Bridge ship collision event multi-level threshold adaptive alarm method and system based on multi-dimensional vibration characteristics

Abstract

The application discloses a bridge-ship collision event multi-level threshold adaptive alarm method and system based on multi-dimensional vibration characteristics, which comprises the following steps: collecting original vibration signals under different vibration events, dividing a training set and performing pretreatment, extracting multi-dimensional vibration characteristics of each vibration signal in the training set, and constructing a ship collision event identification model; collecting a to-be-identified vibration signal and performing pretreatment, and extracting a to-be-identified vibration feature vector; inputting the to-be-identified vibration feature vector into the ship collision event identification model, and outputting a classification probability of the to-be-identified vibration signal belonging to a ship collision event by the model; and establishing a multi-level threshold adaptive alarm mechanism, and issuing a ship collision alarm signal when a first-level alarm, a second-level alarm and a third-level alarm are all triggered within a preset time window. The application extracts high-discrimination multi-dimensional feature indexes from time domain, frequency domain and time-frequency domain, uses a machine learning model for classification, and combines a multi-level threshold adaptive alarm mechanism to realize accurate and reliable alarm for a collision event.

Description

Technical Field

[0001] This application relates to the field of bridge safety monitoring technology, specifically to a multi-level threshold adaptive alarm method and system for bridge-ship collision events based on multi-dimensional vibration characteristics. Background Technology

[0002] With the development of inland waterway shipping and the increasing trend of larger ships, the operational safety of bridges, especially those crossing waterways, is facing increasingly severe challenges. The risk of accidents involving out-of-control or veer-off vessels colliding with bridges is increasing significantly. Such collisions can directly lead to serious damage to the main structure of the bridge or even its collapse, causing major safety accidents and huge economic losses.

[0003] In existing technologies, the main method for monitoring bridge collision risks is a combination of initial vibration threshold assessment and video surveillance. Specifically, accelerometers are typically installed at key parts of the bridge. When the detected vibration amplitude exceeds a preset single threshold, an alarm is triggered, and on-site cameras capture video footage for manual review or identification using image processing technology.

[0004] However, this existing technological approach still has significant limitations in practical applications: On the one hand, the initial judgment logic based on vibration amplitude is too crude, and the false alarm rate remains high. Bridges are subjected to various environmental excitations during daily operation, such as heavy vehicle traffic, strong wind loads, and water flow impacts. The structural vibrations caused by these excitations may overlap with the amplitude of minor or moderate ship collisions in the time domain. Relying solely on simple amplitude thresholds makes it difficult to effectively distinguish between vehicle-induced vibrations and ship collision impacts from a physical mechanism perspective, leading to frequent false alarms and increasing the verification burden on maintenance personnel.

[0005] On the other hand, video capture systems, as a means of verification, have significant blind spots and environmental dependence. Video monitoring is an optical, non-contact sensing method, which is highly susceptible to environmental influences. In adverse weather conditions such as insufficient light at night, heavy fog, or torrential rain, the image quality of cameras will drop significantly or even fail completely, making it impossible to effectively confirm whether a ship collision has occurred through video.

[0006] In summary, although accelerometers have the advantages of low cost and easy deployment, how to combine them with advanced signal processing and pattern recognition technologies to reduce the over-reliance on video verification and directly extract deeper and more distinctive features from vibration signals to achieve intelligent identification and accurate alarm for ship collision events is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] To address the problems existing in the prior art, this application proposes a multi-level threshold adaptive alarm method and system for bridge-ship collision events based on multi-dimensional vibration characteristics. The method extracts differentiated features of vibration signals from the time domain, frequency domain, and time-frequency domain, constructs a highly discriminative multi-dimensional feature index set, and uses a machine learning model for classification. Finally, it achieves accurate and reliable collision event alarm through a multi-level threshold adaptive mechanism.

[0008] Firstly, this application provides a multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics, including: Raw vibration signals under different vibration events are collected, training sets are divided and preprocessed, multi-dimensional vibration features of each vibration signal in the training set are extracted, and a ship collision event recognition model is constructed based on the multi-dimensional vibration features and their corresponding category labels. The vibration signal to be identified is acquired and preprocessed to extract the vibration feature vector to be identified, which includes multiple time-domain feature indicators, at least one frequency-domain feature indicator, and at least one time-frequency-domain feature indicator. The vibration feature vector to be identified is input into the ship collision event identification model, and the classification probability of the vibration signal to be identified belonging to the ship collision event is output. A basic threshold based on a single time-domain feature index, a combined discriminant based on multiple vibration feature indices, and a decision threshold based on a ship collision event recognition model are set respectively. When the single time-domain feature index collected in real time is greater than the basic threshold, a level one alarm is triggered. When the multiple vibration feature indices collected in real time satisfy the combined discriminant, a level two alarm is triggered. When the classification probability is greater than the decision threshold, a level three alarm is triggered. When all three alarms (level one, level two, and level three) are triggered within a preset time window, a ship collision alarm signal is issued.

[0009] In one possible implementation of the first aspect, the step of collecting raw vibration signals under different vibration events, dividing the training set and preprocessing it, extracting multi-dimensional vibration features of each vibration signal in the training set, and constructing a ship collision event recognition model based on the multi-dimensional vibration features and their corresponding category labels includes: Raw vibration signals of the bridge under different vibration events, including vehicle traffic, environmental excitation, and ship collision, are collected. A vibration signal sample database with category labels is constructed based on the raw vibration signals. The vibration signal sample database is divided into a training set, a validation set, and a test set according to a preset ratio. The original vibration signals in the training set are preprocessed to obtain the preprocessed vibration signals. Extract the time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index of the preprocessed vibration signal to construct a multi-dimensional feature index set; Based on the time-domain feature index, frequency-domain feature index, time-frequency-domain feature index and their corresponding category labels of the preprocessed vibration signal, the preset classification model is trained and optimized, and the training and optimization classification model is used as the ship collision event recognition model.

[0010] In one possible implementation of the first aspect, the step of training and optimizing a preset classification model based on the time-domain feature index, frequency-domain feature index, time-frequency-domain feature index, and their corresponding category labels of the preprocessed vibration signal, and using the trained and optimized classification model as a ship collision event recognition model, includes: The multi-dimensional feature index set is used as the input sample of the classification model, and the category label corresponding to the vibration signal is used as the expected output of the classification model. A training algorithm adapted to the classification model is used to iteratively train the model based on the input samples and the expected output to optimize the internal parameters or model structure of the classification model. During the iterative training process, the training process is controlled by a preset convergence condition or the maximum number of iterations, and the optimized classification model is used as the ship collision event recognition model.

[0011] In one possible implementation of the first aspect, the importance / contribution assessment of the time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index is performed on the trained and optimized ship collision event recognition model.

[0012] In one possible implementation of the first aspect, the preset classification model is a random forest model, a gradient boosting decision tree model, or a neural network model.

[0013] In one possible implementation of the first aspect, the time-domain characteristic indicators include peak value, rise time, and waveform factor.

[0014] In one possible implementation of the first aspect, the frequency domain feature includes a spectral centroid.

[0015] In one possible implementation of the first aspect, the time-frequency domain feature includes wavelet packet energy entropy.

[0016] In one possible implementation of the first aspect, the root mean square value of the acceleration of the vibration signal to be identified is calculated in real time; based on the root mean square value of the acceleration, the decision threshold is adaptively updated according to a preset dynamic mapping relationship or proportional coefficient.

[0017] Secondly, this application provides a system for implementing the aforementioned multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics, comprising: The model training module collects raw vibration signals under different vibration events, divides the training set and preprocesses it, extracts multi-dimensional vibration features of each vibration signal in the training set, and constructs a ship collision event recognition model based on the multi-dimensional vibration features and their corresponding category labels. The real-time acquisition module acquires the vibration signal to be identified and performs preprocessing to extract the vibration feature vector to be identified. The event recognition module inputs the vibration feature vector to be identified into the ship collision event recognition model and outputs the classification probability that the vibration signal to be identified belongs to the ship collision event. The multi-level threshold adaptive alarm module sets a basic threshold based on a single time-domain feature index, a combined discriminant based on multiple vibration feature indices, and a decision threshold based on a ship collision event recognition model; it establishes a multi-level threshold adaptive alarm mechanism; and when the first-level alarm, second-level alarm, and third-level alarm are all triggered within a preset time window, a ship collision alarm signal is issued.

[0018] Compared with existing methods, this application has the following advantages: By extracting complementary features from multiple perspectives, including time, frequency, and time-frequency, a multi-dimensional feature index set is constructed that can more comprehensively characterize the essential differences in dynamic response between normal vibration and ship impact vibration, laying the foundation for high-precision classification. The use of machine learning models to fuse multi-dimensional feature indices avoids the limitations of single threshold judgment, significantly improving the recognition accuracy of ship impact events and reducing false alarms and false negatives. The multi-level threshold adaptive mechanism combines the advantages of rapid response and comprehensive judgment, and can dynamically adjust the threshold according to environmental factors (such as background vibration noise) to improve alarm sensitivity and enhance the robustness and reliability of the system under different operating environments. Furthermore, this application has strong engineering adaptability and scalability. It can flexibly configure classification tasks according to the existing layout of bridge monitoring equipment and actual needs. It can achieve rapid early warning of binary classification of "non-ship collision events" and "ship collision events" under the basic sensor configuration. It can also carry out refined multi-class identification including vehicle driving, traffic accidents, environmental excitation and ship collision under the conditions of equipped multi-point sensors and video assistance, so as to meet the differentiated monitoring needs of different bridges. Verification through examples shows that when five specific key feature indicators—peak value, rise time, waveform factor, spectral centroid, and wavelet packet energy entropy—are used and combined with a random forest model, the clarity of the classification boundary when distinguishing between ship impacts and normal vibrations far exceeds that of a single feature or a combination of conventional features. When faced with complex and variable bridge environmental noise, this combination demonstrates excellent robustness and recognition accuracy, effectively solving the technical challenge of balancing high sensitivity and low false alarm rate in existing technologies. Attached Figure Description

[0019] Figure 1 A flowchart of a multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics, provided for the implementation of this application; Figure 2 A flowchart illustrating the construction of a vibration signal sample database and the optimization of model training is provided for embodiments of this application; Figure 3 A confusion matrix based on a verification set is provided for the implementation of this application; Figure 4 A bar chart ranking the importance of vibration characteristic indicators provided for embodiments of this application; Figure 5 A cumulative importance curve of a vibration characteristic index provided for an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the application will now be described in further detail with reference to the accompanying drawings.

[0021] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0022] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should be understood that the term "and / or" used in this article is merely a description of the same field for related objects, indicating that three relationships can exist.

[0024] like Figure 1 As shown, this application provides a multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics, including the following steps: S100: Collect raw vibration signals under different vibration events, divide the training set and preprocess it, extract multi-dimensional vibration features of each vibration signal in the training set, and construct a ship collision event recognition model based on the multi-dimensional vibration features and their corresponding category labels.

[0025] S200: Acquire the vibration signal to be identified, preprocess the vibration signal to be identified to eliminate environmental noise and trend term interference; based on the preprocessed vibration signal to be identified, extract the vibration feature vector to be identified, the vibration feature vector to be identified includes multiple time-domain feature indicators, at least one frequency-domain feature indicator and at least one time-frequency-domain feature indicator. The vibration signal to be identified is collected by a sensing device or a remote non-contact monitoring device installed on the bridge; for example, acceleration sensors are installed on key parts such as the main beam or pier of the bridge.

[0026] S300: Input the vibration feature vector to be identified into the ship collision event identification model, and the ship collision event identification model outputs the classification probability that the vibration signal to be identified belongs to the ship collision event.

[0027] S400: Establish a multi-level threshold adaptive alarm mechanism for ship collision events; A basic threshold based on a single time-domain feature index, a combined discriminant based on multiple vibration feature indices, and a decision threshold based on a ship collision event recognition model are set respectively. When the single time-domain feature index collected in real time is greater than the basic threshold, a level one alarm is triggered. When the multiple vibration feature indices collected in real time satisfy the combined discriminant, a level two alarm is triggered. When the classification probability is greater than the decision threshold, a level three alarm is triggered. When the signals collected in real time are within the same preset time window, and the level one, level two, and level three alarms are all triggered, a ship collision alarm signal is issued.

[0028] like Figure 2 As shown, in one possible implementation, step S100 includes the following: S110: Collect raw vibration signals of the bridge under different vibration events, including but not limited to vehicle traffic, environmental excitation (wind, waves), and ship collision; and construct a vibration signal sample database with category labels based on the raw vibration signals. For example, a fixed time window of 1 to 5 minutes is used. The original vibration signals of the same time window are retrospectively collected or obtained through simulation experiments according to the preset data sampling frequency and then cut off. The time length and number of data points of each segment of the original vibration signal are the same. For each segment of the original vibration signal, its vibration event category is marked. In one embodiment of this application, the category labels of vibration events include "normal vibration" and "ship impact". In this case, the classification model performs a binary classification task. For the binary classification task, considering that the main impact object of the ship is the bridge pier, the acceleration sensor is preferably deployed at the main beam and bridge pier to comprehensively capture vehicle load and traffic conditions while directly obtaining the most representative impact response signal. In another embodiment of this application, in order to more precisely grasp the health status of the bridge, the category labels may include multiple subcategories of events such as "normal vehicle passage", "traffic accident", "environmental stimulation (wind, waves)" and "ship collision". In this case, the classification model performs a multi-classification task. The multi-classification task places higher demands on the accuracy of the samples, aiming to deeply mine and screen out the differentiated features of different types of vibration signals. In particular, for situations such as traffic accidents that occur on the bridge surface but where the sensors are not directly installed on the bridge surface, it is necessary to combine video acquisition and other means for multimodal assistance to screen high-confidence samples and finely label the category labels, thereby solving the problem of sample label establishment in non-contact monitoring scenarios. S120: The vibration signal sample database is divided into a training set, a validation set, and a test set according to a preset ratio. The original vibration signals in the training set are preprocessed to obtain preprocessed vibration signals. The training set is used for iterative training of the classification model. The validation set is used to monitor the model performance during training and optimize the model's hyperparameters. The test set is used to evaluate the alarm accuracy and generalization ability of the classification model after it has been finalized. S130: Extract the time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index of the preprocessed vibration signal to construct a multi-dimensional feature index set; the parameters used in the preprocessing (including but not limited to signal conditioning and denoising parameters, data standardization and normalization parameters) are determined based on the training set and are also used for the unified processing of the validation set, test set, and real-time acquired data in the subsequent process. S140: Based on the time-domain feature index, frequency-domain feature index, time-frequency-domain feature index and their corresponding category labels of the preprocessed vibration signal, the preset classification model is trained and optimized, and the training and optimization classification model is used as the ship collision event recognition model. The ship collision event identification model is embedded into the online monitoring system. During online monitoring, the vibration feature vector to be identified is collected in real time and input into the model, which can output the predicted probability that the current vibration event belongs to a ship collision event.

[0029] In one possible implementation, step S140 includes the following: S141: Use the multi-dimensional feature index set as the input sample of the classification model, and use the category label corresponding to each vibration signal in the dataset as the expected output of the classification model; S142: Using a training algorithm adapted to the classification model, the model is iteratively trained based on the input samples and the expected output to optimize the internal parameters or model structure of the classification model; S143: During the iterative training process, the training process is controlled by a preset convergence condition or the maximum number of iterations, and the optimized classification model is used as the ship collision event recognition model.

[0030] In one possible implementation, the ship collision event recognition model trained and optimized in step S143 is further evaluated for the importance / contribution of feature indicators. Specifically, the importance / contribution score of each time-domain feature indicator, frequency-domain feature indicator, and time-frequency-domain feature indicator in the multi-dimensional feature indicator set to the model classification result is calculated, and the importance / contribution ranking result of the feature indicators is output according to the magnitude of the importance / contribution score. This optimizes the dimension of the vibration feature vector to be identified, assists in the subsequent setting of the combined discriminant formula based on multiple vibration feature indicators, or is used to dynamically adjust the vibration feature indicators included in the combined discriminant formula. For example, the preset classification model is a random forest model, a gradient boosting decision tree model, or a neural network model. Considering that the time domain, frequency domain, and time-frequency domain features extracted in this application have multiple dimensions and large numerical ranges, and that real ship collision event samples are relatively scarce, it is preferred to use a decision tree-based learning model (such as a random forest model or a gradient boosting decision tree model) as the preset classification model.

[0031] The following sections explain the training and optimization processes for the Random Forest model and the Gradient Boosting Decision Tree model, respectively: When using a random forest model as the classification model, the training optimization process is as follows: (1) Create an initialized random forest model using default parameters; (2) Use the time-domain feature index, frequency-domain feature index and time-frequency-domain feature index of each vibration signal in the multi-dimensional feature index set as input samples of the random forest model; (3) Use the category label corresponding to each vibration signal as the expected output of the random forest model; (4) Set performance evaluation indexes (including but not limited to precision, recall, F1 score and accuracy), and use grid search strategy and 5-fold cross-validation in the preset hyperparameter grid space to optimize and evaluate the hyperparameters of the random forest model; until the hyperparameter grid space is traversed or the preset convergence condition is met, the random forest model with the best performance evaluation index value is used as the ship collision event recognition model; After the model training is completed, the importance scores of each time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index are output; the feature indexes are sorted and filtered according to the feature importance scores.

[0032] When using the gradient boosting decision tree model as a classification model, its training and optimization process follows the steps of the random forest model described above. After the model training is completed, the importance score of each feature indicator is output based on the information gain or number of splits of each time-domain feature indicator, frequency-domain feature indicator, and time-frequency-domain feature indicator. The only difference is that when using the random forest model, the preset hyperparameter grid space includes a combination of values ​​for the number of decision trees, the maximum depth of the decision tree, the minimum number of samples required for node splits, and the minimum number of samples for leaf nodes; when using the gradient boosting decision tree model, the preset hyperparameter grid space includes a combination of values ​​for the learning rate, the number of decision trees, the maximum depth of the decision tree, and the subsample ratio.

[0033] In another possible implementation, when the number of ship collision event samples is sufficient and computational resources are adequate, the classification model can also employ a neural network model; the neural network model includes an input layer, at least one hidden layer, and an output layer, and the training and optimization process is as follows: (1) Initialize the weight matrix and bias vector of the neural network model; (2) Input the time-domain feature index, frequency-domain feature index and time-frequency-domain feature index of each vibration signal in the multi-dimensional feature index set into the input layer. After weighted summation and activation function mapping in the hidden layer, the predicted classification result of each vibration signal is obtained in the output layer, including vibration event category and predicted probability value. (3) Calculate the loss function value based on the predicted classification results of each vibration signal and its true category label; the loss function adopts the cross-entropy loss function, the expression of which is: ; Where H is the loss function value; n is the total number of vibration signal samples in the training set, and i is the index number of the vibration signal sample in the training set; m is the total number of vibration event categories, and j is the index number of the vibration event category; y ij Let be the true category label value of the i-th vibration signal belonging to the j-th type of vibration event. Let be the predicted probability value of the i-th vibration signal belonging to the j-th type of vibration event; For example, taking a binary classification task, the total number of vibration event categories is 2, and the category labels include a "normal vibration" label representing a normal state and a "ship collision" label representing an abnormal state; if the i-th vibration signal belongs to the j-th category, then y ij =1, otherwise y ij =0; (4) Based on the loss function value, use the gradient descent algorithm or the adaptive moment estimation algorithm to calculate the gradient of the loss function with respect to the weight matrix and the bias vector through backpropagation, and update the weight matrix and the bias vector according to the gradient; (5) Repeat the aforementioned forward calculation and backward update steps until the loss function value converges, the preset number of iterations is reached, or the performance evaluation index value on the validation set no longer increases, and the trained neural network model is obtained. Optionally, the contribution scores of each time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index to the model classification results can be further calculated. The contribution scores of each feature index can be quantified and output through methods including but not limited to importance analysis, SHAP value attribution analysis, gradient-based attribution analysis, or attention mechanism weight analysis.

[0034] In one possible implementation, the time-domain characteristic indicators are selected as peak value (PV) and rise time (RT); specifically: Acceleration signals, velocity signals, or displacement signals can be selected as the specific manifestation of vibration, with acceleration signals being preferred. When the vibration signal is an acceleration signal, the peak value is the absolute maximum value of the acceleration signal within a preset time window, i.e. , Let N be the acceleration value at the t-th sampling point, and N be the total number of sampling points within the current preset time window; The rise time refers to the time it takes for the amplitude of the acceleration signal to rise from 10% of its peak value to 90%, i.e., RT = t 90% -t 10% , t 10% and t 90% These represent the times when the amplitude of the acceleration signal reaches 10% and 90% of its peak value, respectively. Furthermore, the time-domain characteristic index also includes the waveform factor FF; specifically: The waveform factor is the ratio of the root mean square value of acceleration to the average absolute value of acceleration. RMS is the root mean square value of acceleration. ; In one possible implementation, the frequency domain features include the spectral centroid SC, used to characterize the dominant frequency components of the vibration signal in the frequency domain, reflecting the transient impact characteristics of the impact event; specifically: Define the spectrum of the vibration signal at the k-th frequency point after Fourier transform as follows: f k Let M be the frequency value corresponding to the k-th frequency point, M be the total number of frequency points in the spectrum, and k be the index number of the frequency point. , The frequency value f is the amplitude at the k-th frequency domain point. k With the sampling frequency F of the vibration signal s satisfy , k=1,2,…,M.

[0035] In one possible implementation, the time-frequency domain features include wavelet packet energy entropy (WPEE), used to measure the complexity of the energy distribution of the vibration signal across the frequency band; specifically: The vibration signal is decomposed into L-level wavelet packet decomposition to obtain signal components with K frequency band nodes, where K=2. L ; Calculate the wavelet packet coefficients W for all time sampling points under the r-th frequency band node. r,b The sum of squares gives the energy E of the node in that frequency band. r ,Right now B is the total number of time sampling points under the r-th frequency band node, b is the index number of the time sampling point under the r-th frequency band node, and r is the index number of the frequency band node, b=1,2,…,B; r=1,2,…,K;

[0036] Calculate the energy percentage of the r-th frequency band node based on the energy of each frequency band node. ; Calculate wavelet packet energy entropy .

[0037] In one optional implementation, the multi-level threshold adaptive alarm mechanism includes: Set peak threshold T PV As a baseline threshold; when the real-time collected peak PV is greater than the peak threshold T PV When this occurs, a Level 1 alarm is triggered; Set rise time threshold T RT and / or wavelet packet energy entropy T WPEE The discriminant based on a combination of multiple vibration characteristic indices is specifically: based on real-time acquired peak PV, rise time RT, wavelet packet energy entropy WPEE, and their corresponding threshold T. PV T RT T WPEE At least two vibration characteristic indicators can be selected for combined discrimination. When the combined discrimination formula is satisfied, a level two alarm is triggered; for example, when the real-time collected peak value is greater than the peak value threshold T PV And simultaneously satisfy the condition that the rise time is less than the rise time threshold T. RT When the wavelet packet energy entropy is greater than the wavelet packet energy entropy threshold T WPEE A level 2 alarm is triggered if at least one of the following conditions is met; Set decision threshold T p When the classification probability predicted by the model is greater than the decision threshold T p At that time, a level three alarm was triggered; When all three alarm levels (Level 1, Level 2, and Level 3) are triggered within a preset time window, a ship collision alarm signal is issued.

[0038] In one possible implementation, the peak threshold T PV Rise time threshold T RT Wavelet packet energy entropy T WPEE The acquisition process includes: Visual analysis of historically collected multi-dimensional vibration feature datasets is performed. By plotting scatter plots of various vibration feature indicators, the sample distribution and clustering of normal vibration events such as normal vehicle passage / environmental excitation and ship collision events are displayed. Based on the boundary separation characteristics between sample clusters of different categories in the scatter plot, critical feature values ​​that can effectively distinguish various types of events are extracted manually or through boundary detection algorithms, and these critical values ​​are set as the initial discrimination threshold. Optionally, the peak threshold T PV The value range is 0.01 m / s 2 ~0.012m / s 2 Rise time threshold T RT The value range is 0.15s~0.2s, and the wavelet packet energy entropy threshold T WPEE The value range is 2.2 to 2.8.

[0039] Furthermore, the decision threshold T p The initial value range is 0.8~0.85.

[0040] Furthermore, the vibration signal within the current time window is acquired in real time, and its root mean square acceleration (RMS) value is calculated as a real-time background vibration index. Based on the real-time background vibration index, the decision threshold T used to determine the ship collision event is adjusted according to a preset dynamic mapping relationship or proportional coefficient. p Perform adaptive updates; for example, when the root mean square acceleration (RMS) increases, adjust T... p Adjust the settings appropriately to reduce false alarms caused by environmental interference.

[0041] Based on the same inventive concept, this application also provides a system for implementing the aforementioned multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics, comprising: The model training module collects raw vibration signals under different vibration events, divides the training set and preprocesses it, extracts multi-dimensional vibration features of each vibration signal in the training set, and constructs a ship collision event recognition model based on the multi-dimensional vibration features and their corresponding category labels. The real-time acquisition module acquires the vibration signal to be identified and performs preprocessing to extract the vibration feature vector to be identified. The event recognition module inputs the vibration feature vector to be identified into the ship collision event recognition model and outputs the classification probability that the vibration signal to be identified belongs to the ship collision event. The multi-level threshold adaptive alarm module sets a basic threshold based on a single time-domain feature index, a combined discriminant based on multiple vibration feature indices, and a decision threshold based on a ship collision event recognition model; it establishes a multi-level threshold adaptive alarm mechanism; and when all three levels of alarms (Level 1, Level 2, and Level 3) are triggered within a preset time window, a ship collision alarm signal is issued.

[0042] It should be understood that the division of processing modules in the above system is only a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, the processing modules in the system can be implemented by the processor calling software; for example, the system includes a processor connected to memory, which stores instructions. The processor calls the instructions stored in memory to implement any of the above methods or to implement the functions of each processing module in the system. Here, the processor is a general-purpose processor, and the memory can be internal or external to the system.

[0043] The following detailed explanation of the implementation process of this application, with reference to an application example, illustrates the feasibility and effectiveness of this application.

[0044] I. Bridge Overview and Data Collection:

[0045] This embodiment selects the Xuebuhe Bridge (hereinafter referred to as "Xuebuhe Bridge") on Provincial Highway S265 Zhenli Line as the monitoring object. The Xuebuhe Bridge is 72.12m long, with a center pile number of K52+022, and a design load of Highway-I. Although there is no navigation under the bridge, there are factories near the bridge site, and ships occasionally pass by. Therefore, it is necessary to monitor the bridge span of the navigation channel for ship collisions.

[0046] A single-axis accelerometer was installed on the outer side of the No. 2 cap beam on the right side of the bridge to continuously collect bridge vibration signals at a sampling frequency of 20Hz. In this embodiment, a signal sequence containing typical vehicle vibration and ship collision events was selected for analysis.

[0047] II. Feature Vector Extraction and Dataset Definition: Total sample size: A total of 834 historical vibration events were collected, including 762 non-ship collision events and 72 ship collision events. Time window setting: During the offline training phase, a time window of 1 minute is selected, and each time window contains 1200 data points based on a sampling frequency of 20Hz; Dataset partitioning: The above 834 sets of samples were divided into a training set (500 sets), a test set (167 sets), and a validation set (167 sets) in a ratio of 6:2:2. Construction of a multi-dimensional feature index set: For each vibration signal in the training set, extract five feature indices: peak value, rise time, waveform factor, spectral centroid, and wavelet packet energy entropy. Define the 5-dimensional feature vector of the i-th vibration signal as F. (i) =[PV (i) ,RT (i) ,FF (i) ,SC (i) WPEE (i) ], where PV (i) RT (i) FF (i) SC (i) WPEE (i) Let i = 1, 2, 3, ..., 500, where i is the peak value, rise time, waveform factor, spectral centroid, and wavelet packet energy entropy of the i-th vibration signal. The training set consists of a multi-dimensional feature index set Ftrain={F_train} composed of the aforementioned 500 feature vectors. (1) ,F (2) ,...,F (500)}

[0048] III. Model Training and Parameter Optimization: This embodiment uses a random forest model to construct a ship collision event recognition model; Within a predefined hyperparameter grid space, a grid search strategy and 5-fold cross-validation were used to optimize the model's hyperparameters, resulting in the optimal parameters shown in Table 1. Table 1 Optimal Parameter Table for the Model

[0049] Figure 3 Based on the confusion matrix of the validation set (167 sets), the performance of the model can be summarized as follows: The model processed 167 samples on the test set, correctly classifying 153 + 12 = 165, with an overall accuracy of 165 / (165+2)≈98.8%. For the "non-ship collision" category label, there were 153 samples, all of which were correctly detected, with a recall of 100%, a precision of 153 / (153+2)≈98.71%, and an F1 score of 99.35%. For the "ship collision" category label, there were 14 real ship collision samples, with the model correctly detecting 12 and misclassifying 2 as normal vibration events, resulting in a recall of 12 / 14≈85.71%, a precision of 100%, and an F1 score of 92.31%. The model performance evaluation results based on the validation set are summarized below: Table 2. Model performance evaluation results based on the validation set.

[0050] The model demonstrated excellent performance in both F1 score and overall accuracy, and the validated random forest model was deployed to the online monitoring system.

[0051] like Figure 4 As shown, the importance of each vibration characteristic index is ranked as follows: WPEE > PV > FF > RT > SC. Figure 5 The cumulative importance is calculated based on the order of importance. When the number of vibration characteristic indicators reaches 3, the cumulative importance exceeds 90%. Set peak threshold T PV As a baseline threshold, when the real-time collected peak PV is greater than the peak threshold T PV When this occurs, a Level 1 alarm is triggered; The discriminant based on a combination of multiple vibration characteristic indicators is defined as follows: when the real-time collected peak value is greater than the peak value threshold T... PV And the wavelet packet energy entropy is greater than the wavelet packet energy entropy threshold T WPEE At that time, a level two alarm is triggered.

[0052] IV. Ship Collision Prediction Based on Multi-Level Threshold Alarms: The online monitoring system reads sensor data in real time, repeats the preprocessing and feature extraction process with a time window of 1 second, and collects the peak value PV, rise time RT, waveform factor FF, spectral centroid SC, and wavelet packet energy entropy WPEE of the vibration signal to be identified in real time. The 5-dimensional feature vector is input into the deployed random forest model, which predicts whether a ship collision will occur within the current time window and outputs the predicted category and probability value of the vibration event. Table 3 shows the vibration feature vectors and model prediction results corresponding to the vibration signals collected at two different times: Table 3. Vibration event category prediction results

[0053] The following three-level threshold adaptive alarm mechanism is set: Set peak threshold T PV It is 0.01 m / s 2 Wavelet packet energy entropy threshold T WPEE The decision threshold T is 2.5. p The initial value is 0.85. When the background noise increases (when the root mean square value of acceleration exceeds the preset noise threshold), the decision threshold is increased to 0.90. The alarm triggering process is as follows: Initial amplitude screening: For vibration signal with serial number 1, the peak value PV in its vibration characteristic vector is greater than 0.01 m / s. 2 This triggers a Level 1 alarm and initiates the next judgment process; Feature verification: For vibration signal with sequence number 1, its vibration feature vector simultaneously satisfies peak value PV > 0.01 m / s 2 Furthermore, if the wavelet packet energy entropy WPEE > 2.5, a level 2 alarm is triggered, and the next judgment process is initiated. Model decision: For vibration signal with sequence number 1, the model outputs a ship collision probability > 0.85, triggering a level 3 alarm; Final alarm output: After detecting the triggering of the first-level alarm, the system subsequently detected the vibration signal of the same sequence number triggering the second-level and third-level alarms within two consecutive preset time windows. This was determined to be a valid collision event, and the final ship collision alarm signal was issued. The system simultaneously recorded the complete vibration feature vector data of this event as [0.0347, 0.05, 3.689, 5.1065, 3.7557] (corresponding to [PV, RT, FF, SC, WPEE] respectively) and a prediction probability of 94.07%. The above embodiments are only for illustrating the technical concept and features of this application, and are intended to enable those skilled in the art to understand the content of this application and implement it accordingly. They should not be used to limit the scope of protection of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics, characterized in that, include: Raw vibration signals under different vibration events are collected, training sets are divided and preprocessed, multi-dimensional vibration features of each vibration signal in the training set are extracted, and a ship collision event recognition model is constructed based on the multi-dimensional vibration features and their corresponding category labels using a random forest model. The vibration signal to be identified is acquired and preprocessed to extract a vibration feature vector to be identified, which includes multiple time-domain feature indicators, at least one frequency-domain feature indicator, and at least one time-frequency-domain feature indicator. The time-domain feature indicators include peak value, rise time, and waveform factor. The frequency-domain feature indicators include the spectral centroid. The time-frequency-domain feature indicators include wavelet packet energy entropy. The vibration feature vector to be identified is input into the ship collision event identification model, and the classification probability of the vibration signal to be identified belonging to the ship collision event is output. Set a peak threshold as the base threshold; when the real-time collected peak value is greater than the peak threshold, trigger a level one alarm; set a rise time threshold and / or a wavelet packet energy entropy threshold; when the real-time collected peak value is greater than the peak threshold, and simultaneously satisfies at least one of the following conditions: rise time is less than the rise time threshold and wavelet packet energy entropy is greater than the wavelet packet energy entropy threshold, trigger a level two alarm. A decision threshold is set. When the classification probability is greater than the decision threshold, a level 3 alarm is triggered. When the level 1 alarm, level 2 alarm, and level 3 alarm are all triggered within a preset time window, a ship collision alarm signal is issued.

2. The bridge-ship collision event multi-level threshold adaptive alarm method based on multi-dimensional vibration characteristics according to claim 1, characterized in that, The process involves collecting raw vibration signals under different vibration events, dividing the training set and preprocessing it, extracting multi-dimensional vibration features from each vibration signal in the training set, and constructing a ship collision event recognition model based on the multi-dimensional vibration features and their corresponding category labels using a random forest model. Raw vibration signals of the bridge under different vibration events, including vehicle traffic, environmental excitation, and ship collision, are collected. A vibration signal sample database with category labels is constructed based on the raw vibration signals. The vibration signal sample database is divided into a training set, a validation set, and a test set according to a preset ratio. The original vibration signals in the training set are preprocessed to obtain the preprocessed vibration signals. Extract the time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index of the preprocessed vibration signal to construct a multi-dimensional feature index set; Based on the time-domain feature index, frequency-domain feature index, time-frequency-domain feature index and their corresponding category labels of the preprocessed vibration signal, the random forest model is trained and optimized, and the trained and optimized random forest model is used as the ship collision event recognition model.

3. The bridge-ship collision event multi-level threshold adaptive alarm method based on multi-dimensional vibration characteristics according to claim 2, characterized in that, The random forest model is trained and optimized based on the time-domain feature index, frequency-domain feature index, time-frequency-domain feature index and their corresponding category labels of the preprocessed vibration signal, and the trained and optimized random forest model is used as the ship collision event recognition model, including: Create an initial random forest model; The time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index of each vibration signal in the multi-dimensional feature index set are used as input samples of the random forest model; the category label corresponding to each vibration signal is used as the expected output of the random forest model. Set performance evaluation metrics, and use a grid search strategy and 5-fold cross-validation within a preset hyperparameter grid space to optimize and evaluate the hyperparameters of the random forest model; The random forest model with the best performance evaluation index is selected as the ship collision event recognition model after traversing the hyperparameter grid space or satisfying the preset convergence condition.

4. The bridge-ship collision event multi-level threshold adaptive alarm method based on multi-dimensional vibration characteristics according to claim 3, characterized in that, The importance / contribution of the time-domain feature index, frequency-domain feature index, and time-frequency-domain feature index are evaluated on the trained and optimized ship collision event recognition model.

5. The bridge-ship collision event multi-level threshold adaptive alarm method based on multi-dimensional vibration characteristics according to claim 1, characterized in that, The root mean square value of acceleration of the vibration signal to be identified is calculated in real time; based on the root mean square value of acceleration, the decision threshold is adaptively updated according to a preset dynamic mapping relationship or proportional coefficient.

6. A system for implementing the multi-level threshold adaptive alarm method for bridge-ship collision events based on multi-dimensional vibration characteristics as described in any one of claims 1-5, characterized in that, include: The model training module collects raw vibration signals under different vibration events, divides the training set and preprocesses it, extracts multi-dimensional vibration features of each vibration signal in the training set, and constructs a ship collision event recognition model based on the multi-dimensional vibration features and their corresponding category labels. The real-time acquisition module acquires the vibration signal to be identified and performs preprocessing to extract the vibration feature vector to be identified. The event recognition module inputs the vibration feature vector to be identified into the ship collision event recognition model and outputs the classification probability that the vibration signal to be identified belongs to the ship collision event. The multi-level threshold adaptive alarm module sets a basic threshold based on a single time-domain feature index, a combined discriminant based on multiple vibration feature indices, and a decision threshold based on a ship collision event recognition model; it establishes a multi-level threshold adaptive alarm mechanism; and when the first-level alarm, second-level alarm, and third-level alarm are all triggered within a preset time window, a ship collision alarm signal is issued.

Images