A structural fatigue damage identification method based on acoustic emission and deep learning

Abstract

This invention relates to the field of structural health monitoring and intelligent diagnosis technology, particularly a method for identifying structural fatigue damage based on acoustic emission and deep learning. This method uses an acoustic emission sensor array to collect structural response signals under fatigue loads and inputs them into a CNN-BiLSTM-Attention hybrid deep learning model. The model extracts local time-domain features using a dynamic adaptive convolution kernel, captures long-term temporal dependencies using a bidirectional long-short-term memory network, and focuses on key damage features using a bimodal spatiotemporal attention mechanism. A nonlinear dynamic threshold algorithm based on crack opening is used to divide damage stages. A physical-data fusion training dataset is constructed, and a gradient-sensitive cosine annealing algorithm is used to optimize the learning rate. A noise-resistant adversarial loss function is used to enhance model robustness. By integrating physical features with intelligent algorithms, this method offers advantages such as adaptive noise suppression, strong cross-domain generalization, and high real-time performance.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring and intelligent diagnosis, and specifically to a structural fatigue damage identification method based on acoustic emission and deep learning. Background Art

[0002] Over the course of long-term service, infrastructure such as bridges is susceptible to fatigue damage due to factors such as traffic loads and environmental erosion. This damage can lead to microcracks initiating and propagating, ultimately leading to structural instability and failure. Traditional fatigue damage detection relies on manual inspections and ultrasonic testing, which are inefficient, subjective, and difficult to detect early damage. Furthermore, it lacks real-time monitoring throughout the entire lifecycle.

[0003] Acoustic emission testing, a dynamic nondestructive testing technique, can reveal the internal state of a structure by capturing elastic wave signals released when materials are damaged. It has been widely used for crack monitoring. However, acoustic emission signals are susceptible to interference from environmental noise, and the signal characteristics during fatigue damage evolution are complex (e.g., significant differences in energy and frequency at different stages). Traditional signal processing methods (such as Fourier transforms and wavelet analysis) struggle to effectively extract multi-scale spatiotemporal features and long-term dependencies.

[0004] Deep learning technology has demonstrated powerful capabilities in time series data modeling, but its application in structural health monitoring still faces the following challenges:

[0005] Insufficient adaptability of feature extraction: Traditional convolutional neural networks (CNNs) use fixed-size convolution kernels, which make it difficult to adaptively capture the characteristic scale changes of different damage stages (such as the high-frequency transient signals of microcrack initiation and the low-frequency trend signals of macrocrack extension);

[0006] Weak physical relevance: Existing models mostly rely on single data drivers and lack effective integration with structural physical damage parameters (such as crack opening and cycle number), resulting in a lack of physical interpretability in damage stage division.

[0007] Limited robustness and generalization capabilities: Actual engineering environments are subject to strong noise interference (such as traffic vibration and electromagnetic noise), and the damage characteristics of different bridge structures vary significantly. Traditional models need to be retrained for cross-domain applications, resulting in high deployment costs.

[0008] The contradiction between real-time and lightweight: Complex deep learning models require high computing resources when deployed on edge devices, making it difficult to meet the real-time requirements of online monitoring.

[0009] Therefore, to address the above problems, a structural fatigue damage identification method based on acoustic emission and deep learning is proposed. Summary of the Invention

[0010] The purpose of the present invention is to provide a structural fatigue damage identification method based on acoustic emission and deep learning to solve the problems raised in the above background technology.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] A structural fatigue damage identification method based on acoustic emission and deep learning includes the following steps:

[0013] S1. Acoustic emission signals under fatigue load are collected by an acoustic emission sensor array arranged on the surface of the bridge structure, and the signals are preprocessed, including noise reduction, normalization and segmentation processing;

[0014] S2. Use the sliding window segmentation algorithm to perform time series segmentation on the preprocessed acoustic emission signal, with a window length of 150 sampling points and a step length of 10 sampling points to generate a local time series sample set;

[0015] S3. Construct a CNN-BiLSTM-Attention hybrid deep learning model, wherein the CNN-BiLSTM-Attention hybrid deep learning model includes:

[0016] CNN layer: uses a one-dimensional convolution kernel to extract the local time domain features of the acoustic emission signal, including amplitude mutation and energy peak features;

[0017] BiLSTM layer: captures long-term temporal dependencies in the damage evolution process through a bidirectional long short-term memory network;

[0018] Attention mechanism layer: Adaptively focuses on high-energy signal features at critical injury stages based on a learnable weight matrix;

[0019] S4. The damage stages are divided based on a nonlinear curve of the bridge main crack opening and the number of cycles. The stage labels are determined by the slope mutation points. The labels are aligned with the acoustic emission signal time series to construct a physical-data fusion training dataset.

[0020] S5. Use dynamic learning rate adjustment strategy and early stopping mechanism to train the hybrid model, and combine Dropout and batch normalization operations to improve model robustness;

[0021] S6. Input the real-time collected acoustic emission signals into the trained model and output the damage stage identification results, including crack initiation, stable expansion and unstable failure stages.

[0022] As a preferred solution, the convolution kernel size of the CNN layer is dynamically adjusted, specifically:

[0023] According to the change of the signal energy gradient at the current moment, the convolution kernel size is adjusted in real time, with the adjustment range being 5 to 15 sampling points;

[0024] The convolution kernel weight parameters are updated through online learning, and the bias term is inversely proportional to the local signal-to-noise ratio of the signal;

[0025] The signal energy gradient is constrained by the Sigmoid function to constrain its influence on the convolution kernel.

[0026] As a preferred solution, the weight calculation of the attention mechanism layer is implemented in the following way:

[0027] Generate query vector and key vector based on the hidden state of BiLSTM output;

[0028] In the temporal attention calculation, the logarithmic weighted term of the accumulated acoustic emission energy in the time window is superimposed, and the time window length is 50 sampling points;

[0029] The energy weighting coefficients are optimized via back-propagation to enhance the focusing capability on high-energy damage events.

[0030] As a preferred solution, the damage stage classification method includes:

[0031] Calculate the absolute value of the second-order derivative of the crack opening curve, and trigger the cutoff point judgment when it exceeds the basic threshold of 0.05;

[0032] The threshold of the demarcation point is dynamically adjusted. The adjustment amplitude is proportional to the variance of the crack opening in the last 50 cycles of loading. The maximum variance value is taken from the training set for normalization.

[0033] The interval between adjacent dividing points must meet the constraint of at least 100 cyclic loading times.

[0034] As a preferred solution, the dynamic learning rate adjustment strategy is implemented through the following steps:

[0035] The cosine annealing algorithm is used, with an initial learning rate of 0.001, a minimum learning rate of 0.0001, and a cycle length of 10 training rounds;

[0036] A gradient change sensitivity factor is introduced, and the modulus ratio of the current gradient vector to the previous step gradient is used as the learning rate scaling coefficient. The hyperbolic tangent function is used to suppress the oscillation caused by gradient mutation.

[0037] As a preferred solution, in the pre-processing step, the frequency domain adaptive wavelet packet decomposition algorithm is used for the acoustic emission signal, and the frequency band energy weight calculation includes:

[0038] Perform wavelet packet decomposition on the acoustic emission signal and calculate the energy proportion of each frequency band;

[0039] Apply exponential weight enhancement to the frequency band with a center frequency higher than 50kHz. The weight increases as the difference between the frequency and the threshold increases, and the steepness factor is 0.1.

[0040] The signal segment length is fixed at 150 samples.

[0041] As a preferred solution, the signal fusion method of the acoustic emission sensor array is:

[0042] The spatial attenuation weight is calculated based on the Euclidean distance between the sensor and the damage location estimate, with an attenuation coefficient of 100 mm;

[0043] When superimposing the signals of each sensor, the timing is aligned according to the acoustic wave propagation time delay and multiplied by the sensor calibration amplitude coefficient;

[0044] The distance weight denominator is added with a very small constant 1×10 -5 Prevent division by zero exceptions.

[0045] As a preferred solution, the CNN-BiLSTM-Attention hybrid deep learning model training loss function includes the following three items:

[0046] Category-weighted cross entropy loss, where weights are inversely proportional to the number of category samples;

[0047] Input signal gradient penalty term, with a coefficient of 0.5, is used to suppress noise sensitivity;

[0048] Against the perturbation consistency constraint, a uniformly distributed perturbation is applied to the input signal and the output distribution difference is calculated, with a coefficient of 0.3.

[0049] As a preferred solution, the CNN-BiLSTM-Attention hybrid deep learning model is deployed using a layered knowledge distillation algorithm, including:

[0050] The corresponding layer feature maps of the teacher model and the student model are L2 normalized and the difference loss is calculated;

[0051] The KL divergence of the output distribution is constrained synchronously, and the knowledge distillation weight coefficient is 0.1;

[0052] A very small constant is added to the denominator of the characteristic normalization function to prevent division by zero.

[0053] As a preferred solution, the cross-bridge type generalization method adopts a multi-domain adaptive migration algorithm, including:

[0054] The maximum mean difference between the acoustic emission signal distributions in the source and target domains was calculated, with a domain alignment coefficient of 0.1;

[0055] Apply L2 regularization constraint to the input signal gradient with a coefficient of 0.01 to improve cross-domain stability;

[0056] The training data must cover at least five bridge types, three load spectra, and temperature-humidity coupling environments.

[0057] It can be seen from the technical solutions provided by the present invention that the structural fatigue damage identification method based on acoustic emission and deep learning provided by the present invention has the following beneficial effects:

[0058] 1. Accurate identification and physical correlation of multi-stage damage:

[0059] Full-process feature fusion: Structural response signals are collected through an acoustic emission sensor array. Combined with signal preprocessing, time series segmentation, and a hybrid deep learning model (CNN-BiLSTM-Attention), dynamic identification of all stages of crack initiation, stable expansion, and unstable failure is achieved, preventing complex damage patterns from being missed by a single technology.

[0060] Physics-driven annotation system: Automatically divides damage stages based on the changing trend of crack opening, aligns the acoustic emission signal time series with the physical damage process, and ensures that model training data has both data-driven accuracy and physical interpretability, reducing the subjectivity and errors of manual annotation.

[0061] 2. Strong robustness and generalization ability in complex environments:

[0062] Adaptive noise suppression: Through frequency domain filtering and spatiotemporal co-location algorithms, it effectively suppresses environmental noise (such as traffic vibration and structure-borne noise), improves the signal-to-noise ratio, and ensures the extraction of true damage characteristics in complex engineering environments.

[0063] Cross-scenario migration capability: Using a multi-domain adaptive algorithm, the model can adapt to the differences in damage characteristics of different types of bridges (such as steel bridges and concrete bridges), eliminating the need for separate training for each structure, significantly improving generalization capabilities and deployment efficiency.

[0064] 3. Efficient model training and lightweight deployment:

[0065] Dynamic optimization strategy: Through dynamic learning rate adjustment and early stopping mechanism, it accelerates model convergence and prevents overfitting, shortens training cycle and improves model stability;

[0066] Lightweight and real-time: Knowledge distillation is used to compress the model size. Combined with edge computing technology, the model can be quickly inferred on embedded devices, achieving real-time response to damage signal "collection-analysis-early warning" and meeting the timeliness requirements of online monitoring.

[0067] 4. Intelligent and cost-effective engineering applications:

[0068] Intelligent early warning and decision support: Real-time output and visualization of damage stage identification results, combined with multi-level early warning mechanisms (such as emergency alarms for instability and damage), provide a scientific basis for structural maintenance, transforming passive maintenance into active early warning;

[0069] Full lifecycle cost optimization: Non-contact detection reduces manual intervention and traffic disruption, while the model's self-calibration mechanism reduces subsequent maintenance costs. Suitable for long-term bridge health monitoring, it significantly improves operation and maintenance efficiency and reduces overall costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of a structural fatigue damage identification method based on acoustic emission and deep learning according to the present invention. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0072] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0073] like Figure 1 As shown, an embodiment of the present invention provides a structural fatigue damage identification method based on acoustic emission and deep learning, comprising the following steps:

[0074] S1. Acoustic emission signals under fatigue load are collected by an acoustic emission sensor array arranged on the surface of the bridge structure, and the signals are preprocessed, including noise reduction, normalization and segmentation processing;

[0075] S2. Use the sliding window segmentation algorithm to perform time series segmentation on the preprocessed acoustic emission signal, with a window length of 150 sampling points and a step length of 10 sampling points to generate a local time series sample set;

[0076] S3. Construct a CNN-BiLSTM-Attention hybrid deep learning model, wherein the CNN-BiLSTM-Attention hybrid deep learning model includes:

[0077] CNN layer: uses a one-dimensional convolution kernel to extract the local time domain features of the acoustic emission signal, including amplitude mutation and energy peak features;

[0078] BiLSTM layer: captures long-term temporal dependencies in the damage evolution process through a bidirectional long short-term memory network;

[0079] Attention mechanism layer: Adaptively focuses on high-energy signal features at critical injury stages based on a learnable weight matrix;

[0080] S4. The damage stages are divided based on a nonlinear curve of the bridge main crack opening and the number of cycles. The stage labels are determined by the slope mutation points. The labels are aligned with the acoustic emission signal time series to construct a physical-data fusion training dataset.

[0081] S5. Use dynamic learning rate adjustment strategy and early stopping mechanism to train the hybrid model, and combine Dropout and batch normalization operations to improve model robustness;

[0082] S6. Input the real-time collected acoustic emission signals into the trained model and output the damage stage identification results, including crack initiation, stable expansion and unstable failure stages.

[0083] In this embodiment, the purpose of step S1 of acoustic emission signal acquisition and preprocessing is to obtain acoustic emission signals under structural fatigue load through the sensor array, and provide high-quality time series feature input for the subsequent deep learning model through noise reduction, normalization and segmentation processing. The specific steps are as follows:

[0084] Step S1-1: Acoustic emission sensor array layout and parameter setting:

[0085] Sensor layout strategy: Based on prior knowledge of fatigue damage for bridge structure types (e.g., steel box girder bridges, simply supported concrete girder bridges), N acoustic emission sensors (N ≥ 4, typically 8) are evenly arranged in a grid pattern with 500-1000mm spacing in stress concentration areas (e.g., welds, near supports) and on the surfaces of key load-bearing components. The sensors are secured with magnetic bases, and a coupling agent (e.g., vaseline) is applied to the contact surfaces to reduce acoustic impedance and ensure signal transmission efficiency.

[0086] Sensor parameter configuration:

[0087] Frequency response range: 20kHz-200kHz (covering the characteristic acoustic emission frequency band of fatigue damage of metal / concrete materials);

[0088] Sampling frequency: set to 1MHz (to meet the Nyquist sampling theorem and avoid high-frequency signal aliasing);

[0089] Preamplifier gain: 40dB (enhances the original signal amplitude and suppresses environmental noise);

[0090] Step S1-2: Synchronous signal acquisition under fatigue load:

[0091] Acquisition trigger conditions: When the bridge is subjected to a cyclic loading test (such as a sinusoidal load applied by a vehicle load simulator, with a frequency of 1-5 Hz) or actual traffic load, the acoustic emission acquisition system and the load-displacement monitoring equipment are synchronously triggered;

[0092] Multi-source data synchronization recording:

[0093] Acoustic emission signal: recorded as multi-channel time series data ( is the sensor number, is the sampling point number);

[0094] Physical damage parameters: Synchronous acquisition of the main crack opening of the bridge and the number of loop loads , real-time monitoring by laser displacement meter (accuracy ±0.01mm);

[0095] Single acquisition time: The acquisition time of a single sample corresponds to 100-500 load cycles, ensuring coverage of the complete damage stage from crack initiation and propagation to instability. The total number of sampling points is 150,000-750,000 points / channel (calculated at a 1MHz sampling rate);

[0096] Step S1-3: Signal preprocessing: noise reduction and feature enhancement:

[0097] Frequency domain adaptive wavelet packet decomposition: frequency domain adaptive wavelet packet decomposition algorithm is used to perform 3-layer wavelet packet decomposition on the original signal to generate 8 frequency bands ( ), through the frequency band energy weight formula: ( For the The center frequency of the band, kHz is the damage sensitivity threshold, is the steepness factor, (segment length) automatically suppresses and higher than Environmental noise (such as mechanical vibration and wind noise) is eliminated, and the high-frequency energy components related to damage are retained;

[0098] Spatiotemporal collaborative noise reduction: Combined with the sensor array fusion algorithm (spatiotemporal collaborative positioning algorithm), spatiotemporal filtering is performed on the noise-reduced multi-channel signals:

[0099] (in, For sensors The calibration amplitude coefficient of For sensors and damage location estimate The Euclidean distance of For the sound wave to propagate to the sensor time delay; mm is the spatial attenuation coefficient; To prevent division by zero exception; Indicates sensor Acoustic emission signals collected), through the sensor spatial position Estimated location of damage source Euclidean distance weighting suppresses non-common source noise and improves signal-to-noise ratio (SNR increased by 15-20dB);

[0100] Step S1-4: Signal normalization and timing segmentation:

[0101] Amplitude normalization: Z-score normalization is performed on the single-channel signal after spatiotemporal fusion: (in, is the signal mean, is the standard deviation, and the calculation window takes a single load cycle to eliminate the influence of sensor sensitivity differences and signal propagation attenuation; represents the normalized signal; represents the single-channel acoustic emission signal after spatiotemporal fusion noise reduction processing);

[0102] Sliding window time series segmentation:

[0103] The sliding window algorithm is used to segment the normalized signal with a window length of 150 sampling points (corresponding to 0.15ms) and a step size of 10 sampling points (overlap rate 93%) to generate a local time series sample set. ( is the total number of samples); each sample Corresponding to a signal segment within a time window, and the number of cycles ——Correspondence, which facilitates subsequent time series alignment with physical damage labels (such as crack opening);

[0104] Step S1-5: Pre-processing quality verification:

[0105] Signal-to-noise ratio (SNR) evaluation: Calculate the SNR of the signal before and after preprocessing. The SNR after noise reduction must be ≥ 10 dB to ensure that impairment features (such as energy peaks and burst signals) can be effectively identified. (SNR is the ratio of signal power to noise power, expressed in decibels.)

[0106] Timing alignment accuracy: Through the synchronization check of the load cycle counter and the signal acquisition clock, the timestamp of the segmented samples is ensured to be consistent with the physical damage parameters (such as ) The time error is ≤0.1% of the sampling interval (i.e. ≤1μs, μs is microsecond, the time unit).

[0107] In this embodiment, the purpose of step S2 of time series segmentation and sample set construction is to perform time series structured segmentation on the preprocessed acoustic emission signal using a sliding window algorithm to generate a local time series sample set suitable for deep learning model input while preserving the signal's time dependency and damage feature continuity. The specific steps are as follows:

[0108] Step S2-1: Sliding window parameter setting:

[0109] Window length and step size selection: Based on the time scale of fatigue damage characteristics in acoustic emission signals (e.g., the duration of microcrack initiation signals is approximately 0.1-0.3 ms), the window length is set to 150 sampling points (corresponding to a physical duration of 0.15 ms when the sampling frequency is 1 MHz), ensuring that a single window contains the complete damage event waveform (e.g., the rising edge to the falling edge of the energy pulse);

[0110] The step size is set to 10 sampling points (i.e., 0.01ms), forming a sliding window sequence with a 93% overlap rate. This prevents damage features from being truncated by window boundaries and increases the number of samples (when the original signal length is L, the number of generated samples is approximately L / 10-14, which is 14 times higher than that of non-overlapping segmentation).

[0111] Physical meaning mapping: Each sliding window corresponds to an acoustic emission signal segment within a local time window, and its timestamp is aligned with the number of bridge load cycles n, providing a temporal benchmark for subsequent integration with physical damage stages (such as the mutation point of crack opening);

[0112] Step S2-2: Timing signal frame processing:

[0113] Framing process: for the pre-processed single-channel acoustic emission signal , from the starting sampling point First, generate the first samples :

[0114] , where m=1, 2, …, T (T is the total number of samples), and each sample is a one-dimensional time series vector with a dimension of 1×150;

[0115] Multi-channel signal processing: If a sensor array (e.g. 8 channels) is used, each channel is framed independently to generate a multi-channel sample set. , is the number of channels, which can be subsequently fed into the CNN-BiLSTM model as multidimensional input;

[0116] Step S2-3: Sample label association and quality control:

[0117] Physical damage label mapping:

[0118] Physical damage label mapping: per sample Corresponding to the unique number of load cycles , through the nonlinear dynamic threshold algorithm of step S4, Mapping to injury stage labels (e.g. 0 = initiation period, 1 = stable expansion period, 2 = instability and destruction period); the specific mapping rules are:

[0119] like Located before the mutation point of the second-order derivative of crack opening, the label is "crack initiation";

[0120] If it is between the mutation point and the instability threshold, the label is “stable expansion”;

[0121] After exceeding the instability threshold, the label is “instability failure”;

[0122] Sample quality filtering:

[0123] Eliminate samples whose signal energy is lower than 1 / 3 of the global mean (possibly invalid noise fragments);

[0124] Check label consistency: If the same window contains cycles that span damage stages (for example, the window covers before and after the mutation point), assign labels based on the stage with the largest proportion, or directly remove samples that span stages.

[0125] Step S2-4: Sample set division and storage:

[0126] Dataset partitioning strategy: Divide the generated sample set into training set, validation set and test set in a ratio of 7:2:1:

[0127] Training set: used for model parameter learning, containing balanced samples at different damage stages;

[0128] Validation set: used to adjust hyperparameters (such as learning rate, BatchSize) and monitor model overfitting;

[0129] Test set: used to evaluate the generalization ability of the model, using bridge structure data that did not participate in training;

[0130] Storage format: Samples are stored in a binary format (such as HDF5) containing:

[0131] Input features: (single channel) or (Multi-channel);

[0132] Damage Label: (one-hot encoding or integer labels);

[0133] Metadata: sampling frequency, sensor coordinates, corresponding cycle times wait;

[0134] Step S2-5: Visualization and analysis of time series features:

[0135] Sample feature statistics: Calculate statistics of the training set samples, such as mean, standard deviation, and energy distribution (RMS), to verify the feature differences at different damage stages:

[0136] Crack initiation stage: high frequency low energy pulse (RMS value );

[0137] Stable expansion period: medium frequency energy signal (RMS value ), accompanied by periodic energy fluctuations;

[0138] Instability and destruction period: low frequency high energy burst signal (RMS value ), the energy peak interval is shortened;

[0139] Time series waveform visualization: Randomly select samples from each stage, draw time domain waveforms and spectrum diagrams, and verify whether the sliding window effectively captures the features:

[0140] Time domain diagram: observe amplitude mutation points (such as step signals) and energy envelope;

[0141] Spectrum: Using FFT transform, confirm that the energy in the damage-sensitive frequency band (around 50 kHz) accounts for ≥ 60%.

[0142] In this embodiment, the convolution kernel size of the CNN layer is dynamically adjusted, specifically:

[0143] According to the change of the signal energy gradient at the current moment, the convolution kernel size is adjusted in real time, with the adjustment range being 5 to 15 sampling points;

[0144] The convolution kernel weight parameters are updated through online learning, and the bias term is inversely proportional to the local signal-to-noise ratio of the signal;

[0145] The signal energy gradient constrains its influence weight on the convolution kernel through the Sigmoid function;

[0146] The weight calculation of the attention mechanism layer is achieved in the following way:

[0147] Generate query vector and key vector based on the hidden state of BiLSTM output;

[0148] In the temporal attention calculation, the logarithmic weighted term of the accumulated acoustic emission energy in the time window is superimposed, and the time window length is 50 sampling points;

[0149] The energy weighting coefficients are optimized by back-propagation to enhance the focusing capability on high-energy damage events;

[0150] Furthermore, the purpose of step S3 hybrid deep learning model construction is to design and build a neural network architecture that integrates CNN, BiLSTM and attention mechanism to achieve spatiotemporal feature extraction of acoustic emission signals, long-term dependency modeling and key damage stage focus, providing core algorithm support for damage stage identification. The specific steps are as follows:

[0151] Step S3-1: CNN layer design: dynamic adaptive feature extraction:

[0152] 1.1. One-dimensional convolution kernel configuration:

[0153] Number and size of convolution kernels: 64 dynamic adaptive convolution kernels are set in the first layer, and the initial kernel size range is (sampling point), corresponding to physical duration (measured at 1MHz sampling rate), covering the characteristic time scales from microcrack initiation (nanoseconds) to macrocrack growth (microseconds);

[0154] Dynamic adjustment mechanism: convolution kernel size The signal energy gradient at the current moment Decide:

[0155] When the energy gradient surges (such as sudden damage signals), the kernel size is automatically reduced to 5-8 to capture high-frequency transient features;

[0156] When the energy gradient is gentle (e.g., in the stable expansion phase), the kernel size is increased to 10-15 to extract low-frequency trend features;

[0157] 1.2. Feature map calculation and output:

[0158] Dynamic convolution formula: (in, Dynamic convolution kernel size, For the The convolution kernel is in time step The weight parameters are updated through online learning; is the input signal amplitude; is a time-varying bias term, inversely proportional to the local signal-to-noise ratio; is the Sigmoid function, constraining the influence weight of energy gradient on convolution kernel);

[0159] Output dimension: After convolution operation, a 64-dimensional feature map is generated ( is the time series length), characterizing the local time domain characteristics of the signal (such as amplitude mutation, energy peak position);

[0160] Step S3-2: BiLSTM layer design: long-term temporal dependency modeling:

[0161] 2.1, Bidirectional loop network architecture:

[0162] Number of neurons and layers: A single-layer BiLSTM network with 128 hidden neurons (64 each in the forward and reverse directions) is used to capture the long-range temporal dependencies of acoustic emission signals during damage evolution (e.g., characteristic changes over hundreds of cycles from crack initiation to instability).

[0163] Input and output structure:

[0164] Input: Feature map output by CNN layer , the dimension is [batch size, time series length, 64];

[0165] Output: Bidirectional hidden state sequence , the dimension is [T,128], where 、 are the hidden states of the forward and reverse LSTM at time t respectively;

[0166] 2.2, Capability of capturing temporal features:

[0167] Long-range dependency modeling: Through the gating mechanism (forget gate, input gate, and output gate) of the LSTM unit, the evolution trend of damage characteristics over 1,000 time steps (corresponding to 1 second of physical duration) is memorized, solving the vanishing gradient problem of traditional RNNs.

[0168] Parallel time series processing: A bidirectional structure extracts contextual information from both past (forward) and future (reverse) time series, improving the accuracy of identifying damage stage boundaries (e.g., the transition from stable expansion to unstable failure).

[0169] Step S3-3: Attention mechanism layer design: key feature focus:

[0170] 3.1. Bimodal spatiotemporal attention architecture:

[0171] Query-key-value mechanism:

[0172] Query Vector : Hidden state by BiLSTM Generated by linear transformation, the dimension is ,in, is a learnable weight matrix;

[0173] Key vector matrix : Contains the hidden state projection of all moments, with dimension ;

[0174] Value vector : Directly use the BiLSTM hidden state as the value signal;

[0175] 3.2 Weight calculation and energy coupling:

[0176] Attention weight formula: (in, Time window Sampling point (50 s) the accumulated acoustic emission energy; is the energy-timing coupling coefficient, optimized by back-propagation);

[0177] Dual-modal fusion logic:

[0178] Timing-dependent branches: Compute semantic associations between hidden states to locate historical / future features related to the current injury stage;

[0179] Energy Intensity Branch: Convert the accumulated energy into a weight bias to force the model to focus on high-energy damage events (such as acoustic emission signals during unstable crack propagation);

[0180] 3.3. Feature weighting and output:

[0181] Context vector generation: (in, is time t versus time The attention weight, is the fused context vector; is the total timing length of the input signal);

[0182] Output dimension: Attention layer output ,characterizes the key feature sequence of fusion temporal correlation and energy intensity;

[0183] Step S3-4: Model integration and output layer design:

[0184] 3.4.1. Feature layer connection:

[0185] Cross-layer feature fusion: local features of CNN layer Global features with attention layer Splice to form a composite feature vector: , the dimension is , achieving the complementary features of "local details + global trends";

[0186] 3.4.2 Classifier Design

[0187] Fully connected layer configuration:

[0188] First fully connected layer: 192 neurons, ReLU activation function, used for feature dimensionality reduction and nonlinear transformation;

[0189] Output layer: 3 neurons (corresponding to 3 damage stages), activation function is Softmax, output probability distribution ;

[0190] Loss function: Use anti-noise adversarial loss function:

[0191] (in, It is the class weight to solve the sample imbalance; For true labels, The model predicts that the sample belongs to The probability of the class is in the range of (0,1); , is the anti-noise regularization coefficient; is the input disturbance, which obeys uniform distribution ; is the perturbation-resistant consistency constraint);

[0192] Step S3-5: Model hyperparameter initialization:

[0193] Optimizer selection: AdamW optimizer (with weight decay), initial learning rate , weight decay coefficient , suppress overfitting;

[0194] Regularization configuration:

[0195] Add a Dropout layer after the CNN layer and the fully connected layer with a dropout rate of 0.2;

[0196] Apply batch normalization to the BiLSTM layer weights to accelerate training convergence;

[0197] Input size adaptation: The model input accepts dimensions of [batch size, 150, 1] (single channel signal) or [batch size (Multi-channel signal, is the number of sensors), ensuring fixed-length input by padding or truncation;

[0198] Summary of key technical points:

[0199] Dynamic convolution kernel: breaks through the fixed kernel size limitation of traditional CNN and achieves adaptive capture of feature scales at different damage stages;

[0200] BiLSTM long time series modeling: The bidirectional recurrent structure effectively captures the damage evolution over hundreds of load cycles, outperforming unidirectional RNNs.

[0201] Bimodal attention: Combining temporal dependency and energy intensity, it solves the problem that hidden state-based attention alone cannot locate physical damage events.

[0202] Noise-resistant loss function: Through gradient penalty and adversarial perturbation, the robustness of the model in actual engineering noise environments is improved.

[0203] In this embodiment, the damage stage classification method includes:

[0204] Calculate the absolute value of the second-order derivative of the crack opening curve, and trigger the cutoff point judgment when it exceeds the basic threshold of 0.05;

[0205] The threshold of the demarcation point is dynamically adjusted. The adjustment amplitude is proportional to the variance of the crack opening in the last 50 cycles of loading. The maximum variance value is taken from the training set for normalization.

[0206] The interval between adjacent dividing points must meet the constraint of at least 100 cyclic loading times;

[0207] Furthermore, the purpose of constructing a training dataset for physical-data fusion in step S4 is to achieve objective division of damage stages and data annotation based on the temporal correspondence between the physical damage characteristics (crack opening) of the bridge structure and the acoustic emission signals, thereby constructing a training dataset with both physical interpretability and data-driven capabilities. The specific steps are as follows:

[0208] Step S4-1: Crack opening monitoring and damage evolution curve acquisition:

[0209] 1.1. Physical quantity monitoring solution:

[0210] Monitoring equipment: Use high-precision laser displacement meter (accuracy ± 0.01mm) or digital image correlation (DIC) technology to measure the opening of main cracks in bridges (such as weld cracks in steel box girders and mid-span cracks in concrete beams). Real-time monitoring is carried out, and the sampling frequency is synchronized with the acoustic emission signal (1 time / load cycle);

[0211] Data recording: record the complete fatigue test cycle and the number of cycles , forming a crack opening-cycle number curve , the typical curve shows nonlinear growth, which is divided into the initiation period (slow growth), stable expansion period (linear growth), and unstable destruction period (exponential growth);

[0212] 1.2. Characteristics of damage evolution stages:

[0213] Crack initiation period: (The threshold is usually , corresponding to the initial formation of microcracks, the acoustic emission signal is mainly low-energy high-frequency pulses:

[0214] Stable expansion period: ( is the critical value of instability, such as ), the crack expands at a uniform speed, and the energy of the acoustic emission signal fluctuates periodically;

[0215] Instability and destruction period , the cracks rapidly propagate to structural failure, and the acoustic emission signal shows a high-energy sudden event;

[0216] Step S4-2: Damage stage classification based on nonlinear dynamic threshold:

[0217] 2.1. Second-order derivative mutation point detection:

[0218] Mathematical Principle: Second Derivative of Crack Opening Reflects the change in damage evolution rate. The mutation point corresponds to the boundary of the damage stage. It is determined by a nonlinear dynamic threshold algorithm. The cutoff point determination conditions are: (in, Is the basic threshold used to filter noise interference; For window The variance of the opening within the ,characterizes the degree of local fluctuation; is the maximum variance value of the training set, used for normalization);

[0219] Calculation process:

[0220] right Perform 5-point sliding average filtering to suppress measurement noise;

[0221] Calculate the first derivative With the second derivative ;

[0222] When a point When the above inequality is satisfied, it is determined to be a stage dividing point (such as the initiation-stable expansion dividing point , Stable expansion-instability destruction dividing point );

[0223] 2.2 Physical meaning of dynamic threshold:

[0224] Variance normalization term: Automatically adapt to the damage fluctuation characteristics of different bridge structures (such as the difference in crack growth stability between steel bridges and concrete bridges), avoiding the limitations of fixed thresholds;

[0225] Phase division results:

[0226] Germination period: ;

[0227] Stable expansion period ;

[0228] Instability and destruction period: ( is the total number of cycles);

[0229] Step S4-3: Aligning the acoustic emission signal with the timing of the damage stage:

[0230] 3.1. Sample label mapping:

[0231] Timestamp corresponds to:

[0232] Each acoustic emission signal sample (Generated in step 2, corresponding to the number of cycles )according to The damage stage label is assigned to the interval :

[0233] : Initiation period;

[0234] : Stable expansion period;

[0235] : Instability and destruction period;

[0236] Cross-stage sample processing: If the time window corresponding to the sample covers multiple stage demarcation points (such as lie in loop), then:

[0237] Calculate the proportion of cycles in each stage within the window, and take the stage with the largest proportion as the label;

[0238] Or remove such samples to ensure label uniqueness;

[0239] 3.2 Data balancing strategy:

[0240] Oversampling and undersampling: If the number of samples in a certain stage is too small (for example, the proportion of samples in the unstable and destructive period is less than 10%), the SMOTE algorithm is used to generate synthetic samples, or the majority class (such as the stable expansion period) is randomly undersampled to make the sample ratio of each stage close to 1:1:1;

[0241] Weight distribution:

[0242] Combined category weights ( For the The number of class samples) gives higher weight to the minority class in the loss function to alleviate the impact of class imbalance on model training;

[0243] Step S4-4: Physical-data fusion dataset construction:

[0244] 4.1、Dataset structure:

[0245] Input features: acoustic emission signal time series sample input features (single channel) or (multi-channel);

[0246] Label vector: injury stage label (integer encoding or one-hot encoding);

[0247] Auxiliary physical parameters:

[0248] The number of cycles for each sample

[0249] Crack opening and its first-order and second-order derivatives 、 , used for model interpretability analysis (such as the response of the attention mechanism to physical parameters);

[0250] 4.2 Dataset Verification

[0251] Time series consistency test: Randomly select 100 samples to verify the consistency between the damage stage corresponding to their labels and the characteristics of the acoustic emission signal (for example, samples in the unstable damage period should contain high-energy pulses);

[0252] Physical logic verification: Draw a comparison chart between the model prediction stage and the actual crack opening curve to ensure that the prediction results conform to the unidirectional evolution logic of "initiation → expansion → instability" and there is no reverse prediction.

[0253] In this embodiment, the dynamic learning rate adjustment strategy is implemented by the following steps:

[0254] The cosine annealing algorithm is used, with an initial learning rate of 0.001, a minimum learning rate of 0.0001, and a cycle length of 10 training rounds;

[0255] A gradient change sensitivity factor is introduced, using the modulus ratio of the current gradient vector to the previous step gradient as the learning rate scaling factor, and the hyperbolic tangent function is used to suppress oscillations caused by gradient mutations;

[0256] Furthermore, the purpose of step S5, hybrid model training and robustness optimization, is to efficiently train the CNN-BiLSTM-Attention model through dynamic learning rate adjustment, early stopping mechanism, and regularization technology, thereby improving the model's generalization ability and convergence stability in noisy environments. The specific steps are as follows:

[0257] Step S5-1: Dynamic learning rate adjustment strategy: gradient-sensitive cosine annealing:

[0258] 1.1 Algorithm Principle:

[0259] A gradient-sensitive cosine annealing algorithm is used, combining the periodic learning rate decay of traditional cosine annealing with the gradient change sensitivity factor to suppress training oscillations and accelerate convergence; the update formula is:

[0260]

[0261] (in, 、 are the upper and lower bounds of the learning rate; is the number of iterations in the current cycle; is the maximum number of iterations for each cycle; is the current parameter gradient vector; is a numerical stability constant; is the gradient change sensitivity factor);

[0262] 1.2 Dynamic Adjustment Mechanism:

[0263] Cosine annealing based decay: The learning rate decays in a cosine curve in each cycle, from down to , periodically restart to escape from local optimum;

[0264] Gradient sensitivity factor:

[0265] when (Gradient increases, parameter oscillation may occur), When the value approaches 1, the learning rate decays faster, suppressing unstable updates;

[0266] when (gradient is stationary), The value approaches 0, the learning rate remains high, and the convergence is accelerated;

[0267] Step S5-2: Early stopping mechanism and training process monitoring:

[0268] 2.1、Termination condition setting:

[0269] Validation set performance monitoring: After each training cycle (Epoch), calculate the loss value on the validation set and accuracy ;

[0270] Early stopping trigger condition: If the validation set loss continues cycles (typical ) does not decrease, or the accuracy does not improve, then terminate the training to prevent overfitting;

[0271] 2.2、Visualization of training process:

[0272] Draw a learning curve: record the loss of the training set and the validation set, and the accuracy change trend over time to determine whether the model has converged or is overfitting:

[0273] Ideal state: training loss and validation loss decrease simultaneously, and the gap is small;

[0274] Overfitting warning: The training loss decreases but the validation loss increases, and the regularization parameters need to be adjusted;

[0275] Step S5-3: Regularization technology improves robustness:

[0276] 3.1. Dropout layer application: Insert the Dropout layer after the CNN layer and the BiLSTM layer. The specific configuration is as follows:

[0277] After the CNN layer: dropout rate , randomly discard 20% of the convolution kernel outputs to suppress overfitting of the convolution layer;

[0278] After the BiLSTM layer: dropout rate , randomly discard 30% of the hidden neuron connections to enhance the generalization ability of the recurrent network;

[0279] Batch Normalization: Add a batch normalization layer after each convolutional layer and fully connected layer to standardize the input: (in, 、 is the batch mean and variance; 、 are learnable scaling and offset parameters); this operation alleviates the "internal covariate shift" problem, accelerates training convergence, and allows the use of higher learning rates;

[0280] Step S5-4: Anti-noise adversarial training:

[0281] 4.1. Loss function design:

[0282] Adopting the noise-resistant adversarial loss function, we introduce gradient penalty and adversarial perturbation terms based on the standard cross entropy loss:

[0283] (in, is the category weight; is the gradient penalty coefficient, which suppresses input noise sensitivity; is the anti-disturbance coefficient; To obey the uniform distribution input disturbance);

[0284] Countermeasures against disturbance injection:

[0285] In each training batch, the input signal Superimposed random disturbances , forcing the model to be insensitive to small input changes and improving noise resistance;

[0286] By calculating the gradient norm of the output with respect to the input , penalizes areas where the gradient is too large, forcing the model to learn smooth decision boundaries;

[0287] Step S5-5: Hyperparameter tuning and model validation:

[0288] Hyperparameter search space:

[0289] The main tuning parameters include learning rate (scope , Dropout rate (range 0.2-0.4), batch size (range 32-128), attention coupling coefficient (range 0.1-1.0), etc. The optimal combination is selected by evaluating the validation set performance through random search combined with 5-fold cross validation;

[0290] Final model verification:

[0291] The model's accuracy, precision, recall, and F1 score are calculated using test data that was not used in training. Typical performance indicators require an accuracy rate of ≥85% for the crack initiation phase, ≥90% for the stable propagation phase, and ≥95% for the unstable failure phase.

[0292] Gaussian white noise (SNR = 5dB) was added to the test signal to verify that the reduction in recognition accuracy during the damage phase of the model was ≤10%, ensuring reliability in actual engineering applications.

[0293] Summary of key technical points:

[0294] Dynamic learning rate optimization: By combining the gradient sensitivity factor with cosine annealing, global search and local optimization are balanced, and the training speed is increased by more than 30% compared to a fixed learning rate.

[0295] Early stopping mechanism and regularization: effectively prevent overfitting, and the fluctuation range of validation set loss is controlled within ±5%;

[0296] Noise-resistant adversarial training: significantly improves the model's robustness to input noise, maintaining accuracy ≥ 90% in noisy environments;

[0297] Hyperparameter engineering tuning: Through systematic search and cross-validation, we ensure that the model can converge stably on data from different bridge types.

[0298] In this embodiment, the purpose of step S6, real-time damage stage identification and result output, is to input the real-time collected acoustic emission signals into the trained hybrid deep learning model, realize real-time classification of the fatigue damage stage of the bridge structure through online reasoning, and output visualization results to provide a decision-making basis for structural health monitoring. The specific steps are as follows:

[0299] Step S6-1: Real-time acoustic emission signal acquisition and preprocessing:

[0300] 1.1. Online collection system deployment:

[0301] An acoustic emission sensor array is installed in the designated monitoring area of the bridge (e.g., near the welds of steel box girders and concrete beam supports). The signal is transmitted in real time to an edge computing unit or cloud server via a wired (e.g., Ethernet) or wireless (e.g., 5G) transmission link. The acquisition parameters remain the same as those in the training phase: a sampling frequency of 1 MHz, a preamplifier gain of 40 dB, and a signal-to-noise ratio of ≥ 15 dB.

[0302] 1.2. Streaming data preprocessing:

[0303] Real-time noise reduction: Apply frequency-domain adaptive wavelet packet decomposition algorithm to real-time streaming signals to dynamically suppress environmental noise (such as traffic noise and structural vibration) while retaining the energy of the 50kHz±20kHz damage-sensitive frequency band. Specifically, the frequency band energy weights are calculated as follows:

[0304] (in, For the The wavelet packet coefficients of the frequency band, is the center frequency of the band, kHz is the damage sensitivity threshold, is the steepness factor, is the segment length), suppressing low-weight frequency band noise;

[0305] Standardization: Calculate the mean and standard deviation of the real-time signal through a sliding window and perform Z-score normalization: (in, 、 is the signal mean and standard deviation within the current 10-second window);

[0306] Time series segmentation: Use sliding window algorithm (window length 150 sampling points, step length 10 sampling points) to segment the normalized signal in real time to generate a continuous sample stream ,Each sample corresponds to 0.15ms physical duration, and the timestamp is synchronized with the number of load cycles;

[0307] Step S6-2: Lightweight model online inference:

[0308] 2.1 Model compression and deployment:

[0309] Hierarchical knowledge distillation: The feature maps and output probabilities of the teacher model are transferred to the student model through the hierarchical knowledge distillation algorithm. The specific loss function is:

[0310] (in, 、 is the l-th layer feature map of the teacher / student model, is the feature normalization function, is the distillation weight), the model parameters are reduced by 60%, and it is suitable for embedded devices;

[0311] Inference optimization: Use TensorRT to quantize and compile the model, fuse the convolutional layer and BiLSTM layer into CUDA kernel, and achieve single-sample inference latency of ≤3ms (measured on NVIDIA Jetson devices).

[0312] 2.2 Real-time prediction process:

[0313] Continuous batch inference: Real-time samples are fed into the model in batches of 500 samples (corresponding to a duration of 5 seconds), and the model outputs the probability vectors for each damage stage. ;

[0314] Stage judgment logic:

[0315] Instability and destruction period: If If it lasts for 3 batches (15 seconds), it is considered an emergency state;

[0316] Stable expansion period: If and , it is determined to be in monitoring upgrade state;

[0317] Initiation / Normal: Other situations are considered as the initial stage or normal operation state;

[0318] Step S6-3: Multi-dimensional result visualization:

[0319] 3.1. Real-time monitoring dashboard:

[0320] Dynamic Time Series Graph: Draws a real-time damage stage sequence on the monitoring interface (in the form of a timeline, with green / yellow / red corresponding to initiation / extension / instability, respectively), and superimposes the root mean square (RMS) curve of the acoustic emission signal to intuitively display the relationship between energy changes and stages;

[0321] Key indicators display:

[0322] Real-time prediction probability: , , (accurate to 3 decimal places);

[0323] Physical parameter mapping: Real-time estimation of crack opening through regression model established by historical training data ;

[0324] 3.2. Historical data backtracking:

[0325] Database storage: Real-time samples, prediction results, and raw sensor signals are stored in a time series database (such as TimescaleDB) by timestamp, with a storage frequency of 1 sample per second. This supports distributed expansion and can meet data storage requirements for more than five years.

[0326] Interactive analysis tool: Provides waveform replay function. Users can select any time interval to view the corresponding acoustic emission signal time domain waveform, spectrum diagram and model attention weight distribution to assist in analyzing damage characteristics;

[0327] Step S6-4: Intelligent early warning and model calibration:

[0328] 4.1. Multi-level early warning response:

[0329] Three-level early warning system:

[0330] Level 1 warning (green): In the germination stage, a daily monitoring report is generated, recording the sample energy mean and stage proportion;

[0331] Level 2 warning (yellow): Stable expansion period, triggering SMS notification to monitoring engineers, and recommending weekly manual inspections;

[0332] Level 3 warning (red): In the instability and damage period, an audible and visual alarm is immediately triggered, which is pushed to the bridge management system through the API interface and an emergency response plan is automatically generated;

[0333] 4.2 Model online calibration:

[0334] Physical verification mechanism: Actual crack data is obtained monthly through drone visual inspection or laser crack measuring instruments and compared with the model prediction results. If the error exceeds 20%, the multi-domain adaptive migration algorithm is automatically triggered to fine-tune the model using the new data. The adjustment period is ≤ 2 hours. The specific objective function is:

[0335] (in, is the maximum mean difference between the source domain and the target domain, 、 is the alignment coefficient; is the gradient operator, is the mapping function of the deep learning model (such as CNN-BiLSTM-Attention), Represents the model output with respect to the input signal The gradient of the input signal, that is, the degree to which a small change in each dimension of the input signal affects the model output);

[0336] Cross-device synchronization: When deployed on multiple bridges, a federated learning framework is used to aggregate and update the global model while protecting the data privacy of each bridge, improving cross-domain generalization capabilities.

[0337] Summary of key technical points:

[0338] Real-time architecture: Through edge computing and model compression, the entire process of "collection-processing-prediction" is delayed to ≤ 200ms, meeting the real-time requirements of online monitoring.

[0339] Lightweight model design: Combining knowledge distillation with hardware acceleration enables the model to run on low-power embedded devices, reducing deployment costs;

[0340] Human-machine collaborative decision-making: Through a visual interface and multi-level warning, AI prediction results are combined with human experience to improve the reliability of the monitoring system;

[0341] Self-evolution capability: Based on transfer learning and online calibration, the model can adapt to the damage characteristics of different bridge structures, reducing retraining costs.

[0342] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A structural fatigue damage identification method based on acoustic emission and deep learning, characterized by: The following steps are involved: S1. Acoustic emission signals under fatigue load are collected by an acoustic emission sensor array arranged on the surface of the bridge structure, and the signals are preprocessed, including noise reduction, normalization and segmentation processing; S2. Use the sliding window segmentation algorithm to perform time series segmentation on the preprocessed acoustic emission signal, with a window length of 150 sampling points and a step length of 10 sampling points to generate a local time series sample set; S3. Construct a CNN-BiLSTM-Attention hybrid deep learning model, wherein the CNN-BiLSTM-Attention hybrid deep learning model includes: CNN layer: uses a one-dimensional convolution kernel to extract the local time domain features of the acoustic emission signal, including amplitude mutation and energy peak features; BiLSTM layer: captures long-term temporal dependencies in the damage evolution process through a bidirectional long short-term memory network; Attention mechanism layer: Adaptively focuses on high-energy signal features at critical injury stages based on a learnable weight matrix; S4. The damage stages are divided based on a nonlinear curve of the bridge main crack opening and the number of cycles. The stage labels are determined by the slope mutation points. The labels are aligned with the acoustic emission signal time series to construct a physical-data fusion training dataset. S5. Use dynamic learning rate adjustment strategy and early stopping mechanism to train the hybrid model, and combine Dropout and batch normalization operations to improve model robustness; S6. Input the real-time collected acoustic emission signals into the trained model and output the damage stage identification results, including crack initiation, stable expansion and unstable failure stages.

2. The method for structural fatigue damage identification based on acoustic emission and deep learning according to claim 1, characterized in that: The convolution kernel size of the CNN layer is dynamically adjusted, specifically: According to the change of the signal energy gradient at the current moment, the convolution kernel size is adjusted in real time, with the adjustment range being 5 to 15 sampling points; The convolution kernel weight parameters are updated through online learning, and the bias term is inversely proportional to the local signal-to-noise ratio of the signal; The signal energy gradient is constrained by the Sigmoid function to constrain its influence on the convolution kernel.

3. The method for structural fatigue damage identification based on acoustic emission and deep learning according to claim 1, characterized in that: The weight calculation of the attention mechanism layer is achieved in the following way: Generate query vector and key vector based on the hidden state of BiLSTM output; In the temporal attention calculation, the logarithmic weighted term of the accumulated acoustic emission energy in the time window is superimposed, and the time window length is 50 sampling points; The energy weighting coefficients are optimized via back-propagation to enhance the focusing capability on high-energy damage events.

4. The method for structural fatigue damage identification based on acoustic emission and deep learning according to claim 1, characterized in that: The damage stage division method includes: Calculate the absolute value of the second-order derivative of the crack opening curve, and trigger the cutoff point judgment when it exceeds the basic threshold of 0.05; The threshold of the demarcation point is dynamically adjusted. The adjustment amplitude is proportional to the variance of the crack opening in the last 50 cycles of loading. The maximum variance value is taken from the training set for normalization. The interval between adjacent dividing points must meet the constraint of at least 100 cyclic loading times.

5. The method for structural fatigue damage identification based on acoustic emission and deep learning according to claim 1, characterized in that: The dynamic learning rate adjustment strategy is implemented by the following steps: The cosine annealing algorithm is used, with an initial learning rate of 0.001, a minimum learning rate of 0.0001, and a cycle length of 10 training rounds; A gradient change sensitivity factor is introduced, and the modulus ratio of the current gradient vector to the previous step gradient is used as the learning rate scaling coefficient. The hyperbolic tangent function is used to suppress the oscillation caused by gradient mutation.

6. The method for structural fatigue damage identification based on acoustic emission and deep learning according to claim 1, characterized in that: In the pre-processing step, the frequency domain adaptive wavelet packet decomposition algorithm is used to decompose the acoustic emission signal, and the frequency band energy weight calculation includes: Perform wavelet packet decomposition on the acoustic emission signal and calculate the energy proportion of each frequency band; Apply exponential weight enhancement to the frequency band with a center frequency higher than 50kHz. The weight increases as the difference between the frequency and the threshold increases, and the steepness factor is 0.

1. The signal segment length is fixed at 150 samples.

7. The method for identifying structural fatigue damage based on acoustic emission and deep learning according to claim 1, characterized in that: The signal fusion method of the acoustic emission sensor array is: The spatial attenuation weight is calculated based on the Euclidean distance between the sensor and the damage location estimate, with an attenuation coefficient of 100 mm; When superimposing the signals of each sensor, the timing is aligned according to the acoustic wave propagation time delay and multiplied by the sensor calibration amplitude coefficient; The distance weight denominator is added with a very small constant 1×10 -5 Prevent division by zero exceptions.

8. The method for identifying structural fatigue damage based on acoustic emission and deep learning according to claim 1, characterized in that: The CNN-BiLSTM-Attention hybrid deep learning model training loss function includes the following three items: Category-weighted cross entropy loss, where weights are inversely proportional to the number of category samples; Input signal gradient penalty term, with a coefficient of 0.5, is used to suppress noise sensitivity; Against the perturbation consistency constraint, a uniformly distributed perturbation is applied to the input signal and the output distribution difference is calculated, with a coefficient of 0.

3.

9. The method for identifying structural fatigue damage based on acoustic emission and deep learning according to claim 1, characterized in that: The CNN-BiLSTM-Attention hybrid deep learning model is deployed using a layered knowledge distillation algorithm, including: The corresponding layer feature maps of the teacher model and the student model are L2 normalized and the difference loss is calculated; The KL divergence of the output distribution is constrained synchronously, and the knowledge distillation weight coefficient is 0.1; A very small constant is added to the denominator of the characteristic normalization function to prevent division by zero.

10. The method for identifying structural fatigue damage based on acoustic emission and deep learning according to claim 1, characterized in that: The cross-bridge type generalization method uses a multi-domain adaptive migration algorithm, including: The maximum mean difference between the acoustic emission signal distributions in the source and target domains was calculated, with a domain alignment coefficient of 0.1; Apply L2 regularization constraint to the input signal gradient with a coefficient of 0.01 to improve cross-domain stability; The training data must cover at least five bridge types, three load spectra, and temperature-humidity coupling environments.

Images