and time-frequency clustering, a principal component analysis (PCA) linear dimensionality reduction algorithm is used. This algorithm compresses the feature dimensions to reduce computational redundancy while preserving core stress information. Subsequently, an importance assessment is conducted based on the core criterion of the influence of features on the packaging stress of power devices. By constructing a stress influence correlation model, the correlation between each dimensionality-reduced feature and changes in packaging stress, such as stress increase and stress concentration, is analyzed. The higher the correlation and the stronger the early warning sensitivity to stress anomalies, the higher the importance score of the feature. Finally, a corresponding importance score is generated for each dimensionality reduction result.

[0070] All dimensionality reduction results are sorted in descending order based on importance scores, and the top N feature parameters with the highest scores are selected, where N is a positive integer greater than 0 and less than the total number of dimensionality reduction results. To further eliminate redundant features and focus on core influencing factors, a recursive elimination operation is performed on the top N feature parameters. By gradually removing a feature, its impact on the judgment of the packaging stress change trend is verified: if the removal does not significantly reduce the recognition accuracy of the packaging degradation state, it is considered redundant and should be removed; if the removal significantly reduces the accuracy, it is determined that the feature is sensitive to packaging degradation and should be retained. Finally, the subset of feature parameters most sensitive to packaging degradation is selected to construct an optimized combination of feature parameters.

[0071] The optimized combination of feature parameters is input into a pre-defined deep learning model, such as a Long Short-Term Memory (LSTM) network. The corresponding data of feature parameter-stress evolution in the historical health state of the package are used as training samples. The model parameters are optimized through iterative training, so that the model can accurately learn the mapping relationship between sensitive feature parameters and the package stress degradation process, such as the correlation between the increase of a certain feature parameter and the acceleration of stress concentration. Finally, feature learning results that can accurately represent the corresponding pattern of sensitive feature-stress change are generated.

[0072] Finally, the feature learning results are progressively transferred to the full life cycle stress simulation data: First, the format and dimensions of the feature learning results are aligned and matched with the full life cycle stress simulation data, including stress benchmark values ​​and failure critical stress thresholds for different aging stages, through the data adaptation module; then, using the trained deep learning model, combined with the historical patterns of stress evolution stages and feature changes in the full life cycle stress simulation data, the evolution trend of the current encapsulation stress to the future stage is predicted, including the stress value change curve over time, the time nodes when stress may exceed limits, and the expansion path of stress concentration areas, etc., ultimately generating a complete prediction trend of encapsulation stress evolution.

[0073] In one possible implementation, step S400 further includes: Step S410: Perform stress over-limit analysis based on the predicted trend of the packaging stress evolution to generate a stress over-limit risk index.

[0074] Step S420: Based on the stress over-limit risk index, conduct a risk assessment of the stress change in the predicted trend of the packaging stress evolution, and generate stress change risk data.

[0075] Step S430: Based on the predicted trend of packaging stress evolution, perform cumulative damage assessment on the packaging stress of power devices to generate fatigue damage risk data.

[0076] Step S440: Perform multi-dimensional evaluation and calculation based on the stress over-limit risk index, the stress change risk data, and the fatigue damage risk data to obtain the composite risk index of power device packaging stress.

[0077] Step S450: Calculate the risk contribution based on the composite risk index, allocate weights according to the risk contribution degree, and determine multiple contribution weights.

[0078] Step S460: Divide the composite risk index into multiple levels according to the multiple contribution weights to construct a multi-level early warning signal.

[0079] Specifically, based on the predicted stress values ​​at each future time point in the packaging stress evolution prediction trend, and compared with the stress safety thresholds determined by the historical packaging health status, including the upper limit values ​​of different stress types and the critical values ​​of local areas, the probability, magnitude, and duration of stress values ​​exceeding the threshold at each time point are calculated. These parameters are then converted into a quantifiable stress over-limit risk index through a weighted summation algorithm. The higher the index, the higher the probability and severity of future stress over-limit.

[0080] Based on the stress over-limit risk index, a risk assessment is conducted on the stress change characteristics in the predicted trend of packaged stress evolution, such as the stress rise rate, fluctuation frequency, and abrupt change amplitude. If the stress over-limit risk index is high and accompanied by a sudden increase in stress and high-frequency fluctuations, it is judged as a high-risk change; if the index is low and the stress change is gradual, it is judged as a low-risk change. Finally, the assessment results are converted into stress change risk data that includes risk level, such as high / medium / low, and risk cause, such as thermal stress dominance or mechanical stress dominance.

[0081] Based on the predicted trend of packaging stress evolution, combined with material fatigue damage theory, such as Miner's linear cumulative damage rule, the cumulative damage assessment of power device packaging stress is carried out: the number of stress cycles within the prediction period is statistically analyzed, such as temperature cycles of thermal stress and vibration cycles of mechanical stress, the damage to the packaging material of each stress cycle is calculated and accumulated, and it is determined whether the cumulative damage is close to the critical damage value of material failure, generating fatigue damage risk data that includes the degree of cumulative damage and the remaining life estimate.

[0082] A multi-dimensional assessment algorithm is adopted to integrate and calculate the stress over-limit risk index, stress change risk data, and fatigue damage risk data. First, the three types of data are standardized and converted into a unified score of 0-100. Then, through a weighted average or neural network fusion model, a composite risk index that can comprehensively reflect the overall risk level of the encapsulation stress is output. The higher the score, the higher the overall risk.

[0083] Using the composite risk index as a benchmark, risk contribution calculation models, such as variance decomposition and Shapley value analysis, are employed to decompose the contribution ratios of the three dimensions—stress over-limit risk, stress change risk, and fatigue damage risk—to the composite risk index. For example, by comparing the decrease in the composite index after removing only stress over-limit risk data and the decrease after removing only fatigue damage risk data, the influence of each risk dimension in the composite index is quantified. Subsequently, weights are allocated based on the calculated risk contribution ratios, following the principle that the higher the contribution, the larger the weight. The contribution ratios of each risk dimension are converted into corresponding contribution weights. For example, a 40% contribution from stress over-limit risk corresponds to a weight of 0.4, a 35% contribution from fatigue damage risk corresponds to a weight of 0.35, and a 25% contribution from stress change risk corresponds to a weight of 0.25, ensuring that the total weights are 1. Finally, the contribution weights of each of the three risk dimensions are determined.

[0084] Based on the composite risk index and combined with the contribution weight of each risk dimension, a multi-level classification standard is formulated: for example, if the composite risk index is below 30 and the risk type with high contribution weight is low-level, it is classified as Level 1 warning, low risk; if the index is 30-60 and the risk type with high contribution weight is medium-level, it is classified as Level 2 warning, medium risk; if the index is above 60 or the risk type with high contribution weight is high-level, it is classified as Level 3 warning, high risk. Finally, a multi-level warning signal is constructed that includes warning level, risk-dominant factors, and suggested countermeasures.

[0085] In one possible implementation, step S500 further includes: Step S510: Perform bidirectional analysis between the multi-level early warning signals and the device terminal to construct a bidirectional communication mechanism.

[0086] Step S520: The multi-level early warning signal is sent to the device terminal through the bidirectional communication mechanism to obtain the terminal sending result.

[0087] Step S530: Based on the result sent by the terminal, confirm the receipt of the device terminal and generate a terminal confirmation signal.

[0088] Step S540: The terminal confirmation signal is fed back to the multi-level early warning signal for identification through the bidirectional communication mechanism, and the multi-level early warning identification signal is determined.

[0089] Step S550: Perform cross-validation on the multi-level warning indicator signals to generate validation results.

[0090] Step S560: Based on the verification results, evaluate the credibility of the multi-level early warning indicator signal, generate multiple signal credibility values, update the multi-level early warning signal according to the multiple signal credibility values, and generate a multi-level early warning optimized signal.

[0091] Specifically, based on the established multi-level early warning signals, including information such as early warning level and risk-inducing factors, a two-way analysis mechanism is established with the equipment terminals to which the power devices belong, such as industrial control terminals and operation and maintenance management platforms. On the one hand, the receiving format and feedback timeliness requirements of the terminals for early warning signals are clarified, and on the other hand, the verification standards for the receiving terminals' confirmation information are determined. On this basis, a two-way communication mechanism including a data encryption transmission module, a signal format adaptation module, and a feedback timeout monitoring module is built to ensure the stability and security of signal transmission and feedback.

[0092] Multi-level early warning signals are encapsulated and distributed according to terminal-compatible formats through a two-way communication mechanism, such as using MQTT or industrial Ethernet protocols for transmission. At the same time, distribution status monitoring is initiated: the signal transmission progress is tracked in real time, and key information such as distribution time and terminal address is recorded. After the terminal successfully receives the signal, the received terminal returns a signal and a response information, which is integrated to form a terminal distribution result that includes distribution status, success / failure, terminal reception time, and response content.

[0093] The device terminal's reception status is confirmed based on the terminal's output results: if the output results show that the signal was successfully received and the response information is complete, the terminal reception is deemed valid, and a terminal confirmation signal containing the confirmation time, terminal identifier, and warning signal association ID is generated; if the output results show that the signal transmission failed or no response was received after timeout, the terminal reception is deemed invalid, and a terminal confirmation signal is not generated temporarily.

[0094] The terminal confirmation signal is fed back to the multi-level early warning signal management module through a two-way communication mechanism. The early warning signal is processed according to the rule of receiving confirmation and marking: for early warning signals that have received terminal confirmation signals, a confirmed reception mark is added, such as marking the confirmation timestamp and terminal ID; for early warning signals that have not returned confirmation signals, including transmission failure and timeout, no mark is added, and only the original early warning information is retained. Finally, a multi-level early warning mark signal is formed that includes both marked and unmarked signals.

[0095] A cross-validation system is constructed. At the data layer, the core information contained in the multi-level warning signals, such as the warning level and risk-dominant factors (e.g., excessive thermal stress, fatigue damage), is accurately compared with the real-time power device operating data uploaded by the equipment terminal to determine whether the risk state predicted by the warning signal is consistent with the actual operating conditions of the equipment. At the logic layer, the rigor of the warning signal generation logic is verified. For example, the weight allocation and index calculation process of the warning levels based on the composite risk index are consistent with the preset algorithm. Key parameters, such as the stress safety threshold and fatigue damage coefficient, are the latest calibrated values ​​to eliminate warning deviations caused by algorithm loopholes or parameter errors. At the historical layer, warning records under similar operating conditions are retrieved from the historical database of the equipment terminal. The risk characteristics of the current warning signal, such as the warning triggering time and risk evolution trend, are compared with the similarity to historical effective warnings. If the current signal highly matches the warning patterns that have been verified multiple times in the past, its verification credibility is improved. Based on the combined results of the three-layer verification, each multi-level early warning indicator signal is marked as "verification passed," meaning there are no deviations in the three-layer verification and the early warning is completely consistent with reality, logic, and history; "requires review," meaning there is a slight deviation in a single dimension, such as the actual stress value being close to but not reaching the early warning threshold, requiring further observation; or "verification failed," meaning there are significant deviations in multiple dimensions, such as the early warning risk type being completely inconsistent with the actual fault cause. Finally, a structured verification result is generated.

[0096] Establish a credibility assessment standard corresponding to the verification results. For the multi-level warning signal that "verification passed", a high credibility score of 80-100 points is assigned because there are no problems at the data layer, logic layer and historical layer. For the signal that "requires review", a medium credibility score of 40-79 points is assigned because there is a slight deviation in a single dimension. For the signal that "verification failed", a low credibility score of 0-39 points is assigned because of significant deviations in multiple dimensions. Finally, a specific credibility value is generated for each signal. Subsequently, the original multi-level early warning signals were updated according to their reliability: high-reliability signals retained their original warning level and risk description, with only a reliability score added to enhance their reference value; medium-reliability signals incorporated real-time operating data from equipment terminals, such as the latest stress change trends, to adjust the warning content, for example, slightly adjusting "medium risk" to "low risk," and marking deviation points to prompt maintenance personnel to pay close attention; low-reliability signals were backtracked to the composite risk index calculation stage to verify the original data and algorithm logic for risk dimensions such as stress exceeding limits and fatigue damage, correcting erroneous parameters, such as updating stress safety thresholds, and regenerating the early warning signal to ensure that the optimized signal was consistent with the actual operating conditions. Finally, all updated signals were integrated to form a multi-level optimized early warning signal that combines accuracy, guidance, and reliability marking.

[0097] In one possible implementation, step S560 further includes: Step S561: Based on the verification results and the multi-level early warning indicator signals, perform multi-dimensional fuzzy dynamic analysis, divide multiple dynamic credibility levels for time-series change analysis, and construct a credibility trend chart.

[0098] Step S562: Match the verification results according to the multiple dynamic credibility levels to construct a credibility radar chart.

[0099] Step S563: Using the multi-level early warning indicator signal as an index, the credibility radar map is searched and compared to extract the credibility heat map.

[0100] Step S564: Perform credibility analysis based on the credibility trend chart, the credibility radar chart, and the credibility heatmap to determine the credibility of multiple signals.

[0101] Step S565: Feed back the credibility of the multiple signals to the multi-level early warning signal according to the multiple dynamic credibility levels to perform weight game update and determine multiple weight allocation ratios.

[0102] Step S566: The multi-level early warning signal is tested and optimized according to the multiple weight allocation ratios, and the multi-level early warning optimized signal is...

[0103] Specifically, the cross-validation results—that is, the core information of the multi-level early warning signals (i.e., the early warning level, risk dimension, and terminal feedback status)—are integrated with the results of cross-validation (i.e., verification passed / requires review / failed). A four-dimensional analysis framework is then constructed, consisting of data matching degree, logical rigor, historical similarity, and real-time feedback rate. Fuzzy comprehensive evaluation is used to dynamically quantify each dimension. For example, the "data matching degree" dimension is scored based on the deviation rate between the signal and the actual stress value of the terminal; the "logical rigor" dimension is scored based on the calibration status of the algorithm parameters; the "historical similarity" dimension is scored based on the matching degree with the pattern of historical effective early warnings; and the "real-time feedback rate" dimension is scored based on the timeliness of the terminal's confirmation of the signal. Based on the four-dimensional quantification results, the signal credibility was divided into five dynamic credibility levels using a clustering algorithm: extremely high credibility (90-100 points), high credibility (70-89 points), medium credibility (40-69 points), low credibility (10-39 points), and extremely low credibility (0-9 points). Subsequently, with time as the horizontal axis and the corresponding credibility level score as the vertical axis, the credibility changes of each multi-level warning indicator signal were tracked at different time nodes, such as when the signal was issued, when the terminal confirmed it, and when cross-validation was performed. The trend of credibility rise and fall was captured by time series analysis tools. For example, the credibility of a certain "requires verification" signal increased from 55 points to 82 points after the terminal's real-time data was supplemented. Finally, a credibility trend chart that can intuitively show the time series fluctuation pattern of the credibility of each signal was drawn, providing a time series basis for subsequent multi-dimensional credibility analysis.

[0104] The five dynamic credibility levels were matched with the three dimensions of the verification results. The credibility radar chart was constructed using the scores of each signal in each dimension (e.g., 85 points for the data layer, 90 points for the logic layer, and 75 points for the history layer) as coordinate axes. The length of the ray in each dimension of the chart represents the credibility score of the corresponding dimension. The larger the area of ​​the closed region of the graph, the more balanced the credibility of the signal in multiple dimensions.

[0105] Using the unique ID of the multi-level warning signal, such as the warning number or terminal address, as an index, the radar chart data of the corresponding signal is retrieved in the credibility radar chart database. The scores of each dimension in the radar chart are converted into color gradients through a heat map generation algorithm, such as red representing high scores and blue representing low scores. A credibility heat map that can quickly identify the shortcomings in the credibility of the signal is extracted. For example, if the color of a certain dimension is too blue, the credibility of that dimension is insufficient.

[0106] Key information is extracted from the credibility trend chart to analyze the credibility fluctuation patterns of various multi-level early warning signals at different time points, such as whether they show a continuous upward, stable, or downward trend. Emphasis is placed on score changes after signal issuance, terminal confirmation, and cross-validation to determine the dynamic stability of signal credibility. For example, if a signal's score rises from 60 points at issuance to 85 points after verification, it indicates that its credibility is continuously improving with data supplementation, and a higher base score can be assigned. Next, the credibility radar chart is interpreted to assess the signal's balance across the three core dimensions of data matching, logical rigor, and historical similarity. If the ray lengths of each dimension of the radar chart are similar and the overall closed area is large, and if the scores for each dimension are all above 80 points, it indicates that the signal has no... If a ray in a certain dimension is too short, such as a historical similarity score of only 40 points, then points need to be appropriately deducted from the final credibility score. Subsequently, further detailed analysis is conducted using a credibility heatmap. The credibility shortcomings of the signal in each dimension are identified through color gradients, such as the data matching degree dimension corresponding to the blue area. The credibility score is then quantitatively adjusted based on the deviation magnitude marked on the heatmap, such as a deviation of 15%. Finally, a weighted summation algorithm is used to calculate the specific credibility value of each multi-level warning indicator signal by comprehensively considering the dynamic stability of the trend chart, the multi-dimensional balance of the radar chart, and the deviation of the shortcomings in the heatmap. For example, trend stability accounts for 40%, dimensional balance accounts for 35%, and shortcomings correction accounts for 25%. This process ultimately determines the credibility of multiple signals.

[0107] The credibility of multiple signals is correlated with their corresponding dynamic credibility levels: extremely high credibility, high credibility, medium credibility, low credibility, and extremely low credibility. This clarifies the influence of different signal levels on various risk dimensions, including stress over-limit risk, stress change risk, and fatigue damage risk. For example, if an extremely high credibility signal uses "fatigue damage risk" as its core warning content, it indicates that the prediction accuracy of this risk dimension is higher, and its weight should be increased. Conversely, if a low credibility signal has low credibility due to prediction deviations in "stress change risk," then the weight of this dimension should be reduced. Subsequently, a weighted game-theoretic update model was initiated: based on the initial weights of each risk dimension, such as stress exceeding limits (0.3), stress variation (0.2), and fatigue damage (0.5), iterative adjustments were made according to the support of signals at different confidence levels for each dimension. This included the positive support of extremely high / high confidence signals for dimensions and the negative weakening effect of low / extremely low confidence signals. For example, if 80% of extremely high confidence signals indicated a valid "fatigue damage risk" warning, the weight of that dimension was increased from 0.5 to 0.6; if 30% of low confidence signals were due to prediction errors in "stress variation risk," the weight of that dimension was decreased from 0.2 to 0.15. During the game, the total weight of all risk dimensions was kept constant at 1, and the adjustment magnitude was positively correlated with the signal confidence level, with extremely high confidence signals having a greater impact on weight adjustments than medium confidence signals. After multiple iterations until the weights stabilized, the final weight allocation ratios for each risk dimension in the multi-level warning signals were determined, providing a precise weight basis for subsequent signal verification and optimization.

[0108] The weighted proportions for each risk dimension—stress over-limit risk, stress change risk, and fatigue damage risk—are assigned (e.g., stress over-limit 0.35, stress change 0.15, fatigue damage 0.5). These are then substituted into the original calculation logic of the multi-level early warning signal to re-verify the composite risk index. The focus is on calculating the contribution of high-weight dimensions to the index. For example, with a fatigue damage risk weight of 0.5, the accuracy of its cumulative damage assessment data must be confirmed. If any deviation is found in the original data of a high-weight dimension, such as an outdated fatigue damage coefficient, it must be immediately corrected and the index recalculated to ensure the index matches the weight allocation. Next, the adjusted early warning level is verified using real-time operational data from the equipment terminal, such as the current packaging stress value and temperature load. For instance, if the high-weight fatigue damage risk displays "medium risk," the actual fatigue damage progress reported by the terminal must be compared to determine if the level needs fine-tuning. If the actual progress is slower than the prediction, it is downgraded to "low risk." Simultaneously, for low-weight dimensions, such as stress change risk with a weight of 0.15, although its impact on the early warning level is reduced, the data integrity still needs to be verified to avoid overlooking potential risks due to complete neglect. Finally, the revised composite risk index, the adjusted warning level, and the test results of each dimension are integrated, and the weight allocation is supplemented with an explanation, such as "this warning is dominated by fatigue damage risk, with a weight of 50%", forming a multi-level warning optimization signal that is logically rigorous, fits the actual working conditions, and has a reasonable weight allocation.

[0109] Example 2, based on the same inventive concept as the power device packaging stress monitoring and early warning method in the foregoing examples, such as... Figure 2 As shown, this application provides a power device packaging stress monitoring and early warning system. The system and method embodiments in this application are based on the same inventive concept. The system includes: The multi-source stress data acquisition module 10 is used to construct an integrated stress monitoring network for power devices to collect data during operation and obtain multi-source stress data.

[0110] The bidirectional stress state information acquisition module 20 is used to introduce the historical package health status of power devices, and to perform parallel monitoring of multi-source stress data based on the historical package health status to obtain bidirectional stress state information.

[0111] The prediction trend construction module 30 is used to perform feature analysis based on the bidirectional stress state information, extract multiple feature parameters to predict the evolution of power devices, and construct a packaging stress evolution prediction trend.

[0112] The multi-level early warning signal construction module 40 is used to conduct risk assessment of the power device packaging stress according to the predicted trend of packaging stress evolution, and to construct multi-level early warning signals.

[0113] The multi-level early warning signal update module 50 is used to send the multi-level early warning signal to the device terminal to identify and verify the multi-source stress data, and update the multi-level early warning signal according to the verification result.

[0114] Furthermore, the system is also used to implement the following functions: A distributed sensing array is formed by arranging various micro-sensors inside and on key locations on the surface of the power device package. This distributed sensing array is then integrated via a data bus to construct an integrated stress monitoring network: S1: A Hall-canceled packaged stress sensor is integrated onto the surface or inside the power chip of the power device; S2: A piezoresistive stress sensor is arranged near the lead frame or direct copper-clad substrate of the power device; S3: A fiber Bragg grating sensor is arranged along the heat dissipation path of the power device. Multi-dimensional stress data is acquired through this integrated stress monitoring network: S1: The Hall-canceled packaged stress sensor measures the first mechanical stress of the power device package; S2: A piezoresistive stress sensor monitors the second mechanical stress caused by thermal mismatch; S3: A fiber Bragg grating sensor monitors the thermal stress of the power device during long-term operation. The first mechanical stress, the second mechanical stress, and the thermal stress are then integrated to construct the multi-source stress data.

[0115] Furthermore, the system is also used to implement the following functions: Accelerated aging simulations are performed on power devices throughout their entire lifecycle to obtain full lifecycle stress simulation data. Failure analysis is then performed based on the full lifecycle stress simulation data to construct a failure case library. Maintenance logs of the power devices are retrieved, and the full lifecycle stress simulation data, the failure case library, and the maintenance logs are stored in time-series format to construct a historical package status database. Health analysis is then performed by traversing the historical package statuses to extract the historical package health status of the power devices.

[0116] Furthermore, the system is also used to implement the following functions: A parallel processing architecture for local and global monitoring is constructed. Using the historical encapsulation health status as a reference baseline, the parallel processing architecture performs multi-scale monitoring of the multi-source stress data: S1: Local monitoring of the multi-source stress data is performed based on the historical encapsulation health status to obtain local multi-scale information; S2: Global monitoring of the multi-source stress data is performed based on the historical encapsulation health status to obtain global multi-scale information; the local multi-scale information and the global multi-scale information are spatiotemporally registered to obtain registration data; the local multi-scale information and the global multi-scale information are fused at the feature level according to the registration data to construct an encapsulation stress state map; stress concentration is identified by traversing the encapsulation stress state map to determine key stress concentration areas; the key stress concentration areas are backtracked to the parallel processing architecture for analysis to obtain the bidirectional stress state information.

[0117] Furthermore, the system is also used to implement the following functions: Multiple feature analysis layers are constructed, including a time-domain feature analysis layer, a frequency-domain feature analysis layer, and a time-frequency-domain feature analysis layer. The bidirectional stress state information is synchronized to the time-domain feature analysis layer for calculation to obtain time-domain feature parameters. The bidirectional stress state information is synchronized to the frequency-domain feature analysis layer for calculation to obtain frequency-domain feature parameters. The bidirectional stress state information is synchronized to the time-frequency-domain feature analysis layer for calculation to obtain time-frequency-domain feature parameters. Based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters, dimensionality reduction and filtering are performed to obtain multiple feature parameters, which are then transferred to the full life-cycle stress simulation data for evolution prediction to generate a packaging stress evolution prediction trend.

[0118] Furthermore, the system is also used to implement the following functions: Linear dimensionality reduction is performed based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters. The importance of the dimensionality reduction results is evaluated to generate multiple importance scores. The dimensionality reduction results are then filtered based on the multiple importance scores to extract the top N feature parameters. These top N feature parameters are then recursively eliminated to construct an optimized combination of feature parameters, where N is a positive integer greater than 0 and less than the number of dimensionality reduction results. Deep learning is then performed based on the optimized combination of feature parameters to generate feature learning results. The feature learning results are then progressively transferred to the full lifecycle stress simulation data for evolution prediction to generate the packaging stress evolution prediction trend.

[0119] Furthermore, the system is also used to implement the following functions: Based on the predicted trend of packaging stress evolution, stress exceedance analysis is performed to generate a stress exceedance risk index. Based on the stress exceedance risk index, a risk assessment is conducted on the stress changes within the predicted trend of packaging stress evolution to generate stress change risk data. Based on the predicted trend of packaging stress evolution, a cumulative damage assessment is performed on the power device packaging stress to generate fatigue damage risk data. Multi-dimensional evaluation calculations are performed based on the stress exceedance risk index, the stress change risk data, and the fatigue damage risk data to obtain a composite risk index for power device packaging stress. Based on the composite risk index, risk contribution is calculated, and weights are allocated according to the risk contribution degree to determine multiple contribution weights. The composite risk index is then divided into multiple levels according to the multiple contribution weights to construct a multi-level early warning signal.

[0120] Furthermore, the system is also used to implement the following functions: A two-way communication mechanism is constructed based on bidirectional analysis of the multi-level early warning signals and the device terminal. The multi-level early warning signals are then sent to the device terminal via this mechanism, and a sending result is obtained. The device terminal receives and confirms the signal based on the sending result, generating a terminal confirmation signal. This terminal confirmation signal is then fed back to the multi-level early warning signals for identification, determining the multi-level early warning identification signals. Cross-validation is performed on the multi-level early warning identification signals to generate verification results. Based on the verification results, the credibility of the multi-level early warning identification signals is evaluated, generating multiple signal credibility levels. The multi-level early warning signals are then updated according to these multiple signal credibility levels, generating an optimized multi-level early warning signal.

[0121] Furthermore, the system is also used to implement the following functions: Based on the verification results and the multi-level early warning indicator signals, a multi-dimensional fuzzy dynamic analysis is performed to divide the data into multiple dynamic credibility levels for time-series change analysis, and a credibility trend chart is constructed. The verification results are matched according to the multiple dynamic credibility levels to construct a credibility radar chart. Using the multi-level early warning indicator signals as an index, the credibility radar chart is searched and compared to extract a credibility heatmap. Credibility analysis is performed based on the credibility trend chart, the credibility radar chart, and the credibility heatmap to determine the credibility of multiple signals. The credibility of the multiple signals is fed back to the multi-level early warning signals according to the multiple dynamic credibility levels for weighted game updates to determine multiple weight allocation ratios. The multi-level early warning signals are then tested and optimized according to the multiple weight allocation ratios to form an optimized multi-level early warning signal.

[0122] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0123] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0124] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for monitoring and early warning of stress in power device packaging, characterized in that, The method includes: An integrated stress monitoring network for power devices is constructed to collect operational data and obtain multi-source stress data. The historical package health status of power devices is introduced, and multi-source stress data is monitored in parallel based on the historical package health status to obtain bidirectional stress status information. Based on the bidirectional stress state information, feature analysis is performed to extract multiple feature parameters for power device evolution prediction and to construct a packaging stress evolution prediction trend. Based on the predicted trend of packaging stress evolution, a risk assessment of the packaging stress of power devices is performed, and a multi-level early warning signal is constructed. The multi-level early warning signals are sent to the device terminal to identify and verify the multi-source stress data, and the multi-level early warning signals are updated based on the verification results.

2. The power device packaging stress monitoring and early warning method as described in claim 1, characterized in that, An integrated stress monitoring network for power devices is constructed to collect multi-source stress data during operation. The methods include: A distributed sensing array is formed by arranging multiple micro-sensors inside the package and at key locations on the surface of the power device; An integrated stress monitoring network is constructed by integrating data through the distributed sensor array: S1: A Hall-cancelled packaged stress sensor is used, which is integrated into the surface or inside the power chip of the power device. S2: Place a piezoresistive stress sensor near the lead frame or direct copper-clad substrate of the power device. S3: Arrange fiber Bragg grating sensors on the heat dissipation path of power devices; Multi-dimensional stress data is acquired through the integrated stress monitoring network. S1: Hall-cancelled package stress sensor used to measure the first mechanical stress of a power device package; S2: A piezoresistive stress sensor is used to monitor the second mechanical stress caused by thermal mismatch; S3: Monitor the thermal stress of power devices during long-term operation using fiber optic grating sensors; The first mechanical stress, the second mechanical stress, and the thermal stress are integrated to construct the multi-source stress data.

3. The power device packaging stress monitoring and early warning method as described in claim 1, characterized in that, The process of constructing historical health states includes the following methods: Accelerated aging simulations were performed on the power devices throughout their entire lifecycle to obtain full lifecycle stress simulation data. Failure analysis is performed based on the full life cycle stress simulation data to construct a failure case library; Retrieve the maintenance log of the power device, store the full life cycle stress simulation data, the failure case library, and the maintenance log according to time sequence data, and construct a historical packaging status database; A health analysis is performed by traversing the historical package states to extract the historical package health status of the power devices.

4. The power device packaging stress monitoring and early warning method as described in claim 1, characterized in that, The method involves parallel monitoring of multi-source stress data based on the historical encapsulation health status to obtain bidirectional stress state information, including: Construct a parallel processing architecture for local and global monitoring; Using the historical encapsulation health status as a reference baseline, the multi-source stress data is monitored at multiple scales through the parallel processing architecture: S1: Based on the historical encapsulation health status, perform local monitoring of the multi-source stress data to obtain local multi-scale information; S2: Based on the historical encapsulation health status, perform global monitoring of the multi-source stress data to obtain global multi-scale information; The local multi-scale information and the global multi-scale information are spatiotemporally registered to obtain registration data; Based on the registration data, the local multi-scale information and the global multi-scale information are fused at the feature level to construct an encapsulated stress state map. The stress concentration is identified by traversing the stress state spectrum of the package and key stress concentration areas are determined. The key stress concentration area is traced back to the parallel processing architecture for analysis to obtain the bidirectional stress state information.

5. The power device packaging stress monitoring and early warning method as described in claim 3, characterized in that, Based on the bidirectional stress state information, feature analysis is performed to extract multiple feature parameters for power device evolution prediction, and a packaging stress evolution prediction trend is constructed. The method includes: Multiple feature analysis layers are constructed, including a time-domain feature analysis layer, a frequency-domain feature analysis layer, and a time-frequency-domain feature analysis layer. The bidirectional stress state information is synchronized to the time-domain feature analysis layer for calculation to obtain time-domain feature parameters; The bidirectional stress state information is synchronized to the frequency domain feature analysis layer for calculation to obtain frequency domain feature parameters; The bidirectional stress state information is synchronized to the time-frequency domain feature analysis layer for calculation to obtain time-frequency domain feature parameters. Dimensionality reduction and filtering are performed based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters to obtain multiple feature parameters which are then transferred to the full life-cycle stress simulation data for evolution prediction, generating a packaging stress evolution prediction trend.

6. The power device packaging stress monitoring and early warning method as described in claim 5, characterized in that, Based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters, dimensionality reduction and filtering are performed to obtain multiple feature parameters which are then transferred to the full life-cycle stress simulation data for evolution prediction, generating a predicted trend for package stress evolution. The method includes: Linear dimensionality reduction is performed based on the time-domain feature parameters, the frequency-domain feature parameters, and the time-frequency-domain feature parameters. The importance of the dimensionality reduction result is evaluated, and multiple importance scores are generated. The dimensionality reduction results are filtered based on the multiple importance scores, the top N feature parameters are extracted, the top N feature parameters are recursively eliminated, and an optimized combination of feature parameters is constructed, where N is a positive integer greater than 0 and less than the number of dimensionality reduction results. Deep learning is performed based on the optimized combination of the aforementioned feature parameters to generate feature learning results; The feature learning results are progressively transferred to the full life cycle stress simulation data for evolution prediction, generating the packaging stress evolution prediction trend.

7. The power device packaging stress monitoring and early warning method as described in claim 1, characterized in that, Based on the predicted trend of packaging stress evolution, a risk assessment of the packaging stress of power devices is performed, and a multi-level early warning signal is constructed. The method includes: Based on the predicted trend of packaging stress evolution, stress exceedance analysis is performed to generate a stress exceedance risk index. Based on the stress over-limit risk index, a risk assessment of stress change in the predicted trend of the packaging stress evolution is performed to generate stress change risk data. Based on the predicted trend of packaging stress evolution, cumulative damage assessment of power device packaging stress is performed to generate fatigue damage risk data. A composite risk index for power device packaging stress is obtained by performing multi-dimensional evaluation and calculation based on the stress over-limit risk index, the stress change risk data, and the fatigue damage risk data. Risk contribution is calculated based on the composite risk index, and multiple contribution weights are determined by weighting according to the degree of risk contribution. The composite risk index is divided into multiple levels according to the multiple contribution weights to construct a multi-level early warning signal.

8. The power device packaging stress monitoring and early warning method as described in claim 1, characterized in that, The method involves sending the multi-level early warning signals to the device terminal to identify and verify the multi-source stress data, and updating the multi-level early warning signals based on the verification results. A two-way communication mechanism is constructed based on the bidirectional analysis of the multi-level early warning signals and the device terminal. The multi-level early warning signal is sent to the device terminal through the bidirectional communication mechanism to obtain the terminal sending result. Based on the result sent by the terminal, the device terminal receives and confirms the receipt, and a terminal confirmation signal is generated. The terminal confirmation signal is fed back to the multi-level early warning signal for identification through the bidirectional communication mechanism, thereby determining the multi-level early warning identification signal; Cross-validation is performed on the multi-level warning indicator signals to generate verification results; Based on the verification results, the credibility of the multi-level early warning indicator signal is evaluated, multiple signal credibility values ​​are generated, and the multi-level early warning signal is updated according to the multiple signal credibility values ​​to generate a multi-level early warning optimized signal.

9. The power device packaging stress monitoring and early warning method as described in claim 8, characterized in that, Based on the verification results, the credibility of the multi-level early warning indicator signal is evaluated, generating multiple signal credibility levels. The multi-level early warning signal is then updated according to these multiple signal credibility levels to generate an optimized multi-level early warning signal. The method includes: Based on the verification results and the multi-level early warning indicator signals, multi-dimensional fuzzy dynamic analysis is performed to divide the data into multiple dynamic credibility levels for time-series change analysis, and a credibility trend chart is constructed. The verification results are matched according to the multiple dynamic credibility levels to construct a credibility radar chart; Using the multi-level early warning indicator signal as an index, the credibility radar map is searched and compared to extract the credibility heat map. Based on the credibility trend chart, the credibility radar chart, and the credibility heatmap, credibility analysis is performed to determine the credibility of multiple signals; According to the multiple dynamic credibility levels, the credibility of the multiple signals is fed back to the multi-level early warning signal for weight game update, and multiple weight allocation ratios are determined. The multi-level early warning signals are tested and optimized according to the multiple weight allocation ratios, and the multi-level early warning optimized signals are obtained.

10. A power device packaging stress monitoring and early warning system, characterized in that, The system is used to implement the power device packaging stress monitoring and early warning method according to any one of claims 1-9, the system comprising: The multi-source stress data acquisition module is used to construct an integrated stress monitoring network for power devices to collect operational data and obtain multi-source stress data. The bidirectional stress state information acquisition module is used to introduce the historical package health status of power devices, and to perform parallel monitoring of multi-source stress data based on the historical package health status to obtain bidirectional stress state information. The prediction trend construction module is used to perform feature analysis based on the bidirectional stress state information, extract multiple feature parameters to predict the evolution of power devices, and construct a prediction trend for the evolution of packaging stress. A multi-level early warning signal construction module is used to assess the risk of power device packaging stress according to the predicted trend of packaging stress evolution and construct multi-level early warning signals. The multi-level early warning signal update module is used to send the multi-level early warning signal to the device terminal to identify and verify the multi-source stress data, and update the multi-level early warning signal according to the verification result.

Images