Adaptive fatigue life prediction method for commercial vehicle battery rack facing random vibration

Abstract

An adaptive fatigue life prediction method for commercial vehicle battery racks under random vibration is proposed, relating to the field of battery rack fatigue prediction technology. The adaptive fatigue life prediction method includes: collecting multi-source real-vehicle data on the commercial vehicle battery rack and vehicle operating status, preprocessing it into standardized time-series data; extracting time-domain, frequency-domain, and time-frequency-domain features of road excitation to form a working condition classification feature vector; identifying the current road working condition through a trained time-series classification model and matching it with a corresponding fatigue constitutive model; performing equivalent mapping on cumulative damage during model switching to ensure continuous damage accumulation across working conditions; utilizing a physical information neural network embedded with fatigue constitutive physical constraints, combined with fatigue test and simulation data to identify parameters and uncertainties; constructing an on-board digital twin to assimilate the real-vehicle sensor response, updating loads, constitutive parameters, key node stresses, and cumulative damage, calculating fatigue damage increments, and predicting remaining life.

Description

Technical Field

[0001] This invention relates to the field of battery rack fatigue prediction technology, and more specifically, to an adaptive fatigue life prediction method for commercial vehicle battery racks resistant to random vibration. Background Technology

[0002] As a core component supporting the power batteries of new energy vehicles, commercial vehicle battery racks are widely used in battery swapping systems of various commercial vehicles. When driving on different road conditions such as urban roads, highways, and unpaved roads, the battery racks are subjected to complex and unstable random vibration loads. Long-term service can easily lead to fatigue accumulation damage or even fracture failure in critical parts such as welds and mounting points, seriously affecting driving safety. Therefore, accurate fatigue life prediction of battery racks has become a core requirement for ensuring vehicle operation safety.

[0003] To address this need, existing technologies have proposed several solutions. Some solutions employ a hierarchical design approach combining multi-axis response diagnosis and direction-specific optimization to conduct random vibration response analysis, failure mode identification, and targeted optimization. Other solutions utilize digital twin technology and deep neural network models, constructing virtual entities by collecting physical data in real time through sensors, and integrating physical and virtual data to achieve structural reliability monitoring and early warning. Still other solutions integrate visual and frequency domain data to construct a five-dimensional twin dataset, combining Miller's criterion and a BP neural network model to predict fatigue life. Yet another solution collects measured vibration acceleration and stress data of the battery frame under reinforced road surfaces in a test field, solves the system's inverse transfer function using white noise excitation, obtains the battery frame excitation load based on a dynamic model through virtual iteration, and completes fatigue durability performance analysis by combining the material stress-life curve.

[0004] Existing solutions generally suffer from common shortcomings. Fatigue constitutive models are simplistic and struggle to address the uncertainties of high- and low-cycle fatigue under varying operating conditions, resulting in significant limitations in understanding and selection. Methods for quantifying fatigue uncertainty employ purely data-driven neural network models, but the model training process lacks the integration of physical mechanisms, leading to low prediction accuracy and poor physical consistency. Summary of the Invention

[0005] This invention provides an adaptive fatigue life prediction method for commercial vehicle battery racks subjected to random vibration, in order to improve at least one of the aforementioned technical problems.

[0006] This invention provides an adaptive fatigue life prediction method for commercial vehicle battery racks oriented to random vibration, comprising S1 to S7.

[0007] S1. Collect multi-source real vehicle data on the commercial vehicle battery rack and vehicle operating status, and preprocess the multi-source real vehicle data to obtain standardized time-series data.

[0008] S2. Based on the standardized time-series data, extract the time-domain features, frequency-domain features, and time-frequency-domain features of the road surface excitation to form a working condition classification input feature vector.

[0009] S3. Input the working condition classification feature vector into the trained time series classification model to determine the current road working condition, and match the corresponding fatigue constitutive model in the fatigue constitutive model library according to the current road working condition.

[0010] S4. When the fatigue constitutive model is switched, the cumulative damage before and after the switch is equivalently mapped to maintain the continuity of fatigue damage accumulation across working conditions.

[0011] S5. Using a physical information neural network embedded with fatigue constitutive physical constraints, fatigue constitutive parameters and their uncertainties are identified based on fatigue test data and simulation constraint data.

[0012] S6. Construct an on-board digital twin of the commercial vehicle battery rack and assimilate the real vehicle sensor response to the on-board digital twin to update the load boundary, constitutive parameters, key node stress, and cumulative damage.

[0013] S7. Calculate the fatigue damage increment based on the updated fatigue constitutive model and key node responses, and predict the remaining service life.

[0014] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0015] This invention effectively solves the problem of large prediction errors across working conditions using traditional single models by constructing an adaptive discrimination mechanism for road conditions and an intelligent matching mechanism for fatigue constitutive models. Combined with a unique damage equivalent mapping technique, it achieves seamless fusion and continuous damage accumulation of three different physical mechanism models: Basquin stress-life, Manson-Coffin strain-life, and Chaboche nonlinear kinematic hardening. This fundamentally avoids damage jumps caused by model switching, significantly improving the accuracy of cross-working-condition predictions. Furthermore, this invention introduces a Physical Information Neural Network (PINN) to construct a data-physics hybrid modeling framework. Only 5 to 10 sets of small-sample fatigue test data are needed to achieve high-precision constitutive parameter identification and decoupling of dual uncertainties, greatly reducing test costs and time. By combining a three-layer coupling of geometry, physics, and data in an in-vehicle digital twin and dynamic data assimilation technology, it can assimilate real-vehicle sensor data in real time, update load boundaries and damage status online, and consider load sequence effects based on the improved Miner's rule. It outputs a comprehensive health assessment result that includes point estimation, 95% confidence interval, weak point location, and three-level early warning. This provides technical support for the full life cycle safety management of commercial vehicle battery racks that is far more accurate, reliable, and physically interpretable than existing technologies. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the adaptive fatigue life prediction method.

[0017] Figure 2 This is a flowchart of the three-layer digital twin construction process for the vehicle battery rack.

[0018] Figure 3 This is a flowchart of multi-source sensor data preprocessing and time-frequency domain feature extraction.

[0019] Figure 4 This is a flowchart of road condition identification and adaptive fatigue constitutive matching.

[0020] Figure 5 This is a flowchart of PINN's physical modeling, digital twin construction, and dynamic data assimilation process.

[0021] Figure 6 This is a flowchart for fatigue damage calculation and remaining life prediction. Detailed Implementation

[0022] Please see Figures 1 to 6 This invention provides an adaptive fatigue life prediction method for commercial vehicle battery racks subjected to random vibration, achieving high-precision, adaptive, and dynamic life prediction through steps S1 to S7: First, the system automatically identifies driving conditions based on the time-frequency domain characteristics of road excitation and adaptively matches the optimal fatigue constitutive model according to the condition category, solving the problem of insufficient prediction accuracy of traditional single models across different conditions. Second, a data-physics hybrid model based on Physical Information Neural Network (PINN) is constructed to complete fatigue constitutive parameter identification and uncertainty decoupling under small sample test conditions, significantly reducing dependence on fatigue test data. Third, a digital twin of the vehicle battery rack is constructed, and through real-time assimilation of real vehicle sensor data, dynamic closed-loop updates of fatigue damage accumulation and remaining service life are achieved, ensuring synchronization between the model and the physical entity. Finally, an engineering-deployable online fatigue life prediction method is formed, providing full life-cycle technical support for battery rack structure design optimization, durability verification, and vehicle health early warning.

[0023] The overall process of this method is as follows: Figure 1 As shown, it includes the following nine core components in sequence: real vehicle data acquisition, road excitation feature extraction, working condition adaptive discrimination, fatigue constitutive model intelligent matching, data-physical hybrid modeling based on PINN, construction of digital twin of vehicle battery rack, dynamic data assimilation and model iterative update, fatigue damage calculation, and remaining life prediction and output.

[0024] Figure 3This is a flowchart of multi-source sensor data preprocessing and time-frequency domain feature extraction, demonstrating the complete chain from multi-source sensor signal acquisition, wavelet denoising, resampling, outlier removal, to time-domain, frequency-domain, and time-frequency-domain feature extraction, providing standardized feature vectors for operational condition identification. This corresponds to steps S1 and S2.

[0025] S1. Collect multi-source real-vehicle data on the commercial vehicle battery rack and vehicle operating status, and preprocess the multi-source real-vehicle data to obtain standardized time-series data. Preferably, S1 specifically includes S11 to S14.

[0026] S11. Triaxial accelerometers and strain sensors are installed in the battery pack housing, key load-bearing nodes of the battery rack, and the connection between the vehicle frame and the suspension to collect road excitation vibration signals and battery rack structural strain response signals, and obtain vehicle speed, load status and positioning information through the vehicle bus.

[0027] Specifically, triaxial accelerometers and strain sensors are placed in the battery pack housing, key load-bearing nodes of the battery rack, and the connection between the vehicle frame and suspension to collect road excitation vibration signals and battery rack structural strain response signals in real time. At the same time, auxiliary information such as vehicle speed, load status, and GPS location are obtained through the vehicle's CAN bus.

[0028] S12. The original signal matrix of the multi-source sensing is represented as follows: In the formula: This is the original signal matrix of the multi-source sensing. This refers to the continuous sampling time. For the first The time-domain signal collected by the road sensor. For sensor channel indexes. This represents the number of sensors. This represents the matrix transpose symbol.

[0029] S13, the continuous sampling time satisfies: In the formula: This represents the total sampling time.

[0030] S14. Perform wavelet threshold denoising, unified resampling, and outlier removal on the original multi-source sensor signal matrix, and slice it according to a fixed time window to obtain the standardized time-series data.

[0031] Specifically, the acquired raw signals undergo preprocessing operations such as wavelet threshold denoising, unified resampling to a 1kHz time base, and outlier removal based on the local outlier factor algorithm to form a standardized time series dataset.

[0032] S2. Based on the standardized time-series data, extract the time-domain features, frequency-domain features, and time-frequency-domain features of the road surface excitation to form a working condition classification input feature vector. Preferably, S2 specifically includes S21 to S22.

[0033] S21. Within each time window, using the vertical acceleration sequence at the suspension point as the primary signal, extract multi-dimensional feature vectors, including time-domain and frequency-domain features. Definition: For the preprocessed first One vertical acceleration sample value. This is the index of the sampling points within the time window.

[0034] Temporal features include root mean square, mean, variance, peak factor, and kurtosis.

[0035] The root mean square is: .

[0036] In the formula: The root mean square of the signal within the time window. This represents the number of sampling points within the time window. This is the index of the sampling points within the time window. For the preprocessed first Each vibration signal sample value.

[0037] The mean is expressed as: .

[0038] In the formula: This represents the average signal value within the time window. This represents the number of sampling points within the time window. This is the index of the sampling points within the time window. For the preprocessed first Each vibration signal sample value.

[0039] Variance is expressed as: .

[0040] In the formula: This represents the signal variance within the time window. This represents the number of sampling points within the time window. This is the index of the sampling points within the time window. For the preprocessed first Each vibration signal sample value. This represents the average signal value within the time window.

[0041] The peak factor is: .

[0042] In the formula: This is the peak factor. For indexing sampling points within the time window The operator that retrieves the maximum value.

[0043] The kurtosis is: .

[0044] In the formula: This represents the signal kurtosis within the time window. This represents the average signal value within the time window. This represents the signal variance within the time window.

[0045] The frequency domain characteristics include acceleration power spectral density, band energy ratio, peak frequency, and wavelet packet energy characteristics. The acceleration power spectral density is calculated using the Welch method.

[0046] The acceleration power spectral density is: .

[0047] In the formula: This represents the acceleration power spectral density. For frequency variables. The duration of the signal in the current analysis window. The sampling frequency. This is the Fourier transform of the acceleration signal.

[0048] Extracting the main frequency band position Based on the frequency band energy distribution ratio, the frequency is divided into three bands: 0Hz to 10Hz, 10Hz to 50Hz, and 50Hz to 200Hz. The proportion of energy in each band to the total energy is calculated. It is the lower limit frequency of the main frequency band. This is the upper limit frequency of the main frequency band.

[0049] The energy ratio of the first frequency band is: .

[0050] The energy ratio of the second frequency band is: .

[0051] The energy ratio of the third frequency band is: .

[0052] In the formula: This refers to the proportion of energy in the 0Hz to 10Hz frequency band to the total energy in the 0Hz to 200Hz frequency band. This refers to the proportion of energy in the 10Hz to 50Hz frequency band to the total energy in the 0Hz to 200Hz frequency band. This refers to the proportion of energy in the 50Hz to 200Hz frequency band to the total energy in the 0Hz to 200Hz frequency band.

[0053] The peak frequency is: .

[0054] In the formula: This is the peak frequency at which the power spectral density reaches its maximum value. For frequency variables The operator that takes the maximum value corresponds to the argument variable. This represents the acceleration power spectral density.

[0055] In the time-frequency domain, wavelet packet decomposition energy coefficients for each frequency band are obtained through wavelet packet decomposition, and the signal is then processed. Layer wavelet packet decomposition yields the th Wavelet packet energy of each frequency band. This represents the number of wavelet packet decomposition layers.

[0056] The wavelet packet energy characteristic is: .

[0057] In the formula: For the first Wavelet packet energy of each frequency band. This is the wavelet packet frequency band index. This is the wavelet packet coefficient index. For the first Number of wavelet packet coefficients within each frequency band. For the first Within each frequency band Wavelet packet coefficients.

[0058] The marginal spectrum features are obtained by combining the Hilbert-Huang transform, and the marginal spectrum is obtained after empirical mode decomposition and Hilbert transform. , characterizing the time-frequency distribution of non-stationary excitation.

[0059] S22. The time-domain features, frequency-domain features, and time-frequency-domain features are concatenated to obtain the working condition classification input feature vector.

[0060] All features are combined and concatenated to form the input feature vector for working condition classification.

[0061] .

[0062] In the formula: Input feature vectors for operating condition classification. The root mean square of the signal within the time window. This is the peak factor. This represents the signal kurtosis within the time window. This represents the energy ratio of the first frequency band. This represents the energy ratio of the second frequency band. This represents the energy ratio of the third frequency band. This is the peak frequency. This represents the matrix transpose symbol.

[0063] Figure 4 The flowchart of road condition identification and adaptive fatigue constitutive matching is the core process of the condition-model adaptation of this invention. It shows the complete mechanism of CNN-LSTM condition classification, confidence judgment, temporal voting fusion, as well as matching Basquin model, Manson-Coffin model, Chaboche model according to condition, model switching judgment, damage equivalence mapping, and cumulative damage continuity guarantee.

[0064] S3. Input the working condition classification feature vector into the trained time series classification model to determine the current road working condition, and match the corresponding fatigue constitutive model in the fatigue constitutive model library according to the current road working condition.

[0065] Preferably, S3 includes S31 to S35.

[0066] S31. Input the working condition classification feature vector into the CNN-LSTM classification model, extract the local time-frequency pattern through the convolutional layer, extract the temporal dependency through the LSTM layer, and output the posterior probability of the current time window belonging to different road working conditions through Softmax.

[0067] Convolutional layers extract local time-frequency features: In the formula: This is the output feature vector of the convolutional layer. This is the activation function for the neural network. This is the weight matrix of the convolutional layer. This is the convolution operator. Input feature vectors for operating condition classification. This is the bias vector for the convolutional layer.

[0068] Modeling temporal dependencies using LSTM layers: In the formula: For the first The hidden state vector of LSTM at each time step. This is the LSTM timing step index. This is the mapping function for Long Short-Term Memory (LSTM) networks. For the first The convolutional layer outputs feature vectors for each time step. For the first The hidden state vector of LSTM at each time step.

[0069] The classification score vector output by the fully connected layer is represented as follows: In the formula: This is the classification score vector. Output the weight matrix for the fully connected layer. Output bias vector for fully connected layers.

[0070] No. The Softmax probability of the type of working condition is: In the formula: For the first Posterior probability of a type of working condition. This is an index for operating condition categories. It is a natural exponential function. For the classification score vector, the first... The score for the type of working condition. Use the index for Softmax normalized summation. For the classification score vector, the first... The score for the type of working condition.

[0071] Preferably, the working condition category index is taken as: ,in, Indicates a smooth urban road surface. Indicates a regular highway surface. It indicates an off-road or bumpy road surface.

[0072] Specifically, the extracted feature vectors are input into a pre-trained CNN-LSTM classification model. Convolutional layers are used to extract local time-frequency patterns, LSTM layers are used to model temporal dependencies, and Softmax normalization is used to output the probability distribution of the road condition to which the current time window belongs. The classification results divide the road conditions into three categories: urban smooth road surface, highway conventional road surface, and off-road or bumpy road surface.

[0073] Operating condition category 1 is urban smooth road surface, characterized by low root mean square, kurtosis close to 3, dominant frequency greater than 20Hz, and vibration characteristics of low amplitude, high frequency, and dense cycle.

[0074] Operating condition category 2 is a conventional highway road surface, characterized by moderate root mean square, stable peak factor, and dominant frequency between 5Hz and 20Hz. The vibration characteristics are medium amplitude, medium frequency, and relatively stable load.

[0075] Operating condition category 3 is for off-road or bumpy roads, characterized by a high root mean square, kurtosis greater than 5, energy concentrated in low-frequency impacts, and vibration characteristics of high-amplitude impacts, non-stationary, and significant multiaxial stress.

[0076] S32. The maximum posterior probability is used as the model output confidence level, and the current road surface condition is determined based on the model output confidence level. Specifically, when the model output confidence level is greater than or equal to a preset confidence threshold, the road surface condition with the maximum posterior probability corresponding to the current time window is used as the current road surface condition. When the model output confidence level is less than the preset confidence threshold, time-series voting is performed based on the classification results of multiple consecutive time windows, and the voting result is used as the current road surface condition.

[0077] Preferably, the confidence level of the model output is: In the formula: Output confidence scores for the model. This is the maximum value operator. This represents the posterior probability of a smooth urban road surface. Let be the posterior probability of a conventional highway surface. For off-road or bumpy road surfaces, the posterior probability is given.

[0078] The classification model simultaneously outputs the probability distribution of the operating condition in the current window, and takes the highest probability value as the confidence level. When the confidence level is greater than or equal to 0.7, the system directly adopts the current classification result. When the confidence level is less than 0.7, the system enters the time-series fusion decision mode, taking the classification results of the current window and the two windows before and after it, for a total of 5 consecutive windows, and conducting a majority vote to determine the final operating condition category. If the voting result is a tie, the operating condition category of the previous time window is retained to ensure the continuity and robustness of operating condition discrimination and avoid model switching oscillations caused by frequent false triggers.

[0079] Then, based on the road condition assessment results, the system automatically matches a fatigue constitutive model and damage calculation method suitable for the fatigue mechanism of the condition from the built-in model library: the Basquin stress-life model is matched for smooth urban pavements, and the linear Miner cumulative damage method is used. The Manson-Coffin strain-life model is matched for conventional highway pavements, and the strain-life method is used for damage calculation. The Chaboche nonlinear kinematic hardening constitutive model is matched for off-road or bumpy pavements, combined with the critical plane method, to achieve fatigue damage characterization under multiaxial non-proportional loading.

[0080] S33. When the current road surface condition is a smooth urban road surface, the Basquin stress-life model is matched. The fatigue mechanism of a smooth urban road surface is low-amplitude, high-cycle fatigue.

[0081] The Basquin stress-life model is as follows: .

[0082] In the formula: This is the equivalent stress amplitude. The fatigue index is for the Basquin model. This represents the fatigue life cycle count. These are the material constants for the Basquin model.

[0083] S34. When the current road surface condition is a conventional high-speed road surface, the Manson-Coffin strain-life model is applied. The fatigue mechanism of the conventional high-speed road surface is medium-amplitude elastoplastic fatigue.

[0084] The Manson-Coffin strain lifetime model is as follows: .

[0085] In the formula: This represents the total strain amplitude. This is the fatigue strength coefficient. This is the elastic modulus of the material. The fatigue strength index is the Manson-Coffin model. This is the fatigue ductility coefficient. The fatigue ductility index is the Manson-Coffin model index.

[0086] S35. When the current road surface condition is off-road or bumpy, the Chaboche nonlinear kinematic hardening constitutive model is matched and combined with the critical plane method. The fatigue mechanism of the off-road or bumpy road surface is multiaxial non-proportional fatigue.

[0087] The Chaboche nonlinear kinematic hardening constitutive model is: .

[0088] In the formula: Let be the rate of change of the back stress tensor in the Chaboche model. This represents the kinematic hardening modulus of the Chaboche model. To dynamically recover parameters for the Chaboche model. is the rate of change of the plastic strain tensor. Let be the back stress tensor of the Chaboche model.

[0089] S4. When the fatigue constitutive model is switched, the cumulative damage before and after the switch is equivalently mapped to maintain the continuity of fatigue damage accumulation across working conditions. Preferably, when the number of consecutive time windows of the target working condition meets the model switching condition, the fatigue constitutive model is switched, and the cumulative damage before and after the switch is kept consistent through damage equivalent mapping.

[0090] The system has built-in prior distributions of parameters for each model, and the current cumulative damage value is used when switching. Based on this, a damage equivalence mapping mechanism is employed to handle the transition of cumulative damage: when switching from the uniaxial model to the Chaboche multiaxial model, the equivalent cumulative plastic strain amplitude is inversely calculated using the damage-plastic strain inverse function. When switching back from the Chaboche multiaxial model to the uniaxial model, the multiaxial damage is converted into an equivalent uniaxial stress / strain amplitude based on the maximum damage plane on the critical plane, ensuring the cumulative damage value before and after the switch is accurate. In the formula: The initial cumulative damage value after model switching, compared with the value before switching. Complete consistency ensures physical uniformity in damage accumulation.

[0091] Model switching is triggered by both the work condition classification result and the duration. Let the number of consecutive windows for the target work condition be . ,when Model switching is performed every 30 seconds to avoid frequent switching caused by momentary interference.

[0092] As an exception, when the kurtosis of the vibration signal within the window... or peak factor When a significant impact load is detected, even if the number of consecutive windows for the target working condition does not reach 3, the Chaboche model is temporarily used for damage calculation in the current window to accurately assess the fatigue damage caused by short-term impacts. After the calculation is completed, the model is automatically rolled back to the original model. This represents the number of consecutive windows that represent the target operating condition. This represents the signal kurtosis within the time window. This is the peak factor.

[0093] When switching, the cumulative damage value before model switching is used. Based on this, a damage equivalence mapping mechanism is used to handle cumulative damage transitions, ensuring that the cumulative damage value remains consistent before and after the switchover. In the formula: This represents the cumulative damage value before model switching. This represents the initial cumulative damage value after model switching.

[0094] Based on the cumulative damage value before model switching Based on this, the damage equivalence conversion between the single-axis model and the multi-axis model is realized, and the core mapping relationship is as follows.

[0095] Mapping relationship 1: Single-axis model → Multi-axis model (Basquin / Manson-Coffin → Chaboche).

[0096] by Inverse calculation of the equivalent cumulative plastic strain amplitude under the Chaboche model : .

[0097] .

[0098] In the formula: This represents the cumulative damage value before model switching. This is the mapping function from plastic strain amplitude to damage value under the Chaboche model. This represents the equivalent cumulative plastic strain amplitude. for The inverse mapping function.

[0099] Mapping Relationship 2: Multi-axis model → Single-axis model (Chaboche → Basquin / Manson-Coffin).

[0100] The maximum damage plane is determined based on the critical plane method, and multiaxial damage is converted into an equivalent uniaxial stress amplitude. or equivalent total strain amplitude .

[0101] When switching to the Basquin model, the equivalent fatigue life cycle number is first obtained from the cumulative damage. Then, the equivalent uniaxial stress amplitude is calculated back using the Basquin model. : .

[0102] .

[0103] In the formula: The number of uniaxial fatigue life cycles is obtained from the cumulative damage equivalent. This represents the cumulative damage value before model switching. This is the equivalent uniaxial stress amplitude. These are the material constants for the Basquin model. The fatigue index is for the Basquin model.

[0104] When switching to the Manson-Coffin model, the equivalent total strain amplitude is further calculated from the strain-life equation. : .

[0105] In the formula: This is the equivalent total strain amplitude. This is the fatigue strength coefficient. This is the elastic modulus of the material. The number of uniaxial fatigue life cycles is obtained from the cumulative damage equivalent. The fatigue strength index is the Manson-Coffin model. This is the fatigue ductility coefficient. The fatigue ductility index is the Manson-Coffin model index.

[0106] Through the above equivalent transformation, a continuous mapping from multi-axis damage to uni-axis damage is achieved, ensuring that the cumulative damage value after model switching is consistent with the cumulative damage value before switching, thereby maintaining the physical continuity of the fatigue damage accumulation process.

[0107] It is particularly important to note that the aforementioned damage equivalence mapping mechanism is not only a numerical connection method when switching between different constitutive models, but also the core enabling technology for achieving adaptive fatigue life prediction under different operating conditions in this invention. This mechanism solves the fundamental problem of cross-condition fatigue analysis in the prior art, namely, the inconsistency in the dimensions and physical meaning of damage accumulation under different fatigue mechanisms.

[0108] Through this mechanism, stress amplitude damage under the Basquin model, strain amplitude damage under the Manson-Coffin model, and plastic strain energy damage under the Chaboche model can be seamlessly connected and continuously accumulated under the same accumulation framework, so that the three originally independent fatigue constitutive models can be truly integrated into an organic whole.

[0109] In other words, without this mapping mechanism, condition discrimination and constitutive model switching would degenerate into simple "segmented prediction," failing to form a unified life assessment system. It is precisely the existence of this mechanism that enables the adaptive method of this invention to possess a substantial synergistic effect that transcends "simple combinations of existing technologies."

[0110] Figure 5 The flowchart of PINN physical modeling, digital twin construction, and dynamic data assimilation is the core architecture of the digital twin and virtual-real synchronization of this invention. It shows the closed-loop process of small-sample PINN parameter identification, physical constraint embedding, uncertainty decoupling, geometric twin, physical twin, data twin construction, as well as UKF or PF dynamic data assimilation, and real-time updates of model parameters, damage, and stress. This corresponds to steps S5 and S6.

[0111] S5. Using a physical information neural network embedded with fatigue constitutive physical constraints, fatigue constitutive parameters and their uncertainties are identified based on fatigue test data and simulation constraint data. Preferably, S5 includes S51 to S52.

[0112] S51. Construct a physical information neural network, embed the currently matched fatigue constitutive model as a physical constraint into the loss function, and use fatigue test data as the main driver and simulation constraint data as an auxiliary constraint to output the posterior probability distribution of fatigue constitutive parameters.

[0113] Specifically, a physical information neural network is constructed, with the fatigue constitutive equation directly embedded into the network loss function as a physical constraint. Driven primarily by 5 to 10 sets of small-sample fatigue test data, the network outputs the posterior probability distribution of constitutive parameters. Through joint training of data-driven and physical constraints, this network effectively decouples two types of uncertainty: accidental uncertainty stemming from material dispersion and processing technology discreteness, and cognitive uncertainty arising from model form bias and parameter identification errors. The final result is a fatigue constitutive model considering both uncertainties, significantly improving the accuracy and reliability of constitutive parameter identification under small-sample conditions.

[0114] The input vector of the physical information neural network is: .

[0115] In the formula: This is the PINN input vector. This is the equivalent stress amplitude. This represents the total strain amplitude. It is a multiaxial stress ratio. This is the elastic modulus of the material. This represents the yield strength of the material. For operating conditions. This represents the matrix transpose symbol. The operating conditions labels include urban, highway, and off-road.

[0116] The parameter vector of the Basquin model is: .

[0117] In the formula: This is the parameter vector for the Basquin model. These are the material constants for the Basquin model. The fatigue index is for the Basquin model. This represents the matrix transpose symbol.

[0118] The parameter vector of the Manson-Coffin model is: .

[0119] In the formula: This is the parameter vector for the Manson-Coffin model. This is the fatigue strength coefficient. This is the fatigue ductility coefficient. The fatigue strength index is the Manson-Coffin model. The fatigue ductility index is the Manson-Coffin model index. This represents the matrix transpose symbol.

[0120] The parameter vector of the Chaboche model is: .

[0121] In the formula: This is the parameter vector for the Chaboche model. This represents the kinematic hardening modulus of the Chaboche model. To dynamically recover parameters for the Chaboche model. This represents the isotropic hardening saturation value of the Chaboche model. These are the isotropic hardening rate parameters for the Chaboche model. This represents the matrix transpose symbol.

[0122] The network structure uses 4 to 6 fully connected layers, with 64 to 128 neurons in each layer. Bayesian layers or MCDropout are used to distribute the parameters. A physical constraint layer is added after the output layer.

[0123] The physical constraint loss of the Basquin model is: .

[0124] In the formula: This represents the physical constraint loss for the Basquin model. This represents the number of samples tested in the fatigue test. This is an index for fatigue test samples. For the first The fatigue life was measured in the group test. These are the material constants for the Basquin model. For the first The equivalent stress amplitude of the group of experiments. The fatigue index is for the Basquin model. It is a 2-norm.

[0125] The physical constraint loss of the Manson-Coffin model is: .

[0126] In the formula: This represents the physical constraint loss for the Manson-Coffin model. This represents the number of samples in the fatigue test. This is an index for fatigue test samples. For the first The total strain amplitude of the test group. This is the fatigue strength coefficient. This is the elastic modulus of the material. For the first The fatigue life was measured in the group test. The fatigue strength index is the Manson-Coffin model. This is the fatigue ductility coefficient. The fatigue ductility index is the Manson-Coffin model index. It is a 2-norm.

[0127] The Chaboche model's physical constraint loss is constructed based on the stress-strain hysteresis loop prediction error, quantifying the deviation between the numerical integral and experimental measurements: The physical constraint loss of the Chaboche model is: .

[0128] In the formula: This represents the physical constraint loss of the Chaboche model. This represents the number of samples in the fatigue test. This is an index for fatigue test samples. For PINN to the first The stress tensor predicted by the group of experiments. For the first The stress tensor was measured in the group of experiments. It is a 2-norm.

[0129] The total loss function of the physical information neural network is: .

[0130] In the formula: This is the total loss function for PINN. This refers to the supervised loss of experimental data. These are the physical constraint weighting coefficients. This represents the physical constraint loss corresponding to the fatigue constitutive model.

[0131] The core advantage of the PINN architecture in this invention lies in the introduction of the physical constraint loss function, which is equivalent to adding a large number of virtual constraint samples to the traditional pure data-driven model. Its mathematical essence is that the fatigue constitutive equation holds true over the continuous domain of the input variables. When these physical equations are used as constraints in the loss function, it is equivalent to requiring the output of the neural network across the entire input space to satisfy physical laws.

[0132] From an information theory perspective: the amount of information carried by a physical constraint term can be equivalent to uniformly sampled information within the domain of the physical equation. Group of virtual samples, in which Dimension of the input space It is directly proportional to the strictness of the constraints.

[0133] Regarding the fatigue constitutive model (input dimension) involved in this invention The virtual sample equivalent of a single physical constraint term is approximately on the order of 10^2 to 10^3. This means that even with only 5-10 sets of real fatigue test data, the PINN framework can provide the information equivalent to 50-100 sets of samples required for a pure data-driven model.

[0134] The posterior probability distribution of the fatigue constitutive parameters is: .

[0135] In the formula: The posterior distribution of fatigue constitutive parameters. The vector represents the random vector of fatigue constitutive parameters to be identified. This is a fatigue test dataset. Let be the likelihood function of the experimental data. The prior distribution of fatigue constitutive parameters. is the Bayesian normalization constant.

[0136] S52. Characterize accidental uncertainty through Bayesian layers or MC Dropout, and characterize cognitive uncertainty through multiple independent PINN models with deep integration, to obtain fatigue constitutive parameters that take into account dual uncertainties.

[0137] Cognitive uncertainty is trained using deep ensemble methods Each independent PINN model is used for characterization.

[0138] The variance of cognitive uncertainty is: .

[0139] In the formula: To understand the variance of uncertainty. This represents the number of independent PINN models in the deep integration. Index for deep integration models. For the first The fatigue constitutive parameter vector output by each PINN model. Output the mean values ​​of parameters for the deep ensemble model.

[0140] The mean output parameters of the deep ensemble model are: .

[0141] In the formula: Output the mean values ​​of parameters for the deep ensemble model. This represents the number of independent PINN models in the deep integration. Index for deep integration models. For the first The fatigue constitutive parameter vector output by each PINN model.

[0142] The small sample adaptation strategy requires only 5 to 10 sets of standard fatigue test samples to complete PINN training. To compensate for the lack of training data due to the small sample size, a semi-supervised learning strategy based on physical constraints is adopted. A large amount of load-response simulation data is generated within the target working condition range using finite element simulation, but physical constraint terms are only applied to the loss function without requiring the simulation data to fit the true labels. This ensures that the simulation data is only used to constrain the model output to meet physical laws, avoiding simulation dominance that could cause the model to deviate from the actual material behavior.

[0143] S6. Construct an on-board digital twin of the commercial vehicle battery rack and assimilate the real vehicle sensor response to the on-board digital twin to update the load boundary, constitutive parameters, key node stress, and cumulative damage.

[0144] The S6 constructs an onboard digital twin of the commercial vehicle battery rack, including the following steps.

[0145] A vehicle-mounted digital twin is established, comprising a geometric twin layer, a physical twin layer, and a data twin layer. Parameter synchronization and status transfer are performed between the geometric twin layer, the physical twin layer, and the data twin layer through a unified data interface.

[0146] Specifically, a digital twin of the vehicle battery rack is established, comprising a three-layer structure of geometric twin, physical twin, and data twin. The three-layer model synchronizes parameters and transmits status through a unified data interface, with a synchronization frequency of no less than 10Hz, and the data format adopts Protobuf serialization. The end-to-end latency is controlled within 50ms to meet the real-time prediction requirements of the vehicle.

[0147] The geometric twin layer was established based on the finite element model of the battery rack and the Craig-Bampton modal reduction method was used to reduce its order.

[0148] Specifically, the geometric twin is built using ANSYS or Abaqus to create a refined finite element model with a mesh size of approximately 500,000 to 1,000,000 elements, fully encompassing details such as welds, bolt preload, and rubber bushings. The Craig-Bampton modal reduction method is used for order reduction, retaining 200 modes covering the frequency range of 0Hz to 500Hz. After reduction, the relative error of the stress response at key nodes meets the following requirements: .in, To reduce the relative error of stress response at key nodes in the model.

[0149] The coordinate transformation is as follows: In the formula: This is the structural physical displacement vector. Reduce the modal matrix for Craig-Bampton. Reduce the generalized coordinate vector for Craig-Bampton.

[0150] The reduced-order system dynamics equations are: In the formula: This is the reduced quality matrix. Reduce the second-order time derivative of the Craig-Bampton generalized coordinate vector. This is the reduced stiffness matrix. This is to reduce the external load vector.

[0151] The physical twin layer embeds the currently matched fatigue constitutive model, fatigue constitutive parameters, multiaxial fatigue damage criteria, and cumulative damage model, and calls the damage equivalence mapping mechanism to maintain the continuity of cumulative damage when switching operating conditions.

[0152] The physical twin layer embeds the fatigue constitutive model determined by S3 and the fatigue constitutive parameters identified by S5, and embeds the multiaxial fatigue damage criterion and cumulative damage model.

[0153] The physical twin layer dynamically calls the corresponding constitutive model based on the current working condition label, and maintains the continuity of the cumulative damage value according to the damage equivalence mapping mechanism in step S4.

[0154] The data twin layer interfaces with the actual vehicle sensor system in real time, transmitting CAN bus data via the MQTT protocol and high-frequency vibration and strain data via the gRPC protocol, synchronizing load input and structural response. A high-fidelity reduced-order proxy model (neural network proxy, structured as 3-layer LSTM + 2-layer fully connected layers) is established, with the input being the measured load time history from the sensors and the output being the stress response time history of key nodes.

[0155] The coefficient of determination of the reduced-order proxy model satisfies: Single-step inference time <10ms, achieving millisecond-level structural response prediction. Wherein: The coefficients of determination for the reduced-order surrogate model.

[0156] The geometric twin layer, physical twin layer, and data twin layer achieve real-time coordination through a unified state vector, and the joint state transfer is as follows: .

[0157] In the formula: For the first A joint state vector of geometry, physics, and data layers with a synchronized step size. Synchronize step size index for digital twins. For digital twin synchronization mapping operators. For the first A joint state vector of geometry, physics, and data layers with a synchronized step size. For the first A real vehicle sensor input vector with a synchronous step size.

[0158] The final result is an onboard digital twin that can operate synchronously with the real vehicle and supports online fatigue damage calculation. The flowchart is as follows: Figure 2 As shown.

[0159] S6 assimilates the real vehicle sensing response to the on-board digital twin to update load boundaries, constitutive parameters, critical node stresses, and cumulative damage, including the following steps.

[0160] Unscented Kalman filtering is used as the primary data assimilation algorithm, switching to particle filtering when the system state exhibits strong non-Gaussian characteristics. This assimilates the structural strain and acceleration responses acquired from the actual vehicle into the onboard digital twin. Specifically, unscented Kalman filtering is used as the primary data assimilation algorithm to handle the nonlinear fatigue evolution process, with particle filtering as a backup. The algorithm automatically switches when the system state exhibits strong non-Gaussian characteristics to ensure assimilation accuracy under extreme conditions. The assimilation operation is executed with a fixed 10-second cycle, and an additional assimilation is triggered each time the operating condition changes, ensuring rapid synchronization of key state points.

[0161] The filtered state vector is defined as a fusion vector of load boundary parameters, constitutive parameters, cumulative damage, and key point stress.

[0162] .

[0163] In the formula: For the first The filtered state vector of each assimilation step. Index for the data assimilation step. These are the parameters for correcting the boundary amplitude of the power spectral density load. This is the current fatigue constitutive parameter estimation vector in the data assimilation state vector. This represents the cumulative damage value. This represents the stress state vector at the critical node. This represents the matrix transpose symbol.

[0164] The observation vector is defined as the strain and acceleration response collected from the actual vehicle.

[0165] .

[0166] In the formula: For the first The observation vector of each assimilation step. This is the structural strain response vector collected from the actual vehicle. This is the acceleration response vector collected from the actual vehicle.

[0167] Based on the previously constructed digital twin agent model as the prediction model, state transition equations and observation mapping equations are constructed.

[0168] The state transition equation is: .

[0169] The observation mapping equation is: .

[0170] In the formula: For the first The filtered state vector of each assimilation step. This is the state transition function. For the first The process noise vector of each assimilation step. This is the observation mapping function. For the first The observation noise vector for each assimilation step.

[0171] After each data assimilation, the load boundary conditions, the posterior mean of the fatigue constitutive parameters, the stress time history of key nodes, and the cumulative damage value are simultaneously corrected.

[0172] Specifically, after each assimilation, the system synchronously updates the following: corrects the amplitude coefficient of the power spectral density to update the load boundary conditions; updates the mean of the posterior distribution of fatigue constitutive parameters; recalculates the cumulative damage increment based on the assimilated stress; and updates the stress time history of weak nodes.

[0173] Through the above assimilation mechanism, virtual-real synchronization is achieved, that is, the digital twin model always keeps in line with the actual state of the physical entity, and completes dynamic iteration and closed-loop update.

[0174] In terms of computational deployment, both the unscented Kalman filter and the particle filter are deployed on the vehicle edge unit with a computing power of no less than 20 TOPS, and the single calculation time is less than 100ms, which meets the real-time requirements of the vehicle.

[0175] Figure 6 The flowchart of fatigue damage calculation and remaining life prediction is the final output flow of this invention, showing the complete link of improved Miner damage accumulation, failure determination, Monte Carlo probabilistic life prediction, weak part identification, three-level vehicle warning and cloud platform recording.

[0176] S7. Calculate the fatigue damage increment based on the updated fatigue constitutive model and key node responses, and predict the remaining service life. Preferably, S7 includes S71 to S77.

[0177] S71. Based on the currently matched fatigue constitutive model and the assimilated stress or strain spectrum of key nodes, the improved Miner cumulative damage method, which introduces a load sequence effect correction factor, is used to calculate the damage increment per unit time or per unit mileage.

[0178] S72. The incremental damage per unit time or unit mileage is: In the formula: This represents the incremental damage per unit time or unit mileage. This represents the fatigue life cycle number under the current equivalent stress amplitude. This is the equivalent stress amplitude. This is the load order correction coefficient function. This represents the stress amplitude of the previous load.

[0179] By calibrating the kinematic hardening response using the Chaboche model, damage accumulation can reflect the acceleration effect of the low-to-high load sequence and the hysteresis effect of the high-to-low load sequence.

[0180] S73, Cumulative damage is: In the formula: This represents the cumulative damage value. This is an index for the damage increment sequence. This represents the total number of damage increments already counted in the cumulative damage calculation. For the first The incremental damage per unit time or unit distance.

[0181] The system has a built-in model library and a priori distribution of each model parameter. The switching of the fatigue constitutive model is triggered by the working condition classification result and the duration. The model switching is performed when the target working condition appears continuously for no less than 3 time windows. During the switching, the continuity of cumulative damage is maintained according to the damage equivalence mapping mechanism in step S4.

[0182] S74. When the cumulative damage reaches the failure determination threshold, fatigue failure is determined to have occurred in the corresponding part of the commercial vehicle battery rack.

[0183] S75. Monte Carlo sampling is performed from the posterior distribution of fatigue constitutive parameters, and the probability distribution of remaining service life is obtained by integrating the load uncertainty.

[0184] Specifically, Monte Carlo sampling is performed from the posterior distribution of fatigue constitutive parameters, and the number of Monte Carlo samplings is... For a thousand cycles, the remaining lifetime point estimate is output by comprehensively considering the load uncertainty.

[0185] The posterior distribution of fatigue constitutive parameters is adopted as follows: In the formula: The posterior distribution of fatigue constitutive parameters. The vector represents the random vector of fatigue constitutive parameters to be identified. This is a fatigue test dataset.

[0186] S76, The remaining useful life point estimate is: In the formula: This is a point estimate of the remaining useful life. It is a time-weighted expectation operator based on the current operating conditions.

[0187] S77. Based on the probability distribution of the remaining useful life, output the point estimate of the remaining useful life and the confidence interval.

[0188] Specifically, the system comprehensively considers constitutive parameter uncertainties and load uncertainties, outputting the remaining lifetime distribution with a 95% confidence interval. Finally, it provides a three-level vehicle-mounted warning based on the remaining lifetime: green indicates safety, yellow indicates warning, and red indicates danger.

[0189] The green level corresponds to a remaining lifespan of more than 1000 hours.

[0190] The yellow level corresponds to a remaining lifespan of more than 100 hours and no more than 1000 hours.

[0191] The red level corresponds to a remaining lifespan of no more than 100 hours.

[0192] The cloud platform records the entire lifecycle fatigue evolution curve and abnormal event tracing information. The entire lifecycle fatigue evolution curve is used to characterize the process of damage accumulation over time or mileage. The abnormal event tracing information includes timestamps and state change records for each operating condition switch, model switch, and assimilation trigger, providing data support for structural design and maintenance decisions.

[0193] In terms of weak point identification and output, the system uses a digital twin model to calculate the equivalent stress amplitude and damage rate of all key nodes in real time. Key nodes include welds, bolt holes, and mounting points. The system sorts the weak points in descending order of damage rate and outputs the top three weak points and their damage rates, with the damage rate measured in micro-damage per hour or micro-damage per 100 kilometers.

[0194] Specifically, the equivalent stress amplitude, damage increment, and damage rate of each key node of the battery rack are calculated in real time using the on-board digital twin. These key nodes include welds, bolt holes, and mounting points. The key nodes are sorted according to their damage rate, and the nodes with the highest damage rates are identified as weak points. Warning levels are assigned based on remaining service life: a safety level is output when the remaining service life is greater than a first lifespan threshold; a warning level is output when the remaining service life is no greater than the first lifespan threshold but greater than a second lifespan threshold; and a danger level is output when the remaining service life is no greater than the second lifespan threshold. The entire lifecycle fatigue evolution curve and abnormal event tracing information are recorded. This abnormal event tracing information includes the time and state change information corresponding to operating condition switching, model switching, and data assimilation triggers.

[0195] This embodiment designs an adaptive fatigue constitutive model switching mechanism based on the time-frequency domain features of road excitation: by extracting the time-domain, frequency-domain, and multi-dimensional features of the vibration signal, a road excitation feature vector is constructed. A CNN-LSTM classification model is used to adaptively distinguish between three working conditions: smooth urban roads, conventional highway roads, and bumpy off-road roads. The optimal fatigue constitutive model is automatically matched according to the working condition category: the Basquin stress-life model is matched for urban conditions, the Manson-Coffin strain-life model for highway conditions, and the Chaboche nonlinear kinematic hardening constitutive model combined with the critical plane method is matched for off-road conditions. Model switching is triggered by both the working condition classification result and the duration, and a damage equivalence mapping mechanism is used to ensure the continuity of cumulative damage before and after the switching.

[0196] Model switching is triggered by both the load condition classification result and the duration, and achieves seamless bidirectional connection between the uniaxial model (Basquin / Manson-Coffin) and the multiaxial model (Chaboche) through the damage equivalence mapping mechanism proposed in this invention. The core functions of this mapping mechanism are: unifying damage dimensions by equivalently converting stress amplitude damage, strain amplitude damage, and plastic strain energy damage within the same mathematical framework; ensuring cumulative continuity by guaranteeing strict consistency of the cumulative damage value D before and after switching, avoiding artificial damage jumps caused by model switching; and achieving physical consistency, ensuring that the fatigue accumulation process across load conditions satisfies the fundamental conservation laws of continuum mechanics.

[0197] This damage equivalence mapping mechanism is the "logical glue" of the entire adaptive fatigue life prediction architecture: it couples the three originally independent components—the working condition discrimination module, the constitutive model matching module, and the damage accumulation module—into an organic whole. Without this mechanism, the three modules would degenerate into a serial calling relationship, a simple combination of existing technologies. With the introduction of this mechanism, the three form a two-way data dependency and state feedback loop, generating synergistic effects that exceed the sum of the functions of each module. This achieves unified fatigue life assessment across working conditions and damage mechanisms, which is the essential characteristic that distinguishes this invention from the existing "segmented prediction" methods.

[0198] This embodiment also designs a small-sample data-physical hybrid modeling and uncertainty decoupling method based on PINN: By constructing a Physical Information Neural Network (PINN), the fatigue constitutive equation is used as a physical constraint embedding loss function to establish a hybrid modeling framework jointly driven by data and physical laws. Specific physical constraint loss function forms are constructed for the three fatigue constitutive models: Basquin, Manson-Coffin, and Chaboche. Small-sample fatigue test data (5–10 sets) is used as the main driver, combined with finite element simulation data as a semi-supervised auxiliary, to output the posterior probability distribution of constitutive parameters. Accidental uncertainty is modeled through Bayesian layers or MC Dropout, and cognitive uncertainty is modeled through deep ensemble methods, achieving effective decoupling of the two types of uncertainty.

[0199] This embodiment also designs a three-layer coupled digital twin construction method: a digital twin of the battery rack comprising a geometric twin, a physical twin, and a data twin. The geometric twin layer is based on a refined finite element model and uses the Craig-Bampton modal reduction method to achieve order reduction. The physical twin layer embeds an adaptive constitutive model and a multiaxial fatigue damage criterion (critical surface method). The data twin layer interfaces with the actual vehicle sensor system in real time via the MQTT / gRPC protocol to synchronize load input and structural response. The three-layer model synchronizes parameters and transfers states through a unified data interface.

[0200] This embodiment also designs an online remaining lifetime update method based on dynamic data assimilation: using unscented Kalman filtering (UKF) or particle filtering (PF) algorithms, triggered every 10 seconds or by a change in operating conditions, the structural strain and acceleration response data collected from the actual vehicle are assimilated into the digital twin model. The state vector is constructed using load boundary parameters, constitutive parameters, cumulative damage, and key point stresses, while the actual vehicle strain and acceleration are used as observation vectors. A filtering algorithm is employed to achieve real-time estimation and correction of the state variables. After assimilation, the load boundary conditions, constitutive model parameters, cumulative damage values, and stress states at key locations are updated synchronously, achieving "virtual-real synchronization" and dynamic iteration.

[0201] This embodiment also designs an improved damage accumulation and probabilistic lifetime prediction method considering the load sequence effect: by adopting an improved Miner cumulative damage rule and introducing a load sequence effect correction factor, the acceleration effect of low-high load sequence and the hysteresis effect of high-low load sequence are calibrated based on the kinematic hardening response of the Chaboche model. The continuity of accumulated damage is maintained through a damage equivalence mapping mechanism when the constitutive model switches. Based on the posterior distribution of constitutive parameters, the point estimate of remaining lifetime and the probabilistic lifetime with a 95% confidence interval are output through Monte Carlo sampling. The three critical weak points with the highest damage rates and their damage rates are identified, and real-time early warning of fatigue risk is achieved by combining green, yellow, and red three-level warning levels.

[0202] Existing technologies employ a single fixed constitutive model, resulting in prediction errors of 30%–50% across different road conditions. This invention, through adaptive discrimination of road conditions, automatically matches Basquin, Manson-Coffin, or Chaboche models and uses damage equivalence mapping to ensure continuity of switching, with the expected prediction error across different road conditions controlled within 15%.

[0203] Existing technologies rely on 30–50 sets of fatigue test samples, which are time-consuming and costly. This invention constructs a data-physics hybrid model based on PINN, embedding the fatigue constitutive equation into the loss function as a physical constraint. Training can be completed with only 5–10 sets of small samples, reducing the test sample requirement by more than 70%.

[0204] Existing technologies output single-point lifetime estimates, failing to assess the reliability of the predictions. This invention achieves dual decoupling of accidental and cognitive uncertainties through Bayesian layers and deep ensemble methods, outputting a probabilistic lifetime with a 95% confidence interval, providing quantitative support for risk decision-making.

[0205] Existing technologies rely on offline static prediction or model fixation, which cannot utilize real vehicle data for closed-loop correction. This invention constructs a three-layer digital twin consisting of geometry, physics, and data, and employs the UKF / PF algorithm to dynamically assimilate data every 10 seconds or when the operating condition changes. This process updates load boundaries, constitutive parameters, and cumulative damage in real time, achieving virtual-real synchronization and dynamic iteration.

[0206] Existing technologies employ the traditional Miner linear accumulation rule, neglecting the load sequence effect. This invention introduces a load sequence correction factor, calibrates based on Chaboche kinematic hardening response, reduces damage accumulation error by more than 20% compared to traditional methods, and supports the critical plane method for handling multiaxial fatigue.

[0207] Existing technologies output a single lifetime value or early warning information. This invention outputs point estimates and 95% confidence interval probability lifetimes, the location of the top three weak points with the highest damage rates, and green / yellow / red three-level early warnings. It also uploads the full life cycle fatigue evolution curve and abnormal event tracing to a cloud platform, providing complete data support for design optimization and maintenance decisions.

[0208] Existing multi-model methods typically employ a divide-and-conquer strategy, predicting separately under different operating conditions, making it difficult to interpret the results uniformly. This invention achieves seamless integration of three fatigue constitutive models with different physical mechanisms—Basquin (stress life), Manson-Coffin (strain life), and Chaboche (nonlinear kinematic hardening)—within the same cumulative damage framework through a damage equivalence mapping mechanism.

[0209] This mechanism enables information transferability, state retrospection, and unified prediction. Information transferability: Damage states under previous conditions (such as strain amplitude accumulation under high-speed road conditions) can influence the remaining life calculation of subsequent conditions (such as multi-axis fatigue under off-road conditions) in a physically consistent manner. State retrospection: When the condition switches from multi-axis back to uniaxial, the equivalent uniaxial stress / strain amplitude can be derived through inverse mapping, ensuring the continuity and interpretability of the full life-cycle damage evolution curve. Unified prediction: The final output is a single, physically meaningful cumulative damage value D and remaining life RUL, rather than multiple segmented results that are difficult to integrate. This design concept of "multi-model collaboration rather than multi-model splicing" is a fundamental breakthrough of this invention compared to existing technologies.

[0210] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive fatigue life prediction of a commercial vehicle battery rack for random vibration, characterized in that, include: S1. Collect multi-source real vehicle data on the battery rack and vehicle operating status of commercial vehicles, and preprocess the multi-source real vehicle data to obtain standardized time-series data. S2. Based on the standardized time-series data, extract the time-domain features, frequency-domain features, and time-frequency-domain features of the road surface excitation to form a working condition classification input feature vector; S3. Input the working condition classification input feature vector into the trained time series classification model, determine the current road working condition, and match the corresponding fatigue constitutive model in the fatigue constitutive model library according to the current road working condition. S4. When the fatigue constitutive model is switched, the cumulative damage before and after the switch is equivalently mapped to maintain the continuity of fatigue damage accumulation across working conditions. S5. Using a physical information neural network embedded with fatigue constitutive physical constraints, fatigue constitutive parameters and their uncertainties are identified based on fatigue test data and simulation constraint data. S6. Construct an on-board digital twin of the commercial vehicle battery rack and assimilate the real vehicle sensor response to the on-board digital twin to update the load boundary, constitutive parameters, key node stress and cumulative damage; S7. Calculate the fatigue damage increment based on the updated fatigue constitutive model and key node responses, and predict the remaining service life.

2. A method for adaptive fatigue life prediction of a commercial vehicle battery rack subjected to random vibration according to claim 1, characterized in that S3 include: The working condition classification input feature vector is input into the CNN-LSTM classification model. The local time-frequency pattern is extracted through the convolutional layer, the temporal dependency is extracted through the LSTM layer, and the posterior probability of the current time window belonging to different road working conditions is output through Softmax. The output of the convolutional layer is: ； In the formula: The convolutional layer outputs a feature vector; For neural network activation functions; This is the weight matrix of the convolutional layer; This is the convolution operator; Input feature vectors for operating condition classification; This is the bias vector of the convolutional layer; The output of the LSTM layer is: ； In the formula: For the first The LSTM hidden state vector for each time step; For LSTM timing step index; For Long Short-Term Memory (LSTM) network mapping functions; For the first Each time step corresponds to the output feature vector of the convolutional layer; For the first The LSTM hidden state vector for each time step; The classification score vector is: ； In the formula: This is the classification score vector; Output the weight matrix for the fully connected layer; Output bias vector for fully connected layers; No. The Softmax probability of the type of working condition is: ； In the formula: For the first The posterior probability of a given working condition; Index for operating condition categories; It is a natural exponential function; For the classification score vector, the first... Scores for similar working conditions; Index for Softmax normalized summation; For the classification score vector, the first... Scores for similar working conditions; The maximum posterior probability is used as the model output confidence level, and the current road surface condition is determined based on the model output confidence level.

3. The adaptive fatigue life prediction method for commercial vehicle battery racks accommodating random vibration as described in claim 2, characterized in that, The working condition category index is taken as: ,in, Indicates a smooth urban road surface. Indicates a regular highway surface. Indicates an off-road or bumpy road surface; The model outputs a confidence level of: ； In the formula: Output confidence scores for the model; This is the maximum value operator; The posterior probability of a smooth urban road surface; For the posterior probability of a conventional highway surface; For off-road or bumpy road surfaces; The current road surface condition is determined based on the confidence level output by the model, specifically including: When the confidence level of the model output is greater than or equal to the preset confidence threshold, the maximum posterior probability condition corresponding to the current time window is used as the current road condition. When the confidence level of the model output is less than the preset confidence threshold, time-series voting is performed based on the classification results of multiple consecutive time windows, and the voting result is used as the current road surface condition. S3 also includes: When the current road surface condition is a smooth urban road surface, the Basquin stress-life model is matched. The Basquin stress-life model is as follows: ； In the formula: This is the equivalent stress amplitude; The fatigue index for the Basquin model; This represents the fatigue life cycle count. These are the material constants for the Basquin model; When the current road surface condition is a conventional highway road surface, the Manson-Coffin strain-life model is matched. The Manson-Coffin strain lifetime model is as follows: ； In the formula: This represents the total strain amplitude; This is the fatigue strength coefficient; The elastic modulus of the material; The fatigue strength index of the Manson-Coffin model; It is the fatigue ductility coefficient; The fatigue ductility index of the Manson-Coffin model; When the current road condition is an off-road or bumpy road, the Chaboche nonlinear kinematic hardening constitutive model is matched and combined with the critical plane method; The Chaboche nonlinear kinematic hardening constitutive model is: ； In the formula: For the rate of change of the back stress tensor in the Chaboche model; The kinematic hardening modulus of the Chaboche model; To dynamically recover parameters for the Chaboche model; The rate of change of the plastic strain tensor; For the back stress tensor of the Chaboche model; S4 specifically includes: When the number of consecutive time windows of the target working condition meets the model switching condition, the fatigue constitutive model is switched, and the cumulative damage before and after the switch is kept consistent through damage equivalence mapping.

4. The adaptive fatigue life prediction method for commercial vehicle battery racks oriented towards random vibration as described in claim 1, characterized in that, S5 include: A physical information neural network is constructed, incorporating the currently matched fatigue constitutive model as a physical constraint into the loss function. The fatigue test data serves as the primary driver, while simulation constraint data serves as an auxiliary constraint, outputting the posterior probability distribution of the fatigue constitutive parameters. The input vector of the physical information neural network is: ； In the formula: For PINN input vectors; This is the equivalent stress amplitude; This represents the total strain amplitude; Multiaxial stress ratio; The elastic modulus of the material; The yield strength of the material; For operating condition labels; Represents the matrix transpose symbol; The total loss function of the physical information neural network is: ； In the formula: This is the total loss function for PINN; Supervised loss of experimental data; These are the physical constraint weighting coefficients; This refers to the physical constraint loss corresponding to the fatigue constitutive model; The posterior probability distribution of the fatigue constitutive parameters is: ； In the formula: The posterior distribution of fatigue constitutive parameters; The random vector of fatigue constitutive parameters to be identified; This is a fatigue test dataset; Let be the likelihood function of the experimental data; The prior distribution of fatigue constitutive parameters; This is the Bayesian normalization constant; Accidental uncertainty is represented by Bayesian layers or MC Dropout, and cognitive uncertainty is represented by multiple independent PINN models with deep integration, so as to obtain fatigue constitutive parameters that take into account dual uncertainties.

5. The adaptive fatigue life prediction method for commercial vehicle battery racks accommodating random vibration as described in claim 1, characterized in that, Constructing an onboard digital twin of a commercial vehicle battery rack includes: A vehicle-mounted digital twin containing a geometric twin layer, a physical twin layer, and a data twin layer is established, and parameter synchronization and status transfer are performed between the geometric twin layer, the physical twin layer, and the data twin layer through a unified data interface; The geometric twin layer was established based on the finite element model of the battery rack and the Craig-Bampton modal reduction method was used to reduce its order. The coordinate transformation is as follows: ； In the formula: The physical displacement vector of the structure; Reduce the modal matrix for Craig-Bampton; Reduce the generalized coordinate vector for Craig-Bampton; The reduced-order system dynamics equations are: ； In the formula: This is the reduced mass matrix; Reduce the second-order time derivative of the Craig-Bampton generalized coordinate vector; The reduced stiffness matrix; To reduce the external load vector; The physical twin layer embeds the currently matched fatigue constitutive model, fatigue constitutive parameters, multiaxial fatigue damage criteria, and cumulative damage model, and calls the damage equivalence mapping mechanism to maintain the continuity of cumulative damage when switching operating conditions; The data twin layer interfaces with the actual vehicle sensor system in real time to receive load inputs and structural responses, and uses a reduced-order surrogate model to output the stress response time history of key nodes. The joint state transfer of the geometric twin layer, physical twin layer, and data twin layer is as follows: ； In the formula: For the first A joint state vector of geometry, physics, and data layers with a synchronization step size; Synchronize step size index for digital twins; For digital twin synchronization mapping operators; For the first A joint state vector of geometry, physics, and data layers with a synchronization step size; For the first A real vehicle sensor input vector with a synchronous step size.

6. The adaptive fatigue life prediction method for commercial vehicle battery racks accommodating random vibration as described in claim 5, characterized in that, Assimilate the real-vehicle sensor response into the onboard digital twin to update load boundaries, constitutive parameters, critical node stresses, and cumulative damage, including: Unscented Kalman filtering is used as the main data assimilation algorithm, and particle filtering is switched when the system state exhibits strong non-Gaussianity to assimilate the structural strain response and acceleration response collected from the actual vehicle to the vehicle-mounted digital twin. The filter state vector is: ； In the formula: For the first The filtered state vector of each assimilation step; Index for data assimilation steps; These are parameters for correcting the boundary amplitude of the power spectral density load. This is the current fatigue constitutive parameter estimation vector in the data assimilation state vector; This represents the cumulative damage value. This represents the stress state vector of the critical node. Represents the matrix transpose symbol; The observation vector is: ； In the formula: For the first The observation vector of each assimilation step; The structural strain response vector was collected from the actual vehicle. This is the acceleration response vector collected from the actual vehicle; The state transition equation is: ； In the formula: For the first The filtered state vector of each assimilation step; This is the state transition function; For the first The noise vector of each assimilation step; The observation mapping equation is: ； In the formula: For observation mapping function; For the first The observation noise vector of each assimilation step; After each data assimilation, the load boundary conditions, the posterior mean of the fatigue constitutive parameters, the stress time history of key nodes, and the cumulative damage value are simultaneously corrected.

7. The adaptive fatigue life prediction method for commercial vehicle battery racks accommodating random vibration as described in claim 1, characterized in that, S7 includes: Based on the current matched fatigue constitutive model and the assimilated stress or strain spectrum of key nodes, the improved Miner cumulative damage method, which introduces a load sequence effect correction factor, is used to calculate the damage increment per unit time or per unit mileage. The incremental damage per unit time or unit distance is: In the formula: This represents the incremental damage per unit time or unit distance. This represents the fatigue life cycle number under the current equivalent stress amplitude. This is the equivalent stress amplitude; For load order correction coefficient function; This represents the stress amplitude of the previous load. The cumulative damage is: In the formula: This represents the cumulative damage value. For damage increment sequence index; This represents the total number of damage increments already counted in the cumulative damage calculation; For the first Damage increment per unit time or unit mileage; When the cumulative damage reaches the failure determination threshold, it is determined that fatigue failure has occurred in the corresponding part of the commercial vehicle battery rack. Monte Carlo sampling is performed from the posterior distribution of fatigue constitutive parameters, and the probability distribution of remaining service life is obtained by integrating load uncertainties; The remaining useful life point estimate is: In the formula: This is a point estimate of the remaining useful life. This is a time-weighted expectation operator based on the current operating conditions; Based on the probability distribution of the remaining useful life, output the point estimate of the remaining useful life and the confidence interval.

8. An adaptive fatigue life prediction method for commercial vehicle battery racks accommodating random vibration, as described in any one of claims 1 to 7, characterized in that, S1 specifically includes: Triaxial accelerometers and strain sensors are installed in the battery pack housing, key load-bearing nodes of the battery rack, and the connection between the vehicle frame and the suspension to collect road excitation vibration signals and battery rack structural strain response signals, and obtain vehicle speed, load status and positioning information through the vehicle bus. The original signal matrix from the multi-source sensing is represented as: In the formula: This is the original signal matrix from the multi-source sensing. For continuous sampling time; For the first Time-domain signals acquired by road sensors; For sensor channel index; For the number of sensors; Represents the matrix transpose symbol; The continuous sampling time satisfies: In the formula: Total sampling time; The original signal matrix from the multi-source sensors is subjected to wavelet threshold denoising, unified resampling, and outlier removal, and then sliced ​​according to a fixed time window to obtain the standardized time-series data.

9. An adaptive fatigue life prediction method for commercial vehicle battery racks accommodating random vibration, as described in any one of claims 1 to 7, characterized in that, S2 specifically includes: Within each time window, the vertical acceleration sequence at the suspension point is used as the primary signal to extract the root mean square, peak factor, kurtosis, frequency band energy ratio, peak frequency, and wavelet packet energy features. The root mean square is: ； In the formula: The root mean square of the signal within the time window; This represents the number of sampling points within the time window. Index of sampling points within the time window; For the preprocessed first One vibration signal sample value; The peak factor is: ； In the formula: Peak factor; For indexing sampling points within the time window The operator that retrieves the maximum value; The kurtosis is: ； In the formula: Signal kurtosis within the time window; The average value of the signal within the time window; The variance of the signal within the time window; The acceleration power spectral density is: ； In the formula: The acceleration power spectral density; For frequency variables; The duration of the signal in the current analysis window; The sampling frequency; Fourier transform of the acceleration signal; The frequency band energy ratio is: ； ； ； In the formula: The proportion of energy in the 0Hz to 10Hz frequency band to the total energy in the 0Hz to 200Hz frequency band; The proportion of energy in the 10Hz to 50Hz frequency band to the total energy in the 0Hz to 200Hz frequency band; The proportion of energy in the 50Hz to 200Hz frequency band to the total energy in the 0Hz to 200Hz frequency band; The peak frequency is: ； In the formula: This is the peak frequency at which the power spectral density reaches its maximum value. For frequency variables The operator that takes the maximum value of the corresponding argument variable; The wavelet packet energy characteristic is: ； In the formula: For the first Wavelet packet energy of each frequency band; For wavelet packet frequency band index; For wavelet packet coefficient index; For the first Number of wavelet packet coefficients within each frequency band; For the first Within each frequency band Wavelet packet coefficients; The time-domain features, frequency-domain features, and time-frequency-domain features are concatenated to obtain the working condition classification input feature vector.

Images