A hydrogen storage tank life prediction method based on multi-source perception and machine learning

Abstract

The application discloses a hydrogen storage tank life prediction method based on multi-source sensing and machine learning, and relates to the technical field of hydrogen storage tank life prediction. The method comprises the following steps: S1, during each hydrogen charging and discharging cycle process, pressure signals, strain signals and acoustic emission signals are synchronously collected; S2, based on the measured strain signals, the probability distribution of unmeasurable damage parameters is inversed through a physical information neural network with relevant physical equations as constraints; and S3, based on the time domain variation characteristics of the pressure signals, each cycle is automatically divided into a pressure charging section, a pressure maintaining section, a pressure releasing section and a standing section, different original degradation characteristic values of each stage are extracted and compared with a health state baseline, and a measured degradation deviation is calculated. Through the setting of the physical information neural network and the parameterized probability output layer, the probability distribution of the initial crack size and the hydrogen embrittlement coefficient can be directly inversed from the measured strain signals, and the quantitative sensing and early identification of the unmeasurable parameters are realized.

Description

Technical Field

[0001] This invention relates to the field of hydrogen storage tank life prediction technology, specifically a method for predicting the life of hydrogen storage tanks based on multi-source sensing and machine learning. Background Technology

[0002] Hydrogen storage tanks are core equipment for hydrogen storage and transportation, and their operational safety directly impacts the large-scale application of the hydrogen energy industry. During long-term hydrogen filling and discharging cycles, hydrogen storage tanks are subjected to the coupled effects of alternating pressure, temperature fluctuations, and the hydrogen environment, leading to gradual degradation of material properties. The main failure modes include fatigue cracking of the metal liner, hydrogen embrittlement, and damage accumulation in the composite material layers. Therefore, accurately predicting their remaining lifespan is crucial for preventing failures, ensuring safety, and optimizing operation and maintenance.

[0003] Currently, health management of hydrogen storage tanks mainly relies on periodic offline inspections (such as ultrasound and acoustic emission) and online monitoring based on limited sensors. For example, invention patent CN118362165A discloses a method, device, and system for monitoring the damage state of a high-pressure hydrogen storage container. This scheme measures strain using fiber optic grating sensors, uses classical mechanics formulas to invert macroscopic parameters such as wall thickness, and combines periodic ultrasonic inspections for defect location and safety assessment.

[0004] However, the above-mentioned and similar technical solutions have the following shortcomings when applied to hydrogen storage tank life prediction: relying on directly measurable parameters such as strain and pressure, it is difficult to effectively perceive and quantify key microscopic damage parameters that drive fatigue and hydrogen embrittlement failure, such as initial crack size and hydrogen embrittlement coefficient, resulting in insensitivity to early and hidden damage; the influence of variable operating conditions such as pressurization rate, holding time, and ambient temperature on the material damage accumulation rate is not fully considered in different hydrogen charging and discharging cycles, resulting in poor generalization ability of the evaluation model under variable operating conditions and decreased prediction accuracy; the evaluation relies heavily on triggered or periodic offline detection, making it difficult to achieve continuous and automatic tracking and cumulative calculation of damage in each hydrogen charging and discharging cycle. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the lifespan of hydrogen storage tanks based on multi-source sensing and machine learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning, comprising: S1. Pressure signals, strain signals and acoustic emission signals are collected synchronously during each hydrogen charge-discharge cycle; S2. Based on measured strain signals, the probability distribution of unmeasurable damage parameters is inverted through a physical information neural network constrained by relevant physical equations. S3. Based on the time-domain variation characteristics of the pressure signal, each cycle is automatically divided into a pressurization stage, a pressure holding stage, a pressure release stage, and a resting stage. Different original degradation characteristic values ​​of each stage are extracted and compared with the baseline of the healthy state to calculate the measured degradation deviation. S4. Correct the measured degradation bias based on the probability distribution of unmeasurable damage parameters, and use an attention network to weight and fuse the modified degradation biases of each stage to output the comprehensive damage increment for this cycle. S5. Construct a cumulative damage trajectory based on the historical cumulative damage sequence. Using the complete probability distribution of unmeasurable damage parameters as a priori condition, learn the probability density function of cumulative damage through Gaussian process regression and use the preset quantile as the dynamic lifetime endpoint. Output the remaining number of cycles with confidence interval.

[0007] Furthermore, the unmeasurable damage parameters include the initial crack size and the hydrogen embrittlement coefficient; S2 includes: The measured strain signal of the current cycle is acquired and constructed into a strain time series with equal interval sampling; The strain time sequence is input into a bidirectional long short-term memory network, and the end of the network is connected to a parameterized probability output layer, which is used to directly output the probability distribution parameters of the initial crack size and hydrogen embrittlement coefficient. Using the modified Paris formula that takes into account the hydrogen embrittlement effect as a physical constraint, sampling is performed based on the probability distribution parameters, and the theoretical crack propagation is calculated by substituting into the formula. The theoretical crack propagation is then compared with the reference crack propagation calculated by back-calculation of measured strain to construct a physical constraint loss term. Construct a total loss function that includes data fitting loss, physical constraint loss and regularization loss, wherein the weight of the physical constraint loss is dynamically increased during training according to a preset strategy; A staged transfer learning strategy is adopted to train the network. The network is pre-trained on a public dataset, the feature extraction layer is frozen, and the parameterized probability output layer of the network is fine-tuned online using a single sample from the current loop. Output the joint posterior probability distribution of the initial crack size and hydrogen embrittlement coefficient.

[0008] Furthermore, the method for stage division in S3 includes: The collected pressure signals are smoothed, and the pressure change rate at each sampling point is calculated. Obtain the target hydrogen charging pressure value for this cycle, and based on this value and the initial average pressurization rate, dynamically calculate the starting thresholds for the charging and depressurization stages; When the pressure change rate at several consecutive sampling points exceeds the starting threshold of the pressurization phase, the pressurization phase is determined to begin; when the absolute value of the pressure change rate approaches zero and the pressure value reaches the set proportion of the target pressure value, the pressurization phase is determined to end and the pressure holding phase is determined to begin. When the pressure change rate at several consecutive sampling points is less than the starting threshold of the pressure relief section, the pressure holding section is determined to end and the pressure relief section is determined to begin. When the pressure drops to near the ambient pressure and the absolute value of the pressure change rate approaches zero, the pressure relief phase ends and the settling phase begins. The settling period continues from the end of the depressurization period to the beginning of the pressurization period of the next cycle. If the start of the next cycle is not detected within the preset time window, it continues until the end of this data collection. The duration of each stage is compared with the adaptive time range of historical statistics, and abnormal durations are corrected or the cycle is marked as an abnormal cycle.

[0009] Furthermore, the degradation features extracted in each stage of S3 include: peak strain, maximum strain rate, and strain-pressure data point sequence in the pressurization stage; creep slope, residual strain increment, and strain-time data point sequence in the pressure holding stage; elastic recovery, hysteresis loop area, and plastic strain ratio in the depressurization stage; and zero-point drift, strain relaxation data point sequence, and strain fluctuation signal in the static stage.

[0010] Furthermore, the establishment of the health status baseline in S3 includes: Collect complete sensor data for several consecutive cycles after the hydrogen storage tank is put into use, and select qualified cycles based on preset physical rules to form a healthy candidate cycle set; For each healthy candidate cycle, extract the original degradation feature values ​​for each stage; Extract the operating condition parameters of each healthy candidate cycle, establish a mapping model from the operating condition parameters to each original degradation feature value, correct the original degradation feature value, and obtain the standardized feature value; For each degradation feature, calculate the moving mean and moving standard deviation of its standardized feature value sequence; when the moving statistic meets the preset convergence condition, the moving mean at the time of convergence is determined as the healthy baseline value of the degradation feature, and the moving standard deviation at the time of convergence is recorded as the inherent fluctuation range of the degradation feature in the healthy state. Based on the strain-pressure data point sequence of all qualified cycles in the healthy candidate cycle set, the average value of each point is taken after alignment to generate the standard strain-pressure curve of the pressurization section. The health baseline values ​​of all degradation characteristics, the inherent fluctuation range, and the standard strain-pressure curve together constitute the health status baseline of the hydrogen storage tank.

[0011] Furthermore, S4 includes: Based on the posterior probability distribution of the unmeasurable damage parameters output by S2, the physical probability of occurrence of the measured degradation deviation at each stage extracted by S3 is calculated, and the measured degradation deviation at each stage is corrected according to the physical probability of occurrence to obtain the physically reliable degradation deviation. Input the physical credibility degradation bias, historical cumulative damage sequence, and original signal quality index of each stage of the current loop into the attention network, and output the attention weights corresponding to each stage. The attention weights are used to weight and fuse the physical credibility degradation biases at each stage to calculate the overall damage increment for this cycle, and its uncertainty estimate is calculated.

[0012] Furthermore, the method for calculating the probability of physical occurrence includes: sampling based on the posterior probability distribution of unmeasurable damage parameters, generating a predicted distribution of degradation characteristics at each stage through a physical degradation model; calculating the probability that the measured degradation deviation is located outside the preset central interval of its corresponding predicted distribution as the probability of physical occurrence; the original signal quality index includes at least one of the coefficient of variation of strain signal and the signal-to-noise ratio of acoustic emission signal at the corresponding stage.

[0013] Furthermore, S5 includes: Constructing cumulative damage trajectories based on historical cumulative damage sequences; Based on the posterior probability distribution of unmeasurable damage parameters, a pre-defined physical damage evolution model is driven by Monte Carlo sampling to generate damage evolution sample trajectories, and the expected value of each trajectories is calculated to obtain the expected value sequence, which serves as the mean function of the Gaussian regression process. Using the number of cycles as input, the cumulative damage sequence as observation, the mean function as the prior mean, and the monotonic kernel function as the covariance function, Gaussian process regression is performed to obtain the posterior predicted distribution of cumulative damage corresponding to any future number of cycles. Based on the posterior probability distribution of unmeasurable damage and the posterior prediction distribution of cumulative damage, the failure risk probability of each future cycle is calculated, and the end of life is determined according to the dynamic risk threshold function. Based on the current number of cycles and the end of the lifespan, calculate and output the predicted range of the remaining number of cycles.

[0014] Furthermore, the physical damage evolution model is a numerical integral model of the physical equations describing fatigue crack propagation in a hydrogen environment.

[0015] Compared with the prior art, the beneficial effects of the present invention are: A hydrogen storage tank life prediction method based on multi-source sensing and machine learning is proposed. By setting a dynamic weighted fusion loss function and a two-stage transfer learning training strategy, the method achieves fast and stable convergence domain personalization adaptation of the model under conditions of very few online samples. By setting a dual filtering mechanism that combines probability-based physical occurrence probability correction and signal quality index into the attention mechanism, the reliability of the method is improved in real industrial noise environment.

[0016] Meanwhile, by setting up a physical information neural network and a parameterized probability output layer, the probability distribution of initial crack size and hydrogen embrittlement coefficient can be directly inferred from the measured strain signal, realizing the quantitative perception and early identification of unmeasurable parameters. By establishing a health status baseline decoupled from operating conditions and integrating it with a context-aware attention mechanism, the influence of varying operating conditions on features is effectively removed and adaptive weighting is achieved, improving the prediction accuracy and robustness of the model under different operating conditions. By constructing a full-process online chain from synchronous acquisition of multi-source signals to probabilistic lifetime prediction, the damage status in each hydrogen charge-discharge cycle is continuously tracked, and the remaining lifetime with confidence intervals is output, realizing the improvement from intermittent condition diagnosis to continuous predictive maintenance. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the hydrogen storage tank life prediction method of the present invention; Figure 2 This is a schematic diagram of the process for inverting the probability distribution of unmeasurable damage parameters according to the present invention. Figure 3 This is a schematic diagram of the health status baseline establishment and degradation deviation calculation method of the present invention; Figure 4 This is a schematic diagram of the attention network weighted fusion method for calculating the overall damage increment according to the present invention; Figure 5 This is a schematic diagram of the cumulative damage trajectory construction and remaining life prediction method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the present invention provides a technical solution: a method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning, comprising: S1. During each hydrogen charge / discharge cycle, pressure signals, strain signals, and acoustic emission signals are collected simultaneously.

[0020] It is important to note the following for pressure signal acquisition: An absolute pressure transmitter with a range covering the rated working pressure of the hydrogen storage tank should be used and installed on the main pipeline close to the tank body. Continuous acquisition should be performed at a sampling frequency of at least 10Hz to ensure differentiation between the pressurization, holding, and depressurization stages. For strain signal acquisition: Resistance strain gauges should be attached to the circumferential and axial sides of the outer wall of the hydrogen storage tank, preferentially placed in stress concentration areas (such as the transition zone of the end cap). The strain signal should be amplified by a dynamic strain gauge and synchronously acquired at a sampling frequency of at least 500Hz. For acoustic emission signal acquisition: At least two acoustic emission sensors with resonant frequencies in the range of 100-300kHz should be arranged on the tank surface. A reasonable detection threshold (e.g., 40dB) should be set to suppress environmental noise, and the arrival time, threshold, and energy parameters of acoustic emission events should be continuously acquired. The acquisition of these three signals should be initiated by the same trigger signal to ensure time synchronization.

[0021] The raw data for each loop is stored independently, and the data file must contain an identifiable loop count marker.

[0022] S2. Based on the measured strain signal, the probability distribution of unmeasurable damage parameters is inverted through a physical information neural network constrained by relevant physical equations. The unmeasurable damage parameters include degradation parameters such as initial crack size and hydrogen embrittlement coefficient, which cannot be directly measured by sensors but affect the fatigue life of materials.

[0023] Specifically: The first step is to collect the measured strain signal of the current cycle and construct it into a strain time series with equal interval sampling.

[0024] It is important to note that the measured strain signal of the current k-th cycle is resampled at equal intervals using a fixed sampling frequency (not less than 1 kHz) to obtain a strain time series ε of length T. k The sequence is then normalized. Where, μ ε and σ ε The global mean and standard deviation are calculated based on all historical health cycle strain signals.

[0025] The second step is to input the strain time sequence into a bidirectional long short-term memory network. The end of the network is connected to a parameterized probability output layer, which is used to directly output the probability distribution parameters of the initial crack size and hydrogen embrittlement coefficient.

[0026] It is important to note that a bidirectional long short-term memory (BiLSTM) network is constructed as the core feature encoder, with a parameterized probability output layer connected to its end. This output layer consists of two parallel fully connected sub-networks, which output the probability distribution hyperparameters of two unmeasurable damage parameters: for the initial crack size a0, the log-normal distribution parameter is output; for the hydrogen embrittlement coefficient D... h Output the parameters of its truncated Gaussian distribution (non-negative part).

[0027] The third step involves using the modified Paris formula, which takes into account the hydrogen embrittlement effect, as a physical constraint. Based on the probability distribution parameters, sampling is performed, and the theoretical crack propagation is calculated by substituting the parameters into the formula. This is then compared with the reference crack propagation calculated from the measured strain to construct a physical constraint loss term.

[0028] It is important to note that, in order to ensure that the network output conforms to physical laws, a modified Paris formula considering the hydrogen embrittlement effect is introduced as a differentiable physical constraint: Where C and m are material constants; ΔK is the stress intensity factor amplitude: Where Y is the geometric factor (usually taken as 1.12); Δσ is the stress amplitude converted from the strain amplitude: Where E is the elastic modulus and Δε is the strain amplitude.

[0029] From the probability distribution currently output by the network, sample M1 groups (e.g., 100 groups) of parameters (a0). (i) D h (i) ); for each group of sampling parameters (a0) (i) D h (i) Using this as an initial condition, the theoretical crack propagation Δa in the current k-th cycle is calculated by numerically integrating the aforementioned Paris formula. pred (i) .

[0030] A strain-crack propagation lookup table was established based on fatigue crack propagation test data of standard samples from the same batch of hydrogen storage tanks under simulated operating conditions. Using the strain amplitude Δε as the index, piecewise linear interpolation was employed to obtain the corresponding reference propagation rate, and integration yielded the reference crack propagation amount Δa. ref .

[0031] Physical loss term L p Defined as the root mean square error between the theoretical crack propagation and the reference crack propagation: The fourth step is to construct a total loss function that includes data fitting loss, physical constraint loss, and regularization loss. The weight of the physical constraint loss is dynamically increased during training according to a preset strategy.

[0032] It is important to note the total loss function L: Among them, L d The loss function is the sampled crack propagation amount Δa. perd (i) The response sequence was calculated by converting the crack length-strain relationship and compared with the measured strain time series ε. k Calculate the mean square error. The crack length-strain relationship is pre-established through finite element simulation or calibration experiments, and is a monotonic mapping curve of crack length and strain at measuring points on the outer wall of the hydrogen storage tank. λ(t) is the dynamic weight of the physical constraint loss term, which increases from 0 within the training period t according to the following rules: Where, λ max τ represents the maximum weight (e.g., 10), and τ is the growth time constant (e.g., 50 training epochs). This ensures rapid convergence in the early stages of training, primarily focusing on learning data features, while gradually strengthening the constraints of physical laws in the later stages, guiding the network to find a solution that both fits the data and conforms to physical laws. r To regularize the loss, a KL divergence constraint is applied to the standard deviation of the network output, making it approach a smaller prior distribution (e.g., N(0, 0.1)). 2 )); β is the weighting coefficient (e.g., 0.01).

[0033] Step 5: Train the network using a phased transfer learning strategy. Pre-train the network on a public dataset, freeze the network's feature extraction layer, and fine-tune the network's parameterized probability output layer online using a single sample from the current loop.

[0034] It is important to note that, to address the scarcity of data from individual hydrogen storage tanks during online monitoring, a two-stage transfer learning strategy is employed: Offline pre-training was performed on a publicly available hydrogen-induced fatigue crack propagation dataset (such as a subset of data on relevant materials in a hydrogen environment from the NASA FCG database) using a large number of sample pairs (strain sequence, crack propagation amount) to train the machine output layer of the BiLSTM network, enabling the network to learn a general mapping pattern from strain signals to damage parameters.

[0035] Online individual feature fine-tuning is performed for the specific hydrogen storage tank being monitored. All weights of the pre-trained BiLSTM feature encoder are frozen, preserving its learned general feature extraction capabilities. Only a single sample ε from the current k-th iteration is used. kAs a training set, mini-batch gradient descent training is performed on the parameterized probability output layer at the end; during fine-tuning, the dynamic weights λ(t) are reset and start growing again from 0 according to the above strategy.

[0036] Step 6: Output the joint posterior probability distribution of the initial crack size and hydrogen embrittlement coefficient.

[0037] It is important to note that after online fine-tuning is completed, the preprocessed... The input network is subjected to one forward propagation, and the distribution parameters are obtained from the parameterized probability output layer, which yields the initial crack size a0 and the hydrogen embrittlement coefficient D. h The joint posterior probability distribution.

[0038] Figure 2 This is a flowchart illustrating the method for inverting the probability distribution of unmeasurable damage parameters in this invention. Unlike existing technologies that can only invert macroscopic parameters (such as wall thickness), this invention uses a physical information neural network with a modified Paris formula as a physical constraint to invert the probability distribution of microscopic damage parameters such as initial crack size and hydrogen embrittlement coefficient, which cannot be directly measured. Furthermore, it employs a transfer learning strategy to address the problem of scarce data for individual hydrogen storage tanks.

[0039] S3. Based on the time-domain variation characteristics of the pressure signal, each cycle is automatically divided into a pressurization stage, a pressure holding stage, a pressure release stage, and a resting stage. Different original degradation characteristic values ​​of each stage are extracted and compared with the baseline of the healthy state to calculate the measured degradation deviation.

[0040] The method for dividing the stages includes: The first step is to smooth the collected pressure signals and calculate the pressure change rate at each sampling point.

[0041] It is important to note that the acquired raw pressure signal P(t) undergoes a five-point cubic smoothing filter to suppress high-frequency noise. Its first-order forward difference is calculated as the instantaneous pressure change rate dP / dt: Where Δt is the sampling interval of the pressure signal.

[0042] The second step is to obtain the target hydrogen charging pressure value for this cycle, and based on this value and the initial average pressurization rate, dynamically calculate the starting thresholds for the charging and depressurization stages.

[0043] It is important to note that the target hydrogen charging pressure value P for this cycle should be obtained from the hydrogen storage tank control system. t Calculate the average pressurization rate during the initial pressurization phase. : Where P0 is the initial pressure of the cycle (atmospheric pressure), t0 is the initial time, and ts The pressure first reached 0.1P. t The time point. Dynamically calculate the starting threshold T of the pressurization phase. c and the initial threshold T of the pressure relief section d : The coefficients a1 and a2 are calculated as follows: Wherein, w1 and w2 are preset proportional coefficients, determined through calibration experiments. The calibration method for w1 and w2 is as follows: during the baseline establishment phase of the hydrogen storage tank health status, several typical cycles are selected, and w1 and w2 are adjusted to make the phase division results consistent with the manual annotation results. The w1 and w2 values ​​that result in the highest division accuracy are taken as preset values.

[0044] The third step is to determine the start of the pressure-charging phase when the pressure change rate of several consecutive sampling points exceeds the starting threshold of the pressure-charging phase; and to determine the end of the pressure-charging phase and the start of the pressure-holding phase when the absolute value of the pressure change rate approaches zero and the pressure value reaches the set proportion of the target pressure value.

[0045] It should be noted that when dP / dt > T for a consecutive number of sampling points (e.g., 5), c When the pressure is determined to be at the beginning of the pressurization phase, the first sampling point that meets the conditions is recorded as the starting point of the pressurization phase; when |dP / dt|<ζ (ζ is a small positive number close to zero, such as 0.01) and the current pressure P(t)≥α×P t When α is between 0.95 and 0.98, the pressure-charging phase is considered to have ended. This point is recorded as the end point of the pressure-charging phase and also serves as the start point of the pressure-holding phase.

[0046] Step 4: When the pressure change rate of several consecutive sampling points is less than the starting threshold of the pressure relief section, it is determined that the pressure holding section ends and the pressure relief section begins.

[0047] It should be noted that when dP / dt < T for a consecutive number of sampling points (e.g., 5), d When the pressure holding phase ends, the first sampling point that meets the conditions is recorded as the end point of the pressure holding phase, and it is also used as the starting point of the pressure relief phase.

[0048] Step 5: When the pressure drops to near the ambient pressure and the absolute value of the pressure change rate approaches zero, determine that the pressure relief phase has ended and the settling phase has begun.

[0049] It should be noted that when P(t) ≤ P env And |dP / dt|<ζ(P) envWhen the ambient pressure (which can be set to 0.101 MPa) is reached, the pressure relief section is considered to have ended, and this point is recorded as the end point of the pressure relief section, and also as the starting point of the settling section.

[0050] Step 6: The resting period continues from the end of the depressurization period to the beginning of the pressurization period of the next cycle. If the start of the next cycle is not detected within the preset time window (e.g., 300s), it continues until the end of this data collection.

[0051] Step 7: Compare the duration of each stage with the adaptive time range of historical statistics, correct abnormal durations or mark the cycle as an abnormal cycle.

[0052] It is important to note that the duration of each stage should be calculated and compared with the adaptive time range [μ] of the historical statistics for this hydrogen storage tank. 历史 -3σ 历史 ,μ 历史 +3σ 历史 Compare the values. If the duration of the pressurization or depressurization phase exceeds the range, then according to P... t and The theoretical duration estimated using the ideal gas law is corrected; if the duration of the pressure holding or settling phase exceeds the range, the cycle is directly marked as an abnormal cycle, and its data is not included in subsequent feature extraction and lifetime prediction. This automatically filters out invalid data caused by operational abnormalities or signal loss.

[0053] It should be noted that the above-described method for dividing stages based on the rate of pressure change is only a preferred embodiment of the present invention. The same stage division objective can also be achieved using the second derivative of pressure, pressure difference fraction, pressure threshold segmentation, sliding window pressure variance, and other similar or modified methods.

[0054] like Figure 3 As shown, the present invention provides a method for establishing a health status baseline and calculating degradation deviation; Specifically: Establishing a baseline for health status includes: The first step is to collect complete sensor data for several consecutive cycles (e.g., 50) after the hydrogen storage tank is put into use, and select qualified cycles based on preset physical rules to form a healthy candidate cycle set.

[0055] It should be noted that the physical rule screening includes: the maximum cyclic pressure P. max ∈[0.97P t 1.03P t The duration of each stage is within a reasonable range based on experience (e.g., pressurization stage < 300s, pressure holding stage > 60s, pressure release stage < 300s); the selected cycles constitute a healthy candidate cycle set C.

[0056] The second step is to extract the original degradation feature values ​​for each stage of each healthy candidate cycle.

[0057] It is important to note that for each healthy candidate cycle, the raw degradation feature values ​​for the following stages are extracted: Pressurization section: peak strain, maximum strain rate, strain-pressure data point sequence; Holding pressure segment: creep slope, the linear fit slope of strain versus time during the last 1 / 3 of the holding pressure segment; residual strain increment, the difference between the strain at the end of the holding pressure segment and the strain at the beginning of the holding pressure segment; strain-time data point sequence; The pressure relief section, elastic recovery amount, is the difference between the strain at the beginning of the pressure relief section and the strain at the end of the pressure relief section; hysteresis loop area, is the area enclosed by the pressure relief section and the pressure charging section in this cycle on the strain-pressure plane; plastic strain ratio, is the ratio of the residual strain at the end of the pressure relief section to the peak strain of this cycle. The static settling period, zero-point drift, the difference between the strain at the end of the static settling period and the strain at the beginning of the static settling period; the strain relaxation data point sequence; and the strain fluctuation signal.

[0058] The third step is to extract the operating parameters of each healthy candidate cycle, establish a mapping model from the operating parameters to each original degradation feature value, correct the original degradation feature value, and obtain the standardized feature value.

[0059] It is important to note that the operating parameters include the average hydrogen charging rate and ambient temperature. For each extracted degradation feature, a Gaussian process regression model is constructed, using the operating parameters of all healthy candidate cycles as input and the corresponding original degradation feature value as output, to train the regression model. Using the trained regression model, the original degradation feature values ​​of all healthy candidate cycles are corrected to the preset standard rated operating conditions to obtain standardized feature values. This eliminates the influence of differences in operating conditions such as hydrogen charging rate and ambient temperature between different cycles on the feature values, making the degradation feature values ​​of all healthy candidate cycles comparable.

[0060] Step 4: For each degradation feature, calculate the sliding statistics (such as the sliding mean and sliding standard deviation) of its standardized feature value sequence; when the sliding statistics meet the preset convergence conditions, determine the sliding mean at the time of convergence as the healthy baseline value of the degradation feature, and record the sliding standard deviation at the time of convergence as the inherent fluctuation range of the degradation feature in the healthy state.

[0061] It is important to note that the sliding window size is a fixed value (e.g., 10 cycles). Each time a qualified cycle is added, the oldest cycle is removed, and the sliding statistics within the window are recalculated. Convergence is determined when the number of processed healthy candidate cycles is not less than the preset minimum sample size (e.g., 20 cycles), and within the sliding window of the most recent consecutive cycles (e.g., 10 cycles), the rate of change of the moving mean of the degradation feature is lower than a preset threshold (e.g., 0.5%), and the moving standard deviation is lower than a preset proportion of the moving mean (e.g., 1%).

[0062] Step 5: Based on the sequence of strain-pressure data points of all qualified cycles in the healthy candidate cycle set, align them and take the average value of each point to generate the standard strain-pressure curve of the pressurization section.

[0063] It is important to note that for the strain-pressure data point sequence of each qualified cycle in the healthy candidate cycle set C, the sequence of the first cycle is used as a reference, and the sequences of other cycles are dynamically time-warped and aligned (DTW alignment) to eliminate the phase shift caused by the difference in charging rate between cycles. After alignment, the arithmetic mean of the pressure value corresponding to each strain value is taken to obtain the standard strain-pressure curve, which serves as the curve shape benchmark for the strain-pressure relationship in the charging section.

[0064] Step 6: The health baseline values ​​of all degradation characteristics, the inherent fluctuation range, and the standard strain-pressure curve together constitute the health status baseline of the hydrogen storage tank. Once established, the health status baseline remains fixed as the absolute reference for all subsequent cycle performance degradation.

[0065] The degradation bias calculation is specifically as follows: for the j-th degradation feature value F extracted in the current loop... j The corresponding degradation bias D j The calculation method is as follows: If the degradation characteristic is proportional (e.g., hysteresis loop area ratio), then D j for: Among them, B j is the health baseline value for the j-th degradation characteristic.

[0066] If the degradation characteristic is of the difference type (such as peak strain, residual strain increment), then D j for: Where σ j Let be the standard deviation of the j-th degenerative feature in a healthy state.

[0067] If the degradation characteristic is a time constant or slope type (such as creep slope, relaxation time), then D j for: S4. Correct the measured degradation bias based on the probability distribution of unmeasurable damage parameters. Use an attention network to weight and fuse the modified degradation biases at each stage to output the comprehensive damage increment for this cycle.

[0068] like Figure 4 As shown, this invention provides a method for calculating the comprehensive damage increment using attention network weighted fusion; Specifically: The first step is to calculate the physical probability of occurrence of the measured degradation deviation at each stage extracted by S3 based on the posterior probability distribution of the unmeasurable damage parameters output by S2, and then correct the measured degradation deviation at each stage according to the physical probability of occurrence to obtain the physically reliable degradation deviation.

[0069] It is important to note that the unmeasurable damage parameters output from S2 (initial crack size a0 and hydrogen embrittlement coefficient D) h In the posterior probability distribution of , Monte Carlo sampling is performed to collect M² groups (e.g., 500 groups) of parameter samples. For each sampled parameter (a0... (i) D h (i) By substituting the j-th degradation feature into a pre-defined physical degradation model, the theoretical predicted value of the j-th degradation feature under the current damage state can be calculated. .

[0070] The physical degradation model is based on the modified Paris formula that takes into account the hydrogen embrittlement effect. It calculates the crack propagation through numerical integration and then converts the crack propagation into the theoretical prediction value of the corresponding degradation characteristics at each stage through a pre-established differentiable mapping relationship (such as the crack length and strain characteristic mapping function established through finite element simulation or calibration experiment).

[0071] Repeat the above calculation for all M2 groups to obtain the sample set of theoretical prediction values ​​for the j-th degradation feature, which constitutes the physical prediction distribution of the degradation feature.

[0072] The measured degradation deviation d of the j-th degradation feature calculated by S3 j Calculate its physical occurrence probability P phys，j Calculate the α-quantile Q of the physical prediction distribution of this feature. a and 1-a quantile Q 1-a Where a is a preset value (e.g., a=0.1), and the interval [Q] is... a Q 1-a [ ] represents a preset central interval, encompassing the theoretical range of most likely values ​​for this feature under the existing uncertainties in damage parameters. The physical probability of occurrence is defined as the measured degradation deviation d. j The corresponding measured value of degradation characteristic F jThe probability that a value lies outside the predetermined central interval of its physical prediction distribution. A specific calculation method is as follows: Based on the calculated probability of physical occurrence P phys，j The original measured degradation deviation d j Corrections are made to obtain the physical reliability degradation bias. : in, To correct the strength coefficient (default value is 0.5). When P phys,j When the value is greater than 0.7, the cycle is marked as a candidate cycle for physical anomalies and will be given special attention in subsequent analyses. Its data may be excluded from or reduced in weighting for this weighted fusion.

[0073] The second step involves inputting the physical reliability degradation bias, historical cumulative damage sequence, and original signal quality indicators of each stage of the current loop into the attention network, and outputting the attention weights corresponding to each stage.

[0074] It is important to note that before deploying the method of this invention, the attention network needs to be pre-trained offline using a publicly available, fully labeled experimental dataset of hydrogen storage tank degradation. This dataset should contain multiple sets of signal sequences (strain, pressure, acoustic emission), as well as the true damage increment labels (or cumulative damage labels) for each cycle obtained by experts based on disassembly detection or high-precision model inversion. Following steps S1-S3, the physical reliability degradation deviation vector, the original signal quality index, and the corresponding historical cumulative damage sequence for each cycle are simulated and generated. A lightweight fully connected neural network attention scoring network is constructed. Its input is the input vector of the current cycle, and its output is four unnormalized attention scores. The training objective is to use the attention weights normalized by Softmax to perform a weighted summation of the physical reliability degradation deviation, making the result as close as possible to the true damage increment label of the cycle. Through the above supervised learning, the attention network learns to determine which stage's features contribute the most to the current overall damage increment based on the current deviation pattern, signal quality, and damage evolution history. After training, the weight parameters of the network are fixed.

[0075] For the currently monitored hydrogen storage tank, the following operations are performed during online operation: For the current k-th iteration, obtain the physical reliability degradation deviation for each of the four stages, the historical cumulative damage sequence (padding with zero vectors if the historical sequence is insufficient), and the original signal quality indices (such as strain coefficient of variation and acoustic emission signal-to-noise ratio) for the four stages of the current iteration. Construct an independent input vector for each stage, containing the corrected deviation of that node, the signal quality of that stage, and the historical damage context shared by all stages. Stack the input vectors of the four stages to form an input matrix, which is then input into a pre-trained attention network. The network outputs an attention score for each stage. Softmax normalize all scores to obtain and output the attention weights corresponding to each stage of the current iteration.

[0076] Optionally, after running for a longer period, an online adaptive fine-tuning function can be enabled to better adapt the attention network to the individual characteristics of the current hydrogen storage tank. Specifically, historical data of the current hydrogen storage tank and the corresponding comprehensive damage increment are used as pseudo-labels. The last few layers of the pre-trained attention network are fine-tuned with a small learning rate and a small number of iterations, while the underlying network is frozen.

[0077] The third step is to use the attention weights to perform weighted fusion of the physical credibility degradation biases at each stage, calculate the comprehensive damage increment for this cycle, and calculate its uncertainty estimate.

[0078] It is important to note that the uncertainty estimate U of the overall damage increment is determined based on the dispersion of the posterior probability distribution, the dispersion of the attention weights, and the maximum value of the physical occurrence probability. The specific calculation method is as follows: Where ω1, ω2, and ω3 are weighting coefficients, and σ post Let H(w) be the standard deviation of the posterior probability distribution, and H(w) be the entropy of the attention weights. This represents the maximum probability of physical occurrence at each stage.

[0079] S5. Construct a cumulative damage trajectory based on the historical cumulative damage sequence. Using the complete probability distribution of unmeasurable damage parameters as a priori condition, learn the probability density function of cumulative damage through Gaussian process regression and use the preset quantile as the dynamic lifetime endpoint. Output the remaining number of cycles with confidence interval.

[0080] like Figure 5 As shown, the present invention provides a method for constructing cumulative damage trajectories and predicting remaining lifetime; Specifically: The first step is to construct a cumulative damage trajectory based on historical cumulative damage sequences.

[0081] It should be noted that the incremental damage from the current cycle output by S4 is added to the historical cumulative damage sequence, and a cumulative damage trajectory point set is constructed with the cycle number as the x-axis and the cumulative damage value as the y-axis.

[0082] If there are cycles marked as anomalous in the historical data, their corresponding damage increments will not be included in the construction of the cumulative damage trajectory, and the missing cumulative damage values ​​will be supplemented by linear interpolation.

[0083] The second step involves generating damage evolution sample trajectories based on the posterior probability distribution of unmeasurable damage parameters, using Monte Carlo sampling to drive a pre-defined physical damage evolution model. The expected value of these trajectories at each iteration is then calculated, resulting in a sequence of expected values, which serves as the mean function for the Gaussian regression process. The physical damage evolution model is a numerical integral model of the physical equations describing fatigue crack propagation in a hydrogen environment.

[0084] It is important to note that the average load spectrum under the standard operating conditions used to establish the baseline health status of the hydrogen storage tank is used as input for predicting the future. This load spectrum is the average of the stress intensity factor amplitude sequence of the healthy candidate cycle after operating condition decoupling.

[0085] From the current loop (nth) c In the posterior distribution output by step S2, sample M s Group (e.g., 1000 groups) of parameter samples {(a0) (s) D h (s) For each sample group (a0) (s) D h (s) ), with a0 (s) Let be the initial crack size, and for the nth (n>n) future crack size... c In the next iteration, the current crack size a is... (s) (n-1), D h (s) Substituting the average load spectrum into the modified Paris formula and integrating, the crack propagation Δa for this cycle is calculated. (s) (n), and update the crack size a (s) (n): The cumulative damage value D of the sample in the nth cycle is obtained by using a preset monotonic mapping function (which is calibrated by simulation or experiment to convert the crack size into a cumulative damage value). (s) (n).

[0086] Traverse M sA set of samples is used to obtain a set of cumulative damage sample trajectories. For each future cycle number n, the mean of these sample values ​​is calculated to obtain a mean sequence, which constitutes the prior mean function of Gaussian process regression.

[0087] The third step involves using the number of cycles as input, the cumulative damage sequence (the sequence of ordinate values ​​in the cumulative damage trajectory point set) as observations, the mean function as the prior mean, the monotonic kernel function as the covariance function, and adjusting the observation noise covariance with the uncertainty estimate output by S4, performing Gaussian process regression to obtain the posterior predicted distribution of cumulative damage corresponding to any future number of cycles.

[0088] The fourth step is to calculate the failure risk probability for each future cycle based on the posterior probability distribution of unmeasurable damage and the posterior prediction distribution of cumulative damage, and to determine the end of the lifespan based on the dynamic risk threshold function.

[0089] It is important to note that the failure threshold of a hydrogen storage tank is defined as a fixed critical cumulative damage value D. fail This value is determined based on the design standards, material properties, and safety specifications of the hydrogen storage tank, such as the cumulative damage value corresponding to the critical crack size determined through burst testing or non-destructive testing. Then, the failure risk probability P in the nth cycle is... fail (n) is defined as the probability that the cumulative damage exceeds the failure threshold.

[0090] The standard deviation σ(n) of the predicted cumulative damage distribution for the nth future cycle, obtained through the above steps, is used as a measure of prediction uncertainty. The dynamic risk threshold θ(n) is a monotonically increasing function of the prediction uncertainty σ(n), for example: Where θ0 is the basic risk threshold (e.g., 0.05). The penalty coefficient for uncertainty (>0), σ max It is a preset upper limit value.

[0091] Starting from the next iteration of the current loop, calculate the failure probability P for each future loop. fail (n) and the corresponding dynamic risk threshold θ(n). Condition P will be satisfied. fail The minimum number of cycles n for which (n) ≥ θ(n) is determined as the dynamic lifespan endpoint of the hydrogen storage tank. If P holds for all calculated future cycles... fail If (n) < θ(n), then the hydrogen storage tank is considered safe within the calculated prediction range.

[0092] Step 5: Based on the current number of cycles and the end of the lifespan, calculate and output the predicted range of the remaining number of cycles.

[0093] It is important to note that the remaining cycle count is the end of the lifespan minus the current number of cycles. If the end of the lifespan is lower than the current number of cycles, the hydrogen storage tank is determined to have exceeded its lifespan, and the remaining cycle count is output as 0. Multiple sets of parameters are sampled from the joint posterior probability distribution output by S2, and the above steps are repeated to obtain multiple estimates of the remaining cycle count. The mean of these estimates is used as a point estimate, and a preset quantile is used as the upper and lower limits of the confidence interval to output the predicted interval of the remaining cycle count.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for predicting the lifespan of hydrogen storage tanks based on multi-source sensing and machine learning, characterized in that, include: S1. Pressure signals, strain signals and acoustic emission signals are collected synchronously during each hydrogen charge-discharge cycle; S2. Based on the measured strain signal, the probability distribution of unmeasurable damage parameters is inverted through a physical information neural network constrained by relevant physical equations; the unmeasurable damage parameters include the initial crack size and the hydrogen embrittlement coefficient. S3. Based on the time-domain variation characteristics of the pressure signal, each cycle is automatically divided into a pressurization stage, a pressure holding stage, a pressure release stage, and a resting stage. Different original degradation characteristic values ​​of each stage are extracted and compared with the baseline of the healthy state to calculate the measured degradation deviation. S4. Correct the measured degradation bias based on the probability distribution of unmeasurable damage parameters. Use an attention network to perform weighted fusion of the modified degradation biases at each stage and use the weighted fusion result as the comprehensive damage increment output for this loop. S5. Construct a cumulative damage trajectory based on the historical cumulative damage sequence. With the complete probability distribution of unmeasurable damage parameters as a prior condition, learn the probability density function of cumulative damage through Gaussian process regression and use the preset quantile as the dynamic lifetime endpoint. Output the remaining number of cycles with confidence interval. S4 includes: S401. Based on the posterior probability distribution of the unmeasurable damage parameters output by S2, calculate the physical probability of occurrence of the measured degradation deviation at each stage extracted by S3, and correct the measured degradation deviation at each stage according to the physical probability of occurrence to obtain the physically reliable degradation deviation. S402. Input the physical reliability degradation deviation, historical cumulative damage sequence and original signal quality index of each stage of the current loop into the attention network, and output the attention weights corresponding to each stage. S403. The attention weights are used to perform weighted fusion of the physical credibility degradation deviations at each stage to calculate the comprehensive damage increment for this cycle and to calculate its uncertainty estimate. S5 includes: S501. Construct cumulative damage trajectory based on historical cumulative damage sequence; S502. Based on the posterior probability distribution of unmeasurable damage parameters, a preset physical damage evolution model is driven by Monte Carlo sampling to generate damage evolution sample trajectories, and the expected value of each cycle number is calculated to obtain the expected value sequence, which serves as the mean function of the Gaussian regression process. S503. Using the number of cycles as input, the cumulative damage sequence as the observed value, the mean function as the prior mean, and the monotonic kernel function as the covariance function, perform Gaussian process regression to obtain the posterior predicted distribution of the cumulative damage corresponding to any future number of cycles. S504. Based on the posterior probability distribution of unmeasurable damage and the posterior prediction distribution of cumulative damage, calculate the failure risk probability of each future cycle, and determine the end of life based on the dynamic risk threshold function. S505. Based on the current number of cycles and the end of the lifespan, calculate and output the predicted range of the remaining number of cycles.

2. The method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning according to claim 1, characterized in that: S2 includes: S201. Collect the measured strain signal of the current cycle and construct it into a strain time series with equal interval sampling; S202. Input the strain time sequence into a bidirectional long short-term memory network. The end of the network is connected to a parameterized probability output layer, which is used to directly output the probability distribution parameters of the initial crack size and hydrogen embrittlement coefficient. S203. Using the modified Paris formula that takes into account the hydrogen embrittlement effect as a physical constraint, sampling is performed based on the probability distribution parameters, and the theoretical crack propagation is calculated by substituting into the formula. The theoretical crack propagation is then compared with the reference crack propagation calculated by back-calculation of measured strain to construct a physical constraint loss term. S204. Construct a total loss function that includes data fitting loss, physical constraint loss and regularization loss, wherein the weight of the physical constraint loss is dynamically increased according to a preset strategy during training. S205. The network is trained using a staged transfer learning strategy. The network is pre-trained on a public dataset, the feature extraction layer is frozen, and the parameterized probability output layer of the network is fine-tuned online using a single sample from the current loop. S206. Output the joint posterior probability distribution of the initial crack size and hydrogen embrittlement coefficient.

3. The method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning according to claim 1, characterized in that: The method for phase division in S3 includes: S301. Smooth the collected pressure signal and calculate the pressure change rate at each sampling point; S302. Obtain the target hydrogen charging pressure value for this cycle, and dynamically calculate the starting thresholds for the charging and depressurization stages based on this value and the initial average pressurization rate. S303. When the pressure change rate of several consecutive sampling points exceeds the starting threshold of the pressurization stage, the pressurization stage is determined to begin; when the absolute value of the pressure change rate approaches zero and the pressure value reaches the set proportion of the target pressure value, the pressurization stage is determined to end and the pressure holding stage is determined to begin. S304. When the pressure change rate at several consecutive sampling points is less than the starting threshold of the pressure relief section, the pressure holding section is determined to end and the pressure relief section is determined to begin. S305. When the pressure value drops to near the ambient pressure and the absolute value of the pressure change rate approaches zero, it is determined that the pressure relief phase ends and the settling phase begins. S306. The static period continues from the end of the depressurization period to the beginning of the pressurization period of the next cycle. If the start of the next cycle is not detected within the preset time window, it continues until the end of this data collection. S307. Compare the duration of each stage with the adaptive time range of historical statistics, correct abnormal durations or mark the cycle as an abnormal cycle.

4. The method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning according to claim 1, characterized in that: The original degradation feature values ​​extracted in each stage of S3 include: peak strain, maximum strain rate, and strain-pressure data point sequence extracted in the pressurization stage; creep slope, residual strain increment, and strain-time data point sequence extracted in the pressure holding stage; elastic recovery, hysteresis loop area, and plastic strain ratio extracted in the depressurization stage; and zero-point drift, strain relaxation data point sequence, and strain fluctuation signal extracted in the static stage.

5. The method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning according to claim 1, characterized in that: The establishment of the health status baseline in S3 includes: S308. Collect complete sensor data for several consecutive cycles after the hydrogen storage tank is put into use, and select qualified cycles based on preset physical rules to form a healthy candidate cycle set. S309. For each healthy candidate cycle, extract the original degradation feature values ​​of each stage; S310. Extract the operating condition parameters of each healthy candidate cycle, establish a mapping model from the operating condition parameters to each original degradation feature value, correct the original degradation feature value, and obtain the standardized feature value. S311. For each degradation feature, calculate the moving mean and moving standard deviation of its standardized feature value sequence; when the moving statistic meets the preset convergence condition, determine the moving mean at the time of convergence as the healthy baseline value of the degradation feature, and record the moving standard deviation at the time of convergence as the inherent fluctuation range of the degradation feature in the healthy state. S312. Based on the sequence of strain-pressure data points of all qualified cycles in the healthy candidate cycle set, after alignment, take the average value of each point to generate the standard strain-pressure curve of the pressurization section. S313, the health baseline values ​​of all degradation characteristics, the inherent fluctuation range, and the standard strain-pressure curve together constitute the health status baseline of the hydrogen storage tank.

6. The method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning according to claim 1, characterized in that: The method for calculating the probability of physical occurrence includes: sampling based on the posterior probability distribution of unmeasurable damage parameters, generating a predicted distribution of degradation characteristics at each stage through a physical degradation model; and calculating the probability that the measured degradation deviation is located outside the preset central interval of its corresponding predicted distribution as the probability of physical occurrence.

7. The method for predicting the lifespan of a hydrogen storage tank based on multi-source sensing and machine learning according to claim 1, characterized in that: The physical damage evolution model is a numerical integral model of the physical equations describing fatigue crack propagation in a hydrogen environment.

Images