A bridge expansion joint damage probability diagnosis method based on physical knowledge guided voiceprint features

Abstract

The application discloses a bridge expansion joint damage probability diagnosis method based on physical knowledge guided voiceprint features, and steps are as follows: collecting acoustic response signals generated when a vehicle passes through an expansion joint, obtaining corresponding time-frequency response samples, and further extracting physical features representing the service state of the expansion joint; establishing a physical knowledge constraint according to the influence law of the expansion joint damage state on the vehicle impact response; inputting the time-frequency response samples and the physical features into a fault diagnosis model respectively, and introducing a physical consistency loss function in the model training process, so that the prediction result of the fault diagnosis model is consistent with the physical knowledge constraint; using the trained fault diagnosis model to identify the acoustic response signals of the bridge expansion joint to be detected, and outputting the expansion joint state category, state probability and diagnosis uncertainty. The application can effectively improve the accuracy, robustness and interpretability of the acoustic diagnosis result of the bridge expansion joint, and is suitable for the detection and maintenance evaluation of small and medium-sized bridge expansion joints.

Description

Technical Field

[0001] This invention belongs to the field of bridge structural health monitoring and intelligent detection technology. It combines acoustic signal processing and uncertainty quantification methods to carry out intelligent fault diagnosis of bridge expansion joints guided by physical knowledge. Specifically, it involves a bridge expansion joint damage probability diagnosis method based on physical knowledge-guided acoustic signature features. Background Technology

[0002] Bridge expansion joints are crucial components of the bridge superstructure, designed to accommodate deformation caused by beam temperature changes, concrete shrinkage and creep, and vehicle loads. Expansion joints are constantly exposed to complex environments including vehicle impacts, rainwater erosion, temperature fluctuations, and material aging, making them prone to defects such as rubber strip damage, steel loosening, anchorage zone detachment, localized fractures, and abnormal impact noise. These defects not only affect driving comfort but can also exacerbate pavement damage, beam end concrete deterioration, and the failure of bridge components.

[0003] Current expansion joint inspection methods mainly rely on manual inspection, image recognition, contact vibration testing, and a limited number of long-term monitoring methods. Manual inspection is inefficient and highly subjective; image recognition is easily affected by dirt, obstructions, lighting, and shooting angle; contact vibration testing requires sensor installation, resulting in high on-site deployment and maintenance costs. Vehicles passing over expansion joints generate significant transient acoustic responses. Acoustic detection, with its non-contact, low-cost, and easy-to-deploy characteristics, is suitable for rapid screening of expansion joints in small and medium-sized bridges. In recent years, acoustic diagnostic methods based on machine learning and deep learning can automatically extract discriminative features from acoustic signals, showing potential application in expansion joint condition identification. However, purely data-driven models mainly rely on sample labels and statistical correlations, lacking explicit constraints on vehicle impact and the structural response of expansion joints. When vehicle type, speed, tire condition, environmental noise, or bridge conditions change, the model may produce unstable predictions or even give physically unreasonable diagnostic results. Furthermore, traditional classification models typically only output state categories, making it difficult to quantify the reliability of diagnostic results, which is detrimental to verification and decision-making in engineering maintenance.

[0004] This case arises to address the problems in existing bridge expansion joint fault diagnosis methods, such as acoustic feature extraction relying on experience, lack of physical mechanism constraints in the diagnostic model, unstable identification results due to vehicle excitation differences, and difficulty in quantifying the reliability of diagnostic results. Summary of the Invention

[0005] The purpose of this invention is to provide a bridge expansion joint damage probability diagnosis method based on physical knowledge-guided acoustic signature features. By constructing time-frequency response samples, extracting physical features, establishing physical knowledge constraints, fusing deep features with physical features, and introducing a diagnostic uncertainty discrimination mechanism, an intelligent diagnostic process oriented towards vehicle impact response is formed. This allows the diagnostic model to not only learn discrimination rules from data, but also be constrained by physical knowledge during training and inference, thereby improving the accuracy of expansion joint fault identification, stability across vehicle operating conditions, and reliability of engineering applications.

[0006] To achieve the above objectives, the solution of the present invention is:

[0007] A method for diagnosing bridge expansion joint damage probability based on physical knowledge-guided acoustic signature features includes the following steps:

[0008] Step 1: Install acoustic sensors near the bridge expansion joints to collect the acoustic response signals generated when vehicles pass through the expansion joints, and determine the vehicle passage event based on the short-time energy and impact peak value of the acoustic response signals.

[0009] Step 2: Perform bandpass filtering, vehicle passage event segmentation, amplitude normalization, and short-time Fourier transform on the acoustic response signal to obtain the time-frequency response sample corresponding to the vehicle passing through the expansion joint.

[0010] Step 3: Based on the transient response mechanism generated when a vehicle tire impacts an expansion joint, extract the impact peak value, frequency band energy ratio, dominant frequency position, energy decay rate, and impact duration from the time-frequency response sample as physical characteristics characterizing the service status of the expansion joint.

[0011] Step 4: Based on the influence of expansion joint damage state on vehicle impact response, establish physical knowledge constraints consisting of equivalent stiffness degradation constraints, damping attenuation constraints, and impact energy redistribution constraints.

[0012] Step 5: The fault diagnosis model includes a deep feature extraction network, a physical feature encoding network, a feature fusion layer, and an evidence classification network. The time-frequency response samples are input into the deep feature extraction network, the physical features are input into the physical feature encoding network, and the deep features and physical features are fused and then input into the evidence classification network. A physical consistency loss function is introduced during the model training process to ensure that the prediction results of the fault diagnosis model are consistent with the physical knowledge constraints.

[0013] Step 6: Use the trained fault diagnosis model to identify the acoustic response signal of the bridge expansion joint to be detected, and output the expansion joint state category, state probability and diagnostic uncertainty.

[0014] The specific process of step 1 above is as follows:

[0015] Step 11: Deploy acoustic sensors at the transverse ends of the expansion joint, the inside of the guardrail, or the edge of the bridge deck; the acquired acoustic response signals are... The sampling frequency is ,

[0016]

[0017] In the formula, Represents the acoustic response signal. Indicates the first The sound pressure amplitude or normalized amplitude at each sampling point Indicates the acoustic signal sampling frequency. Indicates the total number of sampling points. Represents a continuous-time variable. Indicates the discrete sampling sequence number;

[0018] Step 12: Identify vehicle passage events from the collected acoustic response signals and calculate the short-time energy function. and short-time peak impact function ,

[0019]

[0020]

[0021] In the formula, Indicates the length of the short-time analysis window. This indicates the sampling point number within the analysis window and the corresponding short-time energy threshold. and short-term impact peak threshold It is determined based on the statistical characteristics of the background noise segment when no vehicles are passing, specifically expressed as follows:

[0022]

[0023]

[0024] In the formula, and Let represent the mean and standard deviation of the short-time energy of the background noise segment, respectively. and These represent the mean and standard deviation of the short-time impact peak value in the background noise segment, respectively. and These represent the adjustment coefficients for the short-time energy threshold and the short-time impact peak threshold, respectively, used to control the sensitivity of vehicle passage event recognition;

[0025] when and At that time, the judgment of the first A vehicle passage event exists near each sampling point; sampling points that consecutively meet the above conditions are merged into the same vehicle passage event, and the corresponding acoustic response segment is extracted based on the start and end positions of the event, denoted as the i-th segment. One vehicle passes through the event sample;

[0026] Step 13, the The peak impact moment of the event involving individual vehicles. Determined according to the following formula,

[0027]

[0028] In the formula, Indicates the first The peak impact moment corresponding to the event of a vehicle passing through. Indicates the first The time interval in which the vehicle passed through the event.

[0029] The specific process of step 2 above is as follows:

[0030] Step 21, based on the peak impact time Centered on the acoustic response signal, a fixed-length segment is extracted to obtain the first... Acoustic segment of a vehicle passing through an expansion joint ,

[0031]

[0032] In the formula, This indicates the length of time intercepted before the peak impact moment. This indicates the length of time taken after the peak impact moment;

[0033] Step 22: Perform bandpass filtering on the acoustic segment.

[0034]

[0035] In the formula, Indicates the first bandpass filter. An acoustic segment, This represents the bandpass filter operator. This indicates the cutoff frequency under bandpass filtering. Indicates the upper cutoff frequency of the bandpass filter;

[0036] Step 23: Normalize the amplitude of the filtered acoustic segment.

[0037]

[0038] In the formula, Represents the normalized i-th An acoustic segment, This indicates the use of extremely small positive numbers to prevent the denominator from being zero; a short-time Fourier transform is used to convert the normalized acoustic segment into a time-frequency response sample.

[0039]

[0040] In the formula, Indicates the first The time-frequency response of each sample. This indicates the time corresponding to the center position of the short-time Fourier transform window. Represents frequency variables. Indicates the sampling point number within the window. This indicates the short-time Fourier transform window length. Represents the window function. It represents the imaginary unit.

[0041] The specific process of step 3 above is as follows: when the vehicle tires pass through the expansion joint, the following physical characteristics are extracted.

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048] In the formula, Indicates the first The peak impact value of each sample Represents the normalized i-th An acoustic segment; Indicates the first The frequency band energy ratio of each sample within the target frequency band and Indicates the lower and upper limits of the target frequency band. Indicates the first Each sample in time and frequency Time-frequency energy at that location This indicates the time corresponding to the center position of the short-time Fourier transform window. Represents frequency variables; Indicates the sampling frequency of the acoustic response signal; Indicates the first The dominant frequency or spectral centroid position of each sample; Indicates the first The envelope signal of each sample, Represents the Hilbert transform; Indicates the first Energy decay rate of each sample Indicates to Over time The slope is obtained by linear fitting of the relationship between the changes. This indicates the duration of the decay rate fitting time. Indicates the first The impact duration of each sample and These represent the start and end times of the impact duration, respectively;

[0049] The above physical characteristics are combined into a physical feature vector. ,

[0050]

[0051] In the formula, Indicates the first The physical feature vector of each sample This represents the vector transpose symbol.

[0052] The above-mentioned Each sample in time and frequency Time-frequency energy at the location According to the following formula,

[0053]

[0054] In the formula, Indicates the first The time-frequency response of each sample.

[0055] The specific process of step 4 above is as follows:

[0056] Step 41: The local response of the vehicle tire impacting the expansion joint is equivalent to a single-degree-of-freedom transient forced vibration process, resulting in...

[0057]

[0058] In the formula, This indicates the local equivalent mass of the expansion joint. This indicates the local equivalent damping of the expansion joint. This indicates the local equivalent stiffness of the expansion joint. This represents the local equivalent displacement response. Indicates the local equivalent velocity response. This represents the local equivalent acceleration response. This refers to the external excitation generated when a vehicle's tires pass through an expansion joint. Additional impact excitation caused by damage to expansion joints;

[0059] Step 42, let the degree of damage to the expansion joint be... , Indicates health status. This indicates a severely damaged state; the corresponding equivalent stiffness and equivalent damping degradation relationship is expressed as follows:

[0060]

[0061]

[0062] In the formula, Indicates the degree of damage to the expansion joint. Represents the equivalent stiffness under healthy conditions. This represents the equivalent damping under healthy conditions. This represents the stiffness degradation coefficient. Indicates the damping degradation coefficient;

[0063] Step 43, the characteristic frequency of the expansion joint is expressed as,

[0064]

[0065] In the formula, Indicates the degree of damage. The equivalent characteristic frequency at time, Represents pi;

[0066] Step 44, Construct the physical damage index It is used to characterize the degree of damage to expansion joints inferred from physical characteristics.

[0067]

[0068]

[0069] In the formula, Indicates the first The physical damage index corresponding to each sample , , , , Represents the standardized physical characteristics. This indicates the weighting of the physical damage index calculation. This represents the Sigmoid mapping function. This represents the input variables for the Sigmoid function.

[0070] The specific process of step 5 above is as follows:

[0071] Step 51, the deep feature extraction network uses time-frequency response samples Input, output deep features Physical feature encoding networks use physical feature vectors As input, output physical coding features ,

[0072]

[0073]

[0074]

[0075] In the formula, The parameter is Deep feature extraction network, The parameter is Physical feature encoding network, Indicates the first Depth features of each sample Indicates the first The physical coding features of each sample This represents the features of the fused samples. Indicates the feature concatenation operator;

[0076] Step 52, the evidence classification network uses fused features As input, output the non-negative evidence value corresponding to each expansion joint state category.

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] In the formula, Indicates the first Non-negative evidence vectors for each state category of each sample This represents the activation function that guarantees a non-negative output. and Let represent the weight matrix and bias vector of the evidence classification network, respectively. Indicates the first Dirichlet distribution parameters of a sample Indicates the first The first sample Dirichlet parameters corresponding to class states Represents the Dirichlet distribution parameters and, Indicates the number of state categories. Indicates the first The sample belongs to the first Predicted probability of class state Indicates the first Diagnostic uncertainty for the nth sample; let the nth sample be... The true label for each sample is , Indicates sample Does it belong to the first Class state; Dirichlet evidence loss is used for evidence classification loss.

[0083]

[0084] In the formula, Indicates loss in evidence classification. Indicates the number of training samples. Indicates the first The first sample The actual label indicator variable for class state, Represents the Digamma function;

[0085] Step 53: To ensure that the damage level output by the constraint model is consistent with the physical feature inference results, define the model-predicted damage index. ,

[0086]

[0087] In the formula, Indicates the first The damage index is obtained by predicting the probability of each sample using the model. This represents the damage level code corresponding to the c-th state, with 0 for a healthy state and 1 for a damaged state; thus, a physical consistency loss function is constructed:

[0088]

[0089] In the formula, Represent the physical consistency loss function; and construct a monotonic constraint loss as follows:

[0090]

[0091] In the formula, This represents the monotonic constraint loss. Represents the linear rectified function. This represents the partial derivative of the model's predicted damage index with respect to the physical characteristics.

[0092] Step 54, define the Dirichlet regularization loss.

[0093]

[0094]

[0095] In the formula, This represents the Dirichlet regularization loss. Indicates the Kullback-Leibler divergence. Represents the Dirichlet distribution. This represents the Dirichlet parameter after removing true class evidence. This represents element-wise multiplication; ultimately, the total loss of the fault diagnosis model is,

[0096]

[0097] In the formula, This represents the total loss function during model training. These represent the weighting coefficients for different loss terms. This represents all learnable parameters in the fault diagnosis model. Indicates the L2 regularization term;

[0098] The model parameters are updated by minimizing the total loss function L, so that the model simultaneously satisfies the sample label constraint, physical consistency constraint, physical feature monotonicity constraint, and evidence uncertainty constraint.

[0099] The specific process of step 6 above is as follows:

[0100] Step 61: The acoustic response signal of the bridge expansion joint to be tested is processed based on steps 1 to 3 to obtain time-frequency response samples and physical characteristics;

[0101] Step 62: Input the time-frequency response samples and physical features into the trained fault diagnosis model to obtain the state category probability. and diagnostic uncertainty ,

[0102]

[0103]

[0104] In the formula, Indicates the first The predicted state category of each sample to be detected. Indicates the first The diagnostic reliability of each sample to be tested. Indicates the first Diagnostic uncertainty of a sample to be tested; when diagnostic uncertainty Greater than the preset threshold When the diagnosis is uncertain, the corresponding sample is marked as requiring manual review; Not greater than the preset threshold At that time, the expansion joint status category, status probability, and diagnostic confidence level are directly output, where a preset threshold is used. The median of diagnostic uncertainty is determined based on the model validation set samples.

[0105] Compared with the prior art, the present invention has the following advantages after adopting the above solution:

[0106] (1) This invention transforms the transient response mechanism of vehicle impact expansion joints into calculable physical characteristics and physical consistency loss, so that the diagnostic model no longer relies solely on sample labels and statistical correlation, thereby improving the physical interpretability of the diagnostic results.

[0107] (2) The present invention characterizes the state of the expansion joint by using the peak impact, frequency band energy ratio, dominant frequency position, energy decay rate and impact duration, thereby reducing the impact of signal differences caused by different vehicle passing events on the diagnostic results.

[0108] (3) The present invention uses an evidence classification network to output state probability and diagnostic uncertainty, enabling the model to identify low-confidence or ambiguous samples, providing a basis for manual review and risk control in bridge maintenance.

[0109] (4) This invention is applicable to non-contact acoustic detection scenarios. The sensor is easy to deploy and can serve the rapid inspection of expansion joints of small and medium-sized bridges, the comparison of the state before and after maintenance, and the daily maintenance assessment. Attached Figure Description

[0110] Figure 1 This is a flowchart of the present invention;

[0111] Figure 2 This is a schematic diagram of physical feature extraction and physical knowledge constraint construction in this invention;

[0112] Figure 3 This is a schematic diagram of the fault diagnosis model structure in this invention;

[0113] Figure 4 This is a schematic diagram of the training framework for the fault diagnosis model guided by physical knowledge in this invention;

[0114] Figure 5 This is a schematic diagram of the diagnostic result output and manual review process in this invention. Detailed Implementation

[0115] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0116] like Figure 1 As shown, this invention proposes a bridge expansion joint damage probability diagnosis method based on physical knowledge-guided acoustic signature features. The main technical process of the implementation scheme is as follows:

[0117] S1. Install acoustic sensors near the bridge expansion joint to collect the acoustic response signals generated when vehicles pass through the expansion joint, and determine the vehicle passage event based on the short-time energy and impact peak of the acoustic response signal.

[0118] Specifically, it includes the following steps:

[0119] S11, the acoustic sensors are deployed at the transverse ends of the expansion joint, the inner side of the guardrail, or the edge of the bridge deck, with a sampling frequency of not less than 16 kHz; the acquired raw acoustic response signal is The sampling frequency is .

[0120] (1)

[0121] In the formula, Indicates the first The sound pressure amplitude or normalized amplitude at each sampling point Indicates the total number of sampling points. Represents a continuous-time variable. Indicates the discrete sampling sequence number;

[0122] S12, Identify vehicle passage events from the continuous acoustic response signals acquired in step S11, and calculate the short-time energy function. and short-time peak impact function :

[0123] (2)

[0124] (3)

[0125] In the formula, Indicates the length of the short-time analysis window. This indicates the sampling point number within the analysis window and the corresponding short-time energy threshold. and short-term impact peak threshold Determined based on the statistical characteristics of the background noise segment when no vehicles are passing, specifically expressed as follows:

[0126] (4)

[0127] (5)

[0128] In the formula, and Let represent the mean and standard deviation of the short-time energy of the background noise segment, respectively. and These represent the mean and standard deviation of the short-time impact peak value in the background noise segment, respectively. and These represent adjustment coefficients for the short-time energy threshold and the short-time impact peak threshold, respectively, used to control the sensitivity of vehicle passage event recognition; when and At that time, the judgment of the first A vehicle passage event exists near each sampling point. Sampling points that consecutively meet the above conditions are merged into a single vehicle passage event, and the corresponding acoustic response segment is extracted based on the start and end positions of the event, denoted as the [number]th [event]. A sample of events involving a single vehicle.

[0129] S13, p. The peak impact moment of the event involving individual vehicles. Determined according to the following formula:

[0130] (6)

[0131] In the formula, Indicates the first The peak impact moment corresponding to the event of a vehicle passing through. Indicates the first The time interval in which the vehicle passed through the event.

[0132] S2, bandpass filtering, vehicle passage event segmentation, amplitude normalization, and short-time Fourier transform are performed on the acoustic response signal to obtain the time-frequency response sample corresponding to the vehicle passing through the expansion joint.

[0133] Specifically, it includes the following steps:

[0134] S21, at the peak moment Centered on the original acoustic response signal, a fixed-length segment is extracted to obtain the first... Acoustic segment of a vehicle passing through an expansion joint :

[0135] (7)

[0136] In the formula, This indicates the length of time intercepted before the peak impact moment. This indicates the length of time taken after the peak impact moment;

[0137] S22 performs bandpass filtering on the acoustic segment to preserve the main response frequency band generated by the vehicle impact expansion joint and suppress low-frequency environmental disturbances and high-frequency random noise:

[0138] (8)

[0139] In the formula, Indicates the first bandpass filter. An acoustic segment, This represents the bandpass filter operator. This indicates the cutoff frequency under bandpass filtering. Indicates the upper cutoff frequency of the bandpass filter;

[0140] S23, Amplitude normalization of the filtered acoustic segment:

[0141] (9)

[0142] In the formula, Represents the normalized i-th An acoustic segment, This prevents extremely small positive numbers with a denominator of zero; the normalized acoustic segment is converted into a time-frequency response sample using a short-time Fourier transform.

[0143] (10)

[0144] In the formula, Indicates the first The time-frequency response of each sample. This indicates the time corresponding to the center position of the short-time Fourier transform window. Represents frequency variables. Indicates the sampling point number within the window. This indicates the short-time Fourier transform window length. Represents the window function. The imaginary unit is represented; the time-frequency energy map in the form of power spectral density is further obtained from the time-frequency response samples. :

[0145] (11)

[0146] In the formula, Indicates the first Each sample in time and frequency Time-frequency energy at a given location.

[0147] S3, see S3. Figure 2Based on the transient response mechanism generated when a vehicle tire impacts an expansion joint, the impact peak value, frequency band energy ratio, dominant frequency position, energy decay rate, and impact duration are extracted from the time-frequency response samples as physical characteristics characterizing the service status of the expansion joint.

[0148] Specifically, when a vehicle tire passes through an expansion joint, a transient impact occurs between the tire, the gap, the beam end, and the expansion joint components. Healthy expansion joints typically exhibit a short impact duration, relatively stable energy distribution, and rapid response decay. Damaged expansion joints, due to loose steel sections, anchorage degradation, or localized voids, typically exhibit increased impact peak value, increased energy in abnormal frequency bands, a shift in dominant frequency, slower decay, and a prolonged impact duration. Based on the above mechanism, the following physical characteristics are extracted:

[0149] (12)

[0150] (13)

[0151] (14)

[0152] (15)

[0153] (16)

[0154] (17)

[0155] In the formula, Indicates the first The peak impact value of each sample Indicates the first The frequency band energy ratio of each sample within the target frequency band and Indicates the lower and upper limits of the target frequency band. Indicates the first The dominant frequency or spectral centroid position of each sample Indicates the first The envelope signal of each sample, Represents the Hilbert transform. Indicates the first Energy decay rate of each sample This indicates the duration of the decay rate fitting time. Indicates the first The impact duration of each sample and These represent the start and end times of the impact duration, respectively. Indicates to Over time The slope is obtained by linearly fitting the changing relationship. The above physical characteristics are then combined to form a physical feature vector. :

[0156] (18)

[0157] In the formula, Indicates the first The physical feature vector of each sample This represents the vector transpose symbol.

[0158] S4, see S4 Figure 2 Based on the influence of expansion joint damage state on vehicle impact response, a physical knowledge constraint consisting of equivalent stiffness degradation constraint, damping attenuation constraint, and impact energy redistribution constraint is established.

[0159] Specifically, it includes the following steps:

[0160] S41, by equating the local response of the vehicle tire impacting the expansion joint to a single-degree-of-freedom transient forced vibration process, we obtain:

[0161] (19)

[0162] In the formula, This indicates the local equivalent mass of the expansion joint. This indicates the local equivalent damping of the expansion joint. This indicates the local equivalent stiffness of the expansion joint. This represents the local equivalent displacement response. Indicates the local equivalent velocity response. This represents the local equivalent acceleration response. This refers to the external excitation generated when a vehicle's tires pass through an expansion joint. Additional impact excitation caused by damage to expansion joints;

[0163] S42, let the degree of damage to the expansion joint be... , Indicates health status. This indicates a severely damaged state. As the damage increases, the local connection stiffness of the expansion joint decreases, its energy dissipation capacity changes, and it triggers a redistribution of impact energy to abnormal frequency bands. The corresponding degradation relationship between equivalent stiffness and equivalent damping is expressed as:

[0164] (20)

[0165] (twenty one)

[0166] In the formula, Indicates the degree of damage to the expansion joint. Represents the equivalent stiffness under healthy conditions. This represents the equivalent damping under healthy conditions. This represents the stiffness degradation coefficient. Indicates the damping degradation coefficient;

[0167] S43, according to the equivalent dynamic relationship, when the equivalent stiffness of the expansion joint decreases, its characteristic frequency shifts. The characteristic frequency can be expressed as:

[0168] (twenty two)

[0169] In the formula, Indicates the degree of damage. The equivalent characteristic frequency at time, It represents pi (π).

[0170] S44, Based on the law of change in impact response, construct the physical damage index. Used to characterize the degree of damage to expansion joints inferred from physical features:

[0171] (twenty three)

[0172] (twenty four)

[0173] In the formula, Indicates the first The physical damage index corresponding to each sample , , , , Represents the standardized physical characteristics. This indicates the weighting of the physical damage index calculation. This represents the Sigmoid mapping function. This represents the input variables of the Sigmoid function; among the above physical damage indices, as the peak impact, frequency band energy ratio, and impact duration increase, Increase; when the main frequency position and energy decay rate decrease. Increase. This relationship constitutes the core constraint guided by physical knowledge in this invention.

[0174] S5, input the time-frequency response sample into the deep feature extraction network, input the physical feature into the physical feature encoding network, and input the fused deep feature and physical feature into the evidence classification network to construct a fault diagnosis model. In the model training process, a physical consistency loss function is introduced to ensure that the prediction results of the fault diagnosis model are consistent with the physical knowledge constraints.

[0175] Specifically, such as Figure 3As shown, the fault diagnosis model includes a deep feature extraction network, a physical feature encoding network, a feature fusion layer, and an evidence classification network. In this embodiment, the time-frequency response samples are normalized and resized to obtain a three-channel time-frequency response map with a size of 224×224×3, which is then input into the deep feature extraction network.

[0176] The deep feature extraction network employs a convolutional neural network structure, comprising a first convolutional layer, a first batch normalization layer, a first activation layer, a first pooling layer, a second convolutional layer, a second batch normalization layer, a second activation layer, a second pooling layer, a global average pooling layer, and a first fully connected layer, connected sequentially. The first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 and the same padding method. The first activation layer uses the ReLU activation function, and the first pooling layer uses max pooling with a size of 2×2 and a stride of 2. The second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1 and the same padding method. The second activation layer uses the ReLU activation function, and the second pooling layer uses max pooling with a size of 2×2 and a stride of 2. After processing by the global average pooling layer and the first fully connected layer containing 128 neurons, 128-dimensional deep features are obtained. .

[0177] The physical feature encoding network uses a five-dimensional physical feature vector. The five-dimensional physical features, as input, include the impact peak value, frequency band energy ratio, dominant frequency position, energy decay rate, and impact duration. The feature encoding network comprises a second fully connected layer, a third activation layer, a third fully connected layer, and a fourth activation layer connected sequentially. The second fully connected layer contains 32 neurons, the third fully connected layer contains 64 neurons, and both the third and fourth activation layers employ the ReLU activation function, outputting 64-dimensional physical encoding features. .

[0178] The feature fusion layer integrates 128-dimensional deep features. With 64-dimensional physical coding features By splicing the features, a 192-dimensional fused feature is obtained. The evidence classification network includes a fourth fully connected layer and a Softplus activation function. The number of output nodes of the fourth fully connected layer corresponds to the number of expansion joint state categories. Similarly, after processing with the Softplus activation function, non-negative evidence vectors corresponding to each state category are obtained. .

[0179] S51, Deep Feature Extraction Network with Time-Frequency Response Samples Input, output deep features Physical feature encoding networks use physical feature vectors As input, output physical coding features :

[0180] (25)

[0181] (26)

[0182] (27)

[0183] In the formula, The parameter is Deep feature extraction network, The parameter is Physical feature encoding network, Indicates the first Depth features of each sample Indicates the first The physical coding features of each sample This represents the features of the fused samples. Indicates the feature concatenation operator;

[0184] S52, Evidence Classification Network with Feature Fusion As input, output the non-negative evidence value corresponding to each expansion joint state category:

[0185] (28)

[0186] (29)

[0187] (30)

[0188] (31)

[0189] (32)

[0190] In the formula, Indicates the first Non-negative evidence vectors for each state category of each sample This represents the activation function that guarantees a non-negative output. and Let represent the weight matrix and bias vector of the evidence classification network, respectively. Indicates the first Dirichlet distribution parameters of a sample Indicates the first The first sample Dirichlet parameters corresponding to class states Represents the Dirichlet distribution parameters and, Indicates the number of state categories. Indicates the first The sample belongs to the first Predicted probability of class state Indicates the first Diagnostic uncertainty for the nth sample; let the nth sample be... The true label for each sample is , Indicates sample Does it belong to the first Class state. The evidence classification loss uses Dirichlet evidence loss:

[0191] (33)

[0192] In the formula, Indicates loss in evidence classification. Indicates the number of training samples. Indicates the first The first sample The actual label indicator variable for class state, Represents the Digamma function;

[0193] S53, to ensure that the damage level output by the model remains consistent with the physical feature inference results, defines the model-predicted damage index. :

[0194] (34)

[0195] In the formula, Indicates the first The damage index is obtained by predicting the probability of each sample using the model. This represents the damage level code corresponding to the c-th state, with 0 for a healthy state and 1 for a damaged state; thus, a physical consistency loss function is constructed:

[0196] (35)

[0197] In the formula, This represents the physical consistency loss function; furthermore, to ensure that the model's response direction to key physical features conforms to physical laws, a monotonic constraint loss is constructed:

[0198] (36)

[0199] In the formula, This represents the monotonic constraint loss. Represents the linear rectified function. This represents the partial derivative of the model's predicted damage index with respect to the physical characteristics.

[0200] S54. To avoid the model generating too much evidence on erroneous samples, Dirichlet regularization loss is introduced:

[0201] (37)

[0202] (38)

[0203] In the formula, This represents the Dirichlet regularization loss. Indicates the Kullback-Leibler divergence. Represents the Dirichlet distribution. This represents the Dirichlet parameter after removing true class evidence. This represents element-wise multiplication; ultimately, the total loss of the fault diagnosis model is:

[0204] (39)

[0205] In the formula, This represents the total loss function during model training. These represent the weighting coefficients for different loss terms. This represents all learnable parameters in the fault diagnosis model. Represents the L2 regularization term; such as Figure 4 As shown, during model training, the total loss function is minimized. Update the model parameters so that the model simultaneously satisfies the sample label constraint, physical consistency constraint, physical feature monotonicity constraint, and evidence uncertainty constraint.

[0206] In the specific training process, the time-frequency response samples, physical feature vectors, and state labels corresponding to the training samples are input into the fault diagnosis model in a preset batch. First, deep features and physical coding features are obtained through a deep feature extraction network and a physical feature encoding network, respectively. Then, through a feature fusion layer and an evidence classification network, the non-negative evidence value, state category probability, and diagnostic uncertainty corresponding to each state category are obtained.

[0207] Secondly, the evidence classification loss is calculated based on the sample state labels; the physical consistency loss is calculated based on the difference between the model-predicted damage index and the physical damage index; the monotonic constraint loss is calculated based on the direction of change of the model-predicted damage index relative to each physical feature; and the Dirichlet regularization loss and L2 regularization term are calculated. The total loss of the fault diagnosis model is obtained by weighted summation of all loss terms.

[0208] Finally, the gradient of the total loss with respect to the model's learnable parameters is calculated through error backpropagation, and the learnable parameters in the deep feature extraction network, physical feature encoding network, and evidence classification network are updated using optimization algorithms. This process is repeated until a preset number of training epochs is reached or the model's loss on the validation set meets the convergence condition, resulting in a successfully trained fault diagnosis model.

[0209] In this embodiment, the Adam optimization algorithm is used, and the initial learning rate is set to 1×10. -4 The batch size is set to 32, and the maximum number of training rounds is set to 100. After each training round, the loss of the model on the validation set is calculated. Training is stopped when the validation loss does not decrease for 10 consecutive rounds, and the model parameters corresponding to the minimum validation loss are saved.

[0210] S6, see S6. Figure 4 The trained fault diagnosis model is used to identify the acoustic response signal of the bridge expansion joint to be tested, and output the expansion joint state category, state probability and diagnostic uncertainty.

[0211] Specifically, for the acoustic response signal of the bridge expansion joint to be detected, the event identification, preprocessing, time-frequency response sample construction, and physical feature extraction are completed according to steps S1 to S3. Then, the time-frequency response samples and physical features are input into the trained fault diagnosis model to obtain the state category probability. and diagnostic uncertainty :

[0212] (40)

[0213] (41)

[0214] In the formula, Indicates the first The predicted state category of each sample to be detected. Indicates the first The diagnostic reliability of each sample to be tested. Indicates the first Diagnostic uncertainty of a single sample to be tested;

[0215] like Figure 5 As shown, when diagnostic uncertainty Greater than the preset threshold When the diagnosis is uncertain, the corresponding sample is marked as requiring manual review; Not greater than the preset threshold At that time, the expansion joint status category, status probability, and diagnostic confidence level are directly output, where a preset threshold is used. The median of diagnostic uncertainty is determined based on the model validation set samples.

[0216] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0217] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for diagnosing bridge expansion joint damage probability based on physical knowledge-guided acoustic signature features, characterized in that: Includes the following steps, Step 1: Install acoustic sensors near the bridge expansion joints to collect the acoustic response signals generated when vehicles pass through the expansion joints, and determine the vehicle passage event based on the short-time energy and impact peak value of the acoustic response signals. Step 2: Perform bandpass filtering, vehicle passage event segmentation, amplitude normalization, and short-time Fourier transform on the acoustic response signal to obtain the time-frequency response sample corresponding to the vehicle passing through the expansion joint. Step 3: Based on the transient response mechanism generated when a vehicle tire impacts an expansion joint, extract the impact peak value, frequency band energy ratio, dominant frequency position, energy decay rate, and impact duration from the time-frequency response sample as physical characteristics characterizing the service status of the expansion joint. Step 4: Based on the influence of expansion joint damage state on vehicle impact response, establish physical knowledge constraints consisting of equivalent stiffness degradation constraints, damping attenuation constraints, and impact energy redistribution constraints. Step 5: The fault diagnosis model includes a deep feature extraction network, a physical feature encoding network, a feature fusion layer, and an evidence classification network. The time-frequency response samples are input into the deep feature extraction network, the physical features are input into the physical feature encoding network, and the deep features and physical features are fused and then input into the evidence classification network. A physical consistency loss function is introduced during the model training process to ensure that the prediction results of the fault diagnosis model are consistent with the physical knowledge constraints. Step 6: Use the trained fault diagnosis model to identify the acoustic response signal of the bridge expansion joint to be tested, and output the expansion joint state category, state probability and diagnostic uncertainty.

2. The method as described in claim 1, characterized in that: The specific process of step 1 is as follows: Step 11: Install acoustic sensors at the transverse ends of the expansion joint, the inside of the guardrail, or the edge of the bridge deck. ， In the formula, Represents the acoustic response signal. Indicates the first The sound pressure amplitude or normalized amplitude at each sampling point Indicates the sampling frequency of the acoustic response signal. Indicates the total number of sampling points. Represents a continuous-time variable. Indicates the discrete sampling sequence number; Step 12: Identify vehicle passage events from the collected acoustic response signals and calculate the short-time energy function. and short-time peak impact function , ， ， In the formula, Indicates the short-time analysis window length. This indicates the sampling point number within the analysis window and the corresponding short-time energy threshold. and short-term impact peak threshold It is determined based on the statistical characteristics of the background noise segment when no vehicles are passing, specifically expressed as follows: ， ， In the formula, and Let represent the mean and standard deviation of the short-time energy of the background noise segment, respectively. and These represent the mean and standard deviation of the short-time impact peak value in the background noise segment, respectively. and These represent the adjustment coefficients for the short-time energy threshold and the short-time impact peak threshold, respectively, used to control the sensitivity of vehicle passage event recognition; when and At that time, the judgment of the first A vehicle passage event exists near each sampling point; sampling points that consecutively meet the above conditions are merged into the same vehicle passage event, and the corresponding acoustic response segment is extracted based on the start and end positions of the event, denoted as the i-th segment. One vehicle passes through the event sample; Step 13, the The peak impact moment of the event involving individual vehicles. Determined according to the following formula, ， In the formula, Indicates the first The peak impact moment corresponding to the event of a vehicle passing through. Indicates the first The time interval in which the vehicle passed through the event.

3. The method as described in claim 1, characterized in that: The specific process of step 2 is as follows: Step 21, based on the peak impact time Centered on the acoustic response signal, a fixed-length segment is extracted to obtain the first... Acoustic segment of a vehicle passing through an expansion joint , ， In the formula, This indicates the length of time intercepted before the peak impact moment. This indicates the length of time taken after the peak impact moment; Step 22: Perform bandpass filtering on the acoustic segment. ， In the formula, Indicates the first bandpass filter. An acoustic segment, This represents the bandpass filter operator. This indicates the cutoff frequency under bandpass filtering. Indicates the upper cutoff frequency of the bandpass filter; Step 23: Normalize the amplitude of the filtered acoustic segment. ， In the formula, Represents the normalized i-th An acoustic segment, This indicates the use of extremely small positive numbers to prevent the denominator from being zero; a short-time Fourier transform is used to convert the normalized acoustic segment into a time-frequency response sample. ， In the formula, Indicates the first The time-frequency response of each sample. This indicates the time corresponding to the center position of the short-time Fourier transform window. Represents frequency variables. Indicates the sampling point number within the window. This indicates the short-time Fourier transform window length. Represents the window function. It represents the imaginary unit.

4. The method as described in claim 1, characterized in that: The specific process of step 3 is as follows: when the vehicle tires pass through the expansion joint, the following physical characteristics are extracted. ， ， ， ， ， ， In the formula, Indicates the first The peak impact value of each sample Represents the normalized i-th An acoustic segment; Indicates the first The frequency band energy ratio of each sample within the target frequency band and Indicates the lower and upper limits of the target frequency band. Indicates the first Each sample in time and frequency Time-frequency energy at that location This indicates the time corresponding to the center position of the short-time Fourier transform window. Represents frequency variables; Indicates the sampling frequency of the acoustic response signal; Indicates the first The dominant frequency or spectral centroid position of each sample; Indicates the first The envelope signal of each sample, Represents the Hilbert transform; Indicates the first Energy decay rate of each sample Indicates to Over time The slope is obtained by linear fitting of the relationship between the changes. Indicates the first The peak impact moment corresponding to the event of a vehicle passing through. This indicates the duration of the decay rate fitting time. Indicates the first The impact duration of each sample and These represent the start and end times of the impact duration, respectively; The above physical characteristics are combined into a physical feature vector. , ， In the formula, Indicates the first The physical feature vector of each sample This represents the vector transpose symbol.

5. The method as described in claim 4, characterized in that: The first Each sample in time and frequency Time-frequency energy at the location According to the following formula, ， In the formula, Indicates the first The time-frequency response of each sample.

6. The method as described in claim 4, characterized in that: The specific process of step 4 is as follows: Step 41: The local response of the vehicle tire impacting the expansion joint is equivalent to a single-degree-of-freedom transient forced vibration process, resulting in... ， In the formula, This indicates the local equivalent mass of the expansion joint. This indicates the local equivalent damping of the expansion joint. This indicates the local equivalent stiffness of the expansion joint. This represents the local equivalent displacement response. Indicates the local equivalent velocity response. This represents the local equivalent acceleration response. This refers to the external excitation generated when a vehicle's tires pass through an expansion joint. Additional impact excitation caused by damage to expansion joints; Step 42, let the degree of damage to the expansion joint be... , Indicates health status. This indicates a severely damaged state; the corresponding equivalent stiffness and equivalent damping degradation relationship is expressed as follows: ， ， In the formula, Indicates the degree of damage to the expansion joint. Represents the equivalent stiffness under healthy conditions. This represents the equivalent damping under healthy conditions. Indicates the stiffness degradation coefficient. Indicates the damping degradation coefficient; Step 43, the characteristic frequency of the expansion joint is expressed as, ， In the formula, Indicates the degree of damage. The equivalent characteristic frequency at time, Represents pi; Step 44, Construct the physical damage index It is used to characterize the degree of damage to expansion joints inferred from physical characteristics. ， ， In the formula, Indicates the first The physical damage index corresponding to each sample , , , , Represents the standardized physical characteristics. This indicates the weighting of the physical damage index calculation. This represents the Sigmoid mapping function. This represents the input variables for the Sigmoid function.

7. The method as described in claim 6, characterized in that: The specific process of step 5 is as follows: Step 51, the deep feature extraction network uses time-frequency response samples Input, output deep features ; Physical feature encoding networks use physical feature vectors As input, output physical coding features , ， ， ， In the formula, The parameter is Deep feature extraction network, The parameter is Physical feature encoding network, Indicates the first Depth features of each sample Indicates the first The physical coding features of each sample This represents the features of the fused samples. Indicates the feature concatenation operator; This indicates the time corresponding to the center position of the short-time Fourier transform window. Represents frequency variables; Step 52, the evidence classification network uses fused features As input, output the non-negative evidence value corresponding to each expansion joint state category. ， ， ， ， ， In the formula, Indicates the first Non-negative evidence vectors for each state category of each sample This represents the activation function that guarantees a non-negative output. and Let represent the weight matrix and bias vector of the evidence classification network, respectively. Indicates the first Dirichlet distribution parameters of a sample Indicates the first The first sample Dirichlet parameters corresponding to class states Represents the parameters of the Dirichlet distribution, Indicates the number of state categories. Indicates the first The sample belongs to the first Predicted probability of class state Indicates the first Diagnostic uncertainty for the nth sample; let the nth sample be... The true label for each sample is , Indicates sample Does it belong to the first Class state; Dirichlet evidence loss is used for evidence classification loss. ， In the formula, Indicates loss in evidence classification. Indicates the number of training samples. Indicates the first The first sample The actual label indicator variable for class state, Represents the Digamma function; Step 53: To ensure that the damage level output by the constraint model is consistent with the physical feature inference results, define the model-predicted damage index. , ， In the formula, Indicates the first The damage index is obtained by predicting the probability of each sample using the model. This represents the damage level code corresponding to the c-th state, with 0 for a healthy state and 1 for a damaged state; thus, a physical consistency loss function is constructed: ， In the formula, Represent the physical consistency loss function; and construct a monotonic constraint loss as follows: ， In the formula, This represents the monotonic constraint loss. Represents the linear rectified function. This represents the partial derivative of the model's predicted damage index with respect to the physical characteristics. Step 54, define the Dirichlet regularization loss. ， ， In the formula, This represents the Dirichlet regularization loss. Indicates the Kullback-Leibler divergence. Represents the Dirichlet distribution. This represents the Dirichlet parameter after removing true class evidence. This represents element-wise multiplication; ultimately, the total loss of the fault diagnosis model is, ， In the formula, This represents the total loss function during model training. These represent the weighting coefficients for different loss terms. This represents all learnable parameters in the fault diagnosis model. Indicates the L2 regularization term; The model parameters are updated by minimizing the total loss function L, so that the model simultaneously satisfies the sample label constraint, physical consistency constraint, physical feature monotonicity constraint, and evidence uncertainty constraint.

8. The method as described in claim 7, characterized in that: The specific process of step 6 is as follows: Step 61: The acoustic response signal of the bridge expansion joint to be tested is processed based on steps 1 to 3 to obtain time-frequency response samples and physical characteristics; Step 62: Input the time-frequency response samples and physical features into the trained fault diagnosis model to obtain the state category probability. and diagnostic uncertainty , ， ， In the formula, Indicates the first The predicted state category of each sample to be detected. Indicates the first The diagnostic reliability of each sample to be tested. Indicates the first Diagnostic uncertainty of a sample to be tested; when diagnostic uncertainty Greater than the preset threshold When the diagnosis is uncertain, the corresponding sample is marked as requiring manual review; Not greater than the preset threshold At that time, the expansion joint status category, status probability, and diagnostic confidence level are directly output, where a preset threshold is used. The median of diagnostic uncertainty is determined based on the model validation set samples.

Images