A main cable amplitude anomaly automatic identification method and system

Abstract

The application discloses a kind of main cable amplitude anomaly automatic identification method and system, it is related to amplitude identification technical field.The sensor collects vibration signal, peak value feature extraction and hilbert transform analysis are carried out to vibration signal, the analytic result is analyzed by principal component analysis dimension reduction, and is arranged according to time and sensor position, respectively obtains peak value sparse matrix, amplitude time series and time delay characteristic matrix;Based on amplitude time series, sub-sequence is divided, and feature extraction is carried out to sub-sequence using time delay long short-term memory network, based on time delay characteristic matrix, correlation weight calculation is carried out by attention mechanism in combination with peak value sparse matrix, and obtain space-time characteristic matrix;Based on space-time characteristic matrix, by affine transformation to amplitude time series, obtain abnormal score, based on analytic result, obtain difference matrix, locate abnormal area by difference matrix, identify abnormal type of abnormal area by abnormal score and correlation weight result, provide accurate amplitude anomaly type identification.

Description

Technical Field

[0001] This invention relates to the technical field of amplitude recognition, and in particular to an automatic method and system for identifying abnormal amplitude of main cable. Background Technology

[0002] Anomaly identification in main cable amplitude is a core task in bridge health monitoring, especially in structures like suspension bridges where the main cable bears significant mechanical loads. The vibration signal of the main cable is often affected by external factors such as wind speed, temperature, and load variations, leading to time delays in the vibration signals collected by sensors. Traditional amplitude anomaly identification methods often rely on simple threshold judgments or feature extraction based on a single sensor, failing to effectively capture the time delay effect and the correlation between sensors, and struggling to handle complex spatiotemporal variations. Therefore, how to identify anomalies by combining multi-sensor data with time delay features and spatiotemporal modeling techniques has become a pressing technical challenge.

[0003] Currently, Chinese invention application CN201410437492.0 discloses a method for identifying engine rotor blade amplitude data based on blade tip timing technology. Its features include: separating the blade amplitude change trend; non-vibration-induced blade displacement including rotor oscillation; shaft torsion; axial movement; centrifugal force-induced blade torsion; mechanical noise; calculating blade amplitude based on rotational speed and rotor diameter when using shaft positioning signals during testing; when the rotor oscillates, the rotor blades follow suit, the blade tip and sensor positions change, and a constant component appears in the blade measurement data; axial movement and shaft torsion cause the sensor to measure the blade tip position, resulting in a change in the blade tip arrival time, but the axial movement frequency is lower than the blade vibration frequency, resulting in a slow change in the blade measurement data. This invention effectively improves the signal-to-noise ratio of blade amplitude measurement, providing data for accurate analysis of blade vibration characteristics based on this amplitude result, significantly advancing blade tip timing analysis technology. However, related technologies struggle to capture time delay effects and the interrelationships between multiple sensors, lack integration of temporal and spatial characteristics, lack spatiotemporal correlation, have low identification efficiency, and lack accuracy. Summary of the Invention

[0004] The technical problem solved by this invention is that related technologies have difficulty capturing time delay effects and the interrelationships between multiple sensors, do not combine time and space features, lack spatiotemporal correlation, have low recognition efficiency, and lack accuracy.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] An automatic identification method for main cable amplitude anomalies includes the following steps:

[0007] Step S1: Collect vibration signals through sensors, extract peak features and perform Hilbert transform analysis on the vibration signals, and perform principal component analysis to reduce the dimensionality of the analysis results. Arrange the extraction results, analysis results and dimensionality reduction results according to time and sensor position to obtain the peak sparse matrix, amplitude time series and time delay feature matrix respectively.

[0008] Step S2: Divide the subsequences based on the amplitude time series, extract features from the subsequences using a time-delayed short-term memory network, calculate the association weights based on the time-delay feature matrix by combining the peak sparse matrix with the attention mechanism, and fuse the feature extraction results and association weight results to obtain the spatiotemporal feature matrix.

[0009] Step S3: Based on the spatiotemporal feature matrix, anomaly scores are obtained by performing an affine transformation on the amplitude time series. A difference matrix is ​​obtained based on the analysis results. Anomaly regions are located according to the difference matrix. The anomaly types of the anomaly regions are identified by the anomaly scores and correlation weight results.

[0010] Preferably, step S1 specifically includes:

[0011] Step S11: Obtain the position information of the sensor and the vibration signal collected by the sensor; perform Hilbert transform on the vibration signal of each sensor to obtain the analytical signal of the vibration signal; calculate the amplitude of the analytical signal to obtain the amplitude time series of each sensor.

[0012] Step S12: Based on the sensor position information and the theoretical wave velocity of the main cable vibration, calculate the theoretical time delay value between every two sensors to form a theoretical time delay matrix.

[0013] The rows and columns of the theoretical time delay matrix correspond to the sensor indices, and the elements in the theoretical time delay matrix represent the theoretical time delay values ​​between the corresponding two sensors.

[0014] Step S13: Based on the amplitude time series, calculate the actual time delay value between each pair of sensors using the cross-correlation analysis method to form the actual time delay matrix;

[0015] Wherein, the rows and columns of the actual time delay matrix correspond to the sensor indexes, and the elements in the actual time delay matrix represent the actual time delay values ​​between the corresponding two sensors;

[0016] Step S14: Flatten and concatenate the theoretical time delay matrix and the actual time delay matrix to obtain an initial time delay feature matrix, and reduce the dimensionality of the initial time delay feature matrix by principal component analysis to obtain the time delay feature matrix.

[0017] Step S15: Perform peak detection on the amplitude time series of each sensor and extract the local maximum value in the vibration signal as the peak feature;

[0018] Step S16: Based on the time delay sampling between every two sensors, extract the corresponding amplitude peak value and construct a peak sparse matrix. In the peak sparse matrix, the corresponding element of the amplitude peak pair between the sensors is non-zero, otherwise it is zero.

[0019] Preferably, feature extraction of the subsequence using a time-extended short-term memory network specifically includes:

[0020] The amplitude time series of each sensor is divided into k subsequences according to the time window. The time delay feature vector of each sensor in the time delay feature matrix is ​​obtained. The corresponding time offset is applied to the subsequence of each sensor to obtain the time-aligned subsequence of each sensor. k is an integer greater than 2.

[0021] After time-series alignment, the subsequence of each sensor and the corresponding time delay feature vector of the sensor are input into the time-delayed short-term memory network. The final hidden state output by the time-delayed short-term memory network is obtained. The final hidden states of all sensors are concatenated to obtain the time feature matrix.

[0022] Preferably, the extended short-term memory network specifically includes:

[0023] The hidden state of the previous time step, the input subsequence of the current time step, and the time delay feature vector corresponding to each sensor are concatenated to form a combined input vector;

[0024] Based on the combined input vector, the input gate signal, forget gate signal, output gate signal, and candidate cell state are computed in parallel through learnable linear transformations and activation functions.

[0025] The input gate signal, forget gate signal, and output gate signal are calculated using the Sigmoid activation function, and the candidate cell state is calculated using the hyperbolic tangent activation function.

[0026] The candidate cell state is calculated using the forget gate signal and the input gate signal to obtain the cell state updated at the current time step. The calculation expression is as follows:

[0027] ;

[0028] in, This represents the cell state after the current time step t and sensor j update. This represents the forget gate signal for the current time step t and sensor j. This represents the cell state at the previous time step t-1 and sensor j, and ⊙ represents element-wise multiplication. This represents the input gate signal for the current time step t and sensor j. This represents the candidate cell state at the current time step t and sensor j;

[0029] The current cell state is adjusted using the output gate signal, and the hidden state at the current time step is output. The calculation expression is as follows:

[0030] ;

[0031] in, This represents the hidden state output by sensor j at the current time step t. This represents the output gate signal of the current time step t and sensor j. Let be the hyperbolic tangent function, and ⊙ denotes element-wise multiplication.

[0032] Preferably, the process of obtaining the spatiotemporal feature matrix specifically includes:

[0033] Based on the time delay feature matrix, the initial correlation weight matrix between sensors is calculated through the attention mechanism. The initial correlation weight matrix is ​​biased by the peak sparse matrix to obtain the correlation weight matrix. The correlation weight matrix is ​​weighted and aggregated with the value matrix to obtain the spatial feature matrix.

[0034] The expression for bias calculation is:

[0035] ;

[0036] For the correlation weight matrix, It is a peak sparse matrix. To adjust the parameters, For querying the matrix, The key matrix, This is the initial correlation weight matrix;

[0037] The rows and columns of the association weight matrix correspond to the sensor indices, and the elements in the association weight matrix represent the association strength between the corresponding two sensors, with the sum of the elements in each row being 1.

[0038] The time feature matrix and the spatial feature matrix are weighted and averaged to form the spatiotemporal feature matrix.

[0039] Preferably, calculating the outlier fraction of the amplitude time series specifically includes:

[0040] The spatiotemporal feature matrix is ​​input into the feedforward neural network, which outputs transformation parameters. Based on the transformation parameters, the vibration time series is subjected to P iterations of affine transformation to obtain the anomaly density function and the cumulative Jacobian determinant. The cumulative Jacobian determinant is the product of the Jacobian determinants of the P iterations.

[0041] Based on the probability density transformation formula, the anomaly probability density of the amplitude time series is calculated, and the expression is as follows:

[0042] ;

[0043] in, For anomaly density function, For cumulative Jacobian determinants, The spatiotemporal characteristic matrix, It is an amplitude time series. It is an affine transformation function;

[0044] The negative logarithm of the anomaly probability density is used as the anomaly score.

[0045] Preferably, the affine transformation operation specifically includes:

[0046] The input vector is randomly divided into a first part and a second part. The first part and the transformation parameters are input into the transformation function, which outputs the scaling and translation parameters. The second part undergoes an affine transformation, and the calculation expression is as follows:

[0047] ;

[0048] ;

[0049] in, The spatiotemporal characteristic matrix, and The input vector is divided into two sub-vectors. For scaling function, It is a translation function. and There are two output sub-vectors;

[0050] The two output sub-vectors are recombine to obtain the output vector;

[0051] The input vector includes:

[0052] The input vector for the first iteration is the vector obtained after flattening the vibration time series data;

[0053] Except for the first iteration, the input vector is the output vector of the affine transformation in the previous iteration;

[0054] The mathematical expression for the Jacobian determinant is:

[0055] ;

[0056] in, scale function The i-th component of the output vector.

[0057] Preferably, the location of abnormal areas specifically includes:

[0058] The difference matrix is ​​obtained by subtracting the actual time delay matrix from the theoretical time delay matrix. Elements in the difference matrix whose related time delay difference values ​​exceed the first comparison threshold are identified. The cable segment between the two sensors corresponding to each element is marked as an abnormal cable segment. If E abnormal cable segments are spatially consecutive, the E abnormal cable segments are merged into an abnormal region, where E is an integer greater than or equal to 2.

[0059] The rows and columns of the difference matrix correspond to sensor indices, and the elements in the difference matrix represent the correlation delay difference values ​​between two corresponding sensors.

[0060] Preferably, identifying the anomaly type in the abnormal region specifically includes:

[0061] Based on the association weight matrix, calculate the average association strength between each sensor and all other sensors;

[0062] Obtain the sensor set that constitutes the abnormal region;

[0063] If the anomaly score exceeds the second comparison threshold, and the average correlation strength of all sensors in the sensor set is higher than the global threshold, then the anomaly region is identified as a global excitation anomaly.

[0064] If the anomaly score exceeds the second comparison threshold, and the correlation strength between two sensors within an anomaly region in the correlation weight matrix is ​​lower than the third comparison threshold, then identification is performed using the difference matrix:

[0065] When the relevant time delay difference values ​​in the difference matrix exhibit a regular anomaly, it is identified as a local structural anomaly.

[0066] When the correlation delay difference values ​​in the difference matrix exhibit irregular discrete anomalies, it is identified as a sensor fault.

[0067] An automatic identification system for main cable amplitude anomalies includes a data acquisition module, a processing module, and an identification module.

[0068] The acquisition module is used to acquire vibration signals through sensors, extract peak features and perform Hilbert transform analysis on the vibration signals, and perform principal component analysis to reduce the dimensionality of the analysis results. The extraction results, analysis results and dimensionality reduction results are arranged according to time and sensor position to obtain the peak sparse matrix, amplitude time series and time delay feature matrix, respectively.

[0069] The processing module is used to divide the subsequence based on the amplitude time series, extract features from the subsequence using a time-delayed short-term memory network, calculate the association weight based on the time-delay feature matrix by combining the peak sparse matrix with the attention mechanism, and fuse the feature extraction results and association weight results to obtain the spatiotemporal feature matrix.

[0070] The identification module is used to obtain anomaly scores by performing an affine transformation on the amplitude time series based on the spatiotemporal feature matrix, obtain a difference matrix based on the analysis results, locate abnormal regions according to the difference matrix, and identify the anomaly type of the abnormal region through the anomaly scores and correlation weight results.

[0071] The beneficial effects of this invention are as follows: This invention extracts time delay features from vibration signals collected by sensors, and combines this with cross-correlation analysis, calculation of theoretical and actual time delay matrices to accurately capture the time delay effects and peak sparsity characteristics between different sensors. By integrating time delay features and peak sparsity characteristics into spatiotemporal features, the combination of these features enables the system to more comprehensively capture dynamic changes in vibration signals, improving the accuracy of anomaly detection. Furthermore, by mapping the amplitude signal to the latent space through affine transformation and adjusting the transformation parameters in conjunction with the spatiotemporal feature matrix, the anomaly probability density is calculated to obtain an anomaly score. This enhances the ability to identify complex anomaly patterns, allows for more flexible adaptation to different types of anomalies, and provides more accurate anomaly type identification. Attached Figure Description

[0072] Figure 1 The following is a basic flowchart of an automatic identification method for main cable amplitude anomalies provided in one embodiment of the present invention. Detailed Implementation

[0073] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0074] Example 1, referring to Figure 1 As an embodiment of the present invention, an automatic identification method for main cable amplitude anomalies is provided, comprising the following steps:

[0075] Step S1: Vibration signals are collected by sensors, peak features are extracted and Hilbert transform is performed on the vibration signals, and principal component analysis is performed to reduce the dimension of the analysis results. The extraction results, analysis results and dimension reduction results are arranged according to time and sensor position to obtain the peak sparse matrix, amplitude time series and time delay feature matrix respectively.

[0076] Step S2: Divide the subsequences based on the amplitude time series, extract features from the subsequences using a time-delayed short-term memory network, calculate the association weights based on the time-delay feature matrix by combining the attention mechanism with the peak sparse matrix, and fuse the feature extraction results and association weight results to obtain the spatiotemporal feature matrix.

[0077] Step S3: Based on the spatiotemporal feature matrix, anomaly scores are obtained by performing an affine transformation on the amplitude time series. The difference matrix is ​​obtained based on the analytical results. Anomaly regions are located based on the difference matrix. The anomaly types of the anomaly regions are identified by the anomaly scores and correlation weight results.

[0078] This invention extracts time-delay features from vibration signals acquired by sensors and, combined with cross-correlation analysis and calculations of theoretical and actual time-delay matrices, accurately captures the time-delay effects and peak sparsity characteristics between different sensors. By integrating time-delay and peak sparsity features into spatiotemporal features, the system can more comprehensively capture dynamic changes in vibration signals, improving the accuracy of anomaly detection. Furthermore, affine transformation maps the amplitude signal to a latent space, and by adjusting transformation parameters based on the spatiotemporal feature matrix, anomaly probability density is calculated to obtain anomaly scores. This enhances the ability to identify complex anomaly patterns, enabling more flexible adaptation to different types of anomalies and providing more accurate anomaly type identification.

[0079] In a specific embodiment, step S1 specifically includes:

[0080] Step S11: Obtain the position information of the sensor and the vibration signal collected by the sensor; perform Hilbert transform on the vibration signal of each sensor to obtain the analytical signal of the vibration signal; calculate the amplitude of the analytical signal to obtain the amplitude time series of each sensor.

[0081] It should be noted that the vibration signal acquired by the sensor is transformed using Hilbert transform to obtain its orthogonal components, which together constitute a complex analytic signal. The amplitude of this analytic signal is the instantaneous amplitude, and its change over time is the desired amplitude time series. This method can effectively remove the high-frequency carrier component from the vibration signal while retaining the low-frequency envelope reflecting energy changes, providing a stable input for subsequent spatiotemporal feature analysis.

[0082] The sensor is an accelerometer.

[0083] Step S12: Based on the sensor position information and the theoretical wave velocity of the main cable vibration, calculate the theoretical time delay value between every two sensors to form a theoretical time delay matrix.

[0084] In this matrix, the rows and columns correspond to the sensor indices, and the elements in the theoretical delay matrix represent the theoretical delay values ​​between the corresponding two sensors.

[0085] Step S13: Based on the amplitude time series, calculate the actual time delay value between each pair of sensors using the cross-correlation analysis method to form the actual time delay matrix;

[0086] The rows and columns of the actual delay matrix correspond to the sensor indices, and the elements in the actual delay matrix represent the actual delay values ​​between the corresponding two sensors.

[0087] Step S14: Flatten and stitch the theoretical time delay matrix and the actual time delay matrix to obtain the initial time delay feature matrix. Then, reduce the dimensionality of the initial time delay feature matrix using principal component analysis to obtain the time delay feature matrix.

[0088] Step S15: Perform peak detection on the amplitude time series of each sensor and extract the local maximum value in the vibration signal as the peak feature;

[0089] Step S16: Based on the time delay sampling between every two sensors, extract the corresponding amplitude peak value and construct a peak sparse matrix. In the peak sparse matrix, the corresponding element of the amplitude peak pair between the sensors is non-zero, otherwise it is zero.

[0090] It should be noted that the theoretical wave velocity of the main cable is calculated using the string vibration theory formula based on the material properties and tension state of the main cable. The calculation expression is as follows:

[0091] ;

[0092] in, For the theoretical wave velocity of vibration, For cable force, This represents the linear density.

[0093] The theoretical delay value is calculated using the following expression:

[0094] ;

[0095] in, This represents the distance between the sensors.

[0096] The actual time delay between each pair of sensors is calculated using cross-correlation analysis, specifically including:

[0097] After preprocessing the amplitude time series of every two sensors, the cross-correlation function is calculated based on the fast Fourier transform. The time lag is determined by finding the peak position of the cross-correlation function, and the time lag is converted into the actual time delay value according to the sampling time interval.

[0098] The time delay feature matrix is ​​obtained by dimensionality reduction of the initial time delay feature matrix using principal component analysis. Specifically, the time delay feature matrix includes:

[0099] After standardizing the initial time delay feature matrix, the top k principal component directions that contribute the most to its covariance matrix are extracted by principal component analysis. The standardized initial time delay feature matrix is ​​then projected onto the principal component directions to reduce the data to k dimensions. The resulting dimensionality-reduced feature matrix is ​​the time delay feature matrix.

[0100] For any two sensors i and j, the actual time delay value calculated in step S13 is used. Find time point t in the amplitude sequence of sensor j, and find time point t in the amplitude sequence of sensor i. Determine the local amplitude peak P near these two time points. i and P j Whether a valid peak pair is formed. In this embodiment, the time difference between the two peaks is less than a preset time tolerance threshold, which is set to half of the data sampling interval. If this condition is met, the two peaks are considered to be caused by the same vibration event, and P is set accordingly. i and P j The average amplitude is used as a non-zero value and filled into the i-th row and j-th column of the peak sparse matrix P. The peak sparse matrix is ​​constructed by traversing all sensor pairs and all time windows.

[0101] Peak detection of the amplitude time series for each sensor specifically includes:

[0102] The amplitude time series is smoothed and filtered to obtain a preprocessed sequence. Based on a threshold-based local maximum search algorithm, the amplitude threshold, minimum peak interval, and peak prominence parameters are set. All local maximum points that meet the conditions are identified and extracted from the preprocessed sequence. The timestamp and amplitude value corresponding to each local maximum point are recorded to obtain the peak characteristics of the sensor.

[0103] The amplitude threshold is determined based on the statistical characteristics of background noise. It is calculated by taking the mean and standard deviation of the amplitude sequence under normal stationary conditions and using the criterion of adding three times the standard deviation to the mean, so as to effectively distinguish between the actual vibration peak and background noise.

[0104] The setting of the minimum peak interval is directly related to the basic vibration frequency of the main cable structure. The dominant frequency is determined by spectrum analysis and converted into vibration period. Then, the period is converted into the number of sample points according to the sensor sampling frequency. Finally, 0.7 of the theoretical interval value is taken as the parameter to ensure that each independent vibration event is identified by only one main peak.

[0105] The peak prominence parameter is obtained by analyzing the prominence distribution of peaks detected in historical normal data, selecting 70% of the prominence distribution as the threshold, and filtering out peaks that are significantly prominent in the local waveform and have complete physical meaning.

[0106] By combining Hilbert transform, theoretical time delay calculation, cross-correlation analysis, and peak detection, the amplitude time series, time delay matrix, and peak sparsity matrix, including fault information, can be systematically extracted from the original vibration signal. Principal component analysis for dimensionality reduction reduces data redundancy and noise interference while preserving core information, providing high-quality, highly representative input features for subsequent anomaly identification models and laying the foundation for the accuracy of the entire method.

[0107] In a specific embodiment, feature extraction of subsequences using a time-extended short-term memory network specifically includes:

[0108] The amplitude time series of each sensor is divided into k subsequences according to the time window. The time delay feature vector of each sensor in the time delay feature matrix is ​​obtained. The corresponding time offset is applied to the subsequence of each sensor to obtain the time-aligned subsequence of each sensor. k is an integer greater than 2.

[0109] It should be noted that the time offset specifically includes: the time delay of sensor B is greater than the time delay of sensor A, that is... If the value is greater than 0, then the subsequence of sensor A will be shifted backward. A unit of time, conversely, if If the value is less than 0, the subsequence of sensor B will be shifted backward, aligning the vibration responses caused by the same excitation event in different sensor data in terms of time sequence. The vibration response caused by the same excitation event refers to the amplitude signal collected by different sensor locations under the same external excitation. Due to the different locations, the vibration responses collected by the sensors will have a time lag. External excitations include changes in wind speed, load, and temperature.

[0110] The number of subsequence partitions k ranges from 4 to 10, balancing temporal locality and computational efficiency.

[0111] By segmenting the amplitude sequence and combining it with the corresponding time delay feature vector, and inputting the segment into a time delay LSTM network, the temporal dynamic characteristics of the vibration signal and the spatial time delay information of the sensor are effectively fused.

[0112] After time-series alignment, the subsequence of each sensor and the corresponding time delay feature vector of the sensor are input into the time-delayed short-term memory network. The final hidden state output by the time-delayed short-term memory network is obtained. The final hidden states of all sensors are concatenated to obtain the time feature matrix.

[0113] In a specific embodiment, the extended short-term memory network specifically includes:

[0114] The hidden state of the previous time step, the input subsequence of the current time step, and the time delay feature vector corresponding to each sensor are concatenated to form a combined input vector;

[0115] Based on the combined input vector, the input gate signal, forget gate signal, output gate signal and candidate cell state are computed in parallel through learnable linear transformation and activation function;

[0116] The input gate signal, forget gate signal, and output gate signal are calculated using the Sigmoid activation function, while the candidate cell state is calculated using the hyperbolic tangent activation function.

[0117] The candidate cell states are calculated using the forget gate and input gate signals to obtain the updated cell states at the current time step. The calculation expression is as follows:

[0118] ;

[0119] in, This represents the cell state after the current time step t and sensor j update. This represents the forget gate signal for the current time step t and sensor j. This represents the cell state at the previous time step t-1 and sensor j, and ⊙ represents element-wise multiplication. This represents the input gate signal for the current time step t and sensor j. This represents the candidate cell state at the current time step t and sensor j;

[0120] It should be noted that, Representing the cell state at the current time step t and after sensor j updates, it is a multi-dimensional vector. It serves as the cell's long-term memory, storing sequence information after selective forgetting and addition. The forgetting gate signal, representing the current time step t and sensor j, is a multi-dimensional vector where each element has a value between 0 and 1 (generated by the sigmoid function). It controls the cell state at the previous time step. The proportion of each corresponding element retained. This represents the cell state at the previous time step t-1 and sensor j, which is a historical storage of long-term memory. ⊙ denotes element-wise multiplication, that is, multiplying corresponding elements of two vectors of the same dimension. It is used to achieve fine-grained control of cell state through gating signals. The input gate signal, representing the current time step t and sensor j, is a multi-dimensional vector, where each element has a value between 0 and 1. It controls the candidate cell state. The proportion of each corresponding element in the new cell state that is written into the new cell state. The candidate cell state at the current time step t and sensor j is represented by a multi-dimensional vector. It is generated by the current input, historical hidden states, and time delay conditions through a linear transformation and a hyperbolic tangent activation function, representing new information that may be added to memory.

[0121] The current cell state is regulated using the output gate signal, and the hidden state at the current time step is output. The calculation expression is as follows:

[0122] ;

[0123] in, This represents the hidden state output by sensor j at the current time step t. This represents the output gate signal of the current time step t and sensor j. Let be the hyperbolic tangent function, and ⊙ denotes element-wise multiplication.

[0124] It should be noted that the calculation of input gate signals, forget gate signals, output gate signals, and candidate cell states all explicitly depend on the time delay embedding vector; the update of cell states is jointly determined by the regulation of old states by the forget gate and the regulation of new candidate states by the input gate; the output of the hidden state is determined by the regulation of the current cell state by the output gate.

[0125] By using the time delay feature vector as one of the gating signals, this network structure can dynamically and adaptively adjust the updates of memory cells and the output of hidden states. This makes the time delay LSTM more sensitive to temporal dynamic changes caused by anomalies.

[0126] In a specific embodiment, the process of obtaining the spatiotemporal feature matrix includes:

[0127] Based on the time delay feature matrix, the initial correlation weight matrix between sensors is calculated through the attention mechanism. The initial correlation weight matrix is ​​biased by the peak sparse matrix to obtain the correlation weight matrix. The correlation weight matrix and the value matrix are weighted and aggregated to obtain the spatial feature matrix.

[0128] The expression for bias calculation is:

[0129] ;

[0130] For the correlation weight matrix, It is a peak sparse matrix. To adjust the parameters, For querying the matrix, The key matrix, This is the initial correlation weight matrix;

[0131] The rows and columns of the association weight matrix correspond to the sensor indices, and the elements in the association weight matrix represent the association strength between the corresponding two sensors, with the sum of the elements in each row being 1.

[0132] It should be noted that calculating the correlation weight matrix between sensors through the attention mechanism specifically includes:

[0133] The latency feature matrix is ​​transformed linearly to obtain the query matrix, key matrix, and value matrix. The product of the transpose of the query matrix and the key matrix is ​​calculated and divided by the scaling factor to obtain the attention score matrix. Each row of the attention score matrix is ​​normalized to obtain the association weight matrix.

[0134] The initial value of the adjustment parameter is set to 1.

[0135] Using the peak sparse matrix directly extracted from actual vibration signals as physical evidence, the initial attention weights calculated through time delay features are biased, guiding the entire calculation process to prioritize strengthening the correlation between sensor nodes with clear physical coupling. This makes the final generated correlation weight matrix not only data-adaptive but also physically interpretable, and can truly reflect the dynamic coupling relationship of the structure in vibration, providing a reliable logical basis for accurately distinguishing different types of anomalies.

[0136] Based on the correlation weight matrix, sensor features are dynamically weighted and aggregated to generate a spatial feature matrix. This process can effectively highlight abnormal spatial patterns whose correlation is weakened due to local damage or failure, while suppressing normal global vibration commonalities and enhancing the detection sensitivity to local structural anomalies and sensor faults. The time features representing the dynamic evolution of the time series are weighted and averaged with the above spatial correlation features to form a unified spatiotemporal feature matrix. The spatiotemporal feature matrix encodes the evolution law of anomalies in the time series and the propagation topology in the sensor network space, providing key input for subsequent steps to perform accurate probabilistic anomaly scoring and localization.

[0137] The time feature matrix and the spatial feature matrix are weighted and averaged to form the spatiotemporal feature matrix.

[0138] It should be noted that the weighted aggregation of the correlation weight matrix and the value matrix specifically includes:

[0139] For each sensor i, the mathematical expression for weighted summation is:

[0140] ;

[0141] in, For spatial feature vectors, The element in the i-th row and j-th column of the association weight matrix A is... This is the value vector for all other sensors.

[0142] The weighted average of the time feature matrix and the spatial feature matrix specifically includes:

[0143] By introducing fusion weight parameters, the spatiotemporal feature matrix is ​​calculated through element-level weighted summation. The calculation expression is as follows:

[0144] ;

[0145] in, The spatiotemporal characteristic matrix, For the fusion weight parameters, where , The time feature matrix, This is the spatial characteristic matrix.

[0146] It is a scalar parameter representing the relative importance weight assigned to time-dynamic features during the final feature fusion. This represents the importance weight assigned to spatial correlation features.

[0147] Set the number of hidden layer units of the LSTM to be equal to the projection dimension of the value matrix in the attention mechanism.

[0148] In a specific embodiment, calculating the outlier score of the amplitude time series includes:

[0149] The spatiotemporal feature matrix is ​​input into the feedforward neural network, which outputs transformation parameters. Based on the transformation parameters, the vibration time series is subjected to P iterations of affine transformation to obtain the anomaly density function and the cumulative Jacobian determinant. The cumulative Jacobian determinant is the product of the Jacobian determinants of the P iterations.

[0150] Based on the probability density transformation formula, the anomaly probability density of the amplitude time series is calculated, and the expression is as follows:

[0151] ;

[0152] in, For anomaly density function, For cumulative Jacobian determinants, The spatiotemporal characteristic matrix, It is an amplitude time series. It is an affine transformation function;

[0153] The negative logarithm of the anomaly probability density is used as the anomaly score.

[0154] It should be noted that when the spatiotemporal feature matrix is ​​input into the feedforward neural network, the output transformation parameters specifically include:

[0155] The spatiotemporal feature matrix is ​​input into a feedforward neural network. The network first compresses the two-dimensional feature matrix into a one-dimensional feature vector through a global average pooling operation. Then, the vector is passed through multiple hidden layers consisting of fully connected layers and nonlinear activation functions for nonlinear transformation and feature abstraction. Finally, the output layer maps the learned high-order spatiotemporal pattern into a set of structured transformation parameter vectors. This vector contains the control information required for each subsequent affine transformation iteration.

[0156] Anomaly score is determined when the input amplitude time series conforms to a normal pattern, meaning that after transformation it should fall within the high-probability region of the basic distribution, and the Jacobian determinant should not deviate extremely from 1; therefore, the anomaly score is low. Conversely, outlier data may map to a low-probability region or cause anomalies in the Jacobian determinant, thus increasing the anomaly score.

[0157] By mapping complex vibration data distributions to simple basis distributions through reversible affine transformations and calculating anomaly probability density using probability density transformation formulas, an anomaly score with clear probabilistic meaning can be directly output. This method achieves semi-supervised anomaly measurement, making anomaly judgment no longer dependent on fixed empirical thresholds, but based on the probability density of the data itself, thus improving the interpretability of anomaly judgment and its adaptability to different anomaly patterns.

[0158] In a specific embodiment, the affine transformation operation specifically includes:

[0159] The input vector is randomly divided into a first part and a second part. The first part and the transformation parameters are input into the transformation function, which outputs the scaling and translation parameters. The second part undergoes an affine transformation, and the calculation expression is as follows:

[0160] ;

[0161] ;

[0162] in, The spatiotemporal characteristic matrix, and The input vector is divided into two sub-vectors. For scaling function, It is a translation function. and There are two output sub-vectors;

[0163] The two output sub-vectors are recombine to obtain the output vector;

[0164] The input vector includes:

[0165] The input vector for the first iteration is the vector obtained after flattening the vibration time series data;

[0166] Except for the first iteration, the input vector is the output vector of the affine transformation in the previous iteration;

[0167] The mathematical expression for the Jacobian determinant is:

[0168] ;

[0169] in, scale function The i-th component of the output vector.

[0170] It should be noted that the transformation function is constructed using a lightweight fully connected neural network. Specifically, this network consists of an input layer, one hidden layer, and an output layer. The hidden layer contains 64 neurons, and the activation function is ReLU. The output layer directly outputs the scale and translation parameters, which have the same dimension as the input vector. The scale parameter is activated by Tanh and then multiplied by a constant to obtain the exponent.

[0171] The first part and the transformation parameters are input into the transformation function, and the output scaling and translation parameters specifically include:

[0172] In each iteration of the affine transformation, the transformation function receives the first part of the current input vector and the corresponding sub-parameters extracted from the transformation parameters output by the feedforward neural network. After concatenating the two into a combined vector, it is fed into another lightweight neural network consisting of a fully connected layer and an activation function for processing. This lightweight neural network outputs a fused vector through nonlinear mapping of the combined features. This fused vector is then divided into two parts according to the dimension, which serve as the scale parameter vector controlling the transformation amplitude and the translation parameter vector determining the transformation offset, respectively. This dynamically generates the specific parameters driving the current affine transformation.

[0173] In the affine transformation process, the Jacobian determinant is used to quantify the impact of the affine transformation on the anomaly probability density. The scaling function is a small neural network that takes the input data of the previous layer as input and outputs a vector with the same dimension as the part to be transformed through nonlinear transformation. The summation of the components in the scaling function reflects the cumulative effect of all scaling components. The concise form of the Jacobian determinant comes from the special structure of the affine transformation, that is, only a part of the input vector is scaled and translated each time, so that the Jacobian matrix of the transformation is in the form of triangular blocks. Thus, the determinant can be directly derived from the product of scaling factors. Each iteration of the affine transformation will generate a corresponding Jacobian determinant. The continuous product of the Jacobian determinants of all iterations constitutes the cumulative Jacobian determinant, which is finally used to calculate the conditional probability density of the amplitude time series through the probability density transformation formula, and then the anomaly score used for anomaly detection is derived.

[0174] In a specific embodiment, the location of the abnormal area specifically includes:

[0175] The difference matrix is ​​obtained by subtracting the actual time delay matrix from the theoretical time delay matrix. Elements in the difference matrix whose relevant time delay difference values ​​exceed the first comparison threshold are identified. The cable segment between the two sensors corresponding to each element is marked as an abnormal cable segment. If E abnormal cable segments are spatially consecutive and adjacent, then the E abnormal cable segments are merged into an abnormal region, where E is an integer greater than or equal to 2.

[0176] In this matrix, the rows and columns correspond to the sensor indices, and the elements in the matrix represent the correlation delay difference between the corresponding two sensors.

[0177] It should be noted that calculating the difference matrix by subtracting the actual time delay matrix from the theoretical time delay matrix specifically involves performing element-wise subtraction between the theoretical and actual time delay matrices. This involves subtracting the corresponding time delay value in the theoretical time delay matrix from the time delay value at each position in the actual time delay matrix, thus obtaining the difference matrix. Each element of this difference matrix quantifies the deviation between the measured wave propagation time delay between sensors and the theoretical time delay based on the health model: a positive value indicates that the wave propagation is slower than expected, a negative value indicates that it is faster than expected, and a value near zero indicates good performance.

[0178] Elements in the difference matrix whose relevant time delay difference values ​​do not exceed the first comparison threshold are identified as normal areas for the cable segment between the two sensors corresponding to the element.

[0179] By comparing the difference matrix between the actual measured time delay and the theoretical time delay, the cable segments between sensor pairs with significant deviations in wave propagation time can be quickly identified, and the detected anomalies can be accurately located to specific physical areas of the main cable, providing a clear target location for subsequent inspection, monitoring and maintenance.

[0180] In a specific embodiment, identifying the anomaly type of the abnormal region specifically includes:

[0181] Based on the association weight matrix, calculate the average association strength between each sensor and all other sensors;

[0182] Obtain the sensor set that constitutes the abnormal region;

[0183] If the anomaly score exceeds the second comparison threshold, and the average correlation strength of all sensors in the sensor set is higher than the global threshold, then the anomaly region is identified as a global excitation anomaly.

[0184] If the anomaly score exceeds the second comparison threshold, and the correlation strength between two sensors within the anomaly region in the correlation weight matrix is ​​lower than the third comparison threshold, then identification is performed using the difference matrix:

[0185] When the correlation delay difference values ​​in the difference matrix exhibit a regular anomaly, it is identified as a local structural anomaly.

[0186] When the correlation delay difference values ​​in the difference matrix exhibit irregular discrete anomalies, it is identified as a sensor fault.

[0187] It should be noted that calculating the average correlation strength between each sensor and all other sensors specifically includes:

[0188] Take the off-diagonal elements of the correlation weight matrix and calculate their arithmetic mean. The calculation expression is as follows:

[0189] ;

[0190] in, This represents the total number of sensors. This represents the average correlation strength.

[0191] The first comparison threshold is to collect multiple sets of main cable amplitude data under normal operating conditions, calculate multiple sets of difference matrices using this method, statistically analyze the distribution characteristics of the absolute values ​​of all elements in the multiple sets of difference matrices, and take the 95th percentile of the distribution characteristics.

[0192] The second comparison threshold is to perform a reversible transformation on the amplitude time series under historical normal operating conditions to obtain a set of benchmark anomaly scores, and calculate the mean and standard deviation of the set of benchmark anomaly scores, and the sum of the mean and three times the standard deviation.

[0193] The global threshold is calculated based on the same historical normal data to obtain multiple sets of correlation weight matrices; for each sensor, its average correlation strength in each correlation weight matrix is ​​calculated, and the value is taken as 90% of the overall mean of the average correlation strength of all sensors.

[0194] The third comparison threshold is 60% of the historical average of the average correlation strength of each individual sensor.

[0195] Regular anomalies are those where, in the difference matrix, multiple sensor pairs involved in the anomaly region have the same sign for their time delay difference values, and their absolute values ​​change with a gradient along the main cable direction.

[0196] Irregular discrete anomalies are characterized by significantly abnormal time delay differences in the anomaly region within the difference matrix, while the time delay differences of other normal cable segments connected to the sensor in this anomaly region are all normal.

[0197] By comprehensively utilizing multiple indicators such as correlation weight matrix, difference matrix, and anomaly score, different types of anomaly root causes can be logically distinguished: global excitation anomalies, local structural anomalies, and sensor malfunctions. This provides engineers with preliminary classification information on the nature and possible causes of anomalies, helping to guide more targeted countermeasures and improving the intelligent diagnostic level of the entire monitoring system.

[0198] Example 2 provides an automatic identification system for main cable amplitude anomalies, including a data acquisition module, a processing module, and an identification module:

[0199] The acquisition module is used to acquire vibration signals through sensors, extract peak features and perform Hilbert transform analysis on the vibration signals, and perform principal component analysis to reduce the dimension of the analysis results. The extraction results, analysis results and dimension reduction results are arranged according to time and sensor location to obtain the peak sparse matrix, amplitude time series and time delay feature matrix, respectively.

[0200] The processing module is used to divide subsequences based on amplitude time series, extract features from subsequences using time-delay short-term memory network, calculate association weights based on time-delay feature matrix and peak sparse matrix through attention mechanism, and fuse feature extraction results and association weight results to obtain spatiotemporal feature matrix;

[0201] The identification module is used to obtain anomaly scores based on the spatiotemporal feature matrix by performing an affine transformation on the amplitude time series, obtain the difference matrix based on the analysis results, locate the abnormal regions according to the difference matrix, and identify the anomaly type of the abnormal region through the anomaly scores and correlation weight results.

[0202] This invention ensures that each module is consistent and alignable in terms of matrix dimensions by uniformly constraining and remapping the dimensions of the delay feature matrix, time feature matrix, and spatial feature matrix. This avoids the problem of algorithm infeasibility caused by high-dimensional expansion, time sequence mismatch, or feature broadcasting conflict, thereby ensuring the feasibility and stability of the overall anomaly identification process at the mathematical and engineering levels.

[0203] This invention extracts time-delay features from vibration signals acquired by sensors and, combined with cross-correlation analysis and calculations of theoretical and actual time-delay matrices, accurately captures the time-delay effects and peak sparsity characteristics between different sensors. By integrating time-delay and peak sparsity features into spatiotemporal features, the system can more comprehensively capture dynamic changes in vibration signals, improving the accuracy of anomaly detection. Furthermore, affine transformation maps the amplitude signal to a latent space, and by adjusting transformation parameters based on the spatiotemporal feature matrix, anomaly probability density is calculated to obtain anomaly scores. This enhances the ability to identify complex anomaly patterns, enabling more flexible adaptation to different types of anomalies and providing more accurate anomaly type identification.

[0204] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0205] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for automatic identification of main cable amplitude anomalies, characterized in that, Includes the following steps: Step S1: Collect vibration signals through sensors, extract peak features and perform Hilbert transform analysis on the vibration signals, and perform principal component analysis to reduce the dimensionality of the analysis results. Arrange the extraction results, analysis results and dimensionality reduction results according to time and sensor position to obtain the peak sparse matrix, amplitude time series and time delay feature matrix respectively. Step S2: Divide the subsequences based on the amplitude time series, extract features from the subsequences using a time-delayed short-term memory network, calculate the association weights based on the time-delay feature matrix by combining the peak sparse matrix with the attention mechanism, and fuse the feature extraction results and association weight results to obtain the spatiotemporal feature matrix. The process of obtaining the spatiotemporal feature matrix specifically includes: Based on the time delay feature matrix, the initial correlation weight matrix between sensors is calculated through the attention mechanism. The initial correlation weight matrix is ​​biased by the peak sparse matrix to obtain the correlation weight matrix. The correlation weight matrix is ​​weighted and aggregated with the value matrix to obtain the spatial feature matrix. The expression for bias calculation is: ； For the correlation weight matrix, It is a peak sparse matrix. To adjust the parameters, For querying the matrix, The key matrix, This is the initial correlation weight matrix; The rows and columns of the association weight matrix correspond to the sensor indices, and the elements in the association weight matrix represent the association strength between the corresponding two sensors, with the sum of the elements in each row being 1. The time feature matrix and the spatial feature matrix are weighted and averaged to form the spatiotemporal feature matrix; Step S3: Based on the spatiotemporal feature matrix, anomaly scores are obtained by performing an affine transformation on the amplitude time series. A difference matrix is ​​obtained based on the analysis results. Anomaly regions are located according to the difference matrix. The anomaly types of the anomaly regions are identified by the anomaly scores and correlation weight results.

2. The method for automatic identification of main cable amplitude anomalies as described in claim 1, characterized in that, Step S1 specifically includes: Step S11: Obtain the position information of the sensor and the vibration signal collected by the sensor; perform Hilbert transform on the vibration signal of each sensor to obtain the analytical signal of the vibration signal; calculate the amplitude of the analytical signal to obtain the amplitude time series of each sensor. Step S12: Based on the sensor position information and the theoretical wave velocity of the main cable vibration, calculate the theoretical time delay value between every two sensors to form a theoretical time delay matrix. The rows and columns of the theoretical time delay matrix correspond to the sensor indices, and the elements in the theoretical time delay matrix represent the theoretical time delay values ​​between the corresponding two sensors. Step S13: Based on the amplitude time series, calculate the actual time delay value between each pair of sensors using the cross-correlation analysis method to form the actual time delay matrix; Wherein, the rows and columns of the actual time delay matrix correspond to the sensor indexes, and the elements in the actual time delay matrix represent the actual time delay values ​​between the corresponding two sensors; Step S14: Flatten and concatenate the theoretical time delay matrix and the actual time delay matrix to obtain an initial time delay feature matrix, and reduce the dimensionality of the initial time delay feature matrix by principal component analysis to obtain the time delay feature matrix. Step S15: Perform peak detection on the amplitude time series of each sensor and extract the local maximum value in the vibration signal as the peak feature; Step S16: Based on the time delay sampling between every two sensors, extract the corresponding amplitude peak value and construct a peak sparse matrix. In the peak sparse matrix, the corresponding element of the amplitude peak pair between the sensors is non-zero, otherwise it is zero.

3. The automatic identification method for main cable amplitude anomalies as described in claim 2, characterized in that, The specific steps for feature extraction from the subsequence using a time-extended short-term memory network include: The amplitude time series of each sensor is divided into k subsequences according to the time window. The time delay feature vector of each sensor in the time delay feature matrix is ​​obtained. The corresponding time offset is applied to the subsequence of each sensor to obtain the time-aligned subsequence of each sensor. k is an integer greater than 2. After time-series alignment, the subsequence of each sensor and the corresponding time delay feature vector of the sensor are input into the time-delayed short-term memory network. The final hidden state output by the time-delayed short-term memory network is obtained. The final hidden states of all sensors are concatenated to obtain the time feature matrix.

4. The automatic identification method for main cable amplitude anomalies as described in claim 3, characterized in that, The time-extended short-term memory network specifically includes: The hidden state of the previous time step, the input subsequence of the current time step, and the time delay feature vector corresponding to each sensor are concatenated to form a combined input vector; Based on the combined input vector, the input gate signal, forget gate signal, output gate signal, and candidate cell state are computed in parallel through learnable linear transformations and activation functions. The input gate signal, forget gate signal, and output gate signal are calculated using the Sigmoid activation function, and the candidate cell state is calculated using the hyperbolic tangent activation function. The candidate cell state is calculated using the forget gate signal and the input gate signal to obtain the cell state updated at the current time step. The calculation expression is as follows: ； in, This represents the cell state after the current time step t and sensor j update. This represents the forget gate signal for the current time step t and sensor j. This represents the cell state at the previous time step t-1 and sensor j, and ⊙ represents element-wise multiplication. This represents the input gate signal for the current time step t and sensor j. This represents the candidate cell state at the current time step t and sensor j; The current cell state is adjusted using the output gate signal, and the hidden state at the current time step is output. The calculation expression is as follows: ； in, This represents the hidden state output by sensor j at the current time step t. This represents the output gate signal of the current time step t and sensor j. Let be the hyperbolic tangent function, and ⊙ denotes element-wise multiplication.

5. The automatic identification method for main cable amplitude anomalies as described in claim 4, characterized in that, Calculating the outlier fraction of the amplitude time series specifically includes: The spatiotemporal feature matrix is ​​input into the feedforward neural network, which outputs transformation parameters. Based on the transformation parameters, the vibration time series is subjected to P iterations of affine transformation to obtain the anomaly density function and the cumulative Jacobian determinant. The cumulative Jacobian determinant is the product of the Jacobian determinants of the P iterations. Based on the probability density transformation formula, the anomaly probability density of the amplitude time series is calculated, and the expression is as follows: ； in, For anomaly density function, For cumulative Jacobian determinants, The spatiotemporal characteristic matrix, It is an amplitude time series. It is an affine transformation function; The negative logarithm of the anomaly probability density is used as the anomaly score.

6. The method for automatic identification of main cable amplitude anomalies as described in claim 5, characterized in that, Affine transformation operations specifically include: The input vector is randomly divided into a first part and a second part. The first part and the transformation parameters are input into the transformation function, which outputs the scaling and translation parameters. The second part undergoes an affine transformation, and the calculation expression is as follows: ； ； in, The spatiotemporal characteristic matrix, and The input vector is divided into two sub-vectors. For scaling function, It is a translation function. and There are two output sub-vectors; The two output sub-vectors are recombine to obtain the output vector; The input vector includes: The input vector for the first iteration is the vector obtained after flattening the vibration time series data; Except for the first iteration, the input vector is the output vector of the affine transformation in the previous iteration; The mathematical expression for the Jacobian determinant is: ； in, scale function The i-th component of the output vector.

7. The automatic identification method for main cable amplitude anomalies as described in claim 6, characterized in that, The specific areas for locating anomalies include: The difference matrix is ​​obtained by subtracting the actual time delay matrix from the theoretical time delay matrix. Elements in the difference matrix whose related time delay difference values ​​exceed the first comparison threshold are identified. The cable segment between the two sensors corresponding to each element is marked as an abnormal cable segment. If E abnormal cable segments are spatially consecutive, the E abnormal cable segments are merged into an abnormal region, where E is an integer greater than or equal to 2. The rows and columns of the difference matrix correspond to sensor indices, and the elements in the difference matrix represent the correlation delay difference values ​​between two corresponding sensors.

8. The method for automatic identification of main cable amplitude anomalies as described in claim 7, characterized in that, Identifying the anomaly type in anomaly regions specifically includes: Based on the association weight matrix, calculate the average association strength between each sensor and all other sensors; Obtain the sensor set that constitutes the abnormal region; If the anomaly score exceeds the second comparison threshold, and the average correlation strength of all sensors in the sensor set is higher than the global threshold, then the anomaly region is identified as a global excitation anomaly. If the anomaly score exceeds the second comparison threshold, and the correlation strength between two sensors within an anomaly region in the correlation weight matrix is ​​lower than the third comparison threshold, then identification is performed using the difference matrix: When the relevant time delay difference values ​​in the difference matrix exhibit a regular anomaly, it is identified as a local structural anomaly. When the correlation delay difference values ​​in the difference matrix exhibit irregular discrete anomalies, it is identified as a sensor fault.

9. An automatic identification system for abnormal amplitude of main cable, characterized in that, It includes a data acquisition module, a processing module, and a recognition module: The acquisition module is used to acquire vibration signals through sensors, extract peak features and perform Hilbert transform analysis on the vibration signals, and perform principal component analysis to reduce the dimensionality of the analysis results. The extraction results, analysis results and dimensionality reduction results are arranged according to time and sensor position to obtain the peak sparse matrix, amplitude time series and time delay feature matrix, respectively. The processing module is used to divide the subsequence based on the amplitude time series, extract features from the subsequence using a time-delayed short-term memory network, calculate the association weight based on the time-delay feature matrix by combining the peak sparse matrix with the attention mechanism, and fuse the feature extraction results and association weight results to obtain the spatiotemporal feature matrix. The process of obtaining the spatiotemporal feature matrix specifically includes: Based on the time delay feature matrix, the initial correlation weight matrix between sensors is calculated through the attention mechanism. The initial correlation weight matrix is ​​biased by the peak sparse matrix to obtain the correlation weight matrix. The correlation weight matrix is ​​weighted and aggregated with the value matrix to obtain the spatial feature matrix. The expression for bias calculation is: ； For the correlation weight matrix, It is a peak sparse matrix. To adjust the parameters, For querying the matrix, The key matrix, This is the initial correlation weight matrix; The rows and columns of the association weight matrix correspond to the sensor indices, and the elements in the association weight matrix represent the association strength between the corresponding two sensors, with the sum of the elements in each row being 1. The time feature matrix and the spatial feature matrix are weighted and averaged to form the spatiotemporal feature matrix; The identification module is used to obtain anomaly scores by performing an affine transformation on the amplitude time series based on the spatiotemporal feature matrix, obtain a difference matrix based on the analysis results, locate abnormal regions according to the difference matrix, and identify the anomaly type of the abnormal region through the anomaly scores and correlation weight results.

Images