Method for evaluating fatigue life of main girder based on sample continuous increment

Abstract

The present application relates to the technical field of engineering structure health monitoring, and discloses a method for main beam fatigue life evaluation based on sample continuous increment, comprising: collecting multi-modal monitoring data, including physical parameters of strain, displacement and acceleration, and external factor data of environmental temperature and humidity and traffic information; performing multi-source data fusion by using an adaptive feature extraction network; constructing a dual knowledge consolidation incremental learning framework to maintain the memory of historical knowledge when processing new data; quantifying the uncertainty of the prediction result based on Bayesian deep learning, and establishing a risk warning mechanism; the present application effectively solves the problem of catastrophic forgetting in the model updating process, maintains a high historical knowledge retention rate; provides a scientific basis for bridge maintenance and repair decision-making, optimizes maintenance resource allocation; has excellent adaptability and generalization ability, and realizes continuous monitoring and evaluation within the whole life cycle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge engineering structure health monitoring, more particularly, it relates to a method for evaluating fatigue life of a main girder based on sample continuous increment. BACKGROUND

[0002] With the rapid development of the transportation system and the wide construction of large-scale bridge engineering, the safety performance and service life of bridge structures have become the focus of the engineering and academic circles. Under the long-term action of complex loads, fatigue damage often occurs in the main girder structure of the bridge, which has the characteristics of accumulation, concealment and suddenness. Once it develops to the critical state, it may lead to the failure or even collapse of the bridge structure. Therefore, it is of great engineering practical significance to accurately evaluate the fatigue life of the main girder of the bridge.

[0003] Traditional fatigue life evaluation of the main girder of the bridge mainly relies on a single data source or a simplified model, which is difficult to fully capture the performance change law of the structure in the actual complex environment. With the development of sensing technology and monitoring system, a large amount of multi-source heterogeneous data is collected and stored, including physical parameters such as strain, displacement and acceleration, as well as external factor data such as environmental temperature and humidity and traffic information. How to make full use of these multi-modal data to build a more accurate and reliable fatigue life evaluation model has become a difficult and hot point in current research.

[0004] In addition, new monitoring data is continuously generated during the service process of the bridge structure, and these incremental data contains important information about the evolution of the structure performance. In the dynamic data environment, the evaluation model needs to have the ability of continuous learning, which can not only absorb the knowledge in the new data, but also maintain the memory of the historical data, so as to realize the real-time updating and precision improvement of the fatigue life evaluation. At the same time, due to the complexity and uncertainty of the structure fatigue process, the reliability quantification of the evaluation result is also a key link to ensure the safe operation of the bridge and the scientific maintenance decision. SUMMARY

[0005] The present application provides a method for evaluating fatigue life of a main girder based on sample continuous increment, which solves the technical problems of insufficient data utilization, poor model adaptability, insufficient uncertainty quantification and difficulty in providing reliable basis for maintenance decision in related technologies.

[0006] The present application provides a method for evaluating fatigue life of a main girder based on sample continuous increment, which includes the following steps:

[0007] The multi-source sensors arranged at the key positions of the main girder are used to collect strain, acceleration, displacement, image, environmental and traffic data, and an innovative adaptive wavelet packet transform denoising method is used to denoise the collected data;

[0008] The data after noise reduction is fused by an innovative multi-dimensional adaptive attention mechanism to generate a fused feature representation.

[0009] Based on the fused features, a multi-dimensional sample representation scoring algorithm is used for dynamic management of the memory bank, and an incremental learning method combining a progressive neural network lateral connection mechanism and feature-level and relation-level dual knowledge distillation is used to update the prediction model.

[0010] Using the updated prediction model, a multi-scale time series fatigue damage model of the environmental self-adaptive damage evolution equation is used for damage evolution calculation, a Bayesian neural network with four-dimensional uncertainty quantification is used for life prediction, and finally four weighted risk index calculations and dynamic threshold adjustments are used for graded early warning.

[0011] In a preferred embodiment, the innovative adaptive wavelet packet transform noise reduction method specifically includes:

[0012] Decomposition layer determination algorithm:

[0013] J = J base + δ SNR + δ λ ;

[0014] wherein J represents the final determined wavelet packet decomposition layer, i.e. the number of times the signal will be recursively decomposed; J base represents the basic decomposition layer; δ SNR represents the signal-to-noise ratio adaptive adjustment factor, which is set according to the signal quality section, and when the signal-to-noise ratio SNR > 15 dB, it is taken as one, when 10 dB ≤ SNR ≤ 15 dB, it is taken as zero, and when SNR < 10 dB, it is taken as -1; δ λ represents the signal complexity adjustment factor, which reflects the signal spectrum complexity improvement Shannon entropy criterion adopts an entropy value calculation method enhanced by a weight factor, and the weight factor is determined by the correlation between the feature and the target signal, and the adjustment coefficient is set to 0.3;

[0015] The semi-soft threshold function uses a three-section piecewise processing method, and uses a gradual processing in the transition interval to avoid the discontinuity of the hard threshold value, and ensures the smoothness of the signal reconstruction.

[0016] In a preferred embodiment, the innovative multi-dimensional adaptive attention mechanism specifically includes:

[0017] The modal quality score is calculated by a sigmoid function, which maps the result of linear transformation of the weight matrix to the quality score interval of zero to one;

[0018] The temperature adaptive adjustment algorithm dynamically adjusts the temperature parameter according to the variance of the query matrix and the key matrix, and the basic temperature parameter is set to 1.0, and the variation coefficient adjustment factor is set to 0.5;

[0019] The enhanced attention weight calculation adopts an improved softmax function, which considers the comprehensive influence of the original attention score, inter-modal correlation and quality score, and introduces a numerical stability parameter to avoid calculation overflow.

[0020] In a preferred embodiment, the multi-dimensional sample representation scoring algorithm specifically includes:

[0021] The comprehensive scoring formula combines four dimensions of diversity, difficulty, uncertainty and time importance, and adopts a dynamic weight distribution mechanism.

[0022] The diversity weight in the dynamic weight distribution is adaptively adjusted according to the current memory bank filling degree, and the maximum capacity of the memory bank is set to 1000 samples.

[0023] The time importance is calculated by an exponential decay function, the time decay parameter is set to 0.01, and the diversity calculation adopts a semantic similarity adjustment factor enhancement algorithm.

[0024] In a preferred embodiment, the progressive neural network lateral connection mechanism specifically includes:

[0025] The layer output calculation considers both the internal connection of the column and the lateral connection of all previous sequences, and realizes the transmission of historical knowledge through weighted summation;

[0026] The basic column maintains the fixed parameters of the initial training, containing five fully connected layers, as the invariant storage of historical knowledge.

[0027] The extended column is dynamically increased with new tasks, each column has the same structure as the basic column, the lateral connection weight matrix transmits historical knowledge, and the learning rate is set to 0.1 times that of the main network.

[0028] In a preferred embodiment, the double knowledge distillation specifically includes:

[0029] The feature-level knowledge distillation loss calculates the Euclidean distance after normalizing all intermediate layer features, maintaining the consistency of new and old models at the feature representation level.

[0030] The relationship knowledge distillation loss maintains the consistency of the relative relationship between samples by calculating the difference in similarity relationship between samples.

[0031] The distillation temperature parameter is set to 2.0, the distillation weight is set to 0.4, the total loss function combines the task loss, feature distillation loss and relationship distillation loss, and the weight ratio is 6:3:1.

[0032] In a preferred embodiment, the environment adaptive damage evolution equation specifically includes:

[0033] Multi-scale damage evolution equation:

[0034]

[0035] wherein, Ds(t) represents the s-th scale damage degree Ds(t) s the rate of change over time t; Ds(t) s Ds(t) represents the s-th scale damage degree at time t, with a value range of [0, 1], 0 representing no damage and 1 representing complete damage; αs s Ds(t) represents the s-th scale damage evolution rate coefficient; βs s Ds(t) represents the s-th scale damage nonlinearity index; σs eff Ds(t) represents the effective stress at time t; γs s Ds(t) represents the s-th scale stress sensitivity index; ηs s Ds(t) represents the scale coupling coefficient; Ds s-1 Ds(t) represents the s-th scale damage degree at time t, reflecting the multi-scale coupling effect; Φs env Ds(t) represents the environmental influence factor; Ψs nonlinear Ds(t) represents the s-th scale damage degree; Ds(t) s Ds(t) represents the nonlinear damage acceleration factor;

[0036] The effective stress calculation combines the square roots of static stress and dynamic stress, the dynamic stress enhancement coefficient is set to 1.2, and the multi-scale comprehensive damage degree weight distribution is macro 0.5, meso 0.3, and micro 0.2.

[0037] In a preferred embodiment, the four-dimensional Bayesian neural network of uncertainty specifically comprises:

[0038] The overall uncertainty is decomposed into four dimensions of random uncertainty, cognitive uncertainty, distribution uncertainty, and time sequence uncertainty, with the addition of the time sequence uncertainty dimension to the traditional three dimensions;

[0039] The cognitive uncertainty is calculated by the Bayesian estimation method, with one hundred times of Monte Carlo sampling combined with the Dropout variance adjustment mechanism, and the adjustment coefficient is set to 0.5;

[0040] The time sequence uncertainty is calculated by weighted time sequence difference, with exponential decay for the time sequence weight, and the periodic adjustment factor considering the twenty-four hour periodic variation law.

[0041] In a preferred embodiment, the risk index calculation and dynamic threshold adjustment specifically comprise:

[0042] The risk index calculation formula is:

[0043]

[0044] wherein, RI represents the risk index; L represents the current measured structural characteristic parameter; L ref represents the reference value of the structural characteristic parameter; denotes the relative degree of deviation of the structural characteristic parameter from the reference value; σ L denotes the standard deviation of the structural characteristic parameter; σ ref denotes the reference value of the standard deviation of the structural characteristic parameter; denotes the relative uncertainty of the structural parameter fluctuation, and the weight is 0.3; D denotes the current cumulative damage degree; D cr denotes the critical damage degree threshold value; denotes the degree of damage close to the critical state, and the weight is 0.2; L t denotes the structural characteristic parameter value at the current time t; L t-1 denotes the structural characteristic parameter value at the previous time t-1; denotes the relative change rate of the structural parameter;

[0045] The dynamic threshold adjustment is corrected according to the traffic importance of the bridge and the load use intensity, and the adjustment coefficients are both 0.1.

[0046] The fourth-level early warning threshold is set as 0.3 for the safety level, 0.5 for the attention level, 0.7 for the warning level and 0.9 for the danger level, corresponding to the annual, semi-annual, quarterly and monthly inspection frequencies.

[0047] In a preferred embodiment, a computer readable storage medium is used to store computer readable instructions capable of running a method for evaluating the fatigue life of a main girder based on a sample continuous increment when the computer readable instructions are read by a computer.

[0048] The beneficial effects of the present application are:

[0049] The bridge main girder fatigue life evaluation method based on a sample continuous increment realizes high-precision prediction of the fatigue life of the bridge main girder through multi-modal data fusion and double-knowledge consolidation incremental learning mechanism. Compared with the traditional evaluation method, the present application significantly improves the prediction accuracy, controls the life prediction error in a smaller range, and significantly narrows the confidence interval of the prediction result, providing a more reliable basis for engineering decision-making.

[0050] In terms of model performance, the present application effectively solves the catastrophic forgetting problem in the model updating process through an innovative incremental learning strategy, and maintains a high historical knowledge retention rate. At the same time, the optimized feature extraction and fusion mechanism significantly enhances the expression ability of the model, enabling it to capture subtle change characteristics in the structural fatigue evolution process, thereby accurately identifying abnormal working conditions and potential risks.

[0051] The application also quantifies the uncertainty of the prediction results through a Bayesian deep learning framework, establishes a warning mechanism based on risk levels, and provides a scientific basis for bridge maintenance and repair decisions. This decision support system based on uncertainty quantification effectively avoids the two extreme cases of over-repair and under-repair, and optimizes the allocation of maintenance resources.

[0052] In addition, the application shows excellent adaptability and generalization ability in actual engineering applications, can adjust the evaluation strategy in real time according to environmental changes and structural performance evolution, and realizes continuous monitoring and evaluation throughout the life cycle. This dynamic evaluation mechanism not only improves the safety of bridge operation, but also provides important technical support for bridge design optimization and life extension. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of the method for evaluating the fatigue life of the main girder based on continuous increments of samples of the application. DETAILED DESCRIPTION

[0054] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the content of this specification. Various processes or components can be omitted, replaced, or added according to need in various examples. In addition, the features described in some examples can also be combined in other examples.

[0055] A method for evaluating the fatigue life of the main girder based on continuous increments of samples is disclosed in at least one embodiment of the application, as shown in Figure 1 includes the following steps:

[0056] Step 1: Collect strain, acceleration, displacement, image, environmental, and traffic data through multi-source sensors deployed at key locations of the main girder, and use an innovative adaptive wavelet packet transform noise reduction method to reduce the collected data.

[0057] Specifically, it includes the following steps:

[0058] Step 1.1, sensor deployment and data collection;

[0059] The sensor network is deployed at key locations of the main girder, specifically:

[0060] Strain gauges (model: BFG-120-2AA): Deployed in stress concentration areas such as the midspan and support areas of the main girder, 3 to 5 along the height of each key section, with a total of not less than 24;

[0061] Acceleration sensors (model: PCB393B12): 3 per span, placed at mid-span and 1 / 4 span, frequency response range 0.15-1000 Hz;

[0062] Displacement meters (model: HS-WY100): placed at mid-span, measurement range 0-100 mm, accuracy 0.01 mm;

[0063] High-definition cameras (resolution 4K): for fixed-point monitoring of suspected crack areas, field of view angle adjustable, equipped with LED auxiliary light source;

[0064] Environmental monitoring equipment: including thermometer (-40℃ to 85℃), hygrometer (0 to 100% RH), anemometer (0 to 60 m / s);

[0065] Traffic monitoring equipment: including vehicle classification counter and vehicle weight detection system.

[0066] The data acquisition system consists of a central data acquisition device, a wireless transmission module, and a cloud storage system, with the following acquisition frequency settings:

[0067] Static data (strain, displacement): collected every hour, each for 1 minute, sampling rate 10 Hz;

[0068] Dynamic data (acceleration): high-frequency collection at fixed time periods (10:00-10:10 and 15:00-15:10) every day, sampling rate 100 Hz;

[0069] Image data: collected once a week, when light conditions are good;

[0070] Environmental data: collected every hour, sampling rate 1 Hz;

[0071] Traffic data: continuously recorded, 24 hours uninterrupted.

[0072] All data is transmitted to the cloud server through a wireless network and stored in a distributed database indexed by time, location, and data type.

[0073] Step 1.2, data preprocessing;

[0074] A combined method based on Isolation Forest and Long Short-Term Memory (LSTM) network is used for anomaly detection:

[0075] Isolation Forest anomaly detection:

[0076] 100 isolation trees are constructed, each tree isolates sample points by randomly splitting the space;

[0077] For each sample point, calculate its anomaly score. The anomaly score calculation steps are as follows:

[0078] Calculate the average path length E(h(x)) of the sample in all isolation trees;

[0079] Calculate the normalization factor c(n), where n is the total number of samples;

[0080] Map the path length to the [0, 1] interval using the exponential conversion function to get the anomaly score;

[0081] When the anomaly score is greater than 0.6, it is determined to be an outlier.

[0082] LSTM anomaly detection:

[0083] Use a double-layer LSTM network (64 units + 32 units) to predict the next time value;

[0084] Calculate the deviation of the actual value and the predicted value, and when the deviation exceeds the mean plus 3 times the standard deviation, it is determined to be abnormal;

[0085] The innovative adaptive wavelet packet transform method is used for data denoising:

[0086] Innovative adaptive decomposition layer determination algorithm:

[0087] Basic decomposition layer calculation:

[0088]

[0089] Where L is the signal length (number of sampling points), is the floor function, J base is the basic decomposition layer, and log is the logarithmic function.

[0090] Signal-to-noise ratio adaptive adjustment factor:

[0091]

[0092] Where SNR is the signal-to-noise ratio; δ SNR is the layer increase or decrease value dynamically adjusted according to the signal-to-noise ratio.

[0093] Signal complexity adjustment factor:

[0094]

[0095] Where λ max and λ min are the maximum and minimum eigenvalues of the signal spectrum, respectively, representing the frequency complexity of the signal, δ λ is the layer increase or decrease value dynamically adjusted according to the signal complexity, and log is the logarithmic function.

[0096] Final innovation formula:

[0097] J = J base + δ SNR + δ λ ;

[0098] where J is the final determined wavelet packet decomposition layer number, considering the signal length, signal-to-noise ratio and signal complexity; J base is the basic decomposition layer number; δ SNR is the layer number increase or decrease value dynamically adjusted according to the signal-to-noise ratio; δ λ is the layer number increase or decrease value dynamically adjusted according to the signal complexity.

[0099] Innovation effect: Compared with the traditional fixed layer number method, the decomposition precision is improved by 15-25%, and the noise suppression ability is enhanced by 30%.

[0100] Improved Shannon entropy criterion is used:

[0101] H improved = -Σ i p i log2(p i )×w i ;

[0102] where H improved represents the improved Shannon entropy value, used to evaluate the information amount of wavelet packet nodes; p i represents the energy probability distribution of the i-th wavelet coefficient; log2(p i ) represents the logarithm function with base 2, used to calculate the information amount; w i represents the weight factor, used to adjust the importance of each coefficient.

[0103] w i = 1 + α × corr(x i ,x target );

[0104] where w i represents the weight factor of the i-th wavelet coefficient; corr(x i ,x target ) represents the correlation coefficient of the i-th wavelet coefficient and the target signal, with a value range of [-1, 1]; α represents the correlation influence adjustment coefficient, with a value of 0.3, controlling the influence degree of correlation on weight.

[0105] Adaptive threshold determination: dynamically adjust the threshold parameter based on the local characteristics of the signal.

[0106] Innovative semi-soft threshold function:

[0107]

[0108] wherein, denotes the wavelet coefficient after threshold processing, i.e. the coefficient value after noise reduction; w i denotes the weight factor of the i-th wavelet coefficient; sign(w i ) denotes the sign function of the coefficient w i , which takes the value 1 when w i > 0 and -1 when w i < 0; |w i | denotes the absolute value of the coefficient w i ; λ i denotes the adaptive threshold parameter, which is dynamically calculated according to the local characteristics of the signal, and is used to distinguish between signal and noise; σ i denotes the local standard deviation of the i-th coefficient, reflecting the local fluctuation degree of the signal.

[0109] The first case: when the absolute value of the coefficient is much larger than the threshold, soft threshold shrinkage processing is adopted;

[0110] The second case: when the absolute value of the coefficient is close to the threshold, nonlinear smooth transition processing is adopted;

[0111] The third case: when the absolute value of the coefficient is much smaller than the threshold, it is considered as noise and is set to zero.

[0112] Step 1.3, feature extraction;

[0113] Extract multi-level features:

[0114] Time domain features: mean, standard deviation, peak value, skewness, kurtosis, root mean square value, peak factor, etc.

[0115] Frequency domain features: main frequency, power spectral density, frequency band energy ratio, etc. are extracted by fast Fourier transform.

[0116] Time-frequency domain features:

[0117] Use continuous wavelet transform to extract time-frequency features;

[0118] Select Morlet wavelet as the basis function;

[0119] Extract wavelet energy, scale energy distribution, wavelet entropy, etc.

[0120] Step 2, for the noise-reduced data, innovative multi-dimensional adaptive attention mechanism is adopted for multi-modal feature fusion to generate fused feature representation;

[0121] Specifically, the following steps are included:

[0122] Step 2.1, innovative multi-dimensional adaptive attention mechanism;

[0123] For the multi-level features extracted in step 1, an innovative multi-dimensional adaptive attention mechanism is used for feature fusion. This is one of the core innovations of the invention, which includes the following breakthrough technologies:

[0124] Innovative technical features:

[0125] Space-time coupled attention: considering both time series characteristics and spatial correlation;

[0126] Quality-aware weight distribution: dynamically adjusting attention weights based on data quality;

[0127] Adaptive temperature regulation: introducing temperature parameters to optimize attention distribution;

[0128] Multi-head collaborative enhancement: enhancing feature expression ability through multi-head attention mechanism.

[0129] Calculate the quality score q m of each modality, the calculation steps are:

[0130] Linear transformation on the feature vector z m of each modality: multiply the feature vector with the weight matrix W q ;

[0131] Add the bias vector b q ;

[0132] Use the sigmoid function to map the result to the interval [0,1[ to get the quality score of the modality.

[0133] Calculate the correlation R ij between any two modalities i and j, the steps are:

[0134] Extract the feature vectors z i and z j of the two modalities;

[0135] Calculate the inner product of the two vectors;

[0136] Calculate the L2 norm of the two vectors respectively;

[0137] Divide the inner product by the product of the two norms to get the cosine similarity as the correlation coefficient.

[0138] Innovative temperature regulation attention weight calculation: for each pair of modalities i and j, use the innovative temperature regulation mechanism to calculate the attention weight A ij :

[0139] Innovative temperature adaptive adjustment algorithm:

[0140] τ = τ base × (1 + β × Var(Q i , K j ));

[0141] where τ base = 1.0 is a base temperature parameter controlling the smoothness of attention distribution, the larger the value the more uniform the attention distribution; β = 0.5 is a variance coefficient adjustment factor used to adjust the influence strength of variance on temperature parameter; Var(Q i , K j ) is the variance index between query matrix Q i and key matrix K j , reflecting the difference degree of feature distribution between two modalities; τ is an adaptive temperature parameter dynamically adjusted according to feature distribution difference, used to optimize attention weight calculation.

[0142] Enhanced attention weight calculation steps:

[0143] Generate query matrix Q i and key matrix K j ;

[0144] Calculate original attention score:

[0145]

[0146] where S raw represents the original attention score matrix between modality i and modality j; Q i represents the query matrix of modality i, containing the feature representation of this modality; represents the transpose of the key matrix of modality j; d k represents the dimension of the key vector, used to scale the attention score to prevent gradient vanishing problem; represents the scaling factor, used to stabilize the numerical range of attention score.

[0147] Introduce relevance and quality enhancement:

[0148] S enhanced = S raw × (1 + 0.5R ij ) × (q i + q j ) / 2;

[0149] where S enhanced represents the attention score after relevance and quality enhancement; R ij represents the correlation coefficient between modality i and modality j, ranging from [-1, 1]; (1 + 0.5R ij ) represents the relevance enhancement factor, which enhances attention when two modalities are positively correlated, and weakens attention when two modalities are negatively correlated; qi and q j denote the quality score of modality i and modality j respectively, ranging from [0, 1]; q i +q j ) / 2 denotes the quality enhancement factor, adjusting the attention weight based on the average quality score of two modalities.

[0150] Apply adaptive temperature regulation:

[0151] S temp = S enhanced / τ;

[0152] where S temp denotes the attention score after temperature regulation; τ denotes the adaptive temperature parameter, controlling the smoothness of attention distribution; smaller τ value will make the attention distribution more concentrated (more "sharp"), and larger τ value will make the attention distribution more uniform (more "smooth").

[0153] Use improved softmax function:

[0154]

[0155] where A ij denotes the final attention weight matrix, used for feature fusion; exp(S temp ) denotes the exponential transformation of the temperature-regulated score; Σ k exp(S temp,k ) denotes the exponential sum of all possible attention scores, used for normalization; ∈ = 1e -8 denotes the numerical stability parameter, preventing calculation errors caused by zero denominator; the normalized A ij value ranges from [0, 1], and the sum of all values is close to 1.

[0156] Innovative effect: compared with traditional attention mechanism, the feature fusion accuracy is improved, and the calculation stability is enhanced.

[0157] Calculate the fused feature vector z fused : the steps are:

[0158] For all modality pairs (i, j), calculate the product of the attention weight A ij and the value matrix V j

[0159] Sum all the results to get the fused feature.

[0160] Step 2.2, multi-head attention and gating mechanism;

[0161] Introduce 8-head attention mechanism: ​

[0162] MultiHead(Z) = Concat(head1, head2,..., head8) W O ;

[0163] where MultiHead(Z) denotes the output feature matrix of multi-head attention, which integrates information from different attention heads; head1, head2, head8 represent the outputs of the 1st, 2nd, and 8th attention heads, respectively, each of which focuses on different aspects of the input data; Concat denotes the concatenation operation, which concatenates the outputs of the 8 attention heads in the feature dimension; W O denotes a learnable output projection matrix, which is used to map the concatenated features to the required output dimension.

[0164] Selective integration using a gating mechanism:

[0165] g m = σ(W g z m +b g );

[0166]

[0167] where g m denotes the gating vector of the mth modality, with a value range of [0, 1], controlling the amount of features passing through; σ denotes the sigmoid activation function, which maps the output to the interval [0, 1]; W g denotes the gating weight matrix, which is used for linear transformation of the feature vector; b g denotes the gating bias vector; z m denotes the feature vector of the mth modality; ⊙ denotes element-wise multiplication (Hadamard product), which realizes the selective passing of features; z gated denotes the fused feature vector after the gating mechanism; M denotes the total number of modalities, representing the number of different types of sensor data.

[0168] Final fused features:

[0169] z final = 0.7·z fused + 0.3·z gated ;

[0170] where z final denotes the final fused feature vector, which serves as the input for subsequent tasks; z fused denotes the feature vector fused through the attention mechanism; z gated denotes the feature vector fused through the gating mechanism.

[0171] Step 3, based on the fusion features, the memory bank is dynamically managed by a multi-dimensional sample representativeness scoring algorithm, and the prediction model is updated by combining an incremental learning method of progressive neural network lateral connection mechanism and feature-level and relation-level double knowledge distillation;

[0172] Specifically, the following steps are included:

[0173] Step 3.1, memory bank dynamic management;

[0174] The fusion features obtained in step 2 are subjected to memory bank dynamic management:

[0175] Innovative multi-dimensional sample representativeness scoring algorithm:

[0176] Dynamic weight distribution mechanism:

[0177]

[0178] w diff (t)=0.3+0.2×ACC recent ;

[0179] w unc (t)=1-w div (t)-w diff (t);

[0180] wherein w div (t) represents a diversity weight coefficient, which dynamically changes over time and decreases when the memory bank is close to full capacity; w diff (t) represents a difficulty weight coefficient, which dynamically changes over time and increases the weight of difficult samples when the model accuracy improves; w unc (t) represents an uncertainty weight coefficient, ensuring that the sum of the three weights is 1, automatically balancing the weights of diversity and difficulty; |M current | represents the number of samples in the current memory bank, used to calculate the filling rate of the memory bank; M max =1000 represents the maximum capacity of the memory bank, limiting the total number of storable samples; ACC recent represents the model accuracy of the last 100 samples, ranging from 0 to 1, used to evaluate the current performance of the model.

[0181] Enhanced scoring formula:

[0182] Score(x i )=w div (t)×Diversity(x i )+w diff (t)×Difficulty(x i )+w unc (t)×Uncertainty(x i)+ γ x Temporal(x i );

[0183] where Score(x i ) represents the comprehensive representativeness score of sample x i , ranging from [0, 1], the higher the score, the more important the sample is to the model training; Diversity(x i ) represents the diversity index of the x i , measuring the difference between the sample and other samples in the memory bank; Difficulty(x i ) represents the difficulty index of sample x i , measuring the difficulty of the current model in predicting the sample; Uncertainty(x i ) represents the uncertainty index of sample x i , measuring the uncertainty of the model in predicting the sample; Temporal(x i ) represents the time decay function, calculating the time freshness of the sample, and the score of the new sample is higher; γ = 0.1 represents the time importance adjustment factor, controlling the influence degree of time factor on the sample score; w div (t), w diff (t), w unc (t) respectively represent the dynamic weight coefficients of the diversity index, the difficulty index and the uncertainty index.

[0184] Innovative algorithm for diversity calculation:

[0185]

[0186] where Diversity(x i ) represents the diversity score of sample x i , ranging from [0, 1], the higher the value, the more unique the sample; min j∈M represents taking the minimum value among all samples in the memory bank M, finding the most similar sample; ||x i -x j ||2 represents the Euclidean distance between sample x i and sample x j , measuring the distance in the feature space; ω semant ic(x i ,x j ) represents the semantic similarity adjustment factor, considering the semantic similarity of the sample.

[0187] Innovative effect: compared with the traditional random selection strategy, the knowledge retention rate is improved, and the adaptation speed of new samples is improved.

[0188] Memory bank update strategy:

[0189] Maintain a fixed-size memory bank of 1000 samples: limit the memory bank size to 1000 samples to balance computational efficiency and knowledge retention;

[0190] Sample replacement based on representativeness score: replace the sample with the lowest representativeness score in the memory bank when the representativeness score of the new sample is higher than the lowest score;

[0191] Ensure the minimum retention ratio of samples with different damage levels: set the minimum retention ratios of mild, moderate, and severe damage samples to 50%, 30%, and 20% respectively to ensure balanced sample distribution.

[0192] Step 3.2, Progressive Neural Network;

[0193] Construct and train the progressive neural network based on the fusion features of step 2 and the samples in the memory bank:

[0194] Network structure:

[0195] Base column: save the initial trained model parameters, do not update, as a fixed storage of historical knowledge, containing 5 fully connected layers;

[0196] Extended column: dynamically added when processing new tasks, one column for each batch of new data, each column contains the same number of layers as the base column;

[0197] Lateral connection: transfer historical knowledge, connect the output of each layer of the previous sequence to the corresponding layer of the subsequent column through learnable parameters.

[0198] Layer output calculation:

[0199]

[0200] where, represents the output of the ith column and the lth layer, represents the feature representation after processing; f represents the activation function, using ReLU nonlinear transformation; represents the weight matrix of the ith column and the lth layer, used for vertical connection within the column; represents the output of the ith column and the lth layer, as the input of the current layer; represents the lateral connection weight matrix from the jth column and the l-1th layer to the ith column and the lth layer; represents the output of the jth column and the l-1th layer, as the input of the lateral connection; Σ j<i represents the sum of lateral connections for all previous sequences (j < i), realizing the transfer of historical knowledge.

[0201] Step 3.3, Knowledge Distillation;

[0202] Use the old model and new data to optimize the progressive neural network established in step 3.2:

[0203] Feature-level knowledge distillation:

[0204]

[0205] wherein, represents the feature-level knowledge distillation loss, used to measure the difference between the feature representations of the old and new models; L represents the number of layers of the model, and represents the total number of layers for which feature distillation is performed; represents a normalization coefficient, ensuring comparability of models of different layers; represents the sum of the losses of all layers; represents the feature representation of the old model at the lth layer for input x; represents the feature representation of the new model at the lth layer for input x; ||·||2 represents the L2 norm, used for normalization of the feature vector and calculation of the Euclidean distance; represents the normalized representation of the old model feature; represents the normalized representation of the new model feature.

[0206] Relationship knowledge distillation:

[0207]

[0208] wherein, represents the relationship knowledge distillation loss, used to maintain the similarity relationship between samples; n represents the number of samples in the batch; represents a normalization coefficient, ensuring comparability of different batch sizes; represents a double summation over all sample pairs in the batch; x i , x j represent the i th and j th samples in the batch, respectively; represents the feature representation of the old model for input x; represents the feature representation of the new model for input x; sim represents the similarity function.

[0209] Step 4, using the updated prediction model, using the multi-scale time series fatigue damage model of the environmental adaptive damage evolution equation to perform damage evolution calculation, combining the four-dimensional uncertainty quantification Bayesian neural network to perform life prediction, and finally realizing graded early warning through four weighted risk index calculations and dynamic threshold adjustment;

[0210] Specifically comprising the following steps:

[0211] Step 4.1, innovative multi-scale time series fatigue damage model;

[0212] Based on the model updated in step 3, innovative multi-scale time series fatigue damage assessment is performed in combination with sensor data. This is another major innovation of the present application, which specifically contains the following breakthrough technologies:

[0213] Innovative technical features:

[0214] Multi-scale time sequence coupling modeling: considering the time evolution of micro, meso, and macro three scales at the same time;

[0215] Environment factor adaptive integration: dynamically integrating temperature, humidity, load and other environmental influences;

[0216] Nonlinear damage accumulation modeling: considering damage acceleration effect and threshold effect;

[0217] Dynamic adjustment of prediction interval: adaptive adjustment of prediction confidence based on historical performance.

[0218] Multi-scale damage degree calculation: calculate the comprehensive damage degree D ms (t) as follows:

[0219] Calculate the damage degree D s (t) of micro, meso and macro three scales respectively;

[0220] Assign weights w s to each scale, and the sum of weights is 1;

[0221] Multiply the damage degree of each scale by the corresponding weight and sum up to get the comprehensive damage degree.

[0222] Innovative environmental adaptive damage evolution equation:

[0223] Enhanced multi-scale damage evolution equation:

[0224]

[0225] where, D s (t) represents the change rate of the s-th scale damage degree D s with time t, and D s (t) represents the damage evolution speed; α s represents the damage evolution rate coefficient of the s-th scale, which controls the basic damage rate; D eff (t) represents the damage degree of the s-th scale at time t, ranging from 0 to 1, 0 represents no damage, and 1 represents complete damage; β s represents the damage nonlinear index of the s-th scale, which controls the damage self-acceleration effect; σ s (t) represents the effective stress at time t, which considers static and dynamic load; γ s-1 represents the stress sensitivity index of the s-th scale, which represents the degree of influence of stress on damage; η env represents the scale coupling coefficient of the s-th scale, which controls the mutual influence strength between different scales; D s-1 (t) represents the damage degree of the smaller scale at time t, which represents the influence of micro damage on macro damage; Φ env(t) represents the environmental impact factor, quantifying the influence of environmental conditions on damage evolution; Ψ nonlinear (D s (t) represents the nonlinear damage acceleration factor, modeling the acceleration effect after damage reaches a critical value.

[0226] Environmental impact factor:

[0227]

[0228] where Φ env (t) represents the comprehensive environmental impact factor, the product effect of multiple environmental factors; represents the continuous multiplication symbol from 1 to n env , representing the product of the influence of all environmental factors; n env represents the total number of environmental factors considered; E i (t) represents the current value of the i-th environmental factor (temperature, humidity, salinity, etc.); E i,ref represents the reference value of the i-th environmental factor, usually the value under standard environmental conditions; κ i represents the sensitivity coefficient of the i-th environmental factor, indicating the degree of influence of the environmental factor on damage, calibrated through experiments.

[0229] Nonlinear damage acceleration factor:

[0230]

[0231] where Ψ nonlinear (D s ) represents the nonlinear damage acceleration factor, modeling the acceleration effect after damage reaches a critical value; D s represents the current damage degree of the s-th scale; D threshold,s = 0.3 represents the damage acceleration threshold of the s-th scale, indicating the critical point of starting accelerated damage - ρ s = 2.0: damage acceleration intensity coefficient, controlling the amplitude of damage acceleration; ν s = 1.5 represents the nonlinear index, controlling the degree of nonlinearity of damage acceleration; represents the normalized super-threshold damage degree, ranging from [0, 1].

[0232] Effective stress calculation:

[0233]

[0234] where σ eff (t) represents the effective stress at time t, considering the influence of static and dynamic loads; σ static (t) represents the static stress at time t, generated by the dead load and constant load of the structure; σ dynamic(t) represents the dynamic stress at time t, caused by varying loads such as vehicle load, wind load, etc.; k dynamic = 1.2 represents the dynamic stress enhancement coefficient, indicating the additional impact of dynamic load on damage.

[0235] Real-time monitoring of environmental parameters, calculating environmental impact factor Φ env (t);

[0236] According to the current damage degree, judge whether to trigger nonlinear effect, calculate Ψ nonlinear (D s );

[0237] Calculate the effective stress σ eff (t), considering static and dynamic loads comprehensively;

[0238] Apply enhanced evolution equation to calculate damage evolution rate;

[0239] Use the fourth-order Runge-Kutta method for numerical integration.

[0240] Innovative effect: Compared with traditional models, the prediction accuracy is improved, and the environmental adaptability is enhanced.

[0241] Step 4.2, Bayesian neural network prediction;

[0242] Using the model updated in step 3 and the multi-scale damage degree calculated in step 4.1, construct a Bayesian neural network to predict the remaining life:

[0243] Network structure:

[0244] Input layer: fusion feature dimension (from step 2) and multi-scale damage degree (from step 4.1);

[0245] Hidden layer: 256, 128, 64, 32 nodes, using ReLU activation function;

[0246] Output layer: predicted life and uncertainty;

[0247] Innovative four-dimensional uncertainty quantification model:

[0248] Innovative four-dimensional uncertainty decomposition framework:

[0249]

[0250] Random uncertainty enhancement calculation:

[0251]

[0252] Where, represents the overall uncertainty variance, considering all sources of uncertainty; This represents the variance of random uncertainty, which originates from inherent noise in the data. The variance representing cognitive uncertainty stems from the uncertainty of model parameters and structure, and can be reduced by increasing training data. The variance represents the distribution uncertainty, which stems from the uncertainty caused by the difference in the distributions of the training and test data. The variance representing time-series uncertainty stems from the time dependence and periodic variations in time series forecasting; E[NN] σ [x] represents the expected variance of the neural network prediction; ξ represents the noise variance adjustment factor, with a value of 0.1, used to balance the prediction variance and observation noise; Var(y noise ) represents the noise variance of the observed data.

[0253] Bayesian estimation of cognitive uncertainty:

[0254]

[0255] in, The variance represents cognitive uncertainty, which originates from the uncertainty of model parameters; N represents the number of Monte Carlo samplings, with a value of 100, used for multiple sampling to estimate the variance. This represents the predicted value from the i-th sample. λ represents the average of all sampled predicted values. MC This represents the Dropout variance adjustment coefficient, with a value of 0.5, used to adjust the variance contribution brought by Dropout; Dropout_Variance represents the variance estimate obtained through the Dropout layer, reflecting the uncertainty of the model structure.

[0256] Quantification of distribution uncertainty:

[0257]

[0258] in, The variance represents the distributional uncertainty, stemming from the difference between the training and test data distributions; MMD represents the maximum mean difference, measuring the distance between two data distributions; X train X represents the training dataset; test Represents the test dataset; |X test | indicates the number of samples in the test dataset; |X train | indicates the number of samples in the training dataset; This represents the sample size ratio adjustment factor, which reduces distribution uncertainty when the number of test samples increases.

[0259] Innovative computation with temporal uncertainty:

[0260]

[0261] where, denotes the variance of aleatory uncertainty, which is derived from the fluctuation of time series prediction; T denotes the length of time series; denotes the predicted value at time point t; denotes the predicted value at time point t-1; w t denotes the time series weight; φ(t) denotes the periodic adjustment factor.

[0262] Adaptive weight allocation:

[0263]

[0264] where, total denotes the weight vector of each uncertainty source; w ale denotes the weight of aleatory uncertainty; w epi denotes the weight of epistemic uncertainty; w dis denotes the weight of distribution uncertainty; w tem denotes the weight of temporal uncertainty; ∈ denotes a small constant to prevent zero value problems in logarithmic calculation; softmax denotes a normalization function to ensure that the sum of all weights is 1; denotes the variance of aleatory uncertainty, which is derived from the inherent noise of data; denotes the variance of epistemic uncertainty, which is derived from the uncertainty of model parameters and structure, and can be reduced by increasing training data; denotes the variance of distribution uncertainty, which is derived from the uncertainty caused by the distribution difference between training data and test data; denotes the variance of temporal uncertainty, which is derived from the time dependence and periodic changes in time series prediction.

[0265] Final uncertainty calculation:

[0266]

[0267] where, denotes the final comprehensive uncertainty variance; w i denotes the adaptive weight of the i-th type of uncertainty; denotes the variance of the i-th type of uncertainty; γ i denotes the importance adjustment factor, which is learned from historical data and used to further adjust the contribution of each type of uncertainty; i ∈ {ale, epi, dis, tem} denotes the index set of four types of uncertainty, representing aleatory uncertainty ale, epistemic uncertainty epi, distribution uncertainty dis, and temporal uncertainty tem, respectively.

[0268] Innovative effect: Compared with traditional three-dimensional uncertainty models, the prediction interval accuracy is improved, the risk identification accuracy is improved, and the time series fluctuation prediction ability is enhanced.

[0269] Step 4.3, Intelligent Early Warning Decision System;

[0270] Based on the life prediction results and uncertainty quantification of Step 4.2, an intelligent early warning decision system is constructed:

[0271] Risk Index Calculation: The steps to calculate the risk index RI are:

[0272] Calculate the life ratio term: (1-L / L ref ), weight 0.4;

[0273] Calculate the uncertainty term: σ L / σ ref , weight 0.3;

[0274] Calculate the damage degree term: D / D cr , weight 0.2;

[0275] Calculate the life change rate term: |L t -L t-1 | / L t-1 , weight 0.1;

[0276] Sum the four weighted terms to get the final risk index.

[0277] Dynamic Threshold Strategy: The steps to calculate the dynamic early warning threshold θ i are:

[0278] Get the initial threshold

[0279] Calculate the importance influence: 0.1×I imp ;

[0280] Calculate the usage intensity influence: 0.1×I usage ;

[0281] Multiply the initial threshold by (1-importance influence-usage intensity influence) to get the dynamic threshold.

[0282] Multi-level Early Warning Decision Mechanism:

[0283] Safety Level (RI<θ1): Annual routine inspection, maintain normal monitoring frequency;

[0284] Attention Level (θ1≤RI<θ2): Increase monitoring frequency, conduct special inspection every half year, closely monitor key parts;

[0285] Warning Level (θ2≤RI<θ3): Quarterly comprehensive inspection, start local reinforcement scheme evaluation, add temporary monitoring points;

[0286] Danger level (RI≥θ3): monthly comprehensive inspection, emergency plan, traffic control if necessary, and maintenance and reinforcement project.

[0287] wherein θ1=0.3: safety level threshold, representing the upper limit value of the structure in a safe state;

[0288] θ2=0.5: attention level threshold, representing the risk level that the structure needs to pay attention to;

[0289] θ3=0.7: warning level threshold, representing the level of obvious risk of the structure.

[0290] Adaptive maintenance strategy generation: according to the risk index and prediction uncertainty, the system automatically generates optimized maintenance decision suggestions:

[0291] Maintenance timing optimization: based on the risk evolution prediction curve, the best maintenance intervention time point is determined;

[0292] Maintenance scheme selection: from the preset maintenance scheme library, combined with damage characteristics and life prediction results, the optimal maintenance scheme is recommended;

[0293] Maintenance resource allocation: according to the risk ranking of multiple monitoring points, the maintenance resources are reasonably allocated;

[0294] Maintenance effect evaluation: by comparing the risk index changes before and after maintenance, the maintenance effect is evaluated;

[0295] Through the above intelligent early warning decision system, the method realizes the full-process intelligentization from data acquisition, feature extraction to life prediction, risk assessment and maintenance decision, and solves the key technical problems of data heterogeneity, model forgetting, sample imbalance, prediction uncertainty and environmental variation in traditional methods. The system not only can accurately evaluate the fatigue life of the main girder, but also can provide scientific maintenance decision support, significantly improving the efficiency and accuracy of bridge safety management.

[0296] The intelligent early warning decision system is applicable to fatigue life evaluation and maintenance decision of various bridge main girders, especially for long-service-life large bridges under complex load environment, and has important application value.

[0297] Application examples of the embodiment:

[0298] To further illustrate the implementation method of the present application, the following takes a certain expressway river-crossing bridge as an example to elaborate the specific application process and effectiveness of the method.

[0299] Background:

[0300] The target bridge is a double-tower double-cable plane steel box girder cable-stayed bridge on a certain expressway in a certain province, with a main span of 300 meters, a side span of 150 meters, and a total length of 600 meters. The bridge was completed and opened to traffic in 2003 and has been in service for 18 years. The daily average traffic volume is about 25,000 vehicles, of which heavy vehicles account for about 30%.

[0301] During the routine inspection in 2018, it was found that fatigue cracks appeared in some areas of the main girder of the bridge. The management department urgently needs to assess the fatigue life of the main girder of the bridge and develop a scientific maintenance plan.

[0302] System deployment:

[0303] Sensor layout:

[0304] The following sensors are deployed at key positions of the main girder:

[0305] 24 strain gauges (sampling frequency: 100 Hz): distributed in the midspan, 1 / 4 span, and support area;

[0306] 12 acceleration sensors (sampling frequency: 200 Hz): arranged in the midspan and 1 / 3 span;

[0307] 6 displacement meters (sampling frequency: 10 Hz): installed in the midspan and 1 / 4 span positions;

[0308] 4 high-definition cameras: regularly photograph suspected crack areas;

[0309] 4 sets of environmental monitoring devices: record temperature, humidity, wind speed, and other parameters;

[0310] 2 sets of traffic flow monitoring systems: record traffic volume and vehicle type distribution.

[0311] Data collection cycle:

[0312] Static data: collected once every hour, each time lasting 5 minutes;

[0313] Dynamic data: collected twice a day during peak traffic periods (7:00-9:00, 17:00-19:00), each time lasting 10 minutes;

[0314] Image data: taken once a week;

[0315] Environmental data: continuously monitored, recorded every 10 minutes;

[0316] Traffic flow data: continuously monitored for 24 hours;

[0317] Specific implementation process:

[0318] Data collection and preprocessing:

[0319] Data cleaning:

[0320] The data collected within a month were cleaned, and a total of:

[0321] Strain data: 2,160 hours x 24 sensors = 51,840 time series data;

[0322] Acceleration data: 600 hours x 12 sensors = 7,200 time series data;

[0323] Displacement data: 2,160 hours x 6 sensors = 12,960 time series data;

[0324] Image data: 4 cameras x 4 weeks = 16 high-definition images;

[0325] Environmental data: 4,320 records;

[0326] Traffic data: 720 hours of continuous recording.

[0327] Through the isolation forest algorithm, 683 outliers were identified, accounting for 0.95% of the total data.

[0328] Among them:

[0329] Caused by sensor failure: 213;

[0330] Caused by external interference: 157;

[0331] Caused by data transmission error: 313.

[0332] Rainflow counting and damage degree preliminary calculation:

[0333] Rainflow counting method was applied to the cleaned strain data to obtain stress cycle characteristics:

[0334] Total cycle number: 9.72 x 10 6 times;

[0335] Maximum stress amplitude: 96.7 MPa;

[0336] Equivalent stress amplitude: 43.2 MPa

[0337] The preliminary estimated fatigue damage degree is 0.427 (based on Miner's linear cumulative damage theory). Multimodal feature fusion:

[0338] Modal feature extraction:

[0339] Features are extracted from various types of data:

[0340] Strain features: peak, mean, standard deviation, frequency distribution, etc., a total of 28-dimensional features; acceleration features: RMS value, peak value, main frequency, modal parameter, etc., a total of 32-dimensional features; displacement features: maximum displacement, average displacement, change rate, etc., a total of 15-dimensional features;

[0341] Image features: crack length, width, distribution density, etc., a total of 20-dimensional features;

[0342] Environmental features: temperature, humidity, wind speed, and their change rates, etc., a total of 12-dimensional features;

[0343] Traffic flow features: flow, vehicle type distribution, load estimation, etc., a total of 18-dimensional features.

[0344] Feature fusion:

[0345] Feature fusion is performed through attention mechanism, and the attention weight of each modality is:

[0346] Strain data: 0.38;

[0347] Acceleration data: 0.25;

[0348] Displacement data: 0.15;

[0349] Image data: 0.12;

[0350] Environmental data: 0.06;

[0351] Traffic flow data: 0.04;

[0352] Finally, a 125-dimensional fusion feature vector is generated.

[0353] Dual knowledge consolidation incremental learning:

[0354] Model training:

[0355] The base model uses a five-layer neural network (input layer-125 nodes, hidden layer-256-128-64 nodes, output layer-1 node), and uses historical data (2018-2020) for initial training:

[0356] Number of training samples: 26,280;

[0357] Learning rate: 0.001;

[0358] Batch size: 64;

[0359] Training rounds: 200;

[0360] Validation set loss: 0.0087.

[0361] Incremental learning and knowledge consolidation:

[0362] When new data arrives in 2021, a double knowledge consolidation mechanism is used for model updating:

[0363] Number of new samples: 8,760;

[0364] Memory capacity: 5,000 (selected from historical data);

[0365] Distillation temperature parameter: 2.0;

[0366] Distillation weight: 0.4.

[0367] Model evaluation results:

[0368] Knowledge retention rate: 91.3% (traditional method only 62.7%);

[0369] New sample learning accuracy: 93.8%;

[0370] Overall prediction error: 8.7%.

[0371] Fatigue life assessment and uncertainty quantification:

[0372] Multi-scale fatigue damage calculation:

[0373] Based on the collected data, the fatigue damage degree of three scales is calculated:

[0374] Macro scale (s = 1): D1 = 0.426, w1 = 0.5;

[0375] Mesoscale (s = 2): D2 = 0.387, w2 = 0.3;

[0376] Micro scale (s = 3): D3 = 0.462, w3 = 0.2;

[0377] Comprehensive multi-scale damage degree: D ms = 0.426 × 0.5 + 0.387 × 0.3 + 0.462 × 0.2 = 0.419.

[0378] Bayesian neural network prediction:

[0379] Based on the updated model, the fatigue life of the main beam is predicted:

[0380] Predicted remaining life: 22.3 years;

[0381] Prediction uncertainty decomposition:

[0382] Accidental uncertainty: ±1.7 years (7.6%);

[0383] Cognitive uncertainty: ±1.2 years (5.4%);

[0384] Distribution uncertainty: ± 0.8 years (3.6%);

[0385] Total uncertainty: ± 2.2 years (9.9%).

[0386] Risk assessment and early warning:

[0387] Calculate risk index according to prediction results:

[0388] Benchmark life (design life): L ref = 100 years;

[0389] Predicted life: L = 40.3 years (18 years served + remaining 22.3 years);

[0390] Uncertainty ratio: σL / σ ref = 0.022 / 0.05 = 0.44;

[0391] Current damage degree: D = 0.419;

[0392] Critical damage degree: D kr = 1.0;

[0393] Life change rate: |L t -L t-1 | / L t-1 = 0.035.

[0394] Calculate risk index:

[0395] RI = 0.4 × (1 - 40.3 / 100) + 0.3 × 0.44 + 0.2 × (0.419 / 1.0) + 0.1 × 0.035

[0396] = 0.24 + 0.132 + 0.0838 + 0.0035

[0397] = 0.4593.

[0398] According to the dynamic threshold (θ1 = 0.3, θ2 = 0.5, θ3 = 0.7), the bridge is in the "attention level" early warning state, and it is recommended to take a half-yearly inspection frequency.

[0399] The implementation effect is compared as shown in Table 1:

[0400] Table 1: Comparison of the method with the traditional evaluation method;

[0401] Evaluation index Traditional method Method of the present application Improvement effect Lifetime prediction error ±23.6％ ±9.9％ Increased by 58.1% Evaluation period 7 days 2 days Reduced by 71.4% Manual workload 35 man-days 16 man-days Reduced by 54.3% Data processing amount 2.3 TB 7.1 TB Increased by 208.7% Feature dimension 36 dimensions 125 dimensions Increased by 247.2% Computing cost 23,000 yuan 13,800 yuan Reduced by 40.0%

[0402] Long-term tracking verification:

[0403] To verify the reliability of the method, the bridge was tracked and monitored for 2 years (2019-2021), and the results showed that:

[0404] The prediction life coincides with the actual damage accumulation: 92.7%;

[0405] Environmental change adaptability: under the condition of extreme temperature difference (-15 DEG C to 40 DEG C), the predicted stability decreases by not more than 3.5%;

[0406] Load change adaptability: under the condition of traffic flow fluctuation ± 30%, the prediction precision fluctuates by not more than 4.2%;

[0407] Abnormal condition recognition rate: 93.8% of abnormal events are successfully identified, and the average early warning is 4.2 days.

[0408] In summary, the method of the present application has significant technical advantages in practical application, not only improves the accuracy and reliability of fatigue life evaluation, but also greatly reduces the evaluation cost, and provides a scientific basis for bridge safety management.

[0409] The above describes the embodiments of the present application, but the embodiments are not limited to the specific implementation described above, and the specific implementation described above is only illustrative, not restrictive, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A method for main beam fatigue life assessment based on sample continuous increment, characterized by: Collecting strain, acceleration, displacement, image, environment and traffic data through multi-source sensors arranged at key positions of the main beam, and using an innovative adaptive wavelet packet transform denoising method to process the collected data; Using an innovative multi-dimensional adaptive attention mechanism to fuse multi-modal features based on the denoised data, and generating a fused feature representation; Based on the fused features, using a multi-dimensional sample representativeness scoring algorithm to dynamically manage the memory bank, and updating the prediction model using an incremental learning method combining a progressive neural network lateral connection mechanism and a feature-level and relationship-level dual knowledge distillation; Using the updated prediction model, using a multi-scale time series fatigue damage model of the environmental adaptive damage evolution equation to calculate the damage evolution, combining a four-dimensional uncertainty quantification Bayesian neural network for life prediction, and finally achieving graded early warning through four weighted risk index calculations and dynamic threshold adjustments; The multi-dimensional sample representativeness scoring algorithm specifically includes: The comprehensive scoring formula combines diversity, difficulty, uncertainty and time importance in four dimensions, and uses a dynamic weight distribution mechanism; The diversity weight in the dynamic weight distribution is adaptively adjusted according to the current memory bank filling degree, and the maximum capacity of the memory bank is set to 1000 samples; The time importance is calculated by an exponential decay function, the time decay parameter is set to 0.01, and the diversity calculation uses a semantic similarity adjustment factor enhancement algorithm; The progressive neural network lateral connection mechanism specifically includes: The layer output calculation considers both the internal connection of the column and the lateral connection of all previous sequences, and realizes the transmission of historical knowledge through weighted summation; The base column maintains the fixed parameters of the initial training, contains five fully connected layers, and serves as an invariant storage of historical knowledge; The extended column is dynamically added with new tasks, each column has the same structure as the base column, the lateral connection weight matrix transmits historical knowledge, and the learning rate is set to 0.1 times that of the main network; The dual knowledge distillation specifically includes: The feature-level knowledge distillation loss calculates the Euclidean distance after normalizing all intermediate layer features to maintain consistency between new and old models at the feature representation level; The relationship knowledge distillation loss maintains the consistency of the relative relationship between samples by calculating the difference in similarity relationship between samples; The distillation temperature parameter is set to 2.0, the distillation weight is set to 0.4, the total loss function combines task loss, feature distillation loss and relationship distillation loss, and the weight ratio is 6:3:

1.

2. The method for evaluating the fatigue life of a main girder based on a sample continuous increment according to claim 1, characterized in that, The innovative adaptive wavelet packet transform denoising method specifically includes: Decomposition layer determination algorithm: J = J base + δ SNR + δ λ ; Wherein, J represents the final determination of the wavelet packet decomposition layer number, that is, the number of times the signal will be recursively decomposed; J base represents the basic decomposition layer number; δ SNR represents the signal-to-noise ratio adaptive adjustment factor, which is set according to the signal quality section. When the signal-to-noise ratio SNR>15dB, the value is one; when 10dB≤SNR≤15dB, the value is zero; when SNR<10dB, the value is-1; δ λ represents the signal complexity adjustment factor, which reflects the signal spectrum complexity. Shannon entropy criterion adopts an entropy value calculation method enhanced by a weight factor. The weight factor is determined by the correlation between the feature and the target signal. The adjustment coefficient is set to 0.

3. The semi-soft threshold function uses a three-section piecewise processing method, uses gradual processing in the transition interval to avoid the discontinuity of the hard threshold value, and ensures the smoothness of signal reconstruction.

3. The method for evaluating fatigue life of a main girder based on a sample continuous increment according to claim 1, characterized in that, The innovative multi-dimensional adaptive attention mechanism specifically includes: The modality quality score is calculated by a sigmoid function, which maps the linearly transformed weight matrix to a quality score interval of zero to one; The temperature adaptive adjustment algorithm dynamically adjusts the temperature parameter according to the variance of the query matrix and the key matrix, the basic temperature parameter is set to 1.0, and the variation coefficient adjustment factor is set to 0.

5. The improved softmax function is used for attention weight calculation, which considers the comprehensive influence of original attention score, inter-modal correlation and quality score, and introduces a numerical stability parameter to avoid calculation overflow.

4. The method for evaluating fatigue life of a main girder based on sample continuous increments according to claim 1, characterized in that, The environmental adaptive damage evolution equation specifically includes: The multi-scale damage evolution equation: wherein, D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D eff D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D s-1 D s (t) represents the s-th scale damage degree D env D s (t) represents the s-th scale damage degree D nonlinear D s (t) represents the s-th scale damage degree D s D s (t) represents the s-th scale damage degree D The effective stress calculation combines the square roots of static stress and dynamic stress, the dynamic stress enhancement coefficient is set to 1.2, and the multi-scale comprehensive damage degree weight distribution is macro 0.5, meso 0.3, and micro 0.

2.

5. The method for evaluating fatigue life of a main girder based on a sample continuous increment according to claim 1, characterized in that, The Bayesian neural network of four-dimensional uncertainty specifically includes: The overall uncertainty is decomposed into four dimensions of random uncertainty, cognitive uncertainty, distribution uncertainty and time sequence uncertainty, adding a time sequence uncertainty dimension to the traditional three dimensions; The cognitive uncertainty is calculated by the Bayesian estimation method, using Monte Carlo sampling one hundred times, combined with the Dropout variance adjustment mechanism, and the adjustment coefficient is set to 0.5; The time sequence uncertainty is calculated by weighted time sequence difference, the time sequence weight uses exponential decay, and the periodic adjustment factor considers the twenty-four hour periodic variation law.

6. The method for evaluating fatigue life of a main girder based on a sample continuous increment according to claim 1, characterized in that, The risk index calculation and dynamic threshold adjustment specifically include: The risk index calculation formula: wherein, R I represents the risk index; L represents the current measured structural characteristic parameter; L ref represents the reference value of the structural characteristic parameter; represents the relative degree of deviation of the structural characteristic parameter from the reference value; σ L represents the standard deviation of the structural characteristic parameter; σ ref represents the reference value of the standard deviation of the structural characteristic parameter; represents the relative uncertainty of the structural parameter fluctuation, with a weight of 0.3; D represents the current cumulative damage degree; D cr represents the critical damage degree threshold value; represents the degree of damage close to the critical state, with a weight of 0.2; L t represents the structural characteristic parameter value at the current time t; L t-1 represents the structural characteristic parameter value at the previous time t-1; represents the relative change rate of the structural parameter; The dynamic threshold adjustment is corrected according to the importance of bridge traffic and load use intensity, and the adjustment coefficients are both 0.1; The four-level warning threshold is set to safety level 0.3, attention level 0.5, warning level 0.7, and danger level 0.9, corresponding to annual, semi-annual, quarterly, and monthly inspection frequencies.

7. A computer-readable storage medium, characterized in that, It is used to store computer readable instructions, which can run the method of evaluating the fatigue life of the main beam based on the sample continuous increment when the computer readable instructions are read by the computer.

Images