Intelligent cockpit hmi multi-dimensional endurance stress accelerated loading test method

Abstract

The application relates to the technical field of equipment testing, and discloses an intelligent cockpit HMI multidimensional endurance stress accelerated loading test method. The method realizes real-time collection of multidimensional state data of the HMI under stress through a multimodal sensor array, constructs a digital twin containing a finite element model, a performance simulation model and a failure physical model, and realizes virtual and real state synchronization by using an extended Kalman filter; failure precursors are identified based on unsupervised anomaly detection, and multidimensional combined stress loading profiles are dynamically generated in combination with reinforcement learning to drive an environmental box, a vibration table and a robot arm to exert composite stress; finally, the failure physical model and an LSTM data-driven model are fused to output a remaining life prediction with a confidence interval through Bayesian averaging. The application significantly improves test efficiency and life prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of equipment testing technology, specifically relating to a multi-dimensional durability stress accelerated loading test method for intelligent cockpit HMI. Background Technology

[0002] With the deep integration of intelligent cockpit human-machine interface (HMI) systems into the automotive field, their functional complexity and user interaction frequency have significantly increased, placing higher demands on the long-term reliability of HMI hardware and software. To verify the durability performance of HMIs throughout their entire lifecycle, the industry commonly adopts accelerated stress loading testing methods, which simulate aging and failure behavior after long-term use by applying mechanical, thermal, or electrical stresses higher than normal use intensity. However, existing accelerated testing systems are mostly based on preset fixed stress spectra and static loading strategies, lacking a perception and feedback mechanism for the real-time response status of the tested object, resulting in the inability to dynamically adjust stress parameters during the testing process to approximate the actual failure boundary.

[0003] The multidimensional durability stress accelerated loading test for intelligent cockpit HMIs focuses on quantitatively evaluating the collaborative degradation behavior of core components such as touch panels, voice recognition modules, display units, and multimodal interaction logic under complex stress fields (such as temperature cycling, vibration and shock, high-frequency clicking, and humid and hot environments). This testing approach aims to induce potential failure modes within a finite time through high-density stress excitation, thereby supporting the construction of life prediction models.

[0004] However, current testing schemes typically embed stress conditions and loading timing into the test script. Even if early abnormal signals are observed during the test (such as a sudden increase in response delay, an increase in false trigger rate, or the initial appearance of microcracks in the material), the test path cannot be reconstructed or the stress combination optimized online.

[0005] Existing technologies suffer from multiple limitations: testing strategies lack adaptability, failing to dynamically adjust stress intensity and dimensional weights based on real-time HMI performance degradation data; digital representation is disconnected from physical testing, lacking an iteratively updatable high-fidelity virtual mapping to invert internal state evolution; and lifetime prediction relies on empirical extrapolation, making it difficult to integrate multi-source sensor data and degradation trajectories to achieve accurate remaining lifetime estimation. These issues are particularly prominent in verification scenarios for next-generation, highly integrated, and highly interactive intelligent cockpits, severely restricting the compression of HMI product development cycles and the improvement of reliability assurance levels. Summary of the Invention

[0006] This invention provides a multi-dimensional durability stress accelerated loading test method for intelligent cockpit human-machine interface, aiming to solve the technical problems in the prior art such as time-consuming accelerated testing process, fixed test scheme, inability to dynamically optimize based on real-time test results, and insufficient accuracy in predicting the lifespan of the human-machine interface throughout its entire life cycle after testing.

[0007] To address the aforementioned technical problems, this invention provides a method for multidimensional durability stress accelerated loading testing of intelligent cockpit human-machine interface. The method utilizes a multimodal sensor array deployed on the hardware of the human-machine interface under test to collect in real time its physical state response data, performance degradation data, and structural integrity characterization data under stress loading.

[0008] Simultaneously, a multiphysics coupled digital twin model is constructed that maps to the tested human-computer interaction interface entity. The digital twin model includes a component-level finite element thermodynamics and structural mechanics model, a system-level performance simulation model, and a degradation mechanism model based on failure physics.

[0009] A state estimation algorithm based on Kalman filtering is used to fuse real-time data collected by the multimodal sensor array with the predicted state of the digital twin model, achieving real-time synchronization between the state of the digital twin model and the state of the physical entity. Furthermore, an anomaly detection model based on unsupervised learning is used to analyze the synchronized digital twin model state data, identifying and quantifying early failure precursor characteristics of key thermal and vibration stress-sensitive components.

[0010] Based on the quantified failure precursor features, a reinforcement learning-based test strategy optimization engine takes the health status of the digital twin model as input and dynamically generates a multi-dimensional combined stress loading profile for the next test cycle. The multi-dimensional combined stress loading profile drives the stress loading device to apply precise environmental and load stresses to the tested human-machine interface.

[0011] Finally, by using the Bayesian model averaging method, the lifespan prediction results based on failure physics in the digital twin model are fused with the lifespan prediction results driven by the measured health index sequence based on the long short-term memory network to generate a probability distribution prediction of the remaining lifespan of the tested human-computer interaction interface with confidence intervals.

[0012] This invention provides a method for accelerated loading testing of multidimensional durability stress on the human-computer interaction interface of a smart clock, comprising:

[0013] A multimodal sensor array deployed on the hardware of the human-computer interaction interface under test is used to collect multidimensional state data of the human-computer interaction interface under test in real time under stress loading.

[0014] Construct and run a digital twin model that maps to the tested human-computer interaction interface entity. The digital twin model includes a preset component-level finite element model, a system-level performance model, and a failure physics model.

[0015] The real-time collected multidimensional state data is fused and the model prediction state of the digital twin model is synchronized to obtain the calibrated real-time digital twin state.

[0016] Based on the real-time digital twin status, assess the health status of the tested human-computer interaction interface and identify early signs of failure.

[0017] Based on the health status and failure precursor characteristics, a multi-dimensional combined stress loading profile for the next test cycle is dynamically generated through a test strategy optimization model.

[0018] Based on the multidimensional combined stress loading profile, the stress loading device is driven to apply composite stress to the human-machine interface under test.

[0019] Based on the real-time digital twin status and historical multidimensional status data, the remaining lifespan prediction results of the tested human-computer interaction interface are continuously updated and output.

[0020] As one embodiment of the present invention, the step of acquiring multi-dimensional state data of the human-computer interaction interface under stress loading in real time through a multi-modal sensor array deployed on the hardware of the human-computer interaction interface under test specifically includes:

[0021] Temperature distribution data is collected by multiple thermocouple sensors attached to the main printed circuit board, display module frame, and key chip packaging surface of the human-machine interface under test.

[0022] Ambient temperature and humidity data are acquired by a temperature and humidity sensor installed in the test environment chamber; vibration response data on the vibration table are collected by a triaxial accelerometer fixed on the mounting bracket of the human-machine interface under test.

[0023] The software proxy program, which communicates with the main processor of the human-machine interface under test, obtains performance parameter data such as the load rate of the central processing unit and the graphics processing unit, the core temperature, the memory usage rate, and the data bus transmission error rate in real time.

[0024] Using an industrial camera and colorimeter facing the human-machine interface display screen under test, the brightness uniformity, chromaticity coordinates, contrast, and pixel defect image data of the display screen are periodically collected; by monitoring the reporting coordinates and response time output by the touch screen controller, the response delay and accuracy drift data of the touch function are obtained.

[0025] Multiple piezoelectric ceramic acoustic emission sensors fixed to the main printed circuit board and display panel structure are used to collect ultrasonic signals released when the structure generates microcracks or delamination under stress. The detection frequency range of the acoustic emission sensors is from 50 kHz to 500 kHz, and the signals are digitized by a data acquisition card with a sampling rate of not less than 5 megasamples per second.

[0026] As one embodiment of the present invention, the construction and operation of the digital twin model mapped to the tested human-computer interaction interface entity specifically includes:

[0027] A refined three-dimensional finite element model of the key thermal stress and vibration stress sensitive components in the human-computer interaction interface under test is established. The key thermal stress and vibration stress sensitive components include the solder joints of the ball grid array packaged chip and the connection between the glass substrate of the display panel and the flexible circuit board.

[0028] The finite element model includes the material's nonlinear constitutive relations, thermal conductivity, thermal expansion coefficient, and elastic modulus as a function of temperature.

[0029] A performance simulation model of the operating system of the human-computer interaction interface under test is established. This model simulates the processor and memory behavior under different software loads through task scheduling queues and resource consumption functions.

[0030] The system integrates multiple analytical models based on failure physics, including the Engelmeier model for calculating the fatigue life of solder joints, the Arrhenius model for the capacitance decay of electrolytic capacitors, and a voltage-time dependent model for the degradation of liquid crystal pixels.

[0031] The finite element model, performance simulation model, and failure physics model are integrated into a unified data framework to form a coupled digital twin that can receive external stress input and simulate the output state response. The data structure of the digital twin is organized using a directed acyclic graph, where nodes represent physical or logical components, and edges represent energy, information, or physical dependencies between components.

[0032] As one embodiment of the present invention, the data fusion and state synchronization of the real-time acquired multidimensional state data with the model prediction state of the digital twin model specifically includes:

[0033] An extended Kalman filter algorithm is employed as the state estimator. The algorithm uses the system state equation of the digital twin model as the prediction model. This system state equation describes the evolution of internal state variables of the model, such as temperature, stress, and strain at each grid node, over time and with input stress. The algorithm uses the measurements from the multimodal sensor array as observation inputs to construct observation equations, which correlate the model state variables with the sensor measurements.

[0034] At each time step, the extended Kalman filter algorithm first predicts the state of the digital twin model at the current time based on the state at the previous time step and the current stress input; then, it calculates the Kalman gain based on the residual between the actual sensor measurement value at the current time step and the measurement value predicted by the model; finally, it uses the Kalman gain to correct the predicted state and outputs the optimal posterior state estimate that integrates the model prediction and the actual measurement information, i.e., the calibrated real-time digital twin state.

[0035] As one embodiment of the present invention, the step of evaluating the health status of the tested human-computer interaction interface and identifying pre-failure characteristics based on the real-time digital twin status specifically includes:

[0036] A set of predefined health indicator vectors are extracted from the real-time digital twin state. The health indicator vectors include the cumulative creep strain energy density of key solder joints, the chromaticity drift of a specific area of ​​the display panel, the highest temperature of the central processing unit under the reference load, and the root mean square value of the acoustic emission signal energy. The health indicator vector sequence of the human-computer interaction interface under test in the initial health stage of the test is used to train an unsupervised anomaly detection model based on a deep convolutional autoencoder.

[0037] During the testing process, the real-time extracted health indicator vector is input into the trained deep convolutional autoencoder to calculate its reconstruction error. When the reconstruction error exceeds the dynamic threshold set based on the 300 sigma principle, it is determined that a failure precursor event has been detected. At the same time, the health indicator vector is input into a pre-trained fault classifier, such as a support vector machine, to locate the root cause of the failure precursor.

[0038] As one embodiment of the present invention, the step of dynamically generating a multi-dimensional combined stress loading profile for the next test cycle based on the health status and failure precursor characteristics through a test strategy optimization model specifically includes:

[0039] A reinforcement learning agent based on proximal policy optimization is used as the test policy optimization model. The state space of the reinforcement learning agent consists of the health indicator vector, the identified failure precursor types, and the test time elapsed; the action space of the reinforcement learning agent is a set of discretized stress loading primitive operations, which include:

[0040] The reward function of the reinforcement learning agent is defined as a composite function that aims to maximize the damage accumulation rate of key thermal stress and vibration stress sensitive components while penalizing the test time. The reward function is the sum of the damage increment of each key thermal stress and vibration stress sensitive component and the preset key weight, and then the product of the test cycle length and the time penalty coefficient. The temperature change rate of the high and low temperature cycle is increased by 5%, the main peak frequency of the random vibration spectrum is shifted by 10 Hz, the relative humidity level is increased by 10 percentage points, and a high-load graphical user interface interaction script is executed for 5 minutes.

[0041] At the end of each test cycle, the reinforcement learning agent selects 6 or a set of stress loading primitive operations based on the current state, and combines them into a multi-dimensional combined stress loading profile for the next test cycle. This profile is sent to the stress loading device in the form of control commands.

[0042] As one embodiment of the present invention, the step of driving the stress loading device to apply composite stress to the tested human-computer interaction interface according to the multidimensional combined stress loading profile specifically includes:

[0043] The stress loading device includes a programmable environmental test chamber, a multi-axis vibration table, and a robotic arm. The programmable environmental test chamber receives temperature and humidity setpoints and rate-of-change commands from the test strategy optimization model, precisely controlling the temperature and humidity environment within the test space. The multi-axis vibration table receives vibration profile commands including spectrum, amplitude, and duration, applying mechanical vibration stress to the human-machine interface under test fixed thereon. The robotic arm is equipped with a flexible stylus with a pressure sensor at its end, receiving screen click coordinate sequences, sliding trajectories, and contact pressure commands to simulate complex user interaction operations, applying electromechanical load stress to the display screen and touch system of the human-machine interface under test.

[0044] As one embodiment of the present invention, the step of continuously updating and outputting the prediction result of the remaining service life of the tested human-computer interaction interface based on the real-time digital twin state and historical multidimensional state data specifically includes:

[0045] The remaining useful life prediction result is generated by fusing the results of two different models. The first model is a failure physics-based prediction model embedded in the digital twin model. Based on real-time updated stress history and component degradation status, it uses analytical formulas such as the Engelmeier model or the Arrhenius model to calculate the theoretical remaining useful life of key thermal and vibration stress-sensitive components. The second model is a sequence data-driven prediction model based on a long short-term memory network. This model takes the time series of historically collected health indicator vectors as input and directly predicts the remaining useful life by learning the pattern of health indicator evolution to final failure in historical test data. Finally, a weighted fusion of the prediction results from the two models is performed using a Bayesian model averaging method. The weights are dynamically adjusted based on the prediction performance of each model within the most recent historical window, with the better-performing model receiving a higher weight. The final output after fusion is a probability density function of the remaining useful life, which includes the expected value and a 95% confidence interval.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] By constructing a high-fidelity digital twin model of the human-computer interaction interface under test and synchronizing it in real time using multimodal sensor data, this invention achieves precise insight into the internal microscopic state of a physical entity.

[0048] This invention introduces a closed-loop adaptive testing strategy optimization mechanism based on reinforcement learning, which enables the test stress to be dynamically and intelligently focused on critical thermal stress and vibration stress-sensitive components that are degrading. This greatly shortens the time required for durability testing without distorting the reproduction of real-world failure mechanisms.

[0049] By integrating a failure physics-based mechanism model with a deep learning-based data-driven model for lifetime prediction, this invention overcomes the limitations of single-model prediction, significantly improves the prediction accuracy and confidence of the reliability of human-computer interaction interfaces throughout their entire lifecycle, and provides unprecedentedly accurate data support for product design optimization and preventive maintenance strategy formulation. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent cockpit HMI multidimensional durability stress accelerated loading test method proposed in this invention;

[0051] Figure 2 This is a schematic diagram of the core principle framework of the adaptive test strategy optimization engine based on digital twins and reinforcement learning in this invention;

[0052] Figure 3This is a logical flowchart of the multimodal sensor data acquisition, digital twin model construction, and state synchronization in this invention.

[0053] Figure 4 This is a flowchart illustrating the logical process of health status assessment and failure precursor identification based on unsupervised learning in this invention.

[0054] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the generation and execution of the multi-dimensional combined stress loading profile in this invention;

[0055] Figure 6 This is a flowchart illustrating the logical flow of the Bayesian model-based fusion prediction mechanism for remaining useful life in this invention. Detailed Implementation

[0056] Please refer to Figures 1 to 6 This invention provides a multi-dimensional durability stress accelerated loading test method for intelligent cockpit human-machine interfaces, aiming to solve the technical problems in existing technologies such as time-consuming accelerated testing processes, fixed test schemes, inability to dynamically optimize based on real-time test results, and insufficient accuracy in predicting the lifespan of the human-machine interface throughout its entire life cycle after testing. The method utilizes a multi-modal sensor array deployed on the hardware of the tested human-machine interface to collect in real-time physical state response data, performance degradation data, and structural integrity characterization data under stress loading.

[0057] Simultaneously, a multiphysics coupled digital twin model is constructed, mapping to the tested human-computer interface entity. This digital twin model includes component-level finite element thermodynamics and structural mechanics models, a system-level performance simulation model, and a degradation mechanism model based on failure physics. Using a Kalman filter-based state estimation algorithm, the real-time data collected by the multimodal sensor array is fused with the predicted state of the digital twin model, achieving real-time synchronization between the state of the digital twin model and the state of the physical entity.

[0058] Furthermore, an anomaly detection model based on unsupervised learning is used to analyze the state data of the synchronized digital twin model, identifying and quantifying early failure precursor features of key thermal and vibration stress-sensitive components. Based on these quantified failure precursor features, a reinforcement learning-based test strategy optimization engine uses the health status of the digital twin model as input to dynamically generate a multi-dimensional combined stress loading profile for the next test cycle. This multi-dimensional combined stress loading profile drives a stress loading device to apply precise environmental and load stresses to the tested human-machine interface.

[0059] Finally, by using the Bayesian model averaging method, the lifespan prediction results based on failure physics in the digital twin model are fused with the lifespan prediction results driven by the measured health index sequence based on the long short-term memory network to generate a probability distribution prediction of the remaining lifespan of the tested human-computer interaction interface with confidence intervals.

[0060] The method includes the following steps:

[0061] S1, by using a multimodal sensor array deployed on the hardware of the human-machine interface under test, multidimensional state data of the human-machine interface under test under stress loading are collected in real time.

[0062] S2, Construct and run a digital twin model that maps to the tested human-computer interaction interface entity, wherein the digital twin model includes a preset component-level finite element model, a system-level performance model, and a failure physics model;

[0063] S3, perform data fusion and state synchronization between the real-time collected multi-dimensional state data and the model prediction state of the digital twin model to obtain the calibrated real-time digital twin state;

[0064] S4. Based on the real-time digital twin status, assess the health status of the tested human-computer interaction interface and identify early signs of failure.

[0065] S5. Based on the health status and failure precursor characteristics, dynamically generate the multi-dimensional combined stress loading profile for the next test cycle through the test strategy optimization model.

[0066] S6, based on the multidimensional combined stress loading profile, drive the stress loading device to apply composite stress to the tested human-machine interface;

[0067] S7. Based on real-time digital twin status and historical multidimensional status data, continuously update and output the prediction result of the remaining service life of the tested human-computer interaction interface.

[0068] In step S1, a multimodal sensor array deployed on the tested human-machine interface hardware is used to collect multidimensional state data of the tested human-machine interface under stress loading in real time. The multimodal sensor array includes multiple thermocouple sensors, temperature and humidity sensors, a triaxial accelerometer, a software agent program, an industrial camera and colorimeter, a touchscreen controller monitoring module, and a piezoelectric ceramic acoustic emission sensor. The thermocouple sensors are attached to the main printed circuit board, display module frame, and key chip package surface of the tested human-machine interface to collect temperature distribution data, with a sampling frequency of no less than 10 times per second.

[0069] The temperature and humidity sensor is installed inside the test environment chamber to acquire ambient temperature and humidity data. Its temperature measurement range is -40℃ to +150℃, and its humidity measurement range is 5% to 98% relative humidity. The triaxial accelerometer is fixed on the mounting bracket of the human-machine interface under test to collect its vibration response data on the vibration table. Its frequency response range covers 0.5 Hz to 5000 Hz.

[0070] The software agent communicates with the main processor of the human-machine interface under test, acquiring performance parameters such as the load rate of the central processing unit and the graphics processing unit, core temperature, memory usage, and data bus transmission error rate in real time. The data refresh cycle is 100 milliseconds. The industrial camera and colorimeter are positioned directly opposite the display screen of the human-machine interface under test, periodically acquiring image data of the screen's brightness uniformity, chromaticity coordinates, contrast ratio, and pixel defects. The image acquisition resolution is 4000×3000 pixels, and the colorimeter accuracy is delta ≤0.005.

[0071] The touchscreen controller monitoring module monitors the reporting coordinates and response time output by the touchscreen controller, acquiring response delay and accuracy drift data for the touch function. The reporting frequency is 120 Hz. The piezoelectric ceramic acoustic emission sensor is fixed to the main printed circuit board and display panel structure, used to collect ultrasonic signals released when the structure generates microcracks or delamination under stress. The detection frequency range of the acoustic emission sensor is 50 kHz to 500 kHz, and its signal is digitized by a data acquisition card with a sampling rate of not less than 5 megasamples per second. The data acquisition card has 16-bit analog-to-digital conversion accuracy.

[0072] In step S2, a digital twin model mapped to the tested human-machine interface entity is constructed and run. The digital twin model includes a component-level finite element model, a system-level performance model, and a failure physics model. The component-level finite element model establishes a refined three-dimensional finite element model for key thermal and vibration stress-sensitive components in the tested human-machine interface. These key thermal and vibration stress-sensitive components include the solder joints of the ball grid array packaged chip and the connection between the glass substrate of the display panel and the flexible circuit board.

[0073] The finite element model uses eight-node hexahedral elements for meshing, with mesh sizes no larger than 0.2 mm in critical regions and no larger than 1.5 mm in non-critical regions. The finite element model includes the material's nonlinear constitutive relations, thermal conductivity, coefficient of thermal expansion, and elastic modulus as a function of temperature. The solder joint material is a tin-silver-copper alloy, whose yield strength decreases exponentially with increasing temperature.

[0074] The system-level performance model simulates the performance behavior of the tested human-computer interface operating system. This model simulates processor and memory behavior under different software loads through a task scheduling queue and a resource occupancy function. The task scheduling queue employs a priority-based preemptive scheduling algorithm, and the resource occupancy function is obtained by fitting the measured processor load and memory bandwidth consumption curves. The failure physics model integrates multiple analytical models based on failure physics, including the Engelmeier model for calculating solder joint fatigue life, the Arrhenius model for electrolytic capacitor capacitance decay, and a voltage-time dependent model for liquid crystal pixel degradation. The expression for the Engelmeier model is:

[0075] ;

[0076] in, For cumulative damage, For the first The range of plastic strain in each cycle The fatigue ductility coefficient, This represents the number of failure cycles. The fatigue strength index. The fatigue ductility index is given. The Arrhenius model expression is:

[0077] ;

[0078] in, For time The capacitance value at that time, This is the initial capacitance value. For activation energy, Boltzmann's constant, This refers to absolute temperature.

[0079] activation energy in the Arrhenius model The methods for obtaining it are as follows:

[0080] Select at least 30 typical samples that are identical in model and production batch to the electrolytic capacitors in the human-machine interface under test, to ensure the consistency and representativeness of the samples.

[0081] Design a multi-temperature-point accelerated aging test scheme, select 4 or more accelerated temperature points higher than the normal operating temperature, and the test duration at each temperature point must ensure that the electrolytic capacitor exhibits quantifiable capacitance decay.

[0082] At each acceleration temperature point, the capacitance data of the electrolytic capacitors were collected periodically. The sampling interval was dynamically adjusted according to the temperature (the sampling interval at high temperature points was 24 hours / time, and the sampling interval at low temperature points was 72 hours / time). At the same time, the capacitance value and the corresponding cumulative aging time at each sampling time were recorded.

[0083] Integral form of the Arrhenius equation The calculation is as follows: , For constants related to the capacitor structure, at different temperature points With reciprocal Perform linear fitting;

[0084] The slope of the straight line obtained from linear fitting Through formula Calculate activation energy ,in Boltzmann constant (value 1.380649 × 10⁻⁶) -23 J / K);

[0085] The effectiveness of the fitting results was verified, requiring a linear correlation coefficient of not less than 0.95. Outlier data points were removed and the fit was refitted.

[0086] Finally, three sets of parallel experiments were selected. The average value is used as the activation energy parameter of this type of electrolytic capacitor in the capacity decay failure mode, and is substituted into the Arrhenius model for lifetime prediction.

[0087] The data structure of the digital twin model is organized using a directed acyclic graph. Nodes in the graph represent physical or logical components, and edges represent energy, information, or physical dependencies between components. Each node stores its current state variables and historical state trajectories, and edges define the transfer functions between state variables.

[0088] In step S3, the real-time acquired multidimensional state data is fused and synchronized with the model prediction state of the digital twin model. An extended Kalman filter algorithm is used as the state estimator. The extended Kalman filter algorithm uses the system state equation of the digital twin model as the prediction model. The system state equation describes the internal state variables of the model, such as the temperature, stress, and strain of each grid node, and their evolution over time and with input stress. The system state equation is expressed as:

[0089] ;

[0090] in, For the first The state vector at time t, It is a nonlinear state transition function. For the first The input stress vector at time t, Let be the process noise vector, following a zero-mean Gaussian distribution. The algorithm uses the measurements from the multimodal sensor array as observation inputs to construct an observation equation. This observation equation relates the model state variables to the sensor measurements and is expressed as:

[0091] ;

[0092] in, For the first The observation vector at time t, For nonlinear observation functions, The observed noise vector follows a zero-mean Gaussian distribution.

[0093] At each time step, the extended Kalman filter algorithm first predicts the state of the digital twin model at the current time step based on the state at the previous time step and the current stress input. Then, it calculates the Kalman gain based on the residual between the actual sensor measurements at the current time step and the model-predicted measurements. Finally, it uses the Kalman gain to correct the predicted state, outputting the optimal posterior state estimate that integrates the model prediction and actual measurement information, i.e., the calibrated real-time digital twin state. The calculation of the Kalman gain considers the real-time estimation of the process noise covariance matrix and the observation noise covariance matrix to adapt to the dynamic changes in noise characteristics during the test.

[0094] In step S4, based on the real-time digital twin state, the health status of the tested human-machine interface is evaluated and pre-failure warning features are identified. A set of predefined health indicator vectors is extracted from the real-time digital twin state. The health indicator vectors include the cumulative creep strain energy density of key solder joints, the chromaticity drift of a specific area of ​​the display panel, the highest temperature of the central processing unit under a reference load, and the root mean square energy value of the acoustic emission signal.

[0095] The cumulative creep strain energy density is obtained by integrating the stress-strain history of the weld area, with the integration step size consistent with the finite element solution step size. The chromaticity drift is defined as the Euclidean distance between the current chromaticity coordinates and the initial chromaticity coordinates. The root mean square value of the acoustic emission signal energy is obtained by averaging the squares of the acoustic emission signal within a sliding time window and then taking the square root; the sliding window length is 10 milliseconds. The health indicator vector sequence of the tested human-computer interface during the initial health phase of the test is used to train an unsupervised anomaly detection model based on a deep convolutional autoencoder.

[0096] The deep convolutional autoencoder comprises four convolutional coding layers and four transposed convolutional decoding layers. The coding layers use 3×3 convolutional kernels with rectified linear units as the activation function, and the decoding layers use bilinear interpolation for upsampling. During testing, the real-time extracted health indicator vector is input into the trained deep convolutional autoencoder, and its reconstruction error is calculated.

[0097] When the reconstruction error exceeds a dynamic threshold set based on the 300-sigma principle, a failure precursor event is detected. The dynamic threshold is calculated based on the mean and standard deviation of the reconstruction errors of the most recent 1000 healthy state samples; the threshold is the mean plus three times the standard deviation. Simultaneously, the health indicator vector is input into a pre-trained fault classifier, such as a support vector machine (SVM), to locate the root cause of the failure precursor. The SVM uses a radial basis function kernel, and its hyperparameters are optimized on a historical failure dataset through cross-validation.

[0098] In step S5, based on the health status and failure precursor characteristics, a multi-dimensional combined stress loading profile for the next test cycle is dynamically generated through a test strategy optimization model. A reinforcement learning agent based on proximal policy optimization is used as the test strategy optimization model. The state space of the reinforcement learning agent consists of the health indicator vector, the identified failure precursor types, and the test time elapsed; the state vector has 25 dimensions.

[0099] The action space of the reinforcement learning agent is a set of discretized stress-loading primitive operations. These primitive operations include: increasing the temperature change rate of high and low temperature cycles by 5%, shifting the main peak frequency of the random vibration spectrum by 10 Hz, increasing the relative humidity level by 10 percentage points, and executing a high-load graphical user interface interaction script for 5 minutes. The action space contains 16 primitive operations, and the agent can select a single operation or a combination of up to 3 operations. The reward function of the reinforcement learning agent is defined as a composite function aimed at maximizing the damage accumulation rate of key thermal stress and vibration stress-sensitive components while penalizing test time. Its mathematical expression is the sum of the products of the damage increment of each key thermal stress and vibration stress-sensitive component and a preset criticality weight, minus the product of the test cycle duration and the time penalty coefficient.

[0100] The critical weights are pre-set based on the importance of each component to the overall vehicle safety: weld joints have a weight of 0.4, display panels have a weight of 0.3, processors have a weight of 0.2, and other components have a weight of 0.1. The time penalty coefficient is 0.05 per hour. At the end of each test cycle, the reinforcement learning agent selects or operates on a set of stress loading primitives based on the current state to form a multi-dimensional combined stress loading profile for the next test cycle. This profile is sent to the stress loading device in the form of control commands. The policy network of the reinforcement learning agent adopts a three-layer fully connected neural network, with each layer containing 256 neurons and the activation function being the hyperbolic tangent function. The value network structure is the same as the policy network.

[0101] In step S6, based on the multidimensional combined stress loading profile, the stress loading device is driven to apply composite stress to the tested human-machine interface. The stress loading device includes a programmable environmental test chamber, a multi-axis vibration table, and a robotic arm. The programmable environmental test chamber receives temperature and humidity setpoints and rate of change commands from the test strategy optimization model, precisely controlling the temperature and humidity environment within the test space. The temperature control accuracy is ±0.5 degrees Celsius, and the humidity control accuracy is ±2% relative humidity.

[0102] The multi-axis vibration table receives vibration profile commands including frequency spectrum, amplitude, and duration, and applies mechanical vibration stress to the human-machine interface under test fixed thereon. The vibration table can be independently controlled in three translational degrees of freedom and three rotational degrees of freedom, with a maximum acceleration of 20 grams. The end effector of the robotic arm is equipped with a flexible stylus with a pressure sensor, which receives screen click coordinate sequences, sliding trajectories, and contact pressure commands to simulate complex user interaction operations, applying electromechanical load stress to the display screen and touch system of the human-machine interface under test. The pressure sensor of the flexible stylus has a range of 0 to 10 Newtons and a resolution of 0.01 Newtons, and the repeatability of the robotic arm is ±0.05 mm.

[0103] In step S7, based on the real-time digital twin state and historical multidimensional state data, the remaining lifespan prediction result of the tested human-computer interaction interface is continuously updated and output. The remaining lifespan prediction result is generated by fusing the results of two different models.

[0104] The first model is a failure physics-based prediction model embedded in the digital twin model. Based on the real-time updated stress history and component degradation state, it uses analytical formulas such as the Engelmeier model or the Arrhenius model to calculate the theoretical remaining life of each key thermal stress and vibration stress-sensitive component.

[0105] The second model is a sequence data-driven prediction model based on long short-term memory networks. This model takes the time series of the health indicator vectors collected in history as input and directly predicts the remaining service life by learning the pattern of the evolution from health indicators to eventual failure in historical test data.

[0106] The Long Short-Term Memory (LSTM) network comprises two hidden layers, each containing 128 memory units. The input sequence has a length of 500 time steps, with each time step corresponding to a health metric vector. Finally, a Bayesian model averaging method is used to weight and fuse the prediction results of the two models. The weights are dynamically adjusted based on the prediction performance of each model within the most recent historical window, with the better-performing model receiving a higher weight.

[0107] The historical window length is the most recent 50 prediction periods, and prediction performance is evaluated using root mean square error. The final output after fusion is a probability density function of the remaining lifetime, which includes the expected value and a 95% confidence interval. The probability density function is generated using a Monte Carlo sampling method, with 10,000 samples taken. Each sample is a weighted random selection based on the prediction results and uncertainties of the two models.

[0108] The method described above, through a closed-loop execution of seven steps, achieves efficient, accurate, and adaptive durability testing and lifespan prediction for the human-machine interface of intelligent cockpits. Throughout the testing process, a multimodal sensor array continuously provides high-dimensional state feedback, a digital twin model acts as a virtual mirror reflecting the internal state of the physical entity in real time, a reinforcement learning engine dynamically adjusts the test stress based on the health status, and finally, a high-confidence lifespan assessment is provided through multi-model fusion prediction. This method significantly reduces the time required for traditional accelerated testing while improving the accuracy of lifespan prediction, providing a novel technical path for the reliability verification of intelligent cockpit products.

Claims

1. A method for accelerated loading testing of multidimensional durability stress in intelligent cockpit HMIs, characterized in that, include: A multimodal sensor array deployed on the hardware of the human-computer interaction interface under test is used to collect multidimensional state data of the human-computer interaction interface under test in real time under stress loading. Construct and run a digital twin model that maps to the tested human-computer interaction interface entity. The digital twin model includes a preset component-level finite element model, a system-level performance model, and a failure physics model. The component-level finite element model establishes a refined three-dimensional finite element model for the key thermal stress and vibration stress sensitive components in the tested human-computer interaction interface. The key thermal stress and vibration stress sensitive components include the solder joints of the ball grid array packaged chip and the connection between the glass substrate of the display panel and the flexible circuit board. The component-level finite element model uses eight-node hexahedral elements for meshing, with a mesh size of no more than 0.2 mm in critical regions and no more than 1.5 mm in non-critical regions. The component-level finite element model includes the nonlinear constitutive relation of the material, thermal conductivity, thermal expansion coefficient, and elastic modulus as a function of temperature. The solder joint material is a tin-silver-copper alloy, and the yield strength decreases exponentially with increasing temperature. The system-level performance model simulates the performance behavior of the tested human-computer interaction interface operating system. The model simulates the processor and memory behavior under different software loads through a task scheduling queue and a resource occupancy function. The task scheduling queue adopts a priority preemptive scheduling algorithm, and the resource occupancy function is obtained by fitting the actual measured processor load and memory bandwidth consumption curves. The failure physics model integrates multiple analytical models based on failure physics, including the Engelmeier model for calculating the fatigue life of solder joints, the Arrhenius model for the capacitance decay of electrolytic capacitors, and a voltage-time dependent model for the degradation of liquid crystal pixels. The component-level finite element model, performance simulation model, and failure physics model are integrated into a unified data framework to form a coupled digital twin that receives external stress input and simulates the output state response. The data structure of the digital twin is organized using a directed acyclic graph, where nodes represent physical or logical components and edges represent energy, information, or physical dependencies between components. The real-time collected multidimensional state data is fused and synchronized with the model prediction state of the digital twin model. An anomaly detection model based on unsupervised learning is used to analyze the synchronized digital twin model state data to identify and quantify the early failure precursor characteristics of key thermal stress and vibration stress sensitive components. Based on the real-time digital twin status, the health status of the tested human-computer interaction interface is assessed and the characteristics of failure precursors are identified. Extract a set of predefined health indicator vectors from the real-time digital twin state. The health indicator vectors include the cumulative creep strain energy density of key solder joints, the chromaticity drift of the display panel area, the highest temperature of the central processing unit under the reference load, and the root mean square value of the acoustic emission signal energy. Based on the aforementioned health status and failure precursor characteristics, a multi-dimensional combined stress loading profile for the next test cycle is dynamically generated through a test strategy optimization model, including: A reinforcement learning agent based on proximal policy optimization is used as the test policy optimization model; The state space of the reinforcement learning agent consists of the health indicator vector, the identified failure precursor types, and the test time. The action space of the reinforcement learning agent is a set of discretized stress loading primitive operations, which include: increasing the temperature change rate of the high and low temperature cycle by 5%, shifting the main peak frequency of the random vibration spectrum by 10 Hz, increasing the relative humidity level by 10 percentage points, and executing a high-load graphical user interface interaction script for 5 minutes. The reward function of the reinforcement learning agent is defined as the sum of the product of the damage increment of each key thermal stress and vibration stress sensitive component and the preset key weight, and then the product of the test period duration and the time penalty coefficient. At the end of each test cycle, the reinforcement learning agent selects or performs a set of stress loading primitive operations based on the current state, and combines them into a multi-dimensional combined stress loading profile for the next test cycle. Based on the multidimensional combined stress loading profile, the stress loading device is driven to apply composite stress to the human-machine interface under test. Based on the real-time digital twin status and historical multidimensional status data, the remaining lifespan prediction results of the tested human-computer interaction interface are continuously updated and output.

2. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 1, characterized in that, A multimodal sensor array deployed on the tested human-machine interface hardware is used to collect multidimensional state data of the tested human-machine interface under stress loading in real time, including: Temperature distribution data is collected by multiple thermocouple sensors attached to the main printed circuit board, display module frame, and key chip packaging surface of the human-machine interface under test. The ambient temperature and humidity data are obtained by using temperature and humidity sensors installed in the test environment chamber; Vibration response data of the human-machine interface under test is collected on the vibration table by a triaxial accelerometer fixed to the mounting bracket of the human-machine interface under test. The software agent program that communicates with the main processor of the human-machine interface under test can obtain the load rate, core temperature, memory usage and data bus transmission error rate of the central processing unit and the graphics processing unit in real time. Using an industrial camera and colorimeter facing the human-machine interface display screen under test, periodically collect image data of the display screen's brightness uniformity, chromaticity coordinates, contrast ratio, and pixel defects. By monitoring the reported coordinates and response time output by the touch screen controller, the response delay and accuracy drift data of the touch function can be obtained. Multiple piezoelectric ceramic acoustic emission sensors fixed to the main printed circuit board and display panel structure are used to collect ultrasonic signals released when the structure generates microcracks or delamination under stress. The detection frequency range of the acoustic emission sensors is 50 kHz to 500 kHz, and the signals are digitized by a data acquisition card with a sampling rate of not less than 5 megasamples per second.

3. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 2, characterized in that, The process of fusing and synchronizing the real-time acquired multidimensional state data with the model prediction state of the digital twin model includes: An extended Kalman filter algorithm is used as the state estimator; The algorithm uses the system state equation of the digital twin model as the prediction model, and the system state equation describes the evolution of the internal state variables of the model with time and input stress. The algorithm uses the measurements from the multimodal sensor array as observation inputs to construct observation equations, which relate the model state variables to the sensor measurements. At each time step, the extended Kalman filter algorithm first predicts the state of the digital twin model at the current time step based on the state at the previous time step and the current stress input. Subsequently, the Kalman gain is calculated based on the residual between the actual sensor measurement at the current moment and the measurement predicted by the model. Finally, the predicted state is corrected using the Kalman gain, and the optimal posterior state estimate, which integrates the model prediction and actual measurement information, is output, i.e., the calibrated real-time digital twin state.

4. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 3, characterized in that, Based on the real-time digital twin status, the health status of the tested human-computer interaction interface is assessed and pre-failure warning signs are identified, including: The health indicator vector sequence of the human-computer interaction interface under test in the initial health stage of the test is used to train an unsupervised anomaly detection model based on a deep convolutional autoencoder. During the testing process, the real-time extracted health indicator vectors are input into the trained deep convolutional autoencoder to calculate its reconstruction error. When the reconstruction error exceeds the dynamic threshold set based on the 300 sigma principle, it is determined that a failure precursor event has been detected. Simultaneously, the health indicator vector is input into a pre-trained fault classifier to locate the root cause component of the failure precursor.

5. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 4, characterized in that, The cumulative creep strain energy density is obtained by integrating the stress-strain history of the weld joint region; The chromaticity drift is defined as the Euclidean distance between the current chromaticity coordinates and the initial chromaticity coordinates; The root mean square value of the energy of the acoustic emission signal is obtained by averaging the square values ​​of the acoustic emission signal within a sliding time window and then taking the square root. The sliding window length is 10 milliseconds. The dynamic threshold is calculated based on the mean and standard deviation of the reconstruction error of the most recent 1000 health status samples, and the threshold is the mean plus 3 times the standard deviation. The fault classifier is a support vector machine with a radial basis function kernel, and its hyperparameters are optimized on a historical failure dataset through cross-validation.

6. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 5, characterized in that, The critical weights are pre-set based on the importance of the components to the overall vehicle safety: weld points have a weight of 0.4, display panels have a weight of 0.3, processors have a weight of 0.2, and other components have a weight of 0.

1. The time penalty coefficient is 0.05 per hour; The policy network of the reinforcement learning agent adopts a three-layer fully connected neural network, with each layer containing 256 neurons and the activation function being the hyperbolic tangent function.

7. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 6, characterized in that, Based on the multidimensional combined stress loading profile, the stress loading device is driven to apply composite stress to the tested human-machine interface, including: The stress loading device includes a programmable environmental test chamber, a multi-axis vibration table, and a robotic arm. The programmable environment test chamber receives temperature and humidity setpoints and change rate instructions from the test strategy optimization model, and precisely controls the temperature and humidity environment in the test space. The multi-axis vibration table receives vibration profile commands containing frequency spectrum, amplitude, and duration, and applies mechanical vibration stress to the human-machine interface under test fixed thereon. The robotic arm is equipped with a flexible stylus with a pressure sensor at its end. It receives screen click coordinate sequences, sliding trajectories, and contact pressure commands to simulate complex user interaction operations and apply electromechanical load stress to the display screen and touch system of the human-computer interaction interface under test.

8. The method for accelerated loading test of multidimensional durability stress in intelligent cockpit HMI according to claim 7, characterized in that, The programmable environmental test chamber has a temperature control accuracy of ±0.5 degrees Celsius and a humidity control accuracy of ±2% relative humidity. The multi-axis vibration table is independently controlled in 3 translational degrees of freedom and 3 rotational degrees of freedom, with a maximum acceleration of 20 grams; The pressure sensor of the flexible stylus has a range of 0 to 10 Newtons and a resolution of 0.01 Newtons. The repeatability of the robotic arm is ±0.05 millimeters.

Images